[AMDGPU] Handle SMRD signed offset immediate

[kerbowa]

Summary

This fixes a few issues related to SMRD offsets. On gfx9 and gfx10 we have a
signed byte offset immediate, however we can overflow into a negative since we
treat it as unsigned.

Also, the SMRD SOFFSET sgpr is an unsigned offset on all subtargets. We
sometimes tried to use negative values here.

Third, S_BUFFER instructions should never use a signed offset immediate.

[kerbowa]

Rebase.

[rampitec]

Please follow lint recommendations on formatting.

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

1275–1281

Should we instead pass here Subtarget and distinguish if it has needed mode right in this function?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1788

Should this just rely on getSMRDEncodedSignedOffset failing based on the subtarget?

[kerbowa]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1788

Sure, I'm just following what's done for the LiteralOffset below but that could be improved in the same way.

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

1275–1281

In that case we would also need to pass something to determine whether it's for S_BUFF since there is no signed offset there.

[rampitec]

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

1275–1281

Ah, right.

[kerbowa]

Address comments.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1788

Typo amd

1798

The GlobalISel patch seems to be missing this part?

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

650

Ugh, is this really different between the buffer and non-buffer forms?

[kerbowa]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1798

I think it's handled here?

AMDGPUInstructionSelector.cpp:2876
if (!GEPInfo.Imm || GEPInfo.Imm < 0 || !isUInt<32>(GEPInfo.Imm))

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

650

Apparently, yes. An alternative would be to disallow negative offsets entirely.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

1798

OK, the use of SignBitIsZero is confusing when it's a known constant. This should also directly check the immediate value

[kerbowa]

Directly check the immediate value.

[arsenm]

LGTM, but would like the same comments copied into the GlobalISel part

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

2876

Should copy the comment here about the unsigned distinction for the immediate

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

if (ByteOffsetNode.getOpcode() == ISD::ZERO_EXTEND) {
         Imm = false;
         return true;
       }
     }
     return false;
   }
 
   SDLoc SL(ByteOffsetNode);
-  GCNSubtarget::Generation Gen = Subtarget->getGeneration();
-  uint64_t ByteOffset = C->getZExtValue();
+  // GFX9 amd GFX10 have signed byte immediate offsets.
+  int64_t ByteOffset = C->getSExtValue();
   Optional<int64_t> EncodedOffset =
-      AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset);
+      AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset, false);
   if (EncodedOffset) {
     Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32);
     Imm = true;
     return true;
   }
 
-  if (Gen == AMDGPUSubtarget::SEA_ISLANDS) {
-    EncodedOffset =
-        AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget, ByteOffset);
-    if (EncodedOffset) {
-      Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32);
-      return true;
-    }
-  }
+  if (!CurDAG->SignBitIsZero(ByteOffsetNode))
+    return false;
+
+  EncodedOffset = AMDGPU::getSMRDEncodedLiteralOffset32(*Subtarget, ByteOffset);
 
   if (!isUInt<32>(ByteOffset) && !isInt<32>(ByteOffset))
     return false;
 
   SDValue C32Bit = CurDAG->getTargetConstant(ByteOffset, SL, MVT::i32);
   Offset = SDValue(
       CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SL, MVT::i32, C32Bit), 0);
 
bool AMDGPUDAGToDAGISel::SelectSMRDSgpr(SDValue Addr, SDValue &SBase,
   bool Imm = false;
   return SelectSMRD(Addr, SBase, Offset, Imm) && !Imm &&
          !isa<ConstantSDNode>(Offset);
 }
 
 bool AMDGPUDAGToDAGISel::SelectSMRDBufferImm(SDValue Addr,
                                              SDValue &Offset) const {
   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Addr)) {
-    if (auto Imm = AMDGPU::getSMRDEncodedOffset(*Subtarget,
-                                                C->getZExtValue())) {
+    // The immediate offset for S_BUFFER instructions is unsigned.
+    if (auto Imm =
+            AMDGPU::getSMRDEncodedOffset(*Subtarget, C->getZExtValue(), true)) {
       Offset = CurDAG->getTargetConstant(*Imm, SDLoc(Addr), MVT::i32);
       return true;
     }
   }
 
   return false;
 }
 

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

 AMDGPUInstructionSelector::selectSmrdImm(MachineOperand &Root) const {
   SmallVector<GEPInfo, 4> AddrInfo;
   getAddrModeInfo(*Root.getParent(), *MRI, AddrInfo);
 
   if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
     return None;
 
