[clang][AMDGPU]: Don't use byval for struct arguments in function ABI

[cfang]

Summary

Byval requires allocating additional stack space, and always requires an implicit copy to be inserted in codegen,
where it can be difficult to optimize. In this work, we use byref/IndirectAliased promotion method instead of
byval with the implicit copy semantics.

[arsenm]

clang/lib/CodeGen/Targets/AMDGPU.cpp

253

Why does this need the type checks? Can this just go under the isIndirect handling?

clang/test/CodeGenOpenCL/addr-space-struct-arg.cl

124

These test checks are pretty thin, I'd like to see the memcpys in the IR. In a pre-commit, can you switch these tests to generated checks?

[arsenm]

Should also get a mention in the release notes (not sure how much ABI detail we have in AMDGPUUsage too)

[jdoerfert]

I would have assumed it makes more sense to have an IR pass that changes the ABI of functions with local linkage. The problem with this is that other languages might still emit byval, etc.

[arsenm]

In D155986#4523635, @jdoerfert wrote:

I would have assumed it makes more sense to have an IR pass that changes the ABI of functions with local linkage. The problem with this is that other languages might still emit byval, etc.

The backend still supports byval

[arsenm]

In D155986#4523635, @jdoerfert wrote:

I would have assumed it makes more sense to have an IR pass that changes the ABI of functions with local linkage. The problem with this is that other languages might still emit byval, etc.

I think this would be good as an alternative optimization, but this wouldn't change the C ABI to avoid byval. I think we don't want byval for whenever we have an object linkable ABI

[jdoerfert]

In D155986#4523674, @arsenm wrote:

In D155986#4523635, @jdoerfert wrote:

I would have assumed it makes more sense to have an IR pass that changes the ABI of functions with local linkage. The problem with this is that other languages might still emit byval, etc.

I think this would be good as an alternative optimization, but this wouldn't change the C ABI to avoid byval. I think we don't want byval for whenever we have an object linkable ABI

I have two points, I should have been more clear:

1. This is dangerous since we change something that is tested. Interoperability might not be the biggest thing yet, but special casing clangs behavior seems a slippery slope towards rust/fortran/Julia/... code not being able to talk to C/C++/HIP/... code when both are compiled for AMDGPUs.
2. We get more bang and less (potential) problems if we transform the IR whenever we can prove that we can, e.g., we see all the call sites. It should be overall better since we target a closed world, at least for the foreseeable future.

[arsenm]

In D155986#4524001, @jdoerfert wrote:

1. This is dangerous since we change something that is tested. Interoperability might not be the biggest thing yet, but special casing clangs behavior seems a slippery slope towards rust/fortran/Julia/... code not being able to talk to C/C++/HIP/... code when both are compiled for AMDGPUs.

It's not special casing clang's behavior. We are defining the C ABI, a thing which currently does not exist. Other languages that want to be compatible with the C ABI have to follow, but they don't have to. This isn't a unique property, and every frontend has to do this for every target. I don't see why we need to carry the albatross of byval simply because llvm didn't well abstract ABI details.

[JonChesterfield]

We do not have a defined ABI. We do not have machine code linking. Calling functions from assembly requires considerable care already. We change things like which registers are live per-function.

The time will come where we are sufficiently constrained to be stuck with inefficiency in calling convention but we are not there yet. Other languages that want to link against IR emitted by clang can still do so, though they'll also need to change to the more efficient convention to do so.

[jdoerfert]

You're saying this is not specializing it but it is. AMDGPU now emits different IR than NVPTX targets. Is that by itself a problem, no. Can I imagine this to be a problem down the line, yes.
That aside, I am not arguing this is by itself wrong. What I'm mostly trying to say is that there is a more generic alternative we should implement instead.
My goal is to get the same effect for all languages targeting AMDGPUs, not only the ones that go through Clang.

Anyway, I'm not blocking this. We'll can deal with the other languages and such later then.

[jdoerfert]

In D155986#4524191, @JonChesterfield wrote:

... though they'll also need to change to the more efficient convention to do so.

Exactly my argument.

[JonChesterfield]

In D155986#4524214, @jdoerfert wrote:

That aside, I am not arguing this is by itself wrong. What I'm mostly trying to say is that there is a more generic alternative we should implement instead.

We should optimise more aggressively within a module. There's an idea in mailing lists to deliberately split functions into local versions with a fast calling convention and externally visible shims used for address escapes. That's a good thing that we should do. It's not this thing.

This is a case where the module external functions can be quicker too. So we should do that. I hear that nvptx does the worse thing and you don't like divergence between the GPU targets. Ptx is the stable ABI there, perhaps we should fix nvptx to match in a later patch.

My goal is to get the same effect for all languages targeting AMDGPUs, not only the ones that go through Clang.

Those languages should also not use the convention in place before this patch. The existing behaviour is a design mistake. This frees them to do something better without generating shims to work around the C ABI.

[arsenm]

In D155986#4524214, @jdoerfert wrote:

You're saying this is not specializing it but it is. AMDGPU now emits different IR than NVPTX targets. Is that by itself a problem, no. Can I imagine this to be a problem down the line, yes.
That aside, I am not arguing this is by itself wrong. What I'm mostly trying to say is that there is a more generic alternative we should implement instead.

It's not an alternative, it's an entirely orthogonal thing. We should have both. If we commit to a real ABI, it is not going to be limited by trying to pretend that target specific IR is portable, when that's never been the case. NVPTX can switch to match this if it wants. If we want similar IR why consolidate in the worse direction

[cfang]

1. Move getIndirectAliased under isAggregateTypeForABI(Ty)
2. update LIT addr-space-struct-arg.cl to check corresponding alloca and memcpy to the struct.

[cfang]

clang/test/CodeGenOpenCL/addr-space-struct-arg.cl

124

Do you mean using "update_cc_test_checks.py" to generate the CHECKs? I am not sure why this does not work as expected. But I am including it as a separate file " addr-space-struct-arg-temp.cl" for reference.

[cfang]

In D155986#4523490, @arsenm wrote:

Should also get a mention in the release notes (not sure how much ABI detail we have in AMDGPUUsage too)

Still trying to figure out where to say what in the document.

[arsenm]

clang/test/CodeGenOpenCL/addr-space-struct-arg.cl

124

Yes. You might need to manually delete the checks that are already there

[yaxunl]

Does this patch cause the callee to skip making a local copy of the struct type argument? What if the callee makes changes to the argument? That is a common use case since users assume the function arguments in C/C++/HIP are passed by value.

[arsenm]

In D155986#4529782, @yaxunl wrote:

Does this patch cause the callee to skip making a local copy of the struct type argument? What if the callee makes changes to the argument? That is a common use case since users assume the function arguments in C/C++/HIP are passed by value.

The callee is supposed to use an explicit memcpy. The explicit memcpy can be optimized away most of the time. The current byval is an invisible copy on the caller side which is difficult to eliminate

[arsenm]

Don't add the temporary test, pre-commit a switch of the existing test so it's easy to see the diff in the review.

[cfang]

add a pre-commit test, CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl, for convenience of review. This test will be added for checking upon commit.

[arsenm]

Need to add to release notes and document in AMDGPUUsage

clang/test/CodeGenOpenCL/byval.cl

0–1

The test name suggests we should test with a different target that does use byval here

[cfang]

add x86 target to check pass by value in byval.cl
remove -enable-var-scope in the pre-commit test.

