Out of range float casts miscompilation

[bjorn3]

I tried this code:

pub fn foo() -> f32 {
    0xffff_ffff_ffff_ffff_ffff_ffff_ffff_ffffu128 as f32
}

I expected to see this happen: Saturating to f32::INFINITY

Instead, this happened: 9223372000000000000.0 is returned

Another case is

pub fn foo() -> f32 {
    1<an absolutely huge amount of zeros>u128 as f32
}

which should give an out of range error, but returns 0.0.

Meta

• What version of Rust GCC were you using, git sha if possible: The current version at godbolt.

[philberty]

I think the code to check for this is in the GCC TREE_OVERFLOW flags

[philberty]

@bjorn3 I have a branch of code to fix this but the part about Saturating to f32::INFINITY I might leave out for the first PR, I have code to catch this case but I wonder is there documentation that I can reference on this saturating to infinity?

[bjorn3]

This is the PR which originallt changed it from UB to saturating: rust-lang/rust#71269