[X86][AVX] Handle vperm2x128 shuffling of a subvector splat.

We already handle "vperm2x128 (ins ?, X, C1), (ins ?, X, C1), 0x31" for shuffling of the upper subvectors, but we weren't dealing with the case when we were splatting the upper subvector from a single source.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -37324,6 +37324,14 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
     SDValue Ins1 = peekThroughBitcasts(N.getOperand(1));
     unsigned Imm = N.getConstantOperandVal(2);
 
+    // Handle subvector splat by tweaking values to match binary concat.
+    // vperm2x128 (ins ?, X, C1), undef, 0x11 ->
+    // vperm2x128 (ins ?, X, C1), (ins ?, X, C1), 0x31 -> concat X, X
+    if (Imm == 0x11 && Ins1.isUndef()) {
+      Imm = 0x31;
+      Ins1 = Ins0;
+    }
+
     if (!(Imm == 0x31 &&
           Ins0.getOpcode() == ISD::INSERT_SUBVECTOR &&
           Ins1.getOpcode() == ISD::INSERT_SUBVECTOR &&

llvm/test/CodeGen/X86/avx-vperm2x128.ll

@@ -130,7 +130,6 @@ define <32 x i8> @shuffle_v32i8_2323_domain(<32 x i8> %a, <32 x i8> %b) nounwind
 ; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
 ; AVX1-NEXT:    vpsubb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: shuffle_v32i8_2323_domain:

llvm/test/CodeGen/X86/vector-shuffle-256-v8.ll

@@ -3098,14 +3098,13 @@ entry:
 define <8 x i32> @add_v8i32_02468ACE_13579BDF(<8 x i32> %a, <8 x i32> %b) {
 ; AVX1-LABEL: add_v8i32_02468ACE_13579BDF:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vphaddd %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm2, %ymm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm2[0],ymm0[0],ymm2[3],ymm0[3]
+; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[0],ymm2[0],ymm0[3],ymm2[3]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2OR512VL-LABEL: add_v8i32_02468ACE_13579BDF:
@@ -3123,14 +3122,13 @@ entry:
 define <8 x i32> @add_v8i32_8ACE0246_9BDF1357(<8 x i32> %a, <8 x i32> %b) {
 ; AVX1-LABEL: add_v8i32_8ACE0246_9BDF1357:
 ; AVX1:       # %bb.0: # %entry
-; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vphaddd %xmm2, %xmm3, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm2, %ymm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX1-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,2,3]
-; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm2[1],ymm0[1],ymm2[2],ymm0[2]
+; AVX1-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1],ymm2[1],ymm0[2],ymm2[2]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2OR512VL-LABEL: add_v8i32_8ACE0246_9BDF1357: