[DAGCombiner] Add more rules to fold shuffles.

This patch adds two new rules to the DAGCombiner:
 1.  shuffle (shuffle A, Undef, M0), B, M1 -> shuffle A, B, M2
 2.  shuffle (shuffle A, Undef, M0), A, M1 -> shuffle A, Undef, M2

We only do this if the combined shuffle is legal for the target.

Example:
;;
define <4 x float> @test(<4 x float> %a, <4 x float> %b) {
  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32><i32 6, i32 0, i32 1, i32 7>
  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32><i32 1, i32 2, i32 4, i32 5>
  ret <4 x i32> %2
}
;;

(using llc -mcpu=corei7 -march=x86-64)
Before, the x86 backend generated:
  pshufd $120, %xmm0, %xmm0
  shufps $-108, %xmm0, %xmm1
  movaps %xmm1, %xmm0

Now the x86 backend generates:
  movsd %xmm1, %xmm0



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213069 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -10780,11 +10780,13 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   }
 
   // Try to fold according to rules:
-  //   shuffle(shuffle A, B, M0), B, M1) -> shuffle(A, B, M2)
+  //   shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
-  //   shuffle(shuffle A, B, M0), A, M1) -> shuffle(A, B, M2)
+  //   shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
+  //   shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
+  //   shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
   // Don't try to fold shuffles with illegal type.
   if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
-      TLI.isTypeLegal(VT)) {
+      N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
     ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
 
     // The incoming shuffle must be of the same type as the result of the
@@ -10795,7 +10797,8 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
     SDValue SV0 = OtherSV->getOperand(0);
     SDValue SV1 = OtherSV->getOperand(1);
     bool HasSameOp0 = N1 == SV0;
-    if (!HasSameOp0 && N1 != SV1)
+    bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
+    if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
       // Early exit.
       return SDValue();
 
@@ -10810,18 +10813,25 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
         continue;
       }
 
-      if (Idx < (int)NumElts)
+      if (Idx < (int)NumElts) {
         Idx = OtherSV->getMaskElt(Idx);
-      else
+        if (IsSV1Undef && Idx >= (int) NumElts)
+          Idx = -1;  // Propagate Undef.
+      } else
         Idx = HasSameOp0 ? Idx - NumElts : Idx;
 
       Mask.push_back(Idx);
     }
 
     // Avoid introducing shuffles with illegal mask.
-    if (TLI.isShuffleMaskLegal(Mask, VT))
+    if (TLI.isShuffleMaskLegal(Mask, VT)) {
+      if (IsSV1Undef)
+        //   shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
+        //   shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
+        return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
       return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
     }
+  }
 
   return SDValue();
 }

test/CodeGen/X86/combine-vec-shuffle-4.ll

@@ -0,0 +1,122 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 | FileCheck %s
+
+; Verify that we fold shuffles according to rule:
+;  (shuffle(shuffle A, Undef, M0), B, M1) -> (shuffle A, B, M2)
+
+define <4 x float> @test1(<4 x float> %a, <4 x float> %b) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 4, i32 2, i32 3, i32 1>
+  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 4, i32 5, i32 1, i32 2>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test1
+; Mask: [4,5,2,3]
+; CHECK: movsd
+; CHECK: ret
+
+define <4 x float> @test2(<4 x float> %a, <4 x float> %b) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 6, i32 0, i32 1, i32 7>
+  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 1, i32 2, i32 4, i32 5>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test2
+; Mask: [0,1,4,5]
+; CHECK: movlhps 
+; CHECK: ret
+
+define <4 x float> @test3(<4 x float> %a, <4 x float> %b) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 0, i32 5, i32 1, i32 7>
+  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 2, i32 4, i32 1>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test3
+; Mask: [0,1,4,u]
+; CHECK: movlhps
+; CHECK: ret
+
+define <4 x float> @test4(<4 x float> %a, <4 x float> %b) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 5, i32 5>
+  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 6, i32 7, i32 0, i32 1>
+  ret <4 x float> %2
+}
+; FIXME: this should be lowered as a single movhlps. However, the backend
+; wrongly thinks that shuffle mask [6,7,2,3] is not legal. Therefore, we
+; end up with the sub-optimal sequence 'movhlps, palignr'.
+; CHECK-LABEL: test4
+; Mask: [6,7,2,3]
+; CHECK: movhlps
+; CHECK: palignr $8
+; CHECK: ret
+
+define <4 x float> @test5(<4 x float> %a, <4 x float> %b) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 0, i32 4, i32 1, i32 3>
+  %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 2, i32 6, i32 7>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test5
+; Mask: [0,1,6,7]
+; CHECK: blendps $12
+; CHECK: ret
+
+; Verify that we fold shuffles according to rule:
+;  (shuffle(shuffle A, Undef, M0), A, M1) -> (shuffle A, Undef, M2)
+
+define <4 x float> @test6(<4 x float> %a) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 4, i32 2, i32 3, i32 1>
+  %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 4, i32 5, i32 1, i32 2>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test6
+; Mask: [0,1,2,3]
+; CHECK-NOT: pshufd
+; CHECK-NOT: shufps
+; CHECK-NOT: movlhps
+; CHECK: ret
+
+define <4 x float> @test7(<4 x float> %a) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 6, i32 0, i32 1, i32 7>
+  %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 1, i32 2, i32 4, i32 5>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test7
+; Mask: [0,1,0,1]
+; CHECK-NOT: pshufd
+; CHECK-NOT: shufps
+; CHECK: movlhps 
+; CHECK-NEXT: ret
+
+define <4 x float> @test8(<4 x float> %a) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 0, i32 5, i32 1, i32 7>
+  %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 2, i32 4, i32 1>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test8
+; Mask: [0,1,0,u]
+; CHECK-NOT: pshufd
+; CHECK-NOT: shufps
+; CHECK: movlhps
+; CHECK-NEXT: ret
+
+define <4 x float> @test9(<4 x float> %a) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 5, i32 5>
+  %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 6, i32 7, i32 0, i32 1>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test9
+; Mask: [2,3,2,3]
+; CHECK-NOT: movlhps
+; CHECK-NOT: palignr
+; CHECK: movhlps
+; CHECK-NEXT: ret
+
+define <4 x float> @test10(<4 x float> %a) {
+  %1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 0, i32 4, i32 1, i32 3>
+  %2 = shufflevector <4 x float> %1, <4 x float> %a, <4 x i32> <i32 0, i32 2, i32 6, i32 7>
+  ret <4 x float> %2
+}
+; CHECK-LABEL: test10
+; Mask: [0,1,2,3]
+; CHECK-NOT: pshufd
+; CHECK-NOT: shufps
+; CHECK-NOT: movlhps
+; CHECK: ret
+