[AMDGPU] Avoid splitting FLAT offsets in unsafe ways

[foad]

Summary

As explained in the comment:

For a FLAT instruction the hardware decides whether to access
global/scratch/shared memory based on the high bits of vaddr,
ignoring the offset field, so we have to ensure that when we add
remainder to vaddr it still points into the same underlying object.
The easiest way to do that is to make sure that we split the offset
into two pieces that are both >= 0 or both <= 0.

In particular FLAT (as opposed to SCRATCH and GLOBAL) instructions have
an unsigned immediate offset field, so we can't use it to help split a
negative offset.

[arsenm]

The mirror change is needed for globalisel

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

This limitation also only needs to be applied if AS == FLAT_ADDRESS

[foad]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

The only "limitation" is that we don't try to split negative offsets if the immediate offset field is unsigned, but you're saying we can do that if AS != FLAT_ADDRESS? What would that mean - that we're using a FLAT instruction but we know statically which part of the address space it is accessing??

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

Correct. This is always the case pre-gfx9 which did not have the "global" flat instructions

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

Actually pre-gfx9 also didn't have flat offsets. However gfx10 does have a bug with flat offsets, so I think it would still be correct to model this correctly. The instruction patterns do accept either (and global instructions are only preferred through pattern priority)

[foad]

The mirror change is needed for globalisel

AMDGPUInstructionSelector::selectFlatOffsetImpl doesn't attempt to split offsets so I don't think there's anything to fix.

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

This limitation also only needs to be applied if AS == FLAT_ADDRESS

I still don't get this. Surely if we're using a FLAT instruction, even if we know which specific address space the programmer is trying to access, we still have to avoid setting vaddr to an address that might point into the wrong aperture.

[foad]

Rebase.
Fix silly mistake in checking for negative offsets.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

My understanding was the aperture only means anything for private or local. If it's a global address, it's neither aperture and behaves as a normal instruction (i.e. there's no aperture for global pointers)

[foad]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

But in that case, you should still avoid making drastic changes to vaddr in case it ends up accidentally pointing *into* one of the apertures, when you wanted a global access. E.g. if you're accessing a global that happens to be just past the end of the private or local aperture.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

I thought Nicolai mentioned this might not be possible?

I guess I don't understand why the aperture is so complicated and not just a couple of flags in the high, unused bits

[sameerds]

In D83394#2142277, @foad wrote:

Rebase.
Fix silly mistake in checking for negative offsets.

It's hard to see through the rebase, but did fixing the negative offset check add more tests? I assuming that the tests in the original patch did not capture this mistake, so it should warrant a new test.

[foad]

In D83394#2143122, @sameerds wrote:

In D83394#2142277, @foad wrote:

Rebase.
Fix silly mistake in checking for negative offsets.

It's hard to see through the rebase, but did fixing the negative offset check add more tests? I assuming that the tests in the original patch did not capture this mistake, so it should warrant a new test.

There are no new tests. All the testing comes from staring at the changes in offset-split-flat.ll and offset-split-global.ll.

In retrospect I should have noticed that there was something wrong with the original patch, because it didn't cause any changes in offset-split-flat.ll.

[foad]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

I don't remember hearing that it wasn't possible. @nhaehnle ?

[foad]

Ping.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

maskTrailingZeros

[foad]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1705

That would round towards -infinity. As the comment says, we're deliberately rounding towards zero instead.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1698

Can you add a fixme to check if this is needed if we know the address isn't FLAT_ADDRESS?

[nhaehnle]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1698

Yes, I'm pretty sure it's needed. If we generate a FLAT instruction, the hardware will check apertures regardless, and you could run into the following scenario:

end_of_scratch_aperture: 0x2'0000'0000
vaddr: 0x1'ffff'fff0
inst_offset: 32

The hardware executes the flat instruction, sees that vaddr falls into the scratch aperture, and executes the instruction as accessing scratch memory.

The only way to fix that would be if we knew that there is never any memory mapped directly after the end of any of the apertures.

[JonChesterfield]

I'm seeing (difficult to minimise) failure modes with gfx10 that look the same as those on gfx9 before this patch. I also see the comment:

"gfx10 does have a bug with flat offsets"

Is there a corresponding patch needed for gfx10?

[foad]

In D83394#2305735, @JonChesterfield wrote:

I'm seeing (difficult to minimise) failure modes with gfx10 that look the same as those on gfx9 before this patch. I also see the comment:

"gfx10 does have a bug with flat offsets"

Is there a corresponding patch needed for gfx10?

This patch affected gfx10 just as much as gfx9, so there is no extra patch required for gfx10.

