Xisto vs Rust

13 Nov 2026, main 4a2f388, on the two AWS VMs. Rust is the language the Benchmarks Game's fastest programs are usually written in, so this is the comparison that says whether xisto's generated C keeps up with a systems language.

Setup

• Rust programs: every Rust program of the official source package (BSD-3), not only the fastest, built as the site builds them: a Cargo project with the crates they use (typed-arena, rayon, toolshed, bumpalo, regex, hashbrown, ...), opt-level = 3, codegen-units = 1, -C target-cpu=native, GMP for pidigits (bench/benchmarksgame/langs.md). rustc 1.99.1. On arm64 the programs that use x86 SIMD do not compile (nbody #5/#7/#9, fannkuch-redux #6, regex-redux #7, pidigits #4).
• Excluded: nbody_7 and nbody_9 compute 1/√x with the approximate rsqrtps instruction and give different doubles; xisto promises Node's doubles, so it is not compared against a different computation.
• xisto: every variant in bench/benchmarksgame/, built with XISTO_CPU=native.
• Command: python3 bench/benchmarksgame/langs.py 3 rust=~/rustbg/target/release, scored with python3 bench/placar.py RESULTS (see How every run is measured).
• Excluded xisto variant: spectralnorm-8 takes 1/d from a float32 division refined to the same double. It is measured (0.422 s on the Xeon, 0.874 s on the Graviton2) but not scored: the problem asks for the double division, and the scoreboard keeps the problem as stated.
• Steal during the round: +229 on the Xeon, +638 on the Graviton2.

Xeon E5-2686 v4 (AWS t2.medium, 2 vCPUs, AVX2)

Source: bench/results/vm-2026-11-13/rust-ubuntu-xeon-2026-11-13-4a2f388.txt.

problem	xisto	Rust	xisto time	Rust time	xisto CPU	Rust CPU	xisto mem	Rust mem	result
fasta	fasta	fasta_7	1.124 s	1.989 s	0.34 s	2.66 s	2.4 MB	2.9 MB	win
regex-redux	regexredux-2	regexredux_6	0.691 s	1.106 s	0.96 s	1.24 s	149.1 MB	150.5 MB	win
k-nucleotide	knucleotide-7 †	knucleotide_6	1.358 s	2.106 s	2.66 s	3.82 s	68.7 MB	138.6 MB	win
mandelbrot	mandelbrot-6	mandelbrot_3 †	1.339 s	1.996 s	2.53 s	3.85 s	7.3 MB	12.1 MB	win
pidigits	pidigits	pidigits_3	0.626 s	0.839 s	0.62 s	0.84 s	2.6 MB	3.0 MB	win
reverse-complement	revcomp-10	revcomp	0.836 s	1.105 s	0.24 s	0.25 s	10.9 MB	123.4 MB	win
fannkuch-redux	fannkuchredux-5	fannkuchredux_6	6.621 s	7.864 s	13.08 s	15.63 s	2.3 MB	2.2 MB	win
binary-trees	binarytrees †	binarytrees_2	1.631 s	1.897 s	3.07 s	3.57 s	70.1 MB	130.2 MB	win
n-body	nbody	nbody_6	2.453 s	2.703 s	2.45 s	2.70 s	2.0 MB	2.0 MB	win
spectral-norm	spectralnorm	spectralnorm_7	0.460 s	0.462 s	0.89 s	0.89 s	2.2 MB	2.4 MB	tie

9 wins, 1 tie, no losses. xisto's peak memory is at or below Rust's on every problem but fannkuch-redux, 2.3 MB against 2.2 MB in this round (2.3 against 2.3–2.4 MB in the other rounds of the day). † Chosen for memory (up to 2% slower with at least 10% less memory, or up to 5% slower with at least 25% less). nbody_9 and nbody_7 are excluded because they compute 1/√x with the approximate rsqrtps instruction and give different doubles.

Graviton2 (AWS t4g.small, 2 vCPUs, arm64)

Source: bench/results/vm-2026-11-13/rust-arm-graviton2-2026-11-13-4a2f388.txt.

problem	xisto	Rust	xisto time	Rust time	xisto CPU	Rust CPU	xisto mem	Rust mem	result
fasta	fasta	fasta_2	0.546 s	1.839 s	0.38 s	3.06 s	1.8 MB	2.5 MB	win
reverse-complement	revcomp-10	revcomp	0.282 s	0.564 s	0.28 s	0.36 s	10.1 MB	123.2 MB	win
fannkuch-redux	fannkuchredux-5	fannkuchredux_4	8.586 s	14.326 s	17.14 s	27.91 s	1.8 MB	1.9 MB	win
regex-redux	regexredux-2	regexredux_6	0.798 s	1.291 s	1.25 s	1.58 s	148.6 MB	149.7 MB	win
k-nucleotide	knucleotide-7	knucleotide_6	2.986 s	4.319 s	5.89 s	7.98 s	64.6 MB	138.3 MB	win
binary-trees	binarytrees-3	binarytrees_2	1.826 s	2.190 s	3.57 s	4.20 s	113.8 MB	130.0 MB	win
pidigits	pidigits	pidigits_3	2.042 s	2.217 s	2.04 s	2.21 s	2.1 MB	2.8 MB	win
n-body	nbody	nbody_2	4.602 s	4.985 s	4.60 s	4.98 s	1.5 MB	1.6 MB	win
mandelbrot	mandelbrot-8	mandelbrot_4	2.627 s	2.792 s	5.21 s	5.49 s	6.9 MB	32.4 MB	win
spectral-norm	spectralnorm-7	spectralnorm_5	1.220 s	1.230 s	2.43 s	2.45 s	1.8 MB	2.1 MB	tie

