Skip to content

Do not use PerformEXTRACTCombine for v8i8 types #81242

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Feb 12, 2024

Conversation

akuegel
Copy link
Member

@akuegel akuegel commented Feb 9, 2024

Same as with v4i8 types, we should not be using PerformEXTRACTCombine for v8i8 types.

Same as with v4i8, we should not be using PerformEXTRACTCombine for v8i8 types.
@akuegel akuegel requested a review from Artem-B February 9, 2024 10:33
@Artem-B
Copy link
Member

Artem-B commented Feb 9, 2024

Something odd is going on here. Skipping MVT::v8i8 here should've resulted in lowering extractelement using 64-bit BFE instruction. Yet we're ending with the code splitting 64-bit input it into two 32-bit words and then extracting bytes from each using 32-bit BFE. It's probably a wash performance-wise, as the actual registers are 32-bit, but it suggests we're missing something. We may need to mark the v8i8 extractelement legal.

Copy link
Member

@Artem-B Artem-B left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM. Puzzling changes in PTX output are still valid, and the test fixes are useful.
I'll take a look at the unexpected lowering separately.

@akuegel akuegel merged commit da9559d into llvm:main Feb 12, 2024
@akuegel akuegel deleted the akuegel-patch-1 branch February 12, 2024 06:31
@Artem-B
Copy link
Member

Artem-B commented Feb 12, 2024

Fix for the lowering is in #81308

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants