[llvm] [X86] Widden binary shuffle to unary for i8/i16 (PR #96414)
Phoebe Wang via llvm-commits
llvm-commits at lists.llvm.org
Sun Jun 23 06:13:28 PDT 2024
https://github.com/phoebewang updated https://github.com/llvm/llvm-project/pull/96414
>From 9cb2fa23a555922bfd3a5e64c1c2535e9ef7bf90 Mon Sep 17 00:00:00 2001
From: Phoebe Wang <phoebe.wang at intel.com>
Date: Sun, 23 Jun 2024 21:07:17 +0800
Subject: [PATCH] [X86] Combine VPERMV3 to VPERMV for i8/i16
VPERM[I,T]2[B,W] are 3 uops on Skylake and Icelake so we try to use VPERMV if possible.
---
llvm/lib/Target/X86/X86ISelLowering.cpp | 27 +++++++++++++
.../X86/avx512-shuffles/partial_permute.ll | 38 +++++++------------
llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll | 3 +-
3 files changed, 42 insertions(+), 26 deletions(-)
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 82d2b301d854e..8c1f6797a1d1e 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -41273,6 +41273,33 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
return SDValue();
}
+ case X86ISD::VPERMV3: {
+ // VPERM[I,T]2[B,W] are 3 uops on Skylake and Icelake so we try to use
+ // VPERMV.
+ if (VT.is512BitVector() || (VT.is256BitVector() && !Subtarget.hasEVEX512()))
+ return SDValue();
+ MVT ElementVT = VT.getVectorElementType();
+ if (ElementVT != MVT::i8 && ElementVT != MVT::i16)
+ return SDValue();
+ SDValue V1 = N.getOperand(0);
+ SDValue V2 = N.getOperand(2);
+ if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+ V1.getConstantOperandVal(1) == 0 &&
+ V2.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
+ V2.getConstantOperandVal(1) == VT.getVectorNumElements() &&
+ V1.getOperand(0) == V2.getOperand(0)) {
+ MVT NVT = VT.getDoubleNumVectorElementsVT();
+ SDValue Mask =
+ DAG.getNode(ISD::INSERT_SUBVECTOR, DL, NVT, DAG.getUNDEF(NVT),
+ N.getOperand(1), DAG.getIntPtrConstant(0, DL));
+ return DAG.getNode(
+ ISD::EXTRACT_SUBVECTOR, DL, VT,
+ DAG.getNode(X86ISD::VPERMV, DL, NVT, Mask, V1.getOperand(0)),
+ DAG.getIntPtrConstant(0, DL));
+ }
+
+ return SDValue();
+ }
default:
return SDValue();
}
diff --git a/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll b/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
index 17ac3e2e6364d..73edceb3c3ede 100644
--- a/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
+++ b/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
@@ -393,9 +393,8 @@ define <16 x i16> @test_masked_z_32xi16_to_16xi16_perm_mask2(<32 x i16> %vec, <1
define <16 x i16> @test_32xi16_to_16xi16_perm_mask3(<32 x i16> %vec) {
; CHECK-LABEL: test_32xi16_to_16xi16_perm_mask3:
; CHECK: # %bb.0:
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm1
-; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm2 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
-; CHECK-NEXT: vpermt2w %ymm1, %ymm2, %ymm0
+; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm1 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
+; CHECK-NEXT: vpermw %zmm0, %zmm1, %zmm0
; CHECK-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
; CHECK-NEXT: retq
