[clang] [Clang] Remove 3-element vector load and store special handling (PR #104661)
Shilei Tian via cfe-commits
cfe-commits at lists.llvm.org
Tue Jan 21 06:17:57 PST 2025
https://github.com/shiltian updated https://github.com/llvm/llvm-project/pull/104661
>From 5729891997dacfac91ad807ddfde60aec44708fb Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Fri, 17 Jan 2025 11:01:49 -0500
Subject: [PATCH 1/2] [Clang] Remove 3-element vector load and store special
handling
Clang uses a long-time special handling of the case where 3 element vector loads
and stores are performed as 4 element, and then a shufflevector is used to extract
the used elements. Odd sized vector codegen should now work reasonably well.
This patch removes this special handling, as well as the compiler argument
`-fpreserve-vec3-type`.
---
clang/include/clang/Basic/CodeGenOptions.def | 3 --
clang/include/clang/Basic/LangOptions.def | 2 +
clang/include/clang/Driver/Options.td | 4 --
clang/lib/Basic/LangOptions.cpp | 2 +
clang/lib/CodeGen/ABIInfo.cpp | 8 +++
clang/lib/CodeGen/ABIInfo.h | 8 +++
clang/lib/CodeGen/CGExpr.cpp | 49 +++++++++---------
clang/lib/CodeGen/Targets/AMDGPU.cpp | 11 ++++
.../test/CodeGenCXX/matrix-vector-bit-int.cpp | 32 ++++++------
clang/test/CodeGenOpenCL/amdgpu-alignment.cl | 8 +--
clang/test/CodeGenOpenCL/preserve_vec3.cl | 51 ++++++++++---------
11 files changed, 101 insertions(+), 77 deletions(-)
diff --git a/clang/include/clang/Basic/CodeGenOptions.def b/clang/include/clang/Basic/CodeGenOptions.def
index 0f4ed13d5f3d8c..1ab8c7fb4d3c33 100644
--- a/clang/include/clang/Basic/CodeGenOptions.def
+++ b/clang/include/clang/Basic/CodeGenOptions.def
@@ -413,9 +413,6 @@ CODEGENOPT(StrictReturn, 1, 1)
/// Whether emit pseudo probes for sample pgo profile collection.
CODEGENOPT(PseudoProbeForProfiling, 1, 0)
-/// Whether 3-component vector type is preserved.
-CODEGENOPT(PreserveVec3Type, 1, 0)
-
CODEGENOPT(NoPLT, 1, 0)
/// Whether to emit all vtables
diff --git a/clang/include/clang/Basic/LangOptions.def b/clang/include/clang/Basic/LangOptions.def
index 3b833240e5b68c..a980be853d53e6 100644
--- a/clang/include/clang/Basic/LangOptions.def
+++ b/clang/include/clang/Basic/LangOptions.def
@@ -532,6 +532,8 @@ BENIGN_LANGOPT(CheckConstexprFunctionBodies, 1, 1,
LANGOPT(BoundsSafety, 1, 0, "Bounds safety extension for C")
+LANGOPT(PreserveVec3Type, 1, 0, "Preserve 3-component vector type")
+
#undef LANGOPT
#undef COMPATIBLE_LANGOPT
#undef BENIGN_LANGOPT
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index d38dd2b4e3cf09..3e752ae86f06d9 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -8240,10 +8240,6 @@ def fhlsl_strict_availability : Flag<["-"], "fhlsl-strict-availability">,
Group<hlsl_Group>,
MarshallingInfoFlag<LangOpts<"HLSLStrictAvailability">>;
-def fpreserve_vec3_type : Flag<["-"], "fpreserve-vec3-type">,
- HelpText<"Preserve 3-component vector type">,
- MarshallingInfoFlag<CodeGenOpts<"PreserveVec3Type">>,
- ImpliedByAnyOf<[hlsl.KeyPath]>;
def fwchar_type_EQ : Joined<["-"], "fwchar-type=">,
