Thread by @japaric_io

The fastest formatter does no formatting at all!

I have created a #rustlang formatter for debugging #embedded apps (logs!) that's ~10 (? hard to quantify) times faster than the standard library's `fmt` module and uses way less bytes of `.text` / `.rodata` (and "bandwidth")

1/

Here I "format" a timestamp and an `f32` value on a 64 MHz ARM CPU w/ FPU disabled in 6us. The formatted string is printed on the host (right)

`core::fmt`-ing the same thing (w/o a xfer) takes 57us. I must say `core`'s float formatting is v good but (still) uses 1KB of stack

2/

My formatter binary encodes the format string + arguments and sends them to the host; the host does the actual formatting

Unlike bin serialization crates there's no need to compile any crate for the host to make formatting work (IOW there's no `Deserialize` trait here)

3/

The reason for that is that all info required to decode & format the received args is stored in ... ELF metadata

Even string literals are stored in ELF metadata! The target sends indices (1-2 bytes) rather than full strings; this reduces `.rodata` usage and xfer bandwidth

4/

L: My formatter provides its own derive(Debug) that formats the sames as `core::fmt`'s

R: But you can also do cool compression tricks like having the target send the contents of a register (4 bytes) and format that on the host side as many bitfields!

5/

All this is just a proof of concept atm and not available for testing though (boo).

P.S. The 6us I quoted in part 2 could, in theory, be brought down to 2us if the data xfer took negligible amount of work but I haven't had much luck reaching the theoretical limit.

6/6