[AMDGPU] Use CAS loop for min/max atomics at system scope

[doru1004]

Summary

Use CAS loop for min/max atomics at system scope. Currently a hardware atomic instruction is being used which does not work at system scope.

[arsenm]

Most of these test changes break the purpose of the tests, which was proper legalization handling of the atomic instructions

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12896–12897

Should use AMDGPU::isFlatGlobalAddrSpace, in principal the same should apply to any other random address space number

12903–12906

move to predicate function to call at the use sites?

llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll

184

This breaks the purpose of this test. These should be updated to have a scope that doesn't require the cas loop

[doru1004]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12896–12897

This check here is just a factoring out of the switch statement since now this check is done in 2 places. Are you implying that this check was wrong or insufficient before?

12903–12906

Could you please give me more details regarding this request? Again this is a check which was factored out of the switch statement because the same check is now done in two different locations.

llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll

184

I agree, this is a good point and I'll try to fix it.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12896–12897

Yes

12903–12906

bool unsafeAtomicsDisabled(const Function &F) {

return ....

}

and call that at the use sites rather than eagerly parsing the attribute

[arsenm]

I think you're best off replacing all changed tests with a specific scope, and introducing new tests specifically for system scoped atomics

[doru1004]

@arsenm I have updated the tests and made one of the two code changes you suggested. I have to do the function still.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12989

I'm pretty sure all of the atomics (except maybe add?) have the same issue, but they don't all need to change at once

[doru1004]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12989

I have the same feeling.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12989

Part of why I think we need D137361 is to get proper lowering for inc/dec for this

[doru1004]

I just updated the patch with the use of the function.

[doru1004]

@arsenm please let me know if you have any further comments.

[arsenm]

LGTM

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12874

static and const&

[doru1004]

Commit: aea5980e26e6a87dab9f8acb10eb3a59dd143cb1

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

if (&Flt == &APFloat::IEEEsingle())
     return DenormMode == DenormalMode::getPreserveSign();
   return DenormMode == DenormalMode::getIEEE();
 }
 
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   unsigned AS = RMW->getPointerAddressSpace();
   if (AS == AMDGPUAS::PRIVATE_ADDRESS)
     return AtomicExpansionKind::NotAtomic;
 
+  auto SSID = RMW->getSyncScopeID();
+
   auto ReportUnsafeHWInst = [&](TargetLowering::AtomicExpansionKind Kind) {
     OptimizationRemarkEmitter ORE(RMW->getFunction());
     LLVMContext &Ctx = RMW->getFunction()->getContext();
     SmallVector<StringRef> SSNs;
     Ctx.getSyncScopeNames(SSNs);
     auto MemScope = SSNs[RMW->getSyncScopeID()].empty()
                         ? "system"
                         : SSNs[RMW->getSyncScopeID()];
     ORE.emit([&]() {
       return OptimizationRemark(DEBUG_TYPE, "Passed", RMW)
              << "Hardware instruction generated for atomic "
              << RMW->getOperationName(RMW->getOperation())
              << " operation at memory scope " << MemScope
              << " due to an unsafe request.";
     });
     return Kind;
   };
 
+  bool HasGlobalOrFlatAS =
+      AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS;
+
+  // The amdgpu-unsafe-fp-atomics attribute enables generation of unsafe
+  // floating point atomic instructions. May generate more efficient code,
+  // but may not respect rounding and denormal modes, and may give incorrect
+  // results for certain memory destinations.
+  bool UnsafeFPAtomicsDisabled =
+      RMW->getFunction()
+          ->getFnAttribute("amdgpu-unsafe-fp-atomics")
+          .getValueAsString() != "true";
+
+  bool HasSystemScope =
+      SSID == SyncScope::System ||
+      SSID == RMW->getContext().getOrInsertSyncScopeID("one-as");
+
   switch (RMW->getOperation()) {
   case AtomicRMWInst::FAdd: {
     Type *Ty = RMW->getType();
 
     if (Ty->isHalfTy())
       return AtomicExpansionKind::CmpXChg;
 
     if (!Ty->isFloatTy() && (!Subtarget->hasGFX90AInsts() || !Ty->isDoubleTy()))
       return AtomicExpansionKind::CmpXChg;
 
-    if ((AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) &&
-        Subtarget->hasAtomicFaddNoRtnInsts()) {
-      // The amdgpu-unsafe-fp-atomics attribute enables generation of unsafe
-      // floating point atomic instructions. May generate more efficient code,
-      // but may not respect rounding and denormal modes, and may give incorrect
-      // results for certain memory destinations.
-      if (RMW->getFunction()
-              ->getFnAttribute("amdgpu-unsafe-fp-atomics")
-              .getValueAsString() != "true")
+    if (HasGlobalOrFlatAS && Subtarget->hasAtomicFaddNoRtnInsts()) {
+      if (UnsafeFPAtomicsDisabled)
         return AtomicExpansionKind::CmpXChg;
 
       // Always expand system scope fp atomics.
-      auto SSID = RMW->getSyncScopeID();
-      if (SSID == SyncScope::System ||
-          SSID == RMW->getContext().getOrInsertSyncScopeID("one-as"))
+      if (HasSystemScope)
         return AtomicExpansionKind::CmpXChg;
 
       if (AS == AMDGPUAS::GLOBAL_ADDRESS && Ty->isFloatTy()) {
         // global atomic fadd f32 no-rtn: gfx908, gfx90a, gfx940, gfx11+.
         if (RMW->use_empty() && Subtarget->hasAtomicFaddNoRtnInsts())
           return ReportUnsafeHWInst(AtomicExpansionKind::None);
         // global atomic fadd f32 rtn: gfx90a, gfx940, gfx11+.
         if (!RMW->use_empty() && Subtarget->hasAtomicFaddRtnInsts())
if (AS == AMDGPUAS::LOCAL_ADDRESS && Subtarget->hasLDSFPAtomicAdd()) {
                          ->getFnAttribute("amdgpu-unsafe-fp-atomics")
                          .getValueAsString() == "true"
                  ? ReportUnsafeHWInst(AtomicExpansionKind::None)
                  : AtomicExpansionKind::CmpXChg;
     }
 
     return AtomicExpansionKind::CmpXChg;
   }
+  case AtomicRMWInst::FMin:
+  case AtomicRMWInst::FMax:
+  case AtomicRMWInst::Min:
+  case AtomicRMWInst::Max:
+  case AtomicRMWInst::UMin:
+  case AtomicRMWInst::UMax: {
+    if (HasGlobalOrFlatAS) {
+      if (RMW->getType()->isFloatTy() && UnsafeFPAtomicsDisabled)
+        return AtomicExpansionKind::CmpXChg;
+
+      // Always expand system scope min/max atomics.
+      if (HasSystemScope)
+        return AtomicExpansionKind::CmpXChg;
+    }
+    break;
+  }
   default:
     break;
   }
 
   return AMDGPUTargetLowering::shouldExpandAtomicRMWInIR(RMW);
 }
 
 TargetLowering::AtomicExpansionKind

llvm/test/CodeGen/AMDGPU/dag-divergence-atomic.ll

; CHECK-NEXT:    s_endpgm
   store float 1.0, float addrspace(1)* %p1
   ret void
 }
 
 define protected amdgpu_kernel void @max(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
 ; CHECK-LABEL: max:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; CHECK-NEXT:    v_mov_b32_e32 v0, 0
-; CHECK-NEXT:    v_mov_b32_e32 v1, 1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    global_atomic_smax v2, v0, v1, s[0:1] glc
-; CHECK-NEXT:    v_mov_b32_e32 v0, s2
-; CHECK-NEXT:    v_mov_b32_e32 v1, s3
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB6_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_max_i32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, 12, v[0:1]
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB6_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw max i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
   store float 1.0, float addrspace(1)* %p1
   ret void
 }
 
 define protected amdgpu_kernel void @min(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
 ; CHECK-LABEL: min:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; CHECK-NEXT:    v_mov_b32_e32 v0, 0
-; CHECK-NEXT:    v_mov_b32_e32 v1, 1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    global_atomic_smin v2, v0, v1, s[0:1] glc
-; CHECK-NEXT:    v_mov_b32_e32 v0, s2
-; CHECK-NEXT:    v_mov_b32_e32 v1, s3
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB7_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_min_i32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, 12, v[0:1]
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB7_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw min i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
   store float 1.0, float addrspace(1)* %p1
   ret void
 }
 
