f32::powi(_,0) broken on raspberry pi (beta and nightly)

[kali]

On a rapsberry pi, rust installed through with rustup:

pi@raspberrypi:~ $ echo 'fn main() { println!("{}", 12345.0f32.powi(0)); }' | rustc +stable - ; ./rust_out
1
pi@raspberrypi:~ $ echo 'fn main() { println!("{}", 12345.0f32.powi(0)); }' | rustc +beta - ; ./rust_out
12345

Stable is correct, but beta and nightly are terribly wrong, so with have a regression. It specifically breaks num-bigint tests on raspberry pi.

[TimNN]

I tried to find the nightlies between which this was introduced, however the best I can do is some time between nightly-2016-08-03 and nightly-2016-10-03, since arm nightlies have been broken in between (the SIGILL + cross compile symbols bugs).

I can get nightly-2016-08-18 to work correctly by specifying -Z orbit=off, although I don't think this is a mir issue, since nightly-2016-08-03 works with -Z orbit=on.

[TimNN]

In case it helps anyone, assembly and ir of the following demo are here: https://gist.github.com/anonymous/a2a13985180e9960579f5f4a5372596a

#[inline(never)]
fn foo() -> f32 {
    12345.0f32.powi(0)
}

fn main() {
    println!("{}", foo());
}

[kali]

Forgot to mention that as far as I can tell, it only impacts the debug builds. -O builds look fine.

[TimNN]

@kali: At least for my example, when compiling with -O (asm / ir) the expression get's const folded by llvm.

Edit: Indeed, when preventing llvm from const folding the powi like below, the error occurs as well with -O.

#[inline(never)]
fn foo(i: i32) -> f32 {
    12345.0f32.powi(i)
}

fn main() {
    let s = String::new();

    println!("{}", foo(s.len() as i32));
}

[kali]

For what it's worth, I can not reproduce it with armv7-apple-ios on current nightly.

[nikomatsakis]

cc @japaric -- maybe related to the builtins stuff?

[brson]

@japaric if you know the fix, can it be backported to beta in the next day or two?

[brson]

Oh, it's related to optimization, so maybe not intrinsics.