[llvm] dec214d - [X86] combineINSERT_SUBVECTOR - generalise insert_subvector(x,extract(broadcast)) -> blend (#140516)
via llvm-commits
llvm-commits at lists.llvm.org
Wed May 21 01:28:56 PDT 2025
Author: Simon Pilgrim
Date: 2025-05-21T09:28:52+01:00
New Revision: dec214d5c638302b92fa1cfe0582531cd58a7f6d
URL: https://github.com/llvm/llvm-project/commit/dec214d5c638302b92fa1cfe0582531cd58a7f6d
DIFF: https://github.com/llvm/llvm-project/commit/dec214d5c638302b92fa1cfe0582531cd58a7f6d.diff
LOG: [X86] combineINSERT_SUBVECTOR - generalise insert_subvector(x,extract(broadcast)) -> blend (#140516)
Don't match against specific broadcast nodes and let isShuffleEquivalent handle it
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/insert-subvector-broadcast.ll
llvm/test/CodeGen/X86/widen_fadd.ll
llvm/test/CodeGen/X86/widen_fdiv.ll
llvm/test/CodeGen/X86/widen_fmul.ll
llvm/test/CodeGen/X86/widen_fsub.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e51d15a4d3825..3d5ef1fc28ec8 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -59284,36 +59284,34 @@ static SDValue combineINSERT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
!(Vec.isUndef() || ISD::isBuildVectorAllZeros(Vec.getNode())))) {
SDValue ExtSrc = SubVec.getOperand(0);
int ExtIdxVal = SubVec.getConstantOperandVal(1);
- if (ExtIdxVal != 0) {
- SmallVector<int, 64> Mask(VecNumElts);
- // First create an identity shuffle mask.
- for (int i = 0; i != VecNumElts; ++i)
- Mask[i] = i;
- // Now insert the extracted portion.
- for (int i = 0; i != SubVecNumElts; ++i)
- Mask[i + IdxVal] = i + ExtIdxVal + VecNumElts;
+ // Create a shuffle mask matching the extraction and insertion.
+ SmallVector<int, 64> Mask(VecNumElts);
+ std::iota(Mask.begin(), Mask.end(), 0);
+ std::iota(Mask.begin() + IdxVal, Mask.begin() + IdxVal + SubVecNumElts,
+ ExtIdxVal + VecNumElts);
+ if (ExtIdxVal != 0)
return DAG.getVectorShuffle(OpVT, dl, Vec, ExtSrc, Mask);
- }
- // If we're broadcasting, see if we can use a blend instead of
- // extract/insert pair. Ensure that the subvector is aligned with the
- // insertion/extractions.
- if ((ExtIdxVal % SubVecNumElts) == 0 && (IdxVal % SubVecNumElts) == 0 &&
- (ExtSrc.getOpcode() == X86ISD::VBROADCAST ||
- ExtSrc.getOpcode() == X86ISD::VBROADCAST_LOAD ||
- (ExtSrc.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD &&
- cast<MemIntrinsicSDNode>(ExtSrc)->getMemoryVT() == SubVecVT))) {
+ // See if we can use a blend instead of extract/insert pair.
+ SmallVector<int, 64> BlendMask(VecNumElts);
+ std::iota(BlendMask.begin(), BlendMask.end(), 0);
+ std::iota(BlendMask.begin() + IdxVal,
+ BlendMask.begin() + IdxVal + SubVecNumElts, VecNumElts + IdxVal);
+ if (isShuffleEquivalent(Mask, BlendMask, Vec, ExtSrc) &&
+ VecNumElts == (2 * SubVecNumElts)) {
+ assert((IdxVal == 0 || IdxVal == SubVecNumElts) &&
+ "Unaligned subvector insertion");
if (OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
- uint64_t BlendMask = IdxVal == 0 ? 0x0F : 0xF0;
SDValue Blend = DAG.getNode(
X86ISD::BLENDI, dl, MVT::v8f32, DAG.getBitcast(MVT::v8f32, Vec),
DAG.getBitcast(MVT::v8f32, ExtSrc),
- DAG.getTargetConstant(BlendMask, dl, MVT::i8));
+ DAG.getTargetConstant(IdxVal == 0 ? 0x0F : 0xF0, dl, MVT::i8));
return DAG.getBitcast(OpVT, Blend);
} else if (OpVT.is512BitVector() && SubVecVT.is256BitVector()) {
- SDValue Lo = DAG.getBitcast(MVT::v8f64, IdxVal == 0 ? ExtSrc : Vec);
- SDValue Hi = DAG.getBitcast(MVT::v8f64, IdxVal == 0 ? Vec : ExtSrc);
+ MVT ShufVT = OpVT.isInteger() ? MVT::v8i64 : MVT::v8f64;
+ SDValue Lo = DAG.getBitcast(ShufVT, IdxVal == 0 ? ExtSrc : Vec);
+ SDValue Hi = DAG.getBitcast(ShufVT, IdxVal == 0 ? Vec : ExtSrc);
SDValue Shuffle =
- DAG.getNode(X86ISD::SHUF128, dl, MVT::v8f64, Lo, Hi,