I am confused about the gfx10 bug. I assume it refers to this:

def FeatureFlatSegmentOffsetBug : SubtargetFeature<"flat-segment-offset-bug",
  "HasFlatSegmentOffsetBug",
  "true",
  "GFX10 bug, inst_offset ignored in flat segment"
>;

But (a) I think this is just the intended behaviour of the hardware, so I wouldn't call it a bug, and (b) I think gfx9 works the same way, otherwise there would have been no need for this patch in the first place!

Update: actually I think there is a gfx10-only hardware bug, but the description of that subtarget feature doesn't describe it very well.

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

if (N0 && N1) {
       const SIInstrInfo *TII = Subtarget->getInstrInfo();
       unsigned AS = findMemSDNode(N)->getAddressSpace();
       if (TII->isLegalFLATOffset(COffsetVal, AS, IsSigned)) {
         Addr = N0;
         OffsetVal = COffsetVal;
       } else {
         // If the offset doesn't fit, put the low bits into the offset field and
         // add the rest.
+        //
+        // For a FLAT instruction the hardware decides whether to access
+        // global/scratch/shared memory based on the high bits of vaddr,
+        // ignoring the offset field, so we have to ensure that when we add
+        // remainder to vaddr it still points into the same underlying object.
+        // The easiest way to do that is to make sure that we split the offset
+        // into two pieces that are both >= 0 or both <= 0.
 
         SDLoc DL(N);
-        uint64_t ImmField;
+        uint64_t RemainderOffset = COffsetVal;
+        uint64_t ImmField = 0;
         const unsigned NumBits = TII->getNumFlatOffsetBits(AS, IsSigned);
         if (IsSigned) {
-          ImmField = SignExtend64(COffsetVal, NumBits);
-
-          // Don't use a negative offset field if the base offset is positive.
-          // Since the scheduler currently relies on the offset field, doing so
-          // could result in strange scheduling decisions.
-
-          // TODO: Should we not do this in the opposite direction as well?
-          if (static_cast<int64_t>(COffsetVal) > 0) {
-            if (static_cast<int64_t>(ImmField) < 0) {
-              const uint64_t OffsetMask =
-                  maskTrailingOnes<uint64_t>(NumBits - 1);
-              ImmField = COffsetVal & OffsetMask;
-            }
-          }
-        } else {
-          // TODO: Should we do this for a negative offset?
-          const uint64_t OffsetMask = maskTrailingOnes<uint64_t>(NumBits);
-          ImmField = COffsetVal & OffsetMask;
-        }
-
-        uint64_t RemainderOffset = COffsetVal - ImmField;
-
+          // Use signed division by a power of two to truncate towards 0.
+          int64_t D = 1LL << (NumBits - 1);
+          RemainderOffset = (static_cast<int64_t>(COffsetVal) / D) * D;
+          ImmField = COffsetVal - RemainderOffset;
+        } else if (static_cast<int64_t>(COffsetVal) >= 0) {
+          ImmField = COffsetVal & maskTrailingOnes<uint64_t>(NumBits);
+          RemainderOffset = COffsetVal - ImmField;
         assert(TII->isLegalFLATOffset(ImmField, AS, IsSigned));
         assert(RemainderOffset + ImmField == COffsetVal);
 
         OffsetVal = ImmField;
 
         // TODO: Should this try to use a scalar add pseudo if the base address
         // is uniform and saddr is usable?
         SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);

llvm/test/CodeGen/AMDGPU/flat-address-space.ll

define amdgpu_kernel void @store_flat_i8_max_offset_p1(i8* %fptr, i8 %x) #0 {
   %fptr.offset = getelementptr inbounds i8, i8* %fptr, i64 4096
   store volatile i8 %x, i8* %fptr.offset
   ret void
 }
 
 ; CHECK-LABEL: {{^}}store_flat_i8_neg_offset:
 ; CIVI: flat_store_byte v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}{{$}}
 
-; GFX9: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff000, v
+; GFX9: v_add_co_u32_e64 v{{[0-9]+}}, vcc, -2, s
 ; GFX9: v_addc_co_u32_e32 v{{[0-9]+}}, vcc, -1,
-; GFX9: flat_store_byte v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset:4094{{$}}
+; GFX9: flat_store_byte v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}{{$}}
 define amdgpu_kernel void @store_flat_i8_neg_offset(i8* %fptr, i8 %x) #0 {
   %fptr.offset = getelementptr inbounds i8, i8* %fptr, i64 -2
   store volatile i8 %x, i8* %fptr.offset
   ret void
 }
 