 define protected amdgpu_kernel void @umax(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
 ; CHECK-LABEL: umax:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; CHECK-NEXT:    v_mov_b32_e32 v0, 0
-; CHECK-NEXT:    v_mov_b32_e32 v1, 1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    global_atomic_umax v2, v0, v1, s[0:1] glc
-; CHECK-NEXT:    v_mov_b32_e32 v0, s2
-; CHECK-NEXT:    v_mov_b32_e32 v1, s3
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB8_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_max_u32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, 12, v[0:1]
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB8_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw umax i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
   store float 1.0, float addrspace(1)* %p1
   ret void
 }
 
 define protected amdgpu_kernel void @umin(i32 addrspace(1)* %p, %S addrspace(1)* %q) {
 ; CHECK-LABEL: umin:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; CHECK-NEXT:    v_mov_b32_e32 v0, 0
-; CHECK-NEXT:    v_mov_b32_e32 v1, 1
+; CHECK-NEXT:    s_mov_b64 s[4:5], 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    global_atomic_umin v2, v0, v1, s[0:1] glc
-; CHECK-NEXT:    v_mov_b32_e32 v0, s2
-; CHECK-NEXT:    v_mov_b32_e32 v1, s3
+; CHECK-NEXT:    s_load_dword s6, s[0:1], 0x0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    v_mov_b32_e32 v0, s6
+; CHECK-NEXT:  .LBB9_1: ; %atomicrmw.start
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    v_mov_b32_e32 v3, v0
+; CHECK-NEXT:    v_min_u32_e32 v2, 1, v3
+; CHECK-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; CHECK-NEXT:    s_waitcnt vmcnt(0)
-; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v2, 12, v[0:1]
+; CHECK-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; CHECK-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CHECK-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    s_cbranch_execnz .LBB9_1
+; CHECK-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    v_mov_b32_e32 v3, s3
+; CHECK-NEXT:    v_mad_u64_u32 v[0:1], s[0:1], v0, 12, v[2:3]
 ; CHECK-NEXT:    v_mov_b32_e32 v2, 1.0
 ; CHECK-NEXT:    global_store_dword v[0:1], v2, off
 ; CHECK-NEXT:    s_endpgm
   %n32 = atomicrmw umin i32 addrspace(1)* %p, i32 1 monotonic
   %n64 = zext i32 %n32 to i64
   %p1 = getelementptr inbounds %S, %S addrspace(1)* %q, i64 %n64, i32 0
   store float 1.0, float addrspace(1)* %p1
   ret void

llvm/test/CodeGen/AMDGPU/flat_atomics.ll

 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s2, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s2, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_max_i32_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s4, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s4, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_ret_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_i32_e32 v0, s2, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i32_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i32_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_ret_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_i32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile max i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_max_i32:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_max_i32_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_max_i32_e32 v0, s2, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_max_i32_e32 v0, s2, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_ret:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile max i32* %out, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_addr64(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i32_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB32_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN2-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB32_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB32_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i32_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_i32_e32 v0, s8, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB33_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i32_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_i32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB33_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_max_i32_ret_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_i32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB33_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile max i32* %ptr, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_offset(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_umax_i32_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s2, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB34_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s2, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB34_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB34_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_umax_i32_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s4, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB35_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s4, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB35_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_ret_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_u32_e32 v0, s2, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB35_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i32_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB36_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB36_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB36_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB37_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB37_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_u32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB37_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile umax i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_umax_i32:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB38_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB38_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB38_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_umax_i32_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_max_u32_e32 v0, s2, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB39_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_max_u32_e32 v0, s2, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB39_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_ret:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_max_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB39_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umax i32* %out, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_addr64(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i32_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB40_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN2-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB40_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_max_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB40_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i32_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_max_u32_e32 v0, s8, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB41_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i32_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_max_u32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB41_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umax_i32_ret_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_max_u32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB41_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile umax i32* %ptr, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_offset(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_min_i32_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s2, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB42_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s2, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB42_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB42_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_min_i32_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s4, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB43_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s4, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB43_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_ret_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_i32_e32 v0, s2, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB43_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i32_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB44_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB44_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB44_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i32_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB45_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB45_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_ret_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_i32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB45_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile min i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_min_i32:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB46_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB46_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB46_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_min_i32_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_min_i32_e32 v0, s2, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB47_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_min_i32_e32 v0, s2, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB47_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_ret:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_i32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB47_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile min i32* %out, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_addr64(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i32_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB48_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN2-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB48_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_i32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB48_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i32_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_i32_e32 v0, s8, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB49_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i32_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_i32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB49_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_min_i32_ret_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_i32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB49_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile min i32* %ptr, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_offset(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_umin_i32_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s2, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB50_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s2, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s3, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB50_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB50_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_umin_i32_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s4, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB51_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s4, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s5, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB51_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_ret_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_u32_e32 v0, s2, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB51_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32* %out, i32 4
   %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i32_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB52_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB52_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB52_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 16
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB53_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 16
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB53_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1] offset:16
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_u32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] offset:16 glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB53_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %gep = getelementptr i32, i32* %ptr, i32 4
   %val = atomicrmw volatile umin i32* %gep, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32(i32* %out, i32 %in) {
 ; GCN1-LABEL: atomic_umin_i32:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB54_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB54_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s3
-; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB54_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret(i32* %out, i32* %out2, i32 %in) {
 ; GCN1-LABEL: atomic_umin_i32_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s0, s[0:1], 0xd
+; GCN1-NEXT:    s_load_dword s2, s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v2, s4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s5
+; GCN1-NEXT:    v_min_u32_e32 v0, s2, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_execnz .LBB55_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s0, s[0:1], 0x34
+; GCN2-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v2, s4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s5
+; GCN2-NEXT:    v_min_u32_e32 v0, s2, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_execnz .LBB55_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_ret:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GCN3-NEXT:    s_mov_b64 s[0:1], 0
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s5
-; GCN3-NEXT:    v_mov_b32_e32 v2, s2
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v2, s4
+; GCN3-NEXT:    v_mov_b32_e32 v3, s5
+; GCN3-NEXT:    v_min_u32_e32 v0, s2, v1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    s_cbranch_execnz .LBB55_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umin i32* %out, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_addr64(i32* %out, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i32_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dword s6, s[0:1], 0xb
+; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dword s4, s[0:1], 0xb
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_add_u32 s0, s6, s0
+; GCN1-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    flat_load_dword v1, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB56_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_add_u32 s0, s6, s0
+; GCN2-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    flat_load_dword v1, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN2-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB56_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GCN3-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GCN3-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GCN3-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GCN3-NEXT:    s_add_u32 s0, s4, s0
-; GCN3-NEXT:    s_addc_u32 s1, s5, s1
+; GCN3-NEXT:    s_add_u32 s0, s6, s0
+; GCN3-NEXT:    s_addc_u32 s1, s7, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s6
+; GCN3-NEXT:    flat_load_dword v1, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_min_u32_e32 v0, s4, v1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB56_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32* %out, i32* %out2, i32 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i32_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN1-NEXT:    s_add_u32 s0, s4, s0
 ; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s8
+; GCN1-NEXT:    flat_load_dword v0, v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
+; GCN1-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v1, v0
+; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_min_u32_e32 v0, s8, v1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN1-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    flat_store_dword v[0:1], v2
+; GCN1-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB57_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v1, s6
+; GCN1-NEXT:    v_mov_b32_e32 v2, s7
+; GCN1-NEXT:    flat_store_dword v[1:2], v0
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i32_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN2-NEXT:    s_add_u32 s0, s4, s0
 ; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s8
+; GCN2-NEXT:    flat_load_dword v0, v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
+; GCN2-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v1, v0
+; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_min_u32_e32 v0, s8, v1
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN2-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    flat_store_dword v[0:1], v2
+; GCN2-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB57_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v1, s6
+; GCN2-NEXT:    v_mov_b32_e32 v2, s7
+; GCN2-NEXT:    flat_store_dword v[1:2], v0
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GCN3-LABEL: atomic_umin_i32_ret_addr64:
 ; GCN3:       ; %bb.0: ; %entry
 ; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GCN3-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN3-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN3-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GCN3-NEXT:    s_add_u32 s0, s4, s0
 ; GCN3-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN3-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN3-NEXT:    v_mov_b32_e32 v1, s1
-; GCN3-NEXT:    v_mov_b32_e32 v2, s8
+; GCN3-NEXT:    flat_load_dword v0, v[0:1]
+; GCN3-NEXT:    s_mov_b64 s[2:3], 0
+; GCN3-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GCN3-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN3-NEXT:    v_mov_b32_e32 v1, v0
+; GCN3-NEXT:    v_mov_b32_e32 v3, s1
+; GCN3-NEXT:    v_mov_b32_e32 v2, s0
+; GCN3-NEXT:    v_min_u32_e32 v0, s8, v1
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
+; GCN3-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
 ; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN3-NEXT:    buffer_wbinvl1_vol
-; GCN3-NEXT:    v_mov_b32_e32 v0, s6
-; GCN3-NEXT:    v_mov_b32_e32 v1, s7
-; GCN3-NEXT:    flat_store_dword v[0:1], v2
+; GCN3-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GCN3-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN3-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    s_cbranch_execnz .LBB57_1
+; GCN3-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN3-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GCN3-NEXT:    v_mov_b32_e32 v1, s6
+; GCN3-NEXT:    v_mov_b32_e32 v2, s7
+; GCN3-NEXT:    flat_store_dword v[1:2], v0
 ; GCN3-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32* %out, i64 %index
   %val = atomicrmw volatile umin i32* %ptr, i32 %in seq_cst
   store i32 %val, i32* %out2
   ret void
 }
 

llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll

entry:
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_offset(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_max_i64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB24_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_max_i64_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    s_add_u32 s0, s0, 32
-; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    s_add_u32 s0, s0, 32
-; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB25_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i64_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB26_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i64_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB27_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile max i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_max_i64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB28_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_max_i64_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB29_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile max i64* %out, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_addr64(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i64_addr64:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_addr64:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB30_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_max_i64_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_max_i64_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB31_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile max i64* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_offset(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_umax_i64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB32_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB32_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_umax_i64_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    s_add_u32 s0, s0, 32
-; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB33_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    s_add_u32 s0, s0, 32
-; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB33_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i64_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB34_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB34_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i64_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB35_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB35_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile umax i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_umax_i64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB36_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB36_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_umax_i64_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB37_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB37_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umax i64* %out, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_addr64(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i64_addr64:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB38_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_addr64:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB38_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umax_i64_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB39_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umax_i64_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB39_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile umax i64* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_offset(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_min_i64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB40_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB40_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_min_i64_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    s_add_u32 s0, s0, 32
-; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB41_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    s_add_u32 s0, s0, 32
-; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB41_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i64_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB42_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB42_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i64_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB43_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB43_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile min i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_min_i64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB44_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB44_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_min_i64_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB45_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB45_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile min i64* %out, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_addr64(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i64_addr64:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB46_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_addr64:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB46_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_min_i64_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB47_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_min_i64_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB47_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile min i64* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_offset(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_umin_i64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB48_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB48_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_umin_i64_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
-; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    s_add_u32 s0, s0, 32
-; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_u32 s2, s4, 32
+; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s2
+; GCN1-NEXT:    v_mov_b32_e32 v1, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s3
+; GCN1-NEXT:    v_mov_b32_e32 v4, s2
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB49_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
-; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    s_add_u32 s0, s0, 32
-; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_u32 s2, s4, 32
+; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s2
+; GCN2-NEXT:    v_mov_b32_e32 v1, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v5, s3
+; GCN2-NEXT:    v_mov_b32_e32 v4, s2
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB49_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64* %out, i64 4
   %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i64_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB50_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB50_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i64_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB51_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB51_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %gep = getelementptr i64, i64* %ptr, i64 4
   %tmp0 = atomicrmw volatile umin i64* %gep, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64(i64* %out, i64 %in) {
 ; GCN1-LABEL: atomic_umin_i64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB52_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB52_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret(i64* %out, i64* %out2, i64 %in) {
 ; GCN1-LABEL: atomic_umin_i64_ret:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s1
+; GCN1-NEXT:    v_mov_b32_e32 v6, s0
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_execnz .LBB53_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_ret:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_mov_b64 s[2:3], 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s4
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s1
+; GCN2-NEXT:    v_mov_b32_e32 v6, s0
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[0:1], v[2:3] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_execnz .LBB53_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umin i64* %out, i64 %in seq_cst
   store i64 %tmp0, i64* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_addr64(i64* %out, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i64_addr64:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s6
-; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN1-NEXT:    s_add_u32 s0, s4, s0
-; GCN1-NEXT:    s_addc_u32 s1, s5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v1, s7
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s4
+; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[4:5], 0
+; GCN1-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s3
+; GCN1-NEXT:    v_mov_b32_e32 v6, s2
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_execnz .LBB54_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_addr64:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s6
-; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; GCN2-NEXT:    s_add_u32 s0, s4, s0
-; GCN2-NEXT:    s_addc_u32 s1, s5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v1, s7
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s4
+; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[2:3], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[4:5], 0
+; GCN2-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s3
+; GCN2-NEXT:    v_mov_b32_e32 v6, s2
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_execnz .LBB54_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64* %out, i64* %out2, i64 %in, i64 %index) {
 ; GCN1-LABEL: atomic_umin_i64_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
 ; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v3, s1
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN1-NEXT:    s_add_u32 s0, s0, s6
+; GCN1-NEXT:    s_addc_u32 s1, s1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s0
+; GCN1-NEXT:    v_mov_b32_e32 v1, s1
+; GCN1-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN1-NEXT:    s_mov_b64 s[6:7], 0
+; GCN1-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GCN1-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN1-NEXT:    v_mov_b32_e32 v3, v1
+; GCN1-NEXT:    v_mov_b32_e32 v2, v0
+; GCN1-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN1-NEXT:    v_mov_b32_e32 v0, s5
+; GCN1-NEXT:    v_mov_b32_e32 v6, s4
+; GCN1-NEXT:    v_mov_b32_e32 v5, s1
+; GCN1-NEXT:    v_mov_b32_e32 v4, s0
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN1-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN1-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN1-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN1-NEXT:    s_cbranch_execnz .LBB55_1
+; GCN1-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN1-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_umin_i64_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
 ; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v3, s1
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GCN2-NEXT:    s_add_u32 s0, s0, s6
+; GCN2-NEXT:    s_addc_u32 s1, s1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s0
+; GCN2-NEXT:    v_mov_b32_e32 v1, s1
+; GCN2-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GCN2-NEXT:    s_mov_b64 s[6:7], 0
+; GCN2-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GCN2-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN2-NEXT:    v_mov_b32_e32 v3, v1
+; GCN2-NEXT:    v_mov_b32_e32 v2, v0
+; GCN2-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; GCN2-NEXT:    v_mov_b32_e32 v0, s5
+; GCN2-NEXT:    v_mov_b32_e32 v6, s4
+; GCN2-NEXT:    v_mov_b32_e32 v5, s1
+; GCN2-NEXT:    v_mov_b32_e32 v4, s0
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v6, v2, vcc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN2-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
+; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GCN2-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GCN2-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GCN2-NEXT:    s_cbranch_execnz .LBB55_1
+; GCN2-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GCN2-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64* %out, i64 %index
   %tmp0 = atomicrmw volatile umin i64* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64* %out2

llvm/test/CodeGen/AMDGPU/global-saddr-atomics.ll

 
 ; --------------------------------------------------------------------------------
 ; atomicrmw max
 ; --------------------------------------------------------------------------------
 
 define amdgpu_ps float @global_max_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB54_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i32_rtn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB54_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i32_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB54_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %rtn = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB55_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i32_rtn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB55_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i32_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] offset:-128 glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB55_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %rtn = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps void @global_max_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB56_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i32_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB56_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i32_nortn:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB56_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %unused = atomicrmw max i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_max_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB57_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i32_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB57_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i32_nortn_neg128:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_i32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB57_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %unused = atomicrmw max i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB58_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB58_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i64_rtn:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB58_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB58_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i64_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB58_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB58_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %rtn = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB59_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB59_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_max_saddr_i64_rtn_neg128:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB59_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB59_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_max_saddr_i64_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB59_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB59_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %rtn = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps void @global_max_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB60_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB60_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i64_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB60_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB60_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i64_nortn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB60_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB60_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %unused = atomicrmw max i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_max_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB61_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB61_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_max_saddr_i64_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB61_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB61_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_max_saddr_i64_nortn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB61_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB61_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %unused = atomicrmw max i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   ret void
 }
 