TODO: document and release note

[cfang]

Add comments in ReleaseNotes and AMDGPUUsage.

[arsenm]

llvm/docs/AMDGPUUsage.rst

13812–13813

s/to allocate memory/for allocating stack memory/

13813

Specify C ABI

13814

copying the value of the struct if modified

[cfang]

llvm/docs/AMDGPUUsage.rst

13813

Do not get what to do to "Specify C ABI"? Can you suggest explicitly? Thanks.

[arsenm]

llvm/docs/AMDGPUUsage.rst

13813

Add the letter C

[cfang]

llvm/docs/AMDGPUUsage.rst

13813

... in function C ABI? Or should we remove "function"?

[arsenm]

llvm/docs/AMDGPUUsage.rst

13813

Can just say C, doesn't really matter if you state function or not. function is implied

[cfang]

Update description in docs as suggested.

[arsenm]

Should look into why noundef was lost, but that can be in a follow up

clang/lib/CodeGen/Targets/AMDGPU.cpp

252

Typo "in stead"

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

257

This lost the noundef, shouldn't lose it

[cfang]

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

257

if (AI.getKind() == ABIArgInfo::Indirect)

return "noundef"

Should we add IndirectAlised check and include in the this same patch? Thanks

[arsenm]

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

257

Don't understand this snippet, the attribute emission presumably comes from somewhere else

[cfang]

Add "noundef" attribute for IndirectAlised.

[cfang]

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

257

In function DetermineNoUndef. "noundef was also missing for kernel byref argument

[arsenm]

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

257

Yes, that should probably also include indirect aliased. You should fix that in a second patch

[JonChesterfield]

Thanks! Happy to see function calls getting cheaper

clang/docs/ReleaseNotes.rst

 Miscellaneous Clang Crashes Fixed
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Target Specific Changes
 -----------------------
 
 AMDGPU Support
 ^^^^^^^^^^^^^^
+- Use pass-by-reference (byref) instead of pass-by-value (byval) for struct
+  arguments in function ABI. Callee is responsible to allocate memory and
+  make a copy of the struct. Note that AMDGPU backend still supports byval
+  for struct arguments.
 
 X86 Support
 ^^^^^^^^^^^
 
 Arm and AArch64 Support
 ^^^^^^^^^^^^^^^^^^^^^^^
 
 Windows Support

clang/lib/CodeGen/CGCall.cpp

         if (Addr.getAlignment() < Align &&
             llvm::getOrEnforceKnownAlignment(V, Align.getAsAlign(), *TD) <
                 Align.getAsAlign()) {
           NeedCopy = true;
         } else if (I->hasLValue()) {
           auto LV = I->getKnownLValue();
           auto AS = LV.getAddressSpace();
 
-          if (!ArgInfo.getIndirectByVal() ||
+          bool isByValOrRef =
+              ArgInfo.isIndirectAliased() || ArgInfo.getIndirectByVal();
+
+          if (!isByValOrRef ||
               (LV.getAlignment() < getContext().getTypeAlignInChars(I->Ty))) {
             NeedCopy = true;
           }
           if (!getLangOpts().OpenCL) {
-            if ((ArgInfo.getIndirectByVal() &&
+            if ((isByValOrRef &&
                 (AS != LangAS::Default &&
                  AS != CGM.getASTAllocaAddressSpace()))) {
               NeedCopy = true;
             }
           }
           // For OpenCL even if RV is located in default or alloca address space
           // we don't want to perform address space cast for it.
-          else if ((ArgInfo.getIndirectByVal() &&
+          else if ((isByValOrRef &&
                     Addr.getType()->getAddressSpace() != IRFuncTy->
                       getParamType(FirstIRArg)->getPointerAddressSpace())) {
             NeedCopy = true;
           }
         }
 
         if (NeedCopy) {
           // Create an aligned temporary, and copy to it.

clang/lib/CodeGen/Targets/AMDGPU.cpp

 
     if (NumRegsLeft > 0) {
       unsigned NumRegs = numRegsForType(Ty);
       if (NumRegsLeft >= NumRegs) {
         NumRegsLeft -= NumRegs;
         return ABIArgInfo::getDirect();
       }
     }
+
+    // Use pass-by-reference in stead of pass-by-value for struct arguments in
+    // function ABI.
+    return ABIArgInfo::getIndirectAliased(
+        getContext().getTypeAlignInChars(Ty),
+        getContext().getTargetAddressSpace(LangAS::opencl_private));
   }
 
   // Otherwise just do the default thing.
   ABIArgInfo ArgInfo = DefaultABIInfo::classifyArgumentType(Ty);
   if (!ArgInfo.isIndirect()) {
     unsigned NumRegs = numRegsForType(Ty);
     NumRegsLeft -= std::min(NumRegs, NumRegsLeft);
   }

clang/test/CodeGenCXX/amdgcn-func-arg.cpp

 
 A g_a;
 B g_b;
 
 void func_with_ref_arg(A &a);
 void func_with_ref_arg(B &b);
 
 // CHECK-LABEL: @_Z22func_with_indirect_arg1A(
-// CHECK-SAME:  ptr addrspace(5) noundef [[ARG:%.*]])
 // CHECK-NEXT:  entry:
-// CHECK-NEXT:    [[INDIRECT_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
+// CHECK-NEXT:    [[A_INDIRECT_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
 // CHECK-NEXT:    [[P:%.*]] = alloca ptr, align 8, addrspace(5)
-// CHECK-NEXT:    [[INDIRECT_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[INDIRECT_ADDR]] to ptr
+// CHECK-NEXT:    [[A_INDIRECT_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A_INDIRECT_ADDR]] to ptr
 // CHECK-NEXT:    [[P_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[P]] to ptr
-// CHECK-NEXT:    store ptr addrspace(5) [[ARG]], ptr [[INDIRECT_ADDR_ASCAST]]
-// CHECK-NEXT:    [[A_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A:%.*]] to ptr
+// CHECK-NEXT:    store ptr addrspace(5) [[A:%.*]], ptr [[A_INDIRECT_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[A_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A]] to ptr
 // CHECK-NEXT:    store ptr [[A_ASCAST]], ptr [[P_ASCAST]], align 8
 // CHECK-NEXT:    ret void
 //
 void func_with_indirect_arg(A a) {
   A *p = &a;
 }
 
 // CHECK-LABEL: @_Z22test_indirect_arg_autov(

 //
 void test_indirect_arg_global() {
   func_with_indirect_arg(g_a);
   func_with_ref_arg(g_a);
 }
 
 // CHECK-LABEL: @_Z19func_with_byval_arg1B(
 // CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[COERCE:%.*]] = alloca [[CLASS_B:%.*]], align 4, addrspace(5)
 // CHECK-NEXT:    [[P:%.*]] = alloca ptr, align 8, addrspace(5)
+// CHECK-NEXT:    [[B:%.*]] = addrspacecast ptr addrspace(5) [[COERCE]] to ptr
 // CHECK-NEXT:    [[P_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[P]] to ptr
-// CHECK-NEXT:    [[B_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[B:%.*]] to ptr
-// CHECK-NEXT:    store ptr [[B_ASCAST]], ptr [[P_ASCAST]], align 8
+// CHECK-NEXT:    call void @llvm.memcpy.p0.p5.i64(ptr align 4 [[B]], ptr addrspace(5) align 4 [[TMP0:%.*]], i64 400, i1 false)
+// CHECK-NEXT:    store ptr [[B]], ptr [[P_ASCAST]], align 8
 // CHECK-NEXT:    ret void
 //
 void func_with_byval_arg(B b) {
   B *p = &b;
 }
 
