[derek-gerstmann]

Summary

Conditional box unions reference previous min/max expressions multiple times.
Across nested specializations, simplification repeatedly traverses the shared
expression DAG as an expanded tree, causing exponential compile-time growth.

Run expression CSE before simplifying conditional min/max bounds so repeated
expressions remain shared through Let bindings.

The union formulas, used predicates, infinity handling, and unconditional
merge path are unchanged.

Regression tests

• conditional_box_merge: compiler-level reproducer with semantic checks.
• bounds_inference_specialized_spans: public-API pipeline reproducer with
  specialization-dependent RDom bounds and an independent pixel oracle.

Both tests have a 60-second timeout.

Validation

The two new regressions and bounds_internal were run on this branch using
Halide's ci-llvm-22 toolchain.

Initial verification against Halide v22.1.0 with LLVM 21.1.8:

• Both stock reproducers exceeded a 30-second limit.
• Patched box merging took approximately 0.07–0.09 seconds.
• Patched public-pipeline compilation took approximately 0.37–0.39 seconds.
• Thirteen selected upstream bounds/specialization correctness tests passed.

Full upstream, cross-platform, and GPU suites have not been run.

• Co-authored-by: Copilot App 223556219+Copilot@users.noreply.github.com

Supersedes #9660 to use the corrected branch name.

[alexreinking]

Have you tested this PR for performance regressions against the apps? CSE has a way of hiding necessary structure from the simplifier.

We really should have a bounds inference fuzzer...

[derek-gerstmann]

Have you tested this PR for performance regressions against the apps? CSE has a way of hiding necessary structure from the simplifier.

Yes, that's what triggered this ... generators in 22.0.0-dev444 that used to take 10-30sec started timing out after 1000's of seconds. Runtime with this fix was the same before the regression got introduced, but testing has only been on my own apps.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 73.01%. Comparing base (dd9d531) to head (6805c39).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #9661      +/-   ##
==========================================
+ Coverage   72.91%   73.01%   +0.10%     
==========================================
  Files         266      266              
  Lines       83341    83345       +4     
  Branches    20468    20470       +2     
==========================================
+ Hits        60765    60855      +90     
+ Misses      16156    16129      -27     
+ Partials     6420     6361      -59     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[alexreinking]

Runtime with this fix was the same before the regression got introduced, but testing has only been on my own apps.

Just to confirm: that's the runtime performance of the app, not compilation time, right?

[derek-gerstmann]

Runtime with this fix was the same before the regression got introduced, but testing has only been on my own apps.

Just to confirm: that's the runtime performance of the app, not compilation time, right?

Correct. I couldn't find a workaround outside of the compiler that didn't introduce a significant runtime performance regression. This PR avoids the compilation time, without impacting the original runtime performance for my specific test cases.