[DAG] Commute shuffle(splat(A,u), shuffle(C,D)) -> shuffle'(shuffle(C,D), splat(A,u))

We only merge shuffles if the inner (LHS) shuffle is a non-splat, so commute these shuffles to improve merging of multiple shuffles.

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -20799,26 +20799,35 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
     }
   }
 
-  // Canonicalize shuffles according to rules:
-  //  shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
-  //  shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
-  //  shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
-  if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
-      N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
-      TLI.isTypeLegal(VT)) {
-    // The incoming shuffle must be of the same type as the result of the
-    // current shuffle.
-    assert(N1->getOperand(0).getValueType() == VT &&
-           "Shuffle types don't match");
-
-    SDValue SV0 = N1->getOperand(0);
-    SDValue SV1 = N1->getOperand(1);
-    bool HasSameOp0 = N0 == SV0;
-    bool IsSV1Undef = SV1.isUndef();
-    if (HasSameOp0 || IsSV1Undef || N0 == SV1)
-      // Commute the operands of this shuffle so that next rule
-      // will trigger.
+  if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
+    // Canonicalize shuffles according to rules:
+    //  shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
+    //  shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
+    //  shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
+    if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
+        N0.getOpcode() != ISD::VECTOR_SHUFFLE) {
+      // The incoming shuffle must be of the same type as the result of the
+      // current shuffle.
+      assert(N1->getOperand(0).getValueType() == VT &&
+             "Shuffle types don't match");
+
+      SDValue SV0 = N1->getOperand(0);
+      SDValue SV1 = N1->getOperand(1);
+      bool HasSameOp0 = N0 == SV0;
+      bool IsSV1Undef = SV1.isUndef();
+      if (HasSameOp0 || IsSV1Undef || N0 == SV1)
+        // Commute the operands of this shuffle so merging below will trigger.
+        return DAG.getCommutedVectorShuffle(*SVN);
+    }
+
+    // Canonicalize splat shuffles to the RHS to improve merging below.
+    //  shuffle(splat(A,u), shuffle(C,D)) -> shuffle'(shuffle(C,D), splat(A,u))
+    if (N0.getOpcode() == ISD::VECTOR_SHUFFLE &&
+        N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
+        cast<ShuffleVectorSDNode>(N0)->isSplat() &&
+        !cast<ShuffleVectorSDNode>(N1)->isSplat()) {
       return DAG.getCommutedVectorShuffle(*SVN);
+    }
   }
 
   // Compute the combined shuffle mask for a shuffle with SV0 as the first

llvm/test/CodeGen/X86/haddsub-undef.ll

@@ -583,17 +583,11 @@ define <4 x float> @add_ps_016(<4 x float> %0, <4 x float> %1) {
 ; SSE-NEXT:    movaps %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-SLOW-LABEL: add_ps_016:
-; AVX-SLOW:       # %bb.0:
-; AVX-SLOW-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
-; AVX-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,0,3,3]
-; AVX-SLOW-NEXT:    retq
-;
-; AVX-FAST-LABEL: add_ps_016:
-; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
-; AVX-FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,0,3,1]
-; AVX-FAST-NEXT:    retq
+; AVX-LABEL: add_ps_016:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,0,3,3]
+; AVX-NEXT:    retq
   %3 = shufflevector <4 x float> %1, <4 x float> %0, <2 x i32> <i32 0, i32 6>
   %4 = shufflevector <4 x float> %1, <4 x float> %0, <2 x i32> <i32 1, i32 7>
   %5 = fadd <2 x float> %3, %4