 // CHECK-LABEL: @_Z19test_byval_arg_autov(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[B:%.*]] = alloca [[CLASS_B:%.*]], align 4, addrspace(5)
 // CHECK-NEXT:    [[AGG_TMP:%.*]] = alloca [[CLASS_B]], align 4, addrspace(5)
 // CHECK-NEXT:    [[B_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[B]] to ptr
 // CHECK-NEXT:    [[AGG_TMP_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[AGG_TMP]] to ptr
 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[AGG_TMP_ASCAST]], ptr align 4 [[B_ASCAST]], i64 400, i1 false)
 // CHECK-NEXT:    [[AGG_TMP_ASCAST_ASCAST:%.*]] = addrspacecast ptr [[AGG_TMP_ASCAST]] to ptr addrspace(5)
-// CHECK-NEXT:    call void @_Z19func_with_byval_arg1B(ptr addrspace(5) noundef byval([[CLASS_B]]) align 4 [[AGG_TMP_ASCAST_ASCAST]])
+// CHECK-NEXT:    call void @_Z19func_with_byval_arg1B(ptr addrspace(5) byref([[CLASS_B]]) align 4 [[AGG_TMP_ASCAST_ASCAST]])
 // CHECK-NEXT:    call void @_Z17func_with_ref_argR1B(ptr noundef nonnull align 4 dereferenceable(400) [[B_ASCAST]])
 // CHECK-NEXT:    ret void
 //
 void test_byval_arg_auto() {
   B b;
   func_with_byval_arg(b);
   func_with_ref_arg(b);
 }
 
 // CHECK-LABEL: @_Z21test_byval_arg_globalv(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[AGG_TMP:%.*]] = alloca [[CLASS_B:%.*]], align 4, addrspace(5)
 // CHECK-NEXT:    [[AGG_TMP_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[AGG_TMP]] to ptr
 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[AGG_TMP_ASCAST]], ptr align 4 addrspacecast (ptr addrspace(1) @g_b to ptr), i64 400, i1 false)
 // CHECK-NEXT:    [[AGG_TMP_ASCAST_ASCAST:%.*]] = addrspacecast ptr [[AGG_TMP_ASCAST]] to ptr addrspace(5)
-// CHECK-NEXT:    call void @_Z19func_with_byval_arg1B(ptr addrspace(5) noundef byval([[CLASS_B]]) align 4 [[AGG_TMP_ASCAST_ASCAST]])
+// CHECK-NEXT:    call void @_Z19func_with_byval_arg1B(ptr addrspace(5) byref([[CLASS_B]]) align 4 [[AGG_TMP_ASCAST_ASCAST]])
 // CHECK-NEXT:    call void @_Z17func_with_ref_argR1B(ptr noundef nonnull align 4 dereferenceable(400) addrspacecast (ptr addrspace(1) @g_b to ptr))
 // CHECK-NEXT:    ret void
 //
 void test_byval_arg_global() {
   func_with_byval_arg(g_b);
   func_with_ref_arg(g_b);
 }

clang/test/CodeGenOpenCL/addr-space-struct-arg.cl

 // AMDGCN: load [9 x i32], ptr addrspace(1)
 // AMDGCN: call %struct.Mat4X4 @foo([9 x i32]
 // AMDGCN: call void @llvm.memcpy.p1.p5.i64(ptr addrspace(1)
 kernel void ker(global Mat3X3 *in, global Mat4X4 *out) {
   out[0] = foo(in[1]);
 }
 
 // X86-LABEL: define{{.*}} void @foo_large(ptr noalias sret(%struct.Mat64X64) align 4 %agg.result, ptr noundef byval(%struct.Mat32X32) align 4 %in)
-// AMDGCN-LABEL: define{{.*}} void @foo_large(ptr addrspace(5) noalias sret(%struct.Mat64X64) align 4 %agg.result, ptr addrspace(5) noundef byval(%struct.Mat32X32) align 4 %in)
+// AMDGCN-LABEL: define{{.*}} void @foo_large(ptr addrspace(5) noalias sret(%struct.Mat64X64) align 4 %agg.result, ptr addrspace(5) byref(%struct.Mat32X32) align 4 %{{.*}}
+// AMDGCN:       %in = alloca %struct.Mat32X32, align 4, addrspace(5)
+// AMDGCN-NEXT:  call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 4 %in, ptr addrspace(5) align 4 %{{.*}}, i64 4096, i1 false)
 Mat64X64 __attribute__((noinline)) foo_large(Mat32X32 in) {
   Mat64X64 out;
   return out;
 }
 
 // ALL-LABEL: define {{.*}} void @ker_large
 // Expect two mem copies: one for the argument "in", and one for
 // the return value.
 // X86: call void @llvm.memcpy.p0.p1.i32(ptr
 // X86: call void @llvm.memcpy.p1.p0.i32(ptr addrspace(1)
 // AMDGCN: call void @llvm.memcpy.p5.p1.i64(ptr addrspace(5)
 // AMDGCN: call void @llvm.memcpy.p1.p5.i64(ptr addrspace(1)
 kernel void ker_large(global Mat32X32 *in, global Mat64X64 *out) {
   out[0] = foo_large(in[1]);
 }
 
 // AMDGCN-LABEL: define{{.*}} void @FuncOneMember(<2 x i32> %u.coerce)
 void FuncOneMember(struct StructOneMember u) {
   u.x = (int2)(0, 0);
 }
 
-// AMDGCN-LABEL: define{{.*}} void @FuncOneLargeMember(ptr addrspace(5) noundef byval(%struct.LargeStructOneMember) align 8 %u)
+// AMDGCN-LABEL: define{{.*}} void @FuncOneLargeMember(ptr addrspace(5) byref(%struct.LargeStructOneMember) align 8 %{{.*}}
+// AMDGCN:  %u = alloca %struct.LargeStructOneMember, align 8, addrspace(5)
+// AMDGCN:  call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 8 %u, ptr addrspace(5) align 8 %{{.*}}, i64 800, i1 false)
 // AMDGCN-NOT: addrspacecast
 // AMDGCN:   store <2 x i32> %{{.*}}, ptr addrspace(5)
 void FuncOneLargeMember(struct LargeStructOneMember u) {
   u.x[0] = (int2)(0, 0);
 }
 
 // AMDGCN20-LABEL: define{{.*}} void @test_indirect_arg_globl()
 // AMDGCN20:  %[[byval_temp:.*]] = alloca %struct.LargeStructOneMember, align 8, addrspace(5)
 // AMDGCN20:  call void @llvm.memcpy.p5.p1.i64(ptr addrspace(5) align 8 %[[byval_temp]], ptr addrspace(1) align 8 @g_s, i64 800, i1 false)
-// AMDGCN20:  call void @FuncOneLargeMember(ptr addrspace(5) noundef byval(%struct.LargeStructOneMember) align 8 %[[byval_temp]])
+// AMDGCN20:  call void @FuncOneLargeMember(ptr addrspace(5) byref(%struct.LargeStructOneMember) align 8 %[[byval_temp]])
 #if (__OPENCL_C_VERSION__ == 200) || (__OPENCL_C_VERSION__ >= 300 && defined(__opencl_c_program_scope_global_variables))
 void test_indirect_arg_globl(void) {
   FuncOneLargeMember(g_s);
 }
 #endif
 
