stage2: `zig cc` compiled `powf` randomly returns NaN on Windows

[jagt]

Zig Version

0.10.0-dev.3653+7152a58c1

Steps to Reproduce

#include <stdio.h>
#include <math.h>

int main() {
    printf("! %f\n", powf(1.389f, 100.0f + 0.1f));

    int acc = 0;
    for (float xx = 0.f; xx < 100.0f; xx += 0.1f)
    {
        float v = powf(1.389f, xx);
        printf("%f: %f\n", xx, v);
        if (acc++ > 3) break;
    }
 
    return 0;
}

Expected Behavior

With clang it's like this:

clang -g -o powtest.exe .\powtest.c
.\powtest.exe
! 192529106468864.000000
0.000000: 1.000000
0.100000: 1.033404
0.200000: 1.067924
0.300000: 1.103597
0.400000: 1.140462

Actual Behavior

zig cc -g -o powtest2.exe .\powtest.c
.\powtest2.exe
! nan
0.000000: 1.000000
0.100000: nan
0.200000: nan
0.300000: nan
0.400000: nan

Note that this is reproducible with the zig.exe from zig win dev kit

Thanks in advance!

[jagt]

I locally toggled verbose linker and seems the pow is from mingwex:

lld-link: Loaded mingwex.lib(pow.obj) for pow
lld-link: Reading msvcrt-os.lib(mingw_lock.obj)
...
lld-link: Reading mingwex.lib(ldexp.obj)
lld-link: Loaded mingwex.lib(ldexp.obj) for ldexp
lld-link: Reading mingwex.lib(powi.obj)
lld-link: Loaded mingwex.lib(powi.obj) for __powi

[kubkon]

This is a miscompilation of stage2 of compiler-rt symbols. Building the same program with stage1 doesn't have this problem:

> zig cc -g -c powtest.c
> zig build-exe powtest.obj -fstage1 -lc
> .\powtest.exe
! 192529106468864.000000
0.000000: 1.000000
0.100000: 1.033404
0.200000: 1.067924
0.300000: 1.103597
0.400000: 1.140462

[jagt]

I spent some time digging and seems it's indeed compiler-rt miscompilation

>zig cc -c -g .\powtest.c
>zig build-exe <path-to-llvm>\compiler-rt\lib\builtins\x86_64\chkstk.S .\powtest.obj -fno-compiler-rt -lc --name powtest3
D:\Dev1\zig\tmp>.\powtest3.exe
@ 192529106468864.000000
0.000000: 1.000000
0.100000: 1.033404
0.200000: 1.067924
0.300000: 1.103597
0.400000: 1.140462

I took a look at the assembly and seems it's log2 returning a nan.