HelpText<"Select underlying type for wchar_t">,
Values<"char,short,int">,
diff --git a/clang/lib/Basic/LangOptions.cpp b/clang/lib/Basic/LangOptions.cpp
index 94caf6a3897bc1..e3037ec819add2 100644
--- a/clang/lib/Basic/LangOptions.cpp
+++ b/clang/lib/Basic/LangOptions.cpp
@@ -208,6 +208,8 @@ void LangOptions::setLangDefaults(LangOptions &Opts, Language Lang,
// OpenCL and HLSL have half keyword
Opts.Half = Opts.OpenCL || Opts.HLSL;
+
+ Opts.PreserveVec3Type = Opts.HLSL;
}
FPOptions FPOptions::defaultWithoutTrailingStorage(const LangOptions &LO) {
diff --git a/clang/lib/CodeGen/ABIInfo.cpp b/clang/lib/CodeGen/ABIInfo.cpp
index 642bca9e8b76da..cda8a494f6c27d 100644
--- a/clang/lib/CodeGen/ABIInfo.cpp
+++ b/clang/lib/CodeGen/ABIInfo.cpp
@@ -236,6 +236,14 @@ void ABIInfo::appendAttributeMangling(StringRef AttrStr,
}
}
+llvm::FixedVectorType *
+ABIInfo::getOptimalVectorMemoryType(llvm::FixedVectorType *T,
+ const LangOptions &Opt) const {
+ if (T->getNumElements() == 3 && !Opt.PreserveVec3Type)
+ return llvm::FixedVectorType::get(T->getElementType(), 4);
+ return T;
+}
+
// Pin the vtable to this file.
SwiftABIInfo::~SwiftABIInfo() = default;
diff --git a/clang/lib/CodeGen/ABIInfo.h b/clang/lib/CodeGen/ABIInfo.h
index b8a8de57e5b971..213e7879c3162b 100644
--- a/clang/lib/CodeGen/ABIInfo.h
+++ b/clang/lib/CodeGen/ABIInfo.h
@@ -20,6 +20,7 @@ class Value;
class LLVMContext;
class DataLayout;
class Type;
+class FixedVectorType;
} // namespace llvm
namespace clang {
@@ -123,6 +124,13 @@ class ABIInfo {
raw_ostream &Out) const;
virtual void appendAttributeMangling(StringRef AttrStr,
raw_ostream &Out) const;
+
+ /// Returns the optimal vector memory type based on the given vector type. For
+ /// example, on certain targets, a vector with 3 elements might be promoted to
+ /// one with 4 elements to improve performance.
+ virtual llvm::FixedVectorType *
+ getOptimalVectorMemoryType(llvm::FixedVectorType *T,
+ const LangOptions &Opt) const;
};
/// Target specific hooks for defining how a type should be passed or returned
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 9a9a8c7f6eae09..78bac4c688dd3c 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -2002,20 +2002,19 @@ llvm::Value *CodeGenFunction::EmitLoadOfScalar(Address Addr, bool Volatile,
return EmitFromMemory(V, Ty);
}
- // Handle vectors of size 3 like size 4 for better performance.
- const llvm::Type *EltTy = Addr.getElementType();
- const auto *VTy = cast<llvm::FixedVectorType>(EltTy);
-
- if (!CGM.getCodeGenOpts().PreserveVec3Type && VTy->getNumElements() == 3) {
-
- llvm::VectorType *vec4Ty =
- llvm::FixedVectorType::get(VTy->getElementType(), 4);
- Address Cast = Addr.withElementType(vec4Ty);
- // Now load value.
- llvm::Value *V = Builder.CreateLoad(Cast, Volatile, "loadVec4");
-
- // Shuffle vector to get vec3.
- V = Builder.CreateShuffleVector(V, ArrayRef<int>{0, 1, 2}, "extractVec");
+ // Handles vectors of sizes that are likely to be expanded to a larger size
+ // to optimize performance.