 // AMDGCN-LABEL: define{{.*}} amdgpu_kernel void @test_indirect_arg_local()
 // AMDGCN: %[[byval_temp:.*]] = alloca %struct.LargeStructOneMember, align 8, addrspace(5)
 // AMDGCN: call void @llvm.memcpy.p5.p3.i64(ptr addrspace(5) align 8 %[[byval_temp]], ptr addrspace(3) align 8 @test_indirect_arg_local.l_s, i64 800, i1 false)
-// AMDGCN: call void @FuncOneLargeMember(ptr addrspace(5) noundef byval(%struct.LargeStructOneMember) align 8 %[[byval_temp]])
+// AMDGCN: call void @FuncOneLargeMember(ptr addrspace(5) byref(%struct.LargeStructOneMember) align 8 %[[byval_temp]])
 kernel void test_indirect_arg_local(void) {
   local struct LargeStructOneMember l_s;
   FuncOneLargeMember(l_s);
 }
 
 // AMDGCN-LABEL: define{{.*}} void @test_indirect_arg_private()
 // AMDGCN: %[[p_s:.*]] = alloca %struct.LargeStructOneMember, align 8, addrspace(5)
 // AMDGCN-NOT: @llvm.memcpy
-// AMDGCN-NEXT: call void @FuncOneLargeMember(ptr addrspace(5) noundef byval(%struct.LargeStructOneMember) align 8 %[[p_s]])
+// AMDGCN-NEXT: call void @FuncOneLargeMember(ptr addrspace(5) byref(%struct.LargeStructOneMember) align 8 %[[p_s]])
 void test_indirect_arg_private(void) {
   struct LargeStructOneMember p_s;
   FuncOneLargeMember(p_s);
 }
 
 // AMDGCN-LABEL: define{{.*}} amdgpu_kernel void @KernelOneMember
 // AMDGCN-SAME:  (<2 x i32> %[[u_coerce:.*]])
 // AMDGCN:  %[[u:.*]] = alloca %struct.StructOneMember, align 8, addrspace(5)
 // AMDGCN:  %[[coerce_dive:.*]] = getelementptr inbounds %struct.StructOneMember, ptr addrspace(5) %[[u]], i32 0, i32 0
 // AMDGCN:  store <2 x i32> %[[u_coerce]], ptr addrspace(5) %[[coerce_dive]]
 // AMDGCN:  call void @FuncOneMember(<2 x i32>
 kernel void KernelOneMember(struct StructOneMember u) {
   FuncOneMember(u);
 }
 
 // SPIR: call void @llvm.memcpy.p0.p1.i32
 // SPIR-NOT: addrspacecast
 kernel void KernelOneMemberSpir(global struct StructOneMember* u) {
   FuncOneMember(*u);
 }
 
 // AMDGCN-LABEL: define{{.*}} amdgpu_kernel void @KernelLargeOneMember(
 // AMDGCN:  %[[U:.*]] = alloca %struct.LargeStructOneMember, align 8, addrspace(5)
 // AMDGCN:  store %struct.LargeStructOneMember %u.coerce, ptr addrspace(5) %[[U]], align 8
-// AMDGCN:  call void @FuncOneLargeMember(ptr addrspace(5) noundef byval(%struct.LargeStructOneMember) align 8 %[[U]])
+// AMDGCN:  call void @FuncOneLargeMember(ptr addrspace(5) byref(%struct.LargeStructOneMember) align 8 %[[U]])
 kernel void KernelLargeOneMember(struct LargeStructOneMember u) {
   FuncOneLargeMember(u);
 }
 
 // AMDGCN-LABEL: define{{.*}} void @FuncTwoMember(<2 x i32> %u.coerce0, <2 x i32> %u.coerce1)
 void FuncTwoMember(struct StructTwoMember u) {
   u.y = (int2)(0, 0);
 }
 
-// AMDGCN-LABEL: define{{.*}} void @FuncLargeTwoMember(ptr addrspace(5) noundef byval(%struct.LargeStructTwoMember) align 8 %u)
+// AMDGCN-LABEL: define dso_local void @FuncLargeTwoMember
+// AMDGCN-SAME: (ptr addrspace(5) byref([[STRUCT_LARGESTRUCTTWOMEMBER:%.*]]) align 8 [[TMP0:%.*]])
+// AMDGCN: %[[U:.*]] = alloca %struct.LargeStructTwoMember, align 8, addrspace(5)
+// AMDGCN: call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 8 %[[U]], ptr addrspace(5) align 8 [[TMP0]], i64 480, i1 false)
 void FuncLargeTwoMember(struct LargeStructTwoMember u) {
   u.y[0] = (int2)(0, 0);
 }
 
 // AMDGCN-LABEL: define{{.*}} amdgpu_kernel void @KernelTwoMember
 // AMDGCN-SAME:  (%struct.StructTwoMember %[[u_coerce:.*]])
 // AMDGCN:  %[[u:.*]] = alloca %struct.StructTwoMember, align 8, addrspace(5)
 // AMDGCN: %[[LD0:.*]] = load <2 x i32>, ptr addrspace(5)
 // AMDGCN: %[[LD1:.*]] = load <2 x i32>, ptr addrspace(5)
 // AMDGCN: call void @FuncTwoMember(<2 x i32> %[[LD0]], <2 x i32> %[[LD1]])
 kernel void KernelTwoMember(struct StructTwoMember u) {
   FuncTwoMember(u);
 }
 
 // AMDGCN-LABEL: define{{.*}} amdgpu_kernel void @KernelLargeTwoMember
 // AMDGCN-SAME:  (%struct.LargeStructTwoMember %[[u_coerce:.*]])
 // AMDGCN:  %[[u:.*]] = alloca %struct.LargeStructTwoMember, align 8, addrspace(5)
 // AMDGCN:  store %struct.LargeStructTwoMember %[[u_coerce]], ptr addrspace(5) %[[u]]
-// AMDGCN:  call void @FuncLargeTwoMember(ptr addrspace(5) noundef byval(%struct.LargeStructTwoMember) align 8 %[[u]])
+// AMDGCN:  call void @FuncLargeTwoMember(ptr addrspace(5) byref(%struct.LargeStructTwoMember) align 8 %[[u]])
 kernel void KernelLargeTwoMember(struct LargeStructTwoMember u) {
   FuncLargeTwoMember(u);
 }

clang/test/CodeGenOpenCL/amdgpu-abi-struct-arg-byref.cl

 // AMDGCN-NEXT:    call void @llvm.memcpy.p1.p5.i64(ptr addrspace(1) align 4 [[ARRAYIDX]], ptr addrspace(5) align 4 [[TMP]], i64 64, i1 false)
 // AMDGCN-NEXT:    ret void
 //
 kernel void ker(global Mat3X3 *in, global Mat4X4 *out) {
   out[0] = foo(in[1]);
 }
 
