Optimize chained bitcasts of the form A->B->A.
Undo r138722 and change isEliminableCastPair to allow this case.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@138756
91177308-0d34-0410-b5e6-
96231b3b80d8
if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
return Constant::getAllOnesValue(DestTy);
if (C->isAllOnesValue() && !DestTy->isX86_MMXTy())
return Constant::getAllOnesValue(DestTy);
- // Bitcast of Bitcast can be done using a single cast.
- ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
- if (CE && CE->getOpcode() == Instruction::BitCast) {
- return ConstantExpr::getBitCast(CE->getOperand(0), DestTy);
- }
-
// The code below only handles casts to vectors currently.
VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
if (DestVTy == 0)
// The code below only handles casts to vectors currently.
VectorType *DestVTy = dyn_cast<VectorType>(DestTy);
if (DestVTy == 0)
if (DestTy == Src->getType())
return ReplaceInstUsesWith(CI, Src);
if (DestTy == Src->getType())
return ReplaceInstUsesWith(CI, Src);
- // Bitcasts are transitive.
- if (BitCastInst* BSrc = dyn_cast<BitCastInst>(Src)) {
- return CastInst::Create(Instruction::BitCast, BSrc->getOperand(0), DestTy);
- }
-
if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
PointerType *SrcPTy = cast<PointerType>(SrcTy);
Type *DstElTy = DstPTy->getElementType();
if (PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
PointerType *SrcPTy = cast<PointerType>(SrcTy);
Type *DstElTy = DstPTy->getElementType();
/// If no such cast is permited, the function returns 0.
unsigned CastInst::isEliminableCastPair(
Instruction::CastOps firstOp, Instruction::CastOps secondOp,
/// If no such cast is permited, the function returns 0.
unsigned CastInst::isEliminableCastPair(
Instruction::CastOps firstOp, Instruction::CastOps secondOp,
- Type *SrcTy, Type *MidTy, Type *DstTy, Type *IntPtrTy)
-{
+ Type *SrcTy, Type *MidTy, Type *DstTy, Type *IntPtrTy) {
// Define the 144 possibilities for these two cast instructions. The values
// in this matrix determine what to do in a given situation and select the
// case in the switch below. The rows correspond to firstOp, the columns
// Define the 144 possibilities for these two cast instructions. The values
// in this matrix determine what to do in a given situation and select the
// case in the switch below. The rows correspond to firstOp, the columns
};
// If either of the casts are a bitcast from scalar to vector, disallow the
};
// If either of the casts are a bitcast from scalar to vector, disallow the
- // merging.
- if ((firstOp == Instruction::BitCast &&
- isa<VectorType>(SrcTy) != isa<VectorType>(MidTy)) ||
- (secondOp == Instruction::BitCast &&
- isa<VectorType>(MidTy) != isa<VectorType>(DstTy)))
- return 0; // Disallowed
+ // merging. However, bitcast of A->B->A are allowed.
+ bool isFirstBitcast = (firstOp == Instruction::BitCast);
+ bool isSecondBitcast = (secondOp == Instruction::BitCast);
+ bool chainedBitcast = (SrcTy == DstTy && isFirstBitcast && isSecondBitcast);
+
+ // Check if any of the bitcasts convert scalars<->vectors.
+ if ((isFirstBitcast && isa<VectorType>(SrcTy) != isa<VectorType>(MidTy)) ||
+ (isSecondBitcast && isa<VectorType>(MidTy) != isa<VectorType>(DstTy)))
+ // Unless we are bitcasing to the original type, disallow optimizations.
+ if (!chainedBitcast) return 0;
int ElimCase = CastResults[firstOp-Instruction::CastOpsBegin]
[secondOp-Instruction::CastOpsBegin];
int ElimCase = CastResults[firstOp-Instruction::CastOpsBegin]
[secondOp-Instruction::CastOpsBegin];
define <4 x float> @test64(<4 x float> %c) nounwind {
%t0 = bitcast <4 x float> %c to <4 x i32>
define <4 x float> @test64(<4 x float> %c) nounwind {
%t0 = bitcast <4 x float> %c to <4 x i32>
- %t1 = bitcast <4 x i32> %t0 to <2 x double>
- %t2 = bitcast <2 x double> %t1 to <4 x float>
- ret <4 x float> %t2
+ %t1 = bitcast <4 x i32> %t0 to <4 x float>
+ ret <4 x float> %t1
; CHECK: @test64
; CHECK-NEXT: ret <4 x float> %c
}
; CHECK: @test64
; CHECK-NEXT: ret <4 x float> %c
}
+define <4 x float> @test65(<4 x float> %c) nounwind {
+ %t0 = bitcast <4 x float> %c to <2 x double>
+ %t1 = bitcast <2 x double> %t0 to <4 x float>
+ ret <4 x float> %t1
+; CHECK: @test65
+; CHECK-NEXT: ret <4 x float> %c
+}
+
+define <2 x float> @test66(<2 x float> %c) nounwind {
+ %t0 = bitcast <2 x float> %c to double
+ %t1 = bitcast double %t0 to <2 x float>
+ ret <2 x float> %t1
+; CHECK: @test66
+; CHECK-NEXT: ret <2 x float> %c
+}
+
define float @test2c() {
ret float extractelement (<2 x float> bitcast (double bitcast (<2 x float> <float -1.000000e+00, float -1.000000e+00> to double) to <2 x float>), i32 0)
; CHECK: @test2c
; CHECK-NOT: extractelement
}
define float @test2c() {
ret float extractelement (<2 x float> bitcast (double bitcast (<2 x float> <float -1.000000e+00, float -1.000000e+00> to double) to <2 x float>), i32 0)
; CHECK: @test2c
; CHECK-NOT: extractelement
}
+
+define i64 @test_mmx(<2 x i32> %c) nounwind {
+ %A = bitcast <2 x i32> %c to x86_mmx
+ %B = bitcast x86_mmx %A to <2 x i32>
+ %C = bitcast <2 x i32> %B to i64
+ ret i64 %C
+; CHECK: @test_mmx
+; CHECK-NOT: x86_mmx
+}
+
+define i64 @test_mmx_const(<2 x i32> %c) nounwind {
+ %A = bitcast <2 x i32> zeroinitializer to x86_mmx
+ %B = bitcast x86_mmx %A to <2 x i32>
+ %C = bitcast <2 x i32> %B to i64
+ ret i64 %C
+; CHECK: @test_mmx_const
+; CHECK-NOT: x86_mmx
+}