pre-commit: PR185517

[zyw-bot]

Link: llvm/llvm-project#185517
Requested by: @andjo403

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@2c1594a
patch: llvm/llvm-project#185517
sha256: 900e83e1ca16f9ccbe4b1a6bea7804bae122869888ae7089e6dca069ecd4ded7
commit: f660110

26 files changed, 2114 insertions(+), 2330 deletions(-)

Improvements:
  instcombine.NumExpand 2628 -> 2630 +0.08%
  simplifycfg.NumSpeculations 393020 -> 393077 +0.01%
  correlated-value-propagation.NumSelects 228289 -> 228290 +0.00%
  instcombine.NumCombined 120233096 -> 120233361 +0.00%
  early-cse.NumCSE 5179463 -> 5179465 +0.00%
  instcombine.NumDeadInst 41840257 -> 41840258 +0.00%
Regressions:
  instcount.NumAndInst 3013907 -> 3013694 -0.01%
  loop-instsimplify.NumSimplified 182159 -> 182158 -0.00%
  simplifycfg.NumSinkCommonInstrs 823224 -> 823223 -0.00%
  func-properties-stats.NumTotalInstructionCount 277389064 -> 277388848 -0.00%
  instcount.TotalInsts 277389064 -> 277388848 -0.00%
  instcount.NumTruncInst 2935520 -> 2935518 -0.00%
  instcombine.NumSunkInst 3380861 -> 3380860 -0.00%
  instcount.NumPHIInst 12156310 -> 12156309 -0.00%

+0 meilisearch-rs/4llghrvh6vzci59o.ll
-1 pola-rs/4lt8y0jyzcqsx8vj5ff1je29c.ll
-1 pola-rs/8iz4w28nkww70r589pc3f77b4.ll
-1 rust-analyzer-rs/53ntv1pfq7fnns7g.ll
-2 gromacs/toppush.ll
-2 just-rs/2sblcsgax6v4zfcc.ll
-2 pola-rs/1ghxbvipwjb7mvkoqdu3h9hr4.ll
-3 rust-analyzer-rs/202uzsghjvp73kcw.ll
-6 pola-rs/3p63alsz0cwi8dlet1k7lar5q.ll
-6 pola-rs/8mjg4tlt628xep4awqt0660mt.ll
-8 pola-rs/55d4mb2xubgpcc6e6wba386pp.ll
-8 pola-rs/druded1wlaptuxbrsbyp1cgur.ll
-9 pola-rs/60trv14ccmssmdmvmdu6vrxge.ll
-11 delta-rs/2braxl0lj34anf5z.ll
-11 pola-rs/3f4qmpfpjkjad9wxa0v7zko9x.ll
-24 pola-rs/59h0xctzreekd1dj94rqcuqk4.ll
-31 pola-rs/3ll6ckwmxs5b00p4hfj8q4ugk.ll
-32 pola-rs/4lreosyeqk7o1vd9fcfoxznlc.ll
-37 pola-rs/5eb1qorc93yw2vdp0woub3vux.ll

[github-actions]

This patch contains a series of targeted, mechanical optimizations across multiple Rust benchmarks (delta-rs, gromacs, just-rs, meilisearch-rs, pola-rs, rust-analyzer-rs), primarily focused on improving the code generation for GenericShunt-related iterator adapters and related control flow. The major changes are:

1. Simplified Null-Check Logic in GenericShunt Return Construction:
   In dozens of GenericShunt functions (e.g., _ZN106_$LT$core..iter..adapters..GenericShunt...), the pattern for constructing the {ptr, ptr} return value was optimized. The original sequence used an icmp ne followed by an and with a condition (%trunc.i) and then a select, which was redundant. It has been replaced with a single icmp eq + select, eliminating the unnecessary and and simplifying the control logic. This improves both readability and potential downstream optimization opportunities.

2. Refinement of Exception Handling and Unwind Block Targets:
   In gromacs/toppush.ll, several invoke instructions had their to label successors updated to point to newly renamed or reordered basic blocks (e.g., %170 instead of %171). Similarly, phi node incoming block references were updated to match new predecessor labels (e.g., %174 → %172). These are consistent with IR reordering or cleanup passes and ensure correct exception propagation semantics.

3. Cleanup of Redundant Truncation and Boolean Folding:
   In just-rs/2sblcsgax6v4zfcc.ll, a now-unneeded trunc nuw i64 %12 to i1 instruction was removed. Its use was replaced by direct boolean composition (or i1 %.demorgan.not, %17) — indicating earlier folding or elimination of the truncation, reducing instruction count and register pressure.

4. Consistent Phi Node and Invoke Target Updates Across meilisearch-rs:
   Multiple phi nodes and invoke unwind/to targets in meilisearch-rs/4llghrvh6vzci59o.ll were updated to reference corrected predecessor blocks (e.g., %747 instead of %748, %702 instead of %703). These reflect internal CFG restructuring (e.g., loop exit splitting, thread merging) and ensure accurate SSA form and exception handling.

5. Elimination of Redundant Intermediate Selects and Phi Inputs:
   Several locations (e.g., meilisearch-rs, pola-rs) replace multi-step selects (e.g., and i1 %A, %B; select i1 %and_result, ...) with direct comparisons and selects. Additionally, some phi nodes had inputs pruned or relabeled to reflect simplified control flow — e.g., removing unreachable or duplicated predecessors — tightening the IR.

All changes are low-level, correctness-preserving transformations typical of LLVM’s optimization pipeline (e.g., InstCombine, SimplifyCFG, SROA). No semantic changes or high-level algorithmic modifications are present.

model: qwen-plus-latest
CompletionUsage(completion_tokens=686, prompt_tokens=73309, total_tokens=73995, completion_tokens_details=None, prompt_tokens_details=None)

[andjo403]

/reviewed

[andjo403]

/close