 // AMDGCN-LABEL: define dso_local void @foo_large
-// AMDGCN-SAME: (ptr addrspace(5) noalias sret([[STRUCT_MAT64X64:%.*]]) align 4 [[AGG_RESULT:%.*]], ptr addrspace(5) noundef byval([[STRUCT_MAT32X32:%.*]]) align 4 [[IN:%.*]]) #[[ATTR0]] {
+// AMDGCN-SAME: (ptr addrspace(5) noalias sret([[STRUCT_MAT64X64:%.*]]) align 4 [[AGG_RESULT:%.*]], ptr addrspace(5) byref([[STRUCT_MAT32X32:%.*]]) align 4 [[TMP0:%.*]]) #[[ATTR0]] {
 // AMDGCN-NEXT:  entry:
+// AMDGCN-NEXT:    [[IN:%.*]] = alloca [[STRUCT_MAT32X32]], align 4, addrspace(5)
+// AMDGCN-NEXT:    call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 4 [[IN]], ptr addrspace(5) align 4 [[TMP0]], i64 4096, i1 false)
 // AMDGCN-NEXT:    ret void
 //
 Mat64X64 __attribute__((noinline)) foo_large(Mat32X32 in) {
   Mat64X64 out;
   return out;
 }
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @ker_large
 // AMDGCN-SAME: (ptr addrspace(1) noundef align 4 [[IN:%.*]], ptr addrspace(1) noundef align 4 [[OUT:%.*]]) #[[ATTR1]] !kernel_arg_addr_space !4 !kernel_arg_access_qual !5 !kernel_arg_type !8 !kernel_arg_base_type !8 !kernel_arg_type_qual !7 {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[IN_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
 // AMDGCN-NEXT:    [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
 // AMDGCN-NEXT:    [[TMP:%.*]] = alloca [[STRUCT_MAT64X64:%.*]], align 4, addrspace(5)
 // AMDGCN-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_MAT32X32:%.*]], align 4, addrspace(5)
 // AMDGCN-NEXT:    store ptr addrspace(1) [[IN]], ptr addrspace(5) [[IN_ADDR]], align 8
 // AMDGCN-NEXT:    store ptr addrspace(1) [[OUT]], ptr addrspace(5) [[OUT_ADDR]], align 8
 // AMDGCN-NEXT:    [[TMP0:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
 // AMDGCN-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [[STRUCT_MAT64X64]], ptr addrspace(1) [[TMP0]], i64 0
 // AMDGCN-NEXT:    [[TMP1:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[IN_ADDR]], align 8
 // AMDGCN-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds [[STRUCT_MAT32X32]], ptr addrspace(1) [[TMP1]], i64 1
 // AMDGCN-NEXT:    call void @llvm.memcpy.p5.p1.i64(ptr addrspace(5) align 4 [[BYVAL_TEMP]], ptr addrspace(1) align 4 [[ARRAYIDX1]], i64 4096, i1 false)
-// AMDGCN-NEXT:    call void @foo_large(ptr addrspace(5) sret([[STRUCT_MAT64X64]]) align 4 [[TMP]], ptr addrspace(5) noundef byval([[STRUCT_MAT32X32]]) align 4 [[BYVAL_TEMP]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @foo_large(ptr addrspace(5) sret([[STRUCT_MAT64X64]]) align 4 [[TMP]], ptr addrspace(5) byref([[STRUCT_MAT32X32]]) align 4 [[BYVAL_TEMP]]) #[[ATTR3]]
 // AMDGCN-NEXT:    call void @llvm.memcpy.p1.p5.i64(ptr addrspace(1) align 4 [[ARRAYIDX]], ptr addrspace(5) align 4 [[TMP]], i64 16384, i1 false)
 // AMDGCN-NEXT:    ret void
 //
 kernel void ker_large(global Mat32X32 *in, global Mat64X64 *out) {
   out[0] = foo_large(in[1]);
 }
 
 // AMDGCN-LABEL: define dso_local void @FuncOneMember

 // AMDGCN-NEXT:    store <2 x i32> [[TMP0]], ptr addrspace(5) [[X]], align 8
 // AMDGCN-NEXT:    ret void
 //
 void FuncOneMember(struct StructOneMember u) {
   u.x = (int2)(0, 0);
 }
 
 // AMDGCN-LABEL: define dso_local void @FuncOneLargeMember
-// AMDGCN-SAME: (ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTONEMEMBER:%.*]]) align 8 [[U:%.*]]) #[[ATTR0]] {
+// AMDGCN-SAME: (ptr addrspace(5) byref([[STRUCT_LARGESTRUCTONEMEMBER:%.*]]) align 8 [[TMP0:%.*]]) #[[ATTR0]] {
 // AMDGCN-NEXT:  entry:
+// AMDGCN-NEXT:    [[U:%.*]] = alloca [[STRUCT_LARGESTRUCTONEMEMBER]], align 8, addrspace(5)
 // AMDGCN-NEXT:    [[DOTCOMPOUNDLITERAL:%.*]] = alloca <2 x i32>, align 8, addrspace(5)
+// AMDGCN-NEXT:    call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 8 [[U]], ptr addrspace(5) align 8 [[TMP0]], i64 800, i1 false)
 // AMDGCN-NEXT:    store <2 x i32> zeroinitializer, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
-// AMDGCN-NEXT:    [[TMP0:%.*]] = load <2 x i32>, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
+// AMDGCN-NEXT:    [[TMP1:%.*]] = load <2 x i32>, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
 // AMDGCN-NEXT:    [[X:%.*]] = getelementptr inbounds [[STRUCT_LARGESTRUCTONEMEMBER]], ptr addrspace(5) [[U]], i32 0, i32 0
 // AMDGCN-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [100 x <2 x i32>], ptr addrspace(5) [[X]], i64 0, i64 0
-// AMDGCN-NEXT:    store <2 x i32> [[TMP0]], ptr addrspace(5) [[ARRAYIDX]], align 8
+// AMDGCN-NEXT:    store <2 x i32> [[TMP1]], ptr addrspace(5) [[ARRAYIDX]], align 8
 // AMDGCN-NEXT:    ret void
 //
 void FuncOneLargeMember(struct LargeStructOneMember u) {
   u.x[0] = (int2)(0, 0);
 }
 
 #if (__OPENCL_C_VERSION__ == 200) || (__OPENCL_C_VERSION__ >= 300 && defined(__opencl_c_program_scope_global_variables))
 // AMDGCN-LABEL: define dso_local void @test_indirect_arg_globl
 // AMDGCN-SAME: () #[[ATTR0]] {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_LARGESTRUCTONEMEMBER:%.*]], align 8, addrspace(5)
 // AMDGCN-NEXT:    call void @llvm.memcpy.p5.p1.i64(ptr addrspace(5) align 8 [[BYVAL_TEMP]], ptr addrspace(1) align 8 @g_s, i64 800, i1 false)
-// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[BYVAL_TEMP]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) byref([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[BYVAL_TEMP]]) #[[ATTR3]]
 // AMDGCN-NEXT:    ret void
 //
 void test_indirect_arg_globl(void) {
   FuncOneLargeMember(g_s);
 }
 #endif
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @test_indirect_arg_local
 // AMDGCN-SAME: () #[[ATTR1]] !kernel_arg_addr_space !9 !kernel_arg_access_qual !9 !kernel_arg_type !9 !kernel_arg_base_type !9 !kernel_arg_type_qual !9 {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_LARGESTRUCTONEMEMBER:%.*]], align 8, addrspace(5)
 // AMDGCN-NEXT:    call void @llvm.memcpy.p5.p3.i64(ptr addrspace(5) align 8 [[BYVAL_TEMP]], ptr addrspace(3) align 8 @test_indirect_arg_local.l_s, i64 800, i1 false)
-// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[BYVAL_TEMP]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) byref([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[BYVAL_TEMP]]) #[[ATTR3]]
 // AMDGCN-NEXT:    ret void
 //
 kernel void test_indirect_arg_local(void) {
   local struct LargeStructOneMember l_s;
   FuncOneLargeMember(l_s);
 }
 
