pre-commit: PR137875

[dtcxzyw]

Link: llvm/llvm-project#137875
Requested by: @dtcxzyw

[dtcxzyw]

Diff mode

runner: ariselab-64c-v2
baseline: llvm/llvm-project@5b29058
patch: llvm/llvm-project#137875
sha256: fc5e1f92a105e7fccd5dd42d546a17c8d9fb210e30fc82d91ec944db435ca8be
commit: 045fd7a

92 files changed, 130027 insertions(+), 130534 deletions(-)

Improvements:
  basicaa.SearchLimitReached 34557 -> 34606 +0.14%
  correlated-value-propagation.NumNNeg 84479 -> 84507 +0.03%
  instcombine.NumDeadInst 35302713 -> 35309491 +0.02%
  correlated-value-propagation.NumNUW 513093 -> 513175 +0.02%
  instcombine.NumCombined 105523566 -> 105528227 +0.00%
  simplifycfg.NumFoldBranchToCommonDest 674931 -> 674955 +0.00%
  early-cse.NumCSELoad 6542893 -> 6543109 +0.00%
  early-cse.NumCSEGEP 11594923 -> 11595258 +0.00%
  jump-threading.NumThreads 2369901 -> 2369954 +0.00%
  globalopt.NumDeleted 853528 -> 853545 +0.00%
Regressions:
  early-cse.NumCSE 3846856 -> 3843521 -0.09%
  instcombine.NumFactor 39349 -> 39336 -0.03%
  correlated-value-propagation.NumAnd 38152 -> 38142 -0.03%
  gvn.NumGVNInstr 139516 -> 139482 -0.02%
  gvn.NumGVNPRE 139516 -> 139482 -0.02%
  bdce.NumRemoved 333537 -> 333507 -0.01%
  licm.NumGEPsHoisted 48404 -> 48402 -0.00%
  instcombine.NumDeadStore 24550 -> 24549 -0.00%
  instcombine.NumSunkInst 2878868 -> 2878755 -0.00%
  count-visits.MaxVisited 43130 -> 43129 -0.00%

21 23 bench/abc/optimized/abc.ll
6 7 bench/abc/optimized/bmcMulti.ll
74 76 bench/abc/optimized/cecSatG2.ll
8 9 bench/abc/optimized/giaCof.ll
36 38 bench/abc/optimized/giaEquiv.ll
57 61 bench/abc/optimized/giaPat2.ll
62 67 bench/abseil-cpp/optimized/float_conversion.ll
42 54 bench/abseil-cpp/optimized/gtest-all.ll
52 64 bench/abseil-cpp/optimized/int128.ll
8 9 bench/fmt/optimized/format-impl-test.ll
64 76 bench/freetype/optimized/sfnt.ll
14 15 bench/hdf5/optimized/H5Fsuper_cache.ll
122 129 bench/imgui/optimized/imgui_draw.ll
164 168 bench/linux/optimized/tree.ll
14 16 bench/llvm/optimized/RISCVMergeBaseOffset.ll
102 105 bench/llvm/optimized/blake3.ll
78 88 bench/meilisearch-rs/optimized/2eqafq69fd46y14a.ll
55 61 bench/meshlab/optimized/gltf_loader.ll
180 176 bench/mold/optimized/arch-m68k.ll
238 234 bench/mold/optimized/arch-sparc64.ll
102 105 bench/mold/optimized/blake3.ll
49 56 bench/openjdk/optimized/hb-ot-name.ll
26 27 bench/openjdk/optimized/utf8.ll
122 127 bench/openssl/optimized/d1_lib.ll
43 46 bench/openssl/optimized/statem_srvr.ll
57 60 bench/php/optimized/pcre2_auto_possess.ll
4 7 bench/php/optimized/pcre2_study.ll
22 29 bench/qemu/optimized/fdt.ll
119 140 bench/qemu/optimized/fdt_sw.ll
128 137 bench/recastnavigation/optimized/imguiRenderGL.ll
1 1 bench/rust-analyzer-rs/optimized/36cskjvhktchzlnh.ll
2 2 bench/rust-analyzer-rs/optimized/5ac99zaxn7b9r9xv.ll
4 4 bench/rust-analyzer-rs/optimized/hf9vzunhg9aziex.ll
6 6 bench/rust-analyzer-rs/optimized/hknx1qr3lu9291s.ll
21 28 bench/spike/optimized/fdt.ll
128 73 bench/spike/optimized/fdt_ro.ll
119 140 bench/spike/optimized/fdt_sw.ll
33 29 bench/wireshark/optimized/netscaler.ll
17 20 bench/wireshark/optimized/packet-atm.ll
10 13 bench/wireshark/optimized/packet-knet.ll
86 89 bench/wireshark/optimized/packet-opensafety.ll
173 189 bench/wireshark/optimized/tvbuff.ll
169 175 bench/wolfssl/optimized/tls.ll
9 11 bench/zed-rs/optimized/38cn6p2m6864jrrxog4mr8xwk.ll
6 8 bench/zed-rs/optimized/7ud3epkhjcjfe38h6hlh4jrau.ll

[github-actions]

Here is a high-level summary of the major changes in the provided LLVM IR diff:

1. Simplified Integer Comparison Logic:

   • In multiple functions (e.g., @Abc_CommandAbc9Pms, @Gia_ManCountConst0PosGia, and others), the transformation replaces a pattern involving shift-and-or (shl + or disjoint) followed by equality check against zero with a simpler or instruction directly. This reduces instruction count and simplifies control flow, improving both readability and execution efficiency.
   • Example: Replaced %72 = shl nsw i32 %47, 1; %73 = or disjoint i32 %72, %45 with %72 = or i32 %47, %45.

2. Phi Node Adjustments for Loop Control:

   • Several functions adjust phi node values in loop headers and exit blocks to reflect updated variable names and new value sources due to code restructuring.
   • For instance, in @Abc_CommandAbc9Pms, .critedge block's phi nodes now reference renamed labels like %._crit_edge instead of outdated ones like %.lr.ph, likely reflecting loop unrolling or restructure.

3. Branch Condition Refinement:

   • The condition logic in branches has been optimized by replacing complex comparisons with more direct ones, often using bitwise OR and EQ/NE checks instead of shifts and masks.
   • Example: In @tt_cmap2_char_index, replaced a multi-step bitshift-comparison sequence with a simple or between two bytes followed by an icmp eq.

4. Memory Load and Phi Node Updates:

   • In several functions (e.g., @H5F__cache_superblock_deserialize, @Gia_GenerateCexesDumpFile), memory loads and pointer calculations have been reordered or simplified. Phi nodes were adjusted accordingly to match new predecessor blocks and variable states.

5. Reduction in Intermediate Instructions:

   • Some functions removed unnecessary intermediate instructions such as redundant truncations, shifts, and insertions, consolidating them into fewer steps while preserving behavior.
   • Example: In @rcu_preempt_deferred_qs_irqrestore, replaced a chain of zext, shl, and or operations with a single or operation on lower bits before final comparison.

These optimizations primarily focus on reducing complexity, removing redundant computation, and streamlining control flow paths. The result is a leaner, more efficient IR that should yield better runtime performance and smaller generated machine code.

model: qwen-plus-latest
CompletionUsage(completion_tokens=545, prompt_tokens=119817, total_tokens=120362, completion_tokens_details=None, prompt_tokens_details=None)