decide whether global addresses should be significant by default

[thestinger]

We have #[address_insignificant] now because it's needed to have LLVM merge constants. I think insignificant addresses should be the default, and TLS should be done with a macro applying an #[address_significant] attribute.

Issue #8957 covers functions, since the situation is different as we already merge them.

[thestinger]

Nominating for the backwards compatible milestone (well defined would work too).

[alexcrichton]

If we do decide to do this, an interesting statistic to see beforehand would be the size difference in the standard libraries or librustc.

When I was dealing with address_insignificant I had to flag all of the statics as internal and private linkage, so this might not be applicable to pub statics, but that too would need further investigation to see what happens.

[thestinger]

One thing to note is that string constants are already merged, so only the address of the slice object itself is significant right now.

[brson]

TLS should not involve fn addresses at all - it was an expedient. Agree we shouldn't make promises about fn addresses if it's useful for optimizations.

[catamorphism]

Accepted backwards-compatible

[nikomatsakis]

does this relate to static muts?

[thestinger]

I thought this was only doable for constants, but it does appear you could tag static mut with unnamed_addr and it wouldn't merge them unless it can prove the content differences between the two are insignificant too.

I don't think it would ever actually result in an optimization with static mut, since a single write would make the significant of the content significance prevent merging.

[brson]

I think addresses should be insignifacant and tls should be done with type_id.

[alexcrichton]

Nominating for closure with the following policy:

• All static mut variables have significant addresses
• All static variables have insignificant addresses unless tagged with #[significant_address]

[brson]

Maybe #[inline(never)] instead of #[significant_adress].

[brson]

Sounds like there's some more to do here. Leaving open.

[brson]

What's the current thinking here?

• All static mut are significant
• All static are insignificant unless they contain certain types (which?) or have #[inline(never)]?