AMDGPU/GlobalISel Check for NoNaNsFPMath in isKnownNeverSNaN

[Petar.Avramovic]

Summary

Check for NoNaNsFPMath function attribute in isKnownNeverSNaN.
Function attributes are in held in 'TargetMachine.Options'.
Among other things, this allows selection of some patterns imported
in D87351 since G_FCANONICALIZE is not generated when isKnownNeverSNaN
returns true in lowerFMinNumMaxNum.

However we notice some incorrect results since function attributes are
not correctly written in TargetMachine.Options when next function is
processed. Look at @v_test_no_global_nnans_med3_f32_pat0_srcmod0, it has
"no-nans-fp-math"="false" but TargetMachine.Options still has it set to
true since first function in test file had this attribute set to true.
This will be fixed by in D87511.

[Petar.Avramovic]

With G_FCANONICALIZE out of the way, we can now select some of the med3 patterns that were imported in D87351.

llvm/test/CodeGen/AMDGPU/GlobalISel/fmed3.ll

69–71

This was not folded because G_FNEG was expected.
TODO: add combine for G_FNEG.

[arsenm]

Probably should add a dedicated test for this

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2658

Should copy the comment from the DAG

// Reset the target options before resetting the optimization
// level below.
// FIXME: This is a horrible hack and should be processed via
// codegen looking at the optimization level explicitly when
// it wants to look at it.

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

5420

I think checking these fields should be buried inside isKnownNeverSNaN

[Petar.Avramovic]

Move change in IRTranslator in another patch, since effects of that change can only be seen when function attributes are inspected during translation of a function that was not first in input file.
Move check for NoNaNsFPMath in isKnownNeverSNaN.
Reduce number of tests, only keep characteristic ones that will be updated when we add some combines and improvements to legalizer. Later switch to existing test in sdag that also checks for pattern commutes.

[foad]

This will be fixed by in D87511.

Wouldn't it make more sense to change the order of the patches in the stack so that D87511 is applied first and this one is rebaed on top of it?

[Petar.Avramovic]

In D87456#2267858, @foad wrote:

This will be fixed by in D87511.

Wouldn't it make more sense to change the order of the patches in the stack so that D87511 is applied first and this one is rebaed on top of it?

Yes, but it is this way because of the regression test. Change is visible in output when there is something that accesses function attributes.

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2658

This does not seem relevant since global-isel does not reset opt level.

[arsenm]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

471–472 ↗

(On Diff #291200)

Should delete this null check in a separate change

474–478 ↗

(On Diff #291200)

Could merge these into getFlag() || TM.Options ...

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

bool IRTranslator::runOnMachineFunction(MachineFunction &CurMF) {
   EntryBuilder->setMF(*MF);
   MRI = &MF->getRegInfo();
   DL = &F.getParent()->getDataLayout();
   ORE = std::make_unique<OptimizationRemarkEmitter>(&F);
   FuncInfo.MF = MF;
   FuncInfo.BPI = nullptr;
   const auto &TLI = *MF->getSubtarget().getTargetLowering();
   const TargetMachine &TM = MF->getTarget();
+  TM.resetTargetOptions(F);
   SL = std::make_unique<GISelSwitchLowering>(this, FuncInfo);
   SL->init(TLI, TM, *DL);
 
   EnableOpts = TM.getOptLevel() != CodeGenOpt::None && !skipFunction(F);
 
   assert(PendingPHIs.empty() && "stale PHIs");
 
   if (!DL->isLittleEndian()) {

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/TargetFrameLowering.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetMachine.h"
 
 #define DEBUG_TYPE "legalizer"
 
 using namespace llvm;
 using namespace LegalizeActions;
 using namespace MIPatternMatch;
 
 /// Try to break down \p OrigTy into \p NarrowTy sized pieces.
LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) {
 
   if (!MI.getFlag(MachineInstr::FmNoNans)) {
     // Insert canonicalizes if it's possible we need to quiet to get correct
     // sNaN behavior.
 
     // Note this must be done here, and not as an optimization combine in the
     // absence of a dedicate quiet-snan instruction as we're using an
     // omni-purpose G_FCANONICALIZE.
+    if (!MI.getMF()->getTarget().Options.NoNaNsFPMath) {
       if (!isKnownNeverSNaN(Src0, MRI))
         Src0 = MIRBuilder.buildFCanonicalize(Ty, Src0, MI.getFlags()).getReg(0);
 
       if (!isKnownNeverSNaN(Src1, MRI))
         Src1 = MIRBuilder.buildFCanonicalize(Ty, Src1, MI.getFlags()).getReg(0);
     }
+  }
 
   // If there are no nans, it's safe to simply replace this with the non-IEEE
   // version.
   MIRBuilder.buildInstr(NewOp, {Dst}, {Src0, Src1}, MI.getFlags());
   MI.eraseFromParent();
   return Legalized;
 }
 

