Speed Benchmarks: Silu vs CPython vs Rust (2026-10-25, v1.9.0)

Generated by python benchmarks/run_benchmarks.py --report
(median of 3 runs per program, including process startup; Apple Silicon;
CPython 3.13 / Rust 1.89 release / clang -O2). For methodology and
per-program attribution see benchmarks/README.md.
Reproduce: python benchmarks/run_benchmarks.py --report.

1. Runtime Speed: 15 programs (incl. Json twin) × 4 planes

Program	Silu Interpreter	CPython	Silu Native	Rust	Native vs Rust
arith	5.532s	0.080s	0.0033s	0.0029s	0.9×
dict	0.269s	0.034s	0.0022s	0.0028s	1.3×
fannkuch	0.049s	0.032s	0.0026s	0.0025s	0.9×
fib	1.814s	0.036s	0.0025s	0.0028s	1.1×
float	3.791s	0.052s	0.0027s	0.0028s	1.0×
jsonparse	1.747s	0.047s	0.0074s	0.0036s	0.5×
jsonparse_json	1.816s	0.042s	0.0079s	—	(Json twin, see RFC 7c)
knucleotide	15.614s	0.889s	0.0110s	0.0148s	1.3×
list	0.479s	0.035s	0.0024s	0.0024s	1.0×
mandelbrot	1.214s	0.041s	0.0024s	0.0027s	1.2×
nbody	61.200s	0.214s	0.0057s	0.0054s	0.9×
spectral	71.253s	0.288s	0.0074s	0.0047s	0.6×※
string	0.553s	0.049s	0.0030s	0.0022s	0.7×
trees	0.873s	0.031s	0.0032s	0.0034s	1.1×
wordcount	6.740s	0.088s	0.0233s	0.0137s	0.6×

**Geometric mean: 0.9× vs Rust (slightly behind overall) | 15.6× vs CPython | 515× vs interpreter**; startup baseline: silu-int 0.038s / cpython 0.028s / binary 0.0041s. 10 of 15 rows at ≥0.9× (dict/knucleotide lead at 1.3×).

Runtime slimming note: LTO + -Wl,-dead_strip strips unreferenced runtime functions at symbol granularity (fib binary 35KB vs 113KB monolithic). The "per-program runtime selection" item is closed — the linker does it for free.

jsonparse five-way comparison (quiet environment, 31 runs, 87KB document, all outputs byte-identical):

Plane	min	Notes
C (clang -O2, isomorphic tree+realloc, benchmarks/c/jsonparse.c)	2.7ms	Median unstable (10-27ms swing, malloc-sensitive)
Rust (serde_json shape)	3.6ms	Stable
Silu class twin	6.5ms	Stable
Silu Json twin	6.7ms	Stable
CPython (pycompat)	42ms
Silu interpreter	1735ms

LTO erased RFC 7c's 1.2ms: the Json twin now matches the class twin (unwrap/tag dispatch inlined into hot loops by -flto). The remaining gap vs Rust/C still attributes to construction (~1.9ms, the J1 three-layer decomposition).

※ spectral high-variance note (v1.14.0 retest): the same binary measured 4.5–7.4ms across three environments (millisecond-level programs swing percentage wildly on ms jitter); and the link-strategy cost was isolated — the same .ll compiled monolithic via clang: 3.86ms (1.17× overtakes Rust), while siluc's "user.o + rt.o" split link: 5.58ms (0.81×): the cross-TU boundary blocks runtime-function inlining. LTO monolithic verified at 3.89ms.

Landed (f475966): LTO link is now the siluc default — rt.o stores LLVM bitcode (one-time 0.15s), the user module compiles to bitcode, linking runs clang -flto -O2 -Wl,-dead_strip (no conflicts). Cost: +20ms per link (0.05 vs 0.03s), siluc total warm stays 0.11-0.12s. Benefit (quiet-environment hand-timed, vs Rust): spectral 4.66→4.15ms, fib 1.08×, dict 1.04×, arith 1.03×; startup baseline halved (4.1→2.1ms).