+ auto *VTy = cast<llvm::FixedVectorType>(Addr.getElementType());
+ auto *NewVecTy =
+ CGM.getABIInfo().getOptimalVectorMemoryType(VTy, getLangOpts());
+
+ if (VTy != NewVecTy) {
+ Address Cast = Addr.withElementType(NewVecTy);
+ llvm::Value *V = Builder.CreateLoad(Cast, Volatile, "loadVecN");
+ unsigned OldNumElements = VTy->getNumElements();
+ SmallVector<int, 4> Mask(OldNumElements);
+ std::iota(Mask.begin(), Mask.end(), 0);
+ V = Builder.CreateShuffleVector(V, Mask, "extractVec");
return EmitFromMemory(V, Ty);
}
}
@@ -2145,21 +2144,21 @@ void CodeGenFunction::EmitStoreOfScalar(llvm::Value *Value, Address Addr,
Addr = Addr.withPointer(Builder.CreateThreadLocalAddress(GV),
NotKnownNonNull);
+ // Handles vectors of sizes that are likely to be expanded to a larger size
+ // to optimize performance.
llvm::Type *SrcTy = Value->getType();
if (const auto *ClangVecTy = Ty->getAs<VectorType>()) {
- auto *VecTy = dyn_cast<llvm::FixedVectorType>(SrcTy);
- if (!CGM.getCodeGenOpts().PreserveVec3Type) {
- // Handle vec3 special.
- if (VecTy && !ClangVecTy->isExtVectorBoolType() &&
- cast<llvm::FixedVectorType>(VecTy)->getNumElements() == 3) {
- // Our source is a vec3, do a shuffle vector to make it a vec4.
- Value = Builder.CreateShuffleVector(Value, ArrayRef<int>{0, 1, 2, -1},
- "extractVec");
- SrcTy = llvm::FixedVectorType::get(VecTy->getElementType(), 4);
+ if (auto *VecTy = dyn_cast<llvm::FixedVectorType>(SrcTy)) {
+ auto *NewVecTy =
+ CGM.getABIInfo().getOptimalVectorMemoryType(VecTy, getLangOpts());
+ if (!ClangVecTy->isExtVectorBoolType() && VecTy != NewVecTy) {
+ SmallVector<int, 4> Mask(NewVecTy->getNumElements(), -1);
+ std::iota(Mask.begin(), Mask.begin() + VecTy->getNumElements(), 0);
+ Value = Builder.CreateShuffleVector(Value, Mask, "extractVec");
+ SrcTy = NewVecTy;
}
- if (Addr.getElementType() != SrcTy) {
+ if (Addr.getElementType() != SrcTy)
Addr = Addr.withElementType(SrcTy);
- }
}
}
diff --git a/clang/lib/CodeGen/Targets/AMDGPU.cpp b/clang/lib/CodeGen/Targets/AMDGPU.cpp
index fa07e68c558356..788eac5f28231e 100644
--- a/clang/lib/CodeGen/Targets/AMDGPU.cpp
+++ b/clang/lib/CodeGen/Targets/AMDGPU.cpp
@@ -52,6 +52,17 @@ class AMDGPUABIInfo final : public DefaultABIInfo {
void computeInfo(CGFunctionInfo &FI) const override;
RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
AggValueSlot Slot) const override;
+
+ llvm::FixedVectorType *
+ getOptimalVectorMemoryType(llvm::FixedVectorType *T,
+ const LangOptions &Opt) const override {
+ // We have legal instructions for 96-bit so 3x32 can be supported.
+ // FIXME: This check should be a subtarget feature as technically SI doesn't
+ // support it.