llvm/test/CodeGen/AMDGPU/GlobalISel/fmed3.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=GFX89 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX9 -check-prefix=GFX89 %s
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_srcmod0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_sub_f32_e32 v2, 0x80000000, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_sub_f32_e32 v0, 0x80000000, v0
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_sub_f32_e32 v1, 0x80000000, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %a.fneg = fsub float -0.0, %a
+  %tmp0 = call float @llvm.minnum.f32(float %a.fneg, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.fneg, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_srcmod1(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod1:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_sub_f32_e32 v3, 0x80000000, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod1:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_sub_f32_e32 v1, 0x80000000, v1
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod1:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_sub_f32_e32 v2, 0x80000000, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %b.fneg = fsub float -0.0, %b
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b.fneg)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b.fneg)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_srcmod2(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod2:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_sub_f32_e32 v4, 0x80000000, v4
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod2:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_sub_f32_e32 v2, 0x80000000, v2
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod2:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_sub_f32_e32 v3, 0x80000000, v3
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %c.fneg = fsub float -0.0, %c
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.fneg)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_srcmod012(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod012:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b32 s2, 0x80000000
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_sub_f32_e32 v2, s2, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_sub_f32_e64 v4, s2, |v4|
+; SI-NEXT:    v_med3_f32 v2, v2, |v3|, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod012:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    s_mov_b32 s2, 0x80000000
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_sub_f32_e32 v4, s2, v7
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_sub_f32_e64 v3, s2, |v3|
+; VI-NEXT:    v_med3_f32 v2, v4, |v2|, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod012:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_mov_b32 s2, 0x80000000
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_sub_f32_e32 v1, s2, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_sub_f32_e64 v3, s2, |v3|
+; GFX9-NEXT:    v_med3_f32 v1, v1, |v2|, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.fneg = fsub float -0.0, %a
+  %b.fabs = call float @llvm.fabs.f32(float %b)
+  %c.fabs = call float @llvm.fabs.f32(float %c)
+  %c.fabs.fneg = fsub float -0.0, %c.fabs
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.fneg, float %b.fabs)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.fneg, float %b.fabs)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.fabs.fneg)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_negabs012(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_negabs012:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b32 s2, 0x80000000
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_sub_f32_e64 v2, s2, |v2|
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_sub_f32_e64 v3, s2, |v3|
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_sub_f32_e64 v4, s2, |v4|
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_negabs012:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    s_mov_b32 s2, 0x80000000
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_sub_f32_e64 v4, s2, |v7|
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_sub_f32_e64 v2, s2, |v2|
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_sub_f32_e64 v3, s2, |v3|
+; VI-NEXT:    v_med3_f32 v2, v4, v2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_negabs012:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_mov_b32 s2, 0x80000000
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_sub_f32_e64 v1, s2, |v1|
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_sub_f32_e64 v2, s2, |v2|
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_sub_f32_e64 v3, s2, |v3|
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.fabs = call float @llvm.fabs.f32(float %a)
+  %a.fabs.fneg = fsub float -0.0, %a.fabs
+  %b.fabs = call float @llvm.fabs.f32(float %b)
+  %b.fabs.fneg = fsub float -0.0, %b.fabs
+  %c.fabs = call float @llvm.fabs.f32(float %c)
+  %c.fabs.fneg = fsub float -0.0, %c.fabs
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.fabs.fneg, float %b.fabs.fneg)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.fabs.fneg, float %b.fabs.fneg)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.fabs.fneg)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_nnan_inputs_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_nnan_inputs_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_f32_e32 v3, 2.0, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v4, 4.0, v4
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_nnan_inputs_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_add_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_nnan_inputs_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.nnan = fadd nnan float %a, 1.0
+  %b.nnan = fadd nnan float %b, 2.0
+  %c.nnan = fadd nnan float %c, 4.0
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.nnan, float %b.nnan)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.nnan, float %b.nnan)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.nnan)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+; 16 combinations
+
+; 0: max(min(x, y), min(max(x, y), z))
+; 1: max(min(x, y), min(max(y, x), z))
+; 2: max(min(x, y), min(z, max(x, y)))
+; 3: max(min(x, y), min(z, max(y, x)))
+; 4: max(min(y, x), min(max(x, y), z))
+; 5: max(min(y, x), min(max(y, x), z))
+; 6: max(min(y, x), min(z, max(x, y)))
+; 7: max(min(y, x), min(z, max(y, x)))
+;
+; + commute outermost max
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat1(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat1:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat1:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat1:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat2(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat2:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat2:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat2:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat3(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat3:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat3:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat3:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat4(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat4:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat4:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat4:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat5(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat5:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat5:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat5:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat6(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat6:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat6:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat6:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat7(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat7:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat7:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat7:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat8(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat8:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat8:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat8:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat9(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat9:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat9:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat9:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat10(