 ; --------------------------------------------------------------------------------
 ; atomicrmw min
 ; --------------------------------------------------------------------------------
 
 define amdgpu_ps float @global_min_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB62_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB62_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i32_rtn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB62_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB62_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i32_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB62_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB62_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %rtn = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB63_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB63_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i32_rtn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB63_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB63_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i32_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] offset:-128 glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB63_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB63_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %rtn = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps void @global_min_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB64_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB64_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i32_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB64_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB64_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i32_nortn:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB64_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB64_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %unused = atomicrmw min i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_min_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB65_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_i32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB65_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i32_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB65_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB65_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i32_nortn_neg128:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB65_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_i32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB65_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %unused = atomicrmw min i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB66_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB66_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i64_rtn:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB66_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB66_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i64_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB66_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB66_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %rtn = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB67_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB67_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_min_saddr_i64_rtn_neg128:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB67_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB67_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_min_saddr_i64_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB67_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB67_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %rtn = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps void @global_min_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB68_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB68_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i64_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB68_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB68_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i64_nortn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB68_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB68_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %unused = atomicrmw min i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_min_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB69_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB69_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_min_saddr_i64_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB69_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB69_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_min_saddr_i64_nortn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB69_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_i64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB69_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %unused = atomicrmw min i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   ret void
 }
 
 ; --------------------------------------------------------------------------------
 ; atomicrmw umax
 ; --------------------------------------------------------------------------------
 
 define amdgpu_ps float @global_umax_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB70_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB70_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_rtn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB70_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB70_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB70_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB70_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB71_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB71_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_rtn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB71_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB71_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] offset:-128 glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB71_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB71_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %rtn = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps void @global_umax_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB72_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB72_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB72_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB72_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_nortn:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB72_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB72_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %unused = atomicrmw umax i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umax_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB73_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB73_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i32_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB73_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB73_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i32_nortn_neg128:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB73_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_max_u32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB73_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %unused = atomicrmw umax i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB74_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB74_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_rtn:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB74_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB74_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB74_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB74_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB75_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB75_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_rtn_neg128:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB75_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB75_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB75_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB75_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %rtn = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps void @global_umax_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB76_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB76_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB76_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB76_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_nortn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB76_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB76_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %unused = atomicrmw umax i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umax_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB77_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB77_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umax_saddr_i64_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB77_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB77_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umax_saddr_i64_nortn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB77_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_gt_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB77_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %unused = atomicrmw umax i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   ret void
 }
 
 ; --------------------------------------------------------------------------------
 ; atomicrmw umin
 ; --------------------------------------------------------------------------------
 
 define amdgpu_ps float @global_umin_saddr_i32_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB78_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB78_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_rtn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB78_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB78_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB78_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB78_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v2
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB79_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB79_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_rtn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    global_load_dword v0, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB79_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB79_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] offset:-128 glc
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    global_load_b32 v0, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB79_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB79_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %rtn = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   %cast.rtn = bitcast i32 %rtn to float
   ret float %cast.rtn
 }
 
 define amdgpu_ps void @global_umin_saddr_i32_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB80_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB80_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB80_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB80_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_nortn:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB80_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB80_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i32 addrspace(1)*
   %unused = atomicrmw umin i32 addrspace(1)* %cast.gep0, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i32 %data) {
 ; GFX9-LABEL: global_umin_saddr_i32_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB81_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_u32_e32 v4, v5, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB81_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i32_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dword v5, v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB81_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB81_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i32_nortn_neg128:
 ; GFX11:       ; %bb.0:
+; GFX11-NEXT:    global_load_b32 v5, v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v2, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB81_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_min_u32_e32 v4, v5, v1
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], v[4:5], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB81_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i32 addrspace(1)*
   %unused = atomicrmw umin i32 addrspace(1)* %cast.gep1, i32 %data seq_cst
   ret void
 }
 
 define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_rtn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB82_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB82_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_rtn:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB82_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB82_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_rtn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB82_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB82_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_rtn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v6, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB83_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v10, v4
+; GFX9-NEXT:    v_mov_b32_e32 v9, v3
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB83_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v0, v3
+; GFX9-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_rtn_neg128:
 ; GFX10:       ; %bb.0:
-; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX10-NEXT:    global_load_dwordx2 v[3:4], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v6, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB83_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v10, v4
+; GFX10-NEXT:    v_mov_b32_e32 v9, v3
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB83_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX10-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    v_mov_b32_e32 v0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_rtn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
+; GFX11-NEXT:    global_load_b64 v[3:4], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v5, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v6, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB83_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v10, v4
+; GFX11-NEXT:    v_mov_b32_e32 v9, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[9:10], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v2, v10, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v1, v9, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[5:6], v[7:10], off offset:-128 glc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[9:10]
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB83_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX11-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    v_mov_b32_e32 v0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v4
 ; GFX11-NEXT:    ; return to shader part epilog
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %rtn = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   %cast.rtn = bitcast i64 %rtn to <2 x float>
   ret <2 x float> %cast.rtn
 }
 
 define amdgpu_ps void @global_umin_saddr_i64_nortn(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_nortn:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB84_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB84_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_nortn:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3]
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB84_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB84_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_nortn:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3]
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB84_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB84_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %cast.gep0 = bitcast i8 addrspace(1)* %gep0 to i64 addrspace(1)*
   %unused = atomicrmw umin i64 addrspace(1)* %cast.gep0, i64 %data seq_cst
   ret void
 }
 
 define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(i8 addrspace(1)* inreg %sbase, i32 %voffset, i64 %data) {
 ; GFX9-LABEL: global_umin_saddr_i64_nortn_neg128:
 ; GFX9:       ; %bb.0:
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX9-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, s2, v0
+; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v3, vcc
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:  .LBB85_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v6, v4
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v5, v3
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB85_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_umin_saddr_i64_nortn_neg128:
 ; GFX10:       ; %bb.0:
+; GFX10-NEXT:    global_load_dwordx2 v[5:6], v0, s[2:3] offset:-128
+; GFX10-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v8, s[0:1], s3, 0, s[0:1]
+; GFX10-NEXT:    s_mov_b64 s[0:1], 0
+; GFX10-NEXT:  .LBB85_1: ; %atomicrmw.start
+; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
-; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    global_atomic_cmpswap_x2 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    buffer_gl0_inv
 ; GFX10-NEXT:    buffer_gl1_inv
+; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX10-NEXT:    v_mov_b32_e32 v6, v4
+; GFX10-NEXT:    v_mov_b32_e32 v5, v3
+; GFX10-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX10-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX10-NEXT:    s_cbranch_execnz .LBB85_1
+; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: global_umin_saddr_i64_nortn_neg128:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3] offset:-128
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_load_b64 v[5:6], v0, s[2:3] offset:-128
+; GFX11-NEXT:    v_add_co_u32 v7, s[0:1], s2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v8, null, s3, 0, s[0:1]
+; GFX11-NEXT:    s_mov_b64 s[0:1], 0
+; GFX11-NEXT:    s_waitcnt_depctr 0xfffe
+; GFX11-NEXT:  .LBB85_1: ; %atomicrmw.start
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_le_u64_e32 vcc, v[5:6], v[1:2]
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v2, v6, vcc
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    global_atomic_cmpswap_b64 v[3:4], v[7:8], v[3:6], off offset:-128 glc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    buffer_gl1_inv
+; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[5:6]
+; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v3
+; GFX11-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_not1_b64 exec, exec, s[0:1]
+; GFX11-NEXT:    s_cbranch_execnz .LBB85_1
+; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX11-NEXT:    s_endpgm
   %zext.offset = zext i32 %voffset to i64
   %gep0 = getelementptr inbounds i8, i8 addrspace(1)* %sbase, i64 %zext.offset
   %gep1 = getelementptr inbounds i8, i8 addrspace(1)* %gep0, i64 -128
   %cast.gep1 = bitcast i8 addrspace(1)* %gep1 to i64 addrspace(1)*
   %unused = atomicrmw umin i64 addrspace(1)* %cast.gep1, i64 %data seq_cst
   ret void
 }

