pre-commit: PR159583

[zyw-bot]

Link: llvm/llvm-project#159583
Requested by: @artagnon

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@d62505f
patch: llvm/llvm-project#159583
sha256: 62e30e2be0fb73d1af8dafb0db36a773ac7c99c301200e36730d8d114fc07bbf
commit: c3bd5c2

2386 files changed, 2978068 insertions(+), 2946009 deletions(-)

Improvements:
  simplifycfg.NumFoldBranchToCommonDest 772694 -> 784888 +1.58%
  indvars.NumElimCmp 57291 -> 57703 +0.72%
  licm.NumLoadStorePromoted 72269 -> 72756 +0.67%
  jump-threading.NumFolds 2649095 -> 2661337 +0.46%
  loop-delete.NumDeleted 115672 -> 116029 +0.31%
  loop-instsimplify.NumSimplified 188191 -> 188463 +0.14%
  loop-rotate.NumInstrsDuplicated 3110950 -> 3114784 +0.12%
  gvn.NumGVNSimpl 4583157 -> 4587461 +0.09%
  simplifycfg.NumSinkCommonCode 374688 -> 374983 +0.08%
  gvn.IsValueFullyAvailableInBlockNumSpeculationsMax 605616 -> 606009 +0.06%
Regressions:
  simple-loop-unswitch.NumSwitches 1968 -> 1873 -4.83%
  simplifycfg.NumFoldValueComparisonIntoPredecessors 522721 -> 505006 -3.39%
  adce.NumBranchesRemoved 23740 -> 23498 -1.02%
  loop-delete.NumBackedgesBroken 45366 -> 44954 -0.91%
  licm.NumPromotionCandidates 611777 -> 607455 -0.71%
  adce.NumRemoved 98698 -> 98373 -0.33%
  memdep.NumCacheDirtyNonLocalPtr 23851 -> 23777 -0.31%
  gvn.NumGVNBlocks 197697 -> 197232 -0.24%
  local.NumPHICSEs 198759 -> 198360 -0.20%
  gvn.NumGVNInstr 155515 -> 155237 -0.18%

+16 lean4/Weekday.ll
+13 z3/add_bounds_tactic.ll
+11 verilator/V3Branch.ll
+8 wolfssl/ssl.ll
+6 abc/fxuUpdate.ll
+4 abc/cbaNtk.ll
+4 abc/fxuSingle.ll
+4 abc/giaMini.ll
+4 abc/pdrIncr.ll
+4 cpython/_posixsubprocess.ll
+4 eastl/TestHash.ll
+4 lean4/Constraint.ll
+4 lean4/WHNF.ll
+4 libquic/heap_profiler_allocation_context_tracker.ll
+4 z3/elim_unconstrained.ll
+4 z3/macro_finder_tactic.ll
+3 boost/bind_stdin_stdout.ll
+3 eastl/TestFixedHash.ll
+3 hermes/Path.ll
+3 verilator/V3Clock.ll
+2 abc/fraMan.ll
+2 cpython/readline.ll
+2 duckdb/tcache.ll
+2 git/mv.ll
+2 hermes/GCBase.ll
+2 lean4/FunInd.ll
+2 openmpi/stacktrace.ll
+2 php/element.ll
+2 wireshark/packet-tls-utils.ll
+2 zed-rs/0fzu5iswc3qhp95jt69713e0a.ll
+1 casadi/feasiblesqpmethod.ll
+1 csmith/FactUnion.ll
+1 eastl/TestIntrusiveHash.ll
+1 folly/EpollBackend.ll
+1 open3d/VoxelGrid.ll
+1 openjdk/classLoader.ll
+1 openssl/ssl_lib.ll
+1 php/zend_signal.ll
+1 qemu/libvhost-user.ll
+1 typst-rs/3kgmqnxcsl3z3n0n.ll
+1 verilator/V3Common.ll
+1 wireshark/drange.ll
+1 wireshark/proto.ll
+1 zed-rs/bqkhvme8kwgfmpydn1hdkle55.ll
+1 zed-rs/etkkpijil810uxaq6a1xhpkj9.ll
+0 assimp/ASEParser.ll
+0 boost/async.ll
+0 box2d/TaskScheduler.ll
+0 clamav/optparser.ll
+0 coreutils-rs/2oz1n2pr10bv8jmi.ll
+0 crow/example.ll
+0 curl/cf-h1-proxy.ll
+0 icu/tznames_impl.ll
+0 opencv/perf_layer.ll
+0 yosys/yosys.ll
-1 cpython/_cursesmodule.ll
-1 darktable/tags.ll
-1 ffmpeg/xcbgrab.ll
-1 hermes/CodeMotion.ll
-1 hermes/InstSimplify.ll
-1 libsodium/utils.ll
-1 openjdk/method.ll
-1 openjdk/mutableNUMASpace.ll
-1 openjdk/threads.ll
-1 redis/pages.ll
-1 typst-rs/2i78fvbm4wocuesi.ll
-1 velox/MmapAllocator.ll
-1 zed-rs/etmluiftgqx5jpyivviyferrj.ll
-2 actix-rs/3sxbtmgc7txd1oqx.ll
-2 foundations-rs/7zdc97sa0ogd517y2re7ok0yw.ll
-2 influxdb-rs/55ksfxb2g8gj9lwh.ll
-2 lean4/TextSync.ll
-2 libquic/shared_memory_posix.ll
-2 lvgl/lv_fs.ll
-2 php/zend_inference.ll
-2 php/zend_shared_alloc.ll
-2 portaudio/pa_unix_util.ll
-2 ruby/thread.ll
-2 zed-rs/3xkvahxsem1hdyf3ef4aq8akj.ll
-3 graphviz/gvplugin.ll
-3 wireshark/addr_resolv.ll
-3 zed-rs/dmgo8iibk9y4hcmil8p9yl5yt.ll
-4 cpython/obmalloc.ll
-4 html5ever-rs/2p0p1zz6gwjy9c4w.ll
-5 nanobind/nb_func.ll
-6 php/zend_hash.ll
-9 node/libnode.node_large_page.ll
-10 cpython/signalmodule.ll
-12 duckdb/pages.ll
-12 freetype/ftbase.ll
-13 jemalloc/pages.ll
-14 openjdk/objectMonitor.ll
-33 mold/gdb-index.cc.X86_64.ll
-49 openssl/err.ll

