Use sqrt intrinsic for floating point sqrt

[brson]

sqrt is still using libc

[huonw]

I'm running tests on a patch that uses intrinsics for sqrt, sin, cos, exp, etc.

[thestinger]

@dbaupp: Nice! It would be awesome to have all of the useful LLVM intrinsics for integers/floats exposed.

[huonw]

Yay! 5x faster:

use core::float::sqrt;

fn main() {
    let mut sum: float = 0.0;
    for core::uint::range(0, 100000000) |i| {
        sum += sqrt(i as float);
    }
    io::println(fmt!("%f", sum));
}

# using cmath
$ rustc --opt-level=3 intrinsics-bench.rs  && time ./intrinsics-bench 
666666661666.567017

real    0m2.729s
user    0m2.712s
sys     0m0.008s

# using intrinsics
$ x86_64-unknown-linux-gnu/stage2/bin/rustc --opt-level=3 intrinsics-bench.rs  && time ./intrinsics-bench 
666666661666.567017

real    0m0.513s
user    0m0.500s
sys     0m0.008s

[brson]

@dbaupp mentioned on reddit that using some of these clobbers the stack canary. I hadn't considered that before and we need to think about how these are implemented. How many of the intrinsics that we are already using are just trotting all over the stack red zone? Perhaps LLVM's intrinsics don't emit the proper stack growth prologues.

[huonw]

(The very brief reddit thread, for future reference.)

Do I continue working on this? (i.e. open a pull request, since I've got the code basically organised.)

The (most) dangerous intrinsics seem to be sin, cos and exp, so I've left sin as libc and added sin_intr for the intrinsic etc.

[brson]

@dbaupp I'd like to know what assembly they generate. Do you have a test case that demonstrates all of them?

[huonw]

@brson: https://github.com/mozilla/rust/pull/5697

[pcwalton]

@huonw The necessary LLVM changes have been made.

However, I didn't realize that intrinsics can clobber the stack. I guess we need to write wrappers with the fixedstacksegment attribute for the problematic intrinsics (sin, cos) as well! This can be done entirely in the Rust compiler now, no need to patch LLVM.

[huonw]

@pcwalton Does fixed_stack_segment work on extern/intrinsic functions? i.e.

#[abi = "rust-intrinsic"]
pub extern "rust-intrinsic" {
   #[fixed_stack_segment]
   pub fn sinf32(x: f32) -> f32;
}

[pcwalton]

@huonw Extern functions automatically require fixed stack segments, if the #[fast_ffi] attribute is used. (Eventually this attribute will no longer be necessary.) However, I don't think that it will work for functions with the rust-intrinsic ABI, because I didn't realize that some intrinsic functions would need it.

[huonw]

@brson #5975 fixed this.

[brson]

Thanks @huonw!