[GVN] Don't coerce vector store via i128+#8574
Merged
Merged
Conversation
GVN forwards a wide stored value to a narrow load from the same addres by bitcasting the stored value to an integer, then trucating to the loaded value's size. This avoids re-loading a value that was just stored. For DXIL this is unsafe if the wider size is greater than 64-bits since DXIL doesn't allow integers larger than 64-bits. This change disables coercing in this case. This shoudln't cause any performance regressions in practice because no existing cases can generate valid DXIL of this form, but it does generate less optimal final output. We could consider more robust load-store optimizations and coersion to integer vectors as an alternative, but at this time the important thing is to make DXC not generate invalid DXIL, which this simplified change does. Assisted by Claude Opus 4.7
V-FEXrt
approved these changes
Jun 22, 2026
Icohedron
approved these changes
Jun 25, 2026
Collaborator
There was a problem hiding this comment.
one nit about the PR description though:
This shouldn't cause any performance regressions in practice because no existing cases can generate valid DXIL of this form, but it does generate less optimal final output.
Is the final output really "less optimal" if the "optimal" is invalid?
Collaborator
Author
Fair that my wording was confusing. The thing that's a bit silly about the DXIL rules here is that there's not really any good reason why i128's aren't legal. In practice i128s and bitwise operations on them (probably not arithmetic) should be trivial to implement and likely allow backends to generate better final code, but DXIL disallows them. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
GVN forwards a wide stored value to a narrow load from the same address by bitcasting the stored value to an integer, then truncating to the loaded value's size. This avoids re-loading a value that was just stored.
For DXIL this is unsafe if the wider size is greater than 64-bits since DXIL doesn't allow integers larger than 64-bits. This change disables coercing in this case. This shouldn't cause any regressions in practice because the current validator fails to allow this code.
We could consider more robust load-store optimizations and coercion to integer vectors as an alternative, but at this time the important thing is to make DXC not generate invalid DXIL, which this simplified change does.
Fixes #8573
Assisted by Claude Opus 4.7