RPCS3/llvm
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm.git synced 2024-12-14 23:48:56 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

[X86] Fix DecodeVPERMVMask to handle cases where the constant pool entry has a different type than the shuffle itself.

Browse Source 
This is especially important for 32-bit targets with 64-bit shuffle elements.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@284453 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Craig Topper 2016-10-18 04:48:33 +00:00
parent 3c59fd410e
commit 63ae3007f1
4 changed files with 26 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
lib/Target/X86/X86ISelLowering.cpp
View File

@ -5091,7 +5091,7 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
break;
}
if (auto *C = getTargetConstantFromNode(MaskNode)) {
DecodeVPERMVMask(C, VT, Mask);
DecodeVPERMVMask(C, MaskEltSize, Mask);
break;
}
return false;

50
lib/Target/X86/X86ShuffleDecodeConstantPool.cpp
View File

@ -275,38 +275,32 @@ void DecodeVPPERMMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask) {
}
}
void DecodeVPERMVMask(const Constant *C, MVT VT,
void DecodeVPERMVMask(const Constant *C, unsigned ElSize,
SmallVectorImpl<int> &ShuffleMask) {
Type *MaskTy = C->getType();
if (MaskTy->isVectorTy()) {
unsigned NumElements = MaskTy->getVectorNumElements();
if (NumElements == VT.getVectorNumElements()) {
unsigned EltMaskSize = Log2_64(NumElements);
for (unsigned i = 0; i < NumElements; ++i) {
Constant *COp = C->getAggregateElement(i);
if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) {
ShuffleMask.clear();
return;
}
if (isa<UndefValue>(COp))
ShuffleMask.push_back(SM_SentinelUndef);
else {
APInt Element = cast<ConstantInt>(COp)->getValue();
Element = Element.getLoBits(EltMaskSize);
ShuffleMask.push_back(Element.getZExtValue());
}
}
unsigned MaskTySize = MaskTy->getPrimitiveSizeInBits();
(void)MaskTySize;
assert((MaskTySize == 128 || MaskTySize == 256 || MaskTySize == 512) &&
"Unexpected vector size.");
assert((ElSize == 8 || ElSize == 16 || ElSize == 32 || ElSize == 64) &&
"Unexpected vector element size.");
// The shuffle mask requires elements the same size as the target.
SmallBitVector UndefElts;
SmallVector<uint64_t, 8> RawMask;
if (!extractConstantMask(C, ElSize, UndefElts, RawMask))
return;
unsigned NumElts = RawMask.size();
for (unsigned i = 0; i != NumElts; ++i) {
if (UndefElts[i]) {
ShuffleMask.push_back(SM_SentinelUndef);
continue;
}
return;
int Index = RawMask[i] & (NumElts - 1);
ShuffleMask.push_back(Index);
}
// Scalar value; just broadcast it
if (!isa<ConstantInt>(C))
return;
uint64_t Element = cast<ConstantInt>(C)->getZExtValue();
int NumElements = VT.getVectorNumElements();
Element &= (1 << NumElements) - 1;
for (int i = 0; i < NumElements; ++i)
ShuffleMask.push_back(Element);
}
void DecodeVPERMV3Mask(const Constant *C, unsigned ElSize,

2
lib/Target/X86/X86ShuffleDecodeConstantPool.h
View File

@ -40,7 +40,7 @@ void DecodeVPERMIL2PMask(const Constant *C, unsigned MatchImm, unsigned ElSize,
void DecodeVPPERMMask(const Constant *C, SmallVectorImpl<int> &ShuffleMask);
/// Decode a VPERM W/D/Q/PS/PD mask from an IR-level vector constant.
void DecodeVPERMVMask(const Constant *C, MVT VT,
void DecodeVPERMVMask(const Constant *C, unsigned ElSize,
SmallVectorImpl<int> &ShuffleMask);
/// Decode a VPERMT2 W/D/Q/PS/PD mask from an IR-level vector constant.

10
test/CodeGen/X86/vector-shuffle-combining-avx512bw.ll
View File

@ -28,10 +28,6 @@ declare <32 x i16> @llvm.x86.avx512.mask.vpermi2var.hi.512(<32 x i16>, <32 x i16
define <8 x double> @combine_permvar_8f64_identity(<8 x double> %x0, <8 x double> %x1) {
; X32-LABEL: combine_permvar_8f64_identity:
; X32: # BB#0:
; X32-NEXT: vmovapd {{.*#+}} zmm1 = [7,0,6,0,5,0,4,0,3,0,2,0,1,0,0,0]
; X32-NEXT: vpermpd %zmm0, %zmm1, %zmm0
; X32-NEXT: vmovapd {{.*#+}} zmm1 = [7,0,14,0,5,0,12,0,3,0,10,0,1,0,8,0]
; X32-NEXT: vpermpd %zmm0, %zmm1, %zmm0
; X32-NEXT: retl
;
; X64-LABEL: combine_permvar_8f64_identity:
@ -703,8 +699,7 @@ define <8 x i64> @combine_permvar_8i64_as_permq_mask(<8 x i64> %x0, <8 x i64> %x
define <8 x double> @combine_permvar_8f64_as_permpd(<8 x double> %x0, <8 x double> %x1) {
; X32-LABEL: combine_permvar_8f64_as_permpd:
; X32: # BB#0:
; X32-NEXT: vmovapd {{.*#+}} zmm1 = <3,0,2,0,1,0,u,u,u,u,6,0,5,0,4,0>
; X32-NEXT: vpermpd %zmm0, %zmm1, %zmm0
; X32-NEXT: vpermpd {{.*#+}} zmm0 = zmm0[3,2,1,0,7,6,5,4]
; X32-NEXT: retl
;
; X64-LABEL: combine_permvar_8f64_as_permpd:
@ -719,8 +714,7 @@ define <8 x double> @combine_permvar_8f64_as_permpd_mask(<8 x double> %x0, <8 x
; X32: # BB#0:
; X32-NEXT: movzbl {{[0-9]+}}(%esp), %eax
; X32-NEXT: kmovd %eax, %k1
; X32-NEXT: vmovapd {{.*#+}} zmm2 = <3,0,2,0,1,0,u,u,u,u,6,0,5,0,4,0>
; X32-NEXT: vpermpd %zmm0, %zmm2, %zmm1 {%k1}
; X32-NEXT: vpermpd {{.*#+}} zmm1 {%k1} = zmm0[3,2,1,0,7,6,5,4]
; X32-NEXT: vmovapd %zmm1, %zmm0
; X32-NEXT: retl
;