pre-commit: PR174355

[zyw-bot]

Link: llvm/llvm-project#174355
Requested by: @nikic

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@5e3b708
patch: llvm/llvm-project#174355
sha256: 8d7562257f0af2b8fead6627d8cf965954d768f582d632e086d096aecc316893
commit: 6dca834

2 files changed, 88 insertions(+), 88 deletions(-)

Improvements:
Regressions:
  correlated-value-propagation.NumNUW 570663 -> 570575 -0.02%

+0 meshlab/filter_screened_poisson.ll

[github-actions]

This patch applies the nsw (no signed wrap) flag to 64 add i32 instructions across the filter_screened_poisson IR function. All changes are localized to the BSplineEvaluationData class’s evaluation logic — specifically in centerValue, cornerValue, and OffsetToIndex methods — where integer additions feed into sign-extensions (sext i32 to i64) used for indexing.

The major changes are:

1. Addition overflow semantics tightened: Each modified add is now marked nsw, signaling that the operation is guaranteed not to overflow signed 32-bit integers. This enables LLVM optimizations (e.g., strength reduction, loop simplification, or elimination of redundant overflow checks) by allowing the optimizer to assume well-defined behavior under signed arithmetic.

2. Consistent pattern across B-spline index computations: The affected adds compute offset-based indices (e.g., %notmask.i.i.iXXX + %some_constant) used in multi-dimensional array or coefficient lookups. The constants (e.g., %65, %103, %945, %1621, etc.) are mostly small offsets or precomputed base indices — suggesting these additions model grid traversal or hierarchical spline evaluation (e.g., child node indexing in octree-like structures).

3. No change to control flow or data dependencies: All modifications are purely attribute additions to existing add instructions; no new instructions, branches, or operands are introduced. The sext and branching structure remain identical — confirming this is a semantic refinement, not a functional change.

4. Broad coverage across evaluator variants: Changes span both Evaluator and ChildEvaluator subroutines, indicating the optimization applies uniformly across different levels of the B-spline evaluation hierarchy — likely reflecting consistent invariants in the underlying mesh/point-cloud processing algorithm.

5. Likely result of improved range analysis or profile-guided optimization: The systematic application of nsw across dozens of similar patterns suggests either (a) a pass (e.g., SimplifyCFG, InstCombine, or LoopVectorize) has proven safe bounds for these adds, or (b) upstream C++ code was annotated (e.g., with __builtin_assume or llvm.assume) or compiled with -fwrapv-incompatible assumptions, enabling stricter IR inference.

In summary: This is a targeted, semantics-preserving optimization that strengthens integer overflow assumptions in B-spline index arithmetic — enabling downstream optimizations while preserving correctness under the established domain constraints.

model: qwen-plus-latest
CompletionUsage(completion_tokens=538, prompt_tokens=20331, total_tokens=20869, completion_tokens_details=None, prompt_tokens_details=None)