Skip to content

[clang][CodeGen] Additional fixes for #114062 #128166

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Feb 27, 2025
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 16 additions & 3 deletions clang/lib/CodeGen/CGCall.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -5163,6 +5163,21 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
}
}
if (IRFunctionArgs.hasSRetArg()) {
// A mismatch between the allocated return value's AS and the target's
// chosen IndirectAS can happen e.g. when passing the this pointer through
// a chain involving stores to / loads from the DefaultAS; we address this
// here, symmetrically with the handling we have for normal pointer args.
if (SRetPtr.getAddressSpace() != RetAI.getIndirectAddrSpace()) {
llvm::Value *V = SRetPtr.getBasePointer();
LangAS SAS = getLangASFromTargetAS(SRetPtr.getAddressSpace());
LangAS DAS = getLangASFromTargetAS(RetAI.getIndirectAddrSpace());
llvm::Type *Ty = llvm::PointerType::get(getLLVMContext(),
RetAI.getIndirectAddrSpace());

SRetPtr = SRetPtr.withPointer(
getTargetHooks().performAddrSpaceCast(*this, V, SAS, DAS, Ty, true),
SRetPtr.isKnownNonNull());
}
IRCallArgs[IRFunctionArgs.getSRetArgNo()] =
getAsNaturalPointerTo(SRetPtr, RetTy);
} else if (RetAI.isInAlloca()) {
Expand Down Expand Up @@ -5394,9 +5409,7 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
V->getType()->isIntegerTy())
V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());

// The only plausible mismatch here would be for pointer address spaces,
// which can happen e.g. when passing a sret arg that is in the AllocaAS
// to a function that takes a pointer to and argument in the DefaultAS.
// The only plausible mismatch here would be for pointer address spaces.
// We assume that the target has a reasonable mapping for the DefaultAS
// (it can be casted to from incoming specific ASes), and insert an AS
// cast to address the mismatch.
Expand Down
11 changes: 2 additions & 9 deletions clang/lib/CodeGen/CGExprAgg.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -302,15 +302,8 @@ void AggExprEmitter::withReturnValueSlot(
llvm::Value *LifetimeSizePtr = nullptr;
llvm::IntrinsicInst *LifetimeStartInst = nullptr;
if (!UseTemp) {
// It is possible for the existing slot we are using directly to have been
// allocated in the correct AS for an indirect return, and then cast to
// the default AS (this is the behaviour of CreateMemTemp), however we know
// that the return address is expected to point to the uncasted AS, hence we
// strip possible pointer casts here.
if (Dest.getAddress().isValid())
RetAddr = Dest.getAddress().withPointer(
Dest.getAddress().getBasePointer()->stripPointerCasts(),
Dest.getAddress().isKnownNonNull());
RetAddr = Dest.getAddress();
RawAddress RetAllocaAddr = RawAddress::invalid();
} else {
RetAddr = CGF.CreateMemTempWithoutCast(RetTy, "tmp");
llvm::TypeSize Size =
Expand Down
16 changes: 16 additions & 0 deletions clang/lib/CodeGen/CGExprScalar.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
#include "clang/Basic/CodeGenOptions.h"
#include "clang/Basic/TargetInfo.h"
#include "llvm/ADT/APFixedPoint.h"
#include "llvm/IR/Argument.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
Expand Down Expand Up @@ -2352,6 +2353,21 @@ Value *ScalarExprEmitter::VisitCastExpr(CastExpr *CE) {
Value *Src = Visit(const_cast<Expr*>(E));
llvm::Type *SrcTy = Src->getType();
llvm::Type *DstTy = ConvertType(DestTy);

// FIXME: this is a gross but seemingly necessary workaround for an issue
// manifesting when a target uses a non-default AS for indirect sret args,
// but the source HLL is generic, wherein a valid C-cast or reinterpret_cast
// on the address of a local struct that gets returned by value yields an
// invalid bitcast from the a pointer to the IndirectAS to a pointer to the
// DefaultAS. We can only do this subversive thing because sret args are
// manufactured and them residing in the IndirectAS is a target specific
// detail, and doing an AS cast here still retains the semantics the user
// expects. It is desirable to remove this iff a better solution is found.
if (auto A = dyn_cast<llvm::Argument>(Src); A && A->hasStructRetAttr())
return CGF.CGM.getTargetCodeGenInfo().performAddrSpaceCast(
CGF, Src, E->getType().getAddressSpace(), DestTy.getAddressSpace(),
DstTy);
Comment on lines +2367 to +2369
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does performAddrSpaceCast take a source value, and the source address space? The source value's type already carries the address space?

Copy link
Contributor Author

@AlexVlx AlexVlx Feb 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It carries the LLVM AS, not the AST AS. Since this is a target hook, I assume the intention is to allow a target to override this interface and take into account language intended AS mappings, rather than how the values being cast got codegen-ed? For example, it is plausible, if not currently done (?) that LangAS::Foo and LangAS::Bar are aliases / map to the same target AS etc., so it'd be possible to simply NOP this and return the source. AFAICS no target does this / overrides the interface though.


assert(
(!SrcTy->isPtrOrPtrVectorTy() || !DstTy->isPtrOrPtrVectorTy() ||
SrcTy->getPointerAddressSpace() == DstTy->getPointerAddressSpace()) &&
Expand Down
4 changes: 2 additions & 2 deletions clang/test/CodeGen/partial-reinitialization2.c
Original file line number Diff line number Diff line change
Expand Up @@ -91,8 +91,8 @@ void test5(void)
// CHECK-LABEL: test6
void test6(void)
{
// CHECK: [[VAR:%[a-z0-9]+]] = alloca
// CHECK: call {{.*}}get456789(ptr {{.*}}sret{{.*}} [[VAR]])
// CHECK: [[LP:%[a-z0-9]+]] = getelementptr{{.*}}%struct.LLP2P2, ptr{{.*}}, i32 0, i32 0
// CHECK: call {{.*}}get456789(ptr {{.*}}sret{{.*}} [[LP]])

