[PATCH] D49829: [X86] Add pattern matching for PMADDUBSW

[Hideki Saito via Phabricator via llvm-commits]

hsaito added a comment.

Sorry, for the late response to the already closed review.

> In this case it looks necessary, but it encourages the belief that its safe to perform vectorization in the DAG. In general we should be relying on the vectorizers to perform this using a proper cost analysis.

Vectorizer to perform a proper cost analysis, I agree 100%.

Vectorizer to perform this part, we can't do that unless we want vectorizer to emit target (in)dependent intrinsics all over the place. As long as we want to keep using sequence of IR instructions, what's more important is to build up common and robust (TTI based?) pattern matchers that can be used at IR level as well as at the DAG level such that what we see at vectorizer can also be captured at DAG Optimizer.


Repository:
  rL LLVM

https://reviews.llvm.org/D49829