Skip to content

Conversation

@zyw-bot
Copy link
Collaborator

@zyw-bot zyw-bot commented Dec 5, 2025

@github-actions github-actions bot mentioned this pull request Dec 5, 2025
@zyw-bot
Copy link
Collaborator Author

zyw-bot commented Dec 5, 2025

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@8f9ef4b
patch: llvm/llvm-project#167771
sha256: d9e2a1e483f2ac737fe6b04cce39282f98ad4f2cae8fda32b872513484db7805
commit: 92858bf

145 files changed, 93096 insertions(+), 93243 deletions(-)

Improvements:
  simplifycfg.NumBitMaps 2245 -> 2246 +0.04%
  simplifycfg.NumLookupTablesHoles 2393 -> 2394 +0.04%
  correlated-value-propagation.NumNonNull 10870047 -> 10874432 +0.04%
  mem2reg.NumDeadAlloca 1963403 -> 1964154 +0.04%
  sroa.NumVectorized 697857 -> 697970 +0.02%
  simple-loop-unswitch.NumCostMultiplierSkipped 17240 -> 17242 +0.01%
  correlated-value-propagation.NumNUW 571062 -> 571121 +0.01%
  correlated-value-propagation.NumAnd 44492 -> 44496 +0.01%
  instcombine.NumPHIsOfInsertValues 26429 -> 26431 +0.01%
  constmerge.NumIdenticalMerged 15670 -> 15671 +0.01%
Regressions:
  bdce.NumRemoved 384283 -> 383488 -0.21%
  vector-combine.NumVecLoad 4291 -> 4288 -0.07%
  gvn.NumGVNLoad 1320135 -> 1319296 -0.06%
  instcombine.NumConstProp 168277 -> 168176 -0.06%
  mem2reg.NumLocalPromoted 549700 -> 549535 -0.03%
  licm.NumBOAssociationsHoisted 3515 -> 3514 -0.03%
  sroa.NumLoadsPredicated 14459 -> 14456 -0.02%
  licm.NumLoadStorePromoted 62000 -> 61989 -0.02%
  local.NumPHICSEs 186111 -> 186080 -0.02%
  gvn.NumGVNSimpl 4415379 -> 4414646 -0.02%

+18 openvdb/Prune.ll
+2 open3d/Geometry3D.ll
+0 assimp/zip.ll
+0 ceres/block_sparse_matrix.ll
+0 cmake/cmComputeLinkDepends.ll
+0 faiss/Clustering.ll
+0 g2o/edge_project_xyz.ll
+0 g2o/optimizable_graph.ll
+0 g2o/optimization_algorithm_dogleg.ll
+0 g2o/slam2d_linear.ll
+0 hdf5/sio_perf.ll
+0 libigl/eigs.ll
+0 llvm/MsgPackDocument.ll
+0 meshlab/edit_align.ll
+0 meshlab/shell.ll
+0 nlohmann_json/diag_on.ll
+0 nori/accel.ll
+0 ockam-rs/1411u8drt798uxi8.ll
+0 opencv/gfluidbackend.ll
+0 pbrt-v4/integrator.ll
+0 pbrt-v4/integrators.ll
+0 php/zip.ll
+0 proj/adams.ll
+0 proj/airocean.ll
+0 proj/datum_set.ll
+0 proj/geod_set.ll
+0 proj/helmert.ll
+0 proj/horner.ll
+0 proj/info.ll
+0 proj/tmerc.ll
+0 quickjs/qjsc.ll
+0 quickjs/quickjs-libc.ll
+0 typst-rs/2d3c2n5y91mtl0x0.ll
+0 typst-rs/d6l9ieo9tcw33dn.ll
+0 wasmtime-rs/4ab4rlryc5h7bf6z.ll
-2 open3d/VoxelGrid.ll
-30 proj/param.ll
-31 typst-rs/2i78fvbm4wocuesi.ll

@github-actions
Copy link
Contributor

github-actions bot commented Dec 5, 2025

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

  1. Phi Node Predecessor Updates: Multiple phi instructions had their predecessor block references updated (e.g., .thread428.thread426, %72%73). These are likely due to control-flow restructuring or loop transformations, ensuring phi nodes correctly reference incoming edges.

  2. Vectorization & Scalar Replacement Improvements: Several functions (e.g., EdgeSE3ProjectXYZ::linearizeOplusEv, LocalIsometry) replaced scalar-to-vector bitcasts and shuffles with direct insertelement/shufflevector patterns using precomputed vector elements—improving clarity and potentially enabling better optimization.

  3. Loop Header & Critical Edge Adjustments: Loop induction variables and critical edge targets were updated across multiple files (e.g., %.lr.ph75%.lr.ph76, %.lr.ph149.i%.lr.ph148.i). This reflects loop canonicalization or unrolling changes affecting loop entry points and back-edges.

  4. Memory Access & Store Type Corrections: In ockam-rs and open3d, <4 x i8> loads/stores were replaced with i32, and double-to-<2 x double> conversions were refined (e.g., using extractelement + fmul instead of bitcast + shuffle). These align types with actual memory layout and improve codegen safety.

  5. Exception Handling & Landing Pad Cleanup: PHI nodes in landing pad blocks (e.g., %.pn222, %.pn23.pn) were updated to reference correct predecessor blocks after exception edge rewrites—ensuring accurate unwind state propagation during stack unwinding.

All changes are mechanical and correctness-preserving, consistent with backend optimizations like loop rotation, SROA refinement, vectorizer improvements, and CFG simplification.

model: qwen-plus-latest
CompletionUsage(completion_tokens=416, prompt_tokens=113219, total_tokens=113635, completion_tokens_details=None, prompt_tokens_details=None)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants