[PATCH] D53823: [X86] In lowerVectorShuffleAsBroadcast, make peeking through CONCAT_VECTORS work correctly if we already walked through a bitcast that changed the element size.

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.
craig.topper added a reviewer: RKSimon.

The CONCAT_VECTORS case was using the original mask element count to determine how to adjust the broadcast index. But if we looked through a bitcast the original mask size doesn't tell us anything about the concat_vectors.

This patch switchs to using the concat_vectors input element count directly instead.

This caused a crash while doing experiments with -mprefer-vector-width=256 with skylake-avx512 on some benchmarks. All the types present in the crash were 256 or 128 bits wide so I'm not sure why -mprefer-vector-width=256 was relevant.

I don't currently have a reduced test case, but I'll see what I can do.


https://reviews.llvm.org/D53823

Files:
  lib/Target/X86/X86ISelLowering.cpp


Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -11239,7 +11239,8 @@
       continue;
     }
     case ISD::CONCAT_VECTORS: {
-      int OperandSize = Mask.size() / V.getNumOperands();
+      int OperandSize =
+          V.getOperand(0).getSimpleValueType().getVectorNumElements();
       V = V.getOperand(BroadcastIdx / OperandSize);
       BroadcastIdx %= OperandSize;
       continue;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D53823.171544.patch
Type: text/x-patch
Size: 531 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181029/f3f4b94e/attachment.bin>