AsRef/AsMut implementations aren't inlined

[WildCryptoFox]

Many AsRef and AsMut implementations in core do not have #[inline] and yet these tiny no-op functions are called from the assembly.

The following assembly maps to AsMut<[T]> for [T]

Any quick search for as_ref or as_mut in convert.rs shows many instances, some with but many without #[inline].

_ZN63_$LT$$RF$mut$u20$T$u20$as$u20$core..convert..AsMut$LT$U$GT$$GT$6as_mut17h47b72d0f8c3b94ffE:
     .cfi_startproc
     movq    (%rdi), %rax
     movq    8(%rdi), %rdx
     retq

Edit: Default::default is another example where more aggressive inlining should be used

[jonas-schievink]

You just need #[inline]. The LLVM hint wouldn't even be needed at all - #[inline] puts the function into all codegen units that reference it, that's all you typically need for small functions to be inlined.

You almost never want #[inline(always)], since it's even respected in debug mode, where it just needlessly impacts compile times.

[WildCryptoFox]

Thanks @jonas-schievink for the correction.

[Olavhaasie]

I would like to do this, If no one else has taken it yet

[sfackler]

LLVM should normally be inlining in release builds these unless you're using multiple codegen units without LTO.

[WildCryptoFox]

@sfackler I haven't tweaked the config for codegen units nor LTO. So whatever the default is.

Edit: Why was the E-easy tag removed? Isn't this just a matter of adding #[inline] or are you saying most instances where we have #[inline] should be replaced with trusting llvm to do its part and fixing whatever relevant that isn't already triggering it in this case?

[sfackler]

I remembered that we generally don't want to add #[inline] on functions that are already eligible for inlining: #52704.

[WildCryptoFox]

Thanks @sfackler. This comment suggests using ThinLTO in release builds incremental builds disabled but obviously that isn't desired as incremental building is very desirable for rapid testing (especially in my case where I need to read the optimized assembly).

[sfackler]

Incremental ThinLTO works now IIRC.

[WildCryptoFox]

Cool. And enabled by default on at least the latest nightly. Should this be closed?

[lcnr]

This seems to be fixed.

See https://rust.godbolt.org/z/j9Ts8P