%res = shufflevector <32 x i16> %vec, <32 x i16> undef, <16 x i32> <i32 1, i32 0, i32 30, i32 5, i32 3, i32 6, i32 25, i32 29, i32 0, i32 13, i32 3, i32 8, i32 7, i32 20, i32 11, i32 5>
@@ -404,11 +403,10 @@ define <16 x i16> @test_32xi16_to_16xi16_perm_mask3(<32 x i16> %vec) {
define <16 x i16> @test_masked_32xi16_to_16xi16_perm_mask3(<32 x i16> %vec, <16 x i16> %vec2, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_32xi16_to_16xi16_perm_mask3:
; CHECK: # %bb.0:
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm3
-; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm4 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
-; CHECK-NEXT: vpermi2w %ymm3, %ymm0, %ymm4
+; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm3 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
+; CHECK-NEXT: vpermw %zmm0, %zmm3, %zmm0
; CHECK-NEXT: vptestnmw %ymm2, %ymm2, %k1
-; CHECK-NEXT: vpblendmw %ymm4, %ymm1, %ymm0 {%k1}
+; CHECK-NEXT: vpblendmw %ymm0, %ymm1, %ymm0 {%k1}
; CHECK-NEXT: retq
%shuf = shufflevector <32 x i16> %vec, <32 x i16> undef, <16 x i32> <i32 1, i32 0, i32 30, i32 5, i32 3, i32 6, i32 25, i32 29, i32 0, i32 13, i32 3, i32 8, i32 7, i32 20, i32 11, i32 5>
%cmp = icmp eq <16 x i16> %mask, zeroinitializer
@@ -419,10 +417,9 @@ define <16 x i16> @test_masked_32xi16_to_16xi16_perm_mask3(<32 x i16> %vec, <16
define <16 x i16> @test_masked_z_32xi16_to_16xi16_perm_mask3(<32 x i16> %vec, <16 x i16> %mask) {
; CHECK-LABEL: test_masked_z_32xi16_to_16xi16_perm_mask3:
; CHECK: # %bb.0:
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm3 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
+; CHECK-NEXT: vpmovsxbw {{.*#+}} ymm2 = [1,0,30,5,3,6,25,29,0,13,3,8,7,20,11,5]
; CHECK-NEXT: vptestnmw %ymm1, %ymm1, %k1
-; CHECK-NEXT: vpermt2w %ymm2, %ymm3, %ymm0 {%k1} {z}
+; CHECK-NEXT: vpermw %zmm0, %zmm2, %zmm0 {%k1} {z}
; CHECK-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0
; CHECK-NEXT: retq
%shuf = shufflevector <32 x i16> %vec, <32 x i16> undef, <16 x i32> <i32 1, i32 0, i32 30, i32 5, i32 3, i32 6, i32 25, i32 29, i32 0, i32 13, i32 3, i32 8, i32 7, i32 20, i32 11, i32 5>
@@ -477,8 +474,7 @@ define <8 x i16> @test_masked_32xi16_to_8xi16_perm_mask1(<32 x i16> %vec, <8 x i
; CHECK-LABEL: test_masked_32xi16_to_8xi16_perm_mask1:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm3 = [1,21,27,10,8,19,14,5]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm4
-; CHECK-NEXT: vpermt2w %ymm4, %ymm3, %ymm0
+; CHECK-NEXT: vpermw %zmm0, %zmm3, %zmm0
; CHECK-NEXT: vptestnmw %xmm2, %xmm2, %k1
; CHECK-NEXT: vpblendmw %xmm0, %xmm1, %xmm0 {%k1}
; CHECK-NEXT: vzeroupper
@@ -493,9 +489,8 @@ define <8 x i16> @test_masked_z_32xi16_to_8xi16_perm_mask1(<32 x i16> %vec, <8 x
; CHECK-LABEL: test_masked_z_32xi16_to_8xi16_perm_mask1:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm2 = [1,21,27,10,8,19,14,5]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm3
; CHECK-NEXT: vptestnmw %xmm1, %xmm1, %k1
-; CHECK-NEXT: vpermt2w %ymm3, %ymm2, %ymm0 {%k1} {z}
+; CHECK-NEXT: vpermw %zmm0, %zmm2, %zmm0 {%k1} {z}
; CHECK-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
@@ -508,8 +503,7 @@ define <8 x i16> @test_masked_32xi16_to_8xi16_perm_mask2(<32 x i16> %vec, <8 x i