   const GEPInfo &GEPInfo = AddrInfo[0];
-  Optional<int64_t> EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm);
+  Optional<int64_t> EncodedImm =
+      AMDGPU::getSMRDEncodedOffset(STI, GEPInfo.Imm, false);
   if (!EncodedImm)
     return None;
 
   unsigned PtrReg = GEPInfo.SgprParts[0];
   return {{
     [=](MachineInstrBuilder &MIB) { MIB.addReg(PtrReg); },
     [=](MachineInstrBuilder &MIB) { MIB.addImm(*EncodedImm); }
   }};
AMDGPUInstructionSelector::selectSmrdSgpr(MachineOperand &Root) const {
   getAddrModeInfo(*MI, *MRI, AddrInfo);
 
   // FIXME: We should shrink the GEP if the offset is known to be <= 32-bits,
   // then we can select all ptr + 32-bit offsets not just immediate offsets.
   if (AddrInfo.empty() || AddrInfo[0].SgprParts.size() != 1)
     return None;
 
   const GEPInfo &GEPInfo = AddrInfo[0];
-  if (!GEPInfo.Imm || !isUInt<32>(GEPInfo.Imm))
+  if (!GEPInfo.Imm || GEPInfo.Imm < 0 || !isUInt<32>(GEPInfo.Imm))
     return None;
 
   // If we make it this far we have a load with an 32-bit immediate offset.
   // It is OK to select this using a sgpr offset, because we have already
   // failed trying to select this load into one of the _IMM variants since
   // the _IMM Patterns are considered before the _SGPR patterns.
   unsigned PtrReg = GEPInfo.SgprParts[0];
   Register OffsetReg = MRI->createVirtualRegister(&AMDGPU::SReg_32RegClass);

 }
 
 InstructionSelector::ComplexRendererFns
 AMDGPUInstructionSelector::selectSMRDBufferImm(MachineOperand &Root) const {
   Optional<uint64_t> OffsetVal = getConstantZext32Val(Root.getReg(), *MRI);
   if (!OffsetVal)
     return {};
 
-  Optional<int64_t> EncodedImm = AMDGPU::getSMRDEncodedOffset(STI, *OffsetVal);
+  Optional<int64_t> EncodedImm =
+      AMDGPU::getSMRDEncodedOffset(STI, *OffsetVal, true);
   if (!EncodedImm)
     return {};
 
   return {{ [=](MachineInstrBuilder &MIB) { MIB.addImm(*EncodedImm); }  }};
 }
 
 InstructionSelector::ComplexRendererFns
 AMDGPUInstructionSelector::selectSMRDBufferImm32(MachineOperand &Root) const {

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi);
 
 bool isArgPassedInSGPR(const Argument *Arg);
 
 /// Convert \p ByteOffset to dwords if the subtarget uses dword SMRD immediate
 /// offsets.
 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST, uint64_t ByteOffset);
 
-/// \returns The encoding that will be used for \p ByteOffset in the SMRD offset
-/// field, or None if it won't fit. This is useful on all subtargets.
+/// \returns The encoding that will be used for \p ByteOffset in the
+/// SMRD offset field, or None if it won't fit. On GFX9 and GFX10
+/// S_LOAD instructions have a signed offset, on other subtargets it is
+/// unsigned. S_BUFFER has an unsigned offset for all subtargets.
 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST,
-                                       int64_t ByteOffset);
+                                       int64_t ByteOffset, bool IsBuffer);
 
 /// \return The encoding that can be used for a 32-bit literal offset in an SMRD
 /// instruction. This is only useful on CI.s
 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST,
                                                 int64_t ByteOffset);
 
 /// \returns true if this offset is small enough to fit in the SMRD
 /// offset field.  \p ByteOffset should be the offset in bytes and

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

default:
     return false;
   }
 }
 
 static bool hasSMEMByteOffset(const MCSubtargetInfo &ST) {
   return isGCN3Encoding(ST) || isGFX10(ST);
 }
 