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat10:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat10:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat10:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat11(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat11:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat11:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat11:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat12(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat12:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat12:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat12:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat13(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat13:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat13:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat13:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat14(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat14:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat14:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v0, v1, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat14:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat15(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat15:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_med3_f32 v2, v3, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat15:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_med3_f32 v0, v1, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat15:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_med3_f32 v1, v2, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %b, float %a)
+  %tmp1 = call float @llvm.maxnum.f32(float %b, float %a)
+  %tmp2 = call float @llvm.minnum.f32(float %c, float %tmp1)
+  %med3 = call float @llvm.maxnum.f32(float %tmp2, float %tmp0)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+; ---------------------------------------------------------------------
+; Negative patterns
+; ---------------------------------------------------------------------
+
+define amdgpu_kernel void @v_test_safe_med3_f32_pat0_multi_use0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_test_safe_med3_f32_pat0_multi_use0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s3, s11
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    buffer_store_dword v5, off, s[0:3], 0
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_safe_med3_f32_pat0_multi_use0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v5
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_safe_med3_f32_pat0_multi_use0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    global_store_dword v[0:1], v4, off
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v3, v3, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  store volatile float %tmp0, float addrspace(1)* undef
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_safe_med3_f32_pat0_multi_use1(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_test_safe_med3_f32_pat0_multi_use1:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s3, s11
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_safe_med3_f32_pat0_multi_use1:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_safe_med3_f32_pat0_multi_use1:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    global_store_dword v[0:1], v1, off
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v3, v3, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  store volatile float %tmp1, float addrspace(1)* undef
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_safe_med3_f32_pat0_multi_use2(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_test_safe_med3_f32_pat0_multi_use2:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b32 s2, -1
+; SI-NEXT:    s_mov_b32 s3, s11
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v5
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
+; SI-NEXT:    s_waitcnt expcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v4, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_safe_med3_f32_pat0_multi_use2:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_safe_med3_f32_pat0_multi_use2:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_f32_e32 v3, v3, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    global_store_dword v[0:1], v1, off
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  store volatile float %tmp2, float addrspace(1)* undef
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+
+define amdgpu_kernel void @v_test_safe_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_test_safe_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_safe_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_safe_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_max_f32_e32 v3, v3, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %tmp0 = call float @llvm.minnum.f32(float %a, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_nnan_inputs_missing0_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_nnan_inputs_missing0_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_f32_e32 v3, 2.0, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v4, 4.0, v4
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_nnan_inputs_missing0_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_add_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; VI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_nnan_inputs_missing0_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.nnan = fadd float %a, 1.0
+  %b.nnan = fadd nnan float %b, 2.0
+  %c.nnan = fadd nnan float %c, 4.0
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.nnan, float %b.nnan)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.nnan, float %b.nnan)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.nnan)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_nnan_inputs_missing1_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_nnan_inputs_missing1_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_f32_e32 v3, 2.0, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v4, 4.0, v4
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_nnan_inputs_missing1_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_add_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_nnan_inputs_missing1_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.nnan = fadd nnan float %a, 1.0
+  %b.nnan = fadd float %b, 2.0
+  %c.nnan = fadd nnan float %c, 4.0
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.nnan, float %b.nnan)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.nnan, float %b.nnan)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.nnan)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_nnan_inputs_missing2_med3_f32_pat0(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_nnan_inputs_missing2_med3_f32_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_add_f32_e32 v3, 2.0, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v4, 4.0, v4
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_nnan_inputs_missing2_med3_f32_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_add_f32_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; VI-NEXT:    v_min_f32_e32 v5, v4, v2
+; VI-NEXT:    v_max_f32_e32 v2, v4, v2
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_nnan_inputs_missing2_med3_f32_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_f32_e32 v2, 2.0, v2
+; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v3, 4.0, v3
+; GFX9-NEXT:    v_max_f32_e32 v2, v3, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+
+  %a.nnan = fadd nnan float %a, 1.0
+  %b.nnan = fadd nnan float %b, 2.0
+  %c.nnan = fadd float %c, 4.0
+
+  %tmp0 = call float @llvm.minnum.f32(float %a.nnan, float %b.nnan)
+  %tmp1 = call float @llvm.maxnum.f32(float %a.nnan, float %b.nnan)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c.nnan)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_med3_f32_pat0_srcmod0_mismatch(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0_mismatch:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_sub_f32_e32 v5, 0x80000000, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_min_f32_e32 v5, v5, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    v_max_f32_e32 v2, v5, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0_mismatch:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v7, v[0:1]
+; VI-NEXT:    flat_load_dword v2, v[2:3]
+; VI-NEXT:    flat_load_dword v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_sub_f32_e32 v4, 0x80000000, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_max_f32_e32 v5, v7, v2