; CHECK-LABEL: test_masked_32xi16_to_8xi16_perm_mask2:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm3 = [15,13,18,16,9,11,26,8]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm4
-; CHECK-NEXT: vpermt2w %ymm4, %ymm3, %ymm0
+; CHECK-NEXT: vpermw %zmm0, %zmm3, %zmm0
; CHECK-NEXT: vptestnmw %xmm2, %xmm2, %k1
; CHECK-NEXT: vpblendmw %xmm0, %xmm1, %xmm0 {%k1}
; CHECK-NEXT: vzeroupper
@@ -524,9 +518,8 @@ define <8 x i16> @test_masked_z_32xi16_to_8xi16_perm_mask2(<32 x i16> %vec, <8 x
; CHECK-LABEL: test_masked_z_32xi16_to_8xi16_perm_mask2:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm2 = [15,13,18,16,9,11,26,8]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm3
; CHECK-NEXT: vptestnmw %xmm1, %xmm1, %k1
-; CHECK-NEXT: vpermt2w %ymm3, %ymm2, %ymm0 {%k1} {z}
+; CHECK-NEXT: vpermw %zmm0, %zmm2, %zmm0 {%k1} {z}
; CHECK-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
@@ -539,8 +532,7 @@ define <8 x i16> @test_32xi16_to_8xi16_perm_mask3(<32 x i16> %vec) {
; CHECK-LABEL: test_32xi16_to_8xi16_perm_mask3:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm1 = [17,0,23,10,1,8,7,30]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; CHECK-NEXT: vpermt2w %ymm2, %ymm1, %ymm0
+; CHECK-NEXT: vpermw %zmm0, %zmm1, %zmm0
; CHECK-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
@@ -551,8 +543,7 @@ define <8 x i16> @test_masked_32xi16_to_8xi16_perm_mask3(<32 x i16> %vec, <8 x i
; CHECK-LABEL: test_masked_32xi16_to_8xi16_perm_mask3:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm3 = [17,0,23,10,1,8,7,30]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm4
-; CHECK-NEXT: vpermt2w %ymm4, %ymm3, %ymm0
+; CHECK-NEXT: vpermw %zmm0, %zmm3, %zmm0
; CHECK-NEXT: vptestnmw %xmm2, %xmm2, %k1
; CHECK-NEXT: vpblendmw %xmm0, %xmm1, %xmm0 {%k1}
; CHECK-NEXT: vzeroupper
@@ -567,9 +558,8 @@ define <8 x i16> @test_masked_z_32xi16_to_8xi16_perm_mask3(<32 x i16> %vec, <8 x
; CHECK-LABEL: test_masked_z_32xi16_to_8xi16_perm_mask3:
; CHECK: # %bb.0:
; CHECK-NEXT: vpmovsxbw {{.*#+}} xmm2 = [17,0,23,10,1,8,7,30]
-; CHECK-NEXT: vextracti64x4 $1, %zmm0, %ymm3
; CHECK-NEXT: vptestnmw %xmm1, %xmm1, %k1
-; CHECK-NEXT: vpermt2w %ymm3, %ymm2, %ymm0 {%k1} {z}
+; CHECK-NEXT: vpermw %zmm0, %zmm2, %zmm0 {%k1} {z}
; CHECK-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; CHECK-NEXT: vzeroupper
; CHECK-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
index a481aaef4257d..95e249984e184 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-512.ll
@@ -348,8 +348,7 @@ define <16 x i8> @trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_
; AVX512VBMIVL-LABEL: trunc_shuffle_v64i8_01_05_09_13_17_21_25_29_33_37_41_45_49_53_57_62:
; AVX512VBMIVL: # %bb.0:
; AVX512VBMIVL-NEXT: vmovdqa {{.*#+}} xmm1 = [1,5,9,13,17,21,25,29,33,37,41,45,49,53,57,62]
-; AVX512VBMIVL-NEXT: vextracti64x4 $1, %zmm0, %ymm2
-; AVX512VBMIVL-NEXT: vpermt2b %ymm2, %ymm1, %ymm0
+; AVX512VBMIVL-NEXT: vpermb %zmm0, %zmm1, %zmm0
; AVX512VBMIVL-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
; AVX512VBMIVL-NEXT: vzeroupper
; AVX512VBMIVL-NEXT: retq
More information about the llvm-commits
mailing list