pre-commit: PR166935

[zyw-bot]

Link: llvm/llvm-project#166935
Requested by: @nikic

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@e5603da
patch: llvm/llvm-project#166935
sha256: 1a920680921980b04377b8b40e07c31fbe0a2f9939e63e21e402edd7973f027a
commit: 9699001

530 files changed, 2734063 insertions(+), 2820512 deletions(-)

Improvements:
  instcombine.NumConstProp 155480 -> 168277 +8.23%
  instcombine.NumDeadInst 41001373 -> 41027879 +0.06%
  instcombine.NumOneIteration 61993817 -> 61999680 +0.01%
  instcombine.NumWorklistIterations 61993817 -> 61999680 +0.01%
  globaldce.NumVariables 91234 -> 91238 +0.00%
  local.NumRemoved 5303564 -> 5303715 +0.00%
  simplifycfg.NumHoistCommonCode 839205 -> 839218 +0.00%
  simplifycfg.NumHoistCommonInstrs 2431936 -> 2431958 +0.00%
  instcombine.NumSunkInst 3381565 -> 3381590 +0.00%
  instcombine.NumCombined 120689279 -> 120689642 +0.00%
Regressions:
  last-run-tracking.NumSkippedPasses 48733699 -> 48727836 -0.01%
  instsimplify.NumSimplified 2467326 -> 2467295 -0.00%
  lcssa.NumLCSSA 15128878 -> 15128817 -0.00%
  sroa.NumAllocasAnalyzed 100370512 -> 100370508 -0.00%

+0 libigl/unproject.ll
-4 linux/af_netlink.ll
-4 linux/ioam6.ll
-4 linux/request.ll
-4 linux/sta_info.ll
-4 linux/util.ll
-4 openjdk/screencast_pipewire.ll
-4 postgres/jsonfuncs.ll
-6 lean4/Assumption.ll
-6 lean4/FVarSubset.ll
-6 lean4/FVarSubst.ll
-6 lean4/FilePath.ll
-6 lean4/InlineCandidate.ll
-6 lean4/MVarRenaming.ll
-6 lean4/Position.ll
-6 lean4/expr.ll
-6 lean4/local_ctx.ll
-6 lean4/sharecommon.ll
-8 abc/cuddZddLin.ll
-8 llvm/Z3CrosscheckVisitor.ll
-8 ruby/compile.ll
-9 lean4/Calc.ll
-9 lean4/tcp.ll
-9 meshlab/filter_parametrization.ll
-11 openjdk/methodData.ll
-12 lean4/AndFlatten.ll
-12 lean4/CaseValues.ll
-12 lean4/CollectAxioms.ll
-12 lean4/CollectFVars.ll
-12 lean4/ConstFolding.ll
-12 lean4/Database.ll
-12 lean4/Delta.ll
-12 lean4/EmbeddedConstraint.ll
-12 lean4/External.ll
-12 lean4/HasConstCache.ll
-12 lean4/InputFileConfig.ll
-12 lean4/InternalExceptionId.ll
-12 lean4/Propagate.ll
-12 lean4/Simp.ll
-12 lean4/TZdb.ll
-12 lean4/TypeName.ll
-12 lean4/WorkspaceConfig.ll
-12 lean4/lcnf.ll
-15 lean4/AutoBound.ll
-15 lean4/ExportAttr.ll
-15 lean4/Links.ll
-18 lean4/ByteArray.ll
-18 lean4/ForEachExprWhere.ll
-18 lean4/PlainDateTime.ll
-18 lean4/ReplaceLevel.ll
-18 lean4/Topological.ll
-18 libigl/vertex_triangle_adjacency.ll
-21 ruby/vm.ll
-24 lean4/AuxDeclCache.ll
-24 lean4/AuxLemma.ll
-24 lean4/CSE.ll
-24 lean4/CheckTactic.ll
-24 lean4/Client.ll
-24 lean4/FindLevelMVar.ll
-24 lean4/GeneralizeTelescope.ll
-24 lean4/InitShutdown.ll
-24 lean4/PrettyPrinter.ll
-24 lean4/RequestCancellation.ll
-24 lean4/inductive.ll
-27 wolfssl/sha.ll
-30 lean4/ElimDeadVars.ll
-30 lean4/Encode.ll
-30 lean4/IndGroupInfo.ll
-30 linux/fatent.ll
-36 lean4/Attach.ll
-36 lean4/EligibleHeaderDecls.ll
-57 pingora-rs/acx590v32pn6eo20biyujqd8g.ll

[github-actions]

1. Optimized reference counting logic in Lean4 IR: Replaced pointer-to-integer bit-checks (ptrtoint + and) with direct 32-bit refcount loads and comparisons, simplifying lean_dec_ref/lean_inc_ref sequences across dozens of Lean4 benchmark files (e.g., AndFlatten.ll, Assumption.ll, Attach.ll). This eliminates unnecessary pointer arithmetic and improves clarity.

2. Refactored hash computation in cuddZddLin.ll: Simplified the key hash calculation by replacing two separate (val << 1) | (ptr & 1) patterns with a single truncation + multiplication (trunc i64 → mul i32 25165834), then combining results with add instead of or disjoint. Reduces instruction count and improves arithmetic locality.

3. Fixed phi node operand indexing in loop exits: Corrected multiple .lcssa.sink and loop-exit phi nodes (e.g., in cuddZddLin.ll at .loopexit466, .loopexit465) to use updated predecessor block labels (e.g., %123, %158) instead of stale ones (%130, %172), ensuring correct SSA value propagation after loop restructuring.

4. Simplified conditional branching in Lean4 dec/ref helpers: Removed redundant ptrtoint/and null-checks before refcount updates, relying instead on direct icmp sgt i32 %ref, 1 checks and updating phi node predecessor lists accordingly (e.g., lean_dec.exit blocks now list only the relevant entry points, dropping unreachable or obsolete predecessors).

5. Consolidated memory access patterns in Lean4 box functions: Standardized field access (e.g., loading refcounts from offset 8 instead of 4, aligning with actual object layouts) and removed duplicate ptrtoint/and null checks before cold decrement paths, improving correctness and reducing code size across Array_mapMUnsafe_map, Array_foldlMUnsafe_fold, and similar boxed helper functions.

model: qwen-plus-latest
CompletionUsage(completion_tokens=457, prompt_tokens=98080, total_tokens=98537, completion_tokens_details=None, prompt_tokens_details=None)

[nikic]

/add-label reviewed
/close