[PATCH] D156910: [clang] Add pragma force_vectorize

Aaron Ballman via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Wed Nov 1 09:34:30 PDT 2023


aaron.ballman added a comment.

In D156910#4602946 <https://reviews.llvm.org/D156910#4602946>, @kitaisreal wrote:

> Hello @aaron.ballman, @erichkeane could you please review this revision ? I wondering if this feature would be useful.

My apologies, this fell off my radar by accident, sorry for the delayed response!

I think the utility you're proposing is valuable in theory, but I'm wondering if this is really the right design for the feature (pragmas are awkward for everyone). e.g., should we use an attribute that appertains only to particular kinds of looping constructs? Should this really be an error as opposed to a warning the user is free to ignore?

It seems we already support `-Rpass-missed=loop-vectorize` and `-Rpass-analysis=loop-vectorize` (https://www.llvm.org/docs/Vectorizers.html#diagnostics) for similar purposes, is there a reason those don't suffice?

I think this is a topic where people may have enough opinions to warrant raising it as an RFC on Discourse (https://discourse.llvm.org/) to get feedback on the design. Note, we've largely switched reviews over to GitHub PRs. We don't usually move a review from Phab to GitHub because we'll lose significant review context, but it may be a reasonable approach here given the unfortunate lack of review thus far. So my recommendation if you're still interested in pursuing this would be to start a discussion on Discourse and then open a new review on GitHub once the design is more settled.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156910/new/

https://reviews.llvm.org/D156910



More information about the cfe-commits mailing list