+; VI-NEXT:    v_min_f32_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_min_f32_e32 v3, v5, v3
+; VI-NEXT:    v_max_f32_e32 v2, v2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_med3_f32_pat0_srcmod0_mismatch:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_sub_f32_e32 v4, 0x80000000, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    v_min_f32_e32 v2, v4, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %a.fneg = fsub float -0.0, %a
+  %tmp0 = call float @llvm.minnum.f32(float %a.fneg, float %b)
+  %tmp1 = call float @llvm.maxnum.f32(float %a, float %b)
+  %tmp2 = call float @llvm.minnum.f32(float %tmp1, float %c)
+  %med3 = call float @llvm.maxnum.f32(float %tmp0, float %tmp2)
+  store float %med3, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_global_nnans_min_max_f32(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr, float addrspace(1)* %cptr) #2 {
+; SI-LABEL: v_test_global_nnans_min_max_f32:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_global_nnans_min_max_f32:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v8, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v8
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v8
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v8
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    flat_load_dword v1, v[2:3]
+; VI-NEXT:    flat_load_dword v2, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v7, s1
+; VI-NEXT:    v_mov_b32_e32 v6, s0
+; VI-NEXT:    v_add_u32_e32 v6, vcc, v6, v8
+; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_max_f32_e32 v0, v0, v1
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_min_f32_e32 v0, v0, v2
+; VI-NEXT:    flat_store_dword v[6:7], v0
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_global_nnans_min_max_f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load volatile float, float addrspace(1)* %gep0
+  %b = load volatile float, float addrspace(1)* %gep1
+  %c = load volatile float, float addrspace(1)* %gep2
+  %max = call float @llvm.maxnum.f32(float %a, float %b)
+  %minmax = call float @llvm.minnum.f32(float %max, float %c)
+  store float %minmax, float addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_test_nnan_input_fmed3_r_i_i_f16(half addrspace(1)* %out, half addrspace(1)* %aptr) #1 {
+; SI-LABEL: v_test_nnan_input_fmed3_r_i_i_f16:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; SI-NEXT:    buffer_load_ushort v2, v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, 1.0
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, 2.0
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_add_f32_e32 v2, v2, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_max_f32_e32 v2, v2, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, 4.0
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[6:7]
+; SI-NEXT:    buffer_store_short v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_test_nnan_input_fmed3_r_i_i_f16:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v2, 1, v0
+; VI-NEXT:    v_max_f16_e64 v4, 2.0, 2.0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_ushort v3, v[0:1]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f16_e32 v2, 1.0, v3
+; VI-NEXT:    v_max_f16_e32 v2, v2, v4
+; VI-NEXT:    v_max_f16_e32 v2, v2, v2
+; VI-NEXT:    v_max_f16_e64 v3, 4.0, 4.0
+; VI-NEXT:    v_min_f16_e32 v2, v2, v3
+; VI-NEXT:    flat_store_short v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_test_nnan_input_fmed3_r_i_i_f16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
+; GFX9-NEXT:    v_max_f16_e64 v2, 2.0, 2.0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_ushort v1, v0, s[2:3]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f16_e32 v1, 1.0, v1
+; GFX9-NEXT:    v_max_f16_e32 v1, v1, v2
+; GFX9-NEXT:    v_max_f16_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f16_e64 v2, 4.0, 4.0
+; GFX9-NEXT:    v_min_f16_e32 v1, v1, v2
+; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
+  %outgep = getelementptr half, half addrspace(1)* %out, i32 %tid
+  %a = load half, half addrspace(1)* %gep0
+  %a.add = fadd nnan half %a, 1.0
+  %max = call half @llvm.maxnum.f16(half %a.add, half 2.0)
+  %med = call half @llvm.minnum.f16(half %max, half 4.0)
+
+  store half %med, half addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @v_nnan_inputs_med3_f16_pat0(half addrspace(1)* %out, half addrspace(1)* %aptr, half addrspace(1)* %bptr, half addrspace(1)* %cptr) #1 {
+; SI-LABEL: v_nnan_inputs_med3_f16_pat0:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0xf000
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, 1.0
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, 2.0
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; SI-NEXT:    buffer_load_ushort v4, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    v_cvt_f32_f16_e32 v5, 4.0
+; SI-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; SI-NEXT:    buffer_load_ushort v6, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; SI-NEXT:    buffer_load_ushort v7, v[0:1], s[8:11], 0 addr64
+; SI-NEXT:    s_waitcnt vmcnt(2)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
+; SI-NEXT:    v_add_f32_e32 v2, v4, v2
+; SI-NEXT:    s_waitcnt vmcnt(1)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v6
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_add_f32_e32 v3, v4, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_cvt_f32_f16_e32 v4, v7
+; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; SI-NEXT:    v_add_f32_e32 v4, v4, v5
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; SI-NEXT:    v_min_f32_e32 v5, v2, v3
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v4
+; SI-NEXT:    v_cvt_f16_f32_e32 v4, v5
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    v_cvt_f32_f16_e32 v3, v4
+; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
+; SI-NEXT:    buffer_store_short v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: v_nnan_inputs_med3_f16_pat0:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v6, 1, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_mov_b32_e32 v3, s5
+; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v6
+; VI-NEXT:    v_mov_b32_e32 v4, s6
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; VI-NEXT:    v_mov_b32_e32 v5, s7
+; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v6
+; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
+; VI-NEXT:    flat_load_ushort v7, v[0:1]
+; VI-NEXT:    flat_load_ushort v2, v[2:3]
+; VI-NEXT:    flat_load_ushort v3, v[4:5]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
+; VI-NEXT:    v_add_f16_e32 v4, 1.0, v7
+; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
+; VI-NEXT:    v_add_f16_e32 v2, 2.0, v2
+; VI-NEXT:    v_min_f16_e32 v5, v4, v2
+; VI-NEXT:    v_max_f16_e32 v2, v4, v2
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f16_e32 v3, 4.0, v3
+; VI-NEXT:    v_max_f16_e32 v2, v2, v2
+; VI-NEXT:    v_min_f16_e32 v2, v2, v3
+; VI-NEXT:    v_max_f16_e32 v3, v5, v5
+; VI-NEXT:    v_max_f16_e32 v2, v2, v2
+; VI-NEXT:    v_max_f16_e32 v2, v3, v2
+; VI-NEXT:    flat_store_short v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: v_nnan_inputs_med3_f16_pat0:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_ushort v1, v0, s[2:3]
+; GFX9-NEXT:    global_load_ushort v2, v0, s[4:5]
+; GFX9-NEXT:    global_load_ushort v3, v0, s[6:7]
+; GFX9-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NEXT:    v_add_f16_e32 v1, 1.0, v1
+; GFX9-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NEXT:    v_add_f16_e32 v2, 2.0, v2
+; GFX9-NEXT:    v_min_f16_e32 v4, v1, v2
+; GFX9-NEXT:    v_max_f16_e32 v1, v1, v2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f16_e32 v3, 4.0, v3
+; GFX9-NEXT:    v_max_f16_e32 v1, v1, v1
+; GFX9-NEXT:    v_min_f16_e32 v1, v1, v3
+; GFX9-NEXT:    v_max_f16_e32 v2, v4, v4
+; GFX9-NEXT:    v_max_f16_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f16_e32 v1, v2, v1
+; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
+  %gep1 = getelementptr half, half addrspace(1)* %bptr, i32 %tid
+  %gep2 = getelementptr half, half addrspace(1)* %cptr, i32 %tid
+  %outgep = getelementptr half, half addrspace(1)* %out, i32 %tid
+  %a = load volatile half, half addrspace(1)* %gep0
+  %b = load volatile half, half addrspace(1)* %gep1
+  %c = load volatile half, half addrspace(1)* %gep2
+
+  %a.nnan = fadd nnan half %a, 1.0
+  %b.nnan = fadd nnan half %b, 2.0
+  %c.nnan = fadd nnan half %c, 4.0
+
+  %tmp0 = call half @llvm.minnum.f16(half %a.nnan, half %b.nnan)
+  %tmp1 = call half @llvm.maxnum.f16(half %a.nnan, half %b.nnan)
+  %tmp2 = call half @llvm.minnum.f16(half %tmp1, half %c.nnan)
+  %med3 = call half @llvm.maxnum.f16(half %tmp0, half %tmp2)
