pre-commit: PR125828

[zyw-bot]

Link: llvm/llvm-project#125828
Requested by: @artagnon

[zyw-bot]

Diff mode

runner: ariselab-64c-docker
baseline: llvm/llvm-project@5835f1e
patch: llvm/llvm-project#125828
sha256: 33d991d3513ea09692d5438c283ad3544574a07b47d2547274f19294db322c95
commit: 5a69592

141 files changed, 70102 insertions(+), 70633 deletions(-)

Improvements:
  licm.NumMinMaxHoisted 778 -> 783 +0.64%
  bdce.NumSExt2ZExt 5466 -> 5468 +0.04%
  dse.NumCFGChecks 651904 -> 652120 +0.03%
  licm.NumBOAssociationsHoisted 7182 -> 7184 +0.03%
  memory-builtins.ObjectVisitorLoad 2476759 -> 2477099 +0.01%
  licm.NumLoadStorePromoted 74279 -> 74288 +0.01%
  dse.NumCFGTries 58313 -> 58319 +0.01%
  indvars.NumElimCmp 57988 -> 57993 +0.01%
  adce.NumBranchesRemoved 24156 -> 24158 +0.01%
  correlated-value-propagation.NumAnd 48454 -> 48458 +0.01%
Regressions:
  loop-rotate.NumInstrsHoisted 1242 -> 1241 -0.08%
  indvars.NumLFTR 337255 -> 337070 -0.05%
  indvars.NumReplaced 72745 -> 72712 -0.05%
  simple-loop-unswitch.NumCostMultiplierSkipped 17325 -> 17320 -0.03%
  correlated-value-propagation.NumNNeg 105567 -> 105537 -0.03%
  loop-simplifycfg.NumLoopBlocksDeleted 7062 -> 7060 -0.03%
  loop-instsimplify.NumSimplified 196750 -> 196705 -0.02%
  correlated-value-propagation.NumSMinMax 9242 -> 9240 -0.02%
  loop-simplifycfg.NumTerminatorsFolded 10643 -> 10641 -0.02%
  gvn.NumGVNEqProp 456567 -> 456487 -0.02%

9 13 bench/abc/optimized/bmcFx.ll
41 47 bench/abc/optimized/wlcAbc.ll
4 6 bench/abc/optimized/wlnWlc.ll
6 8 bench/arrow/optimized/UriParse.ll
8 10 bench/boost/optimized/args_cmd.ll
21 25 bench/bullet3/optimized/btSoftBodyHelpers.ll
67 62 bench/cmake/optimized/cmHexFileConverter.ll
8 11 bench/cvc5/optimized/bags_utils.ll
70 71 bench/eastl/optimized/EAMemory.ll
30 33 bench/ffmpeg/optimized/aacdec.ll
21 20 bench/ffmpeg/optimized/exr.ll
15 16 bench/ffmpeg/optimized/hls_sample_encryption.ll
21 20 bench/ffmpeg/optimized/iamf_writer.ll
27 32 bench/ffmpeg/optimized/matroskadec.ll
24 30 bench/folly/optimized/JSONSchema.ll
35 49 bench/graphviz/optimized/excontext.ll
54 54 bench/graphviz/optimized/gvplugin_vt.ll
54 58 bench/hermes/optimized/hbc-attribute.ll
2 2 bench/icu/optimized/msgfmt.ll
9 14 bench/libcxx/optimized/locale.ll
45 53 bench/libevent/optimized/http.ll
52 58 bench/linux/optimized/indirect.ll
18 29 bench/llvm/optimized/SemaConcept.ll
24 25 bench/llvm/optimized/SemaTemplate.ll
43 47 bench/meshoptimizer/optimized/vertexcodec.ll
4 17 bench/node/optimized/libnode.Protocol.ll
46 48 bench/oiio/optimized/strutil.ll
17 20 bench/opencv/optimized/qrcode_encoder.ll
24 38 bench/openexr/optimized/internal_rle.ll
32 35 bench/openjdk/optimized/mlib_c_ImageCopy.ll
2 4 bench/openjdk/optimized/sharedRuntimeTrig.ll
24 38 bench/openusd/optimized/openexr-c.ll
6 4 bench/php/optimized/ir_ra.ll
74 80 bench/pola-rs/optimized/eo2lit9v8mg9048herjayt2j2.ll
46 52 bench/protobuf/optimized/padding_optimizer.ll
59 66 bench/qemu/optimized/qemu-io-cmds.ll
13 14 bench/qemu/optimized/sdhci-cmd.ll
24 32 bench/quantlib/optimized/date.ll
8 10 bench/recastnavigation/optimized/RecastMeshDetail.ll
8 13 bench/ruby/optimized/prism.ll
30 35 bench/sdl/optimized/SDL_guid.ll
43 45 bench/sqlite/optimized/sqlite3.ll
4 4 bench/sundials/optimized/sundials_hashmap.ll
9 6 bench/tls-rs/optimized/2xhpv4vf07do7r8g.ll
14 16 bench/wireshark/optimized/packet-ieee80211.ll
45 51 bench/wireshark/optimized/packet-mbim.ll
78 86 bench/wireshark/optimized/packet-netlink-sock_diag.ll
12 14 bench/wireshark/optimized/packet-sctp.ll
30 35 bench/wireshark/optimized/packet-ssh.ll
15 21 bench/wireshark/optimized/packet-tls-utils.ll
8 10 bench/wolfssl/optimized/chacha.ll

