[X86] combineConcatVectorOps - add 256-bit concat(shuffle(),shuffle()) handling

Improve IsConcatFree detection to handle splat vector-loads (which can be folded as X86ISD::SUBV_BROADCAST_LOAD).

Fixes #114959

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -56933,6 +56933,11 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
       bool AllConstants = true;
       bool AllSubs = true;
       unsigned VecSize = VT.getSizeInBits();
+      SDValue BC0 = peekThroughBitcasts(SubOps[0].getOperand(Op));
+      if (isa<LoadSDNode>(BC0) && all_of(SubOps, [&](SDValue SubOp) {
+            return BC0 == peekThroughBitcasts(SubOp.getOperand(Op));
+          }))
+        return true;
       for (unsigned I = 0, E = SubOps.size(); I != E; ++I) {
         SDValue BC = peekThroughBitcasts(SubOps[I].getOperand(Op));
         unsigned SubSize = BC.getValueSizeInBits();
@@ -56947,6 +56952,26 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
     };
 
     switch (Op0.getOpcode()) {
+    case ISD::VECTOR_SHUFFLE: {
+      if (NumOps == 2 && VT.is256BitVector() &&
+          (EltSizeInBits >= 32 || Subtarget.hasInt256()) &&
+          (IsConcatFree(VT, Ops, 0) || IsConcatFree(VT, Ops, 1))) {
+        int NumSubElts = Op0.getValueType().getVectorNumElements();
+        SmallVector<int> NewMask;
+        for (int M : cast<ShuffleVectorSDNode>(Ops[0])->getMask()) {
+          M = M >= NumSubElts ? M + NumSubElts : M;
+          NewMask.push_back(M);
+        }
+        for (int M : cast<ShuffleVectorSDNode>(Ops[1])->getMask()) {
+          if (0 <= M)
+            M = (M >= NumSubElts ? M + NumSubElts : M) + NumSubElts;
+          NewMask.push_back(M);
+        }
+        return DAG.getVectorShuffle(VT, DL, ConcatSubOperand(VT, Ops, 0),
+                                    ConcatSubOperand(VT, Ops, 1), NewMask);
+      }
+      break;
+    }
     case X86ISD::VBROADCAST: {
       if (!IsSplat && llvm::all_of(Ops, [](SDValue Op) {
             return Op.getOperand(0).getValueType().is128BitVector();

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

@@ -1810,40 +1810,12 @@ define <4 x double> @broadcast_v4f64_0000_from_v2i64(<2 x i64> %a0) {
 
 ; PR114959
 define <4 x double> @concat_v4f64_0213_broadcasts(ptr %src) {
-; AVX1OR2-LABEL: concat_v4f64_0213_broadcasts:
-; AVX1OR2:       # %bb.0:
-; AVX1OR2-NEXT:    vmovups (%rdi), %xmm0
-; AVX1OR2-NEXT:    vmovups 32(%rdi), %xmm1
-; AVX1OR2-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
-; AVX1OR2-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; AVX1OR2-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
-; AVX1OR2-NEXT:    retq
-;
-; AVX512VL-SLOW-LABEL: concat_v4f64_0213_broadcasts:
-; AVX512VL-SLOW:       # %bb.0:
-; AVX512VL-SLOW-NEXT:    vmovups (%rdi), %xmm0
-; AVX512VL-SLOW-NEXT:    vmovups 32(%rdi), %xmm1
-; AVX512VL-SLOW-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
-; AVX512VL-SLOW-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; AVX512VL-SLOW-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
-; AVX512VL-SLOW-NEXT:    retq
-;
-; AVX512VL-FAST-ALL-LABEL: concat_v4f64_0213_broadcasts:
-; AVX512VL-FAST-ALL:       # %bb.0:
-; AVX512VL-FAST-ALL-NEXT:    vmovupd (%rdi), %xmm1
-; AVX512VL-FAST-ALL-NEXT:    vmovupd 32(%rdi), %xmm2
-; AVX512VL-FAST-ALL-NEXT:    vmovapd {{.*#+}} ymm0 = [0,4,1,5]
-; AVX512VL-FAST-ALL-NEXT:    vpermi2pd %ymm2, %ymm1, %ymm0
-; AVX512VL-FAST-ALL-NEXT:    retq
-;
-; AVX512VL-FAST-PERLANE-LABEL: concat_v4f64_0213_broadcasts:
-; AVX512VL-FAST-PERLANE:       # %bb.0:
-; AVX512VL-FAST-PERLANE-NEXT:    vmovups (%rdi), %xmm0
-; AVX512VL-FAST-PERLANE-NEXT:    vmovups 32(%rdi), %xmm1
-; AVX512VL-FAST-PERLANE-NEXT:    vmovlhps {{.*#+}} xmm2 = xmm0[0],xmm1[0]
-; AVX512VL-FAST-PERLANE-NEXT:    vunpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; AVX512VL-FAST-PERLANE-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
-; AVX512VL-FAST-PERLANE-NEXT:    retq
+; ALL-LABEL: concat_v4f64_0213_broadcasts:
+; ALL:       # %bb.0:
+; ALL-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
+; ALL-NEXT:    vbroadcastf128 {{.*#+}} ymm1 = mem[0,1,0,1]
+; ALL-NEXT:    vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[0],ymm1[3],ymm0[3]
+; ALL-NEXT:    retq
   %src.hi = getelementptr inbounds i8, ptr %src, i64 32
   %lo = load <2 x double>, ptr %src, align 1
   %hi = load <2 x double>, ptr %src.hi, align 1