Skip to content

[AArch64][Clang] Refactor code to emit SVE & SME builtins #70662

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
kmclaughlin-arm merged 3 commits into from
Nov 1, 2023
Merged

[AArch64][Clang] Refactor code to emit SVE & SME builtins #70662

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f00e334
[AArch64][Clang] Refactor code to emit SVE & SME builtins
kmclaughlin-arm Oct 30, 2023
aab1c76
- Rewriting the loop in GetAArch64SVEProcessedOperands
kmclaughlin-arm Oct 31, 2023
bc0572f
- Combined IsTupleGetOrSet & !isa<ScalableVectorType> checks
kmclaughlin-arm Nov 1, 2023
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
157 changes: 85 additions & 72 deletions clang/lib/CodeGen/CGBuiltin.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9571,22 +9571,17 @@ Value *CodeGenFunction::EmitSVEStructStore(const SVETypeFlags &TypeFlags,
Value *BasePtr = Ops[1];

// Does the store have an offset?
if (Ops.size() > 3)
if (Ops.size() > (2 + N))
BasePtr = Builder.CreateGEP(VTy, BasePtr, Ops[2]);

Value *Val = Ops.back();

// The llvm.aarch64.sve.st2/3/4 intrinsics take legal part vectors, so we
// need to break up the tuple vector.
SmallVector<llvm::Value*, 5> Operands;
unsigned MinElts = VTy->getElementCount().getKnownMinValue();
for (unsigned I = 0; I < N; ++I) {
Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts);
Operands.push_back(Builder.CreateExtractVector(VTy, Val, Idx));
}
for (unsigned I = Ops.size() - N; I < Ops.size(); ++I)
Operands.push_back(Ops[I]);
Operands.append({Predicate, BasePtr});

Function *F = CGM.getIntrinsic(IntID, { VTy });

return Builder.CreateCall(F, Operands);
}

Expand Down Expand Up @@ -9893,26 +9888,24 @@ Value *CodeGenFunction::FormSVEBuiltinResult(Value *Call) {
return Call;
}

Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
const CallExpr *E) {
void CodeGenFunction::GetAArch64SVEProcessedOperands(
unsigned BuiltinID, const CallExpr *E, SmallVectorImpl<Value *> &Ops,
SVETypeFlags TypeFlags) {
// Find out if any arguments are required to be integer constant expressions.
unsigned ICEArguments = 0;
ASTContext::GetBuiltinTypeError Error;
getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
assert(Error == ASTContext::GE_None && "Should not codegen an error");

llvm::Type *Ty = ConvertType(E->getType());
if (BuiltinID >= SVE::BI__builtin_sve_reinterpret_s8_s8 &&
BuiltinID <= SVE::BI__builtin_sve_reinterpret_f64_f64) {
Value *Val = EmitScalarExpr(E->getArg(0));
return EmitSVEReinterpret(Val, Ty);
}
// Tuple set/get only requires one insert/extract vector, which is
// created by EmitSVETupleSetOrGet.
bool IsTupleGetOrSet = TypeFlags.isTupleSet() || TypeFlags.isTupleGet();

llvm::SmallVector<Value *, 4> Ops;
for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
if ((ICEArguments & (1 << i)) == 0)
Ops.push_back(EmitScalarExpr(E->getArg(i)));
else {
bool IsICE = ICEArguments & (1 << i);
Value *Arg = EmitScalarExpr(E->getArg(i));

if (IsICE) {
// If this is required to be a constant, constant fold it so that we know
// that the generated intrinsic gets a ConstantInt.
std::optional<llvm::APSInt> Result =
Expand All @@ -9924,12 +9917,56 @@ Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
// immediate requires more than a handful of bits.
*Result = Result->extOrTrunc(32);
Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), *Result));
continue;
}

if (IsTupleGetOrSet) {
Ops.push_back(EmitScalarExpr(E->getArg(i)));
Copy link
Collaborator

@sdesmalen-arm sdesmalen-arm Oct 31, 2023

Choose a reason for hiding this comment

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

This is the same as Ops.push_back(Arg); ? And if so, you can combine it with the case below.