+ if (T->getNumElements() == 3 && getDataLayout().getTypeSizeInBits(T) == 96)
+ return T;
+ return DefaultABIInfo::getOptimalVectorMemoryType(T, Opt);
+ }
};
bool AMDGPUABIInfo::isHomogeneousAggregateBaseType(QualType Ty) const {
diff --git a/clang/test/CodeGenCXX/matrix-vector-bit-int.cpp b/clang/test/CodeGenCXX/matrix-vector-bit-int.cpp
index ffbce9ff8d6f4c..7dc3b6bd598221 100644
--- a/clang/test/CodeGenCXX/matrix-vector-bit-int.cpp
+++ b/clang/test/CodeGenCXX/matrix-vector-bit-int.cpp
@@ -15,14 +15,14 @@ using i512x3x3 = _BitInt(512) __attribute__((matrix_type(3, 3)));
// CHECK-NEXT: [[A:%.*]] = alloca <3 x i8>, align 4
// CHECK-NEXT: [[A_ADDR:%.*]] = alloca <3 x i8>, align 4
// CHECK-NEXT: store i32 [[A_COERCE]], ptr [[A]], align 4
-// CHECK-NEXT: [[LOADVEC4:%.*]] = load <4 x i8>, ptr [[A]], align 4
-// CHECK-NEXT: [[A1:%.*]] = shufflevector <4 x i8> [[LOADVEC4]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x i8>, ptr [[A]], align 4
+// CHECK-NEXT: [[A1:%.*]] = shufflevector <4 x i8> [[LOADVECN]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <3 x i8> [[A1]], <3 x i8> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
// CHECK-NEXT: store <4 x i8> [[EXTRACTVEC]], ptr [[A_ADDR]], align 4
-// CHECK-NEXT: [[LOADVEC42:%.*]] = load <4 x i8>, ptr [[A_ADDR]], align 4
-// CHECK-NEXT: [[EXTRACTVEC3:%.*]] = shufflevector <4 x i8> [[LOADVEC42]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: [[LOADVEC44:%.*]] = load <4 x i8>, ptr [[A_ADDR]], align 4
-// CHECK-NEXT: [[EXTRACTVEC5:%.*]] = shufflevector <4 x i8> [[LOADVEC44]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN2:%.*]] = load <4 x i8>, ptr [[A_ADDR]], align 4
+// CHECK-NEXT: [[EXTRACTVEC3:%.*]] = shufflevector <4 x i8> [[LOADVECN2]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN4:%.*]] = load <4 x i8>, ptr [[A_ADDR]], align 4
+// CHECK-NEXT: [[EXTRACTVEC5:%.*]] = shufflevector <4 x i8> [[LOADVECN4]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
// CHECK-NEXT: [[ADD:%.*]] = add <3 x i8> [[EXTRACTVEC3]], [[EXTRACTVEC5]]
// CHECK-NEXT: store <3 x i8> [[ADD]], ptr [[RETVAL]], align 4
// CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[RETVAL]], align 4
@@ -38,10 +38,10 @@ i8x3 v1(i8x3 a) {
// CHECK-NEXT: [[A_ADDR:%.*]] = alloca <3 x i32>, align 16
// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <3 x i32> [[A]], <3 x i32> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
// CHECK-NEXT: store <4 x i32> [[EXTRACTVEC]], ptr [[A_ADDR]], align 16
-// CHECK-NEXT: [[LOADVEC4:%.*]] = load <4 x i32>, ptr [[A_ADDR]], align 16
-// CHECK-NEXT: [[EXTRACTVEC1:%.*]] = shufflevector <4 x i32> [[LOADVEC4]], <4 x i32> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: [[LOADVEC42:%.*]] = load <4 x i32>, ptr [[A_ADDR]], align 16
-// CHECK-NEXT: [[EXTRACTVEC3:%.*]] = shufflevector <4 x i32> [[LOADVEC42]], <4 x i32> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x i32>, ptr [[A_ADDR]], align 16
+// CHECK-NEXT: [[EXTRACTVEC1:%.*]] = shufflevector <4 x i32> [[LOADVECN]], <4 x i32> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN2:%.*]] = load <4 x i32>, ptr [[A_ADDR]], align 16
+// CHECK-NEXT: [[EXTRACTVEC3:%.*]] = shufflevector <4 x i32> [[LOADVECN2]], <4 x i32> poison, <3 x i32> <i32 0, i32 1, i32 2>
// CHECK-NEXT: [[ADD:%.*]] = add <3 x i32> [[EXTRACTVEC1]], [[EXTRACTVEC3]]
// CHECK-NEXT: ret <3 x i32> [[ADD]]
//
@@ -53,14 +53,14 @@ i32x3 v2(i32x3 a) {
// CHECK-SAME: ptr noundef byval(<3 x i512>) align 256 [[TMP0:%.*]]) #[[ATTR2:[0-9]+]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[A_ADDR:%.*]] = alloca <3 x i512>, align 256
-// CHECK-NEXT: [[LOADVEC4:%.*]] = load <4 x i512>, ptr [[TMP0]], align 256
-// CHECK-NEXT: [[A:%.*]] = shufflevector <4 x i512> [[LOADVEC4]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x i512>, ptr [[TMP0]], align 256
+// CHECK-NEXT: [[A:%.*]] = shufflevector <4 x i512> [[LOADVECN]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <3 x i512> [[A]], <3 x i512> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
// CHECK-NEXT: store <4 x i512> [[EXTRACTVEC]], ptr [[A_ADDR]], align 256
