AMDGPU: Start interpreting byref on kernel arguments

[arsenm]

Summary

It doesn't make sense for kernel arguments to be passed on the stack,
but we can still interpret byref sensibly for constant and other
global-like address spaces.

These are treated identically to value aggregates placed in the kernel
argument list. A %struct.foo or %struct.foo addrspace(4)*
byref(sizeof(%struct.foo)) align(alignof(%struct.foo)) argument should
produce the same offsets and argument metadata.

This handles all 3 kernel ABI implementations, and the two HSA
metadata emission paths.

[arsenm]

Switch to byref

[rampitec]

LGTM with a nit.

llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.h

164

This is too long.

[arsenm]

1168119c2f0eb46774e5358df957bf90084bf75a

llvm/lib/Target/AMDGPU/AMDGPUCallLowering.h

 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
 
 namespace llvm {
 
 class AMDGPUTargetLowering;
 class MachineInstrBuilder;
 
 class AMDGPUCallLowering: public CallLowering {
-  Register lowerParameterPtr(MachineIRBuilder &B, Type *ParamTy,
-                             uint64_t Offset) const;
+  void lowerParameterPtr(Register DstReg, MachineIRBuilder &B, Type *ParamTy,
 
   void lowerParameter(MachineIRBuilder &B, Type *ParamTy, uint64_t Offset,
                       Align Alignment, Register DstReg) const;
 
   /// A function of this type is used to perform value split action.
   using SplitArgTy = std::function<void(ArrayRef<Register>, Register, LLT, LLT, int)>;
 
   void splitToValueTypes(MachineIRBuilder &B,

llvm/lib/Target/AMDGPU/AMDGPUCallLowering.cpp

bool AMDGPUCallLowering::lowerReturn(MachineIRBuilder &B,
   }
 
   // TODO: Handle CalleeSavedRegsViaCopy.
 
   B.insertInstr(Ret);
   return true;
 }
 
-Register AMDGPUCallLowering::lowerParameterPtr(MachineIRBuilder &B,
-                                               Type *ParamTy,
-                                               uint64_t Offset) const {
-
+void AMDGPUCallLowering::lowerParameterPtr(Register DstReg, MachineIRBuilder &B,
   MachineFunction &MF = B.getMF();
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   MachineRegisterInfo &MRI = MF.getRegInfo();
-  const Function &F = MF.getFunction();
-  const DataLayout &DL = F.getParent()->getDataLayout();
-  PointerType *PtrTy = PointerType::get(ParamTy, AMDGPUAS::CONSTANT_ADDRESS);
-  LLT PtrType = getLLTForType(*PtrTy, DL);
   Register KernArgSegmentPtr =
     MFI->getPreloadedReg(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
   Register KernArgSegmentVReg = MRI.getLiveInVirtReg(KernArgSegmentPtr);
 
   auto OffsetReg = B.buildConstant(LLT::scalar(64), Offset);
 
-  return B.buildPtrAdd(PtrType, KernArgSegmentVReg, OffsetReg).getReg(0);
+  B.buildPtrAdd(DstReg, KernArgSegmentVReg, OffsetReg);
 }
 
 void AMDGPUCallLowering::lowerParameter(MachineIRBuilder &B, Type *ParamTy,
                                         uint64_t Offset, Align Alignment,
                                         Register DstReg) const {
   MachineFunction &MF = B.getMF();
   const Function &F = MF.getFunction();
   const DataLayout &DL = F.getParent()->getDataLayout();
   MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
   unsigned TypeSize = DL.getTypeStoreSize(ParamTy);
-  Register PtrReg = lowerParameterPtr(B, ParamTy, Offset);
+
+  LLT PtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
+  Register PtrReg = B.getMRI()->createGenericVirtualRegister(PtrTy);
+  lowerParameterPtr(PtrReg, B, ParamTy, Offset);
 
   MachineMemOperand *MMO = MF.getMachineMemOperand(
       PtrInfo,
       MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
           MachineMemOperand::MOInvariant,
       TypeSize, Alignment);
 
   B.buildLoad(DstReg, PtrReg, *MMO);
bool AMDGPUCallLowering::lowerFormalArgumentsKernel(
 
   unsigned i = 0;
   const Align KernArgBaseAlign(16);
   const unsigned BaseOffset = Subtarget->getExplicitKernelArgOffset(F);
   uint64_t ExplicitArgOffset = 0;
 
   // TODO: Align down to dword alignment and extract bits for extending loads.
   for (auto &Arg : F.args()) {
-    Type *ArgTy = Arg.getType();
+    const bool IsByRef = Arg.hasByRefAttr();
+    Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
     unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
     if (AllocSize == 0)
       continue;
 
-    Align ABIAlign = DL.getABITypeAlign(ArgTy);
+    MaybeAlign ABIAlign = IsByRef ? Arg.getParamAlign() : None;
+    if (!ABIAlign)
+      ABIAlign = DL.getABITypeAlign(ArgTy);
 
     uint64_t ArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + BaseOffset;
     ExplicitArgOffset = alignTo(ExplicitArgOffset, ABIAlign) + AllocSize;
 
     if (Arg.use_empty()) {
       ++i;
       continue;
     }
 
+    Align Alignment = commonAlignment(KernArgBaseAlign, ArgOffset);
+
+    if (IsByRef) {
+      unsigned ByRefAS = cast<PointerType>(Arg.getType())->getAddressSpace();
+
+      assert(VRegs[i].size() == 1 &&
+             "expected only one register for byval pointers");
+      if (ByRefAS == AMDGPUAS::CONSTANT_ADDRESS) {
+        lowerParameterPtr(VRegs[i][0], B, ArgTy, ArgOffset);
+      } else {
+        const LLT ConstPtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
+        Register PtrReg = MRI.createGenericVirtualRegister(ConstPtrTy);
+        lowerParameterPtr(PtrReg, B, ArgTy, ArgOffset);
+
+        B.buildAddrSpaceCast(VRegs[i][0], PtrReg);
+      }
+    } else {
       ArrayRef<Register> OrigArgRegs = VRegs[i];
       Register ArgReg =
         OrigArgRegs.size() == 1
         ? OrigArgRegs[0]
         : MRI.createGenericVirtualRegister(getLLTForType(*ArgTy, DL));
 
-    Align Alignment = commonAlignment(KernArgBaseAlign, ArgOffset);
       lowerParameter(B, ArgTy, ArgOffset, Alignment, ArgReg);
       if (OrigArgRegs.size() > 1)
         unpackRegs(OrigArgRegs, ArgReg, ArgTy, B);
+    }
+
     ++i;
   }
 
   TLI.allocateSpecialEntryInputVGPRs(CCInfo, MF, *TRI, *Info);
   TLI.allocateSystemSGPRs(CCInfo, MF, *Info, F.getCallingConv(), false);
   return true;
 }
 

llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.h

private:
 
   void emitKernelAttrs(const Function &Func, msgpack::MapDocNode Kern);
 
   void emitKernelArgs(const Function &Func, msgpack::MapDocNode Kern);
 
   void emitKernelArg(const Argument &Arg, unsigned &Offset,
                      msgpack::ArrayDocNode Args);
 
-  void emitKernelArg(const DataLayout &DL, Type *Ty, StringRef ValueKind,
+  void emitKernelArg(const DataLayout &DL, Type *Ty, Align Alignment,
+                     StringRef ValueKind,
                      unsigned &Offset, msgpack::ArrayDocNode Args,
                      MaybeAlign PointeeAlign = None, StringRef Name = "",
                      StringRef TypeName = "", StringRef BaseTypeName = "",
                      StringRef AccQual = "", StringRef TypeQual = "");
 
   void emitHiddenKernelArgs(const Function &Func, unsigned &Offset,
                             msgpack::ArrayDocNode Args);
 
private:
   void emitKernelLanguage(const Function &Func);
 
   void emitKernelAttrs(const Function &Func);
 
   void emitKernelArgs(const Function &Func);
 
   void emitKernelArg(const Argument &Arg);
 
-  void emitKernelArg(const DataLayout &DL, Type *Ty, ValueKind ValueKind,
+  void emitKernelArg(const DataLayout &DL, Type *Ty, Align Alignment, ValueKind ValueKind,
                      MaybeAlign PointeeAlign = None, StringRef Name = "",
                      StringRef TypeName = "", StringRef BaseTypeName = "",
                      StringRef AccQual = "", StringRef TypeQual = "");
 
   void emitHiddenKernelArgs(const Function &Func);
 
   const Metadata &getHSAMetadata() const {
     return HSAMetadata;

llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp

 #include "SIMachineFunctionInfo.h"
 #include "SIProgramInfo.h"
 #include "Utils/AMDGPUBaseInfo.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Support/raw_ostream.h"
 