+  store half %med3, half addrspace(1)* %outgep
+  ret void
+}
+
+define amdgpu_kernel void @two_non_inline_constant(float addrspace(1)* %out, float addrspace(1)* %aptr) #1 {
+; SI-LABEL: two_non_inline_constant:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    v_mov_b32_e32 v3, 0x41000000
+; SI-NEXT:    v_mov_b32_e32 v4, 0x41800000
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v2, 0.5, v2
+; SI-NEXT:    v_max_f32_e32 v2, v2, v3
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; SI-NEXT:    v_min_f32_e32 v2, v2, v3
+; SI-NEXT:    s_mov_b64 s[2:3], s[6:7]
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: two_non_inline_constant:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; VI-NEXT:    v_mov_b32_e32 v4, 0x41800000
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_dword v3, v[0:1]
+; VI-NEXT:    v_mov_b32_e32 v0, 0x41000000
+; VI-NEXT:    v_mul_f32_e32 v5, 1.0, v0
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v2, 0.5, v3
+; VI-NEXT:    v_max_f32_e32 v2, v2, v5
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v4
+; VI-NEXT:    v_min_f32_e32 v2, v2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: two_non_inline_constant:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x41000000
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_mov_b32_e32 v3, 0x41800000
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v1, 0.5, v1
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
+; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
+; GFX9-NEXT:    v_max_f32_e32 v2, v3, v3
+; GFX9-NEXT:    v_min_f32_e32 v1, v1, v2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %add = fadd nnan float %a, 0.5
+  %max = call float @llvm.maxnum.f32(float %add, float 8.0)
+  %med = call float @llvm.minnum.f32(float %max, float 16.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; FIXME: Simple stores do not work as a multiple use because they are bitcasted to integer constants.
+define amdgpu_kernel void @one_non_inline_constant(float addrspace(1)* %out, float addrspace(1)* %aptr) #1 {
+; SI-LABEL: one_non_inline_constant:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    v_mul_f32_e64 v2, 1.0, 1.0
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; SI-NEXT:    buffer_load_dword v3, v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    s_mov_b32 s4, 0x41800000
+; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s4
+; SI-NEXT:    s_mov_b64 s[2:3], s[6:7]
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v5, 0.5, v3
+; SI-NEXT:    v_add_f32_e32 v3, s4, v3
+; SI-NEXT:    v_max_f32_e32 v2, v5, v2
+; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; SI-NEXT:    v_min_f32_e32 v2, v2, v4
+; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    buffer_store_dword v3, off, s[4:7], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: one_non_inline_constant:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; VI-NEXT:    v_mul_f32_e64 v4, 1.0, 1.0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_dword v3, v[0:1]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    s_mov_b32 s2, 0x41800000
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_mul_f32_e64 v5, 1.0, s2
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v2, 0.5, v3
+; VI-NEXT:    v_max_f32_e32 v2, v2, v4
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v5
+; VI-NEXT:    v_add_f32_e32 v3, s2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    flat_store_dword v[0:1], v3
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: one_non_inline_constant:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    v_max_f32_e64 v2, 1.0, 1.0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    s_mov_b32 s2, 0x41800000
+; GFX9-NEXT:    v_max_f32_e64 v3, s2, s2
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v4, 0.5, v1
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v2
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_add_f32_e32 v1, s2, v1
+; GFX9-NEXT:    v_min_f32_e32 v2, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v2, s[0:1]
+; GFX9-NEXT:    global_store_dword v[0:1], v1, off
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %add = fadd nnan float %a, 0.5
+  %max = call float @llvm.maxnum.f32(float %add, float 1.0)
+  %med = call float @llvm.minnum.f32(float %max, float 16.0)
+
+  store float %med, float addrspace(1)* %out.gep
+
+  %extra.use = fadd float %a, 16.0
+  store volatile float %extra.use, float addrspace(1)* undef
+  ret void
+}
+
+define amdgpu_kernel void @two_non_inline_constant_multi_use(float addrspace(1)* %out, float addrspace(1)* %aptr) #1 {
+; SI-LABEL: two_non_inline_constant_multi_use:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SI-NEXT:    v_mov_b32_e32 v1, 0
+; SI-NEXT:    s_mov_b32 s6, 0
+; SI-NEXT:    s_mov_b32 s7, 0xf000
+; SI-NEXT:    s_mov_b32 s8, 0x41000000
+; SI-NEXT:    s_mov_b32 s9, 0x41800000
+; SI-NEXT:    s_waitcnt lgkmcnt(0)
+; SI-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; SI-NEXT:    buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
+; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s8
+; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s9
+; SI-NEXT:    s_mov_b64 s[2:3], s[6:7]
+; SI-NEXT:    s_mov_b32 s6, -1
+; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_add_f32_e32 v5, 0.5, v2
+; SI-NEXT:    v_add_f32_e32 v6, s9, v2
+; SI-NEXT:    v_add_f32_e32 v2, s8, v2
+; SI-NEXT:    v_max_f32_e32 v3, v5, v3
+; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
+; SI-NEXT:    v_min_f32_e32 v3, v3, v4
+; SI-NEXT:    buffer_store_dword v3, v[0:1], s[0:3], 0 addr64
+; SI-NEXT:    buffer_store_dword v6, off, s[4:7], 0
+; SI-NEXT:    buffer_store_dword v2, off, s[4:7], 0
+; SI-NEXT:    s_endpgm
+;
+; VI-LABEL: two_non_inline_constant_multi_use:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; VI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; VI-NEXT:    s_waitcnt lgkmcnt(0)
+; VI-NEXT:    v_mov_b32_e32 v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s3
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_dword v3, v[0:1]
+; VI-NEXT:    v_mov_b32_e32 v0, s0
+; VI-NEXT:    s_mov_b32 s2, 0x41000000
+; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v2
+; VI-NEXT:    v_mul_f32_e64 v4, 1.0, s2
+; VI-NEXT:    s_mov_b32 s3, 0x41800000
+; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_mul_f32_e64 v5, 1.0, s3
+; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_add_f32_e32 v2, 0.5, v3
+; VI-NEXT:    v_max_f32_e32 v2, v2, v4
+; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
+; VI-NEXT:    v_min_f32_e32 v2, v2, v5
+; VI-NEXT:    v_add_f32_e32 v6, s3, v3
+; VI-NEXT:    v_add_f32_e32 v3, s2, v3
+; VI-NEXT:    flat_store_dword v[0:1], v2
+; VI-NEXT:    flat_store_dword v[0:1], v6
+; VI-NEXT:    flat_store_dword v[0:1], v3
+; VI-NEXT:    s_endpgm
+;
+; GFX9-LABEL: two_non_inline_constant_multi_use:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NEXT:    s_mov_b32 s2, 0x41000000
+; GFX9-NEXT:    v_max_f32_e64 v2, s2, s2
+; GFX9-NEXT:    s_mov_b32 s3, 0x41800000
+; GFX9-NEXT:    v_max_f32_e64 v3, s3, s3
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    v_add_f32_e32 v4, 0.5, v1
+; GFX9-NEXT:    v_max_f32_e32 v2, v4, v2
+; GFX9-NEXT:    v_max_f32_e32 v2, v2, v2
+; GFX9-NEXT:    v_add_f32_e32 v5, s3, v1
+; GFX9-NEXT:    v_add_f32_e32 v1, s2, v1
+; GFX9-NEXT:    v_min_f32_e32 v2, v2, v3
+; GFX9-NEXT:    global_store_dword v0, v2, s[0:1]
+; GFX9-NEXT:    global_store_dword v[0:1], v5, off
+; GFX9-NEXT:    global_store_dword v[0:1], v1, off
+; GFX9-NEXT:    s_endpgm
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %add = fadd nnan float %a, 0.5
+  %max = call float @llvm.maxnum.f32(float %add, float 8.0)
+  %med = call float @llvm.minnum.f32(float %max, float 16.0)
+
+  store float %med, float addrspace(1)* %out.gep
+
+  %extra.use0 = fadd float %a, 16.0
+  store volatile float %extra.use0, float addrspace(1)* undef
+  %extra.use1 = fadd float %a, 8.0
+  store volatile float %extra.use1, float addrspace(1)* undef
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x() #0
+declare float @llvm.fabs.f32(float) #0
+declare float @llvm.minnum.f32(float, float) #0
+declare float @llvm.maxnum.f32(float, float) #0
+declare double @llvm.minnum.f64(double, double) #0
+declare double @llvm.maxnum.f64(double, double) #0
+declare half @llvm.fabs.f16(half) #0
+declare half @llvm.minnum.f16(half, half) #0
+declare half @llvm.maxnum.f16(half, half) #0
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind "unsafe-fp-math"="false" "no-nans-fp-math"="false" }
+attributes #2 = { nounwind "unsafe-fp-math"="false" "no-nans-fp-math"="true" }

