[X86] Prefer lowerVectorShuffleAsBitMask over using a avx512 masked operation when avx512bw/avx512vl is enabled.

This does require a constant pool load instead of loading an immediate into a gpr, moving to a k register and masking. But its less instructions and more consistent with previous ISAs. It probably opens up more combine opportunities as one of the test cases demonstrates.

llvm-svn: 348018

1 files changed, 5 insertions, 5 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 0035335a2a7..890103dfd9f 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -10175,6 +10175,11 @@ static SDValue lowerVectorShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1,
     assert((VT.is128BitVector() || 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, Zeroable, DAG))
+      return Masked;
+
     if (Subtarget.hasBWI() && Subtarget.hasVLX()) {
       MVT IntegerType =
           MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8));
@@ -10182,11 +10187,6 @@ static SDValue lowerVectorShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1,
       return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG);
     }
 
-    // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
-    if (SDValue Masked =
-            lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable, DAG))
-      return Masked;
-
     // Scale the blend by the number of bytes per element.
     int Scale = VT.getScalarSizeInBits() / 8;