2026-01-03-yklua-profiling

2026-01-03-yklua-profiling.md

Experiment 09: JIT vs Interpreter Time Analysis

This experiment measures the time spent in JIT-compiled code versus the interpreter.

Methodology

JIT Time         = duration_jit_executing
Interpreter Time = wall_clock - jit_executing - trace_mapping - deopting

Benchmark Results

Benchmark	Size	Iters	JIT (s)	JIT %	Interp (s)	Interp %	Total (s)
Bounce	1500	3	2.62	97.37%	0.07	2.52%	2.69
CD	250	3	9.65	97.09%	0.23	2.31%	9.94
DeltaBlue	12000	3	1.76	44.90%	2.03	51.98%	3.91
Havlak	1500	3	20.03	87.51%	2.79	12.18%	22.89
Json	100	3	3.83	95.30%	0.16	4.00%	4.02
List	1500	3	2.46	99.07%	0.02	0.81%	2.48
lulpeg	—	3	3.61	84.77%	0.52	12.27%	4.26
Mandelbrot	500	3	0.37	96.99%	0.01	2.75%	0.39
NBody	250000	3	1.38	97.41%	0.03	2.45%	1.42
Permute	1000	3	2.51	98.28%	0.04	1.45%	2.55
Queens	1000	3	1.44	98.12%	0.02	1.47%	1.47
Richards	100	3	12.38	99.44%	0.05	0.40%	12.45
Sieve	3000	3	1.10	98.88%	0.01	1.12%	1.12
Storage	1000	3	7.14	99.55%	0.03	0.39%	7.18
Towers	600	3	2.52	95.18%	0.12	4.52%	2.65

Analysis: High Interpreter Time Benchmarks

Three benchmarks spend significant time in the interpreter (>10%)

Benchmark	Interp %	Traces Compiled	Trace Executions	Compilation (s)
DeltaBlue	51.98%	132	103,426	4.10
lulpeg	12.27%	818	79,724	13.91
Havlak	12.18%	302	37,006	3.70

Perf Profiles

DeltaBlue (51.98% interpreter)

Top functions by CPU time:

%	Function	Category
11.96	__yk_opt_luaH_getshortstr	Interpreter
9.95	luaV_execute	Interpreter
7.15	__memmove_avx512_unaligned	Memory (libc)
6.31	Inst::decopy_eq (ykrt)	JIT Compiler
3.82	_int_malloc	Memory (libc)
2.43	__yk_trace_14	JIT Trace
1.96	Opt::cse (ykrt)	JIT Compiler
1.71	luaC_step	GC

Highlight: ~22% of time is spent in interpreter functions (luaV_execute, luaH_getshortstr), but deopt time is minimal, only 0.116 seconds.

lulpeg (12.27% interpreter)

Top functions by CPU time:

%	Function	Category
17.67	Inst::decopy_eq (ykrt)	JIT Compiler
8.47	__memmove_avx512_unaligned	Memory (libc)
5.23	Opt::cse (ykrt)	JIT Compiler
3.06	__yk_trace_90	JIT Trace
2.67	__yk_trace_242	JIT Trace
2.30	moveresults_full	Interpreter
2.14	luaV_execute	Interpreter
1.75	str_byte	Interpreter

Highlight: ~23% of time is spent in JIT compiler functions (decopy_eq, cse).

Havlak (12.18% interpreter)

Top functions by CPU time:

%	Function	Category
9.43	__yk_opt_luaH_getshortstr	Interpreter
6.17	_int_malloc	Memory (libc)
6.08	malloc	Memory (libc)
4.65	malloc_consolidate	Memory (libc)
4.23	__yk_opt_propagatemark	GC
4.11	luaC_step	GC
3.68	_int_free_chunk	Memory (libc)
3.03	Inst::decopy_eq (ykrt)	JIT Compiler
2.35	__yk_opt_luaC_step	GC

Highlight: ~20% of time is spent in memory allocation (malloc, free) and ~11% in GC (luaC_step, propagatemark).