AMDGPU/GlobalISel: Select llvm.amdgcn.raw.buffer.store.format

[arsenm]

Summary

This needs special handling due to some subtargets that have a
nonstandard register layout for f16 vectors

Also reject some illegal types on other targets.

[arsenm]

Forgot normal 32-bit case

[kerbowa]

LGTM

[arsenm]

r372292

lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

if (Def->getOpcode() == AMDGPU::G_ADD) {
     // FIXME: matcher should ignore copies
     if (mi_match(Def->getOperand(1).getReg(), MRI, m_Copy(m_ICst(Offset))))
       return {Def->getOperand(0).getReg(), Offset, Def};
   }
 
   return {Reg, 0, Def};
 }
 
+static unsigned getBufferStoreOpcode(LLT Ty,
+                                     const unsigned MemSize,
+                                     const bool Offen) {
+  const int Size = Ty.getSizeInBits();
+  switch (8 * MemSize) {
+  case 8:
+    return Offen ? AMDGPU::BUFFER_STORE_BYTE_OFFEN_exact :
+                   AMDGPU::BUFFER_STORE_BYTE_OFFSET_exact;
+  case 16:
+    return Offen ? AMDGPU::BUFFER_STORE_SHORT_OFFEN_exact :
+                   AMDGPU::BUFFER_STORE_SHORT_OFFSET_exact;
+  default:
+    unsigned Opc = Offen ? AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact :
+                           AMDGPU::BUFFER_STORE_DWORD_OFFSET_exact;
+    if (Size > 32)
+      Opc = AMDGPU::getMUBUFOpcode(Opc, Size / 32);
+    return Opc;
+  }
+}
+
+static unsigned getBufferStoreFormatOpcode(LLT Ty,
+                                           const unsigned MemSize,
+                                           const bool Offen) {
+  bool IsD16Packed = Ty.getScalarSizeInBits() == 16;
+  bool IsD16Unpacked = MemSize < Ty.getSizeInBits();
+  int NumElts = Ty.isVector() ? Ty.getNumElements() : 1;
+
+  if (IsD16Packed) {
+    switch (NumElts) {
+    case 1:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFSET_exact;
+    case 2:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XY_OFFSET_exact;
+    case 3:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_OFFSET_exact;
+    case 4:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_OFFSET_exact;
+    default:
+      return -1;
+    }
+  }
+
+  if (IsD16Unpacked) {
+    switch (NumElts) {
+    case 1:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_X_OFFSET_exact;
+    case 2:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFSET_exact;
+    case 3:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_gfx80_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZ_gfx80_OFFSET_exact;
+    case 4:
+      return Offen ? AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFEN_exact :
+                     AMDGPU::BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFSET_exact;
+    default:
+      return -1;
+    }
+  }
+
+  unsigned Opc = Offen ? AMDGPU::BUFFER_STORE_FORMAT_X_OFFEN_exact :
+                         AMDGPU::BUFFER_STORE_FORMAT_X_OFFSET_exact;
+  return AMDGPU::getMUBUFOpcode(Opc, NumElts);
+}
+
+
 // TODO: Move this to combiner
 // Returns base register, imm offset, total constant offset.
 std::tuple<Register, unsigned, unsigned>
 AMDGPUInstructionSelector::splitBufferOffsets(MachineIRBuilder &B,
                                               Register OrigOffset) const {
   const unsigned MaxImm = 4095;
   Register BaseReg;
   unsigned TotalConstOffset;
bool AMDGPUInstructionSelector::selectStoreIntrinsic(MachineInstr &MI,
   unsigned TotalOffset;
 
   std::tie(VOffset, ImmOffset, TotalOffset) = splitBufferOffsets(B, VOffset);
   if (TotalOffset != 0)
     MMO = MF.getMachineMemOperand(MMO, TotalOffset, MemSize);
 
   const bool Offen = !isZero(VOffset, MRI);
 
