Today I learned some things that make me upset, in Rust: * If you divide by zero with floats, they will return `inf` and not panic * If you `as` cast an `inf` float to `i32`, it will return i32::MAX Probably totally reasonable choices in isolation + practice, but WHEW was that fun to find. - ThreadSky

jamesmunns.com • 376 days ago

Today I learned some things that make me upset, in Rust:

* If you divide by zero with floats, they will return `inf` and not panic
* If you `as` cast an `inf` float to `i32`, it will return i32::MAX

Probably totally reasonable choices in isolation + practice, but WHEW was that fun to find.

Comments

jamesmunns.com•376 days ago

(this is also true even in debug mode, where overflows would usually be panics, btw):

https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=0267399969b01fc9e1aee9d78f9cbd54

blen.blinsay.com•376 days ago

IEEE 754 🫡

jamesmunns.com•376 days ago

yep, I'm familiar. It's the second order interactions that were more surprising to me.

blen.blinsay.com•376 days ago

oh i figured you were, i just wanted to mark your tombstone appropriately

mcheshkov.bsky.social•376 days ago

AFAIR, division by zero leading to infinity without panic is just default exception handling in IEEE754, so okayish. Converting to integer outside of target range is invalid operation, I would expect it to be checked in debug, but LLVM by default assumes no-side-effects floats, so no flags.

jamesmunns.com•375 days ago

Went back and found it, "f64::INFINITY rounds to i32::MAX" is explicitly stated, so it's definitely operating as specified!

https://doc.rust-lang.org/reference/expressions/operator-expr.html?highlight=satu#numeric-cast

dejitarujin.bsky.social•374 days ago

Now I'm curious, what are these values stored as in RAM? Are they just the min/max values for a given data type? Will they still roll over, and compare correctly to literals?

jamesmunns.com•374 days ago

When you cast them, it's actually a conversion function call, so when you use `as` it's doing a conversion, not just reinterpreting the existing bits as a new ttype

mcheshkov.bsky.social•375 days ago

Well, now I’m upset too 😅

mcheshkov.bsky.social•375 days ago

Went to check if there’s a Clippy lint for that cast, found cast_possible_truncations, allow by default. It’s time to become more pedantic!

jamesmunns.com•376 days ago

I'd have to go back and look at the somewhat recent (last 1-2 years) RFC that formalized "float to int" as-casting, to see if they covered it. I'm guessing this is "behaves as specified", even if it was a little surprising to me.

Comments

Posting Rules

Reply