 // AMDGCN-LABEL: define dso_local void @test_indirect_arg_private
 // AMDGCN-SAME: () #[[ATTR0]] {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[P_S:%.*]] = alloca [[STRUCT_LARGESTRUCTONEMEMBER:%.*]], align 8, addrspace(5)
-// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[P_S]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) byref([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[P_S]]) #[[ATTR3]]
 // AMDGCN-NEXT:    ret void
 //
 void test_indirect_arg_private(void) {
   struct LargeStructOneMember p_s;
   FuncOneLargeMember(p_s);
 }
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @KernelOneMember

   FuncOneMember(*u);
 }
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @KernelLargeOneMember
 // AMDGCN-SAME: ([[STRUCT_LARGESTRUCTONEMEMBER:%.*]] [[U_COERCE:%.*]]) #[[ATTR1]] !kernel_arg_addr_space !10 !kernel_arg_access_qual !11 !kernel_arg_type !16 !kernel_arg_base_type !16 !kernel_arg_type_qual !13 {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[U:%.*]] = alloca [[STRUCT_LARGESTRUCTONEMEMBER]], align 8, addrspace(5)
 // AMDGCN-NEXT:    store [[STRUCT_LARGESTRUCTONEMEMBER]] [[U_COERCE]], ptr addrspace(5) [[U]], align 8
-// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[U]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @FuncOneLargeMember(ptr addrspace(5) byref([[STRUCT_LARGESTRUCTONEMEMBER]]) align 8 [[U]]) #[[ATTR3]]
 // AMDGCN-NEXT:    ret void
 //
 kernel void KernelLargeOneMember(struct LargeStructOneMember u) {
   FuncOneLargeMember(u);
 }
 
 // AMDGCN-LABEL: define dso_local void @FuncTwoMember
 // AMDGCN-SAME: (<2 x i32> [[U_COERCE0:%.*]], <2 x i32> [[U_COERCE1:%.*]]) #[[ATTR0]] {

 // AMDGCN-NEXT:    store <2 x i32> [[TMP2]], ptr addrspace(5) [[Y]], align 8
 // AMDGCN-NEXT:    ret void
 //
 void FuncTwoMember(struct StructTwoMember u) {
   u.y = (int2)(0, 0);
 }
 
 // AMDGCN-LABEL: define dso_local void @FuncLargeTwoMember
-// AMDGCN-SAME: (ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTTWOMEMBER:%.*]]) align 8 [[U:%.*]]) #[[ATTR0]] {
+// AMDGCN-SAME: (ptr addrspace(5) byref([[STRUCT_LARGESTRUCTTWOMEMBER:%.*]]) align 8 [[TMP0:%.*]]) #[[ATTR0]] {
 // AMDGCN-NEXT:  entry:
+// AMDGCN-NEXT:    [[U:%.*]] = alloca [[STRUCT_LARGESTRUCTTWOMEMBER]], align 8, addrspace(5)
 // AMDGCN-NEXT:    [[DOTCOMPOUNDLITERAL:%.*]] = alloca <2 x i32>, align 8, addrspace(5)
+// AMDGCN-NEXT:    call void @llvm.memcpy.p5.p5.i64(ptr addrspace(5) align 8 [[U]], ptr addrspace(5) align 8 [[TMP0]], i64 480, i1 false)
 // AMDGCN-NEXT:    store <2 x i32> zeroinitializer, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
-// AMDGCN-NEXT:    [[TMP0:%.*]] = load <2 x i32>, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
+// AMDGCN-NEXT:    [[TMP1:%.*]] = load <2 x i32>, ptr addrspace(5) [[DOTCOMPOUNDLITERAL]], align 8
 // AMDGCN-NEXT:    [[Y:%.*]] = getelementptr inbounds [[STRUCT_LARGESTRUCTTWOMEMBER]], ptr addrspace(5) [[U]], i32 0, i32 1
 // AMDGCN-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [20 x <2 x i32>], ptr addrspace(5) [[Y]], i64 0, i64 0
-// AMDGCN-NEXT:    store <2 x i32> [[TMP0]], ptr addrspace(5) [[ARRAYIDX]], align 8
+// AMDGCN-NEXT:    store <2 x i32> [[TMP1]], ptr addrspace(5) [[ARRAYIDX]], align 8
 // AMDGCN-NEXT:    ret void
 //
 void FuncLargeTwoMember(struct LargeStructTwoMember u) {
   u.y[0] = (int2)(0, 0);
 }
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @KernelTwoMember
 // AMDGCN-SAME: ([[STRUCT_STRUCTTWOMEMBER:%.*]] [[U_COERCE:%.*]]) #[[ATTR1]] !kernel_arg_addr_space !10 !kernel_arg_access_qual !11 !kernel_arg_type !17 !kernel_arg_base_type !17 !kernel_arg_type_qual !13 {

   FuncTwoMember(u);
 }
 
 // AMDGCN-LABEL: define dso_local amdgpu_kernel void @KernelLargeTwoMember
 // AMDGCN-SAME: ([[STRUCT_LARGESTRUCTTWOMEMBER:%.*]] [[U_COERCE:%.*]]) #[[ATTR1]] !kernel_arg_addr_space !10 !kernel_arg_access_qual !11 !kernel_arg_type !18 !kernel_arg_base_type !18 !kernel_arg_type_qual !13 {
 // AMDGCN-NEXT:  entry:
 // AMDGCN-NEXT:    [[U:%.*]] = alloca [[STRUCT_LARGESTRUCTTWOMEMBER]], align 8, addrspace(5)
 // AMDGCN-NEXT:    store [[STRUCT_LARGESTRUCTTWOMEMBER]] [[U_COERCE]], ptr addrspace(5) [[U]], align 8
-// AMDGCN-NEXT:    call void @FuncLargeTwoMember(ptr addrspace(5) noundef byval([[STRUCT_LARGESTRUCTTWOMEMBER]]) align 8 [[U]]) #[[ATTR3]]
+// AMDGCN-NEXT:    call void @FuncLargeTwoMember(ptr addrspace(5) byref([[STRUCT_LARGESTRUCTTWOMEMBER]]) align 8 [[U]]) #[[ATTR3]]
 // AMDGCN-NEXT:    ret void
 //
 kernel void KernelLargeTwoMember(struct LargeStructTwoMember u) {
   FuncLargeTwoMember(u);
 }

clang/test/CodeGenOpenCL/amdgpu-abi-struct-coerce.cl

 {
   flexible_array s = { 0 };
   return s;
 }
 
 // CHECK: define{{.*}} void @func_reg_state_lo(<4 x i32> noundef %arg0, <4 x i32> noundef %arg1, <4 x i32> noundef %arg2, i32 noundef %arg3, i32 %s.coerce0, float %s.coerce1, i32 %s.coerce2)
 void func_reg_state_lo(int4 arg0, int4 arg1, int4 arg2, int arg3, struct_arg_t s) { }
 
