Performance: the compiler and the C it emits

This page is about src/tychoc.c, runtime/tycho_rt.c and the C they emit — how the arena model is kept cheap, and what each mechanism does. For the language's performance story, read performance.md and the thesis.

It carries one reproducible number, the transpile time below. Everything else describes a mechanism rather than timing one. For figures you can reproduce, make bench runs 17 benchmarks and each bench/*/RESULTS.md records its own hardware.

Transpile time

bench/transpile/run.sh measures how long tychoc takes to turn one .ty file into C. It generates a syntax-stable input, verifies the compile succeeds before timing it (a compile that dies is faster than one that works, so an unchecked input reports the failure path as a speed-up), discards a warm-up run, and reports the minimum of N with median and max beside it.

11 ms minimum of 10 timed runs over a 4.2k-line input — flat, no regression across the measured window.

What to run

• make bench runs the 17-benchmark suite. Each bench/*/RESULTS.md records its own hardware and per-workload numbers.
• bench/transpile/run.sh measures transpile time as described above.
• TYCHO_ARENA_STATS=full <program> prints a one-shot residency summary at exit (peak live bytes, bump-allocated bytes, free-list recycle rate, OS-reserved bytes, block-pool reuse, per-statement residency). See debugging.md and performance.md.

What the bounds mean

• Transpile time is bounded by the figure above; make bench fails on regression.
• Memory is bounded per workload by the peak-RSS tables in performance.md: the in-place append (acc = acc + x growing one buffer) holds ~1.5 MB where the unoptimized path needs ~825 MB at the same N, so the make bench guard sits between working and broken.
• Cross-language ratios, not absolute times, are the claim throughout: absolute wall times move with machine, governor, and toolchain. See performance.md.

How the cost is kept down (summary)

The arena model wins by orders of magnitude where it was designed to (the acc = acc + s loop, rewritten to an in-place O(n) append), and pays a per-scope tax everywhere else: every block scope and call does arena_child / arena_reset / arena_free. Four mechanisms keep that tax off the hot path:

• Block pool (runtime/tycho_rt.c): freed blocks go to a process-global free-list instead of free; allocation takes from it first. Per-scope block churn becomes O(1) pointer traffic with no page re-faulting; peak live memory is unchanged.
• Block-retaining reset: a loop's scratch arena keeps its head block and rewinds it, so a one-block-per-iteration loop does zero pool traffic.
• Elided child arenas: if/match blocks use the enclosing scope's arena; escaping values promote to _parent either way. Smaller emitted C, same wall-clock and RSS.
• Borrowed read-only params: a heap-bearing struct parameter the body never mutates is borrowed, not deep-copied — removing the dominant copy cost on compiler-shaped workloads (the symbol-table clone).
• Push-loop fusion: a loop that only pushes to a local array keeps the descriptor in C locals across the loop, closing most of the gap to hand-written C on push-dominated loops.

The tuning diary behind these — profiler notes, blind alleys, and the Ctx-reconstruction and O(n²) fixes — is in internals/runtime-tuning.md.