9 wins, 1 tie, no losses; xisto's peak memory is below Rust's on every problem. Reverse-complement is revcomp-10, the fastest xisto variant here with 10 MB: it reads each sequence backwards one block at a time and moves the newlines to their output places before the reverse, so each line is one run.

The ties

spectral-norm, both machines: the time is the 1.21 billion double divisions 1/d the problem asks for. A C loop doing only those divisions takes the same CPU as both programs (0.88 s on the Xeon, 2.42 s on the Graviton2), so no program can be 5% ahead without caching the matrix (~240 MB) or computing 1/d another way, which the problem does not allow. xisto is ahead by 1–2% on both.

mandelbrot on the Graviton2, a tie until e897e76, is now a win (6% ahead). Two changes:

• mandelbrot-8 runs twelve pixels at a time after a byte with a pixel inside the set, and writes 2·zr·zi as p + p with p = zr·zi. Doubling is exact, so the doubles are the same, and the chain is an addition shorter than a multiplication.
• The compiler's lane planner (src/emit.c) picks the split that puts the most work in vectors: twelve copies become three vectors of four instead of one of eight and four scalars. The C of every other program is unchanged.

The xisto programs

Each name links to its source on GitHub; the original Node programs are in bench/benchmarksgame/node/.

problem	xisto variant	what it does
spectral-norm	spectralnorm (Xeon), spectralnorm-6, spectralnorm-7	the products split in equal shares of rows across threads; the denominator of A carried along each row by additions; #7 joins the 40 products through Atomics counters instead of messages. Same divisions, same order, same doubles as the plain program
n-body	nbody	the plain program with types
fannkuch-redux	fannkuchredux-5	720 ranges to a pool of workers; the permutation as bytes in vector registers (a flip is one shuffle); two permutations' flip chains counted in the same loop
binary-trees	binarytrees, binarytrees-2, binarytrees-3	trees built in parallel (xisto:parallel or a worker pool); nodes of 8 bytes (32-bit references), a statement's temporary trees freed right after it
mandelbrot	mandelbrot-6 (Xeon), mandelbrot-8 (Graviton2)	chunks of rows to a pool of workers; 16 (#6) or 8 and 12 (#8) pixels per call as SIMD vectors; rows written out as they finish
fasta	fasta	the plain program with types; the random sequences from xisto:random
k-nucleotide	knucleotide-7	the sequence at two bits per nucleotide, packed in parallel; small frames counted in arrays indexed by key, the big ones in partitions
reverse-complement	revcomp-10	each sequence read backwards from the file one block at a time, blocks made by workers into a shared ring and written in order (10 MB); the newlines moved to their output places before the reverse, so each line is one run
regex-redux	regexredux-2	the nine counts in a worker while the main thread does the replacements; the 50 MB string shared with the worker, not copied
pidigits	pidigits	the plain program with types (BigInt on xisto's GMP-backed integers)

What changed since 10 Nov

The 10 Nov round had 7 wins, 3 ties on the Xeon and 8 wins, 1 tie, 1 loss (mandelbrot) on the Graviton2, with xisto using more memory than Rust on binary-trees, fasta, k-nucleotide and reverse-complement. Between 10 and 13 Nov:

• Compiler and runtime, for every program: 32-bit references for classes that only point to themselves (the binary-trees node goes from 16 to 8 bytes); objects that cannot outlive one statement go to a region given back to the system after it; the order of SIMD lanes chosen per architecture; a 1 MB stdout pipe on Linux; large strings sent to a worker without a copy; Int32Array and Atomics (wait, notify) as in Node.
• New variants: mandelbrot-6/-7 (lane copies, output written as rows finish), fannkuchredux-5 (two permutations' flips in one loop), knucleotide-7 (2 bits per base), revcomp-9 (each sequence read back to front, one block at a time), regexredux-2 (counts in a worker), spectralnorm-7 (an Atomics barrier), mandelbrot-8 (twelve pixels per run, with the lane planner above), revcomp-10 (one run per line).

Result files

• bench/results/vm-2026-11-13/rust-ubuntu-xeon-2026-11-13-4a2f388.txt
• bench/results/vm-2026-11-13/rust-arm-graviton2-2026-11-13-4a2f388.txt
• Earlier rounds of the same day (*-937cb49, *-4a10af3, *-5296927, *-e5fb066, *-efd0fef, *-e897e76) show each step of the work above.