[PATCH] D75956: [mlir] [VectorOps] Progressively lower vector.outerproduct to LLVM

Tue Mar 10 13:39:53 PDT 2020

aartbik marked 2 inline comments as done.
aartbik added a comment.

In D75956#1915583 <https://reviews.llvm.org/D75956#1915583>, @rriddle wrote:

> Thanks Aart. Is this going to simplify some upcoming patterns or existing ones?

This is supposed to remove the VectorOuterProductOpConversion class completely from ConvertVectorToLLVM.cpp (following the usual MLIR philosophy that optimizations should complement each other, not duplicate each other). I was going to do that removal in this CL already, but than got stuck on some (imho somewhat poorly phrased) linalg tests that (1) insert vector dialect in the test without vector lowing (2) have "dead code" in their tests so that I cannot immediately pull in the vector lowering, since that also does DCE as an intentional side effect.

So the removal + linalg tests will be a follow up CL, so to keep things separate.

================
Comment at: mlir/include/mlir/Dialect/VectorOps/VectorOps.h:68
+
+// TODO(ajcbik): remove when external references are renamed too
+#define populateVectorContractLoweringPatterns                                 \
----------------
rriddle wrote:
> Let's remove this now.
Ok there is only one other place (outside this repo) that I know uses this, so I will take care of that. Removed here now.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D75956/new/

https://reviews.llvm.org/D75956