pre-commit: PR146736

[zyw-bot]

Link: llvm/llvm-project#146736
Requested by: @dtcxzyw

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@e324392
patch: llvm/llvm-project#146736
sha256: 049306b3e5d8228ab37bd2d7d589e9c44c21ac8b77ebf517e5a30a64e75de965
commit: 3de52cf

3 files changed, 10137 insertions(+), 9922 deletions(-)

Improvements:
  early-cse.NumDSE 197393 -> 197399 +0.00%
  globaldce.NumFunctions 368555 -> 368559 +0.00%
  instcombine.NumSunkInst 3564518 -> 3564546 +0.00%
  dse.NumFastStores 1199535 -> 1199541 +0.00%
  early-cse.NumCSELoad 7690407 -> 7690443 +0.00%
  dse.NumGetDomMemoryDefPassed 1410063 -> 1410069 +0.00%
  memdep.NumCacheNonLocalPtr 284131355 -> 284132434 +0.00%
  capture-tracking.NumCapturedBefore 17797514 -> 17797538 +0.00%
  memdep.NumUncacheNonLocalPtr 269977198 -> 269977529 +0.00%
  early-cse.NumCSEGEP 14069017 -> 14069029 +0.00%
Regressions:
  loop-rotate.NumInstrsDuplicated 3186199 -> 3186187 -0.00%
  reassociate.NumChanged 5220676 -> 5220658 -0.00%
  mem2reg.NumPHIInsert 13818916 -> 13818886 -0.00%
  instcombine.NumDeadInst 44226754 -> 44226663 -0.00%
  sroa.NumDeleted 282108450 -> 282108168 -0.00%
  sroa.NumAllocaPartitionUses 288414941 -> 288414701 -0.00%
  sroa.NumNewAllocas 29990864 -> 29990846 -0.00%
  instcombine.NumCombined 130186196 -> 130186141 -0.00%
  inline.NumDeleted 17009651 -> 17009645 -0.00%
  inline.NumInlined 44571441 -> 44571429 -0.00%

95 93 bench/nuklear/optimized/unity.ll

[github-actions]

Here is a summary of the major changes in this LLVM IR diff:

1. Branch Target Modification: A conditional branch (br i1) was updated to target a different label (%350 instead of %348). This likely reflects control flow graph restructuring or improved optimization decisions.

2. Floating-Point Logic Restructuring: The sequence of floating-point operations and vector insertions before the call to nk_draw_text was restructured. Specifically, the logic computing %339, %340, and related values was rewritten to include a comparison with zero and clamping behavior via select. This could represent strength reduction or improved handling of edge cases.

3. Control Flow Reorganization: The layout and numbering of several basic blocks were changed. Notably, labels like %348, %351, %364, etc., were replaced or reordered into %350, %353, %366, etc. This suggests a reorganization of control flow, possibly due to improved switch lowering or block merging.

4. Memory Alignment Fix: Several store instructions were realigned. For example, a store previously aligned to 8 bytes now targets a pointer with 4-byte alignment. These indicate minor but important adjustments for correctness or performance.

5. Register Allocation or PHI Node Adjustments: PHI nodes such as those in %375 now reference updated incoming labels and registers (e.g., %353, %366, %370, %350). This indicates changes in how values are merged at control-flow joins, potentially improving register allocation or reducing spills.

High-Level Overview:

The patch primarily consists of control-flow simplifications, floating-point arithmetic improvements, and minor alignment/memory fixes. It appears to be the result of further optimization passes that cleaned up conditionals, adjusted branching logic, and optimized data flow through basic blocks. The overall structure of the function remains similar, but the changes should improve runtime efficiency and correctness under edge cases.

model: qwen-plus-latest
CompletionUsage(completion_tokens=423, prompt_tokens=6552, total_tokens=6975, completion_tokens_details=None, prompt_tokens_details=None)