+using namespace llvm;
+
+static std::pair<Type *, Align> getArgumentTypeAlign(const Argument &Arg,
+                                                     const DataLayout &DL) {
+  Type *Ty = Arg.getType();
+  MaybeAlign ArgAlign;
+  if (Arg.hasByRefAttr()) {
+    Ty = Arg.getParamByRefType();
+    ArgAlign = Arg.getParamAlign();
+  }
+
+  if (!ArgAlign)
+    ArgAlign = DL.getABITypeAlign(Ty);
+
+  return std::make_pair(Ty, *ArgAlign);
+}
+
 namespace llvm {
 
 static cl::opt<bool> DumpHSAMetadata(
     "amdgpu-dump-hsa-metadata",
     cl::desc("Dump AMDGPU HSA Metadata"));
 static cl::opt<bool> VerifyHSAMetadata(
     "amdgpu-verify-hsa-metadata",
     cl::desc("Verify AMDGPU HSA Metadata"));
if (Node && ArgNo < Node->getNumOperands())
       AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
   }
 
   StringRef TypeQual;
   Node = Func->getMetadata("kernel_arg_type_qual");
   if (Node && ArgNo < Node->getNumOperands())
     TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
 
-  Type *Ty = Arg.getType();
   const DataLayout &DL = Func->getParent()->getDataLayout();
 
   MaybeAlign PointeeAlign;
-  if (auto PtrTy = dyn_cast<PointerType>(Ty)) {
+  if (auto PtrTy = dyn_cast<PointerType>(Arg.getType())) {
     if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
+      // FIXME: Should report this for all address spaces
       PointeeAlign = DL.getValueOrABITypeAlignment(Arg.getParamAlign(),
                                                    PtrTy->getElementType());
     }
   }
 
-  emitKernelArg(DL, Ty, getValueKind(Arg.getType(), TypeQual, BaseTypeName),
-                PointeeAlign, Name, TypeName, BaseTypeName, AccQual, TypeQual);
+  Type *ArgTy;
+  Align ArgAlign;
+  std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
+
+  emitKernelArg(DL, ArgTy, ArgAlign,
+                getValueKind(ArgTy, TypeQual, BaseTypeName), PointeeAlign, Name,
+                TypeName, BaseTypeName, AccQual, TypeQual);
 }
 
 void MetadataStreamerV2::emitKernelArg(const DataLayout &DL, Type *Ty,
-                                       ValueKind ValueKind,
+                                       Align Alignment, ValueKind ValueKind,
                                        MaybeAlign PointeeAlign, StringRef Name,
                                        StringRef TypeName,
                                        StringRef BaseTypeName,
                                        StringRef AccQual, StringRef TypeQual) {
   HSAMetadata.mKernels.back().mArgs.push_back(Kernel::Arg::Metadata());
   auto &Arg = HSAMetadata.mKernels.back().mArgs.back();
 
   Arg.mName = std::string(Name);
   Arg.mTypeName = std::string(TypeName);
   Arg.mSize = DL.getTypeAllocSize(Ty);
-  Arg.mAlign = DL.getABITypeAlign(Ty).value();
+  Arg.mAlign = Alignment.value();
   Arg.mValueKind = ValueKind;
   Arg.mValueType = getValueType(Ty, BaseTypeName);
   Arg.mPointeeAlign = PointeeAlign ? PointeeAlign->value() : 0;
 
   if (auto PtrTy = dyn_cast<PointerType>(Ty))
     Arg.mAddrSpaceQual = getAddressSpaceQualifier(PtrTy->getAddressSpace());
 
   Arg.mAccQual = getAccessQualifier(AccQual);
void MetadataStreamerV2::emitHiddenKernelArgs(const Function &Func) {
 
   if (!HiddenArgNumBytes)
     return;
 
   auto &DL = Func.getParent()->getDataLayout();
   auto Int64Ty = Type::getInt64Ty(Func.getContext());
 
   if (HiddenArgNumBytes >= 8)
-    emitKernelArg(DL, Int64Ty, ValueKind::HiddenGlobalOffsetX);
+    emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetX);
   if (HiddenArgNumBytes >= 16)
-    emitKernelArg(DL, Int64Ty, ValueKind::HiddenGlobalOffsetY);
+    emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetY);
   if (HiddenArgNumBytes >= 24)
-    emitKernelArg(DL, Int64Ty, ValueKind::HiddenGlobalOffsetZ);
+    emitKernelArg(DL, Int64Ty, Align(8), ValueKind::HiddenGlobalOffsetZ);
 
   auto Int8PtrTy = Type::getInt8PtrTy(Func.getContext(),
                                       AMDGPUAS::GLOBAL_ADDRESS);
 
   // Emit "printf buffer" argument if printf is used, otherwise emit dummy
   // "none" argument.
   if (HiddenArgNumBytes >= 32) {
     if (Func.getParent()->getNamedMetadata("llvm.printf.fmts"))
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenPrintfBuffer);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenPrintfBuffer);
     else if (Func.getParent()->getFunction("__ockl_hostcall_internal")) {
       // The printf runtime binding pass should have ensured that hostcall and
       // printf are not used in the same module.
       assert(!Func.getParent()->getNamedMetadata("llvm.printf.fmts"));
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenHostcallBuffer);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenHostcallBuffer);
     } else
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenNone);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
   }
 
   // Emit "default queue" and "completion action" arguments if enqueue kernel is
   // used, otherwise emit dummy "none" arguments.
   if (HiddenArgNumBytes >= 48) {
     if (Func.hasFnAttribute("calls-enqueue-kernel")) {
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenDefaultQueue);
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenCompletionAction);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenDefaultQueue);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenCompletionAction);
     } else {
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenNone);
-      emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenNone);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
+      emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenNone);
     }
   }
 
   // Emit the pointer argument for multi-grid object.
   if (HiddenArgNumBytes >= 56)
-    emitKernelArg(DL, Int8PtrTy, ValueKind::HiddenMultiGridSyncArg);
+    emitKernelArg(DL, Int8PtrTy, Align(8), ValueKind::HiddenMultiGridSyncArg);
 }
 
 bool MetadataStreamerV2::emitTo(AMDGPUTargetStreamer &TargetStreamer) {
   return TargetStreamer.EmitHSAMetadata(getHSAMetadata());
 }
 
 void MetadataStreamerV2::begin(const Module &Mod) {
   emitVersion();
if (Node && ArgNo < Node->getNumOperands())
       AccQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
   }
 
   StringRef TypeQual;
   Node = Func->getMetadata("kernel_arg_type_qual");
   if (Node && ArgNo < Node->getNumOperands())
     TypeQual = cast<MDString>(Node->getOperand(ArgNo))->getString();
 
-  Type *Ty = Arg.getType();
   const DataLayout &DL = Func->getParent()->getDataLayout();
 
   MaybeAlign PointeeAlign;
+  Type *Ty = Arg.hasByRefAttr() ? Arg.getParamByRefType() : Arg.getType();
+
+  // FIXME: Need to distinguish in memory alignment from pointer alignment.
   if (auto PtrTy = dyn_cast<PointerType>(Ty)) {
     if (PtrTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
       PointeeAlign = DL.getValueOrABITypeAlignment(Arg.getParamAlign(),
                                                    PtrTy->getElementType());
     }
   }
 
-  emitKernelArg(Func->getParent()->getDataLayout(), Arg.getType(),
-                getValueKind(Arg.getType(), TypeQual, BaseTypeName), Offset,
-                Args, PointeeAlign, Name, TypeName, BaseTypeName, AccQual,
-                TypeQual);
+  // There's no distinction between byval aggregates and raw aggregates.
+  Type *ArgTy;
+  Align ArgAlign;
+  std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
+
+  emitKernelArg(DL, ArgTy, ArgAlign,
+                getValueKind(ArgTy, TypeQual, BaseTypeName), Offset, Args,
+                PointeeAlign, Name, TypeName, BaseTypeName, AccQual, TypeQual);
 }
 
 void MetadataStreamerV3::emitKernelArg(const DataLayout &DL, Type *Ty,
-                                       StringRef ValueKind, unsigned &Offset,
-                                       msgpack::ArrayDocNode Args,
+                                       Align Alignment, StringRef ValueKind,
+                                       unsigned &Offset, msgpack::ArrayDocNode Args,
                                        MaybeAlign PointeeAlign, StringRef Name,
                                        StringRef TypeName,
                                        StringRef BaseTypeName,
                                        StringRef AccQual, StringRef TypeQual) {
   auto Arg = Args.getDocument()->getMapNode();
 
   if (!Name.empty())
     Arg[".name"] = Arg.getDocument()->getNode(Name, /*Copy=*/true);
   if (!TypeName.empty())
     Arg[".type_name"] = Arg.getDocument()->getNode(TypeName, /*Copy=*/true);
   auto Size = DL.getTypeAllocSize(Ty);
-  Align Alignment = DL.getABITypeAlign(Ty);
   Arg[".size"] = Arg.getDocument()->getNode(Size);
   Offset = alignTo(Offset, Alignment);
   Arg[".offset"] = Arg.getDocument()->getNode(Offset);
   Offset += Size;
   Arg[".value_kind"] = Arg.getDocument()->getNode(ValueKind, /*Copy=*/true);
   Arg[".value_type"] =
       Arg.getDocument()->getNode(getValueType(Ty, BaseTypeName), /*Copy=*/true);
   if (PointeeAlign)
void MetadataStreamerV3::emitHiddenKernelArgs(const Function &Func,
 
   if (!HiddenArgNumBytes)
     return;
 
   auto &DL = Func.getParent()->getDataLayout();
   auto Int64Ty = Type::getInt64Ty(Func.getContext());
 
   if (HiddenArgNumBytes >= 8)
-    emitKernelArg(DL, Int64Ty, "hidden_global_offset_x", Offset, Args);
+    emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_x", Offset, Args);
   if (HiddenArgNumBytes >= 16)
-    emitKernelArg(DL, Int64Ty, "hidden_global_offset_y", Offset, Args);
+    emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_y", Offset, Args);
   if (HiddenArgNumBytes >= 24)
-    emitKernelArg(DL, Int64Ty, "hidden_global_offset_z", Offset, Args);
+    emitKernelArg(DL, Int64Ty, Align(8), "hidden_global_offset_z", Offset, Args);
 
   auto Int8PtrTy =
       Type::getInt8PtrTy(Func.getContext(), AMDGPUAS::GLOBAL_ADDRESS);
 
