Hi rsmith,
With this patch the following code is able to be correctly optimized.
#include <x86intrin.h>
#include <immintrin.h>
#include <avxintrin.h>
#include <avx2intrin.h>
#include <cstdio>
#include <cinttypes>
__m128 bss4( const __m128 *ptr, size_t i, size_t j )
{
float f = ptr[i][j];
return (__m128) { f, f, f, f };
}
Previously an unneeded trunc + zext would be emitted.
http://reviews.llvm.org/D3974
Files:
lib/CodeGen/CGBuiltin.cpp
lib/CodeGen/CGExpr.cpp
lib/CodeGen/CGExprScalar.cpp
test/CodeGen/builtinshufflevector2.c
Index: lib/CodeGen/CGBuiltin.cpp
===================================================================
--- lib/CodeGen/CGBuiltin.cpp
+++ lib/CodeGen/CGBuiltin.cpp
@@ -2493,7 +2493,7 @@
Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
int j = 0;
- ConstantInt *C0 = ConstantInt::get(CGF.Int32Ty, 0);
+ ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
ai != ae; ++ai, ++j) {
llvm::Type *ArgTy = ai->getType();
@@ -2733,7 +2733,7 @@
Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
LoadInst *Ld = Builder.CreateLoad(Ops[0]);
Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue());
- llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
+ llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
return EmitNeonSplat(Ops[0], CI);
}
@@ -3703,16 +3703,16 @@
llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
Op = Builder.CreateBitCast(Op, Int16Ty);
Value *V = UndefValue::get(VTy);
- llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
+ llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
Op = Builder.CreateInsertElement(V, Op, CI);
return Op;
}
Value *CodeGenFunction::vectorWrapScalar8(Value *Op) {
llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
Op = Builder.CreateBitCast(Op, Int8Ty);
Value *V = UndefValue::get(VTy);
- llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
+ llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
Op = Builder.CreateInsertElement(V, Op, CI);
return Op;
}
@@ -3729,7 +3729,7 @@
Ops[1] = vectorWrapScalar8(Ops[1]);
llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8);
Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
- Constant *CI = ConstantInt::get(Int32Ty, 0);
+ Constant *CI = ConstantInt::get(SizeTy, 0);
return Builder.CreateExtractElement(V, CI, "lane0");
}
@@ -3745,7 +3745,7 @@
Ops[1] = vectorWrapScalar16(Ops[1]);
llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4);
Value *V = EmitNeonCall(CGM.getIntrinsic(Int, VTy), Ops, Name);
- Constant *CI = ConstantInt::get(Int32Ty, 0);
+ Constant *CI = ConstantInt::get(SizeTy, 0);
return Builder.CreateExtractElement(V, CI, "lane0");
}
@@ -3950,8 +3950,8 @@
Value *Vec = EmitScalarExpr(E->getArg(0));
// The vector is v2f64, so make sure it's bitcast to that.
Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
- llvm::Value *Idx0 = llvm::ConstantInt::get(Int32Ty, 0);
- llvm::Value *Idx1 = llvm::ConstantInt::get(Int32Ty, 1);
+ llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
+ llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
// Pairwise addition of a v2f64 into a scalar f64.
@@ -3963,8 +3963,8 @@
Value *Vec = EmitScalarExpr(E->getArg(0));
// The vector is v2f64, so make sure it's bitcast to that.
Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
- llvm::Value *Idx0 = llvm::ConstantInt::get(Int32Ty, 0);
- llvm::Value *Idx1 = llvm::ConstantInt::get(Int32Ty, 1);
+ llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
+ llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
// Pairwise addition of a v2f64 into a scalar f64.
@@ -3976,8 +3976,8 @@
Value *Vec = EmitScalarExpr(E->getArg(0));
// The vector is v2f32, so make sure it's bitcast to that.
Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
- llvm::Value *Idx0 = llvm::ConstantInt::get(Int32Ty, 0);
- llvm::Value *Idx1 = llvm::ConstantInt::get(Int32Ty, 1);
+ llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
+ llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
// Pairwise addition of a v2f32 into a scalar f32.
@@ -4228,7 +4228,7 @@
llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
ProductOps, "vqdmlXl");
- Constant *CI = ConstantInt::get(Int32Ty, 0);
+ Constant *CI = ConstantInt::get(SizeTy, 0);
Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
@@ -4325,7 +4325,7 @@
llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4);
Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
ProductOps, "vqdmlXl");
- Constant *CI = ConstantInt::get(Int32Ty, 0);
+ Constant *CI = ConstantInt::get(SizeTy, 0);
Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
Ops.pop_back();
@@ -5232,7 +5232,7 @@
Ty = llvm::PointerType::getUnqual(VTy->getElementType());
Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
Ops[0] = Builder.CreateLoad(Ops[0]);
- llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
+ llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
return EmitNeonSplat(Ops[0], CI);
}
@@ -5611,7 +5611,7 @@
// extract (0, 1)
unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1;
- llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index);
+ llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index);
Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract");
// cast pointer to i64 & store
Index: lib/CodeGen/CGExpr.cpp
===================================================================
--- lib/CodeGen/CGExpr.cpp
+++ lib/CodeGen/CGExpr.cpp
@@ -1331,7 +1331,7 @@
const VectorType *ExprVT = LV.getType()->getAs<VectorType>();
if (!ExprVT) {
unsigned InIdx = getAccessedFieldNo(0, Elts);
- llvm::Value *Elt = llvm::ConstantInt::get(Int32Ty, InIdx);
+ llvm::Value *Elt = llvm::ConstantInt::get(SizeTy, InIdx);
return RValue::get(Builder.CreateExtractElement(Vec, Elt));
}
@@ -1590,7 +1590,7 @@
} else {
// If the Src is a scalar (not a vector) it must be updating one element.
