pre-commit: PR147605

[zyw-bot]

Link: llvm/llvm-project#147605
Requested by: @dtcxzyw

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@6ee3751
patch: llvm/llvm-project#147605
sha256: c56f12279753801b904c86e6d923a701276376202c55c58c71905efa72095c66
commit: c1f6a91

9 files changed, 43997 insertions(+), 43986 deletions(-)

Improvements:
  instcombine.NumSel 32254 -> 32255 +0.00%
Regressions:
  tailcallelim.NumAccumAdded 152 -> 151 -0.66%
  lower-is-constant-intrinsic.IsConstantIntrinsicsHandled 19999 -> 19910 -0.45%
  early-cse.NumCSECVP 104455 -> 104075 -0.36%
  sccp.NumInstReplaced 175637 -> 175094 -0.31%
  simple-loop-unswitch.NumSwitches 2031 -> 2026 -0.25%
  lower-expect-intrinsic.ExpectIntrinsicsHandled 485363 -> 484725 -0.13%
  dse.NumCFGChecks 649040 -> 648268 -0.12%
  local.NumRemoved 5610414 -> 5604374 -0.11%
  correlated-value-propagation.NumUDivURemsNarrowedExpanded 2829 -> 2826 -0.11%
  dse.NumCFGTries 58311 -> 58250 -0.10%

openusd/drawModeStandin.ll fail
php/string.ll fail
php/url.ll fail

7 4 bench/hermes/optimized/Passes.ll
2 3 bench/postgres/optimized/rewriteheap.ll
5 6 bench/recastnavigation/optimized/catch_amalgamated.ll
43 63 bench/ruby/optimized/string.ll
7 6 bench/yosys/optimized/stat.ll
3 3 bench/zed-rs/optimized/eiu35781qwj0wy44b83i3e7bt.ll

[github-actions]

Here is a high-level summary of up to 5 major changes in the provided LLVM IR diffs, ignoring formatting, comments, and non-semantic changes:

1. Addition of freeze for cttz result in Hermes pass logic
   In Passes.ll, the result of llvm.cttz.i32 is now frozen before use, ensuring consistent behavior in cases where the input may be zero or undefined. This improves robustness when computing bit indices from bitmask values.

2. Use of freeze on LHS of umin in Postgres checkpoint logic
   In rewriteheap.ll, the left-hand side operand of llvm.umin.i64 is now explicitly frozen (%8) before comparison. This ensures predictable behavior in presence of undefined or variable inputs.

3. Removal of redundant select chains based on null checks in Ruby string handling
   In several functions in string.ll, chains of select instructions that redundantly handled null pointers were simplified by removing intermediate selections, reducing complexity and improving clarity.

4. Fix of GEP inbounds attribute in Catch string parsing
   In catch_amalgamated.ll, a getelementptr instruction was changed from inbounds to non-inbounds form, possibly due to better analysis indicating that bounds checking wasn't necessary or safe in that context.

5. Improved PHI operand consistency in Yosys RTLIL IdString destructor logic
   In stat.ll, a PHI node that previously used %240 now uses a freeze of that value (%241) across all incoming edges, ensuring consistent SSA form and avoiding potential undefined behavior.

These changes primarily improve correctness, especially in edge cases involving undefined behavior, while also simplifying control flow and reducing redundancy.

model: qwen-plus-latest
CompletionUsage(completion_tokens=375, prompt_tokens=14807, total_tokens=15182, completion_tokens_details=None, prompt_tokens_details=None)