From ef44228acba499cba328ad7ce399be21462e584c Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Tue, 8 Aug 2017 11:03:30 +0000
Subject: [PATCH] [DAGCombiner] Simplify shuffle mask index if the referenced
 input element is UNDEF

Fixes one of the cases in PR34041.

Differential Revision: https://reviews.llvm.org/D36393

llvm-svn: 310344
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 36 +++++++++++++++++++++++++++
 llvm/test/CodeGen/X86/avx-vbroadcast.ll       | 18 +++++---------
 2 files changed, 42 insertions(+), 12 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 4b57787..573002a 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -15112,6 +15112,38 @@ static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
   return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
 }
 
+static SDValue simplifyShuffleMask(ShuffleVectorSDNode *SVN, SDValue N0,
+                                   SDValue N1, SelectionDAG &DAG) {
+  // TODO - handle cases other than BUILD_VECTOR.
+  auto *BV0 = dyn_cast<BuildVectorSDNode>(N0);
+  auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
+  if (!BV0 && !BV1)
+    return SDValue();
+
+  EVT VT = SVN->getValueType(0);
+  unsigned NumElts = VT.getVectorNumElements();
+
+  bool Changed = false;
+  SmallVector<int, 8> NewMask;
+  for (unsigned i = 0; i != NumElts; ++i) {
+    int Idx = SVN->getMaskElt(i);
+    if (BV0 && 0 <= Idx && Idx < (int)NumElts &&
+        BV0->getOperand(Idx).isUndef()) {
+      Changed = true;
+      Idx = -1;
+    } else if (BV1 && Idx > (int)NumElts &&
+               BV1->getOperand(Idx - NumElts).isUndef()) {
+      Changed = true;
+      Idx = -1;
+    }
+    NewMask.push_back(Idx);
+  }
+  if (Changed)
+    return DAG.getVectorShuffle(VT, SDLoc(SVN), N0, N1, NewMask);
+
+  return SDValue();
+}
+
 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
 // or turn a shuffle of a single concat into simpler shuffle then concat.
 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
@@ -15461,6 +15493,10 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
       return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, NewMask);
   }
 
+  // Simplify shuffle mask if a referenced element is UNDEF.
+  if (SDValue V = simplifyShuffleMask(SVN, N0, N1, DAG))
+    return V;
+
   // A shuffle of a single vector that is a splat can always be folded.
   if (auto *N0Shuf = dyn_cast<ShuffleVectorSDNode>(N0))
     if (N1->isUndef() && N0Shuf->isSplat())
diff --git a/llvm/test/CodeGen/X86/avx-vbroadcast.ll b/llvm/test/CodeGen/X86/avx-vbroadcast.ll
index ccb77a2..97035ba 100644
--- a/llvm/test/CodeGen/X86/avx-vbroadcast.ll
+++ b/llvm/test/CodeGen/X86/avx-vbroadcast.ll
@@ -836,14 +836,12 @@ define <4 x double> @broadcast_shuffle_1000(double* %p) {
 ; X32-LABEL: broadcast_shuffle_1000:
 ; X32:       ## BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X32-NEXT:    vbroadcastsd (%eax), %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: broadcast_shuffle_1000:
 ; X64:       ## BB#0:
-; X64-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X64-NEXT:    vbroadcastsd (%rdi), %ymm0
 ; X64-NEXT:    retq
   %1 = load double, double* %p
   %2 = insertelement <2 x double> undef, double %1, i32 0
@@ -855,18 +853,14 @@ define <4 x double> @broadcast_shuffle1032(double* %p) {
 ; X32-LABEL: broadcast_shuffle1032:
 ; X32:       ## BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X32-NEXT:    vmovddup {{.*#+}} xmm1 = xmm0[0,0]
-; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; X32-NEXT:    vpermilpd {{.*#+}} ymm0 = ymm0[1,0,3,2]
+; X32-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
+; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: broadcast_shuffle1032:
 ; X64:       ## BB#0:
-; X64-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; X64-NEXT:    vmovddup {{.*#+}} xmm1 = xmm0[0,0]
-; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; X64-NEXT:    vpermilpd {{.*#+}} ymm0 = ymm0[1,0,3,2]
+; X64-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
+; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-NEXT:    retq
   %1 = load double, double* %p
   %2 = insertelement <2 x double> undef, double %1, i32 1
-- 
2.7.4