[llvm] 70bd80d - [X86] combineTargetShuffle - commute VPERMV3 shuffles so any load is on the RHS

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2024-11-23T15:27:10Z
New Revision: 70bd80dc51b62453210f6203c31ea826dd0675c2

URL: https://github.com/llvm/llvm-project/commit/70bd80dc51b62453210f6203c31ea826dd0675c2
DIFF: https://github.com/llvm/llvm-project/commit/70bd80dc51b62453210f6203c31ea826dd0675c2.diff

LOG: [X86] combineTargetShuffle - commute VPERMV3 shuffles so any load is on the RHS

This helps ensure we lower to VPERMI2/T2 instructions that we can commute the index arg to VPERMT2/I2.

Similar to 1e31a4529244ead9f12abed524f33a48515abee2 to handle cases where the one use load appears after further folding (keep the lowerShuffleWithPERMV version as this can handle the non-VLX widening case as well).

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
    llvm/test/CodeGen/X86/vector-shuffle-v1.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ff2b9199384db4..9048d1d83f1874 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -42252,6 +42252,17 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
                            DAG.getIntPtrConstant(0, DL));
       }
     }
+    SmallVector<SDValue, 2> Ops;
+    SmallVector<int, 32> Mask;
+    if (isShuffleFoldableLoad(N.getOperand(0)) &&
+        !isShuffleFoldableLoad(N.getOperand(2)) &&
+        getTargetShuffleMask(N, /*AllowSentinelZero=*/false, Ops, Mask)) {
+      ShuffleVectorSDNode::commuteMask(Mask);
+      SDValue NewMask = getConstVector(
+          Mask, N.getOperand(1).getSimpleValueType(), DAG, DL, /*IsMask=*/true);
+      return DAG.getNode(X86ISD::VPERMV3, DL, VT, N.getOperand(2), NewMask,
+                         N.getOperand(0));
+    }
     return SDValue();
   }
   default:

diff  --git a/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll b/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
index d07923faa4309f..5078130f180779 100644
--- a/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
+++ b/llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll
@@ -3514,12 +3514,11 @@ define <8 x float> @test_masked_16xfloat_to_8xfloat_perm_mem_mask2(ptr %vp, <8 x
 ; CHECK-FAST-PERLANE-LABEL: test_masked_16xfloat_to_8xfloat_perm_mem_mask2:
 ; CHECK-FAST-PERLANE:       # %bb.0:
 ; CHECK-FAST-PERLANE-NEXT:    vmovaps (%rdi), %xmm2
-; CHECK-FAST-PERLANE-NEXT:    vmovaps 32(%rdi), %ymm3
-; CHECK-FAST-PERLANE-NEXT:    vmovaps {{.*#+}} ymm4 = [9,5,2,3,2,8,8,1]
-; CHECK-FAST-PERLANE-NEXT:    vpermi2ps %ymm2, %ymm3, %ymm4
+; CHECK-FAST-PERLANE-NEXT:    vmovaps {{.*#+}} ymm3 = [1,13,10,11,10,0,0,9]
+; CHECK-FAST-PERLANE-NEXT:    vpermi2ps 32(%rdi), %ymm2, %ymm3
 ; CHECK-FAST-PERLANE-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; CHECK-FAST-PERLANE-NEXT:    vcmpeqps %ymm2, %ymm1, %k1
-; CHECK-FAST-PERLANE-NEXT:    vmovaps %ymm4, %ymm0 {%k1}
+; CHECK-FAST-PERLANE-NEXT:    vmovaps %ymm3, %ymm0 {%k1}
 ; CHECK-FAST-PERLANE-NEXT:    retq
   %vec = load <16 x float>, ptr %vp
   %shuf = shufflevector <16 x float> %vec, <16 x float> undef, <8 x i32> <i32 1, i32 13, i32 10, i32 11, i32 10, i32 0, i32 0, i32 9>
@@ -3542,11 +3541,10 @@ define <8 x float> @test_masked_z_16xfloat_to_8xfloat_perm_mem_mask2(ptr %vp, <8
 ; CHECK-FAST-PERLANE-LABEL: test_masked_z_16xfloat_to_8xfloat_perm_mem_mask2:
 ; CHECK-FAST-PERLANE:       # %bb.0:
 ; CHECK-FAST-PERLANE-NEXT:    vmovaps (%rdi), %xmm2
-; CHECK-FAST-PERLANE-NEXT:    vmovaps 32(%rdi), %ymm3
-; CHECK-FAST-PERLANE-NEXT:    vmovaps {{.*#+}} ymm1 = [9,5,2,3,2,8,8,1]
-; CHECK-FAST-PERLANE-NEXT:    vxorps %xmm4, %xmm4, %xmm4
-; CHECK-FAST-PERLANE-NEXT:    vcmpeqps %ymm4, %ymm0, %k1
-; CHECK-FAST-PERLANE-NEXT:    vpermi2ps %ymm2, %ymm3, %ymm1 {%k1} {z}
+; CHECK-FAST-PERLANE-NEXT:    vmovaps {{.*#+}} ymm1 = [1,13,10,11,10,0,0,9]
+; CHECK-FAST-PERLANE-NEXT:    vxorps %xmm3, %xmm3, %xmm3
+; CHECK-FAST-PERLANE-NEXT:    vcmpeqps %ymm3, %ymm0, %k1
+; CHECK-FAST-PERLANE-NEXT:    vpermi2ps 32(%rdi), %ymm2, %ymm1 {%k1} {z}
 ; CHECK-FAST-PERLANE-NEXT:    vmovaps %ymm1, %ymm0
 ; CHECK-FAST-PERLANE-NEXT:    retq
   %vec = load <16 x float>, ptr %vp

diff  --git a/llvm/test/CodeGen/X86/vector-shuffle-v1.ll b/llvm/test/CodeGen/X86/vector-shuffle-v1.ll
index 6f9b3e94aa68f6..2b89590a0bb419 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-v1.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-v1.ll
@@ -719,10 +719,9 @@ define i8 @shuf8i1__9_6_1_10_3_7_7_1(i8 %a) {
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    kmovw %edi, %k1
 ; AVX512F-NEXT:    vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1
-; AVX512F-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [9,6,1,0,3,7,7,1]
-; AVX512F-NEXT:    vpmovsxbq {{.*#+}} zmm2 = [18446744073709551615,18446744073709551615,0,0,0,0,0,0]
-; AVX512F-NEXT:    vpermt2q %zmm0, %zmm1, %zmm2
-; AVX512F-NEXT:    vptestmq %zmm2, %zmm2, %k0
+; AVX512F-NEXT:    vpmovsxbq {{.*#+}} zmm1 = [1,14,9,8,11,15,15,9]
+; AVX512F-NEXT:    vpermi2q {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
+; AVX512F-NEXT:    vptestmq %zmm1, %zmm1, %k0
 ; AVX512F-NEXT:    kmovw %k0, %eax
 ; AVX512F-NEXT:    # kill: def $al killed $al killed $eax
 ; AVX512F-NEXT:    vzeroupper