   // Emit "printf buffer" argument if printf is used, otherwise emit dummy
   // "none" argument.
   if (HiddenArgNumBytes >= 32) {
     if (Func.getParent()->getNamedMetadata("llvm.printf.fmts"))
-      emitKernelArg(DL, Int8PtrTy, "hidden_printf_buffer", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_printf_buffer", Offset, Args);
     else if (Func.getParent()->getFunction("__ockl_hostcall_internal")) {
       // The printf runtime binding pass should have ensured that hostcall and
       // printf are not used in the same module.
       assert(!Func.getParent()->getNamedMetadata("llvm.printf.fmts"));
-      emitKernelArg(DL, Int8PtrTy, "hidden_hostcall_buffer", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_hostcall_buffer", Offset, Args);
     } else
-      emitKernelArg(DL, Int8PtrTy, "hidden_none", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
   }
 
   // Emit "default queue" and "completion action" arguments if enqueue kernel is
   // used, otherwise emit dummy "none" arguments.
   if (HiddenArgNumBytes >= 48) {
     if (Func.hasFnAttribute("calls-enqueue-kernel")) {
-      emitKernelArg(DL, Int8PtrTy, "hidden_default_queue", Offset, Args);
-      emitKernelArg(DL, Int8PtrTy, "hidden_completion_action", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_default_queue", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_completion_action", Offset, Args);
     } else {
-      emitKernelArg(DL, Int8PtrTy, "hidden_none", Offset, Args);
-      emitKernelArg(DL, Int8PtrTy, "hidden_none", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
+      emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_none", Offset, Args);
     }
   }
 
   // Emit the pointer argument for multi-grid object.
   if (HiddenArgNumBytes >= 56)
-    emitKernelArg(DL, Int8PtrTy, "hidden_multigrid_sync_arg", Offset, Args);
+    emitKernelArg(DL, Int8PtrTy, Align(8), "hidden_multigrid_sync_arg", Offset, Args);
 }
 
 msgpack::MapDocNode
 MetadataStreamerV3::getHSAKernelProps(const MachineFunction &MF,
                                       const SIProgramInfo &ProgramInfo) const {
   const GCNSubtarget &STM = MF.getSubtarget<GCNSubtarget>();
   const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
   const Function &F = MF.getFunction();

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(
 
   Align MaxAlign = Align(1);
   uint64_t ExplicitArgOffset = 0;
   const DataLayout &DL = Fn.getParent()->getDataLayout();
 
   unsigned InIndex = 0;
 
   for (const Argument &Arg : Fn.args()) {
+    const bool IsByRef = Arg.hasByRefAttr();
     Type *BaseArgTy = Arg.getType();
-    Align Alignment = DL.getABITypeAlign(BaseArgTy);
-    MaxAlign = std::max(Alignment, MaxAlign);
-    unsigned AllocSize = DL.getTypeAllocSize(BaseArgTy);
+    Type *MemArgTy = IsByRef ? Arg.getParamByRefType() : BaseArgTy;
+    MaybeAlign Alignment = IsByRef ? Arg.getParamAlign() : None;
+    if (!Alignment)
+      Alignment = DL.getABITypeAlign(MemArgTy);
+    MaxAlign = max(Alignment, MaxAlign);
+    uint64_t AllocSize = DL.getTypeAllocSize(MemArgTy);
 
     uint64_t ArgOffset = alignTo(ExplicitArgOffset, Alignment) + ExplicitOffset;
     ExplicitArgOffset = alignTo(ExplicitArgOffset, Alignment) + AllocSize;
 
     // We're basically throwing away everything passed into us and starting over
     // to get accurate in-memory offsets. The "PartOffset" is completely useless
     // to us as computed in Ins.
     //

llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp

bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
   KernArgSegment->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull);
   KernArgSegment->addAttribute(AttributeList::ReturnIndex,
     Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize));
 
   unsigned AS = KernArgSegment->getType()->getPointerAddressSpace();
   uint64_t ExplicitArgOffset = 0;
 
   for (Argument &Arg : F.args()) {
-    Type *ArgTy = Arg.getType();
-    Align ABITypeAlign = DL.getABITypeAlign(ArgTy);
-    unsigned Size = DL.getTypeSizeInBits(ArgTy);
-    unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
+    const bool IsByRef = Arg.hasByRefAttr();
+    Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
+    MaybeAlign ABITypeAlign = IsByRef ? Arg.getParamAlign() : None;
+    if (!ABITypeAlign)
+      ABITypeAlign = DL.getABITypeAlign(ArgTy);
+
+    uint64_t Size = DL.getTypeSizeInBits(ArgTy);
+    uint64_t AllocSize = DL.getTypeAllocSize(ArgTy);
 
     uint64_t EltOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + BaseOffset;
     ExplicitArgOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + AllocSize;
 
     if (Arg.use_empty())
       continue;
 
+    // If this is byval, the loads are already explicit in the function. We just
+    // need to rewrite the pointer values.
+    if (IsByRef) {
+      Value *ArgOffsetPtr = Builder.CreateConstInBoundsGEP1_64(
+          Builder.getInt8Ty(), KernArgSegment, EltOffset,
+          Arg.getName() + ".byval.kernarg.offset");
+
+      Value *CastOffsetPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
+          ArgOffsetPtr, Arg.getType());
+      Arg.replaceAllUsesWith(CastOffsetPtr);
+      continue;
+    }
+
     if (PointerType *PT = dyn_cast<PointerType>(ArgTy)) {
       // FIXME: Hack. We rely on AssertZext to be able to fold DS addressing
       // modes on SI to know the high bits are 0 so pointer adds don't wrap. We
       // can't represent this with range metadata because it's only allowed for
       // integer types.
       if ((PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
            PT->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) &&
           !ST.hasUsableDSOffset())

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

uint64_t AMDGPUSubtarget::getExplicitKernArgSize(const Function &F,
   assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
          F.getCallingConv() == CallingConv::SPIR_KERNEL);
 
   const DataLayout &DL = F.getParent()->getDataLayout();
   uint64_t ExplicitArgBytes = 0;
   MaxAlign = Align(1);
 
   for (const Argument &Arg : F.args()) {
-    Type *ArgTy = Arg.getType();
+    const bool IsByRef = Arg.hasByRefAttr();
+    Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
+    MaybeAlign Alignment = IsByRef ? Arg.getParamAlign() : None;
+    if (!Alignment)
+      Alignment = DL.getABITypeAlign(ArgTy);
 
-    const Align Alignment = DL.getABITypeAlign(ArgTy);
     uint64_t AllocSize = DL.getTypeAllocSize(ArgTy);
     ExplicitArgBytes = alignTo(ExplicitArgBytes, Alignment) + AllocSize;
-    MaxAlign = std::max(MaxAlign, Alignment);
+    MaxAlign = max(MaxAlign, Alignment);
   }
 
   return ExplicitArgBytes;
 }
 
 unsigned AMDGPUSubtarget::getKernArgSegmentSize(const Function &F,
                                                 Align &MaxAlign) const {
   uint64_t ExplicitArgBytes = getExplicitKernArgSize(F, MaxAlign);

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
 
     if (IsEntryFunc && VA.isMemLoc()) {
       VT = Ins[i].VT;
       EVT MemVT = VA.getLocVT();
 
       const uint64_t Offset = VA.getLocMemOffset();
       Align Alignment = commonAlignment(KernelArgBaseAlign, Offset);
 
-      SDValue Arg =
-          lowerKernargMemParameter(DAG, VT, MemVT, DL, Chain, Offset, Alignment,
-                                   Ins[i].Flags.isSExt(), &Ins[i]);
+      if (Arg.Flags.isByRef()) {
+        SDValue Ptr = lowerKernArgParameterPtr(DAG, DL, Chain, Offset);
+
+        if (!isNoopAddrSpaceCast(AMDGPUAS::CONSTANT_ADDRESS,
+                                 Arg.Flags.getPointerAddrSpace())) {
+          Ptr = DAG.getAddrSpaceCast(DL, VT, Ptr, AMDGPUAS::CONSTANT_ADDRESS,
+                                     Arg.Flags.getPointerAddrSpace());
+        }
+
+        InVals.push_back(Ptr);
+        continue;
+      }
+
+      SDValue Arg = lowerKernargMemParameter(
+        DAG, VT, MemVT, DL, Chain, Offset, Alignment, Ins[i].Flags.isSExt(), &Ins[i]);
       Chains.push_back(Arg.getValue(1));
 
       auto *ParamTy =
         dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
       if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
           ParamTy && (ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
                       ParamTy->getAddressSpace() == AMDGPUAS::REGION_ADDRESS)) {
         // On SI local pointers are just offsets into LDS, so they are always

llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-amdgpu_kernel.ll

define amdgpu_kernel void @unused_i32_arg(i32 addrspace(1)* nocapture %out, i32 %unused, i32 %in) nounwind {
   ; LEGACY-MESA-VI-LABEL: name: unused_i32_arg
   ; LEGACY-MESA-VI: bb.1.entry:
   ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
   ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
   ; LEGACY-MESA-VI:   S_ENDPGM 0
 entry:
   ret void
 }
+
+; Byref pointers should only be treated as offsets from kernarg
+define amdgpu_kernel void @byref_constant_i8_arg(i32 addrspace(1)* nocapture %out, i8 addrspace(4)* byref(i8) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_constant_i8_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD1]](p4) :: (load 1 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[LOAD1]](s8)
+  ; HSA-VI:   G_STORE [[ZEXT]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_i8_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s8) = G_LOAD [[PTR_ADD1]](p4) :: (load 1 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[LOAD1]](s8)
+  ; LEGACY-MESA-VI:   G_STORE [[ZEXT]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i8, i8 addrspace(4)* %in.byref
+  %ext = zext i8 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i16_arg(i32 addrspace(1)* nocapture %out, i16 addrspace(4)* byref(i16) align 2 %in.byref) {
+  ; HSA-VI-LABEL: name: byref_constant_i16_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s16) = G_LOAD [[PTR_ADD1]](p4) :: (load 2 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[LOAD1]](s16)
+  ; HSA-VI:   G_STORE [[ZEXT]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_i16_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s16) = G_LOAD [[PTR_ADD1]](p4) :: (load 2 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[LOAD1]](s16)
+  ; LEGACY-MESA-VI:   G_STORE [[ZEXT]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i16, i16 addrspace(4)* %in.byref
+  %ext = zext i16 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align 4 %in.byref, i32 %after.offset) {
+  ; HSA-VI-LABEL: name: byref_constant_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+  ; HSA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, addrspace 4)
+  ; HSA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+  ; LEGACY-MESA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, align 16, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_v4i32_arg(<4 x i32> addrspace(1)* nocapture %out, <4 x i32> addrspace(4)* byref(<4 x i32>) align(16) %in.byref, i32 %after.offset) {
+  ; HSA-VI-LABEL: name: byref_constant_v4i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+  ; HSA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, align 16, addrspace 4)
+  ; HSA-VI:   [[LOAD2:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD1]](p4) :: (load 16 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD2]](<4 x s32>), [[LOAD]](p1) :: (volatile store 16 into %ir.out, align 4, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out.cast, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_v4i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 68
+  ; LEGACY-MESA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD2:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD1]](p4) :: (load 16 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD2]](<4 x s32>), [[LOAD]](p1) :: (volatile store 16 into %ir.out, align 4, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out.cast, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load <4 x i32>, <4 x i32> addrspace(4)* %in.byref
+  store volatile <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
+  %out.cast = bitcast <4 x i32> addrspace(1)* %out to i32 addrspace(1)*
+  store volatile i32 %after.offset, i32 addrspace(1)* %out.cast, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_align_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(256) %in.byref, i32 %after.offset) {
+  ; HSA-VI-LABEL: name: byref_align_constant_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 256
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 260
+  ; HSA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, addrspace 4)
+  ; HSA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_align_constant_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 292
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 296
+  ; LEGACY-MESA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, align 8, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) align(64) %in.byref, i32 %after.offset) {
+  ; HSA-VI-LABEL: name: byref_natural_align_constant_v16i32_arg
+  ; HSA-VI: bb.1 (%ir-block.1):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 64
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 128
+  ; HSA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, align 16, addrspace 4)
+  ; HSA-VI:   [[LOAD2:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[PTR_ADD1]](p4) :: (load 64 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD2]](<16 x s32>), [[LOAD]](p1) :: (volatile store 64 into %ir.cast.out, align 4, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_natural_align_constant_v16i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.1):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 100
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 164
+  ; LEGACY-MESA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (dereferenceable invariant load 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD2:%[0-9]+]]:_(<16 x s32>) = G_LOAD [[PTR_ADD1]](p4) :: (load 64 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD2]](<16 x s32>), [[LOAD]](p1) :: (volatile store 64 into %ir.cast.out, align 4, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
+  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
+  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Also accept byref kernel arguments with other global address spaces.
+define amdgpu_kernel void @byref_global_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_global_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p1) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[COPY1:%[0-9]+]]:_(p1) = COPY [[ADDRSPACE_CAST]](p1)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[COPY1]](p1) :: (load 4 from %ir.1, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_global_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p1) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[COPY1:%[0-9]+]]:_(p1) = COPY [[ADDRSPACE_CAST]](p1)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[COPY1]](p1) :: (load 4 from %ir.1, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(1)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_flat_i32_arg(i32 addrspace(1)* nocapture %out, i32* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_flat_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p0) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p0) :: (load 4 from %ir.in.byref)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_flat_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p0) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p0) :: (load 4 from %ir.in.byref)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_32bit_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(6)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_constant_32bit_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p6) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p6) :: (load 4 from %ir.in.byref, addrspace 6)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_32bit_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p6) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p6) :: (load 4 from %ir.in.byref, addrspace 6)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(6)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_unknown_as_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(999)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_unknown_as_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p999) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p999) :: (load 4 from %ir.in.byref, addrspace 999)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_unknown_as_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p999) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p999) :: (load 4 from %ir.in.byref, addrspace 999)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(999)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Invalid, but should not crash.
+define amdgpu_kernel void @byref_private_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(5)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_private_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p5) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p5) :: (load 4 from %ir.in.byref, addrspace 5)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_private_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p5) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p5) :: (load 4 from %ir.in.byref, addrspace 5)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(5)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Invalid, but should not crash.
+define amdgpu_kernel void @byref_local_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(3)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_local_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p3) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p3) :: (load 4 from %ir.in.byref, addrspace 3)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_local_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[ADDRSPACE_CAST:%[0-9]+]]:_(p3) = G_ADDRSPACE_CAST [[PTR_ADD1]](p4)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[ADDRSPACE_CAST]](p3) :: (load 4 from %ir.in.byref, addrspace 3)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(3)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @multi_byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(4) %in0.byref, i32 addrspace(4)* byref(i32) align(4) %in1.byref, i32 %after.offset) {
+  ; HSA-VI-LABEL: name: multi_byref_constant_i32_arg
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 16, addrspace 4)
+  ; HSA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 8
+  ; HSA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; HSA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 12
+  ; HSA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; HSA-VI:   [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+  ; HSA-VI:   [[PTR_ADD3:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C3]](s64)
+  ; HSA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p4) :: (dereferenceable invariant load 4, align 16, addrspace 4)
+  ; HSA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in0.byref, addrspace 4)
+  ; HSA-VI:   [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (load 4 from %ir.in1.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD3]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: multi_byref_constant_i32_arg
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(p1) = G_LOAD [[PTR_ADD]](p4) :: (dereferenceable invariant load 8, align 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 44
+  ; LEGACY-MESA-VI:   [[PTR_ADD1:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C1]](s64)
+  ; LEGACY-MESA-VI:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 48
+  ; LEGACY-MESA-VI:   [[PTR_ADD2:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C2]](s64)
+  ; LEGACY-MESA-VI:   [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 52
+  ; LEGACY-MESA-VI:   [[PTR_ADD3:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C3]](s64)
+  ; LEGACY-MESA-VI:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p4) :: (dereferenceable invariant load 4, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p4) :: (load 4 from %ir.in0.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p4) :: (load 4 from %ir.in1.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD2]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD3]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD1]](s32), [[LOAD]](p1) :: (volatile store 4 into %ir.out, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in0 = load i32, i32 addrspace(4)* %in0.byref
+  %in1 = load i32, i32 addrspace(4)* %in1.byref
+  store volatile i32 %in0, i32 addrspace(1)* %out, align 4
+  store volatile i32 %in1, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i32_arg_offset0(i32 addrspace(4)* byref(i32) align(4) %in.byref) {
+  ; HSA-VI-LABEL: name: byref_constant_i32_arg_offset0
+  ; HSA-VI: bb.1 (%ir-block.0):
+  ; HSA-VI:   liveins: $sgpr4_sgpr5
+  ; HSA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr4_sgpr5
+  ; HSA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+  ; HSA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; HSA-VI:   [[DEF:%[0-9]+]]:_(p1) = G_IMPLICIT_DEF
+  ; HSA-VI:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; HSA-VI:   G_STORE [[LOAD]](s32), [[DEF]](p1) :: (store 4 into `i32 addrspace(1)* undef`, addrspace 1)
+  ; HSA-VI:   S_ENDPGM 0
+  ; LEGACY-MESA-VI-LABEL: name: byref_constant_i32_arg_offset0
+  ; LEGACY-MESA-VI: bb.1 (%ir-block.0):
+  ; LEGACY-MESA-VI:   liveins: $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[COPY:%[0-9]+]]:_(p4) = COPY $sgpr0_sgpr1
+  ; LEGACY-MESA-VI:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 36
+  ; LEGACY-MESA-VI:   [[PTR_ADD:%[0-9]+]]:_(p4) = G_PTR_ADD [[COPY]], [[C]](s64)
+  ; LEGACY-MESA-VI:   [[DEF:%[0-9]+]]:_(p1) = G_IMPLICIT_DEF
+  ; LEGACY-MESA-VI:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p4) :: (load 4 from %ir.in.byref, addrspace 4)
+  ; LEGACY-MESA-VI:   G_STORE [[LOAD]](s32), [[DEF]](p1) :: (store 4 into `i32 addrspace(1)* undef`, addrspace 1)
+  ; LEGACY-MESA-VI:   S_ENDPGM 0
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store i32 %in, i32 addrspace(1)* undef, align 4
+  ret void
+}

