[X86][AVX] canonicalizeLaneShuffleWithRepeatedOps - merge VPERMILPD ops with different low/high masks.

Now that PR48908 has been dealt with, we can handle v4f64 permute cases by extracting the low/high lane VPERMILPD masks and creating a new mask based on which lanes are referenced by the VPERM2F128 mask.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -36946,11 +36946,27 @@ static SDValue canonicalizeLaneShuffleWithRepeatedOps(SDValue V,
     return DAG.getBitcast(VT, Res);
   }
   case X86ISD::VPERMILPI:
-    // TODO: Handle v4f64 permutes with different low/high lane masks.
+    // Handle v4f64 permutes with different low/high lane masks by permuting
+    // the permute mask on a lane-by-lane basis.
     if (SrcVT0 == MVT::v4f64) {
-      uint64_t Mask = Src0.getConstantOperandVal(1);
-      if ((Mask & 0x3) != ((Mask >> 2) & 0x3))
-        break;
+      if (Src1.isUndef() || Src0.getOperand(1) == Src1.getOperand(1)) {
+        uint64_t LaneMask = V.getConstantOperandVal(2);
+        uint64_t Mask = Src0.getConstantOperandVal(1);
+        uint64_t LoMask = Mask & 0x3;
+        uint64_t HiMask = (Mask >> 2) & 0x3;
+        uint64_t NewMask = 0;
+        NewMask |= ((LaneMask & 0x02) ? HiMask : LoMask);
+        NewMask |= ((LaneMask & 0x02) ? HiMask : LoMask) << 2;
+        SDValue LHS = Src0.getOperand(0);
+        SDValue RHS =
+            Src1.isUndef() ? DAG.getUNDEF(SrcVT0) : Src1.getOperand(0);
+        SDValue Res = DAG.getNode(X86ISD::VPERM2X128, DL, SrcVT0, LHS, RHS,
+                                  V.getOperand(2));
+        Res = DAG.getNode(SrcOpc0, DL, SrcVT0, Res,
+                          DAG.getTargetConstant(NewMask, DL, MVT::i8));
+        return DAG.getBitcast(VT, Res);
+      }
+      break;
     }
     LLVM_FALLTHROUGH;
   case X86ISD::VSHLI:

llvm/test/CodeGen/X86/vector-shuffle-combining-avx.ll

@@ -442,18 +442,16 @@ 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:    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 {{.*#+}} 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:    vperm2f128 {{.*#+}} ymm5 = ymm0[0,1],ymm2[0,1]
-; 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:    vblendpd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3]
+; 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:    vmovapd %ymm3, (%ecx)
 ; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X86-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]
@@ -515,18 +513,16 @@ 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:    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 {{.*#+}} 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:    vperm2f128 {{.*#+}} ymm5 = ymm0[0,1],ymm2[0,1]
-; 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:    vblendpd {{.*#+}} ymm4 = ymm5[0],ymm4[1],ymm5[2],ymm4[3]
+; 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:    vmovapd %ymm3, (%rsi)
 ; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; X64-AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm2[2,3],ymm1[2,3]