-  unsigned Opc = AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact;
-  switch (8 * MemSize) {
-  case 8:
-    Opc = Offen ? AMDGPU::BUFFER_STORE_BYTE_OFFEN_exact :
-                  AMDGPU::BUFFER_STORE_BYTE_OFFSET_exact;
-    break;
-  case 16:
-    Opc = Offen ? AMDGPU::BUFFER_STORE_SHORT_OFFEN_exact :
-                  AMDGPU::BUFFER_STORE_SHORT_OFFSET_exact;
-    break;
-  default:
-    Opc = Offen ? AMDGPU::BUFFER_STORE_DWORD_OFFEN_exact :
-                  AMDGPU::BUFFER_STORE_DWORD_OFFSET_exact;
-    if (Size > 32)
-      Opc = AMDGPU::getMUBUFOpcode(Opc, Size / 32);
-    break;
-  }
+  int Opc = IsFormat ? getBufferStoreFormatOpcode(Ty, MemSize, Offen) :
+    getBufferStoreOpcode(Ty, MemSize, Offen);
+  if (Opc == -1)
+    return false;
 
   MachineInstrBuilder MIB = B.buildInstr(Opc)
     .addUse(VData);
 
   if (Offen)
     MIB.addUse(VOffset);
 
   MIB.addUse(RSrc)
case Intrinsic::amdgcn_end_cf: {
     I.eraseFromParent();
 
     if (!MRI.getRegClassOrNull(Reg))
       MRI.setRegClass(Reg, TRI.getWaveMaskRegClass());
     return true;
   }
   case Intrinsic::amdgcn_raw_buffer_store:
     return selectStoreIntrinsic(I, false);
+  case Intrinsic::amdgcn_raw_buffer_store_format:
+    return selectStoreIntrinsic(I, true);
   default:
     return selectImpl(I, *CoverageInfo);
   }
 }
 
 bool AMDGPUInstructionSelector::selectG_SELECT(MachineInstr &I) const {
   MachineBasicBlock *BB = I.getParent();
   MachineFunction *MF = BB->getParent();

lib/Target/AMDGPU/AMDGPULegalizerInfo.h

bool legalizePreloadedArgIntrin(
     AMDGPUFunctionArgInfo::PreloadedValue ArgType) const;
 
   bool legalizeFDIVFast(MachineInstr &MI, MachineRegisterInfo &MRI,
                         MachineIRBuilder &B) const;
 
   bool legalizeImplicitArgPtr(MachineInstr &MI, MachineRegisterInfo &MRI,
                               MachineIRBuilder &B) const;
 
+  Register handleD16VData(MachineIRBuilder &B, MachineRegisterInfo &MRI,
+                          Register Reg) const;
   bool legalizeRawBufferStore(MachineInstr &MI, MachineRegisterInfo &MRI,
                               MachineIRBuilder &B, bool IsFormat) const;
   bool legalizeIntrinsic(MachineInstr &MI, MachineRegisterInfo &MRI,
                          MachineIRBuilder &MIRBuilder) const override;
 
 };
 } // End llvm namespace.
 #endif

lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

bool AMDGPULegalizerInfo::legalizeImplicitArgPtr(MachineInstr &MI,
   if (!loadInputValue(KernargPtrReg, B, Arg))
     return false;
 
   B.buildGEP(DstReg, KernargPtrReg, B.buildConstant(IdxTy, Offset).getReg(0));
   MI.eraseFromParent();
   return true;
 }
 
+/// Handle register layout difference for f16 images for some subtargets.
+Register AMDGPULegalizerInfo::handleD16VData(MachineIRBuilder &B,
+                                             MachineRegisterInfo &MRI,
+                                             Register Reg) const {
+  if (!ST.hasUnpackedD16VMem())
+    return Reg;
+
+  const LLT S16 = LLT::scalar(16);
+  const LLT S32 = LLT::scalar(32);
+  LLT StoreVT = MRI.getType(Reg);
+  assert(StoreVT.isVector() && StoreVT.getElementType() == S16);
+
+  auto Unmerge = B.buildUnmerge(S16, Reg);
+
+  SmallVector<Register, 4> WideRegs;
+  for (int I = 0, E = Unmerge->getNumOperands() - 1; I != E; ++I)
+    WideRegs.push_back(B.buildAnyExt(S32, Unmerge.getReg(I)).getReg(0));
+
+  int NumElts = StoreVT.getNumElements();
+
+  return B.buildBuildVector(LLT::vector(NumElts, S32), WideRegs).getReg(0);
+}
+
 bool AMDGPULegalizerInfo::legalizeRawBufferStore(MachineInstr &MI,
                                                  MachineRegisterInfo &MRI,
                                                  MachineIRBuilder &B,
                                                  bool IsFormat) const {
-  assert(!IsFormat && "format buffer stores not implemented");
-
   // TODO: Reject f16 format on targets where unsupported.
   Register VData = MI.getOperand(1).getReg();
   LLT Ty = MRI.getType(VData);
 
