[PATCH] D90719: [DebugInfo] Modify ctor homing as workaround for unconstructed libcxx types

Tue Nov 3 17:41:35 PST 2020

dblaikie added a comment.

In D90719#2372554 <https://reviews.llvm.org/D90719#2372554>, @rnk wrote:

> In D90719#2372463 <https://reviews.llvm.org/D90719#2372463>, @dblaikie wrote:
>
>> Does Chromium need this fixed in clang? Or if it were fixed in libc++ would that be adequate? (does Chromium's build need to work with old libc++s, or does it always build with a libc++ that matches the compiler? (in the latter case, a fix in libc++ would be as good as a fix in clang))
>
> Well, we'd like to make this new type homing behavior the default, and it wouldn't work with old libc++ versions, and there is the general possibility that there is code out there like this libc++ code that has implicit nontrivial constructors that are not used. So, I wouldn't think so much about what's right for Chrome, and more about what's right for Clang.

My understanding is that such code is UB, is that right? - if libc++ was threading some needle/depending on some agreed upon Clang guarantee (even if it was a secret handshake only for libc++) then I'd be more in favor of the "we have to/should support it" but as it seems to stand, I'm inclined towards addressing this by fixing libc++ unless there's evidence of a pervasive dependence on this UB.

> If we do want to fix libc++, I'd warn you that the code is subtle. It takes care to construct the node value in a particular way, and it sets a flag after construction succeeds, which perhaps affects exception safety:
> https://github.com/llvm/llvm-project/blob/master/libcxx/include/__hash_table#L2456
> + @EricWF

Yeah, +1 on getting @EricWF's take on this for sure.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90719/new/

https://reviews.llvm.org/D90719