Add ENABLE_CONSTANT_TRACER prototype (Stage 1 partial) - #260
Open
PhilippGrulich wants to merge 2 commits into
Open
Add ENABLE_CONSTANT_TRACER prototype (Stage 1 partial)#260PhilippGrulich wants to merge 2 commits into
PhilippGrulich wants to merge 2 commits into
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Tracing Benchmark
Details
| Benchmark suite | Current: ac7957e | Previous: ac2c4ea | Ratio |
|---|---|---|---|
e2e_tiered_bc_to_mlir |
49.6816 us (± 11.4293) |
47.7032 us (± 8.68568) |
1.04 |
e2e_single_mlir |
8.23867 ms (± 294.546) |
8.23818 ms (± 491.471) |
1.00 |
ssa_add |
177.991 ns (± 11.6883) |
185.816 ns (± 10.3996) |
0.96 |
ssa_ifThenElse |
466.029 ns (± 29.5821) |
455.998 ns (± 8.73188) |
1.02 |
ssa_deeplyNestedIfElse |
1.19914 us (± 103.988) |
1.1929 us (± 84.0502) |
1.01 |
ssa_loop |
453.106 ns (± 26.3689) |
495.525 ns (± 42.2521) |
0.91 |
ssa_ifInsideLoop |
838.535 ns (± 51.3168) |
923.55 ns (± 58.5963) |
0.91 |
ssa_loopDirectCall |
507.629 ns (± 51.606) |
497.983 ns (± 20.4573) |
1.02 |
ssa_pointerLoop |
596.269 ns (± 35.3296) |
600.749 ns (± 36.2279) |
0.99 |
ssa_staticLoop |
451.68 ns (± 24.8408) |
497.541 ns (± 60.4145) |
0.91 |
ssa_fibonacci |
516.015 ns (± 41.1861) |
520.747 ns (± 31.8873) |
0.99 |
ssa_gcd |
453.194 ns (± 31.439) |
451.525 ns (± 29.293) |
1.00 |
exec_bc_addOne |
35.9851 ns (± 6.12995) |
36.2618 ns (± 5.56148) |
0.99 |
exec_mlir_addOne |
307.967 ns (± 9.30423) |
276.222 ns (± 5.76067) |
1.11 |
exec_cpp_addOne |
4.08613 ns (± 1.0367) |
3.88144 ns (± 0.331571) |
1.05 |
exec_interpreted_addOne |
38.0389 ns (± 1.81467) |
38.0484 ns (± 1.8257) |
1.00 |
ir_add |
631.917 ns (± 33.7744) |
618.115 ns (± 53.3797) |
1.02 |
ir_ifThenElse |
1.51633 us (± 113.587) |
1.44905 us (± 86.8949) |
1.05 |
ir_deeplyNestedIfElse |
3.5321 us (± 459.519) |
3.37592 us (± 239.133) |
1.05 |
ir_loop |
1.59076 us (± 94.4772) |
1.53605 us (± 105.581) |
1.04 |
ir_ifInsideLoop |
2.85155 us (± 216.869) |
2.77156 us (± 196.154) |
1.03 |
ir_loopDirectCall |
1.74458 us (± 187.481) |
1.66853 us (± 129.354) |
1.05 |
ir_pointerLoop |
1.9764 us (± 196.735) |
1.92704 us (± 132.013) |
1.03 |
ir_staticLoop |
1.31866 us (± 214.846) |
1.2645 us (± 80.8019) |
1.04 |
ir_fibonacci |
1.74597 us (± 194.557) |
1.65853 us (± 128.906) |
1.05 |
ir_gcd |
1.41674 us (± 99.0636) |
1.37896 us (± 94.5931) |
1.03 |
ir_nestedIf10 |
7.5986 us (± 487.31) |
7.6512 us (± 799.841) |
0.99 |
ir_nestedIf100 |
