Improve codegen for single argument format_args!

[pickfire]

I was using write! in a lot of places, some I just do write!(writer, " ") but I noticed that the generated assembly is not as good as writer.write_all(b" "). Since we have minimal const fn now, I wonder if we could improve the generated assembly when a single character or a single argument is passed in.

use std::io::{self, Write};

pub fn write_char(buf: &mut Vec<u8>) -> io::Result<()> {
    write!(buf, " ")
}

use std::io::{self, Write};

pub fn write_char(buf: &mut Vec<u8>) -> io::Result<()> {
    buf.write_all(b" ")
}

https://rust.godbolt.org/z/zzEMjd

I also write!() could accept a single byte or character. Example, write!(buf, ' ').

If anyone can mentor, maybe I can take on this part (hopefully without compiling rust).

Meta

rustc --version --verbose:

nigthly

CC @lzutao

[tesuji]

I opened one in #75358

[RalfJung]

This is entirely unrelated to const fn. const fn doesn't change assembly, it just changes which code you can write

const FOO: i32 = { HERE };

However, I agree a string-only format_args! could likely be improved.

Another case that could possibly be improved is

rust/library/core/src/panicking.rs

Line 50 in c94ed5c

panic_fmt(fmt::Arguments::new_v1(&[expr], &[]));

[workingjubilee]

Seems related to #10761

I think the best thing is to try to nudge the write! macro itself to try to handle this, because waiting until it reaches write_fmt seems to damage the codegen, as write_fmt then has more overhead to examine. This should be handle-able at the macro/const eval time, in effect.

[CAD97]

Did fmt::Arguments::as_str improve the situation here? If it's not being used by write! impls yet, it probably can/should be.