Skip to content

[AArch64] Update getCmpSelInstrCost #145568

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.

Sign up for GitHub

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

Jump to bottom
Draft
wants to merge 3 commits into
base: main
Choose a base branch
from
Draft

[AArch64] Update getCmpSelInstrCost #145568

wants to merge 3 commits into from

Conversation

davemgreen
Copy link
Collaborator

@davemgreen davemgreen commented Jun 24, 2025 •
edited
Loading

This currently

  • The AArch64TTIImpl::getCmpSelInstrCost function now tried to handle all cost kinds, not just TCK_RecipThroughput.
  • Updates fcmp costs to correctly model the cost of fcmp's predicates that require expansion into multiple instructions.
  • Adds a target feature for when SVE fcmxx instructions have a lower throughput than the Neon versions.
  • Some of the old select code that assumed they would not be well lowered were removed.

@github-actions GitHub Actions

This comment was marked as off-topic.

Sign in to view
davemgreen added a commit to davemgreen/llvm-project that referenced this pull request Jun 27, 2025
@davemgreen
[AArch64] Enable other cost kinds for getCmpSelInstrCost.
5695f9b 
This removes the CostKind == TCK_RecipThroughput limitation from
getCmpSelInstrCost, allowing it to return more accurate costs for CodeSize and
Lat / SizeLat. Especially for larger vectors under CodeSize, the returned costs
are currently 1, not the legalization cost.

There are further ideas for improvement in llvm#145568, this is the first step.
@davemgreen davemgreen mentioned this pull request Jun 29, 2025
Merged
davemgreen added a commit to davemgreen/llvm-project that referenced this pull request Jul 8, 2025
@davemgreen
[AArch64] Enable other cost kinds for getCmpSelInstrCost.
36567f3 
This removes the CostKind == TCK_RecipThroughput limitation from
getCmpSelInstrCost, allowing it to return more accurate costs for CodeSize and
Lat / SizeLat. Especially for larger vectors under CodeSize, the returned costs
are currently 1, not the legalization cost.

There are further ideas for improvement in llvm#145568, this is the first step.
@davemgreen
[AArch64] Add HasHalfSVEFCmpThroughput
1da81d6 
From the SWOG's for various CPUs, SVE fcmp can have a lower throughput that the
Neon instructions. This adds a cost-model tuning feature to the relevant CPUs
and increase the costs in those cases.  This should lead to fixed-width being
slightly more prefered on those CPUs.

In the long run this information would ideally come straight from the
scheduling model, but that is currently unproven and my experiments with it
have ran into issues that are not yet addressed.
@davemgreen
Grab it from scheduling info. A510 doesn't have the right info
76bbd80
@davemgreen
[AArch64] Update getCmpSelInstrCost
be43c70
@davemgreen davemgreen force-pushed the gh-a64-getCmpSelInstrCost branch from 3daea67 to be43c70 Compare August 12, 2025 09:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@davemgreen