-static bool isLegalSMRDEncodedImmOffset(const MCSubtargetInfo &ST,
+static bool hasSMRDSignedImmOffset(const MCSubtargetInfo &ST) {
+  return isGFX9(ST) || isGFX10(ST);
+}
+
+static bool isLegalSMRDEncodedUnsignedOffset(const MCSubtargetInfo &ST,
                                              int64_t EncodedOffset) {
   return hasSMEMByteOffset(ST) ? isUInt<20>(EncodedOffset)
                                : isUInt<8>(EncodedOffset);
 }
 
 static bool isDwordAligned(uint64_t ByteOffset) {
   return (ByteOffset & 3) == 0;
 }
 
 uint64_t convertSMRDOffsetUnits(const MCSubtargetInfo &ST,
                                 uint64_t ByteOffset) {
   if (hasSMEMByteOffset(ST))
     return ByteOffset;
 
   assert(isDwordAligned(ByteOffset));
   return ByteOffset >> 2;
 }
 
 Optional<int64_t> getSMRDEncodedOffset(const MCSubtargetInfo &ST,
-                                       int64_t ByteOffset) {
+                                       int64_t ByteOffset, bool IsBuffer) {
+  // The signed version is always a byte offset.
+  if (!IsBuffer && hasSMRDSignedImmOffset(ST)) {
+    assert(hasSMEMByteOffset(ST));
+    return isInt<20>(ByteOffset) ? Optional<int64_t>(ByteOffset) : None;
+  }
+
   if (!isDwordAligned(ByteOffset) && !hasSMEMByteOffset(ST))
     return None;
 
   int64_t EncodedOffset = convertSMRDOffsetUnits(ST, ByteOffset);
-  return isLegalSMRDEncodedImmOffset(ST, EncodedOffset) ?
-    Optional<int64_t>(EncodedOffset) : None;
+  return isLegalSMRDEncodedUnsignedOffset(ST, EncodedOffset)
+             ? Optional<int64_t>(EncodedOffset)
+             : None;
 }
 
 Optional<int64_t> getSMRDEncodedLiteralOffset32(const MCSubtargetInfo &ST,
                                                 int64_t ByteOffset) {
-  if (!isDwordAligned(ByteOffset) && !hasSMEMByteOffset(ST))
+  if (!isCI(ST) || !isDwordAligned(ByteOffset))
     return None;
 
-  assert(isCI(ST));
   int64_t EncodedOffset = convertSMRDOffsetUnits(ST, ByteOffset);
   return isUInt<32>(EncodedOffset) ? Optional<int64_t>(EncodedOffset) : None;
 }
 
 // Given Imm, split it into the values to put into the SOffset and ImmOffset
 // fields in an MUBUF instruction. Return false if it is not possible (due to a
 // hardware bug needing a workaround).
 //

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-load-constant.mir

bb.0:
     ; GFX8: liveins: $sgpr0_sgpr1
     ; GFX8: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
     ; GFX8: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[COPY]], 1048575, 0, 0 :: (load 4, addrspace 4)
     ; GFX8: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
     ; GFX10-LABEL: name: load_constant_s32_from_4_gep_1048575
     ; GFX10: liveins: $sgpr0_sgpr1
     ; GFX10: $vcc_hi = IMPLICIT_DEF
     ; GFX10: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
-    ; GFX10: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[COPY]], 1048575, 0, 0 :: (load 4, addrspace 4)
-    ; GFX10: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX10: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 1048575
+    ; GFX10: [[S_LOAD_DWORD_SGPR:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_SGPR [[COPY]], [[S_MOV_B32_]], 0, 0 :: (load 4, addrspace 4)
+    ; GFX10: $sgpr0 = COPY [[S_LOAD_DWORD_SGPR]]
     %0:sgpr(p4) = COPY $sgpr0_sgpr1
     %1:sgpr(s64) = G_CONSTANT i64 1048575
     %2:sgpr(p4) = G_PTR_ADD %0, %1
     %3:sgpr(s32) = G_LOAD %2 :: (load 4, align 4, addrspace 4)
     $sgpr0 = COPY %3
 
 ...
 
bb.0:
     ; GFX10: $sgpr0 = COPY [[S_LOAD_DWORD_SGPR]]
     %0:sgpr(p4) = COPY $sgpr0_sgpr1
     %1:sgpr(s64) = G_CONSTANT i64 1073741823
     %2:sgpr(p4) = G_PTR_ADD %0, %1
     %3:sgpr(s32) = G_LOAD %2 :: (load 4, align 4, addrspace 4)
     $sgpr0 = COPY %3
 