llvm/test/CodeGen/AMDGPU/global_atomics.ll

entry:
   %val = atomicrmw volatile sub i32 addrspace(1)* %ptr, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_offset(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_max_i32_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB27_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB27_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_max_i32_ret_offset:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB28_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_mov_b32 s6, s2
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s6
-; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB28_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] offset:16 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_max_i32_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB29_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB29_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1] offset:16
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_max_i32_ret_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB30_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB30_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] offset:16 glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile max i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_max_i32:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB31_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB31_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_max_i32_ret:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b32 s0, s4
-; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB32_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB32_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB32_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB32_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB32_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile max i32 addrspace(1)* %out, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_max_i32_addr64:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB33_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB33_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB33_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax v[0:1], v2
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB33_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_i32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB33_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_max_i32_ret_addr64:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB34_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB34_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i32_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB34_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB34_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i32_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_i32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB34_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile max i32 addrspace(1)* %ptr, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_offset(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_umax_i32_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB35_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB35_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB35_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB35_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB35_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_umax_i32_ret_offset:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB36_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB36_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_mov_b32 s6, s2
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s6
-; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB36_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB36_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] offset:16 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB36_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umax_i32_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB37_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB37_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB37_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB37_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1] offset:16
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB37_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB38_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB38_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB38_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB38_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] offset:16 glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB38_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile umax i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_umax_i32:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB39_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB39_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB39_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB39_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB39_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_umax_i32_ret:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b32 s0, s4
-; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB40_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB40_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB40_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB40_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB40_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umax i32 addrspace(1)* %out, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umax_i32_addr64:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB41_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB41_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB41_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax v[0:1], v2
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB41_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_max_u32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB41_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umax_i32_ret_addr64:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB42_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_max_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB42_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i32_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB42_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_max_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB42_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i32_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_max_u32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB42_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile umax i32 addrspace(1)* %ptr, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_offset(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_min_i32_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB43_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB43_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB43_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB43_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB43_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_min_i32_ret_offset:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB44_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB44_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_mov_b32 s6, s2
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s6
-; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB44_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB44_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] offset:16 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB44_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_min_i32_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB45_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB45_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB45_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB45_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1] offset:16
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB45_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_min_i32_ret_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB46_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB46_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB46_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB46_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] offset:16 glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB46_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile min i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_min_i32:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB47_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB47_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB47_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB47_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB47_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_min_i32_ret:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b32 s0, s4
-; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB48_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB48_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB48_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB48_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB48_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile min i32 addrspace(1)* %out, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_min_i32_addr64:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB49_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB49_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB49_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin v[0:1], v2
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB49_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_i32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB49_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_min_i32_ret_addr64:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB50_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_i32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB50_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i32_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB50_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_i32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB50_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i32_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_i32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB50_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile min i32 addrspace(1)* %ptr, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_offset(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_umin_i32_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB51_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB51_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
+; VI-NEXT:    s_load_dword s2, s[8:9], 0x10
+; VI-NEXT:    s_add_u32 s0, s8, 16
+; VI-NEXT:    s_addc_u32 s1, s9, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB51_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB51_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:16
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB51_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_umin_i32_ret_offset:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x4
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB52_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB52_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_mov_b32 s6, s2
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s6
-; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x10
+; VI-NEXT:    s_add_u32 s0, s4, 16
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB52_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB52_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_mov_b32 s0, s6
+; VI-NEXT:    s_mov_b32 s1, s7
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] offset:16 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] offset:16 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB52_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i32, i32 addrspace(1)* %out, i64 4
   %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_addr64_offset(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umin_i32_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64 offset:16
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB53_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB53_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin v[0:1], v2
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB53_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB53_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1] offset:16
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB53_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x4
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB54_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 offset:16 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB54_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x10
 ; VI-NEXT:    s_add_u32 s0, s0, 16
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB54_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB54_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x10
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] offset:16 glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB54_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %gep = getelementptr i32, i32 addrspace(1)* %ptr, i64 4
   %val = atomicrmw volatile umin i32 addrspace(1)* %gep, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32(i32 addrspace(1)* %out, i32 %in) {
 ; SI-LABEL: atomic_umin_i32:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dword s4, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SI-NEXT:    s_load_dword s2, s[0:1], 0xb
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[0:1], 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s3
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:  .LBB55_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s2, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; SI-NEXT:    s_cbranch_execnz .LBB55_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dword s4, s[0:1], 0x2c
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
+; VI-NEXT:    s_load_dword s3, s[4:5], 0x0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:  .LBB55_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s2, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[4:7], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB55_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s4
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB55_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in) {
 ; SI-LABEL: atomic_umin_i32_ret:
 ; SI:       ; %bb.0: ; %entry
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
+; SI-NEXT:    s_load_dword s10, s[0:1], 0xd
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SI-NEXT:    s_mov_b64 s[8:9], 0
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
-; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_mov_b32 s0, s4
-; SI-NEXT:    s_mov_b32 s1, s5
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB56_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s10, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_cbranch_execnz .LBB56_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
+; VI-NEXT:    s_load_dword s10, s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[8:9], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
+; VI-NEXT:    s_load_dword s2, s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB56_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s10, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB56_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v1, s2
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:  .LBB56_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s2, v3
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB56_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %val = atomicrmw volatile umin i32 addrspace(1)* %out, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_addr64(i32 addrspace(1)* %out, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umin_i32_addr64:
 ; SI:       ; %bb.0: ; %entry
-; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xd
+; SI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
 ; SI-NEXT:    s_load_dword s6, s[0:1], 0xb
-; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
-; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
-; SI-NEXT:    s_mov_b32 s2, 0
-; SI-NEXT:    v_mov_b32_e32 v2, s6
-; SI-NEXT:    v_mov_b32_e32 v0, s4
-; SI-NEXT:    v_mov_b32_e32 v1, s5
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB57_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s6, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB57_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s6, s[0:1], 0x2c
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB57_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s6, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin v[0:1], v2
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB57_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x24
-; GFX9-NEXT:    s_load_dword s6, s[0:1], 0x2c
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
-; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s6
+; GFX9-NEXT:    s_add_u32 s0, s6, s0
+; GFX9-NEXT:    s_addc_u32 s1, s7, s1
+; GFX9-NEXT:    s_load_dword s5, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s5
+; GFX9-NEXT:  .LBB57_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_min_u32_e32 v0, s4, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB57_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i32_ret_addr64(i32 addrspace(1)* %out, i32 addrspace(1)* %out2, i32 %in, i64 %index) {
 ; SI-LABEL: atomic_umin_i32_ret_addr64:
 ; SI:       ; %bb.0: ; %entry
+; SI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0xf
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; SI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xf
-; SI-NEXT:    s_load_dword s2, s[0:1], 0xd
-; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_load_dword s8, s[0:1], 0xd
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
+; SI-NEXT:    s_add_u32 s0, s4, s0
+; SI-NEXT:    s_addc_u32 s1, s5, s1
+; SI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; SI-NEXT:    s_mov_b64 s[4:5], 0
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v1, s2
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:  .LBB58_1: ; %atomicrmw.start
+; SI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; SI-NEXT:    v_min_u32_e32 v0, s8, v1
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mov_b32_e32 v3, v1
+; SI-NEXT:    v_mov_b32_e32 v2, v0
+; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    buffer_wbinvl1
+; SI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; SI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; SI-NEXT:    v_mov_b32_e32 v1, v2
+; SI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_cbranch_execnz .LBB58_1
+; SI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
+; SI-NEXT:    s_mov_b32 s3, 0xf000
+; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    s_mov_b32 s0, s6
 ; SI-NEXT:    s_mov_b32 s1, s7
-; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
-; SI-NEXT:    s_mov_b32 s6, 0
-; SI-NEXT:    s_mov_b32 s7, s3
-; SI-NEXT:    v_mov_b32_e32 v2, s2
-; SI-NEXT:    v_mov_b32_e32 v0, s8
-; SI-NEXT:    v_mov_b32_e32 v1, s9
-; SI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 glc
-; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    buffer_wbinvl1
-; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i32_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dword s8, s[0:1], 0x34
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; VI-NEXT:    s_add_u32 s0, s4, s0
 ; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_mov_b32_e32 v1, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    s_load_dword s2, s[0:1], 0x0
+; VI-NEXT:    s_mov_b64 s[4:5], 0
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v1, s2
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:  .LBB58_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_min_u32_e32 v0, s8, v1
+; VI-NEXT:    v_mov_b32_e32 v3, v1
+; VI-NEXT:    v_mov_b32_e32 v2, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
+; VI-NEXT:    buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v1, v2
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB58_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; VI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i32_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dword s8, s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 2
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
-; GFX9-NEXT:    v_mov_b32_e32 v1, s8
+; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s4
+; GFX9-NEXT:  .LBB58_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v3, v0
+; GFX9-NEXT:    v_min_u32_e32 v2, s8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] glc
+; GFX9-NEXT:    global_atomic_cmpswap v0, v1, v[2:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
-; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v3
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB58_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    global_store_dword v1, v0, s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i32, i32 addrspace(1)* %out, i64 %index
   %val = atomicrmw volatile umin i32 addrspace(1)* %ptr, i32 %in seq_cst
   store i32 %val, i32 addrspace(1)* %out2
   ret void
 }
 

