[abadams]

Copilot says:

This is an alternative to #9698.

On AArch64 with LLVM 22 and later, LLVM's VectorCombine pass (foldShuffleOfCastops) narrows shuffle(zext x) into zext(shuffle x) across the shuffles Halide uses to split vectors into native-width pieces. This produces illegal types like <4 x i8>, which lower to mov/zip/bic sequences instead of ushll2/umull2. For example, u32(u8) * 1000 and a plain u32(u8) both hit this. #9698 works around this in call_intrin, but that fix relies on cost-model details and misses the plain-zext case. LLVM 21 doesn't have the problem.

This PR skips VectorCombine on ARM targets only, using a pass instrumentation callback on Halide's PassBuilder. It doesn't use the global -disable-vector-combine option, so other targets compiled in the same process are unaffected. The change is small and should backport cleanly to releases pinned to an older LLVM.

Tested with LLVM 23:

• Both cases above now produce ushll/ushll2 + umull(2)/umlal(2), or tbl for the zext-only case.
• Arm apps codegen (camera_pipe, local_laplacian, lens_blur, resize, hannk, ...) with VectorCombine on vs. off: hot inner loops total +35 instructions out of ~74k, which is noise. The largest differences:
  • local_laplacian −33 and depthwise_conv −13.
  • The elementwise ops +27/+29, mostly scalar fmov/mov.
  • camera_pipe +52, all in the top-level function's buffer-validation prologue (the color-correction loop is 4 instructions shorter).
• x86 is unchanged. VectorCombine is still clearly useful there (disabling it costs +664 instructions on AVX2), so it stays enabled for x86.
• correctness_simd_op_check_arm passes, including new sshll2/ushll2 checks.

Checklist

• Tests added or updated (not required for docs, CI config, or typo fixes)
• Documentation updated (if public API changed)
• Python bindings updated (if public API changed)
• Benchmarks are included here if the change is intended to affect performance.
• Commits include AI attribution where applicable (see Code of Conduct)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.95%. Comparing base (f8d34e4) to head (c8047ac).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #9699      +/-   ##
==========================================
+ Coverage   72.89%   72.95%   +0.05%     
==========================================
  Files         266      266              
  Lines       83478    83483       +5     
  Branches    20497    20498       +1     
==========================================
+ Hits        60855    60905      +50     
- Misses      16173    16180       +7     
+ Partials     6450     6398      -52     

☔ 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.