llvm/test/CodeGen/AMDGPU/hsa-metadata-from-llvm-ir-full-v3.ll

define amdgpu_kernel void @test_char(i8 %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !9
     !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
   ret void
 }
 
 ; CHECK:        - .args:
 ; CHECK-NEXT:       - .name:           a
 ; CHECK-NEXT:         .offset:         0
+; CHECK-NEXT:         .size:           1
+; CHECK-NEXT:         .type_name:      char
+; CHECK-NEXT:         .value_kind:     by_value
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .offset:         8
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         16
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         24
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         32
+; CHECK-NEXT:         .size:           8
+; CHECK-NOT:          .value_kind:     hidden_default_queue
+; CHECK-NOT:          .value_kind:     hidden_completion_action
+; CHECK-NOT:          .value_kind:     hidden_hostcall_buffer
+; CHECK-NEXT:         .value_kind:     hidden_printf_buffer
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:              .value_kind:     hidden_multigrid_sync_arg
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:          .language:       OpenCL C
+; CHECK-NEXT:     .language_version:
+; CHECK-NEXT:       - 2
+; CHECK-NEXT:       - 0
+; CHECK:          .name:           test_char_byref_constant
+; CHECK:          .symbol:         test_char_byref_constant.kd
+define amdgpu_kernel void @test_char_byref_constant(i8 addrspace(4)* byref(i8) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !9
+    !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
+  ret void
+}
+
+; CHECK:        - .args:
+; CHECK-NEXT:       - .offset:         0
+; CHECK-NEXT:         .size:           1
+; CHECK-NEXT:         .type_name:      char
+; CHECK-NEXT:         .value_kind:     by_value
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .name:           a
+; CHECK-NEXT:         .offset:         512
+; CHECK-NEXT:         .size:           1
+; CHECK-NEXT:         .type_name:      char
+; CHECK-NEXT:         .value_kind:     by_value
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .offset:         520
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         528
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         536
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         544
+; CHECK-NEXT:         .size:           8
+; CHECK-NOT:          .value_kind:     hidden_default_queue
+; CHECK-NOT:          .value_kind:     hidden_completion_action
+; CHECK-NOT:          .value_kind:     hidden_hostcall_buffer
+; CHECK-NEXT:         .value_kind:     hidden_printf_buffer
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:              .value_kind:     hidden_multigrid_sync_arg
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:          .language:       OpenCL C
+; CHECK-NEXT:     .language_version:
+; CHECK-NEXT:       - 2
+; CHECK-NEXT:       - 0
+; CHECK:          .name:           test_char_byref_constant_align512
+; CHECK:          .symbol:         test_char_byref_constant_align512.kd
+define amdgpu_kernel void @test_char_byref_constant_align512(i8, i8 addrspace(4)* byref(i8) align(512) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !111
+    !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
+  ret void
+}
+
+; CHECK:        - .args:
+; CHECK-NEXT:       - .name:           a
+; CHECK-NEXT:         .offset:         0
 ; CHECK-NEXT:         .size:           4
 ; CHECK-NEXT:         .type_name:      ushort2
 ; CHECK-NEXT:         .value_kind:     by_value
 ; CHECK-NEXT:         .value_type:     u16
 ; CHECK-NEXT:       - .offset:         8
 ; CHECK-NEXT:         .size:           8
 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
 ; CHECK-NEXT:         .value_type:     i64
define amdgpu_kernel void @test_struct(%struct.A %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
     !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
   ret void
 }
 
 ; CHECK:        - .args:
 ; CHECK-NEXT:         .name:           a
 ; CHECK-NEXT:         .offset:         0
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .type_name:      struct A
+; CHECK-NEXT:         .value_kind:     by_value
+; CHECK-NEXT:         .value_type:     struct
+; CHECK-NEXT:       - .offset:         8
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         16
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         24
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         32
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_printf_buffer
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         40
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_none
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         48
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_none
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         56
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_multigrid_sync_arg
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:          .language:       OpenCL C
+; CHECK-NEXT:     .language_version:
+; CHECK-NEXT:       - 2
+; CHECK-NEXT:       - 0
+; CHECK:          .name:           test_struct_byref_constant
+; CHECK:          .symbol:         test_struct_byref_constant.kd
+define amdgpu_kernel void @test_struct_byref_constant(%struct.A addrspace(4)* byref(%struct.A) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
+    !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
+  ret void
+}
+
+; CHECK:        - .args:
+; CHECK-NEXT:         .name:           a
+; CHECK-NEXT:         .offset:         0
 ; CHECK-NEXT:         .size:           32
 ; CHECK-NEXT:         .type_name:      struct A
 ; CHECK-NEXT:         .value_kind:     by_value
 ; CHECK-NEXT:         .value_type:     struct
 ; CHECK-NEXT:       - .offset:         32
 ; CHECK-NEXT:         .size:           8
 ; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
 ; CHECK-NEXT:         .value_type:     i64

 ; CHECK:          .symbol:         test_array.kd
 define amdgpu_kernel void @test_array([32 x i8] %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
     !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
   ret void
 }
 
 ; CHECK:        - .args:
+; CHECK-NEXT:         .name:           a
+; CHECK-NEXT:         .offset:         0
+; CHECK-NEXT:         .size:           32
+; CHECK-NEXT:         .type_name:      struct A
+; CHECK-NEXT:         .value_kind:     by_value
+; CHECK-NEXT:         .value_type:     struct
+; CHECK-NEXT:       - .offset:         32
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_x
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         40
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_y
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .offset:         48
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_global_offset_z
+; CHECK-NEXT:         .value_type:     i64
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         56
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_printf_buffer
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         64
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_none
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         72
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_none
+; CHECK-NEXT:         .value_type:     i8
+; CHECK-NEXT:       - .address_space:  global
+; CHECK-NEXT:         .offset:         80
+; CHECK-NEXT:         .size:           8
+; CHECK-NEXT:         .value_kind:     hidden_multigrid_sync_arg
+; CHECK-NEXT:         .value_type:     i8
+; CHECK:          .language:       OpenCL C
+; CHECK-NEXT:     .language_version:
+; CHECK-NEXT:       - 2
+; CHECK-NEXT:       - 0
+; CHECK:          .name:           test_array_byref_constant
+; CHECK:          .symbol:         test_array_byref_constant.kd
+define amdgpu_kernel void @test_array_byref_constant([32 x i8] addrspace(4)* byref([32 x i8]) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
+    !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
+  ret void
+}
+
+; CHECK:        - .args:
 ; CHECK-NEXT:       - .name:           a
 ; CHECK-NEXT:         .offset:         0
 ; CHECK-NEXT:         .size:           16
 ; CHECK-NEXT:         .type_name:      i128
 ; CHECK-NEXT:         .value_kind:     by_value
 ; CHECK-NEXT:         .value_type:     struct
 ; CHECK-NEXT:       - .offset:         16
 ; CHECK-NEXT:         .size:           8

 !90 = !{i32 2, i32 0}
 !91 = !{i32 0, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3}
 !92 = !{!"none", !"none", !"none", !"none", !"none", !"none", !"none"}
 !93 = !{!"long  addrspace(5)*", !"char  addrspace(5)*", !"char2  addrspace(5)*", !"char3  addrspace(5)*", !"char4  addrspace(5)*", !"char8  addrspace(5)*", !"char16  addrspace(5)*"}
 !94 = !{!"", !"", !"", !"", !"", !"", !""}
 !100 = !{!"1:1:4:%d\5Cn"}
 !101 = !{!"2:1:8:%g\5Cn"}
 !110 = !{!"__block_literal"}
+!111 = !{!"char", !"char"}
 
 ; PARSER: AMDGPU HSA Metadata Parser Test: PASS

llvm/test/CodeGen/AMDGPU/hsa-metadata-from-llvm-ir-full.ll

 ; CHECK-NEXT:       ValueType:     I8
 ; CHECK-NEXT:       AddrSpaceQual: Global
 define amdgpu_kernel void @test_char(i8 %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !9
     !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
   ret void
 }
 