 ...
+
+---
+
+name: load_constant_s32_from_4_gep_negative_1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+body: |
+  bb.0:
+    liveins:  $sgpr0_sgpr1
+
+    ; GFX6-LABEL: name: load_constant_s32_from_4_gep_negative_1
+    ; GFX6: liveins: $sgpr0_sgpr1
+    ; GFX6: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX6: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 -1
+    ; GFX6: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX6: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub0
+    ; GFX6: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX6: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub1
+    ; GFX6: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX6: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX6: [[REG_SEQUENCE:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX6: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX6: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX7-LABEL: name: load_constant_s32_from_4_gep_negative_1
+    ; GFX7: liveins: $sgpr0_sgpr1
+    ; GFX7: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX7: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 -1
+    ; GFX7: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX7: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub0
+    ; GFX7: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX7: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub1
+    ; GFX7: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX7: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX7: [[REG_SEQUENCE:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX7: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX7: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX8-LABEL: name: load_constant_s32_from_4_gep_negative_1
+    ; GFX8: liveins: $sgpr0_sgpr1
+    ; GFX8: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX8: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 -1
+    ; GFX8: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX8: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub0
+    ; GFX8: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX8: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[S_MOV_B64_]].sub1
+    ; GFX8: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX8: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX8: [[REG_SEQUENCE:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX8: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX8: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX10-LABEL: name: load_constant_s32_from_4_gep_negative_1
+    ; GFX10: liveins: $sgpr0_sgpr1
+    ; GFX10: $vcc_hi = IMPLICIT_DEF
+    ; GFX10: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX10: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[COPY]], -1, 0, 0 :: (load 4, addrspace 4)
+    ; GFX10: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    %0:sgpr(p4) = COPY $sgpr0_sgpr1
+    %1:sgpr(s64) = G_CONSTANT i64 -1
+    %2:sgpr(p4) = G_PTR_ADD %0, %1
+    %3:sgpr(s32) = G_LOAD %2 :: (load 4, align 4, addrspace 4)
+    $sgpr0 = COPY %3
+
+...
+
+---
+
+name: load_constant_s32_from_4_gep_negative_524288
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+body: |
+  bb.0:
+    liveins:  $sgpr0_sgpr1
+
+    ; GFX6-LABEL: name: load_constant_s32_from_4_gep_negative_524288
+    ; GFX6: liveins: $sgpr0_sgpr1
+    ; GFX6: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX6: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4294443008
+    ; GFX6: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+    ; GFX6: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[S_MOV_B32_]], %subreg.sub0, [[S_MOV_B32_1]], %subreg.sub1
+    ; GFX6: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX6: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub0
+    ; GFX6: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX6: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub1
+    ; GFX6: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX6: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX6: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX6: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE1]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX6: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX7-LABEL: name: load_constant_s32_from_4_gep_negative_524288
+    ; GFX7: liveins: $sgpr0_sgpr1
+    ; GFX7: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX7: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4294443008
+    ; GFX7: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+    ; GFX7: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[S_MOV_B32_]], %subreg.sub0, [[S_MOV_B32_1]], %subreg.sub1
+    ; GFX7: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX7: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub0
+    ; GFX7: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX7: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub1
+    ; GFX7: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX7: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX7: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX7: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE1]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX7: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX8-LABEL: name: load_constant_s32_from_4_gep_negative_524288
+    ; GFX8: liveins: $sgpr0_sgpr1
+    ; GFX8: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX8: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4294443008
+    ; GFX8: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 -1
+    ; GFX8: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[S_MOV_B32_]], %subreg.sub0, [[S_MOV_B32_1]], %subreg.sub1
+    ; GFX8: [[COPY1:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub0
+    ; GFX8: [[COPY2:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub0
+    ; GFX8: [[COPY3:%[0-9]+]]:sreg_32 = COPY [[COPY]].sub1
+    ; GFX8: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE]].sub1
+    ; GFX8: [[S_ADD_U32_:%[0-9]+]]:sreg_32 = S_ADD_U32 [[COPY1]], [[COPY2]], implicit-def $scc
+    ; GFX8: [[S_ADDC_U32_:%[0-9]+]]:sreg_32 = S_ADDC_U32 [[COPY3]], [[COPY4]], implicit-def $scc, implicit $scc
+    ; GFX8: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[S_ADD_U32_]], %subreg.sub0, [[S_ADDC_U32_]], %subreg.sub1
+    ; GFX8: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[REG_SEQUENCE1]], 0, 0, 0 :: (load 4, addrspace 4)
+    ; GFX8: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    ; GFX10-LABEL: name: load_constant_s32_from_4_gep_negative_524288
+    ; GFX10: liveins: $sgpr0_sgpr1
+    ; GFX10: $vcc_hi = IMPLICIT_DEF
+    ; GFX10: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GFX10: [[S_LOAD_DWORD_IMM:%[0-9]+]]:sreg_32_xm0_xexec = S_LOAD_DWORD_IMM [[COPY]], -524288, 0, 0 :: (load 4, addrspace 4)
+    ; GFX10: $sgpr0 = COPY [[S_LOAD_DWORD_IMM]]
+    %0:sgpr(p4) = COPY $sgpr0_sgpr1
+    %1:sgpr(s64) = G_CONSTANT i64 -524288
+    %2:sgpr(p4) = G_PTR_ADD %0, %1
+    %3:sgpr(s32) = G_LOAD %2 :: (load 4, align 4, addrspace 4)
+    $sgpr0 = COPY %3
+
+...

