AffineCFG: check lockstep dependences on the parallel check's relations

[wsmoses]

Lockstep check: use the same dependence relations as the parallel check.

isLoopMemoryLockStepExecutable, which the raiser asks before running a constant-count loop in lock step, checked dependences with MLIR's checkMemrefAccessDependence. Meanwhile isLoopMemoryParallel builds its own access relations through InvariantTerms:

• loop-invariant products of symbols are abstracted (AffineCFG: prove loops parallel through loop-invariant products of symbols #3369);
• domains stop at the affine scope and read bounds defined above it as symbols (AffineCFG: prove kernel loops parallel whatever surrounds the kernel #3380);
• the in-bounds accesses of a nest contribute facts (AffineCFG: use that a nest's accesses stay in bounds at its corners #3400);
• so do the conditional branches that dominate it within the function (AffineCFG: use what the checks on the way to a nest establish #3404, now on main). For example, a verify whose failing side never returns.

So an access whose index multiplies two symbols, or one under an scf.if, failed the lockstep check outright, even where the parallel check could analyze it.

The lockstep check now builds the same relations and calls checkAccessDependence, so all of these facts reach it too.

Where the guard facts reach lockstep. The raiser checks lockstep on outlined kernel functions. There, a host-side MFEM_VERIFY is not in the function being raised, so its facts reach only affine-cfg's parallel check, which runs before outlining. A guard inside the raised function does reach the lockstep check. As before, accesses to different memrefs, or two reads, have no dependence. Its rules for which dependences lockstep execution may break are unchanged.

Measured with affine-cfg, canonicalize and raising (enable_lockstep_for=true) on all 143 MFEM translation units. The raiser emits 14 fewer stablehlo.whiles (11,077 → 11,063), in bilininteg_dgtrace_ea (−5), bilininteg_mass_ea (−5), bilininteg_convection_ea (−3) and bilininteg_diffusion_ea (−1). No file gains any, and none fail.

Tests:

• New raising/affine_to_stablehlo_lockstep_invariant_terms.mlir (full-line goldens), which fails on main: a 3-trip loop copying at an index offset by a product of two symbols now runs in lock step, as one gather and one scatter, instead of a while.
• raising/affine_to_stablehlo_if_masking.mlir: @test_if_else_masking (per element, if (m[i]) a[i] *= 2 else a[i] *= a[i]) now runs in lock step as two whole-tensor selects instead of a while. That is correct: each iteration reads and writes only its own element.

Lit: every affinecfg* test, plus the raising tests above, against this branch rebased on main (which includes #3404 and #3406).

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 29.61%. Comparing base (9409410) to head (4b0f2f4).
⚠️ Report is 29 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3405   +/-   ##
=======================================
  Coverage   29.61%   29.61%           
=======================================
  Files         240      240           
  Lines       48460    48467    +7     
=======================================
+ Hits        14351    14354    +3     
- Misses      34109    34113    +4     

☔ 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.