pre-commit: PR183082

[zyw-bot]

Link: llvm/llvm-project#183082
Requested by: @SavchenkoValeriy

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@2e40d2b
patch: llvm/llvm-project#183082
sha256: 47245178c16d1833d47ceac6065946c14c6fed8281842455c6363765f1fe3443
commit: 46e9730

891 files changed, 270695 insertions(+), 270382 deletions(-)

Improvements:
  licm.NumBOAssociationsHoisted 3463 -> 7217 +108.40%
  licm.NumHoisted 5265744 -> 5269024 +0.06%
  correlated-value-propagation.NumSubNSW 77839 -> 77848 +0.01%
  instcombine.NegatorNumValuesVisited 21488110 -> 21490204 +0.01%
  instcombine.NegatorTotalNegationsAttempted 20572494 -> 20574477 +0.01%
  sccp.NumInstRemoved 1926014 -> 1926180 +0.01%
  correlated-value-propagation.NumCmps 270347 -> 270359 +0.00%
  correlated-value-propagation.NumSubNW 114432 -> 114437 +0.00%
  instcombine.NumDeadInst 40975323 -> 40977084 +0.00%
  scalar-evolution.NumExitCountsNotComputed 11927133 -> 11927596 +0.00%
Regressions:
  licm.NumAddSubHoisted 1907 -> 1903 -0.21%
  reassociate.NumChanged 4922532 -> 4921506 -0.02%
  correlated-value-propagation.NumMinMax 10113 -> 10111 -0.02%
  licm.NumGEPsHoisted 23765 -> 23761 -0.02%
  correlated-value-propagation.NumSubNUW 37055 -> 37051 -0.01%
  indvars.NumElimExt 298039 -> 298030 -0.00%
  instcombine.NumSunkInst 3383726 -> 3383636 -0.00%
  instcount.NumSubInst 1915428 -> 1915399 -0.00%
  instsimplify.NumReassoc 813644 -> 813637 -0.00%
  sccp.NumInstReplaced 131997 -> 131996 -0.00%

+9 velox/FirstLastValue.ll
+8 nlohmann_json/unit-json_pointer.ll
+8 nlohmann_json/unit-regression2.ll
+8 sentencepiece/bpe_model_trainer.ll
+8 sentencepiece/builder.ll
+8 z3/model.ll
+8 z3/theory_seq.ll
+7 cmake/cmSystemTools.ll
+6 linux/dmi_scan.ll
+6 stockfish/thread.ll
+4 abseil-cpp/fnmatch.ll
+4 abseil-cpp/str_split.ll
+4 abseil-cpp/stripping_test.ll
+4 cmake/cmQtAutoGenerator.ll
+4 mold/glob.ll
+4 nix/remote-store.ll
+4 node/libnode.node_contextify.ll
+4 pocketpy/io.ll
+4 velox/DateTimeFormatter.ll
+4 velox/StringFunctions.ll
+4 yalantinglibs/EnumFieldGenerator.ll
+3 clamav/bytecode_api.ll
+3 faiss/PolysemousTraining.ll
+3 ffmpeg/hlsenc.ll
+3 hyperscan/tamarama.ll
+3 libquic/ssl_test.ll
+3 libquic/string16.ll
+3 linux/extents.ll
+3 llama.cpp/ggml-opt.ll
+3 llvm/BitcastBuffer.ll
+3 openspiel/tiny_bridge.ll
+3 php/shared_alloc_mmap.ll
+3 php/zend_opcode.ll
+3 postgres/parse_relation.ll
+3 wireshark/packet-gsm_a_bssmap.ll
+1 ffmpeg/bmvvideo.ll
+0 abc/cecSynth.ll
+0 abseil-cpp/spinlock_test_common.ll
+0 arrow/UriQuery.ll
+0 arrow/datetime.ll
+0 arrow/uri.ll
+0 clamav/pe.ll
+0 coreutils-rs/4dx3xgc2q0yp2q7n.ll
+0 curl/tool_cb_dbg.ll
+0 darktable/identify_tools.ll
+0 eastl/EATextUtil.ll
+0 fmt/format-impl-test.ll
+0 git/ref-cache.ll
+0 gromacs/mpiinfo.ll
+0 grpc/http_filters_plugin.ll
+0 hdf5/H5HFdbg.ll
+0 hyperscan/rose_build_convert.ll
+0 influxdb-rs/1s5ibqekmzia6ibu.ll
+0 just-rs/4mdvpwvrpdu4jonv.ll
+0 libcxx/nodebug_on_aliases.ll
+0 libcxx/string.ll
+0 lief/AArch64PAuth.ll
+0 linux/ohci-hcd.ll
+0 meilisearch-rs/2bqmf34tdvo23w8l.ll
+0 meshlab/arap.ll
+0 meshoptimizer/vertexfilter.ll
+0 minetest/dungeongen.ll
+0 open3d/AdvancedIndexing.ll
+0 open3d/FileGLTF.ll
+0 openexr/parse_header.ll
+0 openjdk/hb-ot-shaper-khmer.ll
+0 openssl/bn_lib.ll
+0 openssl/cipher_aes_cbc_hmac_sha1_hw.ll
+0 openusd/aom_scale.ll
+0 openusd/openexr-c.ll
+0 postgres/inv_api.ll
+0 protobuf/importer.ll
+0 protobuf/php_generator.ll
+0 redis/t_hash.ll
+0 ruff-rs/8rgebevtg1rphkf0xv4zhooko.ll
+0 rustfmt-rs/3xcdaapyewyrfogi.ll
+0 slurm/gres_sock_list.ll
+0 tev/Image.ll
+0 typst-rs/5z4no3nnr5v1s13.ll
+0 vcpkg/format.ll
+0 velox/VectorFuzzer.ll
+0 wireshark/packet-ipsec.ll
+0 wireshark/packet-xmcp.ll
+0 wolfssl/sp_int.ll
+0 z3/upolynomial.ll
+0 zed-rs/5qf95fx0pqn6duuq4hkly3hl0.ll
-1 opencv/tldEnsembleClassifier.ll
-2 rust-analyzer-rs/36cskjvhktchzlnh.ll
-2 wireshark/androiddump.ll

