pre-commit: PR126423

[dtcxzyw]

Link: llvm/llvm-project#126423
Requested by: @nikic

[dtcxzyw]

runner: ariselab-64c-v2
baseline: llvm/llvm-project@3d14000
patch: llvm/llvm-project#126423
sha256: df7d4a34083494b12a196b32b7d759c87de009d14c4df0277524a247c9bc6814
commit: ea0861b

4 files changed, 161 insertions(+), 164 deletions(-)

Improvements:
  adce.NumRemoved 91012 -> 91015 +0.00%
  jump-threading.NumFolds 2086517 -> 2086518 +0.00%
  instcombine.NumCombined 99678365 -> 99678368 +0.00%
  last-run-tracking.NumSkippedPasses 46000114 -> 46000115 +0.00%
Regressions:
  correlated-value-propagation.NumAnd 37422 -> 37420 -0.01%
  correlated-value-propagation.NumSExt 37965 -> 37963 -0.01%
  simplifycfg.NumHoistCommonCode 608542 -> 608540 -0.00%
  correlated-value-propagation.NumPhis 994865 -> 994862 -0.00%
  simplifycfg.NumHoistCommonInstrs 1771199 -> 1771197 -0.00%
  instcombine.NumSunkInst 2723811 -> 2723809 -0.00%
  jump-threading.NumThreads 2236141 -> 2236140 -0.00%
  local.NumRemoved 3758933 -> 3758932 -0.00%
  gvn.NumGVNInstr 7678929 -> 7678928 -0.00%
  instcombine.NumDeadInst 33143068 -> 33143065 -0.00%

27 31 bench/llvm/optimized/X86RegisterBankInfo.ll
1 1 bench/meshoptimizer/optimized/indexcodec.ll
21 20 bench/opencv/optimized/find_ellipses.ll

[github-actions]

Summary of Changes in LLVM IR Diff

This patch introduces several modifications across three different LLVM IR files. Below is a summary of the major changes:

1. Optimization in X86RegisterBankInfo.ll:

   • Removed redundant instructions: The call to llvm.lifetime.start.p0 and related comparisons (icmp, select) have been eliminated, simplifying the control flow.
   • Restructured phi nodes: A new phi node %.pre-phi75 has been introduced to handle values from multiple predecessors, improving clarity and reducing redundancy.
   • Adjusted switch statements: The labels in the switch instruction for %43 have been reorganized, with some labels being removed or reordered (42, 43, etc.), aligning better with the updated control flow.
   • Simplified phi node merging: The final phi node %.0 now incorporates fewer preds, streamlining the logic and reducing unnecessary computations.

2. Truncation Optimization in indexcodec.ll:

   • Added nuw (No Unsigned Wrap) flag: In the truncation of %indvars.iv.i185 to an 8-bit integer, the nuw flag ensures that the operation will not result in unsigned overflow, enhancing precision and safety of the computation.

3. Control Flow Refinement in find_ellipses.ll:

   • Updated branching logic: Several branch instructions have been modified to use updated labels (226 instead of 225), ensuring consistency in the flow of execution.
   • Improved pointer arithmetic: The calculation involving %217 and %218 has been adjusted to use zext operations more effectively, potentially improving performance by avoiding unnecessary sign extensions.
   • Streamlined phi nodes: The phi node %.6.us now merges values from fewer predecessors, simplifying the data dependency graph and making the logic easier to follow.
   • Reordered label usage: Some labels have been reused or repurposed (226), reducing duplication and improving readability.

High-Level Overview

The overall goal of these changes appears to be optimization-focused, targeting both computational efficiency and code clarity. By removing redundant instructions, restructuring control flow, and refining data dependencies, the patch aims to improve the generated machine code's performance and maintainability. These modifications likely result in reduced instruction count, better register utilization, and improved runtime behavior when compiled to native code.

model: qwen-plus-latest
CompletionUsage(completion_tokens=508, prompt_tokens=4430, total_tokens=4938, completion_tokens_details=None, prompt_tokens_details=None)

[nikic]

Regression due to lack of sext vs zext nneg CSE.