implement floor and ceil in assembly on i586

[folkertdev]

fixes #837

The assembly is based on

• https://github.com/NetBSD/src/blob/20433927938987dd64c8f6aa46904b7aca3fa39e/lib/libm/arch/i387/s_floor.S
• https://github.com/NetBSD/src/blob/20433927938987dd64c8f6aa46904b7aca3fa39e/lib/libm/arch/i387/s_ceil.S

Which both state

/*
 * Written by J.T. Conklin <jtc@NetBSD.org>.
 * Public domain.
 */

Which I believe means we're good in terms of licensing.

[tgross35]

This is awesome, thank you for implementing it!

There is one problem, our libm MSRV unfortunately doesn't support naked functions. I'll be bumping it in the near future but I don't think it will be that high; how hard is this to do with inline asm? Probably adds a few instructions because the loads stack access becomes opaque, but that would still be faster than the slow existing implementation.

[tgross35]

Fix for the red CI in #979 btw

[folkertdev]

how hard is this to do with inline asm?

Well, I don't know this target and its assembly that well. So using global_asm! would be much easier (for me, anyway) and more straightforward to upgrade once the MSRV does support naked functions.

I think the only downside is that the symbols might be visible from the outside? Is that a problem?

[folkertdev]

actually we can prevent even that with a trick that Björn showed me recently using sym

[folkertdev]

from what I can find this is a 2024 edition thing

error: extern blocks must be unsafe
  --> builtins-shim/../compiler-builtins/src/math/../../../libm/src/math/arch/i586.rs:23:1
   |
23 | / extern "cdecl" {
24 | |     fn ceil_helper(_: f64) -> f64;
25 | |     fn floor_helper(_: f64) -> f64;
26 | | }
   | |_^

but rust 1.63 won't compile unsafe extern right?

[tgross35]

I think the only downside is that the symbols might be visible from the outside? Is that a problem?

I think it may be, also conflicts (though we could make it weak). But this should be pretty easy to move to inline asm; I think something like this would work https://rust.godbolt.org/z/P98rKEW7P, which is reasonably close to the original (not sure if there's a way to turn pointers to locals into stack offsets, to avoid the leal s)

[folkertdev]

not sure if there's a way to turn pointers to locals into stack offsets, to avoid the leal s

That's probably fine, we can fix that properly once naked functions are available

---

we now get errors like this?

  stderr ───
    Spaced Musl sinf arg 1/1: 800 iterations (800 total)

    thread 'musl_quickspace_sinf' panicked at libm-test/tests/compare_built_musl.rs:26:50:
    called `Result::unwrap()` on an `Err` value: 
        input:    (-239880100.0,)
        as hex:   (-0x1.c988f4p+27,)
        as bits:  (0xcd64c47a,)
        expected: -0.17352822            -0x1.6362c4p-3 0xbe31b162
        actual:   -0.17353132            -0x1.636464p-3 0xbe31b232

[quaternic]

You're clobbering the saved control word on the stack with the modified one, so it isn't restored correctly.

I think there's an argument for using a fixed control word for the frndint, but I'll have to come back to this later.

[tgross35]

Sorry, yeah, as @quaternic mentioned this needs a second stack slot. As a microopt this could be let mut cw_stash = [0u16; 2]; and then these accesses can be -4({stash_ptr}) so we save a register vs. two different u16 locals

[tgross35]

Actually even better, let mut cw_stash = MaybeUninit::<[u16; 2]>::uninit(); to save the zero init instruction

[folkertdev]

I got this to work

https://rust.godbolt.org/z/44Tjcx3bb

the -4 stuff didn't work (also, for u16, why -4?)