2. Compile Speed: siluc vs rustc vs clang

Compiler	fib (s/program)	Notes
siluc (C driver + siluc1)	0.136s	One command, zero Python; runtime .o cached
rustc -C opt-level=3	0.128s	Direct call, no cache (three consecutive compiles constant + zero disk residue, verified)
Python pipeline (selfhost_build)	0.178s	Orchestration is Python; the compiler itself is still siluc1
Python compiler (silu build)	0.497s	Full Python implementation; for development/debugging
clang -O2 (C twin)	0.06s	Reference

siluc vs rustc gap is 1.06×; the remainder is entirely the clang user-segment + link proper (the pipeline's ~0.12s is the clang floor). cargo's cache is artifact-level (cold 0.274s / warm 0.050s), similar to our .o cache but coarser-grained.

3. Speed Evolution Trail (what worked)

Date	Milestone	jsonparse	Full-suite geomean vs Rust
10-14	Baseline (pre-v1.0)	78ms (0.06×)	0.9× (Rust slightly ahead)
10-24	Print specialization: static scalars/strings skip pyobj boxing	—	—
10-25	dead_strip linking (binary size 3×↓)	—	—
10-25	Match-class pattern narrowing: case JInt(n) payload static binding	78→75ms	—
10-25	Borrowed str parameters: fixed per-recursion deep copy of the whole document	78.7→8.3ms (9.5×)	—
10-25	Fresh report (all fixes)	10.7ms (0.4×)	0.9× (parity range)

Two attributed remaining gaps:

• **jsonparse 0.6× (10-27 third round: 8.0-8.5ms vs rust 4.5-4.8ms, post

materialization pooling)**: three-layer attribution (construction 12% / materialization the bulk) and fix chain — v1.11.4's pool was initially attached to the zero-call rawstr_char_at (its 0.8ms gain was noise, corrected); the real hot path was s[i]'s view→RAWSTR materialization (rawstr_new_copy: 2 mallocs + memcpy + SIMD code-point count, ~62ns/char). This round connected materialization to rawstr_char1 (one-byte code points serve from a 256-entry shared pool, cap=-1 immortal sentinel): scan probe 9.9→4.7ms, **at parity with the hand-written ord() idiom (62→~5ns/char), jsonparse 9.8→8.0-8.5ms. Inline strategy measured and rejected**: clang -O2 already inlines small runtime functions (verified in assembly); the remaining single char1 call with alwaysinline actually regressed 4.64→4.91ms (icache damage). **Remaining gap ≈ construction cost 1.9ms (pure-runtime 5.2 vs 3.5ms, gap 1.7ms)** — sum-types (RFC S1-S4) promoted from "secondary" to the closing critical path: post-materialization construction share rose from 12% to ~24%.

• wordcount 0.6× (10-27 retest): v1.11.3's subscript type inference

eliminated w = vocab[i]'s pyobj boxing (.ll call sequence verified), but exposed a deeper layer: rawlist_str_get has copy semantics (slots own copies) — each read mallocs a new block + frees the old slot, two heap operations actually costlier than single boxing. A/B measured (old .ll 21.2ms vs new 24.3ms, 15-run min each). The fix = borrowed reads (immutable-element subscripts return borrowed pointers) or buffer reuse (rawstr_copy_into, Rust's clone_from approach) — the same immutable-value copy-semantics family as jsonparse's char access, treatable together.

4. Methodology Notes

• Four planes, verify before timing: each program first confirms

interpreter/CPython/compiled-artifact/Rust-twin outputs are identical; otherwise the row is marked MISMATCH and excluded.

• CPython plane goes through the pycompat shim (Silu classes use positional

constructors, differing from bare CPython semantics); Rust twins are hand-mirrored per program and output-compared.

• Compile timing is median of 3; startup baseline separately deducted

(4.3ms silu / 2.3ms rust empty program).

• Per-item attribution uses ablation + macOS sample profiling (example:

jsonparse's 70ms hot zone was sampled and pinned to parameter deep-copy; fixed to 8.3ms).