Revert "Persist valid rematerialisations from slice creation" #8395
Conversation
This reverts commit e3cf2ad.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting
| for (Value v : slice) { | ||
| auto layoutIt = layout.find(v); | ||
| assert(layoutIt != layout.end()); | ||
| // If we found a valid rematerialization for this value while constructing | ||
| // the slice, use that. | ||
| if (Value remat = existingRemats.lookup({v, layoutIt->second})) { | ||
| assert(getRematValue(v, layoutIt->second) == remat && "remat mismatch"); | ||
| // If we already have a remat value for this value, use it. | ||
| if (Value remat = getRematValue(v, layoutIt->second)) { | ||
| mapping.map(v, remat); | ||
| valuesWithExistingRemat.insert(v); | ||
| continue; |
There was a problem hiding this comment.
While rewriting a slice, the code now calls getRematValue directly to skip rematerializing a value whenever any remat is registered. The callback in getConvertBackwardSlice only filters remats that dominate the current use, but that information is no longer carried forward to the rewrite step after the existingRemats map was removed. As a result, rewriteSlice can map a value to a remat that lives in a different control path and does not dominate the rewritten convert, yielding invalid IR (uses before defs) or incorrect rematerialisation. The earlier implementation only reused remats recorded during slice construction, so the dominance constraint was preserved.
Useful? React with 👍 / 👎.
…-lang#8395) Reverts triton-lang#8292 as it causes functional regressions
Reverts #8292 as it causes functional regressions
…-lang#8395) Reverts triton-lang#8292 as it causes functional regressions
Reverts #8292 as it causes functional regressions