llvm/test/CodeGen/AMDGPU/global_atomics_i64.ll

entry:
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_offset(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_max_i64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB24_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB24_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB24_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB24_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB24_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_max_i64_ret_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB25_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_mov_b32 s6, s2
-; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB25_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB25_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB25_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB25_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_max_i64_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB26_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB26_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB26_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB26_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB26_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_max_i64_ret_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB27_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB27_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB27_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB27_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB27_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_max_i64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s4, s0
-; CI-NEXT:    s_mov_b32 s5, s1
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB28_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s7, 0xf000
-; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s4, s0
-; VI-NEXT:    s_mov_b32 s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB28_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB28_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB28_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB28_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_max_i64_ret:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s0, s4
-; CI-NEXT:    s_mov_b32 s1, s5
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB29_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smax_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB29_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB29_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB29_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB29_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %out, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_max_i64_addr64:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB30_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_addr64:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB30_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB30_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_addr64:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB30_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB30_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_max_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_max_i64_ret_addr64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB31_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_max_i64_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB31_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB31_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_max_i64_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v2, v[0:1], s[0:1] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB31_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB31_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile max i64 addrspace(1)* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_offset(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_umax_i64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB32_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB32_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB32_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB32_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB32_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB32_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_umax_i64_ret_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB33_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB33_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_mov_b32 s6, s2
-; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB33_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB33_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB33_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB33_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umax_i64_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB34_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB34_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB34_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB34_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB34_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB34_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umax_i64_ret_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB35_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB35_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB35_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB35_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB35_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB35_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_umax_i64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s4, s0
-; CI-NEXT:    s_mov_b32 s5, s1
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB36_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB36_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s7, 0xf000
-; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s4, s0
-; VI-NEXT:    s_mov_b32 s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB36_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB36_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB36_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB36_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_umax_i64_ret:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s0, s4
-; CI-NEXT:    s_mov_b32 s1, s5
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB37_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB37_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umax_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB37_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB37_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB37_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB37_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %out, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umax_i64_addr64:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB38_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB38_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_addr64:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB38_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB38_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_addr64:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB38_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB38_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umax_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umax_i64_ret_addr64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umax_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB39_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB39_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umax_i64_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umax_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB39_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB39_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umax_i64_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v2, v[0:1], s[0:1] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB39_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_lt_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB39_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile umax i64 addrspace(1)* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_offset(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_min_i64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB40_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB40_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB40_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB40_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB40_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB40_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_min_i64_ret_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB41_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB41_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_mov_b32 s6, s2
-; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB41_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB41_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB41_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB41_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_min_i64_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB42_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB42_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB42_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB42_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB42_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB42_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_min_i64_ret_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB43_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB43_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB43_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB43_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB43_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB43_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_min_i64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s4, s0
-; CI-NEXT:    s_mov_b32 s5, s1
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB44_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB44_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s7, 0xf000
-; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s4, s0
-; VI-NEXT:    s_mov_b32 s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB44_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB44_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB44_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB44_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_min_i64_ret:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s0, s4
-; CI-NEXT:    s_mov_b32 s1, s5
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB45_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB45_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_smin_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB45_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB45_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB45_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB45_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %out, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_min_i64_addr64:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB46_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB46_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_addr64:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB46_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB46_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_addr64:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB46_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB46_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_min_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_min_i64_ret_addr64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_smin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB47_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB47_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_min_i64_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_smin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB47_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB47_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_min_i64_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v2, v[0:1], s[0:1] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB47_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_i64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB47_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile min i64 addrspace(1)* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_offset(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_umin_i64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x8
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB48_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB48_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b64 s[8:9], 0
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_mov_b32 s3, 0xf000
-; VI-NEXT:    s_mov_b32 s2, -1
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 offset:32
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[10:11], s[0:1], 0x20
+; VI-NEXT:    s_add_u32 s4, s0, 32
+; VI-NEXT:    s_addc_u32 s5, s1, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s10
+; VI-NEXT:    v_mov_b32_e32 v3, s11
+; VI-NEXT:  .LBB48_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[8:9], vcc, s[8:9]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[8:9]
+; VI-NEXT:    s_cbranch_execnz .LBB48_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB48_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB48_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_umin_i64_ret_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB49_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB49_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_mov_b32 s6, s2
-; CI-NEXT:    s_mov_b32 s7, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_ret_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s0, s4, 32
+; VI-NEXT:    s_addc_u32 s1, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB49_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB49_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_mov_b32 s6, s2
-; VI-NEXT:    s_mov_b32 s7, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0 offset:32 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_ret_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[4:5] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB49_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB49_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, i64 addrspace(1)* %out, i64 4
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_addr64_offset(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umin_i64_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[4:7], 0 addr64 offset:32
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB50_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB50_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x20
+; VI-NEXT:    s_add_u32 s4, s4, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s5, s5, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB50_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB50_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB50_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1] offset:32
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] offset:32 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB50_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_addr64_offset(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umin_i64_ret_addr64_offset:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 offset:32 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x8
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB51_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 offset:32 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB51_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_ret_addr64_offset:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    s_add_u32 s0, s0, 32
-; VI-NEXT:    s_addc_u32 s1, s1, 0
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s6, s0, s6
+; VI-NEXT:    s_addc_u32 s7, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[6:7], 0x20
+; VI-NEXT:    s_add_u32 s8, s6, 32
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_addc_u32 s9, s7, 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB51_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB51_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_ret_addr64_offset:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[0:1] offset:32 glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x20
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB51_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] offset:32 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB51_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %gep = getelementptr i64, i64 addrspace(1)* %ptr, i64 4
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %gep, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64(i64 addrspace(1)* %out, i64 %in) {
 ; CI-LABEL: atomic_umin_i64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
-; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b64 s[4:5], 0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s4, s0
-; CI-NEXT:    s_mov_b32 s5, s1
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; CI-NEXT:    s_mov_b32 s6, s2
+; CI-NEXT:    s_mov_b32 s7, s3
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB52_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s7
+; CI-NEXT:    v_mov_b32_e32 v4, s6
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; CI-NEXT:    s_cbranch_execnz .LBB52_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; VI-NEXT:    s_mov_b32 s7, 0xf000
-; VI-NEXT:    s_mov_b32 s6, -1
+; VI-NEXT:    s_mov_b64 s[4:5], 0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s4, s0
-; VI-NEXT:    s_mov_b32 s5, s1
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[4:7], 0
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; VI-NEXT:    s_mov_b32 s6, s2
+; VI-NEXT:    s_mov_b32 s7, s3
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB52_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s7
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; VI-NEXT:    s_cbranch_execnz .LBB52_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], 0
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s6
+; GFX9-NEXT:    v_mov_b32_e32 v3, s7
+; GFX9-NEXT:  .LBB52_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v5, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[4:5]
+; GFX9-NEXT:    s_cbranch_execnz .LBB52_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in) {
 ; CI-LABEL: atomic_umin_i64_ret:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; CI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
+; CI-NEXT:    s_mov_b64 s[10:11], 0
 ; CI-NEXT:    s_mov_b32 s3, 0xf000
 ; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_mov_b32 s0, s4
-; CI-NEXT:    s_mov_b32 s1, s5
-; CI-NEXT:    v_mov_b32_e32 v0, s8
-; CI-NEXT:    v_mov_b32_e32 v1, s9
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s12
+; CI-NEXT:    v_mov_b32_e32 v3, s13
+; CI-NEXT:  .LBB53_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s9
+; CI-NEXT:    v_mov_b32_e32 v4, s8
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_cbranch_execnz .LBB53_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; CI-NEXT:    s_mov_b32 s3, 0xf000
+; CI-NEXT:    s_mov_b32 s2, -1
 ; CI-NEXT:    s_mov_b32 s0, s6
 ; CI-NEXT:    s_mov_b32 s1, s7
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_ret:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; VI-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x34
+; VI-NEXT:    s_mov_b64 s[10:11], 0
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_mov_b32 s0, s4
-; VI-NEXT:    s_mov_b32 s1, s5
-; VI-NEXT:    v_mov_b32_e32 v0, s8
-; VI-NEXT:    v_mov_b32_e32 v1, s9
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    buffer_atomic_umin_x2 v[0:1], off, s[0:3], 0 glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_load_dwordx2 s[12:13], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], s[4:5]
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s12
+; VI-NEXT:    v_mov_b32_e32 v3, s13
+; VI-NEXT:  .LBB53_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s9
+; VI-NEXT:    v_mov_b32_e32 v4, s8
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[10:11], vcc, s[10:11]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_cbranch_execnz .LBB53_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[10:11]
+; VI-NEXT:    s_mov_b32 s3, 0xf000
+; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_mov_b32 s0, s6
 ; VI-NEXT:    s_mov_b32 s1, s7
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_ret:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NEXT:    s_mov_b64 s[0:1], 0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s2
-; GFX9-NEXT:    v_mov_b32_e32 v1, s3
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[4:5] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB53_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s3
+; GFX9-NEXT:    v_mov_b32_e32 v1, s2
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[4:5] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    s_cbranch_execnz .LBB53_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[6:7]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %out, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_addr64(i64 addrspace(1)* %out, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umin_i64_addr64:
 ; CI:       ; %bb.0: ; %entry
-; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
-; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s6
-; CI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; CI-NEXT:    v_mov_b32_e32 v3, s1
-; CI-NEXT:    v_mov_b32_e32 v1, s7
+; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0xd
+; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    s_mov_b32 s6, 0
-; CI-NEXT:    v_mov_b32_e32 v2, s0
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[4:7], 0 addr64
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; CI-NEXT:    s_add_u32 s4, s0, s4
+; CI-NEXT:    s_addc_u32 s5, s1, s5
+; CI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s8
+; CI-NEXT:    v_mov_b32_e32 v3, s9
+; CI-NEXT:  .LBB54_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s3
+; CI-NEXT:    v_mov_b32_e32 v4, s2
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB54_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_addr64:
 ; VI:       ; %bb.0: ; %entry
-; VI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
-; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s7, 0xf000
+; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s6
-; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 3
-; VI-NEXT:    s_add_u32 s0, s4, s0
-; VI-NEXT:    s_addc_u32 s1, s5, s1
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v1, s7
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin_x2 v[2:3], v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
+; VI-NEXT:    s_add_u32 s4, s0, s4
+; VI-NEXT:    s_addc_u32 s5, s1, s5
+; VI-NEXT:    s_load_dwordx2 s[8:9], s[4:5], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s8
+; VI-NEXT:    v_mov_b32_e32 v3, s9
+; VI-NEXT:  .LBB54_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[2:3], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s3
+; VI-NEXT:    v_mov_b32_e32 v4, s2
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[4:7], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB54_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_addr64:
 ; GFX9:       ; %bb.0: ; %entry
-; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s6
 ; GFX9-NEXT:    s_lshl_b64 s[0:1], s[2:3], 3
 ; GFX9-NEXT:    s_add_u32 s0, s4, s0
-; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_addc_u32 s1, s5, s1
+; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[2:3], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v2, s4
+; GFX9-NEXT:    v_mov_b32_e32 v3, s5
+; GFX9-NEXT:  .LBB54_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[6:7], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s7
+; GFX9-NEXT:    v_mov_b32_e32 v5, s6
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v5, v2, vcc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v2, v[0:1], s[0:1]
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX9-NEXT:    s_cbranch_execnz .LBB54_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
   ret void
 }
 
