Skip to content

[amdgpu] Pass variadic arguments without splitting #94083

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.

Sign up for GitHub

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 1 commit into from
Jun 4, 2024
Merged

[amdgpu] Pass variadic arguments without splitting #94083

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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/Targets/AMDGPU.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,8 @@ class AMDGPUABIInfo final : public DefaultABIInfo {

ABIArgInfo classifyReturnType(QualType RetTy) const;
ABIArgInfo classifyKernelArgumentType(QualType Ty) const;
ABIArgInfo classifyArgumentType(QualType Ty, unsigned &NumRegsLeft) const;
ABIArgInfo classifyArgumentType(QualType Ty, bool Variadic,
unsigned &NumRegsLeft) const;

void computeInfo(CGFunctionInfo &FI) const override;
Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
Expand Down Expand Up @@ -103,12 +104,16 @@ void AMDGPUABIInfo::computeInfo(CGFunctionInfo &FI) const {
if (!getCXXABI().classifyReturnType(FI))
FI.getReturnInfo() = classifyReturnType(FI.getReturnType());

unsigned ArgumentIndex = 0;
const unsigned numFixedArguments = FI.getNumRequiredArgs();

unsigned NumRegsLeft = MaxNumRegsForArgsRet;
for (auto &Arg : FI.arguments()) {
if (CC == llvm::CallingConv::AMDGPU_KERNEL) {
Arg.info = classifyKernelArgumentType(Arg.type);
} else {
Arg.info = classifyArgumentType(Arg.type, NumRegsLeft);
bool FixedArgument = ArgumentIndex++ < numFixedArguments;
Arg.info = classifyArgumentType(Arg.type, !FixedArgument, NumRegsLeft);
}
}
}
Expand Down Expand Up @@ -197,12 +202,20 @@ ABIArgInfo AMDGPUABIInfo::classifyKernelArgumentType(QualType Ty) const {
return ABIArgInfo::getDirect(LTy, 0, nullptr, false);
}

ABIArgInfo AMDGPUABIInfo::classifyArgumentType(QualType Ty,
ABIArgInfo AMDGPUABIInfo::classifyArgumentType(QualType Ty, bool Variadic,
unsigned &NumRegsLeft) const {
assert(NumRegsLeft <= MaxNumRegsForArgsRet && "register estimate underflow");

Ty = useFirstFieldIfTransparentUnion(Ty);

if (Variadic) {
Copy link
Contributor

Choose a reason for hiding this comment

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

Wouldn't we still want to follow the isAggregateTypeForABI rules? Large structs should still go through byref?

Copy link
Contributor

Choose a reason for hiding this comment

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

The bigger concern is using giant aggregate types is not great IR, and not all that well supported. a 65k element array will crash in SelectionDAG for example

return ABIArgInfo::getDirect(/*T=*/nullptr,
/*Offset=*/0,
/*Padding=*/nullptr,
/*CanBeFlattened=*/false,
/*Align=*/0);
}

if (isAggregateTypeForABI(Ty)) {
// Records with non-trivial destructors/copy-constructors should not be
// passed by value.
Expand Down
293 changes: 293 additions & 0 deletions clang/test/CodeGen/amdgpu-variadic-call.c
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,293 @@
// REQUIRES: amdgpu-registered-target
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature
// RUN: %clang_cc1 -cc1 -std=c23 -triple amdgcn-amd-amdhsa -emit-llvm -O1 %s -o - | FileCheck %s

void sink_0(...);
void sink_1(int, ...);
void sink_2(double, int, ...);

// Simple scalar values

// CHECK-LABEL: define {{[^@]+}}@zero_varargs
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0() #[[ATTR2:[0-9]+]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void zero_varargs(int f0, double f1)
{
sink_0();
sink_1(f0);
sink_2(f1, f0);
}

// CHECK-LABEL: define {{[^@]+}}@one_i32
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i32 noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(i32 noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], i32 noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], i32 noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_i32(int f0, double f1, int v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_ptr
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], ptr noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(ptr noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], ptr noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], ptr noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_ptr(int f0, double f1, void* v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_f64
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], double noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(double noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], double noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], double noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_f64(int f0, double f1, double v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}


// C has various type promotion rules for variadics

// CHECK-LABEL: define {{[^@]+}}@one_i8
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i8 noundef signext [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[CONV:%.*]] = sext i8 [[V0]] to i32
// CHECK-NEXT: tail call void (...) @sink_0(i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_i8(int f0, double f1, char v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_i16
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i16 noundef signext [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[CONV:%.*]] = sext i16 [[V0]] to i32
// CHECK-NEXT: tail call void (...) @sink_0(i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], i32 noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_i16(int f0, double f1, short v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_f32
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], float noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[CONV:%.*]] = fpext float [[V0]] to double
// CHECK-NEXT: tail call void (...) @sink_0(double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_f32(int f0, double f1, float v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}


// Various half types. _Float16 is passed as half and __fp16 as double

// CHECK-LABEL: define {{[^@]+}}@one_f16a
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], half noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(half noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], half noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], half noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_f16a(int f0, double f1, _Float16 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_f16b
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], half noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[CONV:%.*]] = fpext half [[V0]] to double
// CHECK-NEXT: tail call void (...) @sink_0(double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], double noundef [[CONV]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_f16b(int f0, double f1, __fp16 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// CHECK-LABEL: define {{[^@]+}}@one_f16c
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], bfloat noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(bfloat noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], bfloat noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], bfloat noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_f16c(int f0, double f1, __bf16 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// Simple composites

typedef struct
{
double x0;
double x1;
} pair_f64;