-// CHECK-NEXT: [[LOADVEC41:%.*]] = load <4 x i512>, ptr [[A_ADDR]], align 256
-// CHECK-NEXT: [[EXTRACTVEC2:%.*]] = shufflevector <4 x i512> [[LOADVEC41]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: [[LOADVEC43:%.*]] = load <4 x i512>, ptr [[A_ADDR]], align 256
-// CHECK-NEXT: [[EXTRACTVEC4:%.*]] = shufflevector <4 x i512> [[LOADVEC43]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN1:%.*]] = load <4 x i512>, ptr [[A_ADDR]], align 256
+// CHECK-NEXT: [[EXTRACTVEC2:%.*]] = shufflevector <4 x i512> [[LOADVECN1]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
+// CHECK-NEXT: [[LOADVECN3:%.*]] = load <4 x i512>, ptr [[A_ADDR]], align 256
+// CHECK-NEXT: [[EXTRACTVEC4:%.*]] = shufflevector <4 x i512> [[LOADVECN3]], <4 x i512> poison, <3 x i32> <i32 0, i32 1, i32 2>
// CHECK-NEXT: [[ADD:%.*]] = add <3 x i512> [[EXTRACTVEC2]], [[EXTRACTVEC4]]
// CHECK-NEXT: ret <3 x i512> [[ADD]]
//
diff --git a/clang/test/CodeGenOpenCL/amdgpu-alignment.cl b/clang/test/CodeGenOpenCL/amdgpu-alignment.cl
index 8f57713fe1f041..3c2653bf34124f 100644
--- a/clang/test/CodeGenOpenCL/amdgpu-alignment.cl
+++ b/clang/test/CodeGenOpenCL/amdgpu-alignment.cl
@@ -106,7 +106,7 @@ typedef double __attribute__((ext_vector_type(16))) double16;
// CHECK: store volatile <16 x i16> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v16i16, align 32
// CHECK: store volatile i32 0, ptr addrspace(3) @local_memory_alignment_global.lds_i32, align 4
// CHECK: store volatile <2 x i32> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v2i32, align 8
-// CHECK: store volatile <4 x i32> <i32 0, i32 0, i32 0, i32 undef>, ptr addrspace(3) @local_memory_alignment_global.lds_v3i32, align 16
+// CHECK: store volatile <3 x i32> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v3i32, align 16
// CHECK: store volatile <4 x i32> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v4i32, align 16
// CHECK: store volatile <8 x i32> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v8i32, align 32
// CHECK: store volatile <16 x i32> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v16i32, align 64
@@ -124,7 +124,7 @@ typedef double __attribute__((ext_vector_type(16))) double16;
// CHECK: store volatile <16 x half> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v16f16, align 32
// CHECK: store volatile float 0.000000e+00, ptr addrspace(3) @local_memory_alignment_global.lds_f32, align 4
// CHECK: store volatile <2 x float> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v2f32, align 8
-// CHECK: store volatile <4 x float> <float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float undef>, ptr addrspace(3) @local_memory_alignment_global.lds_v3f32, align 16
+// CHECK: store volatile <3 x float> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v3f32, align 16
// CHECK: store volatile <4 x float> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v4f32, align 16
// CHECK: store volatile <8 x float> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v8f32, align 32
// CHECK: store volatile <16 x float> zeroinitializer, ptr addrspace(3) @local_memory_alignment_global.lds_v16f32, align 64
@@ -393,7 +393,7 @@ kernel void local_memory_alignment_arg(
// CHECK: store volatile <16 x i16> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 32
// CHECK: store volatile i32 0, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 4
// CHECK: store volatile <2 x i32> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 8
-// CHECK: store volatile <4 x i32> <i32 0, i32 0, i32 0, i32 undef>, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
+// CHECK: store volatile <3 x i32> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
// CHECK: store volatile <4 x i32> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
// CHECK: store volatile <8 x i32> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 32
// CHECK: store volatile <16 x i32> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 64
@@ -411,7 +411,7 @@ kernel void local_memory_alignment_arg(
// CHECK: store volatile <16 x half> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 32
// CHECK: store volatile float 0.000000e+00, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 4
