search

edef1c / libfringe

a Rust library implementing safe, lightweight context switches, without relying on kernel services
https://edef1c.github.io/libfringe
Apache License 2.0
512 stars 31 forks source link

No swap() inline #60

Closed MarkSwanson closed 8 years ago

MarkSwanson commented 8 years ago

This seems to have fixed the relocation troubles that occur when building a dso with libfringe.

edef1c commented 8 years ago

libfringe's speed advantage comes entirely from inlining. This is literally the most core part of our implementation technique — ie this won't be merged. The LLVM assembler appears to be broken, and I'm okay with workarounds for that, but this defeats the very purpose of libfringe.

MarkSwanson commented 8 years ago

I submit to you that LLVM is smart enough to still inline - the benchmarks seem to imply this. Maybe because the Fn is small enough it meets some LLVM metric to inline anyway?

I'm hoping you're open to explaining in a bit more detail.

I'm guessing you're worried that because there is no explicit #[inline(always)] maybe LLVM won't auto-inline it after some future update? The metrics might change, Rustc might do something different...?

Thanks.

edef1c commented 8 years ago

#[inline(always)] precisely states our intentions here, so I'd rather keep it. If it still inlines without it, but the attribute causes it to behave differently, that's a rustc bug.