[PATCH] D135238: [clang] adds copy-constructible type-trait builtins

Thu Oct 6 17:08:41 PDT 2022

cjdb added inline comments.

================
Comment at: clang/include/clang/Basic/TokenKinds.def:528
+TYPE_TRAIT_1(__is_nothrow_copy_constructible, IsNothrowCopyConstructible, KEYCXX)
+TYPE_TRAIT_1(__is_trivially_copy_constructible, IsTriviallyCopyConstructible, KEYCXX)
 TYPE_TRAIT_2(__reference_binds_to_temporary, ReferenceBindsToTemporary, KEYCXX)
----------------
jwakely wrote:
> cjdb wrote:
> > ldionne wrote:
> > > cjdb wrote:
> > > > erichkeane wrote:
> > > > > royjacobson wrote:
> > > > > > erichkeane wrote:
> > > > > > > cjdb wrote:
> > > > > > > > erichkeane wrote:
> > > > > > > > > So this one is a whole 'thing'.  The Clang definition of 'trivially copy constructible' is a few DRs behind.  We should probably discuss this with libcxx to make sure use of this wouldn't be broken.
> > > > > > > > I'd prefer to get those DRs in before finalising D135238 and subsequent ones. Do you know the DR numbers I should be looking at, or should I just poke npaperbot?
> > > > > > > Not off the top of my head, Aaron and I both poked at it at one point trying to get trivially constructible right at one point, but I think we both gave up due to the ABI/versioning concerns.
> > > > > > Maybe DR1734? Although it's about the trivially copyable trait, not trivially copy constructible. 
> > > > > > 
> > > > > Yeah, I think that was the DR, that number sounds familiar.
> > > > The `__is_trivially_*` traits were, in part, what inspired the Great Split of D116208. I could remove them for now and revisit once I rip my hair out over these DRs, if that would substantially improve the chances of these commits landing (other commentary notwithstanding).
> > > I am not sure I see a problem with the "triviality" part of this -- we already use a compiler builtin for `std::is_trivially_constructible`, so I would expect either this patch is fine, or we already have a latent bug in libc++.
> > > 
> > > I think I can echo @philnik's comment about this not necessarily providing the biggest benefit since our implementation of `std::is_trivially_copy_constructible` is a fairly trivial wrapper on top of `__is_trivially_constructible`, but I wouldn't object to the patch on that basis. I think it would probably be possible to instead provide a set of basis builtin operations that we can then build all of the library type traits on top of -- that would provide the highest bang-for-our-buck ratio.
> > > 
> > > At the same time, there's something kind of enticing in the consistency of defining every single type trait as a builtin, without exception. If that's the end goal, I think that would also be neat and we'd likely regroup all of our type traits under a single header, since each of them would literally be a one liner.
> > > 
> > > There's also the question of whether GCC provides these builtins -- if they don't and if they don't have plans to, then we'd actually need to add complexity in libc++ to support both, which we would be unlikely to do given that there's probably not a huge compile-time performance benefit.
> > > 
> > > TLDR, I think the two questions that can help gauge how much interest there will be from libc++ to use this are:
> > > 
> > > 1. Is the plan to provide *all* the type traits as builtins?
> > > 2. Will GCC implement them?
> > > 
> > > That being said, libc++ might not be the only potential user of these builtins, so I wouldn't necessarily make it a hard requirement to satisfy us.
> > > 
> > > I think I can echo @philnik's comment about this not necessarily providing the biggest benefit since our implementation of `std::is_trivially_copy_constructible` is a fairly trivial wrapper on top of `__is_trivially_constructible`, but I wouldn't object to the patch on that basis.
> > 
> > I haven't had time to do anything properly in the way of benchmarking, but after looking at @philnik's quoted code, I see that I'd naively addressed `__is_constructible(T, T const&)`, forgetting that `__add_lvalue_reference` would've fixed that issue.
> > 
> > > 1. Is the plan to provide *all* the type traits as builtins?
> > 
> > Yes, with the possible exception of `enable_if` and `add_const` etc. (see D116203 for why the qualifier ones aren't already in). The hardest ones will probably be `common_type`, `common_reference`, `*invocable*`, and `*swappable*`. The former two depend on technology that doesn't exist in Clang yet, and the latter two are likely hard due there not being prior art.
> > 
> > > 2. Will GCC implement them?
> > 
> > @jwakely do you know if there can be cross-compiler synergy here?
> > 
> > 
> Which traits are you asking about, just the __is_{,trivially,nothrow}_copy_constructible ones? Or all type traits?
> 
> Either way, I think the answer is no. We already use __is_{,trivially,nothrow}_constructible for the copy-constructible traits, and it works fine. I'm not aware of any problems with it. So I don't see any benefit in adding these ones. As already mentioned, it creates *more* code to maintain when using these traits, because we still have to support older versions of other compilers that don't have all the new traits yet.
> 
> We've added traits for a few more things recently (`__remove_cv`, `__remove_cvref`, `__remove_reference`) but they can only be used in limited ways (you can't use them directly to implement `std::remove_cv_t` for example, because they don't mangle to the same thing, which is observable). So I'm sceptical about the benefits of implementing *every* trait as a built-in. There are some which are bottlenecks and worth doing, and some which aren't.
> Which traits are you asking about, just the __is_{,trivially,nothrow}_copy_constructible ones? Or all type traits?
> 
> Either way, I think the answer is no. We already use __is_{,trivially,nothrow}_constructible for the copy-constructible traits, and it works fine. I'm not aware of any problems with it. So I don't see any benefit in adding these ones. As already mentioned, it creates *more* code to maintain when using these traits, because we still have to support older versions of other compilers that don't have all the new traits yet.
> 

Thanks for filling in the knowledge-gap here. I think libc++ only supports the most recent two releases of Clang (and I think only the latest release of GCC), so that library code would have an expiration date of around most eighteen months. So I'm clearer on the libstdc++ side of things, how far back do you normally support?

> We've added traits for a few more things recently (`__remove_cv`, `__remove_cvref`, `__remove_reference`) but they can only be used in limited ways (you can't use them directly to implement `std::remove_cv_t` for example, because they don't mangle to the same thing, which is observable).

Interesting. Is this a libstdc++ backwards-compat issue, or does it burrow into the Itanium ABI, which has mangling conventions for builtin type transformers?

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135238/new/

https://reviews.llvm.org/D135238