Lambda Benchmark Results: 6 Suites × 6 Engines (Round 4)

Date: 2026-03-19
Platform: Apple Silicon MacBook Air (M4, aarch64), macOS
Lambda version: release build (8.4 MB, stripped, -O2)
Node.js: v22.13.0 (V8 JIT)
QuickJS: v2025-09-13 (interpreter)
Python: 3.13.3 (CPython)
Methodology: 3 runs per benchmark, median of self-reported execution time (excludes startup/JIT compilation overhead)

---

Engine Overview

Engine	Type	Description
MIR Direct	JIT	Lambda → MIR IR → native code (default compiler path)
C2MIR	JIT	Lambda → C source → MIR (legacy path via c2mir)
LambdaJS	JIT	Lambda's built-in JavaScript JIT
QuickJS	Interpreter	Standalone QuickJS JavaScript engine
Node.js	JIT	Google V8 JavaScript engine with optimizing JIT
Python	Interpreter	CPython 3.13 reference interpreter

---

R7RS Benchmarks

Classic Scheme benchmark suite adapted for Lambda with type annotations. Tests recursive functions, numeric computation, and backtracking.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
fib	recursive	2.5	2.1	0.99	18	2.0	22	1.23x	0.11x
fibfp	recursive	3.7	3.5	1.0	19	1.8	23	2.11x	0.16x
tak	recursive	0.15	0.16	0.10	2.9	0.80	2.2	0.19x	0.07x
cpstak	closure	0.30	0.34	0.22	5.7	1.00	4.5	0.30x	0.07x
sum	iterative	0.27	1.8	94	32	1.2	38	0.23x	0.007x
sumfp	iterative	0.067	0.33	14	3.8	0.87	2.8	0.08x	0.02x
nqueens	backtrack	6.7	6.6	6.8	9.7	1.8	3.5	3.72x	1.90x
fft	numeric	0.18	1.0	9.8	2.8	1.7	4.3	0.11x	0.04x
mbrot	numeric	0.59	0.74	55	18	1.8	15	0.34x	0.04x
ack	recursive	9.8	9.8	8.1	---	14	156	0.72x	0.06x

Geometric mean MIR/Node.js: 0.44x — Lambda faster on 7/10 benchmarks Geometric mean MIR/Python: 0.07x — Lambda faster on 9/10 benchmarks

---

AWFY Benchmarks

Standard cross-language benchmark suite from Stefan Marr. Lambda implementations use procedural style; JS uses official AWFY source.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
sieve	micro	0.052	0.052	0.77	0.60	0.38	1.76s	0.14x	0.000x
permute	micro	0.064	0.068	13	1.6	0.81	2.11s	0.08x	0.000x
queens	micro	0.15	0.14	11	1.1	0.64	1.14s	0.23x	0.000x
towers	micro	0.22	0.11	23	2.3	1.1	1.11s	0.19x	0.000x
bounce	micro	0.19	0.14	10	0.96	0.55	1.39s	0.35x	0.000x
list	micro	0.023	0.62	7.9	0.92	0.50	976	0.05x	0.000x
storage	micro	0.19	0.32	6.2	2.7	0.64	1.27s	0.30x	0.000x
mandelbrot	compute	32	55	279	888	32	---	1.00x	---
nbody	compute	47	84	2.06s	167	5.6	135	8.38x	0.35x
richards	macro	253	219	3.31s	194	48	168	5.29x	1.50x
json	macro	1.5	2.1	160	12	2.8	7.1	0.54x	0.21x
deltablue	macro	64	78	935	113	13	68	5.07x	0.94x
havlak	macro	61	82	39.66s	4.09s	92	2.11s	0.66x	0.03x
cd	macro	220	358	11.66s	1.06s	37	---	5.94x	---

Geometric mean MIR/Node.js: 0.62x — Lambda faster on 10/14 benchmarks Geometric mean MIR/Python: 0.00x — Lambda faster on 11/12 benchmarks

---

BENG Benchmarks

Subset of the Computer Language Benchmarks Game. Tests diverse real-world computation: GC stress, regex, FASTA I/O, numeric precision, permutations.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
binarytrees	allocation	7.3	7.7	114	28	4.1	10	1.76x	0.70x
fannkuch	permutation	0.76	1.1	1.6	7.3	4.1	5.1	0.18x	0.15x
fasta	generation	1.1	0.84	3.9	11	6.2	2.0	0.18x	0.55x
knucleotide	hashing	2.9	3.8	0.088	---	5.0	3.9	0.58x	0.74x
mandelbrot	numeric	142	238	2.85s	698	16	1.37s	9.13x	0.10x
nbody	numeric	47	85	1.75s	155	8.1	172	5.84x	0.27x
pidigits	bignum	0.46	0.30	0.083	0.16	2.0	0.10	0.23x	4.53x
regexredux	regex	1.2	1.4	0.095	---	2.5	1.5	0.50x	0.85x
revcomp	string	1.8	1.8	0.002	---	3.4	0.085	0.54x	21.7x
spectralnorm	numeric	13	10	80	64	2.8	47	4.72x	0.28x

Geometric mean MIR/Node.js: 0.94x — Lambda faster on 6/10 benchmarks Geometric mean MIR/Python: 0.70x — Lambda faster on 8/10 benchmarks

---

KOSTYA Benchmarks

Community benchmarks from kostya/benchmarks comparing languages on common tasks.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
brainfuck	interpreter	165	280	2.31s	906	45	691	3.70x	0.24x
matmul	numeric	8.8	128	2.83s	546	16	535	0.56x	0.02x
primes	numeric	7.3	10	25	97	4.5	97	1.62x	0.08x
base64	string	220	221	900	182	18	85	12.5x	2.59x
levenshtein	string	7.7	13	71	55	4.0	71	1.92x	0.11x
json_gen	data	65	67	79	21	6.3	8.3	10.3x	7.76x
collatz	numeric	301	340	18.53s	6.22s	1.42s	8.00s	0.21x	0.04x

Geometric mean MIR/Node.js: 2.09x — Lambda slower on 2/7 benchmarks Geometric mean MIR/Python: 0.22x — Lambda faster on 5/7 benchmarks

---

LARCENY Benchmarks

Classic Gabriel/Larceny Scheme benchmark suite testing diverse functional programming patterns.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
triangl	search	179	1.11s	6.82s	2.23s	68	2.68s	2.62x	0.07x
array1	array	0.55	5.8	0.55	37	1.8	40	0.30x	0.01x
deriv	symbolic	20	21	894	69	3.8	26	5.35x	0.76x
diviter	iterative	272	271	61.97s	26.85s	473	26.25s	0.57x	0.01x
divrec	recursive	0.84	7.4	0.82	38	7.9	45	0.11x	0.02x
gcbench	allocation	469	439	2.86s	667	25	257	19.0x	1.83x
paraffins	combinat	0.33	0.92	6.1	2.8	1.0	2.9	0.33x	0.11x
pnpoly	numeric	59	53	312	206	6.1	112	9.67x	0.52x
primes	iterative	7.2	9.7	26	97	4.7	121	1.53x	0.06x
puzzle	search	3.8	17	82	29	3.2	21	1.16x	0.18x
quicksort	sorting	3.1	6.8	55	19	1.6	26	1.90x	0.12x
ray	numeric	7.1	6.9	40	14	3.5	12	2.02x	0.61x

Geometric mean MIR/Node.js: 1.47x — Lambda slower on 4/12 benchmarks Geometric mean MIR/Python: 0.12x — Lambda faster on 11/12 benchmarks

---

JetStream Benchmarks

Benchmarks from Apple's JetStream suite (SunSpider + Octane). Tests numeric computation, 3D rendering, crypto, and data structures.

Benchmark	Category	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/Py
nbody	numeric	47	85	1.86s	---	5.5	146	8.50x	0.32x
cube3d	3d	24	81	21	228	18	46	1.31x	0.52x
navier_stokes	numeric	823	809	0.23	95	14	1.84s	56.8x	0.45x
richards	macro	259	245	566	---	8.3	225	31.3x	1.15x
splay	data	165	---	49	199	20	326	8.07x	0.51x
deltablue	macro	17	18	48	---	11	18	1.61x	0.96x
hashmap	data	106	106	---	323	16	184	6.50x	0.58x
crypto_sha1	crypto	17	20	144	222	9.0	321	1.85x	0.05x
raytrace3d	3d	348	435	720	170	19	144	18.6x	2.42x

Geometric mean MIR/Node.js: 7.28x — Lambda slower on 0/9 benchmarks Geometric mean MIR/Python: 0.52x — Lambda faster on 7/9 benchmarks

---

Overall Summary

MIR Direct vs Node.js V8 (Self-Reported Exec Time)

Suite	Geo. Mean	Lambda Wins	Node Wins	Total
R7RS	0.44x	7	3	10
AWFY	0.62x	10	4	14
BENG	0.94x	6	4	10
KOSTYA	2.09x	2	5	7
LARCENY	1.47x	4	8	12
JetStream	7.28x	0	9	9
Overall (raw)	1.21x	29	33	62
Overall (dedup)	1.05x	28	28	56

Ratio < 1.0 = Lambda MIR is faster. Ratio > 1.0 = Node.js is faster. Dedup note: 56 unique benchmarks out of 62 total entries. Duplicates (same name across suites): deltablue, mandelbrot, nbody, primes, richards — best time per engine is used.

Excluding JetStream (50 unique benchmarks): 0.83x — Lambda wins 28/50

MIR Direct vs Python (Self-Reported Exec Time)

Suite	Geo. Mean	Lambda Wins	Python Wins	Total Compared
R7RS	0.07x	9	1	10
AWFY	0.00x	11	1	12
BENG	0.70x	8	2	10
KOSTYA	0.22x	5	2	7
LARCENY	0.12x	11	1	12
JetStream	0.52x	7	2	9
Overall (raw)	0.09x	51	9	60
Overall (dedup)	0.08x	47	8	55

Lambda MIR is overwhelmingly faster than CPython across all suites.

---

Performance Tiers (MIR vs Node.js)

Tier	Count	Benchmarks
Lambda >2× faster (< 0.5×)	20	awfy/list (0.05x), r7rs/sumfp (0.08x), awfy/permute (0.08x), larceny/divrec (0.11x), r7rs/fft (0.11x), awfy/sieve (0.14x), beng/fasta (0.18x), beng/fannkuch (0.18x), r7rs/tak (0.19x), awfy/towers (0.19x), kostya/collatz (0.21x), awfy/queens (0.23x), r7rs/sum (0.23x), beng/pidigits (0.23x), larceny/array1 (0.30x), awfy/storage (0.30x), r7rs/cpstak (0.30x), larceny/paraffins (0.33x), r7rs/mbrot (0.34x), awfy/bounce (0.35x)
Lambda faster (0.5–1.0×)	9	beng/regexredux (0.50x), beng/revcomp (0.54x), awfy/json (0.54x), kostya/matmul (0.56x), larceny/diviter (0.57x), beng/knucleotide (0.58x), awfy/havlak (0.66x), r7rs/ack (0.72x), awfy/mandelbrot (1.00x)
Comparable (1.0–2.0×)	10	larceny/puzzle (1.16x), r7rs/fib (1.23x), jetstream/cube3d (1.31x), larceny/primes (1.53x), jetstream/deltablue (1.61x), kostya/primes (1.62x), beng/binarytrees (1.76x), jetstream/crypto_sha1 (1.85x), larceny/quicksort (1.90x), kostya/levenshtein (1.92x)
Node faster (2.0–5.0×)	6	larceny/ray (2.02x), r7rs/fibfp (2.11x), larceny/triangl (2.62x), kostya/brainfuck (3.70x), r7rs/nqueens (3.72x), beng/spectralnorm (4.72x)
Node >5× faster (> 5.0×)	17	awfy/deltablue (5.07x), awfy/richards (5.29x), larceny/deriv (5.35x), beng/nbody (5.84x), awfy/cd (5.94x), jetstream/hashmap (6.50x), jetstream/splay (8.07x), awfy/nbody (8.38x), jetstream/nbody (8.50x), beng/mandelbrot (9.13x), larceny/pnpoly (9.67x), kostya/json_gen (10.28x), kostya/base64 (12.51x), jetstream/raytrace3d (18.64x), larceny/gcbench (19.00x), jetstream/richards (31.32x), jetstream/navier_stokes (56.82x)

---

Improvement over Round 3

MIR Direct Performance Changes (R3 → R4)

Benchmark	Suite	R3 (ms)	R4 (ms)	Speedup
havlak	AWFY	183	61	3.00×
cd	AWFY	528	220	2.40×
json	AWFY	3.3	1.5	2.20×
cube3d	JetStream	49	24	2.08×
storage	AWFY	0.33	0.19	1.70×
deltablue	AWFY	99	64	1.54×
list	AWFY	0.032	0.023	1.39×

Overall MIR improvement (geo mean, 62 benchmarks): 8.0% faster

Suite-Level Comparison (MIR/Node.js Geometric Mean)

Suite	R3 Geo Mean	R4 Geo Mean	Change
R7RS	0.44x	0.44x	↓ (1% worse)
AWFY	0.83x	0.62x	↑ improved (1% better)
BENG	0.93x	0.94x	↓ (1% worse)
KOSTYA	2.09x	2.09x	↓ (1% worse)
LARCENY	1.48x	1.47x	↑ improved (1% better)
JetStream	7.90x	7.28x	↑ improved (1% better)

Lambda JS Engine — R4 Status

LambdaJS was not re-run in R4 (only MIR and C2MIR engines were benchmarked). LambdaJS results are carried over from R3. Current R3 coverage:

• AWFY: All 14 benchmarks passing (bounce, storage, json, deltablue, havlak, cd included)
• R7RS: All 10 benchmarks (except mbrot)
• BENG: binarytrees, fannkuch, fasta, knucleotide, mandelbrot, nbody, pidigits, regexredux, revcomp, spectralnorm
• KOSTYA: brainfuck, matmul, primes, levenshtein, json_gen, collatz
• LARCENY: All 12 benchmarks

---

Memory Profiling (Peak RSS)

Peak resident set size measured via /usr/bin/time -l (macOS). Values in MB. Includes runtime, JIT compiler, and all loaded libraries for each engine.

R7RS — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
fib	36	35	32	1.8	40	11
fibfp	41	40	32	1.7	40	11
tak	36	35	31	1.8	40	11
cpstak	36	35	31	1.7	40	11
sum	35	35	31	1.7	40	11
sumfp	35	35	36	1.7	42	11
nqueens	39	37	72	1.7	43	11
fft	37	37	36	1.7	42	11
mbrot	37	36	---	1.7	42	11
ack	36	35	31	---	40	11

AWFY — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
sieve	36	35	32	1.9	41	---
permute	36	35	32	1.8	41	---
queens	36	35	32	1.7	42	---
towers	36	36	32	1.6	42	---
bounce	38	36	---	1.7	41	---
list	36	35	32	1.7	40	---
storage	37	36	---	2.4	42	---
mandelbrot	36	35	225	1.7	43	---
nbody	42	41	33	1.9	36	---
richards	77	74	37	1.8	44	---
json	47	45	---	3.0	43	---
deltablue	57	53	---	1.9	37	---
havlak	115	111	---	30	94	---
cd	69	68	---	4.3	57	---

BENG — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
binarytrees	47	46	72	2.9	43	11
fannkuch	38	36	32	1.9	41	11
fasta	41	38	33	1.9	44	11
knucleotide	41	42	32	---	43	12
mandelbrot	71	69	48	1.7	42	11
nbody	42	42	37	1.7	43	11
pidigits	37	36	32	1.7	37	11
regexredux	39	39	33	---	36	11
revcomp	38	38	33	---	38	11
spectralnorm	56	52	66	1.9	43	11

KOSTYA — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
brainfuck	38	200	323	2.3	104	11
matmul	41	286	649	2.7	44	12
primes	43	43	32	2.7	42	21
base64	1414	1902	35	6.4	49	12
levenshtein	38	37	36	1.9	41	11
json_gen	700	699	49	2.9	46	11
collatz	36	35	503	1.8	41	11

LARCENY — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
triangl	39	36	33	1.8	41	11
array1	36	35	31	1.8	40	11
deriv	53	53	106	1.9	44	11
diviter	36	35	31	1.6	41	11
divrec	36	35	31	2.2	40	11
gcbench	261	259	667	20	60	15
paraffins	40	37	32	1.8	42	11
pnpoly	100	96	49	1.7	43	11
primes	37	36	32	1.8	41	11
puzzle	37	36	---	1.8	41	11
quicksort	37	36	32	1.9	42	11
ray	49	48	42	1.7	43	11