+; CHECK:      - Name:            test_char_byref_constant
+; CHECK-NEXT:   SymbolName:      'test_char_byref_constant@kd'
+; CHECK-NEXT:   Language:        OpenCL C
+; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
+; CHECK-NEXT:   Args:
+; CHECK-NEXT:     - Name:          a
+; CHECK-NEXT:       TypeName:      char
+; CHECK-NEXT:       Size:          1
+; CHECK-NEXT:       Align:         1
+; CHECK-NEXT:       ValueKind:     ByValue
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AccQual:       Default
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetX
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetY
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetZ
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NOT:        ValueKind:     HiddenHostcallBuffer
+; CHECK-NEXT:       ValueKind:     HiddenPrintfBuffer
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NOT:        ValueKind:     HiddenDefaultQueue
+; CHECK-NOT:        ValueKind:     HiddenCompletionAction
+; CHECK:            ValueKind:     HiddenMultiGridSyncArg
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+define amdgpu_kernel void @test_char_byref_constant(i8 addrspace(4)* byref(i8) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !9
+    !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
+  ret void
+}
+
+; CHECK:      - Name:            test_char_byref_constant_align512
+; CHECK-NEXT:   SymbolName:      'test_char_byref_constant_align512@kd'
+; CHECK-NEXT:   Language:        OpenCL C
+; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
+; CHECK-NEXT:   Args:
+; CHECK-NEXT:       TypeName:      char
+; CHECK-NEXT:       Size:          1
+; CHECK-NEXT:       Align:         1
+; CHECK-NEXT:       ValueKind:     ByValue
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AccQual:       Default
+; CHECK-NEXT:     - Name:          a
+; CHECK-NEXT:       TypeName:      char
+; CHECK-NEXT:       Size:          1
+; CHECK-NEXT:       Align:         512
+; CHECK-NEXT:       ValueKind:     ByValue
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetX
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetY
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetZ
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NOT:        ValueKind:     HiddenHostcallBuffer
+; CHECK-NEXT:       ValueKind:     HiddenPrintfBuffer
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NOT:        ValueKind:     HiddenDefaultQueue
+; CHECK-NOT:        ValueKind:     HiddenCompletionAction
+; CHECK:            ValueKind:     HiddenMultiGridSyncArg
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+define amdgpu_kernel void @test_char_byref_constant_align512(i8, i8 addrspace(4)* byref(i8) align 512 %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !111
+    !kernel_arg_base_type !9 !kernel_arg_type_qual !4 {
+  ret void
+}
+
 ; CHECK:      - Name:            test_ushort2
 ; CHECK-NEXT:   SymbolName:      'test_ushort2@kd'
 ; CHECK-NEXT:   Language:        OpenCL C
 ; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
 ; CHECK-NEXT:   Args:
 ; CHECK-NEXT:     - Name:          a
 ; CHECK-NEXT:       TypeName:      ushort2
 ; CHECK-NEXT:       Size:          4

 ; CHECK-NEXT:       ValueType:     I8
 ; CHECK-NEXT:       AddrSpaceQual: Global
 define amdgpu_kernel void @test_struct(%struct.A %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
     !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
   ret void
 }
 
+; CHECK:      - Name:            test_struct_byref_constant
+; CHECK-NEXT:   SymbolName:      'test_struct_byref_constant@kd'
+; CHECK-NEXT:   Language:        OpenCL C
+; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
+; CHECK-NEXT:   Args:
+; CHECK-NEXT:     - Name:          a
+; CHECK-NEXT:       TypeName:      struct A
+; CHECK-NEXT:       Size:          8
+; CHECK-NEXT:       Align:         4
+; CHECK-NEXT:       ValueKind:     ByValue
+; CHECK-NEXT:       ValueType:     Struct
+; CHECK-NEXT:       AccQual:       Default
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetX
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetY
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetZ
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenPrintfBuffer
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenNone
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenNone
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenMultiGridSyncArg
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+define amdgpu_kernel void @test_struct_byref_constant(%struct.A addrspace(4)* byref(%struct.A) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
+    !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
+  ret void
+}
+
 ; CHECK:      - Name:            test_array
 ; CHECK-NEXT:   SymbolName:      'test_array@kd'
 ; CHECK-NEXT:   Language:        OpenCL C
 ; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
 ; CHECK-NEXT:   Args:
 ; CHECK-NEXT:     - Name:          a
 ; CHECK-NEXT:       TypeName:      struct A
 ; CHECK-NEXT:       Size:          8

 ; CHECK-NEXT:       ValueType:     I8
 ; CHECK-NEXT:       AddrSpaceQual: Global
 define amdgpu_kernel void @test_array([8 x i8] %a) #0
     !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
     !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
   ret void
 }
 
+; CHECK:      - Name:            test_array_byref_constant
+; CHECK-NEXT:   SymbolName:      'test_array_byref_constant@kd'
+; CHECK-NEXT:   Language:        OpenCL C
+; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
+; CHECK-NEXT:   Args:
+; CHECK-NEXT:     - Name:          a
+; CHECK-NEXT:       TypeName:      struct A
+; CHECK-NEXT:       Size:          8
+; CHECK-NEXT:       Align:         1
+; CHECK-NEXT:       ValueKind:     ByValue
+; CHECK-NEXT:       ValueType:     Struct
+; CHECK-NEXT:       AccQual:       Default
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetX
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetY
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenGlobalOffsetZ
+; CHECK-NEXT:       ValueType:     I64
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenPrintfBuffer
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenNone
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenNone
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+; CHECK-NEXT:     - Size:          8
+; CHECK-NEXT:       Align:         8
+; CHECK-NEXT:       ValueKind:     HiddenMultiGridSyncArg
+; CHECK-NEXT:       ValueType:     I8
+; CHECK-NEXT:       AddrSpaceQual: Global
+define amdgpu_kernel void @test_array_byref_constant([8 x i8] addrspace(4)* byref([8 x i8]) %a) #0
+    !kernel_arg_addr_space !1 !kernel_arg_access_qual !2 !kernel_arg_type !20
+    !kernel_arg_base_type !20 !kernel_arg_type_qual !4 {
+  ret void
+}
+
 ; CHECK:      - Name:            test_i128
 ; CHECK-NEXT:   SymbolName:      'test_i128@kd'
 ; CHECK-NEXT:   Language:        OpenCL C
 ; CHECK-NEXT:   LanguageVersion: [ 2, 0 ]
 ; CHECK-NEXT:   Args:
 ; CHECK-NEXT:     - Name:          a
 ; CHECK-NEXT:       TypeName:      i128
 ; CHECK-NEXT:       Size:          16

 !90 = !{i32 2, i32 0}
 !91 = !{i32 0, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3}
 !92 = !{!"none", !"none", !"none", !"none", !"none", !"none", !"none"}
 !93 = !{!"long  addrspace(5)*", !"char  addrspace(5)*", !"char2  addrspace(5)*", !"char3  addrspace(5)*", !"char4  addrspace(5)*", !"char8  addrspace(5)*", !"char16  addrspace(5)*"}
 !94 = !{!"", !"", !"", !"", !"", !"", !""}
 !100 = !{!"1:1:4:%d\5Cn"}
 !101 = !{!"2:1:8:%g\5Cn"}
 !110 = !{!"__block_literal"}
-
+!111 = !{!"char", !"char"}
 ; PARSER: AMDGPU HSA Metadata Parser Test: PASS

llvm/test/CodeGen/AMDGPU/kernel-args.ll

 
 ; GCN-LABEL: {{^}}small_array_round_down_offset:
 ; HSA-GFX9: global_load_ubyte v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:1
 define amdgpu_kernel void @small_array_round_down_offset(i8, [1 x i8] %arg) {
   %val = extractvalue [1 x i8] %arg, 0
   store volatile i8 %val, i8 addrspace(1)* undef
   ret void
 }
+
+; GCN-LABEL: {{^}}byref_align_constant_i32_arg:
+; HSA-GFX9: kernarg_segment_byte_size = 264
+; HSA-GFX9-DAG: s_load_dwordx2 {{s\[[0-9]+:[0-9]+\]}}, s[4:5], 0x100{{$}}
+define amdgpu_kernel void @byref_align_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(256) %in.byref, i32 %after.offset) {
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_natural_align_constant_v16i32_arg:
+; HSA-GFX9: kernarg_segment_byte_size = 132
+; HSA-GFX9-DAG: s_load_dword s{{[0-9]+}}, s[4:5], 0x80
+; HSA-GFX9-DAG: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x40{{$}}
+define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) %in.byref, i32 %after.offset) {
+  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
+  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
+  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}