91.5981 us (± 4.86431) |
93.3698 us (± 6.73228) |
0.98 |
ir_chainedIf10 |
12.4957 us (± 942.577) |
12.4575 us (± 863.742) |
1.00 |
ir_chainedIf100 |
183.386 us (± 9.94714) |
183.581 us (± 16.7128) |
1.00 |
exec_mlir_add |
10.9265 ns (± 1.22541) |
10.822 ns (± 0.936004) |
1.01 |
exec_mlir_fibonacci |
15.8631 us (± 1.87754) |
13.7916 us (± 1.74485) |
1.15 |
exec_mlir_sum |
538.244 us (± 27.0618) |
515.967 us (± 16.084) |
1.04 |
exec_cpp_add |
4.81543 ns (± 0.996424) |
4.6159 ns (± 0.528671) |
1.04 |
exec_cpp_fibonacci |
94.6739 us (± 7.14187) |
97.6672 us (± 10.8249) |
0.97 |
exec_cpp_sum |
35.8817 ms (± 221.876) |
35.9244 ms (± 434.181) |
1.00 |
exec_bc_add |
43.8306 ns (± 5.07487) |
44.8834 ns (± 7.90089) |
0.98 |
exec_bc_fibonacci |
839.887 us (± 15.5265) |
842.734 us (± 20.2036) |
1.00 |
exec_bc_sum |
178.513 ms (± 1.60754) |
178.8 ms (± 3.10469) |
1.00 |
exec_asmjit_add |
3.50799 ns (± 0.465687) |
3.49445 ns (± 0.375379) |
1.00 |
exec_asmjit_fibonacci |
21.368 us (± 1.88768) |
21.6426 us (± 2.74605) |
0.99 |
exec_asmjit_sum |
4.65507 ms (± 95.0337) |
4.71937 ms (± 21.7272) |
0.99 |
exec_bc_add_noRegAlloc |
43.5893 ns (± 4.34073) |
43.9791 ns (± 4.89944) |
0.99 |
exec_bc_add_regAlloc |
43.8246 ns (± 4.82559) |
43.5639 ns (± 2.86818) |
1.01 |
exec_bc_fibonacci_noRegAlloc |
841.07 us (± 14.1067) |
861.571 us (± 98.4839) |
0.98 |
exec_bc_fibonacci_regAlloc |
845.345 us (± 15.2232) |
839.407 us (± 13.4705) |
1.01 |
exec_bc_sum_noRegAlloc |
178.559 ms (± 1.05316) |
179.223 ms (± 2.02788) |
1.00 |
exec_bc_sum_regAlloc |
178.727 ms (± 538.677) |
178.601 ms (± 3.45719) |
1.00 |
comp_mlir_add |
8.09733 ms (± 195.508) |
8.03653 ms (± 44.1051) |
1.01 |
comp_mlir_ifThenElse |
8.87835 ms (± 187.319) |
8.61342 ms (± 54.5514) |
1.03 |
comp_mlir_deeplyNestedIfElse |
7.80659 ms (± 198.487) |
7.5802 ms (± 40.4471) |
1.03 |
comp_mlir_loop |
9.98918 ms (± 357.061) |
9.48877 ms (± 103.05) |
1.05 |
comp_mlir_ifInsideLoop |
31.5971 ms (± 471.105) |
30.5007 ms (± 230.085) |
1.04 |
comp_mlir_loopDirectCall |
14.6247 ms (± 285.779) |
14.0043 ms (± 46.9025) |
1.04 |
comp_mlir_pointerLoop |
31.0957 ms (± 876.226) |
29.5874 ms (± 103.344) |
1.05 |
comp_mlir_staticLoop |
7.72811 ms (± 285.271) |
7.47096 ms (± 38.3264) |
1.03 |
comp_mlir_fibonacci |
13.4716 ms (± 453.298) |
12.6628 ms (± 38.156) |
1.06 |
comp_mlir_gcd |
12.3361 ms (± 370.663) |
11.6741 ms (± 63.0358) |
1.06 |
comp_mlir_nestedIf10 |
13.7389 ms (± 436.582) |