 ; CHECK-LABEL: {{^}}load_flat_i8_max_offset:
 ; CIVI: flat_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}{{$}}
define amdgpu_kernel void @load_flat_i8_max_offset_p1(i8* %fptr) #0 {
   %fptr.offset = getelementptr inbounds i8, i8* %fptr, i64 4096
   %val = load volatile i8, i8* %fptr.offset
   ret void
 }
 
 ; CHECK-LABEL: {{^}}load_flat_i8_neg_offset:
 ; CIVI: flat_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}{{$}}
 
-; GFX9: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff000, v
+; GFX9: v_add_co_u32_e64 v{{[0-9]+}}, vcc, -2, s
 ; GFX9: v_addc_co_u32_e32 v{{[0-9]+}}, vcc, -1,
-; GFX9: flat_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}} offset:4094{{$}}
+; GFX9: flat_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}{{$}}
 define amdgpu_kernel void @load_flat_i8_neg_offset(i8* %fptr) #0 {
   %fptr.offset = getelementptr inbounds i8, i8* %fptr, i64 -2
   %val = load volatile i8, i8* %fptr.offset
   ret void
 }
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind convergent }

llvm/test/CodeGen/AMDGPU/offset-split-flat.ll

; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %load = load i8, i8* %gep, align 4
   ret i8 %load
 }
 
 define i8 @flat_inst_valu_offset_neg_11bit_max(i8* %p) {
 ; GFX9-LABEL: flat_inst_valu_offset_neg_11bit_max:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfffff000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2048
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: flat_inst_valu_offset_neg_11bit_max:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0xfffff800, v0
; GFX10-NEXT:    s_setpc_b64 s[30:31]
   ret i8 %load
 }
 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2047
 define i8 @flat_inst_valu_offset_64bit_11bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_valu_offset_64bit_11bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x7ff, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2047
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: flat_inst_valu_offset_64bit_11bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x7ff, v0
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
 ; GFX10-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr i8, i8* %p, i64 -9223372036854773761
   %load = load i8, i8* %gep, align 4
   ret i8 %load
 }
 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2048
 define i8 @flat_inst_valu_offset_64bit_11bit_neg_high_split1(i8* %p) {
 ; GFX9-LABEL: flat_inst_valu_offset_64bit_11bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x800, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2048
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: flat_inst_valu_offset_64bit_11bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x800, v0
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
 ; GFX10-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr i8, i8* %p, i64 -9223372036854773760
   %load = load i8, i8* %gep, align 4
   ret i8 %load
 }
 
 ; Fill 12-bit low-bits, negative high bits (1ull << 63) | 4095
 define i8 @flat_inst_valu_offset_64bit_12bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_valu_offset_64bit_12bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfff, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:4095
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: flat_inst_valu_offset_64bit_12bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0xfff, v0
; GFX10-NEXT:    s_setpc_b64 s[30:31]
   ret i8 %load
 }
 
 ; Fill 13-bit low-bits, negative high bits (1ull << 63) | 8191
 define i8 @flat_inst_valu_offset_64bit_13bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_valu_offset_64bit_13bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1fff, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:4095
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: flat_inst_valu_offset_64bit_13bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x1fff, v0

 
 define amdgpu_kernel void @flat_inst_salu_offset_neg_11bit_max(i8* %p) {
 ; GFX9-LABEL: flat_inst_salu_offset_neg_11bit_max:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfffff000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfffff800, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, -1, v1, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2048
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    flat_store_byte v[0:1], v0
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: flat_inst_salu_offset_neg_11bit_max:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2047
 define amdgpu_kernel void @flat_inst_salu_offset_64bit_11bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_salu_offset_64bit_11bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x7ff, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2047
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    flat_store_byte v[0:1], v0
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: flat_inst_salu_offset_64bit_11bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2048
 define amdgpu_kernel void @flat_inst_salu_offset_64bit_11bit_neg_high_split1(i8* %p) {
 ; GFX9-LABEL: flat_inst_salu_offset_64bit_11bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x800, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:2048
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    flat_store_byte v[0:1], v0
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: flat_inst_salu_offset_64bit_11bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