continue;
}

if (!isa<ScalableVectorType>(Arg->getType())) {
Ops.push_back(Arg);
continue;
}

auto *VTy = cast<ScalableVectorType>(Arg->getType());
unsigned MinElts = VTy->getMinNumElements();
bool IsPred = VTy->getElementType()->isIntegerTy(1);
unsigned N = (MinElts * VTy->getScalarSizeInBits()) / (IsPred ? 16 : 128);

if (N == 1) {
Ops.push_back(Arg);
continue;
}

for (unsigned I = 0; I < N; ++I) {
Value *Idx = ConstantInt::get(CGM.Int64Ty, (I * MinElts) / N);
auto *NewVTy =
ScalableVectorType::get(VTy->getElementType(), MinElts / N);
Ops.push_back(Builder.CreateExtractVector(NewVTy, Arg, Idx));
}
}

return;
Copy link
Collaborator

@sdesmalen-arm sdesmalen-arm Oct 31, 2023

Choose a reason for hiding this comment

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

return is unnecessary for a function returning void.

}

Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
const CallExpr *E) {
llvm::Type *Ty = ConvertType(E->getType());
if (BuiltinID >= SVE::BI__builtin_sve_reinterpret_s8_s8 &&
BuiltinID <= SVE::BI__builtin_sve_reinterpret_f64_f64) {
Value *Val = EmitScalarExpr(E->getArg(0));
return EmitSVEReinterpret(Val, Ty);
}

auto *Builtin = findARMVectorIntrinsicInMap(AArch64SVEIntrinsicMap, BuiltinID,
AArch64SVEIntrinsicsProvenSorted);

llvm::SmallVector<Value *, 4> Ops;
SVETypeFlags TypeFlags(Builtin->TypeModifier);
GetAArch64SVEProcessedOperands(BuiltinID, E, Ops, TypeFlags);

if (TypeFlags.isLoad())
return EmitSVEMaskedLoad(E, Ty, Ops, Builtin->LLVMIntrinsic,
TypeFlags.isZExtReturn());
Expand All @@ -9943,14 +9980,14 @@ Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
return EmitSVEPrefetchLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isGatherPrefetch())
return EmitSVEGatherPrefetch(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isStructLoad())
return EmitSVEStructLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isStructStore())
return EmitSVEStructStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isStructLoad())
return EmitSVEStructLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isStructStore())
return EmitSVEStructStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isTupleSet() || TypeFlags.isTupleGet())
return EmitSVETupleSetOrGet(TypeFlags, Ty, Ops);
return EmitSVETupleSetOrGet(TypeFlags, Ty, Ops);
else if (TypeFlags.isTupleCreate())
return EmitSVETupleCreate(TypeFlags, Ty, Ops);
return EmitSVETupleCreate(TypeFlags, Ty, Ops);
else if (TypeFlags.isUndef())
return UndefValue::get(Ty);
else if (Builtin->LLVMIntrinsic != 0) {
Expand Down Expand Up @@ -10202,13 +10239,8 @@ Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
case SVE::BI__builtin_sve_svtbl2_f64: {
SVETypeFlags TF(Builtin->TypeModifier);
auto VTy = cast<llvm::ScalableVectorType>(getSVEType(TF));
Value *V0 = Builder.CreateExtractVector(VTy, Ops[0],
ConstantInt::get(CGM.Int64Ty, 0));
unsigned MinElts = VTy->getMinNumElements();
Value *V1 = Builder.CreateExtractVector(
VTy, Ops[0], ConstantInt::get(CGM.Int64Ty, MinElts));
Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_tbl2, VTy);
return Builder.CreateCall(F, {V0, V1, Ops[1]});
return Builder.CreateCall(F, Ops);
}

case SVE::BI__builtin_sve_svset_neonq_s8:
Expand Down Expand Up @@ -10266,35 +10298,13 @@ Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,

