Enhancement: add trait getPointerBitmap to help precise scanning

[rainers]

This helps the precise GC (dlang/druntime#1022) to generate the required data without adding to the compile time.

[ghost]

Holy mother of god. :D

[rainers]

As usual with ugly code, it was wrong ;-)

[ghost]

Could you add a description either in the commit or the PR what exactly this does? A dlang.org pull will be needed documenting this trait as well (maybe just do that instead of documenting it in the commit).

[rainers]

Could you add a description either in the commit or the PR what exactly this does? A dlang.org pull will be needed documenting this trait as well (maybe just do that instead of documenting it in the commit).

I've added a comment to the code. I'll follow up with a documentation PR if this appears to have a chance of getting accepted.

[ghost]

I have to admit that it's a bit obscure. But if you believe it will help in implementing a precise GC, I'm all for it. There's no way to implement this in the library instead, right?

[rainers]

This trait is a direct translation of the RTInfo in the precise GC: https://github.com/rainers/druntime/blob/gc_precise_nov14/src/gc/rtinfo.d

The problem with the library implementation is that it adds considerably to the compilation time (more than 20% for some projects). Having the proper RTInfo always available should have minimal impact on build time, especially in case you don't use the precise GC.

[ghost]

Ok, you got me convinced. Will auto-merge soon, and hoping for that dlang pull as well. ;)

[rainers]

and hoping for that dlang pull as well. ;)

dlang/dlang.org#718

[ghost]

Auto-merge toggled on

[rainers]

Thanks!

[MartinNowak]

Make sense to have this as a compiler intrinsic.

[MartinNowak]

Just move that into main please, no other dmd test uses unittest.

[rainers]

I have copied that from another test, but I shouldn't have used traits_getUnittests.d for that :-)
The bad thing about main() and building without -unittest is that asserts are a noop for all the -release builds.

[MartinNowak]

Have you tested this with nested classes and structs?

[rainers]

Have you tested this with nested classes and structs?

TBH not with this trait. I'll add some test cases.

[rainers]

I'll add some test cases.

See #4196

[llucax]

Wow, we (mihails :P) just removed code using using a very similar approach in an old patch from @dsimcha in the draft of the port of the concurrent GC. Maybe it can be added right back :)

[9rnsr]

Why this PR is merged without @WalterBright 's approval?

[ghost]

Because it's an enhancement that doesn't change the language, it only introduces a trait (there are no stolen keywords). And it seems important enough to have as a trait, especially since @rainers is one of the few people who actually know a thing or two about working on GCs.

Also, I'm not in the mood to wait for Walter to respond 5 months later for what is an uncontroversial feature.

[9rnsr]

@andralex OK, I understand.

[ghost]

I mean I may have shotgun-pulled this, but GC is a high point of focus right now, we need to get this stuff moving. :)