   B.setInstr(MI);
 
   const LLT S32 = LLT::scalar(32);
   const LLT S16 = LLT::scalar(16);
 
   // Fixup illegal register types for i8 stores.
   if (Ty == LLT::scalar(8) || Ty == S16) {
     Register AnyExt = B.buildAnyExt(LLT::scalar(32), VData).getReg(0);
     MI.getOperand(1).setReg(AnyExt);
     return true;
   }
 
   if (Ty.isVector()) {
-    if (Ty.getElementType() == S16 && Ty.getNumElements() <= 4)
+    if (Ty.getElementType() == S16 && Ty.getNumElements() <= 4) {
+      if (IsFormat)
+        MI.getOperand(1).setReg(handleD16VData(B, MRI, VData));
       return true;
+    }
 
     return Ty.getElementType() == S32 && Ty.getNumElements() <= 4;
   }
 
   return Ty == S32;
 }
 
 bool AMDGPULegalizerInfo::legalizeIntrinsic(MachineInstr &MI,
return legalizePreloadedArgIntrin(
       MI, MRI, B, AMDGPUFunctionArgInfo::IMPLICIT_BUFFER_PTR);
   case Intrinsic::amdgcn_dispatch_id:
     return legalizePreloadedArgIntrin(MI, MRI, B,
                                       AMDGPUFunctionArgInfo::DISPATCH_ID);
   case Intrinsic::amdgcn_fdiv_fast:
     return legalizeFDIVFast(MI, MRI, B);
   case Intrinsic::amdgcn_raw_buffer_store:
     return legalizeRawBufferStore(MI, MRI, B, false);
+  case Intrinsic::amdgcn_raw_buffer_store_format:
+    return legalizeRawBufferStore(MI, MRI, B, true);
   default:
     return true;
   }
 