12.8187 ms (± 55.0301) |
1.07 |
comp_mlir_nestedIf100 |
27.2215 ms (± 426.761) |
27.225 ms (± 87.2712) |
1.00 |
comp_mlir_chainedIf10 |
12.307 ms (± 341.335) |
11.8168 ms (± 116.022) |
1.04 |
comp_mlir_chainedIf100 |
23.2382 ms (± 659.182) |
22.1996 ms (± 240.353) |
1.05 |
comp_cpp_add |
25.0971 ms (± 719.14) |
24.5321 ms (± 480.572) |
1.02 |
comp_cpp_ifThenElse |
25.479 ms (± 603.278) |
25.2727 ms (± 558.226) |
1.01 |
comp_cpp_deeplyNestedIfElse |
26.3087 ms (± 688.859) |
26.2665 ms (± 459.988) |
1.00 |
comp_cpp_loop |
25.8364 ms (± 508.337) |
25.2988 ms (± 344.351) |
1.02 |
comp_cpp_ifInsideLoop |
27.0255 ms (± 592.473) |
26.4559 ms (± 789.37) |
1.02 |
comp_cpp_loopDirectCall |
26.9702 ms (± 1.35278) |
25.0943 ms (± 221.508) |
1.07 |
comp_cpp_pointerLoop |
26.397 ms (± 805.398) |
25.222 ms (± 258.79) |
1.05 |
comp_cpp_staticLoop |
24.8856 ms (± 717.374) |
24.5928 ms (± 309.914) |
1.01 |
comp_cpp_fibonacci |
25.85 ms (± 764.112) |
24.9754 ms (± 293.062) |
1.04 |
comp_cpp_gcd |
25.8512 ms (± 612.924) |
25.2159 ms (± 502.157) |
1.03 |
comp_cpp_nestedIf10 |
30.6915 ms (± 704.766) |
28.5753 ms (± 693.59) |
1.07 |
comp_cpp_nestedIf100 |
62.1937 ms (± 1.59868) |
61.7031 ms (± 527.471) |
1.01 |
comp_cpp_chainedIf10 |
31.5669 ms (± 1.10673) |
30.694 ms (± 425.766) |
1.03 |
comp_cpp_chainedIf100 |
91.8308 ms (± 1.00526) |
91.4516 ms (± 673.024) |
1.00 |
comp_bc_add |
14.855 us (± 2.21368) |
14.3597 us (± 2.48149) |
1.03 |
comp_bc_ifThenElse |
18.7942 us (± 2.98858) |
18.1495 us (± 2.61377) |
1.04 |
comp_bc_deeplyNestedIfElse |
22.997 us (± 3.22877) |
21.8467 us (± 3.49218) |
1.05 |
comp_bc_loop |
19.3884 us (± 4.63735) |
18.5722 us (± 4.77224) |
1.04 |
comp_bc_ifInsideLoop |
22.2572 us (± 3.72876) |
21.1077 us (± 3.93291) |
1.05 |
comp_bc_loopDirectCall |
19.9262 us (± 3.87713) |
18.8034 us (± 3.15769) |
1.06 |
comp_bc_pointerLoop |
20.5619 us (± 3.35869) |
20.003 us (± 3.56989) |
1.03 |
comp_bc_staticLoop |
17.2778 us (± 3.03127) |
17.0675 us (± 3.19641) |
1.01 |
comp_bc_fibonacci |
19.1415 us (± 3.14582) |
18.8155 us (± 3.99054) |
1.02 |
comp_bc_gcd |
18.7145 us (± 3.72348) |
18.1727 us (± 3.2952) |
1.03 |
comp_bc_nestedIf10 |
36.2962 us (± 4.60828) |
35.1955 us (± 4.58262) |
1.03 |
comp_bc_nestedIf100 |
201.406 us (± 13.7971) |
197.212 us (± 11.5333) |
1.02 |
comp_bc_chainedIf10 |
52.1207 us (± 7.41229) |
50.988 us (± 9.82885) |
1.02 |
comp_bc_chainedIf100 |
303.043 us (± 13.7926) |