[github-actions]

Here is a summary of the most significant changes in the provided LLVM IR diffs, focusing on up to 5 major transformations and ignoring non-interesting changes such as formatting, comments, and reordering:

---

1. Loop Condition Refinement Using icmp samesign

In several functions (e.g., bmcFx.ll, Wlc_NtkPrintInputInfo.ll, exr.ll, hls_sample_encryption.ll, matroskadec.ll, etc.), loop exit conditions have been refined using icmp samesign ult or icmp samesign uge instead of comparing against a computed trip count. This suggests a transformation that avoids explicit computation of trip counts by checking whether the induction variable has reached a certain bound under the same sign assumptions.

• Example:

  %exitcond.not = icmp samesign ult i64 %indvars.iv.next.i.i.us, %wide.trip.count.i.i.us
  br i1 %exitcond.not, label %59, label %_ZL23ConvertMotorolaSrecLinePKcP8_IO_FILE.exit.us, !llvm.loop !12

  becomes:

  %exitcond.not = icmp eq i64 %indvars.iv.next31.i.i.us, %wide.trip.count.i.i.us)

High-level impact: Improves loop control logic by eliminating unnecessary trip count calculations and relying on bounds checks, potentially enabling better optimization opportunities and reducing code size.

---

2. Removal of Redundant llvm.umin Intrinsic Usage

Multiple instances of @llvm.umin.i32 and @llvm.umin.i64 are removed and replaced with more direct uses of zext/trunc combinations or simpler arithmetic. For example, in wlcAbc.ll and SemaTemplate.ll, explicit minimum value computations between two values are replaced with pointer difference or other comparisons.

• Example:

  %umin.i = tail call i32 @llvm.umin.i32(i32 %217, i32 %218)

  becomes:

  %217 = zext i32 %2 to i64
  %218 = zext nneg i32 %.val37.i to i64

High-level impact: Eliminates calls to intrinsic min functions where they can be expressed via truncation or zero extension, improving performance and reducing reliance on helper intrinsics.

---

3. Simplified PHI Nodes in Loops

Several loops had their PHI nodes simplified by removing redundant incoming edges or restructuring them to match updated control flow. For instance, in args_cmd.ll, Wlc_NtkGetInv.ll, and SemaConcept.ll, PHI instructions were adjusted after control flow changes or loop exits.

• Example:

  %.01230.i.i.i = phi i64 [ %.01230.i.i.i.add, %35 ], [ 0, %.lr.ph.preheader.i.i.i ]

  becomes:

  %.01230.i.i.i = phi i64 [ %.01230.i.i.i.add, %34 ], [ 0, %.lr.ph.preheader.i.i.i ]

High-level impact: Reflects structural simplifications in loop predecessors, often resulting from merging blocks or changing branch targets, which helps reduce complexity and improve clarity for downstream optimizations.

---

4. Switch Statement Target Adjustments

In several files (UriParse.ll, hls_sample_encryption.ll, matroskadec.ll), switch statements had their default labels changed or target labels reordered. This often corresponds to CFG cleanup or edge removal due to unreachable block elimination or sink threading.

• Example:

  switch i8 %80, label %._crit_edge.loopexit.i.i.us_crit_edge [
    i8 48, label %96
    ...
  ]

  becomes:

  switch i8 %80, label %._crit_edge.loopexit.i.i.us_crit_edge [
    i8 48, label %95
    ...
  ]

High-level impact: Indicates improved control flow graph (CFG) structure through dead code elimination and block reorganization.

---

5. Use of tail call and Inline Simplification of abs and bswap Intrinsics

In Wlc_NtkPrintInputInfo.ll, Wlc_NtkGetInv.ll, and aaccdec.ll, calls to @llvm.abs.i32 and @llvm.bswap.i32 are either made tail call or their arguments are folded into other expressions. Additionally, some conditional branches based on abs were simplified.

• Example:

  %smax = tail call i32 @llvm.smax.i32(i32 %.val67, i32 %.val68)
  %22 = add nuw nsw i32 %smax, 1

  becomes:

  %23 = zext nneg i32 %.val67 to i64

High-level impact: Suggests better use of intrinsic properties and potential constant folding, leading to reduced instruction count and improved execution efficiency.

---

Conclusion

The overall patch appears to focus on loop condition simplification, removal of redundant intrinsics, and control flow graph restructuring. These changes likely result from late-stage optimizations like LoopUnroll, TailCallElim, or DeadCodeElimination passes, aiming to improve performance and clean up generated IR by removing intermediate variables and replacing complex conditions with simpler ones.

model: qwen-plus-latest
CompletionUsage(completion_tokens=1252, prompt_tokens=105700, total_tokens=106952, completion_tokens_details=None, prompt_tokens_details=None)