Skip to content

Add nautilus-profiling plugin for runtime performance analysis - #244

Open
PhilippGrulich wants to merge 1 commit into
mainfrom
claude/nautilus-profiling-plugin-AGwCD
Open

Add nautilus-profiling plugin for runtime performance analysis#244
PhilippGrulich wants to merge 1 commit into
mainfrom
claude/nautilus-profiling-plugin-AGwCD

Conversation

@PhilippGrulich

@PhilippGrulich PhilippGrulich commented Apr 12, 2026

Copy link
Copy Markdown
Member

New plugin that provides cross-backend function profiling for Nautilus:

  • ProfiledFunction wrapper with zero overhead when profiling inactive
  • Thread-local call stacks and ring buffers for lock-free hot path
  • Chrome Trace Format export (viewable in Perfetto/chrome://tracing)
  • Folded stacks export (for Brendan Gregg's flamegraph.pl)
  • Perf map writer (/tmp/perf-.map) for JIT symbol resolution
  • Sampling support to reduce overhead for high-frequency calls
  • Optional CPU cycle collection via rdtsc (x86)

https://claude.ai/code/session_01SLkb9NXtrjb4QAkiUTTMf9

github-actions[bot]
github-actions Bot reviewed Apr 12, 2026
View reviewed changes

@github-actions github-actions Bot left a comment •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Tracing Benchmark

