[X86] combineTargetShuffle - canonicalize vperm2x128(x,x)/vperm2x128(undef,x) -> vperm2x128(x,undef)

Improves fold matching for future patches.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -42649,7 +42649,15 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
   case X86ISD::VPERM2X128: {
     SDValue LHS = N->getOperand(0);
     SDValue RHS = N->getOperand(1);
-    unsigned Imm = N.getConstantOperandVal(2);
+    unsigned Imm = N.getConstantOperandVal(2) & 255;
+
+    // Canonicalize unary/repeated operands to LHS.
+    if (LHS.isUndef() && !RHS.isUndef())
+      return DAG.getNode(X86ISD::VPERM2X128, DL, VT, RHS, LHS,
+                         DAG.getTargetConstant(Imm ^ 0x22, DL, MVT::i8));
+    if (LHS == RHS)
+      return DAG.getNode(X86ISD::VPERM2X128, DL, VT, LHS, DAG.getUNDEF(VT),
+                         DAG.getTargetConstant(Imm & ~0x22, DL, MVT::i8));
 
     // Fold vperm2x128(bitcast(x),bitcast(y),c) -> bitcast(vperm2x128(x,y,c)).
     if (LHS.getOpcode() == ISD::BITCAST &&

llvm/test/CodeGen/X86/matrix-multiply.ll

@@ -119,9 +119,9 @@ define <4 x double> @test_mul2x2_f64(<4 x double> %a0, <4 x double> %a1) nounwin
 ;
 ; AVX1-LABEL: test_mul2x2_f64:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vshufpd {{.*#+}} ymm2 = ymm1[1,1,3,3]
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm0[2,3,2,3]
-; AVX1-NEXT:    vmulpd %ymm2, %ymm3, %ymm2
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm2 = ymm0[2,3,2,3]
+; AVX1-NEXT:    vshufpd {{.*#+}} ymm3 = ymm1[1,1,3,3]
+; AVX1-NEXT:    vmulpd %ymm3, %ymm2, %ymm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    vmovddup {{.*#+}} ymm1 = ymm1[0,0,2,2]
 ; AVX1-NEXT:    vmulpd %ymm1, %ymm0, %ymm0