pre-commit: PR165975

[zyw-bot]

Link: llvm/llvm-project#165975
Requested by: @dtcxzyw

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@c35a726
patch: llvm/llvm-project#165975
sha256: 816126057ae146f737a6d2c8085ac0dc931e86178e96175da6ddd200f6ee2dd8
commit: 52ebb98

219 files changed, 80863 insertions(+), 80998 deletions(-)

Improvements:
  bdce.NumSimplified 5815 -> 5816 +0.02%
  instcount.NumAndInst 3015055 -> 3015165 +0.00%
  correlated-value-propagation.NumShlNUW 152313 -> 152315 +0.00%
  correlated-value-propagation.NumShlNW 268748 -> 268751 +0.00%
  correlated-value-propagation.NumShlNSW 116435 -> 116436 +0.00%
  instcount.NumOrInst 1142721 -> 1142729 +0.00%
  instcombine.NumConstProp 155104 -> 155105 +0.00%
  bdce.NumRemoved 378277 -> 378279 +0.00%
  instcount.NumSubInst 1915033 -> 1915040 +0.00%
  early-cse.NumCSE 5181730 -> 5181745 +0.00%
Regressions:
  indvars.NumSameSign 40231 -> 40226 -0.01%
  instcount.NumShlInst 1394829 -> 1394676 -0.01%
  correlated-value-propagation.NumAddNSW 215589 -> 215567 -0.01%
  instsimplify.NumExpand 291246 -> 291219 -0.01%
  correlated-value-propagation.NumAddNW 426447 -> 426421 -0.01%
  instcombine.NumDeadInst 40959700 -> 40957453 -0.01%
  instcount.NumLShrInst 1229089 -> 1229027 -0.01%
  correlated-value-propagation.NumNSW 564262 -> 564241 -0.00%
  instcombine.NumReassoc 277247 -> 277238 -0.00%
  instsimplify.NumReassoc 813651 -> 813630 -0.00%

+3 wasmtime-rs/4bsmuvpz9r22ks1w.ll
+2 eastl/EACallback.ll
+1 bdwgc/gc.ll
+1 duckdb/ub_duckdb_storage_compression_dictionary.ll
+1 harfbuzz/hb-ot-cff1-table.ll
+1 harfbuzz/hb-subset-plan.ll
+1 hermes/HadesGC.ll
+1 lean4/ByteArray.ll
+1 openjdk/hb-ot-cff1-table.ll
+1 openjdk/hb-ot-cff2-table.ll
+1 openjdk/hb-ot-font.ll
+1 postgres/spell.ll
+1 rocksdb/memtable.ll
+0 abc/amapUniq.ll
+0 abc/mapperTree.ll
+0 abc/satSolver.ll
+0 bullet3/btDeformableBodySolver.ll
+0 c3c/semantic_analyser.ll
+0 cmake/System.ll
+0 cmake/cmJSONState.ll
+0 darktable/FujiDecompressor.ll
+0 darktable/aspect_ratio.ll
+0 folly/Benchmark.ll
+0 folly/EventBase.ll
+0 folly/EventBaseLocal.ll
+0 folly/ThreadedExecutor.ll
+0 graphviz/emit.ll
+0 graphviz/node.ll
+0 image-rs/4srzh4wujeew249y.ll
+0 libigl/intersect_other.ll
+0 libigl/projected_cdt.ll
+0 opencv/gfluidimgproc_func.dispatch.ll
+0 opencv/stereosgbm.ll
+0 openssl/kmac_prov.ll
+0 openssl/wrap128.ll
+0 postgres/slru.ll
+0 proxygen/HTTP2PriorityQueue.ll
+0 proxygen/HTTPSession.ll
+0 ruby/hash.ll
+0 wireshark/packet-dcerpc-netlogon.ll
-1 icu/ucnv_bld.ll
-1 libwebp/lossless_sse2.ll
-1 linux/tree.ll
-1 linux/xstate.ll
-1 llvm/InstructionCombining.ll
-1 llvm/StatepointLowering.ll
-1 luau/IrRegAllocX64.ll
-1 sdl/SDL_sysjoystick.ll
-1 wasmtime-rs/1bz8ofzmdclmzhoz.ll
-2 ffmpeg/cavs.ll
-2 ffmpeg/dvbsubdec.ll
-2 ffmpeg/mobiclip.ll
-2 libquic/stack.ll
-2 openspiel/dynamic_routing.ll
-2 openspiel/spiel.ll
-3 linux/intel_vdsc.ll
-3 llvm/ScaledNumber.ll
-3 pbrt-v4/color.ll
-3 postgres/gram.ll
-4 ffmpeg/webp.ll
-5 ruff-rs/9ezhgv3vaoku7b96fwwr4f701.ll
-7 libphonenumber/geocoding_data_test.ll
-7 openspiel/best_response.ll
-8 abc/mpmPre.ll
-8 darktable/ArwDecoder.ll
-9 libphonenumber/generate_geocoding_data.ll

[github-actions]

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

1. Reordering of bit manipulation operations: In multiple files (e.g., amapUniq.ll, mapperTree.ll, satSolver.ll, folly/*.ll), sequences of and, shr, and add operations were reordered — typically moving and after arithmetic or combining shifts with masks — preserving semantics while improving instruction scheduling or enabling later optimizations.

2. Refinement of pointer arithmetic and memory alignment logic: Several benchmarks (cmake/System.ll, eastl/EACallback.ll, bdwgc/gc.ll, duckdb/...ll) replaced naive pointer subtraction/addition with explicit masking (e.g., and i64 ..., -8 or -16) to compute aligned offsets, improving safety and clarity of buffer size calculations and memmove/memcpy bounds.

3. Fixes and simplifications in phi node operand matching: Across many files (mpmPre.ll, graphviz/emit.ll, libphonenumber/...ll, llvm/InstructionCombining.ll), phi node incoming value blocks were corrected to reference the actual predecessor basic blocks, not stale or mislabeled ones (e.g., %20 → %19, %98 → %97). This ensures correct SSA form and dataflow.

4. Optimization of color/alpha channel arithmetic: In image-processing code (dvbsubdec.ll, image-rs/...ll, webp.ll, libwebp/lossless_sse2.ll), RGB/ARGB bitfield extraction and blending was restructured — e.g., replacing lshr + and pairs with direct and-then-add patterns — to reduce instruction count and better expose vectorization opportunities (notably shifting from /2 to /4 in averaging).

5. Cleanup of dead or redundant control flow: Several functions (aspect_ratio.ll, harfbuzz/...ll, icu/ucnv_bld.ll) removed unnecessary basic blocks (e.g., 132:, 154:), merged switch cases, or simplified conditionals (e.g., replacing icmp ult with icmp eq after masking), reducing CFG complexity without altering behavior.

These changes reflect ongoing compiler-driven optimization passes (e.g., InstCombine, SimplifyCFG, LoopVectorizer) and manual tuning for correctness, performance, and maintainability — especially around low-level bit manipulation, memory layout, and SSA integrity.

model: qwen-plus-latest
CompletionUsage(completion_tokens=531, prompt_tokens=119999, total_tokens=120530, completion_tokens_details=None, prompt_tokens_details=None)