306.036 us (± 11.3522) |
0.99 |
comp_asmjit_add |
19.814 us (± 3.60815) |
20.6396 us (± 3.7317) |
0.96 |
comp_asmjit_ifThenElse |
32.548 us (± 5.01094) |
32.008 us (± 5.75999) |
1.02 |
comp_asmjit_deeplyNestedIfElse |
54.6524 us (± 8.07573) |
54.3531 us (± 8.14402) |
1.01 |
comp_asmjit_loop |
34.101 us (± 5.32475) |
33.7759 us (± 4.23803) |
1.01 |
comp_asmjit_ifInsideLoop |
56.3058 us (± 11.5638) |
53.9387 us (± 5.94154) |
1.04 |
comp_asmjit_loopDirectCall |
45.236 us (± 10.1102) |
44.0242 us (± 8.85063) |
1.03 |
comp_asmjit_pointerLoop |
46.3312 us (± 7.93411) |
46.0983 us (± 8.87218) |
1.01 |
comp_asmjit_staticLoop |
27.0939 us (± 4.20994) |
27.1825 us (± 4.14008) |
1.00 |
comp_asmjit_fibonacci |
41.3349 us (± 8.10415) |
41.5607 us (± 7.09007) |
0.99 |
comp_asmjit_gcd |
33.4159 us (± 4.89637) |
34.0827 us (± 5.38372) |
0.98 |
comp_asmjit_nestedIf10 |
100.642 us (± 10.9382) |
100.924 us (± 9.12832) |
1.00 |
comp_asmjit_nestedIf100 |
1.02162 ms (± 13.6663) |
1.02018 ms (± 11.7866) |
1.00 |
comp_asmjit_chainedIf10 |
151.903 us (± 14.2155) |
152.283 us (± 13.5281) |
1.00 |
comp_asmjit_chainedIf100 |
2.14853 ms (± 32.2963) |
2.13238 ms (± 35.6257) |
1.01 |
tiered_compile_addOne |
54.1884 us (± 13.4109) |
51.9377 us (± 13.0686) |
1.04 |
single_compile_mlir_addOne |
6.37705 ms (± 218.012) |
6.10715 ms (± 63.2407) |
1.04 |
single_compile_cpp_addOne |
24.6656 ms (± 509.553) |
24.4534 ms (± 271.946) |
1.01 |
single_compile_bc_addOne |
69.2658 us (± 21.9383) |
51.9263 us (± 8.48295) |
1.33 |
tiered_compile_sumLoop |
75.0086 us (± 11.7772) |
73.8566 us (± 13.3747) |
1.02 |
single_compile_mlir_sumLoop |
8.69136 ms (± 274.968) |
8.15025 ms (± 122.033) |
1.07 |
single_compile_cpp_sumLoop |
26.0165 ms (± 872.992) |
25.6265 ms (± 346.062) |
1.02 |
single_compile_bc_sumLoop |
77.0162 us (± 14.9975) |
73.7859 us (± 9.66704) |
1.04 |
trace_add |
2.32013 us (± 213.105) |
2.27443 us (± 166.248) |
1.02 |
completing_trace_add |
2.2852 us (± 199.618) |
2.30772 us (± 203.032) |
0.99 |
trace_ifThenElse |
8.78943 us (± 1.19485) |
8.79821 us (± 706.962) |
1.00 |
completing_trace_ifThenElse |
4.55249 us (± 455.198) |
4.5873 us (± 359.51) |
0.99 |
trace_deeplyNestedIfElse |
26.29 us (± 3.12233) |
25.9509 us (± 1.83273) |
1.01 |
completing_trace_deeplyNestedIfElse |
12.7999 us (± 1.55701) |
12.719 us (± 1.13015) |
1.01 |
trace_loop |
8.72374 us (± 1.26144) |
8.60986 us (± 668.278) |
1.01 |
completing_trace_loop |