unsigned InIdx = getAccessedFieldNo(0, Elts);
- llvm::Value *Elt = llvm::ConstantInt::get(Int32Ty, InIdx);
+ llvm::Value *Elt = llvm::ConstantInt::get(SizeTy, InIdx);
Vec = Builder.CreateInsertElement(Vec, SrcVal, Elt);
}
@@ -2307,7 +2307,6 @@
// Emit the vector as an lvalue to get its address.
LValue LHS = EmitLValue(E->getBase());
assert(LHS.isSimple() && "Can only subscript lvalue vectors here!");
- Idx = Builder.CreateIntCast(Idx, Int32Ty, IdxSigned, "vidx");
return LValue::MakeVectorElt(LHS.getAddress(), Idx,
E->getBase()->getType(), LHS.getAlignment());
}
Index: lib/CodeGen/CGExprScalar.cpp
===================================================================
--- lib/CodeGen/CGExprScalar.cpp
+++ lib/CodeGen/CGExprScalar.cpp
@@ -940,9 +940,8 @@
MTy->getNumElements());
Value* NewV = llvm::UndefValue::get(RTy);
for (unsigned i = 0, e = MTy->getNumElements(); i != e; ++i) {
- Value *IIndx = Builder.getInt32(i);
+ Value *IIndx = llvm::ConstantInt::get(CGF.SizeTy, i);
Value *Indx = Builder.CreateExtractElement(Mask, IIndx, "shuf_idx");
- Indx = Builder.CreateZExt(Indx, CGF.Int32Ty, "idx_zext");
Value *VExt = Builder.CreateExtractElement(LHS, Indx, "shuf_elt");
NewV = Builder.CreateInsertElement(NewV, VExt, IIndx, "shuf_ins");
@@ -1073,8 +1072,6 @@
if (CGF.SanOpts->ArrayBounds)
CGF.EmitBoundsCheck(E, E->getBase(), Idx, IdxTy, /*Accessed*/true);
- bool IdxSigned = IdxTy->isSignedIntegerOrEnumerationType();
- Idx = Builder.CreateIntCast(Idx, CGF.Int32Ty, IdxSigned, "vecidxcast");
return Builder.CreateExtractElement(Base, Idx, "vecext");
}
Index: test/CodeGen/builtinshufflevector2.c
===================================================================
--- test/CodeGen/builtinshufflevector2.c
+++ test/CodeGen/builtinshufflevector2.c
@@ -6,23 +6,23 @@
// CHECK-LABEL: define void @clang_shufflevector_v_v(
void clang_shufflevector_v_v( float4* A, float4 x, uint4 mask ) {
// CHECK: [[MASK:%.*]] = and <4 x i32> {{%.*}}, <i32 3, i32 3, i32 3, i32 3>
-// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i32 0
-// CHECK: [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i32 [[I]]
+// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i{{[0-9]+}} 0
+// CHECK: [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i{{[0-9]+}} [[I]]
//
// Here is where ToT Clang code generation makes a mistake.
// It uses [[I]] as the insertion index instead of 0.
// Similarly on the remaining insertelement.
-// CHECK: [[V:%[a-zA-Z0-9._]+]] = insertelement <4 x float> undef, float [[E]], i32 0
+// CHECK: [[V:%[a-zA-Z0-9._]+]] = insertelement <4 x float> undef, float [[E]], i{{[0-9]+}} 0
-// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i32 1
-// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i32 [[I]]
-// CHECK: [[V2:%.*]] = insertelement <4 x float> [[V]], float [[E]], i32 1
-// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i32 2
-// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i32 [[I]]
-// CHECK: [[V3:%.*]] = insertelement <4 x float> [[V2]], float [[E]], i32 2
-// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i32 3
-// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i32 [[I]]
-// CHECK: [[V4:%.*]] = insertelement <4 x float> [[V3]], float [[E]], i32 3
+// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i{{[0-9]+}} 1
+// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i{{[0-9]+}} [[I]]
+// CHECK: [[V2:%.*]] = insertelement <4 x float> [[V]], float [[E]], i{{[0-9]+}} 1
+// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i{{[0-9]+}} 2
+// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i{{[0-9]+}} [[I]]
+// CHECK: [[V3:%.*]] = insertelement <4 x float> [[V2]], float [[E]], i{{[0-9]+}} 2
+// CHECK: [[I:%.*]] = extractelement <4 x i32> [[MASK]], i{{[0-9]+}} 3
+// CHECK: [[E:%.*]] = extractelement <4 x float> [[X]], i{{[0-9]+}} [[I]]
+// CHECK: [[V4:%.*]] = insertelement <4 x float> [[V3]], float [[E]], i{{[0-9]+}} 3
// CHECK: store <4 x float> [[V4]], <4 x float>* {{%.*}},
*A = __builtin_shufflevector( x, mask );
}
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