 
 ; Fill 12-bit low-bits, negative high bits (1ull << 63) | 4095
 define amdgpu_kernel void @flat_inst_salu_offset_64bit_12bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_salu_offset_64bit_12bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0xfff, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:4095
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    flat_store_byte v[0:1], v0
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: flat_inst_salu_offset_64bit_12bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

 define amdgpu_kernel void @flat_inst_salu_offset_64bit_13bit_neg_high_split0(i8* %p) {
 ; GFX9-LABEL: flat_inst_salu_offset_64bit_13bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1fff, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    flat_load_ubyte v0, v[0:1] offset:4095
+; GFX9-NEXT:    flat_load_ubyte v0, v[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    flat_store_byte v[0:1], v0
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: flat_inst_salu_offset_64bit_13bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

llvm/test/CodeGen/AMDGPU/offset-split-global.ll

; GFX10-NEXT:    s_setpc_b64 s[30:31]
   ret i8 %load
 }
 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2047
 define i8 @global_inst_valu_offset_64bit_11bit_neg_high_split0(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_valu_offset_64bit_11bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:2047
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2049
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_inst_valu_offset_64bit_11bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0, v0
+; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x800, v0
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
-; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:2047
+; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:-1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr i8, i8 addrspace(1)* %p, i64 -9223372036854773761
   %load = load i8, i8 addrspace(1)* %gep, align 4
   ret i8 %load
 }
 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2048
 define i8 @global_inst_valu_offset_64bit_11bit_neg_high_split1(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_valu_offset_64bit_11bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:2048
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2048
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_inst_valu_offset_64bit_11bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x1000, v0
+; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x800, v0
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
-; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2048
+; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr i8, i8 addrspace(1)* %p, i64 -9223372036854773760
   %load = load i8, i8 addrspace(1)* %gep, align 4
   ret i8 %load
 }
 
 ; Fill 12-bit low-bits, negative high bits (1ull << 63) | 4095
 define i8 @global_inst_valu_offset_64bit_12bit_neg_high_split0(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_valu_offset_64bit_12bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:4095
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_inst_valu_offset_64bit_12bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x1000, v0
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
 ; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:-1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr i8, i8 addrspace(1)* %p, i64 -9223372036854771713
   %load = load i8, i8 addrspace(1)* %gep, align 4
   ret i8 %load
 }
 
 ; Fill 12-bit low-bits, negative high bits (1ull << 63) | 4096
 define i8 @global_inst_valu_offset_64bit_12bit_neg_high_split1(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_valu_offset_64bit_12bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x2000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_bfrev_b32_e32 v2, 1
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-4096
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_inst_valu_offset_64bit_12bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x1000, v0

 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2047
 define amdgpu_kernel void @global_inst_salu_offset_64bit_11bit_neg_high_split0(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_salu_offset_64bit_11bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:2047
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2049
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_inst_salu_offset_64bit_11bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    v_mov_b32_e32 v1, s1
-; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0, s0
+; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x800, s0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
-; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:2047
+; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:-1
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX10-NEXT:    s_endpgm
   %gep = getelementptr i8, i8 addrspace(1)* %p, i64 -9223372036854773761
   %load = load volatile i8, i8 addrspace(1)* %gep, align 1
   store i8 %load, i8 addrspace(1)* undef
   ret void
 }
 
 ; Fill 11-bit low-bits, negative high bits (1ull << 63) | 2048
 define amdgpu_kernel void @global_inst_salu_offset_64bit_11bit_neg_high_split1(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_salu_offset_64bit_11bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:2048
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2048
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_inst_salu_offset_64bit_11bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    v_mov_b32_e32 v1, s1
-; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x1000, s0
+; GFX10-NEXT:    v_add_co_u32_e64 v0, vcc_lo, 0x800, s0
 ; GFX10-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0x80000000, v1, vcc_lo
-; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off offset:-2048
+; GFX10-NEXT:    global_load_ubyte v0, v[0:1], off
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX10-NEXT:    s_endpgm
   %gep = getelementptr i8, i8 addrspace(1)* %p, i64 -9223372036854773760
   %load = load volatile i8, i8 addrspace(1)* %gep, align 1
   store i8 %load, i8 addrspace(1)* undef
   ret void
 }
 
 ; Fill 12-bit low-bits, negative high bits (1ull << 63) | 4095
 define amdgpu_kernel void @global_inst_salu_offset_64bit_12bit_neg_high_split0(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_salu_offset_64bit_12bit_neg_high_split0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e64 v0, vcc, 0, s0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:4095
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-1
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_inst_salu_offset_64bit_12bit_neg_high_split0:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

 define amdgpu_kernel void @global_inst_salu_offset_64bit_12bit_neg_high_split1(i8 addrspace(1)* %p) {
 ; GFX9-LABEL: global_inst_salu_offset_64bit_12bit_neg_high_split1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX9-NEXT:    v_bfrev_b32_e32 v1, 1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s1
-; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x2000, v0
+; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
 ; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v2, vcc
-; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off offset:-4096
+; GFX9-NEXT:    global_load_ubyte v0, v[0:1], off
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    global_store_byte v[0:1], v0, off
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: global_inst_salu_offset_64bit_12bit_neg_high_split1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX10-NEXT:    ; implicit-def: $vcc_hi