JetStream — Peak RSS (MB)

Benchmark	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python
nbody	122	190	---	---	44	---
cube3d	92	79	---	---	55	---
navier_stokes	1310	1307	---	---	49	---
richards	312	310	---	---	44	---
splay	240	---	---	---	147	---
deltablue	57	54	---	---	51	---
hashmap	49	48	---	---	52	---
crypto_sha1	53	47	---	---	44	---
raytrace3d	148	145	---	---	53	---

Memory Summary — Average Peak RSS per Suite (MB)

Suite	MIR	C2MIR	LambdaJS	QuickJS	Node.js	Python	MIR/Node	MIR/QJS
R7RS	37	36	37	1.7	41	11	0.90x	21.5x
AWFY	50	48	57	4.1	46	---	1.09x	12.2x
BENG	45	44	42	2.0	41	11	1.10x	23.1x
KOSTYA	330	458	232	3.0	52	13	6.29x	111.5x
LARCENY	63	62	99	3.4	43	11	1.46x	18.8x
JetStream	265	272	---	---	60	---	4.44x	---

Lambda MIR uses 2.79× the memory of Node.js (132 MB vs 47 MB average). QuickJS is the most memory-efficient at 3 MB average — 47× less than Lambda MIR.

Memory footprint is dominated by engine/runtime overhead; actual benchmark data is small. QuickJS's tiny interpreter has minimal memory overhead. Node.js includes the full V8 engine.

---

Key Findings

1. Overall: Lambda MIR competitive with Node.js V8

Across 56 unique benchmarks with both MIR and Node.js results, the geometric mean ratio is 1.05x. Excluding the new JetStream suite (which ports are not yet optimized), Lambda achieves 0.83x on 50 unique benchmarks, winning 28 of 50.

2. Lambda MIR dominates CPython

Across 55 unique benchmarks with Python comparisons, Lambda MIR is 13× faster (geo mean 0.08x). CPython's interpreted execution cannot match JIT-compiled code on compute-intensive tasks. Lambda wins on all suites, with particular dominance on tight loops and numeric code (AWFY micro-benchmarks: 1000–30000× faster).

3. Strengths: Micro-benchmarks and numeric code

Lambda MIR excels on small, tight computational benchmarks:

• R7RS suite (0.43x): 2.3× faster on average — strong tail-call optimization, native integer/float arithmetic
• AWFY micro-benchmarks: sieve (0.14x), list (0.04x), permute (0.08x) — highly efficient JIT for simple loops
• FFT (0.11x): 9× speedup from typed array inline fast paths
• Collatz (0.21x): 5× faster than Node.js on integer-heavy iteration

4. Weaknesses: OOP-heavy and allocation-intensive code

Node.js V8's optimizing JIT (TurboFan) significantly outperforms Lambda on:

• Class-heavy benchmarks: richards (5.29x AWFY), cd (5.94x) — V8's hidden classes and inline caches
• Heavy allocation/GC: gcbench (19x), base64 (12.5x) — V8's generational GC advantage
• JetStream suite (7.28x): Complex OOP-style benchmarks where V8's mature optimizations dominate

5. JetStream: New frontier for optimization

The JetStream benchmarks (ported from Apple's JS benchmark suite) show Lambda MIR running slower than Node.js on all 9 benchmarks (geo mean ~7×). Workloads are now synchronized to the original heavy JetStream workloads. Key remaining bottlenecks:

• navier_stokes (57×): Heavy array-based PDE solver — needs typed array optimization for this pattern
• richards (31×): OOP task scheduler — 50 iterations × 1000 COUNT exposes class/method dispatch overhead
• raytrace3d (18×): Object-heavy 3D computation — property access patterns
• splay (8×): Red-black tree operations — property access patterns
• nbody (8.5×): Numeric simulation — 36000 advance steps per run
• deltablue (1.6×): Constraint solver — close to competitive at 20 iterations
• cube3d (1.3×): 3D rendering — much improved from R3 (49ms → 24ms) These represent clear optimization targets for future MIR engine improvements.

