r? @Mark-Simulacrum

(rust-highfive has picked a reviewer for you, use r? to override)

-1 isn’t a u8 is it? This failing test is a bit of a curveball.

I guess it's some shenanigans with type inference. Before it was being inferred as i32, but now that it's being inferred as u8. Consider this example (playground), without the specialized impl of Foo for u8 it compiles fine, with it it throws an error.

Type inference breaking changes are in principle allowed by the stability rules, but you could try routing the default implementation (impl<T: fmt::Display + ?Sized> ToString for T {)) through a private specialization trait instead. Or try implementing it for all integer types.

Ok, well if the easiest thing is to make the rest faster I'm gaim with that. Double or quits. That would completely close the associated ticket.

If you prefer splitting the PR then you could first specialize for i32 since that's picked over the other integer types anyway.

r? @KodrAus since it sounds like there's inference breakage here so likely needs T-libs decision

The new assembly looks pretty unfortunate to me FWIW, even though it does seem better than what was there before.

I don't want to break anything. Will update this PR with a more comprehensive solution.

"The new assembly looks pretty unfortunate to me" - really interested in this. I'm just getting into understanding the asm side of things. I've found a construction that allows the String with_capacity to be set to the exact length without the vec growth asm being included but it does add an extra 20+ instructions. I guess memory versus instructions is a tradeoff. At the moment b'a'.to_string().capacity() returns 8 so whether it's 3 or the exact size it's a lot smaller in memory?

Oh, I was mainly thinking that there's several multiplications in the generated code - it feels like a u8 should be convertible to decimal without going through that overhead, but I'm not really familiar with the state of the art here.

I tried lots of things and couldn't get it below 65ns (The original is around 85ns for me). Then I removed everything and just returned String::with_capacity(3) - took 65ns. Alas there's no way that I can avoid that allocation so I guess we should optimise for min code size and trying to match the capacity closely.
(edit) I switched to jemalloc and it's now about 22ns. That makes it a bit easier to tell if changes are making a perf difference.

Sometimes it looks like the compiler can figure out when there's no possibility of growth in a vec and avoid generating the instructions but quite often it can't tell: both these approches trigger it including the grow: https://godbolt.org/z/zxG7hK
Is that just something that's fine becaue it gets amortised over a big program or will it dump out those instructions every time it's inlined?

This is probably the best example of catching it in the act:

https://godbolt.org/z/ePvYrs

Digging into this MIR seems the same as well as LLIR so alas this seems to be in llvm that it is deciding to do something very different. (Edit: Am now wondering if the above differences are just godbolt having an off day or if it's really that different.)

Well godbolt ( https://rust.godbolt.org/z/Keb5ao ) at least seems happy with the i8 impl. It's about as good as I can get it and it's pretty fast - the compulsary alloc consumes the vast amount of the time, but with jemalloc it's over twice as fast as the old way and with the system allocator it seems to shave off 25%.

If anyone has any other sugestions of things I've not tried yet please let me know, but if not I think its ready for another review.

Using a vec for at most 4 bytes (since it's all ascii) seems quite complicated. Have you tried operating on a local array instead and only allocating the output string once the length is known?

Looks like the String-based version optimizes fine when MIR optimizations are disabled, which suggests that it's the same problem as #82801.
(Dropped the shrink_to_fit, since it generates quite a bit of assembly and wasn't a fair comparison.)

Just to make sure there's no misunderstanding: It's not my approval you need to land this. I'm just a random dude without any power in that regard.
Maybe update the PR's title now that it covers i8 in addition to u8. Can't comment on the stability attributes. The code looks fine to me.

That said, it would be cool if we could do this and have the compiler optimize it well:

impl ToString for i8 {
    #[inline]
    fn to_string(&self) -> String {
        let mut buf = String::with_capacity(4);
        if self.is_negative() {
            buf.push('-');
        }
        buf.push_str(&self.unsigned_abs().to_string());
        buf
    }
}

Alas, it can't do that (yet?).

@LingMan thank you for all your help on this PR, it's been really appreciated! Frankly I think to_string() is technically a little broken. Ideally something like write_to_string(&mut String) would be better. That way you could do something like you suggest above nicely but as it is that would be two allocations.

@Mark-Simulacrum this one's ready for review at your leasure.

r? @Amanieu

Knowing what I have recently learned I can probably shave a bit off these timings by pulling the adds a little earlier in the function. Give me an evening to revisit this.

No maybe not. I think that's as good as it gets.

That's great! Do you think you could implement this for other integer sizes too?

Idk, the way the larger integer types are currently handled makes sense - the two at a time digits makes sense - it's just at the short end that the general implemention seems overkill. Mostly I'm depressed that anything done here is dwarfed by the alloc that we can't ever avoid due to the signature. Rather I'm focusing on the other end - parsing ints which isn't bounded by allocation times.

@bors r+

📌 Commit 05330aa has been approved by Amanieu

⌛ Testing commit 05330aa with merge 8a8ed07...

☀️ Test successful - checks-actions
Approved by: Amanieu
Pushing 8a8ed07 to master...