llvm/test/CodeGen/AMDGPU/promote-constOffset-to-imm.ll

 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ;
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-4096
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-4096
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:2048
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-4096
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:2048
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off
 ;
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 ; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 entry:
   %call = tail call i64 @_Z13get_global_idj(i32 0)
   %conv = and i64 %call, 255
   %a0 = shl i64 %call, 17
   %idx.ext11 = and i64 %a0, 4261412864
   %add.ptr12 = getelementptr inbounds i8, i8 addrspace(1)* %buffer, i64 %idx.ext11
   %a1 = bitcast i8 addrspace(1)* %add.ptr12 to i64 addrspace(1)*
   %add.ptr6 = getelementptr inbounds i64, i64 addrspace(1)* %a1, i64 %conv
entry:
   ret void
 }
 
 define hidden amdgpu_kernel void @negativeoffset(i8 addrspace(1)* nocapture %buffer) {
 ; GCN-LABEL: negativeoffset:
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ; GFX8:    flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
 ;
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:2048
-; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
+; GFX9:    global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 ;
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off offset:-2048
-; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
+; GFX10:   global_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}], off{{$}}
 entry:
   %call = tail call i64 @_Z13get_global_idj(i32 0) #2
   %conv = and i64 %call, 255
   %0 = shl i64 %call, 7
   %idx.ext11 = and i64 %0, 4294934528
   %add.ptr12 = getelementptr inbounds i8, i8 addrspace(1)* %buffer, i64 %idx.ext11
   %buffer_head = bitcast i8 addrspace(1)* %add.ptr12 to i64 addrspace(1)*
 

llvm/test/CodeGen/AMDGPU/store-hi16.ll

entry:
   ret void
 }
 
 ; GCN-LABEL: {{^}}store_flat_hi_v2i16_neg_offset:
 ; GCN: s_waitcnt
 ; GFX803: v_add{{(_co)?}}_{{i|u}}32_e32
 ; GFX803: v_addc_u32_e32
 
-; GFX9-DAG: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff000, v
+; GFX9-DAG: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff802, v
 ; GFX9-DAG: v_addc_co_u32_e32 v{{[0-9]+}}, vcc, -1, v
 
 ; GFX906-DAG: v_lshrrev_b32_e32
-; GFX906: flat_store_short v[0:1], v2 offset:2050{{$}}
+; GFX906: flat_store_short v[0:1], v2{{$}}
 
-; GFX900-NEXT: flat_store_short_d16_hi v[0:1], v2 offset:2050{{$}}
+; GFX900-NEXT: flat_store_short_d16_hi v[0:1], v2{{$}}
 ; GFX803: flat_store_short v[0:1], v2{{$}}
 ; GCN-NEXT: s_waitcnt
 ; GCN-NEXT: s_setpc_b64
 define void @store_flat_hi_v2i16_neg_offset(i16* %out, i32 %arg) #0 {
 entry:
   %value = bitcast i32 %arg to <2 x i16>
   %hi = extractelement <2 x i16> %value, i32 1
   %gep = getelementptr inbounds i16, i16* %out, i64 -1023

 }
 
 ; GCN-LABEL: {{^}}store_flat_hi_v2i16_i8_neg_offset:
 ; GCN: s_waitcnt
 
 ; GFX803-DAG: v_add_u32_e32
 ; GFX803-DAG: v_addc_u32_e32
 
-; GFX9-DAG: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff000, v
+; GFX9-DAG: v_add_co_u32_e32 v{{[0-9]+}}, vcc, 0xfffff001, v
 ; GFX9-DAG: v_addc_co_u32_e32 v{{[0-9]+}}, vcc, -1, v{{[0-9]+}}, vcc
 
-; GFX900-NEXT: flat_store_byte_d16_hi v[0:1], v2 offset:1{{$}}
+; GFX900-NEXT: flat_store_byte_d16_hi v[0:1], v2{{$}}
 
 ; GFX906-DAG: v_lshrrev_b32_e32 v2, 16, v2
-; GFX906: flat_store_byte v[0:1], v2 offset:1{{$}}
+; GFX906: flat_store_byte v[0:1], v2{{$}}
 
 ; GFX803-DAG: v_lshrrev_b32_e32 v2, 16, v2
 ; GFX803: flat_store_byte v[0:1], v2{{$}}
 
 ; GCN-NEXT: s_waitcnt
 ; GCN-NEXT: s_setpc_b64
 define void @store_flat_hi_v2i16_i8_neg_offset(i8* %out, i32 %arg) #0 {
 entry: