[X86] combineBlendOfPermutes - don't introduce lane-crossing permutes without AVX2 support.

Fixes llvm#91433

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -40078,10 +40078,10 @@ static SDValue combineCommutableSHUFP(SDValue N, MVT VT, const SDLoc &DL,
 
 // Attempt to fold BLEND(PERMUTE(X),PERMUTE(Y)) -> PERMUTE(BLEND(X,Y))
 // iff we don't demand the same element index for both X and Y.
-static SDValue combineBlendOfPermutes(MVT VT, SDValue N0, SDValue N1,
-                                      ArrayRef<int> BlendMask,
-                                      const APInt &DemandedElts,
-                                      SelectionDAG &DAG, const SDLoc &DL) {
+static SDValue
+combineBlendOfPermutes(MVT VT, SDValue N0, SDValue N1, ArrayRef<int> BlendMask,
+                       const APInt &DemandedElts, SelectionDAG &DAG,
+                       const X86Subtarget &Subtarget, const SDLoc &DL) {
   assert(isBlendOrUndef(BlendMask) && "Blend shuffle expected");
   if (!N0.hasOneUse() || !N1.hasOneUse())
     return SDValue();
@@ -40156,6 +40156,11 @@ static SDValue combineBlendOfPermutes(MVT VT, SDValue N0, SDValue N1,
       return SDValue();
   }
 
+  // Don't introduce lane-crossing permutes without AVX2.
+  if (VT.is256BitVector() && !Subtarget.hasAVX2() &&
+      isLaneCrossingShuffleMask(128, VT.getScalarSizeInBits(), NewPermuteMask))
+    return SDValue();
+
   SDValue NewBlend =
       DAG.getVectorShuffle(VT, DL, DAG.getBitcast(VT, Ops0[0]),
                            DAG.getBitcast(VT, Ops1[0]), NewBlendMask);
@@ -41918,9 +41923,9 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
   case X86ISD::BLENDI: {
     SmallVector<int, 16> BlendMask;
     DecodeBLENDMask(NumElts, Op.getConstantOperandVal(2), BlendMask);
-    if (SDValue R = combineBlendOfPermutes(VT.getSimpleVT(), Op.getOperand(0),
-                                           Op.getOperand(1), BlendMask,
-                                           DemandedElts, TLO.DAG, SDLoc(Op)))
+    if (SDValue R = combineBlendOfPermutes(
+            VT.getSimpleVT(), Op.getOperand(0), Op.getOperand(1), BlendMask,
+            DemandedElts, TLO.DAG, Subtarget, SDLoc(Op)))
       return TLO.CombineTo(Op, R);
     break;
   }

llvm/test/CodeGen/X86/vector-shuffle-256-v4.ll

@@ -699,6 +699,28 @@ define <4 x double> @shuffle_v4f64_0437(<4 x double> %a, <4 x double> %b) {
   ret <4 x double> %shuffle
 }
 
+; PR91433
+define <4 x double> @shuffle_v4f64_2303(<4 x double> %a) {
+; AVX1-LABEL: shuffle_v4f64_2303:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,2,3]
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2,3],ymm0[4,5],ymm1[6,7]
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: shuffle_v4f64_2303:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,3]
+; AVX2-NEXT:    retq
+;
+; AVX512VL-LABEL: shuffle_v4f64_2303:
+; AVX512VL:       # %bb.0:
+; AVX512VL-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,3]
+; AVX512VL-NEXT:    retq
+  %shuffle = shufflevector <4 x double> %a, <4 x double> poison, <4 x i32> <i32 2, i32 3, i32 0, i32 3>
+  ret <4 x double> %shuffle
+}
+
 define <4 x double> @shuffle_v4f64_0z3z(<4 x double> %a, <4 x double> %b) {
 ; ALL-LABEL: shuffle_v4f64_0z3z:
 ; ALL:       # %bb.0:

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

@@ -308,8 +308,9 @@ define <4 x float> @combine_vpermilvar_4f32_as_insertps(<4 x float> %a0) {
 define <8 x i32> @combine_blend_of_permutes_v8i32(<4 x i64> %a0, <4 x i64> %a1) {
 ; AVX1-LABEL: combine_blend_of_permutes_v8i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7]
 ; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
+; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[3,4],ymm0[5,6],ymm1[7]
 ; AVX1-NEXT:    ret{{[l|q]}}
 ;
 ; AVX2-LABEL: combine_blend_of_permutes_v8i32: