use tbaa.struct to inform LLVM about struct/enum padding

[thestinger]

This will allow LLVM to be lazier when copying the contents.

[catamorphism]

Visiting for bug triage. Performance improvements are always good, but this one doesn't look like a release blocker.

[pcwalton]

#10613 is an example of how this bites us.

[brson]

Can anybody mentor this?

[emberian]

I'm working on this.

[erickt]

@cmr blogged about trying to implement this here.

[Aatch]

Investigation in to how LLVM uses tbaa.struct suggests that it doesn't actually help with padding, it just informs LLVM of the types inside what would otherwise be an opaque pointer to help with alias analysis.

If/When we get actual TBAA, outputing tbaa.struct metadata may be useful, but my investigations have shown that LLVM doesn't use it for avoiding padding, etc.

Maybe somebody with more knowledge on LLVM could confirm? @dotdash?

[dotdash]

I came to the same conclusion when I looked at the LLVM code.
Am 20.01.2015 21:15 schrieb "James Miller" notifications@github.com:

Investigation in to how LLVM uses tbaa.struct suggests that it doesn't
actually help with padding, it just informs LLVM of the types inside what
would otherwise be an opaque pointer to help with alias analysis.

If/When we get actual TBAA, outputing tbaa.struct metadata may be useful,
but my investigations have shown that LLVM doesn't use it for avoiding
padding, etc.

Maybe somebody with more knowledge on LLVM could confirm? @dotdash
https://github.com/dotdash?

—
Reply to this email directly or view it on GitHub
#6736 (comment).

[thestinger]

It used to use it but it was lost after various rewrites of the relevant passes. It could be taught to make use of it again, of course.

[nox]

Hey, psst. @rust-lang/wg-codegen

[nox]

Also cc @rust-lang/compiler.

[nox]

Given @cmr's blog post and @Aatch and @dotdash comments which seem to indicate that it doesn't actually help with padding, I suggest we just close this issue.

[oli-obk]

I agree, without a concrete case of something that we want to improve, this issue is not really actionable