Merging r372886:

rotateright · tstellar · commit 7a140f4a293e · 2019-11-12T13:57:03.000-08:00
------------------------------------------------------------------------
r372886 | spatel | 2019-09-25 08:08:33 -0700 (Wed, 25 Sep 2019) | 7 lines

[DAGCombiner] add one-use restriction to vector transform with cheap extract

We might be able to do better on the example in the test,
but in general, we should not scalarize a splatted vector
binop if there are other uses of the binop. Otherwise, we
can end up with code as we had - a scalar op that is
redundant with a vector op.
------------------------------------------------------------------------
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18878,7 +18878,7 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   // build_vector.
   if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
     int SplatIndex = SVN->getSplatIndex();
-    if (TLI.isExtractVecEltCheap(VT, SplatIndex) &&
+    if (N0.hasOneUse() && TLI.isExtractVecEltCheap(VT, SplatIndex) &&
         TLI.isBinOp(N0.getOpcode()) && N0.getNode()->getNumValues() == 1) {
       // splat (vector_bo L, R), Index -->
       // splat (scalar_bo (extelt L, Index), (extelt R, Index))
diff --git a/llvm/test/CodeGen/X86/scalarize-fp.ll b/llvm/test/CodeGen/X86/scalarize-fp.ll
@@ -777,21 +777,18 @@ define <8 x float> @splat0_fdiv_const_op0_v8f32(<8 x float> %vx) {
 define <4 x float> @multi_use_binop(<4 x float> %x, <4 x float> %y) {
 ; SSE-LABEL: multi_use_binop:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movaps %xmm0, %xmm2
-; SSE-NEXT:    mulps %xmm1, %xmm2
-; SSE-NEXT:    mulss %xmm1, %xmm0
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0,0,0]
-; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm2[0]
+; SSE-NEXT:    mulps %xmm1, %xmm0
+; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1,2,0]
 ; SSE-NEXT:    addps %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: multi_use_binop:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmulps %xmm1, %xmm0, %xmm2
-; AVX-NEXT:    vmulss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,0,0]
-; AVX-NEXT:    vmovddup {{.*#+}} xmm1 = xmm2[0,0]
-; AVX-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vmulps %xmm1, %xmm0, %xmm0
+; AVX-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[0,1,2,0]
+; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %mul = fmul <4 x float> %x, %y
   %mul0 = shufflevector <4 x float> %mul, <4 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 undef, i32 0>