+ DAG.getNode(X86ISD::SHUF128, dl, ShufVT, Lo, Hi,
getV4X86ShuffleImm8ForMask({0, 1, 2, 3}, dl, DAG));
return DAG.getBitcast(OpVT, Shuffle);
}
diff --git a/llvm/test/CodeGen/X86/insert-subvector-broadcast.ll b/llvm/test/CodeGen/X86/insert-subvector-broadcast.ll
index 9b35857804022..0717a01ac2abc 100644
--- a/llvm/test/CodeGen/X86/insert-subvector-broadcast.ll
+++ b/llvm/test/CodeGen/X86/insert-subvector-broadcast.ll
@@ -7,9 +7,8 @@ define void @insert_subvector_broadcast_as_blend() {
; CHECK-NEXT: movq (%rax), %rax
; CHECK-NEXT: incq %rax
; CHECK-NEXT: vpbroadcastq %rax, %zmm0
-; CHECK-NEXT: vpslldq {{.*#+}} xmm1 = zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7]
-; CHECK-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm1
-; CHECK-NEXT: vshuff64x2 {{.*#+}} zmm1 = zmm1[0,1,2,3],zmm0[4,5,6,7]
+; CHECK-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; CHECK-NEXT: valignq {{.*#+}} zmm1 = zmm1[7],zmm0[0,1,2,3,4,5,6]
; CHECK-NEXT: vpcmpltq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm1, %k0
; CHECK-NEXT: vpcmpltq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k1
; CHECK-NEXT: kunpckbw %k0, %k1, %k1
diff --git a/llvm/test/CodeGen/X86/widen_fadd.ll b/llvm/test/CodeGen/X86/widen_fadd.ll
index c3700189d3d0e..f8cde4cf223a7 100644
--- a/llvm/test/CodeGen/X86/widen_fadd.ll
+++ b/llvm/test/CodeGen/X86/widen_fadd.ll
@@ -221,7 +221,7 @@ define void @widen_fadd_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {
; AVX512F-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2
; AVX512F-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpermt2pd %zmm2, %zmm5, %zmm0
-; AVX512F-NEXT: vinsertf64x4 $0, %ymm0, %zmm4, %zmm0
+; AVX512F-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-NEXT: vmovupd %zmm0, (%rdx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/widen_fdiv.ll b/llvm/test/CodeGen/X86/widen_fdiv.ll
index 2d9e3f60bee46..fdf895921ca67 100644
--- a/llvm/test/CodeGen/X86/widen_fdiv.ll
+++ b/llvm/test/CodeGen/X86/widen_fdiv.ll
@@ -182,7 +182,7 @@ define void @widen_fdiv_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {
; AVX512F-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2
; AVX512F-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpermt2pd %zmm2, %zmm5, %zmm0
-; AVX512F-NEXT: vinsertf64x4 $0, %ymm0, %zmm4, %zmm0
+; AVX512F-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-NEXT: vmovupd %zmm0, (%rdx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/widen_fmul.ll b/llvm/test/CodeGen/X86/widen_fmul.ll
index 6c3e0ff5a9bcd..16baa068fc24f 100644
--- a/llvm/test/CodeGen/X86/widen_fmul.ll
+++ b/llvm/test/CodeGen/X86/widen_fmul.ll
@@ -221,7 +221,7 @@ define void @widen_fmul_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {
; AVX512F-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2
; AVX512F-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpermt2pd %zmm2, %zmm5, %zmm0
-; AVX512F-NEXT: vinsertf64x4 $0, %ymm0, %zmm4, %zmm0
+; AVX512F-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-NEXT: vmovupd %zmm0, (%rdx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/widen_fsub.ll b/llvm/test/CodeGen/X86/widen_fsub.ll
index 7405d9b7b1c65..8dcd887ab4144 100644
--- a/llvm/test/CodeGen/X86/widen_fsub.ll
+++ b/llvm/test/CodeGen/X86/widen_fsub.ll
@@ -221,7 +221,7 @@ define void @widen_fsub_v2f32_v16f32(ptr %a0, ptr %b0, ptr %c0) {
; AVX512F-NEXT: vinsertf32x4 $1, %xmm3, %zmm2, %zmm2
; AVX512F-NEXT: vinsertf32x4 $1, %xmm1, %zmm0, %zmm0
; AVX512F-NEXT: vpermt2pd %zmm2, %zmm5, %zmm0
-; AVX512F-NEXT: vinsertf64x4 $0, %ymm0, %zmm4, %zmm0
+; AVX512F-NEXT: vshuff64x2 {{.*#+}} zmm0 = zmm0[0,1,2,3],zmm4[4,5,6,7]
; AVX512F-NEXT: vmovupd %zmm0, (%rdx)
; AVX512F-NEXT: vzeroupper
; AVX512F-NEXT: retq
More information about the llvm-commits
mailing list