// CHECK: store volatile <2 x float> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 8
-// CHECK: store volatile <4 x float> <float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float undef>, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
+// CHECK: store volatile <3 x float> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
// CHECK: store volatile <4 x float> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 16
// CHECK: store volatile <8 x float> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 32
// CHECK: store volatile <16 x float> zeroinitializer, ptr addrspace(5) %arraydecay{{[0-9]+}}, align 64
diff --git a/clang/test/CodeGenOpenCL/preserve_vec3.cl b/clang/test/CodeGenOpenCL/preserve_vec3.cl
index c84effe0c4b6e3..747cc301feff6c 100644
--- a/clang/test/CodeGenOpenCL/preserve_vec3.cl
+++ b/clang/test/CodeGenOpenCL/preserve_vec3.cl
@@ -1,5 +1,5 @@
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
-// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown -fpreserve-vec3-type | FileCheck %s
+// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown | FileCheck %s
typedef char char3 __attribute__((ext_vector_type(3)));
typedef char char8 __attribute__((ext_vector_type(8)));
@@ -9,10 +9,11 @@ typedef float float3 __attribute__((ext_vector_type(3)));
typedef float float4 __attribute__((ext_vector_type(4)));
// CHECK-LABEL: define dso_local spir_kernel void @foo(
-// CHECK-SAME: ptr addrspace(1) nocapture noundef readonly align 16 [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 12)) [[B:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] !kernel_arg_addr_space [[META3:![0-9]+]] !kernel_arg_access_qual [[META4:![0-9]+]] !kernel_arg_type [[META5:![0-9]+]] !kernel_arg_base_type [[META6:![0-9]+]] !kernel_arg_type_qual [[META7:![0-9]+]] {
+// CHECK-SAME: ptr addrspace(1) nocapture noundef readonly align 16 [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 16)) [[B:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] !kernel_arg_addr_space [[META3:![0-9]+]] !kernel_arg_access_qual [[META4:![0-9]+]] !kernel_arg_type [[META5:![0-9]+]] !kernel_arg_base_type [[META6:![0-9]+]] !kernel_arg_type_qual [[META7:![0-9]+]] {
// CHECK-NEXT: [[ENTRY:.*:]]
-// CHECK-NEXT: [[TMP0:%.*]] = load <3 x float>, ptr addrspace(1) [[A]], align 16, !tbaa [[TBAA8:![0-9]+]]
-// CHECK-NEXT: store <3 x float> [[TMP0]], ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x float>, ptr addrspace(1) [[A]], align 16
+// CHECK-NEXT: [[EXTRACTVEC1:%.*]] = shufflevector <4 x float> [[LOADVECN]], <4 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x float> [[EXTRACTVEC1]], ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8:![0-9]+]]
// CHECK-NEXT: ret void
//
void kernel foo(global float3 *a, global float3 *b) {
@@ -20,11 +21,11 @@ void kernel foo(global float3 *a, global float3 *b) {
}
// CHECK-LABEL: define dso_local spir_kernel void @float4_to_float3(
-// CHECK-SAME: ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 12)) [[A:%.*]], ptr addrspace(1) nocapture noundef readonly align 16 [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META11:![0-9]+]] !kernel_arg_base_type [[META12:![0-9]+]] !kernel_arg_type_qual [[META7]] {
+// CHECK-SAME: ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 16)) [[A:%.*]], ptr addrspace(1) nocapture noundef readonly align 16 [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META11:![0-9]+]] !kernel_arg_base_type [[META12:![0-9]+]] !kernel_arg_type_qual [[META7]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[TMP0:%.*]] = load <4 x float>, ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8]]
-// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <4 x float> [[TMP0]], <4 x float> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: store <3 x float> [[ASTYPE]], ptr addrspace(1) [[A]], align 16, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <4 x float> [[TMP0]], <4 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x float> [[EXTRACTVEC]], ptr addrspace(1) [[A]], align 16, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
