pre-commit: PR157865

[zyw-bot]

Link: llvm/llvm-project#157865
Requested by: @XChy

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@b7e31e7
patch: llvm/llvm-project#157865
sha256: f53a051156046b61bdc722f6935c5e1b938d7edad4e4578b862dc6dd803b93f5
commit: 8007070

3 files changed, 5 insertions(+), 5 deletions(-)

Improvements:
Regressions:
  instcombine.NumCombined 125678718 -> 125657478 -0.02%
  correlated-value-propagation.NumNNeg 99658 -> 99656 -0.00%

4 4 bench/abseil-cpp/optimized/crc.ll
1 1 bench/rocksdb/optimized/xxhash.ll

[github-actions]

The patch introduces a single, consistent change across multiple LLVM IR files: the addition of the nneg (non-negative) attribute to zext (zero-extend) instructions.

Summary of Changes:

1. Addition of nneg flag to zext operations:
   In several locations across crc.ll (Abseil CRC implementation) and xxhash.ll (RocksDB XXH3 hash implementation), zext i32 %value to i64 is updated to zext nneg i32 %value to i64. This indicates that the input value is guaranteed to be non-negative, allowing the optimizer to make stronger assumptions about the value range.

2. Multiple instances in loop exit blocks:
   In crc.ll, four instances of this change occur in different preheader blocks (_exit.preheader, exit62.preheader, etc.), all involving values derived from an OR operation with 1 (or disjoint ..., 1). Since the input %i is OR’d with 1, it ensures the result is at least 1 (i.e., positive), justifying the nneg property.

3. One instance in XXH3 hash computation:
   In xxhash.ll, one zext is similarly updated in the ROCKSDB_XXH3_128bits function. The extended value (%101) comes from a bitwise XOR but is used in a context where its non-negativity can be proven, enabling the nneg annotation.

4. Optimization-enabling semantic hint:
   The nneg attribute does not change runtime behavior but informs LLVM’s optimizer that the extended value is ≥0. This can improve optimization outcomes, such as eliminating redundant checks or enabling better constant propagation and range analysis.

5. Pattern consistency across benchmarks:
   The changes reflect a broader pattern of inserting range-knowledge attributes into performance-critical hashing code, likely resulting from profile-guided optimization or manual IR refinement to expose optimization opportunities.

These modifications are low-level, correctness-preserving annotations that enhance optimization potential by expressing known value properties. They do not alter control flow or data computation but help the backend generate more efficient machine code.

model: qwen-plus-latest
CompletionUsage(completion_tokens=474, prompt_tokens=1260, total_tokens=1734, completion_tokens_details=None, prompt_tokens_details=None)