Details
Benchmark suite Current: 2fccf55 Previous: 735a59d Ratio
e2e_tiered_bc_to_mlir 54.7671 us (± 12.073) 44.7721 us (± 16.4703) 1.22
e2e_single_mlir 8.216 ms (± 195.965) 9.16088 ms (± 188.959) 0.90
ir_add 683.408 ns (± 125.492) 557.603 ns (± 33.4202) 1.23
ir_ifThenElse 1.53666 us (± 203.108) 1.35697 us (± 125.581) 1.13
ir_deeplyNestedIfElse 3.53577 us (± 533.464) 3.25431 us (± 120.92) 1.09
ir_loop 1.6294 us (± 199.26) 1.48915 us (± 149.204) 1.09
ir_ifInsideLoop 2.96122 us (± 442.026) 2.71554 us (± 269.14) 1.09
ir_loopDirectCall 1.75895 us (± 200.625) 1.53102 us (± 46.9752) 1.15
ir_pointerLoop 2.04121 us (± 241.925) 1.97107 us (± 430.214) 1.04
ir_staticLoop 1.385 us (± 186.946) 1.21683 us (± 80.1593) 1.14
ir_fibonacci 1.74644 us (± 190.629) 1.57054 us (± 100.473) 1.11
ir_gcd 1.4915 us (± 261.968) 1.30244 us (± 80.1681) 1.15
ir_nestedIf10 7.80092 us (± 759.417) 7.07715 us (± 494.947) 1.10
ir_nestedIf100 90.2452 us (± 5.46478) 94.1299 us (± 6.63792) 0.96
ir_chainedIf10 12.5194 us (± 995.477) 11.6931 us (± 684.358) 1.07
ir_chainedIf100 180.376 us (± 8.65112) 178.212 us (± 10.231) 1.01
ssa_add 198.299 ns (± 21.9527) 178.589 ns (± 8.61098) 1.11
ssa_ifThenElse 494.836 ns (± 65.537) 432.511 ns (± 24.8694) 1.14
ssa_deeplyNestedIfElse 1.24183 us (± 138.063) 1.12092 us (± 62.1812) 1.11
ssa_loop 532.229 ns (± 62.3371) 454.465 ns (± 22.6714) 1.17
ssa_ifInsideLoop 949.79 ns (± 111.829) 914.873 ns (± 156.953) 1.04
ssa_loopDirectCall 525.267 ns (± 51.9665) 463.089 ns (± 27.2138) 1.13
ssa_pointerLoop 615.26 ns (± 61.8665) 554.161 ns (± 36.2962) 1.11
ssa_staticLoop 486.748 ns (± 67.996) 430.496 ns (± 17.6849) 1.13
ssa_fibonacci 548.787 ns (± 72.4053) 475.635 ns (± 26.8476) 1.15
ssa_gcd 472.001 ns (± 44.8091) 422.707 ns (± 24.5879) 1.12
comp_mlir_add 8.16241 ms (± 141.623) 9.42323 ms (± 413.446) 0.87
comp_mlir_ifThenElse 8.78392 ms (± 179.918) 10.0186 ms (± 356.845) 0.88
comp_mlir_deeplyNestedIfElse 7.82057 ms (± 279.79) 9.30722 ms (± 253.08) 0.84
comp_mlir_loop 9.89583 ms (± 221.832) 11.6633 ms (± 259.204) 0.85
comp_mlir_ifInsideLoop 31.5907 ms (± 537.955) 33.8733 ms (± 651.356) 0.93
comp_mlir_loopDirectCall 14.5314 ms (± 279.57) 16.3832 ms (± 447.685) 0.89
comp_mlir_pointerLoop 30.7105 ms (± 521.05) 32.7066 ms (± 451.19) 0.94
comp_mlir_staticLoop 7.75208 ms (± 161.867) 8.53315 ms (± 175.692) 0.91
comp_mlir_fibonacci 13.2057 ms (± 225.043) 14.6574 ms (± 312.126) 0.90
comp_mlir_gcd 12.1243 ms (± 231.481) 13.9072 ms (± 1.1126) 0.87
comp_mlir_nestedIf10 14.2804 ms (± 585.905) 14.7447 ms (± 361.456) 0.97
comp_mlir_nestedIf100 27.774 ms (± 416.369) 29.0962 ms (± 380.159) 0.95
comp_mlir_chainedIf10 12.4629 ms (± 292.477) 14.0527 ms (± 1.03727) 0.89
comp_mlir_chainedIf100 23.2097 ms (± 597.686) 24.8764 ms (± 347.559) 0.93
comp_cpp_add 25.0594 ms (± 550.437) 28.8601 ms (± 464.566) 0.87
comp_cpp_ifThenElse 25.8434 ms (± 634.298) 29.4161 ms (± 503.621) 0.88
comp_cpp_deeplyNestedIfElse 26.8055 ms (± 652.32) 30.3349 ms (± 470.032) 0.88
comp_cpp_loop 26.3085 ms (± 785.009) 29.5386 ms (± 452.283) 0.89
comp_cpp_ifInsideLoop 27.1847 ms (± 710.565) 30.6408 ms (± 1.1838) 0.89
comp_cpp_loopDirectCall 26.7895 ms (± 880.228) 29.968 ms (± 534.834) 0.89
comp_cpp_pointerLoop 26.2453 ms (± 567.182) 30.0626 ms (± 649.42) 0.87
comp_cpp_staticLoop 26.3035 ms (± 1.09779) 29.3365 ms (± 625) 0.90