// CHECK-LABEL: define {{[^@]+}}@one_pair_f64
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], double [[V0_COERCE0:%.*]], double [[V0_COERCE1:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [[STRUCT_PAIR_F64:%.*]] poison, double [[V0_COERCE0]], 0
// CHECK-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [[STRUCT_PAIR_F64]] [[DOTFCA_0_INSERT]], double [[V0_COERCE1]], 1
// CHECK-NEXT: tail call void (...) @sink_0([[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], [[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], [[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_pair_f64(int f0, double f1, pair_f64 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

typedef double v2f64 __attribute__((ext_vector_type(2)));

// CHECK-LABEL: define {{[^@]+}}@one_pair_v2f64
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], <2 x double> noundef [[V0:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(<2 x double> noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], <2 x double> noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], <2 x double> noundef [[V0]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_pair_v2f64(int f0, double f1, v2f64 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

typedef union
{
float x0;
int x1;
} union_f32_i32;

// CHECK-LABEL: define {{[^@]+}}@one_pair_union_f32_i32
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i32 [[V0_COERCE:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[TMP0:%.*]] = bitcast i32 [[V0_COERCE]] to float
// CHECK-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [[UNION_UNION_F32_I32:%.*]] poison, float [[TMP0]], 0
// CHECK-NEXT: tail call void (...) @sink_0([[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], [[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], [[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_pair_union_f32_i32(int f0, double f1, union_f32_i32 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

typedef union
{
int x0;
float x1;
} transparent_union_f32_i32 __attribute__((transparent_union));

// CHECK-LABEL: define {{[^@]+}}@one_pair_transparent_union_f32_i32
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i32 [[V0_COERCE:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [[UNION_TRANSPARENT_UNION_F32_I32:%.*]] poison, i32 [[V0_COERCE]], 0
// CHECK-NEXT: tail call void (...) @sink_0([[UNION_TRANSPARENT_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], [[UNION_TRANSPARENT_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], [[UNION_TRANSPARENT_UNION_F32_I32]] [[DOTFCA_0_INSERT]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void one_pair_transparent_union_f32_i32(int f0, double f1, transparent_union_f32_i32 v0)
{
sink_0(v0);
sink_1(f0, v0);
sink_2(f1, f0, v0);
}

// Passing multiple values in the variadic pack

// CHECK-LABEL: define {{[^@]+}}@multiple_one
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], i32 noundef [[V0:%.*]], double noundef [[V1:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: tail call void (...) @sink_0(i32 noundef [[V0]], double noundef [[V1]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], i32 noundef [[V0]], double noundef [[V1]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], i32 noundef [[V0]], double noundef [[V1]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void multiple_one(int f0, double f1, int v0, double v1)
{
sink_0(v0, v1);
sink_1(f0, v0, v1);
sink_2(f1, f0, v0, v1);
}

// CHECK-LABEL: define {{[^@]+}}@multiple_two
// CHECK-SAME: (i32 noundef [[F0:%.*]], double noundef [[F1:%.*]], double [[V0_COERCE0:%.*]], double [[V0_COERCE1:%.*]], float noundef [[V1:%.*]], i32 [[V2_COERCE:%.*]], i32 noundef [[V3:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-NEXT: entry:
// CHECK-NEXT: [[TMP0:%.*]] = bitcast i32 [[V2_COERCE]] to float
// CHECK-NEXT: [[CONV:%.*]] = fpext float [[V1]] to double
// CHECK-NEXT: [[DOTFCA_0_INSERT16:%.*]] = insertvalue [[STRUCT_PAIR_F64:%.*]] poison, double [[V0_COERCE0]], 0
// CHECK-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue [[STRUCT_PAIR_F64]] [[DOTFCA_0_INSERT16]], double [[V0_COERCE1]], 1
// CHECK-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue [[UNION_UNION_F32_I32:%.*]] poison, float [[TMP0]], 0
// CHECK-NEXT: tail call void (...) @sink_0([[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]], double noundef [[CONV]], [[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]], i32 noundef [[V3]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (i32, ...) @sink_1(i32 noundef [[F0]], [[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]], double noundef [[CONV]], [[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]], i32 noundef [[V3]]) #[[ATTR2]]
// CHECK-NEXT: tail call void (double, i32, ...) @sink_2(double noundef [[F1]], i32 noundef [[F0]], [[STRUCT_PAIR_F64]] [[DOTFCA_1_INSERT]], double noundef [[CONV]], [[UNION_UNION_F32_I32]] [[DOTFCA_0_INSERT]], i32 noundef [[V3]]) #[[ATTR2]]
// CHECK-NEXT: ret void
//
void multiple_two(int f0, double f1, pair_f64 v0, float v1, union_f32_i32 v2, int v3)
{
sink_0(v0, v1, v2, v3);
sink_1(f0, v0, v1, v2, v3);
sink_2(f1, f0, v0, v1, v2, v3);
}
Copy link
Contributor

Choose a reason for hiding this comment

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

Also test some structs that are so big they wouldn't be passed directly?

Copy link
Contributor

Choose a reason for hiding this comment

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

Missing big tests?

Loading