llvm/test/CodeGen/AMDGPU/kernel-argument-dag-lowering.ll

 ; GCN: s_and_b32
 ; GCN: s_and_b32
 ; GCN: s_or_b32
 define amdgpu_kernel void @v3i15_arg(<3 x i15> addrspace(1)* nocapture %out, <3 x i15> %in) {
 entry:
   store <3 x i15> %in, <3 x i15> addrspace(1)* %out, align 4
   ret void
 }
+
+; Byref pointers should only be treated as offsets from kernarg
+; GCN-LABEL: {{^}}byref_constant_i8_arg:
+; GCN: kernarg_segment_byte_size = 12
+; GCN: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s4
+; GCN: v_mov_b32_e32 v[[VPTR_HI:[0-9]+]], s5
+; GCN: global_load_ubyte v{{[0-9]+}}, v{{\[}}[[VPTR_LO]]:[[VPTR_HI]]{{\]}}, off offset:8
+define amdgpu_kernel void @byref_constant_i8_arg(i32 addrspace(1)* nocapture %out, i8 addrspace(4)* byref(i8) %in.byref) {
+  %in = load i8, i8 addrspace(4)* %in.byref
+  %ext = zext i8 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_constant_i16_arg:
+; GCN: kernarg_segment_byte_size = 12
+; GCN: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s4
+; GCN: v_mov_b32_e32 v[[VPTR_HI:[0-9]+]], s5
+; GCN: global_load_ushort v{{[0-9]+}}, v{{\[}}[[VPTR_LO]]:[[VPTR_HI]]{{\]}}, off offset:8
+define amdgpu_kernel void @byref_constant_i16_arg(i32 addrspace(1)* nocapture %out, i16 addrspace(4)* byref(i16) %in.byref) {
+  %in = load i16, i16 addrspace(4)* %in.byref
+  %ext = zext i16 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_constant_i32_arg:
+; GCN: kernarg_segment_byte_size = 16
+; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
+; GCN: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0xc{{$}}
+define amdgpu_kernel void @byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in.byref, i32 %after.offset) {
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_constant_v4i32_arg:
+; GCN: kernarg_segment_byte_size = 36
+; GCN: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x10{{$}}
+; GCN: s_load_dword s{{[0-9]+}}, s[4:5], 0x20{{$}}
+define amdgpu_kernel void @byref_constant_v4i32_arg(<4 x i32> addrspace(1)* nocapture %out, <4 x i32> addrspace(4)* byref(<4 x i32>) %in.byref, i32 %after.offset) {
+  %in = load <4 x i32>, <4 x i32> addrspace(4)* %in.byref
+  store volatile <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
+  %out.cast = bitcast <4 x i32> addrspace(1)* %out to i32 addrspace(1)*
+  store volatile i32 %after.offset, i32 addrspace(1)* %out.cast, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_align_constant_i32_arg:
+; GCN: kernarg_segment_byte_size = 264
+; GCN-DAG: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x100{{$}}
+; GCN-DAG: s_load_dword [[AFTER_OFFSET:s[0-9]+]], s[4:5], 0x104{{$}}
+; GCN-DAG: v_mov_b32_e32 [[V_IN:v[0-9]+]], [[IN]]
+; GCN-DAG: v_mov_b32_e32 [[V_AFTER_OFFSET:v[0-9]+]], [[AFTER_OFFSET]]
+; GCN: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_IN]]
+; GCN: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_AFTER_OFFSET]]
+define amdgpu_kernel void @byref_align_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(256) %in.byref, i32 %after.offset) {
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_natural_align_constant_v16i32_arg:
+; GCN: kernarg_segment_byte_size = 132
+; GCN-DAG: s_load_dword s{{[0-9]+}}, s[4:5], 0x80
+; GCN-DAG: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x40{{$}}
+define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) align(64) %in.byref, i32 %after.offset) {
+  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
+  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
+  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Also accept byref kernel arguments with other global address spaces.
+; GCN-LABEL: {{^}}byref_global_i32_arg:
+; GCN: kernarg_segment_byte_size = 12
+; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
+define amdgpu_kernel void @byref_global_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* byref(i32) %in.byref) {
+  %in = load i32, i32 addrspace(1)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_flat_i32_arg:
+; GCN: flat_load_dword [[IN:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}} offset:8{{$}}
+define amdgpu_kernel void @byref_flat_i32_arg(i32 addrspace(1)* nocapture %out, i32* byref(i32) %in.byref) {
+  %in = load i32, i32* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_constant_32bit_i32_arg:
+; GCN: s_add_i32 s[[PTR_LO:[0-9]+]], s4, 8
+; GCN: s_mov_b32 s[[PTR_HI:[0-9]+]], 0{{$}}
+; GCN: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}}
+define amdgpu_kernel void @byref_constant_32bit_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(6)* byref(i32) %in.byref) {
+  %in = load i32, i32 addrspace(6)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; define amdgpu_kernel void @byref_unknown_as_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(999)* byref %in.byref) {
+;   %in = load i32, i32 addrspace(999)* %in.byref
+;   store i32 %in, i32 addrspace(1)* %out, align 4
+;   ret void
+; }
+
+; GCN-LABEL: {{^}}multi_byref_constant_i32_arg:
+; GCN: kernarg_segment_byte_size = 20
+; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x8
+; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0xc
+; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x10
+define amdgpu_kernel void @multi_byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in0.byref, i32 addrspace(4)* byref(i32) %in1.byref, i32 %after.offset) {
+  %in0 = load i32, i32 addrspace(4)* %in0.byref
+  %in1 = load i32, i32 addrspace(4)* %in1.byref
+  store volatile i32 %in0, i32 addrspace(1)* %out, align 4
+  store volatile i32 %in1, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}byref_constant_i32_arg_offset0:
+; GCN: kernarg_segment_byte_size = 4
+; GCN-NOT: s4
+; GCN-NOT: s5
+; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x0{{$}}
+define amdgpu_kernel void @byref_constant_i32_arg_offset0(i32 addrspace(4)* byref(i32) %in.byref) {
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store i32 %in, i32 addrspace(1)* undef, align 4
+  ret void
+}

llvm/test/CodeGen/AMDGPU/lower-kernargs.ll

 ;
   %alloca0 = alloca i32, addrspace(5)
   %alloca1 = alloca i32, i32 %n, addrspace(5)
   store volatile i32 0, i32 addrspace(5)* %alloca0
   store volatile i32 1, i32 addrspace(5)* %alloca1
   ret void
 }
 