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

 ; FIXME: Need to add support for mubuf stores to enable this on SI.
 ; XUN: llc < %s -march=amdgcn -mcpu=tahiti -show-mc-encoding -verify-machineinstrs -global-isel | FileCheck --check-prefix=SI --check-prefix=GCN --check-prefix=SIVI %s
-; RUN: llc < %s -march=amdgcn -mcpu=bonaire -show-mc-encoding -verify-machineinstrs -global-isel | FileCheck --check-prefix=CI --check-prefix=GCN %s
-; RUN: llc < %s -march=amdgcn -mcpu=tonga -show-mc-encoding -verify-machineinstrs -global-isel | FileCheck --check-prefix=VI --check-prefix=GCN --check-prefix=SIVI %s
+; RUN: llc < %s -march=amdgcn -mcpu=bonaire -show-mc-encoding -verify-machineinstrs -global-isel | FileCheck --check-prefix=CI --check-prefix=GCN --check-prefix=SICIVI %s
+; RUN: llc < %s -march=amdgcn -mcpu=tonga -show-mc-encoding -verify-machineinstrs -global-isel | FileCheck --check-prefix=VI --check-prefix=GCN --check-prefix=SIVI --check-prefix=SICIVI %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -show-mc-encoding -verify-machineinstrs -global-isel < %s | FileCheck -check-prefix=GFX9 -check-prefix=GFX9_10 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
+; RUN: llc -march=amdgcn -mcpu=gfx1010 -show-mc-encoding -verify-machineinstrs -global-isel < %s | FileCheck -check-prefix=GFX10 -check-prefix=GFX9_10 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
 
 ; SMRD load with an immediate offset.
 ; GCN-LABEL: {{^}}smrd0:
 ; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x1 ; encoding: [0x01
-; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4
+; VIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4
 define amdgpu_kernel void @smrd0(i32 addrspace(4)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(4)* %ptr, i64 1
   %1 = load i32, i32 addrspace(4)* %0
   store i32 %1, i32 addrspace(1)* undef
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset.
 ; GCN-LABEL: {{^}}smrd1:
 ; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff,0x{{[0-9]+[137]}}
-; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
+; VIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
 define amdgpu_kernel void @smrd1(i32 addrspace(4)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(4)* %ptr, i64 255
   %1 = load i32, i32 addrspace(4)* %0
   store i32 %1, i32 addrspace(1)* undef
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate.
 ; GCN-LABEL: {{^}}smrd2:
 ; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
 ; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
 ; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
-; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
+; VIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
 ; GCN: s_endpgm
 define amdgpu_kernel void @smrd2(i32 addrspace(4)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(4)* %ptr, i64 256
   %1 = load i32, i32 addrspace(4)* %0
   store i32 %1, i32 addrspace(1)* undef
   ret void
 }

 }
 
 ; SMRD load with the largest possible immediate offset on VI
 ; GCN-LABEL: {{^}}smrd4:
 ; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
 ; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 ; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
 ; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
+; GFX9_10: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
+; GFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 define amdgpu_kernel void @smrd4(i32 addrspace(4)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(4)* %ptr, i64 262143
   %1 = load i32, i32 addrspace(4)* %0
   store i32 %1, i32 addrspace(1)* undef
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate on VI
 ; GCN-LABEL: {{^}}smrd5:
-; SIVI: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
-; SIVI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
+; SIVIGFX9_10: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
+; SIVIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 ; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; GCN: s_endpgm
 define amdgpu_kernel void @smrd5(i32 addrspace(4)* %ptr) {
 entry:
   %0 = getelementptr i32, i32 addrspace(4)* %ptr, i64 262144
   %1 = load i32, i32 addrspace(4)* %0
   store i32 %1, i32 addrspace(1)* undef
   ret void
 }
 