   return true;
 }

test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.store.format.f16.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -stop-after=instruction-select -o - %s | FileCheck -check-prefix=UNPACKED %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx810 -stop-after=instruction-select -o - %s | FileCheck -check-prefix=PACKED %s
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_f16(<4 x i32> inreg %rsrc, half %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_f16
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_X_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 2 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_f16
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_X_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 2 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.f16(half %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY7]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16(<4 x i32> inreg %rsrc, <4 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY8]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_1:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY9]], [[COPY5]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1, [[COPY5]], %subreg.sub2, [[V_LSHRREV_B32_e64_1]], %subreg.sub3
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1, $vgpr2
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; PACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XYZW_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.v4f16(<4 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+; Make sure unpack code is emitted outside of loop
+define amdgpu_ps void @raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16(<4 x i32> %rsrc, <4 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   successors: %bb.2(0x80000000)
+  ; UNPACKED:   liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6
+  ; UNPACKED:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY8]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_1:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY9]], [[COPY5]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1, [[COPY5]], %subreg.sub2, [[V_LSHRREV_B32_e64_1]], %subreg.sub3
+  ; UNPACKED:   [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
+  ; UNPACKED:   [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
+  ; UNPACKED:   [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+  ; UNPACKED: bb.2:
+  ; UNPACKED:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub0, implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub1, implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; UNPACKED:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY10]], implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub0, implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub1, implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
+  ; UNPACKED:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY11]], implicit $exec
+  ; UNPACKED:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+  ; UNPACKED:   [[REG_SEQUENCE4:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE4]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; UNPACKED:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+  ; UNPACKED:   S_CBRANCH_EXECNZ %bb.2, implicit $exec
+  ; UNPACKED: bb.3:
+  ; UNPACKED:   successors: %bb.4(0x80000000)
+  ; UNPACKED:   $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+  ; UNPACKED: bb.4:
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   successors: %bb.2(0x80000000)
+  ; PACKED:   liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6
+  ; PACKED:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; PACKED:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
+  ; PACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
+  ; PACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
+  ; PACKED:   [[COPY8:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
+  ; PACKED:   [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
+  ; PACKED:   [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+  ; PACKED: bb.2:
+  ; PACKED:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
+  ; PACKED:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY8]].sub0, implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY8]].sub1, implicit $exec
+  ; PACKED:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; PACKED:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY8]], implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY9]].sub0, implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY9]].sub1, implicit $exec
+  ; PACKED:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
+  ; PACKED:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY9]], implicit $exec
+  ; PACKED:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+  ; PACKED:   [[REG_SEQUENCE4:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XYZW_OFFEN_exact [[REG_SEQUENCE1]], [[COPY6]], [[REG_SEQUENCE4]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; PACKED:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+  ; PACKED:   S_CBRANCH_EXECNZ %bb.2, implicit $exec
+  ; PACKED: bb.3:
+  ; PACKED:   successors: %bb.4(0x80000000)
+  ; PACKED:   $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+  ; PACKED: bb.4:
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.v4f16(<4 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4095(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4095
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095
+  ; UNPACKED:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY6]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], [[COPY5]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4095
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4095
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 4095, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4096(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4096
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
+  ; UNPACKED:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY6]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], [[COPY5]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset4096
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 4096
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], [[COPY5]], [[REG_SEQUENCE]], [[S_MOV_B32_]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 4096, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_16(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_16
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   %11:vgpr_32, dead %21:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY8]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_16
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; PACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; PACKED:   %11:vgpr_32, dead %13:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  %voffset.add = add i32 %voffset, 16
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4095(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4095
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4095
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   %11:vgpr_32, dead %22:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; UNPACKED:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY8]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4095
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4095
+  ; PACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; PACKED:   %11:vgpr_32, dead %13:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  %voffset.add = add i32 %voffset, 4095
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4096(<4 x i32> inreg %rsrc, <2 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4096
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; UNPACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4096