comp_cpp_fibonacci 26.1082 ms (± 617.409) 30.3049 ms (± 1.0683) 0.86
comp_cpp_gcd 25.7528 ms (± 634.316) 30.1386 ms (± 655.731) 0.85
comp_cpp_nestedIf10 28.9192 ms (± 664.198) 33.1127 ms (± 1.59488) 0.87
comp_cpp_nestedIf100 62.9123 ms (± 1.60961) 66.1697 ms (± 1.99571) 0.95
comp_cpp_chainedIf10 31.7168 ms (± 544.477) 35.0051 ms (± 569.074) 0.91
comp_cpp_chainedIf100 92.52 ms (± 936.298) 95.6542 ms (± 1.27066) 0.97
comp_bc_add 14.5435 us (± 2.21486) 14.4189 us (± 2.37199) 1.01
comp_bc_ifThenElse 18.7079 us (± 3.37375) 16.7978 us (± 3.7847) 1.11
comp_bc_deeplyNestedIfElse 23.3502 us (± 4.11475) 20.1859 us (± 4.35147) 1.16
comp_bc_loop 18.5939 us (± 3.35349) 16.8307 us (± 3.90376) 1.10
comp_bc_ifInsideLoop 21.316 us (± 3.50596) 18.9231 us (± 3.849) 1.13
comp_bc_loopDirectCall 19.2566 us (± 3.68133) 17.0217 us (± 3.51388) 1.13
comp_bc_pointerLoop 20.6824 us (± 4.86347) 18.0283 us (± 4.01563) 1.15
comp_bc_staticLoop 16.9912 us (± 2.82385) 16.5494 us (± 3.68146) 1.03
comp_bc_fibonacci 18.7936 us (± 3.33194) 17.0462 us (± 3.76914) 1.10
comp_bc_gcd 19.3061 us (± 5.00465) 15.9395 us (± 2.47334) 1.21
comp_bc_nestedIf10 35.9529 us (± 4.75701) 31.0134 us (± 4.42009) 1.16
comp_bc_nestedIf100 201.857 us (± 10.1483) 191.712 us (± 9.52673) 1.05
comp_bc_chainedIf10 52.5269 us (± 8.73991) 43.141 us (± 8.89736) 1.22
comp_bc_chainedIf100 311.203 us (± 14.6285) 309.419 us (± 12.6364) 1.01
comp_asmjit_add 21.3213 us (± 4.7054) 17.5231 us (± 5.7044) 1.22
comp_asmjit_ifThenElse 33.0105 us (± 4.40255) 26.185 us (± 5.62526) 1.26
comp_asmjit_deeplyNestedIfElse 57.2958 us (± 10.4535) 47.5287 us (± 10.7073) 1.21
comp_asmjit_loop 35.9778 us (± 5.8488) 28.2969 us (± 6.06829) 1.27
comp_asmjit_ifInsideLoop 57.4703 us (± 10.2007) 47.413 us (± 12.3118) 1.21
comp_asmjit_loopDirectCall 65.6188 us (± 16.4463) 30.9936 us (± 6.48682) 2.12
comp_asmjit_pointerLoop 48.2705 us (± 9.42333) 33.2829 us (± 5.89897) 1.45
comp_asmjit_staticLoop 28.4792 us (± 5.07002) 23.4582 us (± 4.88987) 1.21
comp_asmjit_fibonacci 43.1285 us (± 8.03749) 30.2905 us (± 5.68839) 1.42
comp_asmjit_gcd 34.4759 us (± 4.84178) 28.1555 us (± 6.13342) 1.22
comp_asmjit_nestedIf10 103.171 us (± 11.6038) 96.044 us (± 15.4869) 1.07
comp_asmjit_nestedIf100 1.03411 ms (± 40.8132) 1.04677 ms (± 51.1612) 0.99
comp_asmjit_chainedIf10 154.8 us (± 15.8851) 148.825 us (± 16.3448) 1.04
comp_asmjit_chainedIf100 2.14084 ms (± 27.6337) 2.19743 ms (± 43.3502) 0.97
exec_mlir_add 11.0682 ns (± 1.01233) 11.1087 ns (± 0.853075) 1.00
exec_mlir_fibonacci 14.2285 us (± 1.40998) 18.7581 us (± 2.52184) 0.76
exec_mlir_sum 557.915 us (± 34.7983) 565.098 us (± 56.3939) 0.99
exec_cpp_add 4.78506 ns (± 0.919231) 4.67595 ns (± 1.06637) 1.02
exec_cpp_fibonacci 97.0579 us (± 9.67916) 109.488 us (± 8.76714) 0.89
exec_cpp_sum 36.0829 ms (± 857.458) 23.5994 ms (± 113.601) 1.53
exec_bc_add 45.6111 ns (± 7.86091) 44.5227 ns (± 6.90048) 1.02
exec_bc_fibonacci 787.658 us (± 14.7907) 626.695 us (± 14.1954) 1.26
exec_bc_sum 168.61 ms (± 263.39) 143.596 ms (± 1.5291) 1.17
exec_asmjit_add 3.5426 ns (± 0.433836) 3.16976 ns (± 0.210314) 1.12
exec_asmjit_fibonacci 20.6619 us (± 1.74046) 22.4194 us (± 2.2047) 0.92
exec_asmjit_sum 4.84839 ms (± 24.9235) 5.28048 ms (± 29.4129) 0.92
exec_bc_add_noRegAlloc 43.8871 ns (± 4.30266) 43.4498 ns (± 2.08094) 1.01
