Implement {float}::asin_acos(), {float}::asin(), and {float}::acos()

[fu5ha]

Fixes #44

[fu5ha]

Due to the tests for this PR hitting an obscure bug relating to f32::from_bits() quieting a "signalling nan" causing an off-by-one-bit (that could have happened at any other time elsewhere technically i guess), it depends on #43 to be in first in order to pass tests on i586 hardware (non-simd codepath)

[MaulingMonkey]

Huzzah for test coverage!

Cross referencing rust-lang/rust#73328 for upstream information about x not necessairly round-tripping through f32::from_bits(x).to_bits(). I'm not entirely sure if it's considered a codegen bug, or if f32::from_bits just needs better documentation, but considering how deep into the bowels of LLVM behavior you'd need to go to fix this, I wouldn't hold my breath about Rust's stdlib behavior getting changed anytime soon.

A concrete example for posterity:

fn main() {
    let bits : u32 = 0x7f8d9fa6; // some negative signaling (x86/ARM) NaN
    let f = f32::from_bits(bits);
    assert_eq!(bits, f.to_bits()); // will fail, now equal to 0x7fcd9fa6
}

And some annotated disassembly of f32::from_bits:

    #[stable(feature = "float_bits_conv", since = "1.20.0")]
    #[inline]
    pub fn from_bits(v: u32) -> Self {
007F3300  sub         esp,8  
007F3303  mov         eax,dword ptr [esp+0Ch]                                   EAX := 7F8D9FA6
        // SAFETY: `u32` is a plain old datatype so we can always transmute from it
        // It turns out the safety issues with sNaN were overblown! Hooray!
        unsafe { mem::transmute(v) }
007F3307  mov         dword ptr [esp+4],eax                                     *(unsigned int*)(ESP+4),x := 0x7f8d9fa6     Expected
007F330B  fld         dword ptr [esp+4]                                         ST0 = 1#QNAN 
007F330F  fstp        dword ptr [esp]                                           *(unsigned int*)(ESP+0),x := 0x7fcd9fa6     High bit of the mantissa set
    }
007F3312  fld         dword ptr [esp]  
007F3315  add         esp,8  
007F3318  ret  

(EDIT: And a discussion thread for posterity too: https://discord.com/channels/273534239310479360/335502453371961344/746445211260944444)

[fu5ha]

@Lokathor i rebased this on the avx changes and added impls for those types as well, and I also #[cfg]'d out the tests that don't pass for now with FIXME's to put them back in once #43 lands

[Lokathor]

whoops forgot about this.

we good here?

[fu5ha]

Believe so :)

…

On Thu, Aug 27, 2020, 7:51 PM Lokathor ***@***.***> wrote: whoops forgot about this. we good here? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#48 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGYXH26CEN2KG7NCTOUKYTSC4LT5ANCNFSM4QHRXXCA> .