 define amdgpu_kernel void @atomic_umin_i64_ret_addr64(i64 addrspace(1)* %out, i64 addrspace(1)* %out2, i64 %in, i64 %index) {
 ; CI-LABEL: atomic_umin_i64_ret_addr64:
 ; CI:       ; %bb.0: ; %entry
 ; CI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b32 s10, -1
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s4
-; CI-NEXT:    v_mov_b32_e32 v1, s5
-; CI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; CI-NEXT:    v_mov_b32_e32 v2, s4
-; CI-NEXT:    s_mov_b32 s8, s2
-; CI-NEXT:    s_mov_b32 s9, s3
-; CI-NEXT:    s_mov_b32 s2, 0
-; CI-NEXT:    s_mov_b32 s3, s11
-; CI-NEXT:    v_mov_b32_e32 v3, s5
-; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CI-NEXT:    buffer_atomic_umin_x2 v[0:1], v[2:3], s[0:3], 0 addr64 glc
-; CI-NEXT:    s_waitcnt vmcnt(0)
-; CI-NEXT:    buffer_wbinvl1_vol
-; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[8:11], 0
+; CI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; CI-NEXT:    s_add_u32 s8, s0, s6
+; CI-NEXT:    s_addc_u32 s9, s1, s7
+; CI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; CI-NEXT:    s_mov_b64 s[0:1], 0
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v2, s6
+; CI-NEXT:    v_mov_b32_e32 v3, s7
+; CI-NEXT:  .LBB55_1: ; %atomicrmw.start
+; CI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v0, s5
+; CI-NEXT:    v_mov_b32_e32 v4, s4
+; CI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; CI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; CI-NEXT:    v_mov_b32_e32 v7, v3
+; CI-NEXT:    v_mov_b32_e32 v6, v2
+; CI-NEXT:    v_mov_b32_e32 v5, v1
+; CI-NEXT:    v_mov_b32_e32 v4, v0
+; CI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    buffer_wbinvl1_vol
+; CI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, v4
+; CI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, v5
+; CI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_cbranch_execnz .LBB55_1
+; CI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; CI-NEXT:    s_or_b64 exec, exec, s[0:1]
+; CI-NEXT:    s_mov_b32 s7, 0xf000
+; CI-NEXT:    s_mov_b32 s6, -1
+; CI-NEXT:    s_mov_b32 s4, s2
+; CI-NEXT:    s_mov_b32 s5, s3
+; CI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: atomic_umin_i64_ret_addr64:
 ; VI:       ; %bb.0: ; %entry
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    s_mov_b32 s11, 0xf000
+; VI-NEXT:    s_mov_b32 s10, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s4
-; VI-NEXT:    v_mov_b32_e32 v1, s5
-; VI-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; VI-NEXT:    s_add_u32 s0, s0, s4
-; VI-NEXT:    s_addc_u32 s1, s1, s5
-; VI-NEXT:    v_mov_b32_e32 v3, s1
-; VI-NEXT:    v_mov_b32_e32 v2, s0
-; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    flat_atomic_umin_x2 v[0:1], v[2:3], v[0:1] glc
-; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; VI-NEXT:    s_add_u32 s8, s0, s6
+; VI-NEXT:    s_addc_u32 s9, s1, s7
+; VI-NEXT:    s_load_dwordx2 s[6:7], s[8:9], 0x0
+; VI-NEXT:    s_mov_b64 s[0:1], 0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v2, s6
+; VI-NEXT:    v_mov_b32_e32 v3, s7
+; VI-NEXT:  .LBB55_1: ; %atomicrmw.start
+; VI-NEXT:    ; =>This Inner Loop Header: Depth=1
+; VI-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v0, s5
+; VI-NEXT:    v_mov_b32_e32 v4, s4
+; VI-NEXT:    v_cndmask_b32_e32 v1, v0, v3, vcc
+; VI-NEXT:    v_cndmask_b32_e32 v0, v4, v2, vcc
+; VI-NEXT:    v_mov_b32_e32 v7, v3
+; VI-NEXT:    v_mov_b32_e32 v6, v2
+; VI-NEXT:    v_mov_b32_e32 v5, v1
+; VI-NEXT:    v_mov_b32_e32 v4, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    buffer_atomic_cmpswap_x2 v[4:7], off, s[8:11], 0 glc
+; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    buffer_wbinvl1_vol
+; VI-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, v4
+; VI-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, v5
+; VI-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; VI-NEXT:    s_cbranch_execnz .LBB55_1
+; VI-NEXT:  ; %bb.2: ; %atomicrmw.end
+; VI-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
-; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
+; VI-NEXT:    buffer_store_dwordx2 v[4:5], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: atomic_umin_i64_ret_addr64:
 ; GFX9:       ; %bb.0: ; %entry
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_mov_b32_e32 v0, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, s5
-; GFX9-NEXT:    s_lshl_b64 s[4:5], s[6:7], 3
-; GFX9-NEXT:    s_add_u32 s0, s0, s4
-; GFX9-NEXT:    s_addc_u32 s1, s1, s5
-; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v2, v[0:1], s[0:1] glc
-; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    s_lshl_b64 s[6:7], s[6:7], 3
+; GFX9-NEXT:    s_add_u32 s0, s0, s6
+; GFX9-NEXT:    s_addc_u32 s1, s1, s7
+; GFX9-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s8
+; GFX9-NEXT:    v_mov_b32_e32 v1, s9
+; GFX9-NEXT:  .LBB55_1: ; %atomicrmw.start
+; GFX9-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT:    v_mov_b32_e32 v6, v1
+; GFX9-NEXT:    v_mov_b32_e32 v5, v0
+; GFX9-NEXT:    v_cmp_ge_u64_e32 vcc, s[4:5], v[5:6]
+; GFX9-NEXT:    v_mov_b32_e32 v0, s5
+; GFX9-NEXT:    v_mov_b32_e32 v1, s4
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v6, vcc
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v1, v5, vcc
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v2, v[3:6], s[0:1] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    buffer_wbinvl1_vol
+; GFX9-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[5:6]
+; GFX9-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
+; GFX9-NEXT:    s_andn2_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    s_cbranch_execnz .LBB55_1
+; GFX9-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX9-NEXT:    s_or_b64 exec, exec, s[6:7]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
 ; GFX9-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, i64 addrspace(1)* %out, i64 %index
   %tmp0 = atomicrmw volatile umin i64 addrspace(1)* %ptr, i64 %in seq_cst
   store i64 %tmp0, i64 addrspace(1)* %out2
   ret void
 }

