EnzymeHLOOpt: ScatterIndicesAreUnique reads a constant without copying

[wsmoses]

ScatterIndicesAreUnique's constant-indices path built every index tuple through a recursive lambda copying a SmallVector per element, then a std::set of vectors — and since a scatter whose indices are not distinct makes the pattern fail, the whole enumeration ran again on every visit of that scatter by the greedy driver. It also read shape[indexVectorDim] for scalar indices (index_vector_dim == rank), one past the end.

On MFEM's ex1 (star mesh, order 3, PA) one exec-time module (PADiffusionSetup's, six scatters with a splat index constant of 20480x32x32x1) spent 56 s in it — the bulk of the 44 s run was this one optimization (perf: 50% of all samples in the pattern's SmallVector churn); the module now takes 0.1 s through the runtime pattern list.

Now: a splat is distinct iff there is one scatter point; otherwise the values are read once into a flat array, the tuples laid out point by point (the index-vector dimension handled whether real or implicit), and distinctness is a sort plus an adjacent compare.

Test scatter_indices_unique_constant.mlir: distinct, repeated, two-component tuples along the leading dimension, scalar indices (full-line goldens). On a 200k-element random permutation the verdicts match the old code.

🤖 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 29.96%. Comparing base (2d3ff23) to head (50e9dfb).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3439      +/-   ##
==========================================
+ Coverage   29.63%   29.96%   +0.33%     
==========================================
  Files         240      240              
  Lines       48496    48546      +50     
==========================================
+ Hits        14371    14548     +177     
+ Misses      34125    33998     -127     

☔ 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.

[wsmoses]

MFEM GPU unit suite (74 tests) with the runtime built from Enzyme-JAX main + #3437 #3439 #3440 #3441 #3442 and Reactant.jl #3423 + #3425, over objects from main + the open affine-cfg PRs (#3436 #3438 among them): 74/74 (sweep final55, 2026-10-08 17:38).

ex1 (Poisson, order 3, PA, PCG to 1e-12), timed solves in one process (JIT excluded), native CUDA for reference:

• star (185k dofs, 400 iterations): 0.64 s (native 0.047 s) — the morning's state ran ~70 iterations per 600 s
• fichera (802k dofs, 270 iterations): 1.08 s (native 0.14 s)