JIT: Graph-based loop inversion

[amanasifkhalid]

Succeeds #113709 (GitHub won't let me reopen it). Part of #107749 and #108901. Fixes #50204.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull Request Overview

This pull request implements graph-based loop inversion in the JIT, aiming to improve IR transformation behavior and diagnostics.

• In fgopt.cpp, the logic now returns true to indicate IR modifications in edge cases, and error handling for stmt cloning has been shifted to an assertion.
• In fgdiagnostic.cpp, a more descriptive debug message is added for loop exits with non-loop predecessors.
• In compiler.h, the API is updated to use a more specific function (optTryInvertWhileLoop) with a refined parameter type.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
src/coreclr/jit/fgopt.cpp	Updated control flow for IR modification detection and error handling
src/coreclr/jit/fgdiagnostic.cpp	Enhanced loop diagnostic messaging
src/coreclr/jit/compiler.h	Updated API for loop inversion with adjusted parameter type

Comments suppressed due to low confidence (2)

src/coreclr/jit/fgopt.cpp:2713

• Changing the response for non-RelOp conditions from returning false to returning true alters the behavior of IR modifications. Verify that this change is intentional and that downstream code correctly handles the modified IR state.

if (!condTree->OperIsCompare())

src/coreclr/jit/compiler.h:7123

• The API change from 'optInvertWhileLoop(BasicBlock* block)' to 'optTryInvertWhileLoop(FlowGraphNaturalLoop* loop)' requires ensuring that all call sites and semantic expectations are updated accordingly.

bool optTryInvertWhileLoop(FlowGraphNaturalLoop* loop);

[amanasifkhalid]

Here's aspnet:

• Oddly, the relationship between more inversion and loops found is flipped.
• Again, the gap between 80 and 100 looks interesting here.

A size limit of 100 seems like a good starting point. That should help trim out the runaway cloning for a few large loops in libraries_tests.run. It looks like anything lower might meaningfully reduce existing cloning.

[amanasifkhalid]

Diffs are now a net size reduction. Unsurprisingly, we have lots of size increases from finding new inversion candidates, as well as size decreases from bailing on large loops previously inverted by the lexical implementation. Whether we are doing more or less inversion overall depends on the collection. For example:

• In aspnet, the number of loops found drops modestly from 41768 to 41481, and the number of loops inverted increases from 15939 to 16281.
• In benchmarks.run_pgo, the number of loops found drops from 91864 to 86981, and the number of loops inverted decreases from 36156 to 31067.

@AndyAyersMS @jakobbotsch do these diffs seem like too much of a swing in the opposite direction?

[AndyAyersMS]

@AndyAyersMS @jakobbotsch do these diffs seem like too much of a swing in the opposite direction?

I think the TP cost is ok here. For CQ I don't know what experiments to try other than running a lot of benchmarks, and for that the lab data will be a better guide. P6 snap looks like 6/13 so we have time to get data back and change our minds.

So let's try this and see.

[amanasifkhalid]

@AndyAyersMS I think that's a reasonable plan, though we'll want #115850 merged in around the same time to see improvements in the list enumeration scenarios we're targeting. That change will incur a similarly large number of diffs; are you ok trying these back-to-back before the next snap?

[AndyAyersMS]

@AndyAyersMS I think that's a reasonable plan, though we'll want #115850 merged in around the same time to see improvements in the list enumeration scenarios we're targeting. That change will incur a similarly large number of diffs; are you ok trying these back-to-back before the next snap?

Maybe not immeidately back to back -- a day apart?

[amanasifkhalid]

a day apart?

Sure, sounds reasonable.

[AndyAyersMS]

Improvements:

• [Perf] Linux/x64: 23 Improvements on 6/4/2025 5:12:34 PM +00:00 perf-autofiling-issues#57262
• [Perf] Windows/x64: 19 Improvements on 6/4/2025 5:12:34 PM +00:00 perf-autofiling-issues#57219
• [Perf] Windows/x64: 53 Improvements on 6/4/2025 5:12:34 PM +00:00 perf-autofiling-issues#57137
• [Perf] Linux/x64: 53 Improvements on 6/4/2025 5:12:34 PM +00:00 perf-autofiling-issues#57147
• [Perf] Linux/arm64: 31 Improvements on 6/4/2025 4:11:30 PM +00:00 perf-autofiling-issues#57439
• [Perf] Windows/arm64: 11 Improvements on 6/4/2025 4:11:30 PM +00:00 perf-autofiling-issues#57547