4.73607 us (± 574.716) |
4.69067 us (± 293.036) |
1.01 |
trace_ifInsideLoop |
17.4882 us (± 2.62388) |
17.1019 us (± 1.72681) |
1.02 |
completing_trace_ifInsideLoop |
8.69817 us (± 1.03793) |
8.49425 us (± 815.985) |
1.02 |
trace_loopDirectCall |
9.00565 us (± 818.836) |
8.87258 us (± 695.472) |
1.01 |
completing_trace_loopDirectCall |
4.96096 us (± 533.35) |
4.8027 us (± 315.653) |
1.03 |
trace_pointerLoop |
13.1579 us (± 1.46027) |
14.3937 us (± 1.20265) |
0.91 |
completing_trace_pointerLoop |
9.95885 us (± 1.20702) |
10.4807 us (± 1.59059) |
0.95 |
trace_staticLoop |
7.63257 us (± 873.921) |
7.39065 us (± 849.79) |
1.03 |
completing_trace_staticLoop |
7.66787 us (± 783.017) |
7.4703 us (± 715.621) |
1.03 |
trace_fibonacci |
10.3169 us (± 1.84159) |
10.163 us (± 1.71855) |
1.02 |
completing_trace_fibonacci |
6.24658 us (± 843.983) |
6.05346 us (± 710.97) |
1.03 |
trace_gcd |
8.13646 us (± 1.30721) |
7.99856 us (± 1.09167) |
1.02 |
completing_trace_gcd |
4.0229 us (± 413.402) |
3.98571 us (± 405.803) |
1.01 |
trace_nestedIf10 |
39.2338 us (± 6.67466) |
38.6445 us (± 6.35823) |
1.02 |
completing_trace_nestedIf10 |
38.7864 us (± 6.15003) |
39.1952 us (± 6.89562) |
0.99 |
trace_nestedIf100 |
1.36722 ms (± 20.738) |
1.36513 ms (± 19.5193) |
1.00 |
completing_trace_nestedIf100 |
1.39942 ms (± 20.8884) |
1.38142 ms (± 29.3291) |
1.01 |
trace_chainedIf10 |
99.7868 us (± 9.12703) |
100.633 us (± 10.9977) |
0.99 |
completing_trace_chainedIf10 |
49.9673 us (± 7.11158) |
49.6066 us (± 7.49433) |
1.01 |
trace_chainedIf100 |
4.46633 ms (± 90.4112) |
4.45201 ms (± 36.6404) |
1.00 |
completing_trace_chainedIf100 |
2.23488 ms (± 63.0638) |
2.27281 ms (± 21.1455) |
0.98 |
This comment was automatically generated by workflow using github-action-benchmark.
PhilippGrulich
force-pushed
the
claude/constant-value-tracer-BTcHn
branch
9 times, most recently
from
April 18, 2026 09:58
af6d74e to
695f469
Compare
Bundles the trace-time partial-evaluation plugin (ENABLE_CONSTANT_TRACER,
default off) and the TraceContextRegistry that swaps tracer selection
from a hard-coded if/else onto a registry lookup mirroring
CompilationBackendRegistry.
Core:
- TraceContextRegistry (nautilus/tracing/): Meyer's-singleton that maps
a trace-mode name to a TraceContextBase instance. Core registers
"lazyTracing" and "exceptionBasedTracing" in the registry constructor;
the PE plugin's static-init registers "partialEvaluation".
