[PATCH] D22814: [X86][SSE] Optimize the truncation of vector comparison results with PACKSS

[Michael Kuperstein via llvm-commits]

mkuper accepted this revision.
mkuper added a comment.
This revision is now accepted and ready to land.

LGTM


================
Comment at: lib/Target/X86/X86ISelLowering.cpp:14150
@@ +14149,3 @@
+  // No truncation required, we might get here due to recursive calls.
+  if (SrcVT == DstVT)
+    return In;
----------------
Ah, ok, this makes sense.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:14185
@@ +14184,3 @@
+    Lo = DAG.getBitcast(MVT::v16i16, Lo);
+    Hi = DAG.getBitcast(MVT::v16i16, Hi);
+    SDValue Res = DAG.getNode(X86ISD::PACKSS, DL, MVT::v32i8, Lo, Hi);
----------------
Oh, I just completely missed that in the definition of the VEX.256-encoded versions.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:14200
@@ +14199,3 @@
+    return truncateVectorCompareWithPACKSS(DstVT, Res, DL, DAG, Subtarget);
+  }
+
----------------
Never mind, I completely misread the code (hence the greater -> smaller comment).
Forget I said anything.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:14283
@@ +14282,3 @@
+  // we are saturating+packing zero/all bits only.
+  auto IsPackableComparison = [](SDValue V) {
+    unsigned Opcode = V.getOpcode();
----------------
Sounds good.


Repository:
  rL LLVM

https://reviews.llvm.org/D22814