[X86][SSE42] Lower v2i64/v4i64 ASHR(X, 63) as PCMPGTQ(0, X)

Similar to what we do for vXi8 ASHR(X, 7), use SSE42's PCMPGTQ to splat the sign instead of using the PSRAD+PSHUFD.

Avoiding bitcasts this improves combines that utilize computeNumSignBits, permits memory folding and reduces pipe pressure. Although it does require a second register, given that this is a (cheap) zero register the impact is minimal.

Differential Revision: https://reviews.llvm.org/D32973

llvm-svn: 302525

1 files changed, 8 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ec04f284039..421bfce3c0c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -21741,6 +21741,14 @@ static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
     MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
     SDValue Ex = DAG.getBitcast(ExVT, R);
 
+    // ashr(R, 63) === cmp_slt(R, 0)
+    if (ShiftAmt == 63 && Subtarget.hasSSE42()) {
+      assert((VT != MVT::v4i64 || Subtarget.hasInt256()) &&
+             "Unsupported PCMPGT op");
+      return DAG.getNode(X86ISD::PCMPGT, dl, VT,
+                         getZeroVector(VT, Subtarget, DAG, dl), R);
+    }
+
     if (ShiftAmt >= 32) {
       // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
       SDValue Upper =