[X86][AVX] canonicalizeLaneShuffleWithRepeatedOps - don't merge VPERMILPD ops with different low/high masks.

Unlike VPERMILPS, VPERMILPD can have non-repeating masks in each 128-bit subvector, we weren't accounting for this when folding vperm2f128(vpermilpd(x,c),vpermilpd(y,c)) -> vpermilpd(vperm2f128(x,y),c).

I'm intending to add support for this but wanted to get a minimal fix in first for merging into 12.xx.

Fixes PR48908

(cherry picked from commit 6663330bc8c84a75ea092272297b557bfc310380)

2 files changed, 30 insertions, 19 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 0dd20235aa3c..6b816c710f98 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36916,11 +36916,18 @@ static SDValue canonicalizeLaneShuffleWithRepeatedOps(SDValue V,
     Res = DAG.getNode(SrcOpc0, DL, SrcVT0, DAG.getBitcast(SrcVT0, Res));
     return DAG.getBitcast(VT, Res);
   }
+  case X86ISD::VPERMILPI:
+    // TODO: Handle v4f64 permutes with different low/high lane masks.
+    if (SrcVT0 == MVT::v4f64) {
+      uint64_t Mask = Src0.getConstantOperandVal(1);
+      if ((Mask & 0x3) != ((Mask >> 2) & 0x3))
+        break;
+    }
+    LLVM_FALLTHROUGH;
   case X86ISD::VSHLI:
   case X86ISD::VSRLI:
   case X86ISD::VSRAI:
   case X86ISD::PSHUFD:
-  case X86ISD::VPERMILPI:
     if (Src1.isUndef() || Src0.getOperand(1) == Src1.getOperand(1)) {
       SDValue LHS = DAG.getBitcast(VT, Src0.getOperand(0));
       SDValue RHS =
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
index 3da83b25d363..1a1153d0e886 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll
@@ -442,16 +442,18 @@ define void @PR48908(<4 x double> %v0, <4 x double> %v1, <4 x double> %v2, <4 x
 ; X86-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X86-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %edx
-; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm1[2,3],ymm2[0,1]
-; X86-AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm4
-; X86-AVX1-NEXT:    vshufpd {{.*#+}} ymm3 = ymm4[1],ymm3[0],ymm4[2],ymm3[3]
-; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm1[0,1],ymm0[0,1]
+; X86-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm3
+; X86-AVX1-NEXT:    vpermilpd {{.*#+}} ymm3 = ymm3[0,1,2,2]
+; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm1[2,3],ymm2[0,1]
+; X86-AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm5
+; X86-AVX1-NEXT:    vshufpd {{.*#+}} ymm4 = ymm5[1],ymm4[0],ymm5[2],ymm4[3]
 ; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm0[0,1],ymm2[0,1]
-; X86-AVX1-NEXT:    vshufpd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[2]
-; X86-AVX1-NEXT:    vmovapd %ymm4, (%edx)
-; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm3[2,3,0,1]
-; X86-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm3[0,1],ymm0[2],ymm3[3]
-; X86-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[3]
+; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm3[2,3,0,1]
+; X86-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3]
+; X86-AVX1-NEXT:    vmovapd %ymm3, (%edx)
+; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm4[2,3,0,1]
+; X86-AVX1-NEXT:    vblendpd {{.*#+}} ymm4 = ymm4[0,1],ymm0[2],ymm4[3]
+; X86-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3]
 ; X86-AVX1-NEXT:    vmovapd %ymm3, (%ecx)
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
@@ -513,16 +515,18 @@ define void @PR48908(<4 x double> %v0, <4 x double> %v1, <4 x double> %v2, <4 x
 ;
 ; X64-AVX1-LABEL: PR48908:
 ; X64-AVX1:       # %bb.0:
-; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm1[2,3],ymm2[0,1]
-; X64-AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm4
-; X64-AVX1-NEXT:    vshufpd {{.*#+}} ymm3 = ymm4[1],ymm3[0],ymm4[2],ymm3[3]
-; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm1[0,1],ymm0[0,1]
+; X64-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm3
+; X64-AVX1-NEXT:    vpermilpd {{.*#+}} ymm3 = ymm3[0,1,2,2]
+; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm1[2,3],ymm2[0,1]
+; X64-AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm5
+; X64-AVX1-NEXT:    vshufpd {{.*#+}} ymm4 = ymm5[1],ymm4[0],ymm5[2],ymm4[3]
 ; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm5 = ymm0[0,1],ymm2[0,1]
-; X64-AVX1-NEXT:    vshufpd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[2]
-; X64-AVX1-NEXT:    vmovapd %ymm4, (%rdi)
-; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm3[2,3,0,1]
-; X64-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm3[0,1],ymm0[2],ymm3[3]
-; X64-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm3[0],ymm4[1],ymm3[2],ymm4[3]
+; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm3[2,3,0,1]
+; X64-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm5[0],ymm3[1],ymm5[2],ymm3[3]
+; X64-AVX1-NEXT:    vmovapd %ymm3, (%rdi)
+; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm4[2,3,0,1]
+; X64-AVX1-NEXT:    vblendpd {{.*#+}} ymm4 = ymm4[0,1],ymm0[2],ymm4[3]
+; X64-AVX1-NEXT:    vblendpd {{.*#+}} ymm3 = ymm4[0],ymm3[1],ymm4[2],ymm3[3]
 ; X64-AVX1-NEXT:    vmovapd %ymm3, (%rsi)
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]