void kernel float4_to_float3(global float3 *a, global float4 *b) {
@@ -34,8 +35,8 @@ void kernel float4_to_float3(global float3 *a, global float4 *b) {
// CHECK-LABEL: define dso_local spir_kernel void @float3_to_float4(
// CHECK-SAME: ptr addrspace(1) nocapture noundef readonly align 16 [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 16)) [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META11]] !kernel_arg_base_type [[META12]] !kernel_arg_type_qual [[META7]] {
// CHECK-NEXT: [[ENTRY:.*:]]
-// CHECK-NEXT: [[TMP0:%.*]] = load <3 x float>, ptr addrspace(1) [[A]], align 16, !tbaa [[TBAA8]]
-// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <3 x float> [[TMP0]], <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x float>, ptr addrspace(1) [[A]], align 16
+// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <4 x float> [[LOADVECN]], <4 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
// CHECK-NEXT: store <4 x float> [[ASTYPE]], ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
@@ -46,9 +47,9 @@ void kernel float3_to_float4(global float3 *a, global float4 *b) {
// CHECK-LABEL: define dso_local spir_kernel void @float3_to_double2(
// CHECK-SAME: ptr addrspace(1) nocapture noundef readonly align 16 [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly align 16 initializes((0, 16)) [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META13:![0-9]+]] !kernel_arg_base_type [[META14:![0-9]+]] !kernel_arg_type_qual [[META7]] {
// CHECK-NEXT: [[ENTRY:.*:]]
-// CHECK-NEXT: [[TMP0:%.*]] = load <3 x float>, ptr addrspace(1) [[A]], align 16, !tbaa [[TBAA8]]
-// CHECK-NEXT: [[TMP1:%.*]] = shufflevector <3 x float> [[TMP0]], <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
-// CHECK-NEXT: store <4 x float> [[TMP1]], ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[LOADVECN:%.*]] = load <4 x float>, ptr addrspace(1) [[A]], align 16
+// CHECK-NEXT: [[TMP0:%.*]] = shufflevector <4 x float> [[LOADVECN]], <4 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[B]], align 16, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
void kernel float3_to_double2(global float3 *a, global double2 *b) {
@@ -56,11 +57,11 @@ void kernel float3_to_double2(global float3 *a, global double2 *b) {
}
// CHECK-LABEL: define dso_local spir_kernel void @char8_to_short3(
-// CHECK-SAME: ptr addrspace(1) nocapture noundef writeonly align 8 initializes((0, 6)) [[A:%.*]], ptr addrspace(1) nocapture noundef readonly align 8 [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META15:![0-9]+]] !kernel_arg_base_type [[META16:![0-9]+]] !kernel_arg_type_qual [[META7]] {
+// CHECK-SAME: ptr addrspace(1) nocapture noundef writeonly align 8 initializes((0, 8)) [[A:%.*]], ptr addrspace(1) nocapture noundef readonly align 8 [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !kernel_arg_addr_space [[META3]] !kernel_arg_access_qual [[META4]] !kernel_arg_type [[META15:![0-9]+]] !kernel_arg_base_type [[META16:![0-9]+]] !kernel_arg_type_qual [[META7]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[TMP0:%.*]] = load <4 x i16>, ptr addrspace(1) [[B]], align 8, !tbaa [[TBAA8]]
-// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <4 x i16> [[TMP0]], <4 x i16> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: store <3 x i16> [[ASTYPE]], ptr addrspace(1) [[A]], align 8, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <4 x i16> [[TMP0]], <4 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x i16> [[EXTRACTVEC]], ptr addrspace(1) [[A]], align 8, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
void kernel char8_to_short3(global short3 *a, global char8 *b) {
@@ -70,8 +71,8 @@ void kernel char8_to_short3(global short3 *a, global char8 *b) {
// CHECK-LABEL: define dso_local spir_func void @from_char3(
// CHECK-SAME: <3 x i8> noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 4)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1:[0-9]+]] {
// CHECK-NEXT: [[ENTRY:.*:]]
-// CHECK-NEXT: [[TMP0:%.*]] = shufflevector <3 x i8> [[A]], <3 x i8> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
