Microbenchmarks comparing:

• Register-CC version of Astil Forth in AOT mode.
• Register-CC version of Astil Forth in JIT mode.
• Stack-CC version of Astil Forth in JIT mode.
• Gforth.
• C (Clang).
• JS (Bun).
• Lua (LuaJIT).
• Common Lisp (SBCL).
• Erlang/BEAM.
• Python (PyPy and CPython).
• Occasionally some other langs.

Summary: in these very limited microbenchmarks, the reg-CC implementation of Astil Forth trounces VM interpreters, approximates other JITs, and vaguely approaches Clang C with -O2 (x1.5-x2).

Naming:

• _aot -- Astil reg-CC as AOT-compiled executable.
• _jit -- Astil reg-CC in JIT mode.
• _stack -- Astil stack-CC in JIT mode.

Each _jit and _stack benchmark includes the cost of bootstrapping the entire language first. Reg-CC takes longer to bootstrap because it has more features.

In Erlang: baseline_erlang_single, CPU-oriented benchmarks, and "cat" use one BEAM scheduler (+S 1:1). TCP benchmark uses default scheduler count.

Notes:

• All measurements were done on M3 Pro.
• CPU microbenchmarks are sensitive to code layout and instruction selection. Small source changes can shift CPU frontend, cache, and branch-prediction behavior; on M3 Pro, we have seen cosmetic-looking changes move results by up to ≈10% or up to 10ms depending on benchmark runtime. Avoid over-generalizing differences in that range.
• The suite records wall time, total CPU time, and peak memory usage. GC-based engines can spend substantial CPU on background threads.
• In many benchmarks, startup time skews the measurement. Adjust them by the "baseline" metrics when comparing.

clang 22.1.4

gforth 0.7.3

luajit 2.1.1753364724
luv 1.52.1

bun 1.3.10

sbcl 2.6.4

zig 0.16.0

go version go1.25.1 darwin/arm64

openjdk version "25" 2025-09-16

Erlang/OTP 28 [erts-16.4] [source] [64-bit] [smp:12:12] [ds:12:12:10] [async-threads:1] [jit] [dtrace]

[pypy 7.3.17 with gcc apple llvm 16.0.0 (clang-1600.0.26.4)]
python 3.10.14 (39dc8d3c85a7, nov 09 2024, 22:49:03)

python 3.14.4

Erlang/OTP 28 [erts-16.4] [source] [64-bit] [smp:12:12] [ds:12:12:10] [async-threads:1] [jit] [dtrace]

C and Astil use uint128. Other languages use actual bigints.

Copies a warmed 512 MiB file and 512 MiB stdin (1 GiB total).

Measures 4096 concurrent connections with 32 one-byte request/echo exchanges per connection.

Wall time includes Python TCP driver work. After every connection closes, the driver kills and reaps the idle server; this work is also included in wall time. CPU time and peak mem/RSS measure only the server subprocess.

Erlang uses +sbwt none; it dramatically reduces active-N server CPU without changing the default scheduler count.

This benchmark is noisier than others, especially when using pthreads. Results vary more between reruns.

Results are sorted by total user+kernel CPU time.

Each implementation is measured 5 times.