replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[InstCombine][SSE] Added support to VPERMILVAR to shuffle combine to accept UNDEF elements. 

llvm-svn: 268204
This commit is contained in:
Simon Pilgrim 2016-05-01 20:22:42 +00:00
parent cc7f567b6a
commit eeacc40e27
2 changed files with 36 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
View File

@ -604,7 +604,7 @@ static Value *simplifyX86pshufb(const IntrinsicInst &II,
"Unexpected number of elements in shuffle mask!");
// Construct a shuffle mask from constant integers or UNDEFs.
Constant *Indexes[32] = { NULL };
Constant *Indexes[32] = {NULL};
// Each byte in the shuffle control mask forms an index to permute the
// corresponding byte in the destination operand.
@ -644,39 +644,46 @@ static Value *simplifyX86vpermilvar(const IntrinsicInst &II,
if (!V)
return nullptr;
unsigned Size = cast<VectorType>(V->getType())->getNumElements();
assert(Size == 8 || Size == 4 || Size == 2);
auto *MaskEltTy = Type::getInt32Ty(II.getContext());
unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
assert(NumElts == 8 || NumElts == 4 || NumElts == 2);
// Initialize the resulting shuffle mask to all zeroes.
uint32_t Indexes[8] = { 0 };
// Construct a shuffle mask from constant integers or UNDEFs.
Constant *Indexes[8] = {NULL};
// The intrinsics only read one or two bits, clear the rest.
for (unsigned I = 0; I < Size; ++I) {
for (unsigned I = 0; I < NumElts; ++I) {
Constant *COp = V->getAggregateElement(I);
if (!COp || !isa<ConstantInt>(COp))
if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
return nullptr;
int32_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue() & 0x3;
if (isa<UndefValue>(COp)) {
Indexes[I] = UndefValue::get(MaskEltTy);
continue;
}
APInt Index = cast<ConstantInt>(COp)->getValue();
Index = Index.zextOrTrunc(32).getLoBits(2);
// The PD variants uses bit 1 to select per-lane element index, so
// shift down to convert to generic shuffle mask index.
if (II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
Index >>= 1;
Indexes[I] = Index;
Index = Index.lshr(1);
// The _256 variants are a bit trickier since the mask bits always index
// into the corresponding 128 half. In order to convert to a generic
// shuffle, we have to make that explicit.
if ((II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) &&
((NumElts / 2) <= I)) {
Index += APInt(32, NumElts / 2);
}
Indexes[I] = ConstantInt::get(MaskEltTy, Index);
}
// The _256 variants are a bit trickier since the mask bits always index
// into the corresponding 128 half. In order to convert to a generic
// shuffle, we have to make that explicit.
if (II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
II.getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
for (unsigned I = Size / 2; I < Size; ++I)
Indexes[I] += Size / 2;
}
auto ShuffleMask =
ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts));
auto V1 = II.getArgOperand(0);
auto V2 = UndefValue::get(V1->getType());
return Builder.CreateShuffleVector(V1, V2, ShuffleMask);

16
llvm/test/Transforms/InstCombine/x86-avx.ll
View File

@ -117,8 +117,8 @@ define <4 x double> @test_vpermilvar_pd_256(<4 x double> %v) {
define <4 x float> @undef_test_vpermilvar_ps(<4 x float> %v) {
; CHECK-LABEL: @undef_test_vpermilvar_ps(
; CHECK-NEXT: [[A:%.*]] = tail call <4 x float> @llvm.x86.avx.vpermilvar.ps(<4 x float> %v, <4 x i32> <i32 undef, i32 2, i32 1, i32 undef>)
; CHECK-NEXT: ret <4 x float> [[A]]
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 undef, i32 2, i32 1, i32 undef>
; CHECK-NEXT: ret <4 x float> [[TMP1]]
;
%a = tail call <4 x float> @llvm.x86.avx.vpermilvar.ps(<4 x float> %v, <4 x i32> <i32 undef, i32 2, i32 1, i32 undef>)
ret <4 x float> %a
@ -126,8 +126,8 @@ define <4 x float> @undef_test_vpermilvar_ps(<4 x float> %v) {
define <8 x float> @undef_test_vpermilvar_ps_256(<8 x float> %v) {
; CHECK-LABEL: @undef_test_vpermilvar_ps_256(
; CHECK-NEXT: [[A:%.*]] = tail call <8 x float> @llvm.x86.avx.vpermilvar.ps.256(<8 x float> %v, <8 x i32> <i32 undef, i32 6, i32 5, i32 undef, i32 3, i32 2, i32 1, i32 0>)
; CHECK-NEXT: ret <8 x float> [[A]]
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <8 x float> %v, <8 x float> undef, <8 x i32> <i32 undef, i32 2, i32 1, i32 undef, i32 7, i32 6, i32 5, i32 4>
; CHECK-NEXT: ret <8 x float> [[TMP1]]
;
%a = tail call <8 x float> @llvm.x86.avx.vpermilvar.ps.256(<8 x float> %v, <8 x i32> <i32 undef, i32 6, i32 5, i32 undef, i32 3, i32 2, i32 1, i32 0>)
ret <8 x float> %a
@ -135,8 +135,8 @@ define <8 x float> @undef_test_vpermilvar_ps_256(<8 x float> %v) {
define <2 x double> @undef_test_vpermilvar_pd(<2 x double> %v) {
; CHECK-LABEL: @undef_test_vpermilvar_pd(
; CHECK-NEXT: [[A:%.*]] = tail call <2 x double> @llvm.x86.avx.vpermilvar.pd(<2 x double> %v, <2 x i64> <i64 undef, i64 0>)
; CHECK-NEXT: ret <2 x double> [[A]]
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x double> %v, <2 x double> undef, <2 x i32> <i32 undef, i32 0>
; CHECK-NEXT: ret <2 x double> [[TMP1]]
;
%a = tail call <2 x double> @llvm.x86.avx.vpermilvar.pd(<2 x double> %v, <2 x i64> <i64 undef, i64 0>)
ret <2 x double> %a
@ -144,8 +144,8 @@ define <2 x double> @undef_test_vpermilvar_pd(<2 x double> %v) {
define <4 x double> @undef_test_vpermilvar_pd_256(<4 x double> %v) {
; CHECK-LABEL: @undef_test_vpermilvar_pd_256(
; CHECK-NEXT: [[A:%.*]] = tail call <4 x double> @llvm.x86.avx.vpermilvar.pd.256(<4 x double> %v, <4 x i64> <i64 undef, i64 1, i64 2, i64 undef>)
; CHECK-NEXT: ret <4 x double> [[A]]
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x double> %v, <4 x double> undef, <4 x i32> <i32 undef, i32 0, i32 3, i32 undef>
; CHECK-NEXT: ret <4 x double> [[TMP1]]
;
%a = tail call <4 x double> @llvm.x86.avx.vpermilvar.pd.256(<4 x double> %v, <4 x i64> <i64 undef, i64 1, i64 2, i64 undef>)
ret <4 x double> %a