LegacyCompiler::compileToIR picks the mode from engine.traceMode and
dispatches via the instance's virtual Trace() — no core-side
awareness of the PE subclass.
- TraceContextBase grows a pure-virtual Trace(...). The existing
LazyTraceContext::Trace and ExceptionBasedTraceContext::Trace move
from static to virtual overrides; registry-owned instances are
dormant factories whose Trace() forwards to a thread_local worker
of the same concrete type.
- TracingInterface grows a virtual isPartialEvaluationMode() — PE's
PELazyTraceContext::onActivate/onDeactivate use it to flip the PE
runtime flag on while PE is the active tracer, so mode selection is
the sole source of truth for whether LazyTracedRef's fold fast paths
and the Constant registry are live.
PE plugin (plugins/partial_evaluation/, ENABLE_CONSTANT_TRACER=ON):
- PELazyTraceContext : public LazyTraceContext — single direct
subclass (the earlier CRTP PETraceContextMixin over both
LazyTraceContext and ExceptionBasedTraceContext is retired).
Overrides the eight trace ops that can split or merge the trace
with pe::materializeAllConstants() + base call; overrides
onActivate/onDeactivate/beforeInnerFunction for counter reset,
Constant-registry sweep, and PE-flag toggling.
- LazyTracedRef<T>, Constant registry, Stage 2 stratified widening
(per-site unroll_scope + blunt iteration cap), observability
counters, assume_stable RAII, example/DemoConstantFolding.cpp.
- ConstantTracerTest: fold correctness, Stage 2 widening behavior,
exceptionBasedTracing-mode sanity.
Testing:
- forEachBackend / forEachBackendWithTraceMode iterate trace-mode
names from TraceContextRegistry. Under ENABLE_CONSTANT_TRACER=ON the
default mode is "partialEvaluation" and TracingTest iterates only
that mode so the committed golden-trace overrides (780 files under
test/data/constant-tracer/) match. ReferenceDumpHelper now takes an
explicit `usePartialEvaluationOverride` flag instead of sniffing
ENABLE_CONSTANT_TRACER.
- Plugin tests for specialization and inlining iterate
constantTracerModes() unchanged (currently {false}) so those call
sites compile without edits; a follow-up can switch them to mode
iteration once they care.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…tics Add virtual TraceContextBase::collectStatistics(CompilationStatistics&) no-op hook; PE overrides it to populate constantTracer.* keys. LegacyCompiler::compileToIR now calls the virtual instead of the #ifdef ENABLE_CONSTANT_TRACER block, removing the last PE-specific ifdef from core. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PhilippGrulich
force-pushed
the
claude/constant-value-tracer-BTcHn
branch
from
April 19, 2026 16:30
695f469 to
ac7957e
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Introduces LazyTracedRef, a lazy state holder for val that is either
a pure compile-time constant (no trace footprint, no aliveVars entry) or
a materialized TypedValueRef. Materialization happens on the first
implicit conversion to TypedValueRef, so existing code that passes
value.stateto trace*Op functions keeps working unchanged.When ENABLE_CONSTANT_TRACER is ON (default OFF), val and
val:
(dead-branch elimination),
time anyone reads
stateas a TypedValueRef.When OFF, val behavior is byte-identical to before; all 165 existing
tests still pass with the option disabled.
KNOWN INCOMPLETE — the design's "no elision crosses a basic-block
boundary" invariant is documented in LazyTracedRef.hpp but NOT yet
enforced. Programs that re-assign a previously-Constant val inside a
branch (e.g. chainedIf100) currently produce traces missing the
required phi for that var, so several existing tests fail with the
option ON. The fix path (val-registry that materialize()s every live
Constant on traceBool) is sketched in the same comment.
Stage 2 (loop-aware widening via Snapshot stratification) is not
implemented; it requires restructuring SymbolicExecutionContext and
is documented as future work.
https://claude.ai/code/session_01H6BejBjwWYmoEXG6Vt6on4