Introduce matching patterns for vbroadcast AVX instruction. The idea is to

match splats in the form (splat (scalar_to_vector (load ...))) whenever
the load can be folded. All the logic and instruction emission is
working but because of PR8156, there are no ways to match loads, cause
they can never be folded for splats. Thus, the tests are XFAILed, but
I've tested and exercised all the logic using a relaxed version for
checking the foldable loads, as if the bug was already fixed. This
should work out of the box once PR8156 gets fixed since MayFoldLoad will
work as expected.

llvm-svn: 137810

1 files changed, 47 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 42d42da08d2..b330a4f9809 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6151,6 +6151,48 @@ static inline unsigned getVPERMILOpcode(EVT VT) {
   return 0;
 }
 
+/// isVectorBroadcast - Check if the node chain is suitable to be xformed to
+/// a vbroadcast node. The nodes are suitable whenever we can fold a load coming
+/// from a 32 or 64 bit scalar. Update Op to the desired load to be folded.
+static bool isVectorBroadcast(SDValue &Op) {
+  EVT VT = Op.getValueType();
+  bool Is256 = VT.getSizeInBits() == 256;
+
+  assert((VT.getSizeInBits() == 128 || Is256) &&
+         "Unsupported type for vbroadcast node");
+
+  SDValue V = Op;
+  if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
+    V = V.getOperand(0);
+
+  if (Is256 && !(V.hasOneUse() &&
+                 V.getOpcode() == ISD::INSERT_SUBVECTOR &&
+                 V.getOperand(0).getOpcode() == ISD::UNDEF))
+    return false;
+
+  if (Is256)
+    V = V.getOperand(1);
+  if (V.hasOneUse() && V.getOpcode() != ISD::SCALAR_TO_VECTOR)
+    return false;
+
+  // Check the source scalar_to_vector type. 256-bit broadcasts are
+  // supported for 32/64-bit sizes, while 128-bit ones are only supported
+  // for 32-bit scalars.
+  unsigned ScalarSize = V.getOperand(0).getValueType().getSizeInBits();
+  if (ScalarSize != 32 && ScalarSize != 64)
+    return false;
+  if (!Is256 && ScalarSize == 64)
+    return false;
+
+  V = V.getOperand(0);
+  if (!MayFoldLoad(V))
+    return false;
+
+  // Return the load node
+  Op = V;
+  return true;
+}
+
 static
 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
                                const TargetLowering &TLI,
@@ -6174,6 +6216,10 @@ SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
     if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
       return Op;
 
+    // Use vbroadcast whenever the splat comes from a foldable load
+    if (Subtarget->hasAVX() && isVectorBroadcast(V1))
+      return DAG.getNode(X86ISD::VBROADCAST, dl, VT, V1);
+
     // Handle splats by matching through known shuffle masks
     if (VT.is128BitVector() && NumElem <= 4)
       return SDValue();
@@ -10189,6 +10235,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::PUNPCKHWD:          return "X86ISD::PUNPCKHWD";
   case X86ISD::PUNPCKHDQ:          return "X86ISD::PUNPCKHDQ";
   case X86ISD::PUNPCKHQDQ:         return "X86ISD::PUNPCKHQDQ";
+  case X86ISD::VBROADCAST:         return "X86ISD::VBROADCAST";
   case X86ISD::VPERMILPS:          return "X86ISD::VPERMILPS";
   case X86ISD::VPERMILPSY:         return "X86ISD::VPERMILPSY";
   case X86ISD::VPERMILPD:          return "X86ISD::VPERMILPD";