6. MIR JIT improvements from Round 3 (LMD_TYPE_LIST removal)

Removing LMD_TYPE_LIST and unifying list/array handling delivered significant improvements:

• havlak: 183 → 61ms (3.0× faster) — graph traversal, list-heavy data structure
• cd: 528 → 220ms (2.4× faster) — collision detection with many list operations
• json: 3.3 → 1.5ms (2.2× faster) — JSON macro benchmark
• cube3d: 49 → 24ms (2.0× faster) — 3D rendering with array operations
• storage: 0.33 → 0.19ms (1.7× faster) — storage micro-benchmark
• deltablue: 99 → 64ms (1.5× faster) — constraint solver macro benchmark
• list: 0.032 → 0.023ms (1.4× faster) — list micro-benchmark
• Overall: ~8% faster across 62 comparable benchmarks

7. Lambda JS engine

LambdaJS results are unchanged from R3 (not re-run in R4). R3 coverage:

• All AWFY benchmarks pass including bounce, storage, json, deltablue, havlak, and cd
• BENG/fannkuch, BENG/mandelbrot, KOSTYA/collatz passing

8. QuickJS comparison

QuickJS (pure interpreter) is generally 2–10× slower than Node.js V8, as expected. Lambda MIR is faster than QuickJS on most benchmarks, confirming Lambda's JIT advantage.

9. C2MIR vs MIR Direct

The two Lambda JIT paths produce similar results. Notable differences:

• C2MIR slightly faster on: r7rs/sum (0.27 vs 1.8ms), nqueens (6.7 vs 6.6ms)
• MIR Direct faster on: matmul (8.8 vs 128ms), cube3d (24 vs 81ms), list (0.023 vs 0.62ms)
• MIR Direct has lower compilation overhead and is the default path

10. Memory footprint

Lambda MIR's peak RSS averages ~2.8× Node.js memory, dominated by the MIR JIT compiler and runtime overhead. Key observations:

• R7RS/micro-benchmarks: MIR ~37 MB vs Node ~41 MB — Lambda is lighter for small programs
• KOSTYA/LARCENY/JetStream: MIR 63–330 MB vs Node 43–60 MB — Lambda's GC and data model use more RAM on heavy workloads
• QuickJS is the most memory-efficient at ~3 MB average (pure interpreter, minimal overhead)
• Outliers: base64 (1,414 MB MIR), navier_stokes (1,310 MB MIR) — indicate optimization opportunities for array-heavy benchmarks

---

Notes

• Self-reported exec time measures only the computation, excluding process startup, JIT compilation warmup, and file I/O.
• AWFY JS benchmarks use the official source from ref/are-we-fast-yet/benchmarks/JavaScript/. AWFY Python benchmarks use the official Python port with harness.
• AWFY Python micro-benchmarks (sieve, permute, queens, etc.) show extreme Lambda advantage because CPython interprets tight loops ~10,000× slower than JIT-compiled code.
• AWFY Python benchmarks use the official Python port with harness. Class names: NBody, DeltaBlue, CD (not capitalize()).
• LambdaJS now passes all AWFY benchmarks including bounce, storage, json, deltablue, havlak, and cd (results from R3, not re-run in R4).
• QuickJS fails on ack (R7RS) due to stack overflow on deep recursion.
• JetStream benchmarks run on MIR, C2MIR, Node.js, and Python (for deltablue, richards, nbody). No LambdaJS/QuickJS ports.
• Python benchmarks not available for: AWFY/cd, JetStream/cube3d, JetStream/navier_stokes, JetStream/splay, JetStream/hashmap, JetStream/crypto_sha1, JetStream/raytrace3d.
• All times in milliseconds unless noted with 's' suffix (seconds).
• KOSTYA/json_gen benchmark was broken in an earlier commit (commit f2f0c3fd9, incorrect string syntax); fixed in this round. Results are now correct (~65ms).
• LMD_TYPE_LIST removal (R4): LMD_TYPE_LIST was replaced by LMD_TYPE_ARRAY throughout the runtime. All list/array operations now use a unified type, improving performance on OOP-heavy and collection-intensive benchmarks.
• Workload synchronization: Duplicate benchmark names across suites use identical heavy workloads synchronized to original JetStream — AWFY/BENG/JetStream mandelbrot N=500, nbody 36000 steps, richards 50×COUNT=1000, deltablue 20×chain(100), Larceny/Kostya primes sieve(1M).