Value *CodeGenFunction::EmitAArch64SMEBuiltinExpr(unsigned BuiltinID,
const CallExpr *E) {
// Find out if any arguments are required to be integer constant expressions.
unsigned ICEArguments = 0;
ASTContext::GetBuiltinTypeError Error;
getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
assert(Error == ASTContext::GE_None && "Should not codegen an error");

llvm::Type *Ty = ConvertType(E->getType());
llvm::SmallVector<Value *, 4> Ops;
for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
if ((ICEArguments & (1 << i)) == 0)
Ops.push_back(EmitScalarExpr(E->getArg(i)));
else {
// If this is required to be a constant, constant fold it so that we know
// that the generated intrinsic gets a ConstantInt.
std::optional<llvm::APSInt> Result =
E->getArg(i)->getIntegerConstantExpr(getContext());
assert(Result && "Expected argument to be a constant");

// Immediates for SVE llvm intrinsics are always 32bit. We can safely
// truncate because the immediate has been range checked and no valid
// immediate requires more than a handful of bits.
*Result = Result->extOrTrunc(32);
Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), *Result));
}
}

auto *Builtin = findARMVectorIntrinsicInMap(AArch64SMEIntrinsicMap, BuiltinID,
AArch64SMEIntrinsicsProvenSorted);

llvm::SmallVector<Value *, 4> Ops;
SVETypeFlags TypeFlags(Builtin->TypeModifier);
GetAArch64SVEProcessedOperands(BuiltinID, E, Ops, TypeFlags);

if (TypeFlags.isLoad() || TypeFlags.isStore())
return EmitSMELd1St1(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (TypeFlags.isReadZA() || TypeFlags.isWriteZA())
Expand All @@ -10307,21 +10317,24 @@ Value *CodeGenFunction::EmitAArch64SMEBuiltinExpr(unsigned BuiltinID,
BuiltinID == SME::BI__builtin_sme_svldr_za ||
BuiltinID == SME::BI__builtin_sme_svstr_za)
return EmitSMELdrStr(TypeFlags, Ops, Builtin->LLVMIntrinsic);
else if (Builtin->LLVMIntrinsic != 0) {
// Predicates must match the main datatype.
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
if (auto PredTy = dyn_cast<llvm::VectorType>(Ops[i]->getType()))
if (PredTy->getElementType()->isIntegerTy(1))
Ops[i] = EmitSVEPredicateCast(Ops[i], getSVEType(TypeFlags));

Function *F = CGM.getIntrinsic(Builtin->LLVMIntrinsic,
getSVEOverloadTypes(TypeFlags, Ty, Ops));
Value *Call = Builder.CreateCall(F, Ops);
return Call;
}
// Should not happen!
if (Builtin->LLVMIntrinsic == 0)
return nullptr;

/// Should not happen
return nullptr;
// Predicates must match the main datatype.
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
if (auto PredTy = dyn_cast<llvm::VectorType>(Ops[i]->getType()))
if (PredTy->getElementType()->isIntegerTy(1))
Ops[i] = EmitSVEPredicateCast(Ops[i], getSVEType(TypeFlags));

Function *F =
TypeFlags.isOverloadNone()
? CGM.getIntrinsic(Builtin->LLVMIntrinsic)
: CGM.getIntrinsic(Builtin->LLVMIntrinsic, {getSVEType(TypeFlags)});
Value *Call = Builder.CreateCall(F, Ops);

return FormSVEBuiltinResult(Call);
}

Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
Expand Down
5 changes: 5 additions & 0 deletions clang/lib/CodeGen/CodeGenFunction.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4311,6 +4311,11 @@ class CodeGenFunction : public CodeGenTypeCache {
llvm::Value *EmitSMELdrStr(const SVETypeFlags &TypeFlags,
llvm::SmallVectorImpl<llvm::Value *> &Ops,
unsigned IntID);

void GetAArch64SVEProcessedOperands(unsigned BuiltinID, const CallExpr *E,
SmallVectorImpl<llvm::Value *> &Ops,
SVETypeFlags TypeFlags);

llvm::Value *EmitAArch64SMEBuiltinExpr(unsigned BuiltinID, const CallExpr *E);

