[X86] Generate VPSHUFB for in-place v16i16 shuffles

This used to resort to splitting the 256-bit operation into two 128-bit
shuffles and then recombining the results.

Fixes <rdar://problem/16167303>

llvm-svn: 204735

2 files changed, 47 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 14d7be31306..ec923ba7aa2 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6576,6 +6576,25 @@ LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
   return NewV;
 }
 
+/// \brief v16i16 shuffles
+///
+/// FIXME: We only support generation of a single pshufb currently.  We can
+/// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
+/// well (e.g 2 x pshufb + 1 x por).
+static SDValue
+LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
+  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+  SDValue V1 = SVOp->getOperand(0);
+  SDValue V2 = SVOp->getOperand(1);
+  SDLoc dl(SVOp);
+
+  if (V2.getOpcode() != ISD::UNDEF)
+    return SDValue();
+
+  SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
+  return getPSHUFB(MaskVals, V1, dl, DAG);
+}
+
 // v16i8 shuffles - Prefer shuffles in the following order:
 // 1. [ssse3] 1 x pshufb
 // 2. [ssse3] 2 x pshufb + 1 x por
@@ -7635,6 +7654,12 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
       return NewOp;
   }
 
+  if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
+    SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
+    if (NewOp.getNode())
+      return NewOp;
+  }
+
   if (VT == MVT::v16i8) {
     SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
     if (NewOp.getNode())
diff --git a/llvm/test/CodeGen/X86/vec_shuffle-40.ll b/llvm/test/CodeGen/X86/vec_shuffle-40.ll
new file mode 100644
index 00000000000..75b45e3df11
--- /dev/null
+++ b/llvm/test/CodeGen/X86/vec_shuffle-40.ll
@@ -0,0 +1,22 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core-avx2 | FileCheck %s
+
+define void @shuffle_v16i16(<16 x i16>* %a) {
+; CHECK-LABEL: shuffle_v16i16:
+; CHECK: vpshufb {{.*}}%ymm
+; CHECK-NOT: vpshufb {{.*}}%xmm
+entry:
+  %0 = load <16 x i16>* %a, align 32
+  %shuffle = shufflevector <16 x i16> %0, <16 x i16> undef, <16 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7, i32 9, i32 9, i32 11, i32 11, i32 13, i32 13, i32 15, i32 15>
+  store <16 x i16> %shuffle, <16 x i16>* %a, align 32
+  ret void
+}
+
+define void @shuffle_v16i16_lanecrossing(<16 x i16>* %a) {
+; CHECK-LABEL: shuffle_v16i16_lanecrossing:
+; CHECK-NOT: vpshufb {{.*}}%ymm
+entry:
+  %0 = load <16 x i16>* %a, align 32
+  %shuffle = shufflevector <16 x i16> %0, <16 x i16> undef, <16 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 13, i32 7, i32 7, i32 9, i32 9, i32 11, i32 11, i32 13, i32 13, i32 15, i32 15>
+  store <16 x i16> %shuffle, <16 x i16>* %a, align 32
+  ret void
+}