[X86] Lower vselect with constant condition to vector_shuffle even with AVX512 instructions.

Summary:
It's better to use our shuffle lowering code to handle these than loading an immediate into a k-register.

It really feels like this should be a DAG combine optimization rather than a lowering operation, but that's a problem for another day.

Reviewers: RKSimon, delena, zvi

Reviewed By: delena

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D38932

llvm-svn: 315849

1 files changed, 5 insertions, 5 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index c08d79663fa..f155de8ef02 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -14077,16 +14077,16 @@ SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
       ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
     return SDValue();
 
-  // If this VSELECT has a vector if i1 as a mask, it will be directly matched
-  // with patterns on the mask registers on AVX-512.
-  if (Op->getOperand(0).getValueType().getScalarSizeInBits() == 1)
-    return Op;
-
   // Try to lower this to a blend-style vector shuffle. This can handle all
   // constant condition cases.
   if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
     return BlendOp;
 
+  // If this VSELECT has a vector if i1 as a mask, it will be directly matched
+  // with patterns on the mask registers on AVX-512.
+  if (Op->getOperand(0).getValueType().getScalarSizeInBits() == 1)
+    return Op;
+
   // Variable blends are only legal from SSE4.1 onward.
   if (!Subtarget.hasSSE41())
     return SDValue();