exec_bc_add_regAlloc 44.6625 ns (± 6.27319) 43.0869 ns (± 1.60579) 1.04
exec_bc_fibonacci_noRegAlloc 787.423 us (± 16.4042) 631.805 us (± 12.0476) 1.25
exec_bc_fibonacci_regAlloc 788.138 us (± 12.4609) 626.084 us (± 11.7745) 1.26
exec_bc_sum_noRegAlloc 168.646 ms (± 1.10532) 144.361 ms (± 7.34656) 1.17
exec_bc_sum_regAlloc 168.925 ms (± 1.62293) 143.149 ms (± 1.08876) 1.18
tiered_compile_addOne 53.2616 us (± 13.052) 43.1655 us (± 15.6613) 1.23
single_compile_mlir_addOne 6.53973 ms (± 837.363) 7.30884 ms (± 250.756) 0.89
single_compile_cpp_addOne 25.7984 ms (± 668.839) 29.3191 ms (± 604.247) 0.88
single_compile_bc_addOne 55.4063 us (± 12.5207) 47.631 us (± 23.1615) 1.16
tiered_compile_sumLoop 74.7299 us (± 10.0684) 60.9415 us (± 18.3227) 1.23
single_compile_mlir_sumLoop 8.54146 ms (± 298.643) 9.44764 ms (± 302.585) 0.90
single_compile_cpp_sumLoop 25.9572 ms (± 858.66) 30.3033 ms (± 749.906) 0.86
single_compile_bc_sumLoop 77.0379 us (± 13.5931) 62.316 us (± 18.6408) 1.24
exec_bc_addOne 37.5175 ns (± 7.47004) 35.8115 ns (± 3.14464) 1.05
exec_mlir_addOne 289.423 ns (± 12.719) 318.969 ns (± 13.122) 0.91
exec_cpp_addOne 4.11486 ns (± 0.937765) 4.15126 ns (± 0.438887) 0.99
exec_interpreted_addOne 41.4778 ns (± 4.22045) 38.4164 ns (± 1.75957) 1.08
trace_add 2.40625 us (± 247.496) 2.06394 us (± 181.932) 1.17
completing_trace_add 2.42015 us (± 314.091) 2.09441 us (± 173.694) 1.16
trace_ifThenElse 9.05026 us (± 1.39162) 7.82154 us (± 1.27749) 1.16
completing_trace_ifThenElse 4.79966 us (± 734.403) 4.15494 us (± 494.679) 1.16
trace_deeplyNestedIfElse 26.8847 us (± 3.47424) 23.5052 us (± 3.7184) 1.14
completing_trace_deeplyNestedIfElse 13.2808 us (± 1.7771) 11.5912 us (± 1.68569) 1.15
trace_loop 9.11383 us (± 1.43192) 7.70984 us (± 1.09486) 1.18
completing_trace_loop 4.9523 us (± 763.863) 4.27197 us (± 452.073) 1.16
trace_ifInsideLoop 17.7908 us (± 2.97482) 15.3213 us (± 2.10195) 1.16
completing_trace_ifInsideLoop 8.94958 us (± 1.76726) 7.73302 us (± 938.568) 1.16
trace_loopDirectCall 9.1779 us (± 1.41379) 7.80253 us (± 1.15647) 1.18
completing_trace_loopDirectCall 5.06418 us (± 889.091) 4.2795 us (± 509.818) 1.18
trace_pointerLoop 14.7854 us (± 2.63526) 12.6363 us (± 2.33328) 1.17
completing_trace_pointerLoop 10.9273 us (± 1.61604) 8.99261 us (± 1.44436) 1.22
trace_staticLoop 7.84281 us (± 1.07616) 7.04359 us (± 755.434) 1.11
completing_trace_staticLoop 8.02091 us (± 1.07411) 7.09734 us (± 735.226) 1.13
trace_fibonacci 10.726 us (± 1.9064) 9.02193 us (± 1.46526) 1.19
completing_trace_fibonacci 6.68872 us (± 1.09886) 5.49339 us (± 665.791) 1.22
trace_gcd 8.36616 us (± 1.34929) 8.90115 us (± 2.65556) 0.94
completing_trace_gcd 4.1841 us (± 538.403) 3.57162 us (± 411.759) 1.17
trace_nestedIf10 39.7454 us (± 7.88893) 36.5071 us (± 7.32195) 1.09
completing_trace_nestedIf10 40.5467 us (± 7.91072) 36.8142 us (± 7.18954) 1.10
trace_nestedIf100 1.35962 ms (± 64.4087) 1.56252 ms (± 247.341) 0.87
completing_trace_nestedIf100 1.37936 ms (± 55.947) 1.4543 ms (± 35.5561) 0.95
trace_chainedIf10 100.314 us (± 10.0364) 92.2442 us (± 11.1303) 1.09
completing_trace_chainedIf10 51.8227 us (± 9.51509) 46.815 us (± 8.88196) 1.11
trace_chainedIf100 4.43399 ms (± 49.7087) 4.4846 ms (± 225.461) 0.99
completing_trace_chainedIf100 2.30073 ms (± 104.487) 2.19662 ms (± 55.2405) 1.05

