JIT: Avoid inverting already-inverted loops that are IV tested #119540
Conversation
- Follow unique preds when locating the init statement for the iterator variable - Generalize zero-trip detection by similarly following unique preds
Graph-based loop inversion works by inverting the loop such that we do not jump into the loop to immediately test a condition. However, we are not detecting the case where the loop is already inverted because its latch is an exit that tests a different condition. If the loop has multiple exits that means we may end up creating a suboptimal latch condition. Detect this case based on the pattern-based IV analysis that we already have.
github-actions
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
@EgorBot -intel --filter "SeekUnroll" |
|
@EgorBot -intel --filter "QuickSortSpan" |
|
@EgorBot -intel --filter "QuickSortSpan" --disasm |
|
On benchmarks.run_pgo we go from 7970 -> 7584 loops inverted, 1445 -> 1442 loops cloned, 342 -> 411 loops unrolled. Need to look at some of these loop cloning differences to spot check whether they make sense or not. |
There was a problem hiding this comment.
Pull Request Overview
This PR improves JIT loop inversion optimization by avoiding the inversion of loops that are already optimally inverted with IV (induction variable) tests. The change prevents the creation of suboptimal latch conditions when a loop has multiple exits and one exit is already serving as an effective latch with an IV test.
Key changes:
- Add detection logic to identify already-inverted loops with IV tests before attempting loop inversion
- Remove unused parameter from
optExtractTestIncrmethod to simplify the interface - Update all call sites to use the simplified method signature
Reviewed Changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| src/coreclr/jit/optimizer.cpp | Add IV test detection logic to skip loop inversion and remove unused header parameter from optExtractTestIncr |
| src/coreclr/jit/flowgraph.cpp | Update call to optExtractTestIncr with simplified signature |
| src/coreclr/jit/compiler.h | Update method declaration to remove unused header parameter |
|
The difference in loop cloning is just from a context where the base replays cleanly while the diff (after differences in inversion) no longer replays cleanly. In reality the partial replay of the diff does show 3 loop clonings, but those do not get recorded once it misses. cc @dotnet/jit-contrib PTAL @AndyAyersMS Diffs. As mentioned above we go from 7970 -> 7584 loops inverted (because we now detect they are already inverted with an "IV looking" condition) and 342 -> 411 loops unrolled. |
2 participants
Graph-based loop inversion works by inverting the loop such that we do not jump into the loop to immediately test a condition. However, we are not detecting the case where the loop is already inverted because its latch is an exit that tests a different condition. If the loop has multiple exits that means we may end up creating a suboptimal latch condition.
Detect this case based on the pattern-based IV analysis that we already have.
This fixes #119277 and likely fixes a bunch of the regressions tracked by #116754.
This PR is based on #119537. We may seriously consider whether a dumbed down version of this can be backported to .NET 10. For example, if the latch is an exit that looks like a simple IV increment, then skip loop inversion.