llvm/test/CodeGen/AMDGPU/GlobalISel/frem.ll

 ; CI:       ; %bb.0:
 ; 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_mov_b32 s10, -1
 ; CI-NEXT:    s_mov_b32 s11, 0xf000
 ; CI-NEXT:    s_mov_b64 s[2:3], s[10:11]
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
 ; CI-NEXT:    s_mov_b64 s[8:9], s[6:7]
-; CI-NEXT:    buffer_load_ushort v0, off, s[8:11], 0
 ; CI-NEXT:    buffer_load_ushort v1, off, s[0:3], 0 offset:8
+; CI-NEXT:    buffer_load_ushort v0, off, s[8:11], 0
+; CI-NEXT:    s_mov_b64 s[6:7], s[10:11]
 ; CI-NEXT:    s_waitcnt vmcnt(1)
-; CI-NEXT:    v_cvt_f32_f16_e32 v0, v0
-; CI-NEXT:    s_waitcnt vmcnt(0)
 ; CI-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; CI-NEXT:    v_div_scale_f32 v2, s[0:1], v1, v1, v0
-; CI-NEXT:    v_div_scale_f32 v3, vcc, v0, v1, v0
-; CI-NEXT:    v_rcp_f32_e32 v4, v2
-; CI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
-; CI-NEXT:    v_fma_f32 v5, -v2, v4, 1.0
-; CI-NEXT:    v_fma_f32 v4, v5, v4, v4
-; CI-NEXT:    v_mul_f32_e32 v5, v3, v4
-; CI-NEXT:    v_fma_f32 v6, -v2, v5, v3
-; CI-NEXT:    v_fma_f32 v5, v6, v4, v5
-; CI-NEXT:    v_fma_f32 v2, -v2, v5, v3
-; CI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
-; CI-NEXT:    v_div_fmas_f32 v2, v2, v4, v5
-; CI-NEXT:    s_mov_b64 s[6:7], s[10:11]
-; CI-NEXT:    v_div_fixup_f32 v2, v2, v1, v0
+; CI-NEXT:    s_waitcnt vmcnt(0)
+; CI-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; CI-NEXT:    v_rcp_f32_e32 v2, v1
+; CI-NEXT:    v_mul_f32_e32 v2, v0, v2
 ; CI-NEXT:    v_trunc_f32_e32 v2, v2
 ; CI-NEXT:    v_fma_f32 v0, -v2, v1, v0
-; CI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0
 ; CI-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; CI-NEXT:    buffer_store_short v0, off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: unsafe_frem_f16:
 ; VI:       ; %bb.0:
 ; 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_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v0, s6
 ; VI-NEXT:    s_add_u32 s0, s8, 8
 ; VI-NEXT:    v_mov_b32_e32 v1, s7
 ; VI-NEXT:    s_addc_u32 s1, s9, 0
 ; VI-NEXT:    flat_load_ushort v2, v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_load_ushort v0, v[0:1]
-; VI-NEXT:    s_waitcnt vmcnt(1) lgkmcnt(1)
-; VI-NEXT:    v_cvt_f32_f16_e32 v1, v2
 ; VI-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; VI-NEXT:    v_cvt_f32_f16_e32 v3, v0
-; VI-NEXT:    v_rcp_f32_e32 v3, v3
-; VI-NEXT:    v_mul_f32_e32 v1, v1, v3
-; VI-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; VI-NEXT:    v_div_fixup_f16 v1, v1, v0, v2
+; VI-NEXT:    v_rcp_f16_e32 v1, v0
+; VI-NEXT:    v_mul_f16_e32 v1, v2, v1
 ; VI-NEXT:    v_trunc_f16_e32 v1, v1
 ; VI-NEXT:    v_fma_f16 v2, -v1, v0, v2
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
 ; VI-NEXT:    v_mov_b32_e32 v1, s5
 ; VI-NEXT:    flat_store_short v[0:1], v2
 ; VI-NEXT:    s_endpgm
    %gep2 = getelementptr half, half addrspace(1)* %in2, i32 4
    %r0 = load half, half addrspace(1)* %in1, align 4