llvm::Value *EmitAArch64BuiltinExpr(unsigned BuiltinID, const CallExpr *E,
Expand Down
36 changes: 18 additions & 18 deletions clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_st2-bfloat.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,18 +16,18 @@
#endif
// CHECK-LABEL: @test_svst2_bf16(
// CHECK-NEXT: entry:
// CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x bfloat> [[TMP2]], <vscale x 8 x i1> [[TMP0]], ptr [[BASE:%.*]])
// CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP0]], <vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x i1> [[TMP2]], ptr [[BASE:%.*]])
// CHECK-NEXT: ret void
//
// CPP-CHECK-LABEL: @_Z15test_svst2_bf16u10__SVBool_tPu6__bf1614svbfloat16x2_t(
// CPP-CHECK-NEXT: entry:
// CPP-CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CPP-CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CPP-CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CPP-CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x bfloat> [[TMP2]], <vscale x 8 x i1> [[TMP0]], ptr [[BASE:%.*]])
// CPP-CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CPP-CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CPP-CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CPP-CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP0]], <vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x i1> [[TMP2]], ptr [[BASE:%.*]])
// CPP-CHECK-NEXT: ret void
//
void test_svst2_bf16(svbool_t pg, bfloat16_t *base, svbfloat16x2_t data)
Expand All @@ -37,20 +37,20 @@ void test_svst2_bf16(svbool_t pg, bfloat16_t *base, svbfloat16x2_t data)

// CHECK-LABEL: @test_svst2_vnum_bf16(
// CHECK-NEXT: entry:
// CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CHECK-NEXT: [[TMP1:%.*]] = getelementptr <vscale x 8 x bfloat>, ptr [[BASE:%.*]], i64 [[VNUM:%.*]]
// CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CHECK-NEXT: [[TMP3:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP2]], <vscale x 8 x bfloat> [[TMP3]], <vscale x 8 x i1> [[TMP0]], ptr [[TMP1]])
// CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CHECK-NEXT: [[TMP3:%.*]] = getelementptr <vscale x 8 x bfloat>, ptr [[BASE:%.*]], i64 [[VNUM:%.*]]
// CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP0]], <vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x i1> [[TMP2]], ptr [[TMP3]])
// CHECK-NEXT: ret void
//
// CPP-CHECK-LABEL: @_Z20test_svst2_vnum_bf16u10__SVBool_tPu6__bf16l14svbfloat16x2_t(
// CPP-CHECK-NEXT: entry:
// CPP-CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CPP-CHECK-NEXT: [[TMP1:%.*]] = getelementptr <vscale x 8 x bfloat>, ptr [[BASE:%.*]], i64 [[VNUM:%.*]]
// CPP-CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CPP-CHECK-NEXT: [[TMP3:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CPP-CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP2]], <vscale x 8 x bfloat> [[TMP3]], <vscale x 8 x i1> [[TMP0]], ptr [[TMP1]])
// CPP-CHECK-NEXT: [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA:%.*]], i64 0)
// CPP-CHECK-NEXT: [[TMP1:%.*]] = tail call <vscale x 8 x bfloat> @llvm.vector.extract.nxv8bf16.nxv16bf16(<vscale x 16 x bfloat> [[DATA]], i64 8)
// CPP-CHECK-NEXT: [[TMP2:%.*]] = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> [[PG:%.*]])
// CPP-CHECK-NEXT: [[TMP3:%.*]] = getelementptr <vscale x 8 x bfloat>, ptr [[BASE:%.*]], i64 [[VNUM:%.*]]
// CPP-CHECK-NEXT: tail call void @llvm.aarch64.sve.st2.nxv8bf16(<vscale x 8 x bfloat> [[TMP0]], <vscale x 8 x bfloat> [[TMP1]], <vscale x 8 x i1> [[TMP2]], ptr [[TMP3]])
// CPP-CHECK-NEXT: ret void
//
void test_svst2_vnum_bf16(svbool_t pg, bfloat16_t *base, int64_t vnum, svbfloat16x2_t data)
Expand Down
Loading