TBAA: `-new-struct-path-tbaa `status

IR & Optimizations
1

-new-struct-path-tbaa is work from ~6Y ago (although I think the discussion started ~10Y ago ) authored by @kosarev with fixes from ~1Y ago by @brunodf. It creates !tbaa metadata for structs and members where it wasn’t available before - in particular improving the optimizations for strong-typedefs, making them essentially zero-cost abstractions as they are advertised to be. Godbolt toy example here.

I really hope this work can stop hiding behind the -new-struct-path-tbaa switch. I built and tested some projects with it and personally I consider it stable. Is there any planned route forward with this?

1 Like
2

I would propose you make a patch to turn it on. We can use @nikic’s compile time tracker to see the impact. If it works alright, we can turn it on. If we see something in the meantime, we can investigate further.

3

Unfortunately when enabling it by default TBAAMetadataTest fails - it expects a different !tbaa layout. I couldn’t figure out the semantics of the new !tbaa - I can just change the expected result to match. Should I?

4

We’d need that for sure. We should test more software and see if the new encoding itself is causing problems.

5

@jdoerfert I finished fixing the regular tbaa test and setting up a matching new-path-struct-tbaa test, but came across what seems like a regression, for virtual base classes. The following code is from TBAAMetadataTest.VirtualBase:

    struct Base {
      int f32;
    };

    struct Derived : public virtual Base {
      short f16;
    };

    void func(Base *B, Derived *D) {
      B->f32 = 14;
      D->f16 = 35;
      D->f32 = 77;
    }

Regular tbaa includes tbaa attribute for Derived, while new-struct-path-tbaa doesn’t :

image
image991×203 20.4 KB

You can check interactively at this godbolt:

The final assembly for this particular code is identical AND virtual base classes are pretty rare, so I’m not sure if this is a blocking issue.
I suggest:
(1) I’ll submit a patch with just the new tests for new-struct-path-tbaa (with matching expected results, i.e without this missing Derived tbaa metadata).
(2) Maybe I’ll open an issue for this regression? Do you think it has a chance of receiving attention? (Who should I @mention here or at github for this?)
WDYT?