[github-actions]

The provided diff shows a series of changes across multiple LLVM IR files, primarily involving transformations of pointer comparison and branching logic. Below is a summary of up to five major changes observed in the patch:

1. Replacement of ptrtoint + switch with icmp + br patterns:
   Multiple instances replace sequences that convert a pointer to an integer (ptrtoint) followed by a switch on magic values (e.g., 0, -1, 1, 2) with direct pointer comparisons using icmp eq/ne ptr and conditional branches (br i1). This improves type safety and avoids potential issues with pointer-to-integer conversions.

2. Refinement of null and sentinel pointer checks:
   Several functions previously used unsigned less-than comparisons (icmp ult ptr %ptr, inttoptr 2) to detect null or special sentinel pointers. These have been replaced with explicit equality checks against null and specific sentinels (e.g., inttoptr (i64 1 to ptr)), often combined with and/or conditions for clarity and correctness.

3. Introduction of combined condition checks using and/or of boolean expressions:
   Instead of chaining multiple basic blocks for individual pointer checks, the updated code combines conditions using and or or of icmp results. For example, checking both null and -1 is now done via %or.cond = or i1 %is_null, %is_minus_one, streamlining control flow.

4. Phi node predecessor list updates due to control flow restructuring:
   As branching logic is restructured (e.g., merging paths or changing successor blocks), phi nodes are updated to reflect new control flow predecessors. This includes reordering or adding/removing entries in phi instructions to maintain SSA form correctness.

5. Improved handling of tagged or sentinel pointers in data structures:
   In several benchmarks (e.g., Lean4, Eastl, Folly), the changes reflect better optimization around tagged pointers (common in functional or high-performance C++ code), replacing fragile arithmetic-based checks with safe pointer comparisons, which may enable further optimizations or improve debuggability.

These changes collectively suggest a refactoring aimed at improving code safety, clarity, and potentially enabling further optimization by avoiding lossy ptrtoint operations and expressing intent more directly through typed comparisons.

model: qwen-plus-latest
CompletionUsage(completion_tokens=498, prompt_tokens=100943, total_tokens=101441, completion_tokens_details=None, prompt_tokens_details=None)

[artagnon]

/close