[github-actions]

This patch applies a set of loop-invariant code motion (LICM) and arithmetic strength-reduction optimizations across many LLVM IR files. The core transformation is the extraction of invariant subtractions—specifically expressions of the form x - c or c - x where c is constant and x is loop-invariant—and replacing them with precomputed invariant.op = ... values used later in loop bodies.

The major changes are:

1. Loop-Invariant Subtraction Hoisting: In numerous loops, subtraction operations involving loop-invariant operands (e.g., %397 - %.val203, %397 - %6, %258 - %223) are hoisted out of loops and computed once into new %invariant.op variables. This eliminates redundant arithmetic on every iteration.

2. Strength Reduction of Indexing Arithmetic: Many sub + add sequences (e.g., sub i32 %a, %b; add i32 %..., 1) are replaced by a single add using the precomputed invariant (e.g., add i32 %a, %invariant.op). This reduces instruction count and improves ILP.

3. Control-Flow Graph (CFG) Restructuring for LICM: Several functions introduce new preheader blocks (e.g., .lr.ph.split.preheader, .preheader5.preheader, .lr.ph149.preheader) to enable safe hoisting of invariants before loop entry, ensuring correctness under all control paths.

4. Phi Node Updates & Block Renaming: PHI nodes in loop headers and exits are updated to reflect new predecessor blocks (e.g., changing %59 → %54 or %60 → %55), and branch targets are adjusted accordingly (e.g., br i1 ..., label %36 instead of %39). These are mechanical but critical for maintaining SSA form after CFG changes.

5. Consistent Pattern in String/Buffer Search Loops: A recurring pattern appears across fnmatch.ll, str_split.ll, stripping_test.ll, cmQtAutoGenerator.ll, nlohmann_json.ll, etc.: the memchr call’s length argument was previously computed as (%len - %offset) + 1; now it's (%len + %invariant.op) where %invariant.op = 1 - %offset. This simplifies the address arithmetic and exposes more optimization opportunities.

These changes collectively improve performance by reducing loop-carried dependencies, decreasing dynamic instruction count, and enabling further downstream optimizations like vectorization or better register allocation.

model: qwen-plus-latest
CompletionUsage(completion_tokens=557, prompt_tokens=112845, total_tokens=113402, completion_tokens_details=None, prompt_tokens_details=None)

[nikic]

Looks like we have a bunch of regressions like this where the icmp is no longer optimized away...