[X86][SSE] Use bitmasks instead of shuffles where possible.

VPAND is a lot faster than VPSHUFB and VPBLENDVB - this patch ensures we attempt to lower to a basic bitmask before lowering to the slower byte shuffle/blend instructions.

Split off from D11518.

Differential Revision: http://reviews.llvm.org/D11541

llvm-svn: 243395

1 files changed, 8 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 10904775179..b3672c3216d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6663,6 +6663,10 @@ static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
     assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
            "256-bit byte-blends require AVX2 support!");
 
+    // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
+    if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
+      return Masked;
+
     // Scale the blend by the number of bytes per element.
     int Scale = VT.getScalarSizeInBits() / 8;
 
@@ -9071,6 +9075,10 @@ static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
       return V;
   }
 
+  if (SDValue Masked =
+          lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
+    return Masked;
+
   // Use dedicated unpack instructions for masks that match their pattern.
   if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
                                          0, 16, 1, 17, 2, 18, 3, 19,