f32::NAN as i128 different between miri and LLVM

[bjorn3]

fn main() {
    let f = std::f32::NAN;
    assert_eq!(f as i128, 0x7fffffffffffffffffffffffffffffffi128); // LLVM
    assert_eq!(std::f32::NAN as i128, 0i128); // miri
}

[ollie27]

Duplicate of #10184

[bjorn3]

It doesn't seem like the LLVM case is caused by UB, as the compiler-builtins intrinsic returns 0x7fffffffffffffffffffffffffffffffi128 too.

[ollie27]

Well it definitely is UB. I don't think it's really defined anywhere what the compiler-builtins intrinsic should return in this case so what it returns doesn't really mean anything.

[bjorn3]

In that case the saturating-float-casts test shouldnt test for a specific value being returned.

[ollie27]

That test needs to be compiled with -Z saturating-float-casts:

rust/src/test/ui/numbers-arithmetic/saturating-float-casts.rs

Line 3 in c798dff

// compile-flags: -Z saturating-float-casts

See #10184 for more information.