llvm/test/CodeGen/AMDGPU/move-to-valu-atomicrmw.ll

 ; GCN-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
 ; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    s_mov_b64 s[8:9], s[6:7]
 ; GCN-NEXT:    buffer_load_dwordx2 v[1:2], v[1:2], s[8:11], 0 addr64 glc
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v0
 ; GCN-NEXT:    s_and_saveexec_b64 s[2:3], vcc
-; GCN-NEXT:    s_cbranch_execz .LBB0_2
+; GCN-NEXT:    s_cbranch_execz .LBB0_4
 ; GCN-NEXT:  ; %bb.1: ; %atomic
-; GCN-NEXT:    s_load_dword s0, s[0:1], 0xf
 ; GCN-NEXT:    s_mov_b32 s8, s10
 ; GCN-NEXT:    s_mov_b32 s9, s10
-; GCN-NEXT:    s_mov_b32 s6, -1
-; GCN-NEXT:    s_mov_b32 s7, s11
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    buffer_load_dword v4, v[1:2], s[8:11], 0 addr64 offset:400
+; GCN-NEXT:    s_load_dword s2, s[0:1], 0xf
+; GCN-NEXT:    s_mov_b64 s[0:1], 0
+; GCN-NEXT:  .LBB0_2: ; %atomicrmw.start
+; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-NEXT:    buffer_atomic_smax v0, v[1:2], s[8:11], 0 addr64 offset:400 glc
+; GCN-NEXT:    v_max_i32_e32 v3, s2, v4
+; GCN-NEXT:    s_waitcnt expcnt(0)
+; GCN-NEXT:    v_mov_b32_e32 v6, v4
+; GCN-NEXT:    v_mov_b32_e32 v5, v3
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    buffer_atomic_cmpswap v[5:6], v[1:2], s[8:11], 0 addr64 offset:400 glc
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    buffer_wbinvl1
-; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
-; GCN-NEXT:  .LBB0_2: ; %exit
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v4
+; GCN-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v4, v5
+; GCN-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_cbranch_execnz .LBB0_2
+; GCN-NEXT:  ; %bb.3: ; %atomicrmw.end
+; GCN-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_mov_b32 s7, 0xf000
+; GCN-NEXT:    s_mov_b32 s6, -1
+; GCN-NEXT:    buffer_store_dword v5, off, s[4:7], 0
+; GCN-NEXT:  .LBB0_4: ; %exit
 ; GCN-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %tid.gep = getelementptr ptr addrspace(1), ptr addrspace(1) %in, i32 %tid
   %ptr = load volatile ptr addrspace(1), ptr addrspace(1) %tid.gep
   %xor = xor i32 %tid, 1
   %cmp = icmp ne i32 %xor, 0
   br i1 %cmp, label %atomic, label %exit
 

 ; GCN-NEXT:    s_mov_b32 s6, 0
 ; GCN-NEXT:    v_lshlrev_b32_e32 v1, 3, v0
 ; GCN-NEXT:    v_mov_b32_e32 v2, 0
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    buffer_load_dwordx2 v[1:2], v[1:2], s[4:7], 0 addr64 glc
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    v_cmp_ne_u32_e32 vcc, 1, v0
 ; GCN-NEXT:    s_and_saveexec_b64 s[2:3], vcc
-; GCN-NEXT:    s_cbranch_execz .LBB1_2
+; GCN-NEXT:    s_cbranch_execz .LBB1_3
 ; GCN-NEXT:  ; %bb.1: ; %atomic
-; GCN-NEXT:    s_load_dword s0, s[0:1], 0xf
 ; GCN-NEXT:    s_mov_b32 s4, s6
 ; GCN-NEXT:    s_mov_b32 s5, s6
-; GCN-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    buffer_load_dword v4, v[1:2], s[4:7], 0 addr64 offset:400
+; GCN-NEXT:    s_load_dword s2, s[0:1], 0xf
+; GCN-NEXT:    s_mov_b64 s[0:1], 0
+; GCN-NEXT:  .LBB1_2: ; %atomicrmw.start
+; GCN-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_max_i32_e32 v3, s2, v4
+; GCN-NEXT:    s_waitcnt expcnt(0)
+; GCN-NEXT:    v_mov_b32_e32 v6, v4
+; GCN-NEXT:    v_mov_b32_e32 v5, v3
 ; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-NEXT:    buffer_atomic_smax v0, v[1:2], s[4:7], 0 addr64 offset:400
+; GCN-NEXT:    buffer_atomic_cmpswap v[5:6], v[1:2], s[4:7], 0 addr64 offset:400 glc
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
 ; GCN-NEXT:    buffer_wbinvl1
-; GCN-NEXT:  .LBB1_2: ; %exit
+; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v4
+; GCN-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v4, v5
+; GCN-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GCN-NEXT:    s_cbranch_execnz .LBB1_2
+; GCN-NEXT:  .LBB1_3: ; %exit
 ; GCN-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %tid.gep = getelementptr ptr addrspace(1), ptr addrspace(1) %in, i32 %tid
   %ptr = load volatile ptr addrspace(1), ptr addrspace(1) %tid.gep
   %xor = xor i32 %tid, 1
   %cmp = icmp ne i32 %xor, 0
   br i1 %cmp, label %atomic, label %exit