DAGCombiner: Fold a shuffle on CONCAT_VECTORS into a new CONCAT_VECTORS if possible.

This pattern occurs in SROA output due to the way vector arguments are lowered
on ARM.

The testcase from PR15525 now compiles into this, which is better than the code
we got with the old scalarrepl:
_Store:
	ldr.w	r9, [sp]
	vmov	d17, r3, r9
	vmov	d16, r1, r2
	vst1.8	{d16, d17}, [r0]
	bx	lr

Differential Revision: http://llvm-reviews.chandlerc.com/D647

llvm-svn: 179106

2 files changed, 72 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index eb160957501..44e212aef00 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -9125,6 +9125,44 @@ SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
   return SDValue();
 }
 
+// Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
+static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
+  EVT VT = N->getValueType(0);
+  unsigned NumElts = VT.getVectorNumElements();
+
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+
+  SmallVector<SDValue, 4> Ops;
+  EVT ConcatVT = N0.getOperand(0).getValueType();
+  unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
+  unsigned NumConcats = NumElts / NumElemsPerConcat;
+
+  // Look at every vector that's inserted. We're looking for exact
+  // subvector-sized copies from a concatenated vector
+  for (unsigned I = 0; I != NumConcats; ++I) {
+    // Make sure we're dealing with a copy.
+    unsigned Begin = I * NumElemsPerConcat;
+    if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
+      return SDValue();
+
+    for (unsigned J = 1; J != NumElemsPerConcat; ++J) {
+      if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
+        return SDValue();
+    }
+
+    unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
+    if (FirstElt < N0.getNumOperands())
+      Ops.push_back(N0.getOperand(FirstElt));
+    else
+      Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
+  }
+
+  return DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT, Ops.data(),
+                     Ops.size());
+}
+
 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   EVT VT = N->getValueType(0);
   unsigned NumElts = VT.getVectorNumElements();
@@ -9226,6 +9264,17 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
     }
   }
 
+  if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
+      Level < AfterLegalizeVectorOps &&
+      (N1.getOpcode() == ISD::UNDEF ||
+      (N1.getOpcode() == ISD::CONCAT_VECTORS &&
+       N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
+    SDValue V = partitionShuffleOfConcats(N, DAG);
+
+    if (V.getNode())
+      return V;
+  }
+
   // If this shuffle node is simply a swizzle of another shuffle node,
   // and it reverses the swizzle of the previous shuffle then we can
   // optimize shuffle(shuffle(x, undef), undef) -> x.
diff --git a/llvm/test/CodeGen/ARM/dagcombine-concatvector.ll b/llvm/test/CodeGen/ARM/dagcombine-concatvector.ll
new file mode 100644
index 00000000000..e9e0fe3239a
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/dagcombine-concatvector.ll
@@ -0,0 +1,23 @@
+; RUN: llc < %s -mtriple=thumbv7s-apple-ios3.0.0 | FileCheck %s
+
+; PR15525
+; CHECK: test1:
+; CHECK: ldr.w	[[REG:r[0-9]+]], [sp]
+; CHECK-NEXT: vmov	{{d[0-9]+}}, r1, r2
+; CHECK-NEXT: vmov	{{d[0-9]+}}, r3, [[REG]]
+; CHECK-NEXT: vst1.8	{{{d[0-9]+}}, {{d[0-9]+}}}, [r0]
+; CHECK-NEXT: bx	lr
+define void @test1(i8* %arg, [4 x i64] %vec.coerce) {
+bb:
+  %tmp = extractvalue [4 x i64] %vec.coerce, 0
+  %tmp2 = bitcast i64 %tmp to <8 x i8>
+  %tmp3 = shufflevector <8 x i8> %tmp2, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %tmp4 = extractvalue [4 x i64] %vec.coerce, 1
+  %tmp5 = bitcast i64 %tmp4 to <8 x i8>
+  %tmp6 = shufflevector <8 x i8> %tmp5, <8 x i8> undef, <16 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %tmp7 = shufflevector <16 x i8> %tmp6, <16 x i8> %tmp3, <16 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  tail call void @llvm.arm.neon.vst1.v16i8(i8* %arg, <16 x i8> %tmp7, i32 2)
+  ret void
+}
+
+declare void @llvm.arm.neon.vst1.v16i8(i8*, <16 x i8>, i32)