M4 max shows unexpected results for nightly format_into() vs itoa crate

[sammysheep]

@GuillaumeGomez This is feedback for rust-lang/rust#142098.

On an Apple M4 Max with rustc 1.90.0-nightly (855e0fe46 2025-07-11) the results look like this:

test bench::all_digits::itoa ... bench:          51.07 ns/iter (+/- 1.89)
test bench::all_digits::std1 ... bench:         436.27 ns/iter (+/- 30.34)
test bench::all_digits::std2 ... bench:         582.99 ns/iter (+/- 49.70)
test bench::all_digits::std3 ... bench:          64.49 ns/iter (+/- 3.20)
test bench::first_999::itoa  ... bench:         500.48 ns/iter (+/- 25.55)
test bench::first_999::std1  ... bench:      15,732.48 ns/iter (+/- 1,030.35)
test bench::first_999::std2  ... bench:      22,634.60 ns/iter (+/- 1,111.68)
test bench::first_999::std3  ... bench:       1,346.81 ns/iter (+/- 19.22)
test bench::first_9::itoa    ... bench:           2.23 ns/iter (+/- 0.02)
test bench::first_9::std1    ... bench:         149.36 ns/iter (+/- 5.87)
test bench::first_9::std2    ... bench:         221.03 ns/iter (+/- 8.88)
test bench::first_9::std3    ... bench:          11.39 ns/iter (+/- 0.32)

• std1 is to_string()
• std2 is format!()
• std3 is NumBuffer (nightly feature)
• itoa is David Tolnay's crate

When I bench things on real world code I don't see quite such a difference between itoa and the nightly feature (std3), but I wanted to report it in case it is not an artifact of the test.

Thanks so much for doing this work!

[GuillaumeGomez]

Much appreciated!

To be noted: to_string should be almost as fast as std3 normally now.

[sammysheep]

Astute observation! It seems std1 and std2 methods mostly benchmark the system allocator. Having added a bench on simply boxing the number:

test bench::all_digits::boxit ... bench:         334.97 ns/iter (+/- 10.73)
test bench::all_digits::itoa  ... bench:          51.38 ns/iter (+/- 4.72)
test bench::all_digits::std1  ... bench:         419.51 ns/iter (+/- 20.29)
test bench::all_digits::std2  ... bench:         579.47 ns/iter (+/- 26.05)
test bench::all_digits::std3  ... bench:          69.86 ns/iter (+/- 6.14)
test bench::first_999::boxit  ... bench:      12,167.12 ns/iter (+/- 453.17)
test bench::first_999::itoa   ... bench:         483.45 ns/iter (+/- 24.42)
test bench::first_999::std1   ... bench:      15,280.83 ns/iter (+/- 1,284.51)
test bench::first_999::std2   ... bench:      21,753.19 ns/iter (+/- 685.77)
test bench::first_999::std3   ... bench:       1,325.85 ns/iter (+/- 12.67)
test bench::first_9::boxit    ... bench:         119.58 ns/iter (+/- 3.60)
test bench::first_9::itoa     ... bench:           2.22 ns/iter (+/- 0.04)
test bench::first_9::std1     ... bench:         144.31 ns/iter (+/- 6.10)
test bench::first_9::std2     ... bench:         211.74 ns/iter (+/- 10.85)
test bench::first_9::std3     ... bench:          11.35 ns/iter (+/- 0.09)

Then here is a table if I subtract out the boxit benchmark for std1/std2:

Test	Method	Time (ns/iter)	(+/-)
all_digits	itoa	51.38	4.72
all_digits	std1	84.54*	*
all_digits	std2	244.5*	*
all_digits	std3	69.86	6.14
first_999	itoa	483.45	24.42
first_999	std1	3,113.71*	*
first_999	std2	9,586.07*	*
first_999	std3	1,325.85	12.67
first_9	itoa	2.22	0.04
first_9	std1	24.73*	*
first_9	std2	92.16*	*
first_9	std3	11.35	0.09

• Subtracts boxit

So yes, to_string() has much better actual number conversion than format!() and closer to format_into().

[GuillaumeGomez]

Thanks for double-checking and for the numbers!