-// CHECK: define{{.*}} void @func_reg_state_hi(<4 x i32> noundef %arg0, <4 x i32> noundef %arg1, <4 x i32> noundef %arg2, i32 noundef %arg3, i32 noundef %arg4, ptr addrspace(5) nocapture noundef byval(%struct.struct_arg) align 4 %s)
+// CHECK: define{{.*}} void @func_reg_state_hi(<4 x i32> noundef %arg0, <4 x i32> noundef %arg1, <4 x i32> noundef %arg2, i32 noundef %arg3, i32 noundef %arg4, ptr addrspace(5) nocapture byref(%struct.struct_arg) align 4 %{{.*}})
 void func_reg_state_hi(int4 arg0, int4 arg1, int4 arg2, int arg3, int arg4, struct_arg_t s) { }
 
 // XXX - Why don't the inner structs flatten?
-// CHECK: define{{.*}} void @func_reg_state_num_regs_nested_struct(<4 x i32> noundef %arg0, i32 noundef %arg1, i32 %arg2.coerce0, %struct.nested %arg2.coerce1, i32 %arg3.coerce0, %struct.nested %arg3.coerce1, ptr addrspace(5) nocapture noundef byval(%struct.num_regs_nested_struct) align 8 %arg4)
+// CHECK: define{{.*}} void @func_reg_state_num_regs_nested_struct(<4 x i32> noundef %arg0, i32 noundef %arg1, i32 %arg2.coerce0, %struct.nested %arg2.coerce1, i32 %arg3.coerce0, %struct.nested %arg3.coerce1, ptr addrspace(5) nocapture byref(%struct.num_regs_nested_struct) align 8 %{{.*}})
 void func_reg_state_num_regs_nested_struct(int4 arg0, int arg1, num_regs_nested_struct arg2, num_regs_nested_struct arg3, num_regs_nested_struct arg4) { }
 
 // CHECK: define{{.*}} void @func_double_nested_struct_arg(<4 x i32> noundef %arg0, i32 noundef %arg1, i32 %arg2.coerce0, %struct.double_nested %arg2.coerce1, i16 %arg2.coerce2)
 void func_double_nested_struct_arg(int4 arg0, int arg1, double_nested_struct arg2) { }
 
 // CHECK: define{{.*}} %struct.double_nested_struct @func_double_nested_struct_ret(<4 x i32> noundef %arg0, i32 noundef %arg1)
 double_nested_struct func_double_nested_struct_ret(int4 arg0, int arg1) {
   double_nested_struct s = { 0 };
   return s;
 }
 
 // CHECK: define{{.*}} void @func_large_struct_padding_arg_direct(i8 %arg.coerce0, i32 %arg.coerce1, i8 %arg.coerce2, i32 %arg.coerce3, i8 %arg.coerce4, i8 %arg.coerce5, i16 %arg.coerce6, i16 %arg.coerce7, [3 x i8] %arg.coerce8, i64 %arg.coerce9, i32 %arg.coerce10, i8 %arg.coerce11, i32 %arg.coerce12, i16 %arg.coerce13, i8 %arg.coerce14)
 void func_large_struct_padding_arg_direct(large_struct_padding arg) { }
 
-// CHECK: define{{.*}} void @func_large_struct_padding_arg_store(ptr addrspace(1) nocapture noundef writeonly %out, ptr addrspace(5) nocapture noundef readonly byval(%struct.large_struct_padding) align 8 %arg)
+// CHECK: define{{.*}} void @func_large_struct_padding_arg_store(ptr addrspace(1) nocapture noundef writeonly %out, ptr addrspace(5) nocapture readonly byref(%struct.large_struct_padding) align 8 %{{.*}})
 void func_large_struct_padding_arg_store(global large_struct_padding* out, large_struct_padding arg) {
   *out = arg;
 }
 
 // CHECK: define{{.*}} void @v3i32_reg_count(<3 x i32> noundef %arg1, <3 x i32> noundef %arg2, <3 x i32> noundef %arg3, <3 x i32> noundef %arg4, i32 %arg5.coerce0, float %arg5.coerce1, i32 %arg5.coerce2)
 void v3i32_reg_count(int3 arg1, int3 arg2, int3 arg3, int3 arg4, struct_arg_t arg5) { }
 
 // Function signature from blender, nothing should be passed byval. The v3i32
 // should not count as 4 passed registers.
 // CHECK: define{{.*}} void @v3i32_pair_reg_count(ptr addrspace(5) nocapture noundef %arg0, <3 x i32> %arg1.coerce0, <3 x i32> %arg1.coerce1, <3 x i32> noundef %arg2, <3 x i32> %arg3.coerce0, <3 x i32> %arg3.coerce1, <3 x i32> noundef %arg4, float noundef %arg5)
 void v3i32_pair_reg_count(int3_pair *arg0, int3_pair arg1, int3 arg2, int3_pair arg3, int3 arg4, float arg5) { }
 
 // Each short4 should fit pack into 2 registers.
 // CHECK: define{{.*}} void @v4i16_reg_count(<4 x i16> noundef %arg0, <4 x i16> noundef %arg1, <4 x i16> noundef %arg2, <4 x i16> noundef %arg3, <4 x i16> noundef %arg4, <4 x i16> noundef %arg5, i32 %arg6.coerce0, i32 %arg6.coerce1, i32 %arg6.coerce2, i32 %arg6.coerce3)
 void v4i16_reg_count(short4 arg0, short4 arg1, short4 arg2, short4 arg3,
                      short4 arg4, short4 arg5, struct_4regs arg6) { }
 
-// CHECK: define{{.*}} void @v4i16_pair_reg_count_over(<4 x i16> noundef %arg0, <4 x i16> noundef %arg1, <4 x i16> noundef %arg2, <4 x i16> noundef %arg3, <4 x i16> noundef %arg4, <4 x i16> noundef %arg5, <4 x i16> noundef %arg6, ptr addrspace(5) nocapture noundef byval(%struct.struct_4regs) align 4 %arg7)
+// CHECK: define{{.*}} void @v4i16_pair_reg_count_over(<4 x i16> noundef %arg0, <4 x i16> noundef %arg1, <4 x i16> noundef %arg2, <4 x i16> noundef %arg3, <4 x i16> noundef %arg4, <4 x i16> noundef %arg5, <4 x i16> noundef %arg6, ptr addrspace(5) nocapture byref(%struct.struct_4regs) align 4 %{{.*}})
 void v4i16_pair_reg_count_over(short4 arg0, short4 arg1, short4 arg2, short4 arg3,
                                short4 arg4, short4 arg5, short4 arg6, struct_4regs arg7) { }
 
 // CHECK: define{{.*}} void @v3i16_reg_count(<3 x i16> noundef %arg0, <3 x i16> noundef %arg1, <3 x i16> noundef %arg2, <3 x i16> noundef %arg3, <3 x i16> noundef %arg4, <3 x i16> noundef %arg5, i32 %arg6.coerce0, i32 %arg6.coerce1, i32 %arg6.coerce2, i32 %arg6.coerce3)
 void v3i16_reg_count(short3 arg0, short3 arg1, short3 arg2, short3 arg3,
                      short3 arg4, short3 arg5, struct_4regs arg6) { }
 