 }
 
 define amdgpu_kernel void @unsafe_frem_f32(float addrspace(1)* %out, float addrspace(1)* %in1, float addrspace(1)* %in2) #1 {
 ; CI-LABEL: unsafe_frem_f32:
 ; CI:       ; %bb.0:
 ; 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_waitcnt lgkmcnt(0)
-; CI-NEXT:    s_load_dword s2, s[6:7], 0x0
-; CI-NEXT:    s_load_dword s0, s[8:9], 0x4
-; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s0
-; CI-NEXT:    v_div_scale_f32 v1, s[0:1], v0, v0, s2
-; CI-NEXT:    v_div_scale_f32 v2, vcc, s2, v0, s2
-; CI-NEXT:    v_rcp_f32_e32 v3, v1
-; CI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
-; CI-NEXT:    v_fma_f32 v4, -v1, v3, 1.0
-; CI-NEXT:    v_fma_f32 v3, v4, v3, v3
-; CI-NEXT:    v_mul_f32_e32 v4, v2, v3
-; CI-NEXT:    v_fma_f32 v5, -v1, v4, v2
-; CI-NEXT:    v_fma_f32 v4, v5, v3, v4
-; CI-NEXT:    v_fma_f32 v1, -v1, v4, v2
-; CI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
-; CI-NEXT:    v_div_fmas_f32 v1, v1, v3, v4
+; CI-NEXT:    s_load_dword s0, s[6:7], 0x0
+; CI-NEXT:    s_load_dword s1, s[8:9], 0x4
 ; CI-NEXT:    s_mov_b32 s6, -1
 ; CI-NEXT:    s_mov_b32 s7, 0xf000
-; CI-NEXT:    v_div_fixup_f32 v1, v1, v0, s2
-; CI-NEXT:    v_trunc_f32_e32 v1, v1
-; CI-NEXT:    v_fma_f32 v0, -v1, v0, s2
+; CI-NEXT:    s_waitcnt lgkmcnt(0)
+; CI-NEXT:    v_mov_b32_e32 v1, s0
+; CI-NEXT:    v_rcp_f32_e32 v0, s1
+; CI-NEXT:    v_mul_f32_e32 v0, s0, v0
+; CI-NEXT:    v_trunc_f32_e32 v0, v0
+; CI-NEXT:    v_fma_f32 v0, -v0, s1, v1
 ; CI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: unsafe_frem_f32:
 ; VI:       ; %bb.0:
 ; 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_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_load_dword s2, s[6:7], 0x0
-; VI-NEXT:    s_load_dword s0, s[8:9], 0x10
+; VI-NEXT:    s_load_dword s0, s[6:7], 0x0
+; VI-NEXT:    s_load_dword s1, s[8:9], 0x10
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_div_scale_f32 v1, s[0:1], v0, v0, s2
-; VI-NEXT:    v_div_scale_f32 v2, vcc, s2, v0, s2
-; VI-NEXT:    v_rcp_f32_e32 v3, v1
-; VI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3
-; VI-NEXT:    v_fma_f32 v4, -v1, v3, 1.0
-; VI-NEXT:    v_fma_f32 v3, v4, v3, v3
-; VI-NEXT:    v_mul_f32_e32 v4, v2, v3
-; VI-NEXT:    v_fma_f32 v5, -v1, v4, v2
-; VI-NEXT:    v_fma_f32 v4, v5, v3, v4
-; VI-NEXT:    v_fma_f32 v1, -v1, v4, v2
-; VI-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0
-; VI-NEXT:    v_div_fmas_f32 v1, v1, v3, v4
-; VI-NEXT:    v_div_fixup_f32 v1, v1, v0, s2
-; VI-NEXT:    v_trunc_f32_e32 v1, v1
-; VI-NEXT:    v_fma_f32 v2, -v1, v0, s2
+; VI-NEXT:    v_mov_b32_e32 v1, s0
+; VI-NEXT:    v_rcp_f32_e32 v0, s1
+; VI-NEXT:    v_mul_f32_e32 v0, s0, v0
+; VI-NEXT:    v_trunc_f32_e32 v0, v0
+; VI-NEXT:    v_fma_f32 v2, -v0, s1, v1
 ; VI-NEXT:    v_mov_b32_e32 v0, s4
 ; VI-NEXT:    v_mov_b32_e32 v1, s5
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
    %gep2 = getelementptr float, float addrspace(1)* %in2, i32 4
    %r0 = load float, float addrspace(1)* %in1, align 4
    %r1 = load float, float addrspace(1)* %gep2, align 4
    %r2 = frem float %r0, %r1

