pre-commit: PR184246

[zyw-bot]

Link: llvm/llvm-project#184246
Requested by: @andjo403

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@ebe3c1e
patch: llvm/llvm-project#184246
sha256: 9e45465e198cff46a18fadba650560618613935a529fe902d5343a5a5ed6186f
commit: 3c8d9be

7 files changed, 7810 insertions(+), 7855 deletions(-)

Improvements:
  aggressive-instcombine.NumAnyOrAllBitsSet 70 -> 79 +12.86%
  indvars.NumInvariantCmp 335 -> 337 +0.60%
  simple-loop-unswitch.NumCostMultiplierSkipped 17241 -> 17243 +0.01%
  instcombine.NumFactor 46658 -> 46662 +0.01%
  instcount.NumFreezeInst 55945 -> 55947 +0.00%
  loop-instsimplify.NumSimplified 182375 -> 182381 +0.00%
  indvars.NumReplaced 69496 -> 69498 +0.00%
  simple-loop-unswitch.NumBranches 104946 -> 104948 +0.00%
  correlated-value-propagation.NumAddNUW 210865 -> 210868 +0.00%
  instcombine.NumReassoc 277696 -> 277698 +0.00%
Regressions:
  reassociate.NumChanged 4922851 -> 4922720 -0.00%
  instcount.NumOrInst 1142633 -> 1142603 -0.00%
  instcount.NumLShrInst 1229099 -> 1229069 -0.00%
  instcount.NumTruncInst 2935508 -> 2935490 -0.00%
  instcount.NumXorInst 539999 -> 539996 -0.00%
  local.NumPHICSEs 185363 -> 185362 -0.00%
  instcount.NumShlInst 1394878 -> 1394874 -0.00%
  simplifycfg.NumSinkCommonCode 376289 -> 376288 -0.00%
  loop-rotate.NumInstrsDuplicated 3015029 -> 3015021 -0.00%
  simplifycfg.NumSinkCommonInstrs 823145 -> 823143 -0.00%

-3 llvm/SMTConstraintManager.ll
-26 fish-rs/2jegnuo3b7uy1oojff5kt8alk.ll

[github-actions]

Here's a concise summary of the major changes in this LLVM IR diff:

1. Simplified bit-field extraction and condition logic in _ZN4fish12input_common16InputEventQueuer9parse_csi...:
   The original code performed multiple lshr/trunc operations to extract individual byte fields from an i40 value (%.sroa.3.0.extract.shift21, %.sroa.6.0.extract.shift23, etc.), followed by complex boolean logic involving or, xor, and icmp ne. The patch eliminates most of those extractions and replaces the condition with a single masked and i40 %9, 4311744769 (≈ 0x100A0001) and icmp eq against 4294967296 (0x10000000), significantly reducing instruction count and control flow complexity.

2. Elimination of redundant phi node inputs and basic blocks:
   The .sroa.414.0 phi node is now fed only from %3 and %15 (previously %3 and %17), and the branch target labels are renumbered and merged (e.g., %16 → %12, %17 → %15). This reflects CFG simplification — likely due to dead code elimination or more aggressive jump-threading after the logic rewrite.

3. Optimized bit-packing for KeyEvent::with_shifted_codepoint:
   The multi-step or disjoint sequence building a 32-bit value from shifted/truncated fields is replaced with a single and i32 %13, -16711681 (i.e., 0xFF00FFFF) followed by or disjoint with the shifted i8→i32 value. This preserves the intended bit layout (masking out bits 8–15) while removing 5+ intermediate instructions and unnecessary truncations/shifting.

4. Consistent transformation across duplicated function bodies:
   Identical optimizations are applied to two nearly identical copies of the same function (at offsets ~39582 and ~39631), indicating the change targets a common code pattern—likely resulting from inlining or template instantiation—and was applied uniformly post-optimization.

5. Cleanup in SMTConstraintManager::checkModel:
   In SMTConstraintManager.ll, the unused lshr i16 %25, 8 (.sroa.5.0.extract.shift) is removed. The branching logic around %28 == 257 / == 256 is streamlined: redundant trunc/xor/and chains are replaced with direct and+icmp pairs, and the CFG is flattened (e.g., %35 → %33, %39 → %37), reducing block count and improving predictability.

Overall, these changes reflect aggressive peephole and DAG-combining optimizations—replacing verbose, multi-instruction bit manipulations with compact, semantically equivalent patterns—leading to smaller, faster, and more analyzable IR.

model: qwen-plus-latest
CompletionUsage(completion_tokens=711, prompt_tokens=4756, total_tokens=5467, completion_tokens_details=None, prompt_tokens_details=None)

[andjo403]

/add-label reviewed

[andjo403]

/close