-// CHECK-NEXT: store <4 x i8> [[TMP0]], ptr addrspace(1) [[OUT]], align 4, !tbaa [[TBAA17:![0-9]+]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <3 x i8> [[A]], <3 x i8> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x i8> [[EXTRACTVEC]], ptr addrspace(1) [[OUT]], align 4, !tbaa [[TBAA17:![0-9]+]]
// CHECK-NEXT: ret void
//
void from_char3(char3 a, global int *out) {
@@ -81,8 +82,8 @@ void from_char3(char3 a, global int *out) {
// CHECK-LABEL: define dso_local spir_func void @from_short3(
// CHECK-SAME: <3 x i16> noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 8)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1]] {
// CHECK-NEXT: [[ENTRY:.*:]]
-// CHECK-NEXT: [[TMP0:%.*]] = shufflevector <3 x i16> [[A]], <3 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
-// CHECK-NEXT: store <4 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[TBAA19:![0-9]+]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <3 x i16> [[A]], <3 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x i16> [[EXTRACTVEC]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[TBAA19:![0-9]+]]
// CHECK-NEXT: ret void
//
void from_short3(short3 a, global long *out) {
@@ -90,11 +91,11 @@ void from_short3(short3 a, global long *out) {
}
// CHECK-LABEL: define dso_local spir_func void @scalar_to_char3(
-// CHECK-SAME: i32 noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 3)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1]] {
+// CHECK-SAME: i32 noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 4)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[TMP0:%.*]] = bitcast i32 [[A]] to <4 x i8>
-// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <4 x i8> [[TMP0]], <4 x i8> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: store <3 x i8> [[ASTYPE]], ptr addrspace(1) [[OUT]], align 4, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <4 x i8> [[TMP0]], <4 x i8> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x i8> [[EXTRACTVEC]], ptr addrspace(1) [[OUT]], align 4, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
void scalar_to_char3(int a, global char3 *out) {
@@ -102,11 +103,11 @@ void scalar_to_char3(int a, global char3 *out) {
}
// CHECK-LABEL: define dso_local spir_func void @scalar_to_short3(
-// CHECK-SAME: i64 noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 6)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1]] {
+// CHECK-SAME: i64 noundef [[A:%.*]], ptr addrspace(1) nocapture noundef writeonly initializes((0, 8)) [[OUT:%.*]]) local_unnamed_addr #[[ATTR1]] {
// CHECK-NEXT: [[ENTRY:.*:]]
// CHECK-NEXT: [[TMP0:%.*]] = bitcast i64 [[A]] to <4 x i16>
-// CHECK-NEXT: [[ASTYPE:%.*]] = shufflevector <4 x i16> [[TMP0]], <4 x i16> poison, <3 x i32> <i32 0, i32 1, i32 2>
-// CHECK-NEXT: store <3 x i16> [[ASTYPE]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[TBAA8]]
+// CHECK-NEXT: [[EXTRACTVEC:%.*]] = shufflevector <4 x i16> [[TMP0]], <4 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
+// CHECK-NEXT: store <4 x i16> [[EXTRACTVEC]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[TBAA8]]
// CHECK-NEXT: ret void
//
void scalar_to_short3(long a, global short3 *out) {
>From 3898bdc545214cf8e9de8db2c333a0d910808637 Mon Sep 17 00:00:00 2001
From: Shilei Tian <i at tianshilei.me>
Date: Tue, 21 Jan 2025 09:17:46 -0500
Subject: [PATCH 2/2] Update clang/lib/CodeGen/CGExpr.cpp
Co-authored-by: Matt Arsenault <Matthew.Arsenault at amd.com>
---
clang/lib/CodeGen/CGExpr.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 78bac4c688dd3c..d6881b1327e2da 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -2152,7 +2152,7 @@ void CodeGenFunction::EmitStoreOfScalar(llvm::Value *Value, Address Addr,
auto *NewVecTy =
CGM.getABIInfo().getOptimalVectorMemoryType(VecTy, getLangOpts());
if (!ClangVecTy->isExtVectorBoolType() && VecTy != NewVecTy) {
- SmallVector<int, 4> Mask(NewVecTy->getNumElements(), -1);
+ SmallVector<int, 16> Mask(NewVecTy->getNumElements(), -1);
std::iota(Mask.begin(), Mask.begin() + VecTy->getNumElements(), 0);
Value = Builder.CreateShuffleVector(Value, Mask, "extractVec");
SrcTy = NewVecTy;
More information about the cfe-commits
mailing list