 ; CI-LABEL: unsafe_frem_f64:
 ; CI:       ; %bb.0:
 ; 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_waitcnt lgkmcnt(0)
 ; CI-NEXT:    s_load_dwordx2 s[0:1], s[6:7], 0x0
 ; CI-NEXT:    s_load_dwordx2 s[2:3], s[8:9], 0x0
 ; CI-NEXT:    s_waitcnt lgkmcnt(0)
-; CI-NEXT:    v_mov_b32_e32 v0, s2
-; CI-NEXT:    v_mov_b32_e32 v1, s3
-; CI-NEXT:    v_div_scale_f64 v[2:3], s[2:3], v[0:1], v[0:1], s[0:1]
-; CI-NEXT:    v_div_scale_f64 v[8:9], vcc, s[0:1], v[0:1], s[0:1]
-; CI-NEXT:    v_rcp_f64_e32 v[4:5], v[2:3]
-; CI-NEXT:    v_fma_f64 v[6:7], -v[2:3], v[4:5], 1.0
-; CI-NEXT:    v_fma_f64 v[4:5], v[4:5], v[6:7], v[4:5]
-; CI-NEXT:    v_fma_f64 v[6:7], -v[2:3], v[4:5], 1.0
-; CI-NEXT:    v_fma_f64 v[4:5], v[4:5], v[6:7], v[4:5]
-; CI-NEXT:    v_mul_f64 v[6:7], v[8:9], v[4:5]
-; CI-NEXT:    v_fma_f64 v[2:3], -v[2:3], v[6:7], v[8:9]
-; CI-NEXT:    v_div_fmas_f64 v[2:3], v[2:3], v[4:5], v[6:7]
-; CI-NEXT:    v_div_fixup_f64 v[2:3], v[2:3], v[0:1], s[0:1]
-; CI-NEXT:    v_trunc_f64_e32 v[2:3], v[2:3]
-; CI-NEXT:    v_fma_f64 v[0:1], -v[2:3], v[0:1], s[0:1]
+; CI-NEXT:    v_mov_b32_e32 v3, s1
+; CI-NEXT:    v_rcp_f64_e32 v[0:1], s[2:3]
+; CI-NEXT:    v_mov_b32_e32 v2, s0
+; CI-NEXT:    v_mul_f64 v[0:1], s[0:1], v[0:1]
+; CI-NEXT:    v_trunc_f64_e32 v[0:1], v[0:1]
+; CI-NEXT:    v_fma_f64 v[0:1], -v[0:1], s[2:3], v[2:3]
 ; CI-NEXT:    v_mov_b32_e32 v2, s4
 ; CI-NEXT:    v_mov_b32_e32 v3, s5
 ; CI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; CI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: unsafe_frem_f64:
 ; VI:       ; %bb.0:
 ; 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_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_load_dwordx2 s[0:1], s[6:7], 0x0
 ; VI-NEXT:    s_load_dwordx2 s[2:3], s[8:9], 0x0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    v_mov_b32_e32 v0, s2
-; VI-NEXT:    v_mov_b32_e32 v1, s3
-; VI-NEXT:    v_div_scale_f64 v[2:3], s[2:3], v[0:1], v[0:1], s[0:1]
-; VI-NEXT:    v_div_scale_f64 v[8:9], vcc, s[0:1], v[0:1], s[0:1]
-; VI-NEXT:    v_rcp_f64_e32 v[4:5], v[2:3]
-; VI-NEXT:    v_fma_f64 v[6:7], -v[2:3], v[4:5], 1.0
-; VI-NEXT:    v_fma_f64 v[4:5], v[4:5], v[6:7], v[4:5]
-; VI-NEXT:    v_fma_f64 v[6:7], -v[2:3], v[4:5], 1.0
-; VI-NEXT:    v_fma_f64 v[4:5], v[4:5], v[6:7], v[4:5]
-; VI-NEXT:    v_mul_f64 v[6:7], v[8:9], v[4:5]
-; VI-NEXT:    v_fma_f64 v[2:3], -v[2:3], v[6:7], v[8:9]
-; VI-NEXT:    v_div_fmas_f64 v[2:3], v[2:3], v[4:5], v[6:7]
-; VI-NEXT:    v_div_fixup_f64 v[2:3], v[2:3], v[0:1], s[0:1]
-; VI-NEXT:    v_trunc_f64_e32 v[2:3], v[2:3]
-; VI-NEXT:    v_fma_f64 v[0:1], -v[2:3], v[0:1], s[0:1]
+; VI-NEXT:    v_mov_b32_e32 v3, s1
+; VI-NEXT:    v_rcp_f64_e32 v[0:1], s[2:3]
+; VI-NEXT:    v_mov_b32_e32 v2, s0
+; VI-NEXT:    v_mul_f64 v[0:1], s[0:1], v[0:1]
+; VI-NEXT:    v_trunc_f64_e32 v[0:1], v[0:1]
+; VI-NEXT:    v_fma_f64 v[0:1], -v[0:1], s[2:3], v[2:3]
 ; VI-NEXT:    v_mov_b32_e32 v2, s4
 ; VI-NEXT:    v_mov_b32_e32 v3, s5
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
                              double addrspace(1)* %in2) #1 {
    %r0 = load double, double addrspace(1)* %in1, align 8
    %r1 = load double, double addrspace(1)* %in2, align 8
    %r2 = frem double %r0, %r1