int: implement signed comparison for unsigned integers

[xobs]

Implement signed comparison for unsigned integers, which is very similar to unsigned integers but uses to_signed() rather than to_unsigned(). This fixes a case where niche optimization causes the discriminant to sometimes compare signed and unsigned values.

This was caught when running a test suite for an embedded device compiled under the gcc backend. A reproducer when compiling under -O for x86-64 is:

use std::hint::black_box;

#[repr(u8)]
enum Inner { L0 = 0, H255 = 255 }
enum O { A(Inner), B, C }

#[inline(never)]
fn which(o: &O) -> &'static str {
    match o { O::A(_) => "a", O::B => "b", O::C => "c" }
}

fn main() {
    assert_eq!(which(black_box(&O::A(Inner::H255))), "a"); // 0xFF
    println!("all ok");
}

The problem only shows up with -O, not by default.

thread 'main' (1223) panicked at /home/seancross/work/testcase/testcase-3.rs:13:5:
assertion `left == right` failed
  left: "c"
 right: "a"

[xobs]

Let me know if you'd like me to open an issue in the rust repo, or if there's a place I can add a testcase here. I'm manually driving rustc due to #930.

I have a few other minor fixes, and I can send PRs with them once I know how you'd prefer them formatted.

[antoyo]

Let me know if you'd like me to open an issue in the rust repo, or if there's a place I can add a testcase here. I'm manually driving rustc due to #930.

No need to open an issue: you can open a PR right away.
You can add a test in this directory.

I have a few other minor fixes, and I can send PRs with them once I know how you'd prefer them formatted.

Anything else you need to know about my preferences here?

Thanks for your contribution!

[xobs]

That should be it. Now I know how to add and run tests, I can use that to make sure the changes I push actually fix things.

There are a few bigger changes I've accumulated (LTO improvements, link section fixes, clobber support), and a couple of them require patches to gcc (and consequently gccjit support), but I'll start with these small patches first.

[xobs]

Hmm... The test doesn't seem to work without int128. Or on m68k. Think that is a real bug, or should I gate the test on x86 only?

[antoyo]

Hmm... The test doesn't seem to work without int128. Or on m68k. Think that is a real bug, or should I gate the test on x86 only?

It seems like a real bug to me.
Without int128, the error is:

gcc_jit_context_new_cast: cannot cast param0[(int)1] from type: __int64_t  __attribute__((aligned(8))) to type: __int64_t  __attribute__((aligned(8)))[2]

which means it seems you try to cast a low/high value to the complete 128-bit representation (which is an array of i64).
I guess the code needs to be adjusted to call to_signed on the inner element type instead of the array type.

[xobs]

Yep, sorry, I just needed to figure out how to use gcc-15-without-int128.deb. Now I see I just need to extract it somewhere and point config.toml:gcc-path at it.

Hopefully works now. So happy CI and test coverage exists.

[antoyo]

One last thing, then we can merge this:

View changes since this review

[antoyo]

Thanks for your contribution!