Benchmarks

Milo compiles via LLVM with -O2 — the same backend as C and Rust. On most workloads it lands within noise of C.

Results

MiloCGo

matmul 512×512

Milo

12ms

C

12.8ms

Go

13.2ms

binarytrees depth 18

Milo

3ms

C

3.9ms

Go

10.5ms

quicksort 2M f64

Milo

34.7ms

C

35.7ms

Go

34.7ms

startup empty main

Milo

1.2ms

C

1.2ms

Go

1.5ms

stringops 100k concat

Milo

3.2ms

C

3.1ms

Go

6.5ms

fib(42)

Milo

20.8ms

C

18.4ms

Go

21.6ms

sieve to 10M

Milo

2.5ms

C

2.1ms

Go

3.4ms

maplookup 100k

Milo

4.4ms

C

3.3ms

Go

5ms

grep -c 5MB

Milo

5.5ms

C

2.1ms

Go

4ms

json parse+walk 1MB

Milo

7.1ms

C

1.6ms*

Go

9.7ms

Apple M-series, macOS. *C uses yyjson; Go and Milo use their stdlibs.

Notes

• Values < 1.0x mean Milo is faster than C (usually within measurement noise)
• 4 benchmarks match or beat C; 3 more within 20%
• Slower entries (grep, maplookup) have known hot spots — not fundamental limits
• Go's GC overhead shows clearly in allocation-heavy benchmarks (binarytrees 3.5x slower)
• JSON: flat pool parser with source-offset strings — ~6 allocations for any document size. Beats Go; remaining gap vs C is stdlib-vs-yyjson

Reproduce

./benchmarks/run.sh