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   %11:vgpr_32, dead %22:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; UNPACKED:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY8]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XY_gfx80_OFFEN_exact [[REG_SEQUENCE1]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__sgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v2f16_soffset_add_4096
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   liveins: $sgpr2, $sgpr3, $sgpr4, $sgpr5, $sgpr6, $vgpr0, $vgpr1
+  ; PACKED:   [[COPY:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr2
+  ; PACKED:   [[COPY1:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr3
+  ; PACKED:   [[COPY2:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr4
+  ; PACKED:   [[COPY3:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr5
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr6
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4096
+  ; PACKED:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; PACKED:   %11:vgpr_32, dead %13:sreg_64_xexec = V_ADD_I32_e64 [[COPY5]], [[COPY7]], 0, implicit $exec
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XY_OFFEN_exact [[COPY4]], %11, [[REG_SEQUENCE]], [[COPY6]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   S_ENDPGM 0
+  %voffset.add = add i32 %voffset, 4096
+  call void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
+  ret void
+}
+
+
+; Check what happens with offset add inside a waterfall loop
+define amdgpu_ps void @raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16_add_4096(<4 x i32> %rsrc, <4 x half> %val, i32 %voffset, i32 inreg %soffset) {
+  ; UNPACKED-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16_add_4096
+  ; UNPACKED: bb.1 (%ir-block.0):
+  ; UNPACKED:   successors: %bb.2(0x80000000)
+  ; UNPACKED:   liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6
+  ; UNPACKED:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; UNPACKED:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; UNPACKED:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; UNPACKED:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; UNPACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
+  ; UNPACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
+  ; UNPACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
+  ; UNPACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; UNPACKED:   [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; UNPACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4096
+  ; UNPACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; UNPACKED:   %13:vgpr_32, dead %47:sreg_64_xexec = V_ADD_I32_e64 [[COPY6]], [[COPY8]], 0, implicit $exec
+  ; UNPACKED:   [[S_MOV_B32_1:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 16
+  ; UNPACKED:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY9]], [[COPY4]], implicit $exec
+  ; UNPACKED:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
+  ; UNPACKED:   [[V_LSHRREV_B32_e64_1:%[0-9]+]]:vgpr_32 = V_LSHRREV_B32_e64 [[COPY10]], [[COPY5]], implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[V_LSHRREV_B32_e64_]], %subreg.sub1, [[COPY5]], %subreg.sub2, [[V_LSHRREV_B32_e64_1]], %subreg.sub3
+  ; UNPACKED:   [[COPY11:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
+  ; UNPACKED:   [[COPY12:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
+  ; UNPACKED:   [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+  ; UNPACKED: bb.2:
+  ; UNPACKED:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub0, implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY11]].sub1, implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; UNPACKED:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY11]], implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY12]].sub0, implicit $exec
+  ; UNPACKED:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY12]].sub1, implicit $exec
+  ; UNPACKED:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
+  ; UNPACKED:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY12]], implicit $exec
+  ; UNPACKED:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+  ; UNPACKED:   [[REG_SEQUENCE4:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
+  ; UNPACKED:   BUFFER_STORE_FORMAT_D16_XYZW_gfx80_OFFEN_exact [[REG_SEQUENCE1]], %13, [[REG_SEQUENCE4]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; UNPACKED:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; UNPACKED:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+  ; UNPACKED:   S_CBRANCH_EXECNZ %bb.2, implicit $exec
+  ; UNPACKED: bb.3:
+  ; UNPACKED:   successors: %bb.4(0x80000000)
+  ; UNPACKED:   $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+  ; UNPACKED: bb.4:
+  ; UNPACKED:   S_ENDPGM 0
+  ; PACKED-LABEL: name: raw_buffer_store_format__vgpr_rsrc__vgpr_val__vgpr_voffset__sgpr_soffset_v4f16_add_4096
+  ; PACKED: bb.1 (%ir-block.0):
+  ; PACKED:   successors: %bb.2(0x80000000)
+  ; PACKED:   liveins: $sgpr2, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6
+  ; PACKED:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; PACKED:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; PACKED:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; PACKED:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; PACKED:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr4
+  ; PACKED:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr5
+  ; PACKED:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr6
+  ; PACKED:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; PACKED:   [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; PACKED:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY5]], %subreg.sub1
+  ; PACKED:   [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0 = S_MOV_B32 4096
+  ; PACKED:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_]]
+  ; PACKED:   %13:vgpr_32, dead %31:sreg_64_xexec = V_ADD_I32_e64 [[COPY6]], [[COPY8]], 0, implicit $exec
+  ; PACKED:   [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
+  ; PACKED:   [[COPY10:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub2_sub3
+  ; PACKED:   [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+  ; PACKED: bb.2:
+  ; PACKED:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
+  ; PACKED:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY9]].sub0, implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY9]].sub1, implicit $exec
+  ; PACKED:   [[REG_SEQUENCE2:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; PACKED:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE2]], [[COPY9]], implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub0, implicit $exec
+  ; PACKED:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0 = V_READFIRSTLANE_B32 [[COPY10]].sub1, implicit $exec
+  ; PACKED:   [[REG_SEQUENCE3:%[0-9]+]]:sreg_64_xexec = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
+  ; PACKED:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[COPY10]], implicit $exec
+  ; PACKED:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+  ; PACKED:   [[REG_SEQUENCE4:%[0-9]+]]:sreg_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
+  ; PACKED:   BUFFER_STORE_FORMAT_D16_XYZW_OFFEN_exact [[REG_SEQUENCE1]], %13, [[REG_SEQUENCE4]], [[COPY7]], 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into custom TargetCustom7, align 1, addrspace 4)
+  ; PACKED:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; PACKED:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+  ; PACKED:   S_CBRANCH_EXECNZ %bb.2, implicit $exec
+  ; PACKED: bb.3:
+  ; PACKED:   successors: %bb.4(0x80000000)
+  ; PACKED:   $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+  ; PACKED: bb.4:
+  ; PACKED:   S_ENDPGM 0
+  %voffset.add = add i32 %voffset, 4096
+  call void @llvm.amdgcn.raw.buffer.store.format.v4f16(<4 x half> %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
+  ret void
+}
+
+declare void @llvm.amdgcn.raw.buffer.store.format.f16(half, <4 x i32>, i32, i32, i32 immarg)
+declare void @llvm.amdgcn.raw.buffer.store.format.v2f16(<2 x half>, <4 x i32>, i32, i32, i32 immarg)
+declare void @llvm.amdgcn.raw.buffer.store.format.v4f16(<4 x half>, <4 x i32>, i32, i32, i32 immarg)