This comment was automatically generated by workflow using github-action-benchmark.

@PhilippGrulich
PhilippGrulich force-pushed the claude/nautilus-profiling-plugin-AGwCD branch 4 times, most recently from 9f18d08 to 728157f Compare April 19, 2026 07:06
@PhilippGrulich @claude
Add profiling plugin with engine-level IR pass registration
2fccf55 
Introduces a `plugins/profiling/` plugin that auto-instruments every
compiled function with wall-clock timing and (on Linux) hardware
performance-counter sampling via `perf_event_open`.

Plugin pass infrastructure:
- `nautilus::compiler::ir::IRPass` (in `compiler/ir/passes/IRPass.hpp`)
  is now the single base class for both built-in and plugin passes;
  `nautilus::IRPass` is a friendly alias for plugin code.
- `JITCompiler::addIRPass(unique_ptr<IRPass>)` is part of the abstract
  interface. `LegacyCompiler` stores plugin passes as `shared_ptr` and
  appends them to the same `IRPassManager` that runs built-in passes,
  so plugin passes share dump/verify/statistics handling.
- `TieredJITCompiler::addIRPass` delegates to its inner `LegacyCompiler`
  — plugin passes now work in both tier-0 and tier-1 (replacing the
  previous "TieredJITCompiler does not support plugin IR passes" throw).
- `IRPassManager::addPass` switched to `shared_ptr<IRPass>`; built-in
  callers pass `make_shared<...>` (refcount=1, semantically equivalent
  to unique_ptr), plugin passes flow through unchanged.
- `NautilusEngine::addIRPass()` is the public entry point; defined
  out-of-line in `Engine.cpp` so the public header keeps `IRPass`
  forward-declared and `NautilusModule` stays a pure function registry.

Profiling plugin:
- `ProfilingIRPass` walks every `FunctionOperation` and emits
  `ProfileRegion::runtimeStart` / `runtimeStop` proxy calls at entry
  and every return site.
- `Profiler` / `ProfileRegion` provide thread-safe accounting for call
  counts, total/min/max latency and up to four HW event counters; the
  profiler reports to stdout or Chrome-trace JSON on destruction.
- `Profiler::enableForEngine(engine)` is the user-facing entry point —
  installs the IR pass on the engine so every subsequent module the
  engine compiles is instrumented automatically.
- Example (`DemoProfilingPlugin.cpp`) and full test suite
  (`ProfilingTest.cpp`, 10 cases / 1039 assertions) covering timing,
  multi-function modules, and the Linux perf-counter paths.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
@PhilippGrulich
PhilippGrulich force-pushed the claude/nautilus-profiling-plugin-AGwCD branch from 728157f to 2fccf55 Compare April 19, 2026 19:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@github-actions github-actions[bot] github-actions[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@PhilippGrulich