try, compiled: Ruby CLI to a 705KB native binary

[tobi]

Ruby try compiled all the way to a 705 KB native binary with Spinel (Matz's AOT Ruby→C compiler). Same TUI, same 399 specs as MRI, startup in under a millisecond instead of ~70 ms.

--help is 73× faster. A first TUI frame is ~50–58×. The 1.0 MiB unstripped build shrinks to 705 KB with -Os + strip.

No patches to Spinel or CRuby. ruby try.rb and the gem keep working. make native emits dist/try.c and links dist/try.

Watch it

Same session on MRI and on the native binary: type spin, cursor into the word, insert x → sxpin, Ctrl-A/E, Backspace, Ctrl-U, filter tui, rename, Esc.

MRI (ruby try.rb)

Native (dist/try)

Playable: MRI asciinema · native asciinema

Numbers

hyperfine 1.19.0, --shell=none -w 3 -r 20, 2026-08-13 16:04 EDT. Ruby 3.3.8 vs Spinel 79b1aba8bdef (PR 3906 so IO#tty? / #winsize work). Native won every case.

Case	MRI	native	Speedup
--help	69.4 ± 4.8 ms	0.956 ± 0.096 ms	72.6×
--version	70.8 ± 6.3 ms	0.959 ± 0.125 ms	73.8×
no args	68.7 ± 4.3 ms	0.982 ± 0.068 ms	69.9×
init	96.5 ± 9.4 ms	21.2 ± 1.3 ms	4.5×
TUI empty	74.0 ± 8.4 ms	1.28 ± 0.14 ms	57.8×
TUI 5 dirs	72.0 ± 6.0 ms	1.38 ± 0.08 ms	52.1×
TUI --and-keys=beta	72.4 ± 5.9 ms	2.70 ± 0.35 ms	26.8×
TUI exec beta	74.6 ± 8.1 ms	1.49 ± 0.16 ms	50.2×
clone script	70.3 ± 6.2 ms	0.959 ± 0.103 ms	73.3×
worktree script	72.9 ± 8.6 ms	1.01 ± 0.09 ms	71.9×
scan 1000 dirs	80.3 ± 7.3 ms	8.57 ± 0.69 ms	9.4×
filter 1000 dirs	93.9 ± 5.5 ms	40.9 ± 4.8 ms	2.3×

init is the outlier because native still forks ps to detect fish vs bash (~20 ms). Filtering 1000 dirs is whole-process (scan + four keystrokes + re-renders), not a microbench of the matcher. MRI never wins: interpreter boot is the MRI cost.

make native SPINEL=/path/to/spinel
./dist/try --help
eval "$(./dist/try init)"
make native-test

399/399 specs on both ruby try.rb and dist/try. Stack: this PR, then #131 (strip / -Os), then #133 (lint + MRI vs native compare).

Why the Ruby had to change

Spinel compiles a closed world. try is a small stdlib CLI, but it used a few things AOT does not have yet (FileUtils, IO#raw, method_missing, Integer#times, …). Those became MRI-compatible shims in TryCompat so the same source still runs on Ruby.

Full gap list is in the first version of this description and in the diff. Remaining: no live SIGWINCH on native, raw mode is stty, non-tty winsize is [0,0].

[tobi]

Benchmark (hyperfine, MRI vs native)

Moved here from the TUI PR so this stack stays Spinel-only. Measured 2026-08-13 16:04 EDT. hyperfine 1.19.0, --shell=none -w 3 -r 20. Ruby 3.3.8, Spinel 79b1aba8bdef (PR 3906). Native at measurement was ~1.0 MiB unstripped; after -Os+strip (#131) dist/try is 705 KB. Native won every case.

Case	MRI	native	Speedup
--help	69.4 ± 4.8 ms	0.956 ± 0.096 ms	72.6×
--version	70.8 ± 6.3 ms	0.959 ± 0.125 ms	73.8×
no args (exit 2)	68.7 ± 4.3 ms	0.982 ± 0.068 ms	69.9×
init	96.5 ± 9.4 ms	21.2 ± 1.3 ms	4.5×
TUI empty	74.0 ± 8.4 ms	1.28 ± 0.14 ms	57.8×
TUI 5 dirs	72.0 ± 6.0 ms	1.38 ± 0.08 ms	52.1×
TUI --and-keys=beta	72.4 ± 5.9 ms	2.70 ± 0.35 ms	26.8×
TUI exec beta	74.6 ± 8.1 ms	1.49 ± 0.16 ms	50.2×
clone script	70.3 ± 6.2 ms	0.959 ± 0.103 ms	73.3×
worktree script	72.9 ± 8.6 ms	1.01 ± 0.09 ms	71.9×
scan 1000 dirs	80.3 ± 7.3 ms	8.57 ± 0.69 ms	9.4×
filter 1000 dirs	93.9 ± 5.5 ms	40.9 ± 4.8 ms	2.3×

init still forks ps for fish vs bash (~20 ms on native).

TUI demos (MRI + native) are still on branch tui-input-field:

• MRI gif
• native gif