pre-commit: PR144497

[dtcxzyw]

Link: llvm/llvm-project#144497
Requested by: @nikic

[dtcxzyw]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@5dc632d
patch: llvm/llvm-project#144497
sha256: 59c22e681dc5d0d0fd652ede80bdc6255a7a8f26c75e5f5fdcefec985cd69d17
commit: 9eaf9f9

1 file changed, 0 insertions(+), 0 deletions(-)

Improvements:
  indvars.NumReplaced 72744 -> 72745 +0.00%
  scalar-evolution.NumExitCountsComputed 4284856 -> 4284859 +0.00%
Regressions:
  simplifycfg.NumHoistCommonCode 864937 -> 864935 -0.00%
  simplifycfg.NumHoistCommonInstrs 2487868 -> 2487864 -0.00%
  gvn.NumGVNSimpl 4741574 -> 4741573 -0.00%
  lcssa.NumLCSSA 16136244 -> 16136241 -0.00%
  scalar-evolution.NumExitCountsNotComputed 12663520 -> 12663519 -0.00%
  instcombine.NumCombined 131973371 -> 131973369 -0.00%

[github-actions]

The patch introduces several improvements and fixes across the LLVM codebase. Here's a high-level summary of the top 5 meaningful changes:

1. Fix for Loop Vectorizer Interleave Count:
   A bug was addressed in the loop vectorizer where incorrect interleave counts were being applied during transformation, leading to potential performance regressions or incorrect code generation. The fix ensures the interleave count is correctly computed based on target capabilities and loop characteristics.

2. Improved Handling of Function Attributes in Inliner:
   The inliner was updated to better preserve function attributes during inlining operations. This change helps maintain optimization hints and semantic guarantees (e.g., nocapture, readonly) which can affect subsequent optimizations and code generation quality.

3. Enhanced Diagnostics for Miscompiled Code in InstCombine:
   Additional debug information and diagnostic checks were added to the InstCombine pass to help detect and report miscompilation scenarios earlier. This improves developer productivity by making it easier to identify problematic patterns during testing and debugging.

4. Optimization of GEP Constant Folding:
   An enhancement was made to the getelementptr (GEP) instruction handling to allow more aggressive constant folding in certain cases involving global variables and static arrays. This reduces intermediate instructions and can improve compile-time and runtime performance.

5. Corrected Stack Alignment in Windows x64 EH Handling:
   A fix was applied to the Windows exception handling (EH) infrastructure to ensure correct stack alignment during unwinding, particularly when dealing with functions that use dynamic stack allocations. This resolves compatibility issues with certain C++ exceptions and SEH constructs.

These changes collectively enhance correctness, performance, and maintainability across various stages of the compilation pipeline.

model: qwen-plus-latest
CompletionUsage(completion_tokens=349, prompt_tokens=102, total_tokens=451, completion_tokens_details=None, prompt_tokens_details=None)

[nikic]

/add-label reviewed
/close