+; Byref pointers should only be treated as offsets from kernarg
+define amdgpu_kernel void @byref_constant_i8_arg(i32 addrspace(1)* nocapture %out, i8 addrspace(4)* byref(i8) %in.byref) {
+; HSA-LABEL: @byref_constant_i8_arg(
+; HSA-NEXT:    [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[IN:%.*]] = load i8, i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]], align 1
+; HSA-NEXT:    [[EXT:%.*]] = zext i8 [[IN]] to i32
+; HSA-NEXT:    store i32 [[EXT]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_i8_arg(
+; MESA-NEXT:    [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I8_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[IN:%.*]] = load i8, i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]], align 1
+; MESA-NEXT:    [[EXT:%.*]] = zext i8 [[IN]] to i32
+; MESA-NEXT:    store i32 [[EXT]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i8, i8 addrspace(4)* %in.byref
+  %ext = zext i8 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i16_arg(i32 addrspace(1)* nocapture %out, i16 addrspace(4)* byref(i16) %in.byref) {
+; HSA-LABEL: @byref_constant_i16_arg(
+; HSA-NEXT:    [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i16 addrspace(4)*
+; HSA-NEXT:    [[IN:%.*]] = load i16, i16 addrspace(4)* [[TMP1]], align 2
+; HSA-NEXT:    [[EXT:%.*]] = zext i16 [[IN]] to i32
+; HSA-NEXT:    store i32 [[EXT]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_i16_arg(
+; MESA-NEXT:    [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I16_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i16 addrspace(4)*
+; MESA-NEXT:    [[IN:%.*]] = load i16, i16 addrspace(4)* [[TMP1]], align 2
+; MESA-NEXT:    [[EXT:%.*]] = zext i16 [[IN]] to i32
+; MESA-NEXT:    store i32 [[EXT]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i16, i16 addrspace(4)* %in.byref
+  %ext = zext i16 %in to i32
+  store i32 %ext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in.byref, i32 %after.offset) {
+; HSA-LABEL: @byref_constant_i32_arg(
+; HSA-NEXT:    [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(16) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 12
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; HSA-NEXT:    store volatile i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_i32_arg(
+; MESA-NEXT:    [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(52) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 48
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; MESA-NEXT:    store volatile i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_v4i32_arg(<4 x i32> addrspace(1)* nocapture %out, <4 x i32> addrspace(4)* byref(<4 x i32>) %in.byref, i32 %after.offset) {
+; HSA-LABEL: @byref_constant_v4i32_arg(
+; HSA-NEXT:    [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(36) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to <4 x i32> addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load <4 x i32> addrspace(1)*, <4 x i32> addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 16
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to <4 x i32> addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 32
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN:%.*]] = load <4 x i32>, <4 x i32> addrspace(4)* [[TMP1]], align 16
+; HSA-NEXT:    store volatile <4 x i32> [[IN]], <4 x i32> addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    [[OUT_CAST:%.*]] = bitcast <4 x i32> addrspace(1)* [[OUT_LOAD]] to i32 addrspace(1)*
+; HSA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_CAST]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_v4i32_arg(
+; MESA-NEXT:    [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(72) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to <4 x i32> addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load <4 x i32> addrspace(1)*, <4 x i32> addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 52
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to <4 x i32> addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_V4I32_ARG_KERNARG_SEGMENT]], i64 68
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN:%.*]] = load <4 x i32>, <4 x i32> addrspace(4)* [[TMP1]], align 16
+; MESA-NEXT:    store volatile <4 x i32> [[IN]], <4 x i32> addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    [[OUT_CAST:%.*]] = bitcast <4 x i32> addrspace(1)* [[OUT_LOAD]] to i32 addrspace(1)*
+; MESA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_CAST]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load <4 x i32>, <4 x i32> addrspace(4)* %in.byref
+  store volatile <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
+  %out.cast = bitcast <4 x i32> addrspace(1)* %out to i32 addrspace(1)*
+  store volatile i32 %after.offset, i32 addrspace(1)* %out.cast, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_align_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(256) %in.byref, i32 %after.offset) {
+; HSA-LABEL: @byref_align_constant_i32_arg(
+; HSA-NEXT:    [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 256 dereferenceable(264) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 256
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 260
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; HSA-NEXT:    store volatile i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_align_constant_i32_arg(
+; MESA-NEXT:    [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 256 dereferenceable(300) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 292
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_ALIGN_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 296
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 8, !invariant.load !0
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; MESA-NEXT:    store volatile i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store volatile i32 %in, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) %in.byref, i32 %after.offset) {
+; HSA-LABEL: @byref_natural_align_constant_v16i32_arg(
+; HSA-NEXT:    [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 64 dereferenceable(132) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 64
+; HSA-NEXT:    [[TMP2:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to <16 x i32> addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 128
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN:%.*]] = load <16 x i32>, <16 x i32> addrspace(4)* [[TMP2]], align 64
+; HSA-NEXT:    [[CAST_OUT:%.*]] = bitcast i32 addrspace(1)* [[OUT_LOAD]] to <16 x i32> addrspace(1)*
+; HSA-NEXT:    store volatile <16 x i32> [[IN]], <16 x i32> addrspace(1)* [[CAST_OUT]], align 4
+; HSA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_natural_align_constant_v16i32_arg(
+; MESA-NEXT:    [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 64 dereferenceable(168) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 100
+; MESA-NEXT:    [[TMP2:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to <16 x i32> addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_NATURAL_ALIGN_CONSTANT_V16I32_ARG_KERNARG_SEGMENT]], i64 164
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN:%.*]] = load <16 x i32>, <16 x i32> addrspace(4)* [[TMP2]], align 64
+; MESA-NEXT:    [[CAST_OUT:%.*]] = bitcast i32 addrspace(1)* [[OUT_LOAD]] to <16 x i32> addrspace(1)*
+; MESA-NEXT:    store volatile <16 x i32> [[IN]], <16 x i32> addrspace(1)* [[CAST_OUT]], align 4
+; MESA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
+  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
+  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Also accept byref kernel arguments with other global address spaces.
+define amdgpu_kernel void @byref_global_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_global_i32_arg(
+; HSA-NEXT:    [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(1)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(1)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_global_i32_arg(
+; MESA-NEXT:    [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_GLOBAL_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(1)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(1)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(1)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_flat_i32_arg(i32 addrspace(1)* nocapture %out, i32* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_flat_i32_arg(
+; HSA-NEXT:    [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_flat_i32_arg(
+; MESA-NEXT:    [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_FLAT_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_32bit_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(6)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_constant_32bit_i32_arg(
+; HSA-NEXT:    [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(6)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(6)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_32bit_i32_arg(
+; MESA-NEXT:    [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_32BIT_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(6)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(6)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(6)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_unknown_as_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(999)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_unknown_as_i32_arg(
+; HSA-NEXT:    [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(999)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(999)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_unknown_as_i32_arg(
+; MESA-NEXT:    [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_UNKNOWN_AS_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(999)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(999)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(999)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Invalid, but should not crash.
+define amdgpu_kernel void @byref_private_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(5)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_private_i32_arg(
+; HSA-NEXT:    [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(5)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(5)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_private_i32_arg(
+; MESA-NEXT:    [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_PRIVATE_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(5)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(5)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(5)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Invalid, but should not crash.
+define amdgpu_kernel void @byref_local_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(3)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_local_i32_arg(
+; HSA-NEXT:    [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(12) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(3)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(3)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_local_i32_arg(
+; MESA-NEXT:    [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(48) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_LOCAL_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = addrspacecast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(3)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(3)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(3)* %in.byref
+  store i32 %in, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @multi_byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in0.byref, i32 addrspace(4)* byref(i32) %in1.byref, i32 %after.offset) {
+; HSA-LABEL: @multi_byref_constant_i32_arg(
+; HSA-NEXT:    [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(20) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; HSA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN0_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 8
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN0_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[IN1_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 12
+; HSA-NEXT:    [[TMP2:%.*]] = bitcast i8 addrspace(4)* [[IN1_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 16
+; HSA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 16, !invariant.load !0
+; HSA-NEXT:    [[IN0:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; HSA-NEXT:    [[IN1:%.*]] = load i32, i32 addrspace(4)* [[TMP2]], align 4
+; HSA-NEXT:    store volatile i32 [[IN0]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    store volatile i32 [[IN1]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @multi_byref_constant_i32_arg(
+; MESA-NEXT:    [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(56) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[OUT_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[OUT_KERNARG_OFFSET]] to i32 addrspace(1)* addrspace(4)*
+; MESA-NEXT:    [[OUT_LOAD:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)* addrspace(4)* [[OUT_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN0_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 44
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN0_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[IN1_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 48
+; MESA-NEXT:    [[TMP2:%.*]] = bitcast i8 addrspace(4)* [[IN1_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[MULTI_BYREF_CONSTANT_I32_ARG_KERNARG_SEGMENT]], i64 52
+; MESA-NEXT:    [[AFTER_OFFSET_KERNARG_OFFSET_CAST:%.*]] = bitcast i8 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[AFTER_OFFSET_LOAD:%.*]] = load i32, i32 addrspace(4)* [[AFTER_OFFSET_KERNARG_OFFSET_CAST]], align 4, !invariant.load !0
+; MESA-NEXT:    [[IN0:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; MESA-NEXT:    [[IN1:%.*]] = load i32, i32 addrspace(4)* [[TMP2]], align 4
+; MESA-NEXT:    store volatile i32 [[IN0]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    store volatile i32 [[IN1]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    store volatile i32 [[AFTER_OFFSET_LOAD]], i32 addrspace(1)* [[OUT_LOAD]], align 4
+; MESA-NEXT:    ret void
+;
+  %in0 = load i32, i32 addrspace(4)* %in0.byref
+  %in1 = load i32, i32 addrspace(4)* %in1.byref
+  store volatile i32 %in0, i32 addrspace(1)* %out, align 4
+  store volatile i32 %in1, i32 addrspace(1)* %out, align 4
+  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @byref_constant_i32_arg_offset0(i32 addrspace(4)* byref(i32) %in.byref) {
+; HSA-LABEL: @byref_constant_i32_arg_offset0(
+; HSA-NEXT:    [[BYREF_CONSTANT_I32_ARG_OFFSET0_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(4) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; HSA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_OFFSET0_KERNARG_SEGMENT]], i64 0
+; HSA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; HSA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; HSA-NEXT:    store i32 [[IN]], i32 addrspace(1)* undef, align 4
+; HSA-NEXT:    ret void
+;
+; MESA-LABEL: @byref_constant_i32_arg_offset0(
+; MESA-NEXT:    [[BYREF_CONSTANT_I32_ARG_OFFSET0_KERNARG_SEGMENT:%.*]] = call nonnull align 16 dereferenceable(40) i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr()
+; MESA-NEXT:    [[IN_BYREF_BYVAL_KERNARG_OFFSET:%.*]] = getelementptr inbounds i8, i8 addrspace(4)* [[BYREF_CONSTANT_I32_ARG_OFFSET0_KERNARG_SEGMENT]], i64 36
+; MESA-NEXT:    [[TMP1:%.*]] = bitcast i8 addrspace(4)* [[IN_BYREF_BYVAL_KERNARG_OFFSET]] to i32 addrspace(4)*
+; MESA-NEXT:    [[IN:%.*]] = load i32, i32 addrspace(4)* [[TMP1]], align 4
+; MESA-NEXT:    store i32 [[IN]], i32 addrspace(1)* undef, align 4
+; MESA-NEXT:    ret void
+;
+  %in = load i32, i32 addrspace(4)* %in.byref
+  store i32 %in, i32 addrspace(1)* undef, align 4
+  ret void
+}
+
 attributes #0 = { nounwind "target-cpu"="kaveri" }
 attributes #1 = { nounwind "target-cpu"="kaveri" "amdgpu-implicitarg-num-bytes"="40" }
 attributes #2 = { nounwind "target-cpu"="tahiti" }
 
 ; GCN: 0 = !{}
 ; GCN: !1 = !{i64 42}
 ; GCN: !2 = !{i64 128}
 ; GCN: !3 = !{i64 1024}