// CHECK: [[CALL:%[a-z0-9]+]] = call {{.*}}@get235()
// CHECK: store{{.*}}[[CALL]], {{.*}}[[TMP0:%[a-z0-9.]+]]
Expand Down
32 changes: 32 additions & 0 deletions clang/test/CodeGenCXX/sret_cast_with_nonzero_alloca_as.cpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,32 @@
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa-gnu -target-cpu gfx900 -emit-llvm -o - %s | FileCheck %s

struct X { int z[17]; };

// CHECK-LABEL: define dso_local void @_Z3foocc(
// CHECK-SAME: ptr addrspace(5) dead_on_unwind noalias writable sret([[STRUCT_X:%.*]]) align 4 [[AGG_RESULT:%.*]], i8 noundef signext [[X:%.*]], i8 noundef signext [[Y:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[X_ADDR:%.*]] = alloca i8, align 1, addrspace(5)
// CHECK-NEXT: [[Y_ADDR:%.*]] = alloca i8, align 1, addrspace(5)
// CHECK-NEXT: [[X_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[X_ADDR]] to ptr
// CHECK-NEXT: [[Y_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[Y_ADDR]] to ptr
// CHECK-NEXT: store i8 [[X]], ptr [[X_ADDR_ASCAST]], align 1
// CHECK-NEXT: store i8 [[Y]], ptr [[Y_ADDR_ASCAST]], align 1
// CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[X_ADDR_ASCAST]], align 1
// CHECK-NEXT: [[AGG_RESULT_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[AGG_RESULT]] to ptr
// CHECK-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i8, ptr [[AGG_RESULT_ASCAST]], i64 1
// CHECK-NEXT: store i8 [[TMP0]], ptr [[ADD_PTR]], align 1
// CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[Y_ADDR_ASCAST]], align 1
// CHECK-NEXT: [[AGG_RESULT_ASCAST1:%.*]] = addrspacecast ptr addrspace(5) [[AGG_RESULT]] to ptr
// CHECK-NEXT: [[ADD_PTR2:%.*]] = getelementptr inbounds i8, ptr [[AGG_RESULT_ASCAST1]], i64 2
// CHECK-NEXT: store i8 [[TMP1]], ptr [[ADD_PTR2]], align 1
// CHECK-NEXT: ret void
//
X foo(char x, char y) {
X r;

*(((char*)&r) + 1) = x;
*(reinterpret_cast<char*>(&r) + 2) = y;

return r;
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does the analogous byval / byref case need to be tested?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe those are fine since they're not implicit / manufactured and they already get AS-casted to generic as part of normal handling: https://gcc.godbolt.org/z/df3TE8d19.

38 changes: 38 additions & 0 deletions clang/test/OpenMP/amdgcn_sret_ctor.cpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --include-generated-funcs --version 5
// RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple amdgcn-amd-amdhsa -emit-llvm %s -fopenmp-is-target-device -o - | FileCheck %s

#pragma omp begin declare target
struct S {
~S();
};
S s();
struct E {
S foo;
E() noexcept;
};
E::E() noexcept : foo(s()) {}
#pragma omp end declare target
// CHECK-LABEL: define hidden void @_ZN1EC2Ev(
// CHECK-SAME: ptr noundef nonnull align 1 dereferenceable(1) [[THIS:%.*]]) unnamed_addr #[[ATTR0:[0-9]+]] align 2 {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
// CHECK-NEXT: [[THIS_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[THIS_ADDR]] to ptr
// CHECK-NEXT: store ptr [[THIS]], ptr [[THIS_ADDR_ASCAST]], align 8
// CHECK-NEXT: [[THIS1:%.*]] = load ptr, ptr [[THIS_ADDR_ASCAST]], align 8
// CHECK-NEXT: [[THIS1_ASCAST:%.*]] = addrspacecast ptr [[THIS1]] to ptr addrspace(5)
// CHECK-NEXT: call void @_Z1sv(ptr addrspace(5) dead_on_unwind writable sret([[STRUCT_S:%.*]]) align 1 [[THIS1_ASCAST]]) #[[ATTR2:[0-9]+]]
// CHECK-NEXT: ret void
//
// CHECK-LABEL: declare void @_Z1sv(
// CHECK-SAME: ptr addrspace(5) dead_on_unwind writable sret([[STRUCT_S]]) align 1) #[[ATTR1:[0-9]+]]
//
// CHECK-LABEL: define hidden void @_ZN1EC1Ev(
// CHECK-SAME: ptr noundef nonnull align 1 dereferenceable(1) [[THIS:%.*]]) unnamed_addr #[[ATTR0]] align 2 {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[THIS_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
// CHECK-NEXT: [[THIS_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[THIS_ADDR]] to ptr
// CHECK-NEXT: store ptr [[THIS]], ptr [[THIS_ADDR_ASCAST]], align 8
// CHECK-NEXT: [[THIS1:%.*]] = load ptr, ptr [[THIS_ADDR_ASCAST]], align 8
// CHECK-NEXT: call void @_ZN1EC2Ev(ptr noundef nonnull align 1 dereferenceable(1) [[THIS1]]) #[[ATTR3:[0-9]+]]
// CHECK-NEXT: ret void
//