+; GFX9_10 can use a signed immediate byte offset
+; GCN-LABEL: {{^}}smrd6:
+; SICIVI: s_add_u32 s{{[0-9]}}, s{{[0-9]}}, -4
+; SICIVI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x0
+; GFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xfffffffc
+define amdgpu_kernel void @smrd6(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
+entry:
+  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 -1
+  %tmp1 = load i32, i32 addrspace(4)* %tmp
+  store i32 %tmp1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Don't use a negative SGPR offset
+; GCN-LABEL: {{^}}smrd7:
+; GCN: s_add_u32 s{{[0-9]}}, s{{[0-9]}}, 0xffe00000
+; SICIVI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x0
+; GFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x0
+define amdgpu_kernel void @smrd7(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
+entry:
+  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 -524288
+  %tmp1 = load i32, i32 addrspace(4)* %tmp
+  store i32 %tmp1, i32 addrspace(1)* %out
+  ret void
+}

llvm/test/CodeGen/AMDGPU/smrd.ll

 ; RUN: llc -march=amdgcn -mcpu=tahiti  -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=SI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=SICI -check-prefix=SIVIGFX9_10 %s
 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=CI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=SICI %s
 ; RUN: llc -march=amdgcn -mcpu=tonga   -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=VI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx900  -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=GFX9 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
-; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=GFX10 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
+; RUN: llc -march=amdgcn -mcpu=gfx900  -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=GFX9 -check-prefix=GFX9_10 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
+; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=GFX10 -check-prefix=GFX9_10 -check-prefix=GCN -check-prefix=VIGFX9_10 -check-prefix=SIVIGFX9_10  %s
 
 ; SMRD load with an immediate offset.
 ; GCN-LABEL: {{^}}smrd0:
 ; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x1 ; encoding: [0x01
 ; VIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4
 define amdgpu_kernel void @smrd0(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
 entry:
   %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 1
entry:
   ret void
 }
 
 ; SMRD load with the largest possible immediate offset on VI
 ; GCN-LABEL: {{^}}smrd4:
 ; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
 ; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 ; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
-; VIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
+; VI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
+; GFX9_10: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
+; GFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 define amdgpu_kernel void @smrd4(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
 entry:
   %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 262143
   %tmp1 = load i32, i32 addrspace(4)* %tmp
   store i32 %tmp1, i32 addrspace(1)* %out
   ret void
 }
 
 ; SMRD load with an offset greater than the largest possible immediate on VI
 ; GCN-LABEL: {{^}}smrd5:
 ; SIVIGFX9_10: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
 ; SIVIGFX9_10: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
 ; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
 ; GCN: s_endpgm
 define amdgpu_kernel void @smrd5(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
 entry:
   %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 262144
   %tmp1 = load i32, i32 addrspace(4)* %tmp
   store i32 %tmp1, i32 addrspace(1)* %out
   ret void
 }
 
+; GFX9_10 can use a signed immediate byte offset
+; GCN-LABEL: {{^}}smrd6:
+; SICIVI: s_add_u32 s{{[0-9]}}, s{{[0-9]}}, -4
+; SICIVI: s_load_dword s{{[0-9]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x0
+; GFX9_10: s_load_dword s{{[0-9]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xfffffffc
+define amdgpu_kernel void @smrd6(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
+entry:
+  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 -1
+  %tmp1 = load i32, i32 addrspace(4)* %tmp
+  store i32 %tmp1, i32 addrspace(1)* %out
+  ret void
+}
+
+; Don't use a negative SGPR offset
+; GCN-LABEL: {{^}}smrd7:
+; GCN: s_add_u32 s{{[0-9]}}, s{{[0-9]}}, 0xffe00000
+; SICIVI: s_load_dword s{{[0-9]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x0
+; GFX9_10: s_load_dword s{{[0-9]}}, s{{\[[0-9]+:[0-9]+\]}}, 0x0
+define amdgpu_kernel void @smrd7(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
+entry:
+  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 -524288
+  %tmp1 = load i32, i32 addrspace(4)* %tmp
+  store i32 %tmp1, i32 addrspace(1)* %out
+  ret void
+}
+
 ; GCN-LABEL: {{^}}smrd_hazard:
 ; GCN-DAG: s_mov_b32 s3, 3
 ; GCN-DAG: s_mov_b32 s2, 2
 ; GCN-DAG: s_mov_b32 s1, 1
 ; GCN-DAG: s_mov_b32 s0, 0
 ; SI-NEXT: nop 3
 ; GFX10-NEXT: ; implicit-def: $vcc_hi
 ; GCN-NEXT: s_buffer_load_dword s0, s[0:3], 0x0