-// CHECK: define{{.*}} void @v3i16_reg_count_over(<3 x i16> noundef %arg0, <3 x i16> noundef %arg1, <3 x i16> noundef %arg2, <3 x i16> noundef %arg3, <3 x i16> noundef %arg4, <3 x i16> noundef %arg5, <3 x i16> noundef %arg6, ptr addrspace(5) nocapture noundef byval(%struct.struct_4regs) align 4 %arg7)
+// CHECK: define{{.*}} void @v3i16_reg_count_over(<3 x i16> noundef %arg0, <3 x i16> noundef %arg1, <3 x i16> noundef %arg2, <3 x i16> noundef %arg3, <3 x i16> noundef %arg4, <3 x i16> noundef %arg5, <3 x i16> noundef %arg6, ptr addrspace(5) nocapture byref(%struct.struct_4regs) align 4 %{{.*}})
 void v3i16_reg_count_over(short3 arg0, short3 arg1, short3 arg2, short3 arg3,
                           short3 arg4, short3 arg5, short3 arg6, struct_4regs arg7) { }
 
 // CHECK: define{{.*}} void @v2i16_reg_count(<2 x i16> noundef %arg0, <2 x i16> noundef %arg1, <2 x i16> noundef %arg2, <2 x i16> noundef %arg3, <2 x i16> noundef %arg4, <2 x i16> noundef %arg5, <2 x i16> noundef %arg6, <2 x i16> noundef %arg7, <2 x i16> noundef %arg8, <2 x i16> noundef %arg9, <2 x i16> noundef %arg10, <2 x i16> noundef %arg11, i32 %arg13.coerce0, i32 %arg13.coerce1, i32 %arg13.coerce2, i32 %arg13.coerce3)
 void v2i16_reg_count(short2 arg0, short2 arg1, short2 arg2, short2 arg3,
                      short2 arg4, short2 arg5, short2 arg6, short2 arg7,
                      short2 arg8, short2 arg9, short2 arg10, short2 arg11,
                      struct_4regs arg13) { }
 
-// CHECK: define{{.*}} void @v2i16_reg_count_over(<2 x i16> noundef %arg0, <2 x i16> noundef %arg1, <2 x i16> noundef %arg2, <2 x i16> noundef %arg3, <2 x i16> noundef %arg4, <2 x i16> noundef %arg5, <2 x i16> noundef %arg6, <2 x i16> noundef %arg7, <2 x i16> noundef %arg8, <2 x i16> noundef %arg9, <2 x i16> noundef %arg10, <2 x i16> noundef %arg11, <2 x i16> noundef %arg12, ptr addrspace(5) nocapture noundef byval(%struct.struct_4regs) align 4 %arg13)
+// CHECK: define{{.*}} void @v2i16_reg_count_over(<2 x i16> noundef %arg0, <2 x i16> noundef %arg1, <2 x i16> noundef %arg2, <2 x i16> noundef %arg3, <2 x i16> noundef %arg4, <2 x i16> noundef %arg5, <2 x i16> noundef %arg6, <2 x i16> noundef %arg7, <2 x i16> noundef %arg8, <2 x i16> noundef %arg9, <2 x i16> noundef %arg10, <2 x i16> noundef %arg11, <2 x i16> noundef %arg12, ptr addrspace(5) nocapture byref(%struct.struct_4regs) align 4 %{{.*}})
 void v2i16_reg_count_over(short2 arg0, short2 arg1, short2 arg2, short2 arg3,
                           short2 arg4, short2 arg5, short2 arg6, short2 arg7,
                           short2 arg8, short2 arg9, short2 arg10, short2 arg11,
                           short2 arg12, struct_4regs arg13) { }
 
 // CHECK: define{{.*}} void @v2i8_reg_count(<2 x i8> noundef %arg0, <2 x i8> noundef %arg1, <2 x i8> noundef %arg2, <2 x i8> noundef %arg3, <2 x i8> noundef %arg4, <2 x i8> noundef %arg5, i32 %arg6.coerce0, i32 %arg6.coerce1, i32 %arg6.coerce2, i32 %arg6.coerce3)
 void v2i8_reg_count(char2 arg0, char2 arg1, char2 arg2, char2 arg3,
                     char2 arg4, char2 arg5, struct_4regs arg6) { }
 
-// CHECK: define{{.*}} void @v2i8_reg_count_over(<2 x i8> noundef %arg0, <2 x i8> noundef %arg1, <2 x i8> noundef %arg2, <2 x i8> noundef %arg3, <2 x i8> noundef %arg4, <2 x i8> noundef %arg5, i32 noundef %arg6, ptr addrspace(5) nocapture noundef byval(%struct.struct_4regs) align 4 %arg7)
+// CHECK: define{{.*}} void @v2i8_reg_count_over(<2 x i8> noundef %arg0, <2 x i8> noundef %arg1, <2 x i8> noundef %arg2, <2 x i8> noundef %arg3, <2 x i8> noundef %arg4, <2 x i8> noundef %arg5, i32 noundef %arg6, ptr addrspace(5) nocapture byref(%struct.struct_4regs) align 4 %{{.*}})
 void v2i8_reg_count_over(char2 arg0, char2 arg1, char2 arg2, char2 arg3,
                          char2 arg4, char2 arg5, int arg6, struct_4regs arg7) { }
 
 // CHECK: define{{.*}} void @num_regs_left_64bit_aggregate(<4 x i32> noundef %arg0, <4 x i32> noundef %arg1, <4 x i32> noundef %arg2, <3 x i32> noundef %arg3, [2 x i32] %arg4.coerce, i32 noundef %arg5)
 void num_regs_left_64bit_aggregate(int4 arg0, int4 arg1, int4 arg2, int3 arg3, struct_char_x8 arg4, int arg5) { }

clang/test/CodeGenOpenCL/byval.cl

-// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn %s | FileCheck %s
-
+// RUN: %clang_cc1 -emit-llvm -o - -triple i686-pc-darwin %s | FileCheck -check-prefix=X86 %s
+// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn %s | FileCheck -check-prefix=AMDGCN %s
 struct A {
   int x[100];
 };
 
 int f(struct A a);
 
 int g() {
   struct A a;
-  // CHECK: call i32 @f(ptr addrspace(5) noundef byval{{.*}}%a)
+  // X86:    call i32 @f(ptr noundef nonnull byval(%struct.A) align 4 %a)
+  // AMDGCN: call i32 @f(ptr addrspace(5) byref{{.*}}%a)
   return f(a);
 }
 
-// CHECK: declare i32 @f(ptr addrspace(5) noundef byval{{.*}})
+// X86:   declare i32 @f(ptr noundef byval(%struct.A) align 4)
+// AMDGCN: declare i32 @f(ptr addrspace(5) byref{{.*}})

llvm/docs/AMDGPUUsage.rst

 
     If the function calls another function, it will place any stack allocated
     arguments after the last local allocation and adjust SGPR32 to the address
     after the last local allocation.
 
 9.  All other registers are unspecified.
 10. Any necessary ``s_waitcnt`` has been performed to ensure memory is available
     to the function.
+11: Use pass-by-reference (byref) instead of pass-by-value (byval) for struct
+    arguments in function ABI. Callee is responsible to allocate memory and
+    make a copy of the struct. Note that the backend still supports byval for
+    struct arguments.
 
 On exit from a function:
 
 1.  VGPR0-31 and SGPR4-29 are used to pass function result arguments as
     described below. Any registers used are considered clobbered registers.
 2.  The following registers are preserved and have the same value as on entry:
 
     * FLAT_SCRATCH