[PATCH] D128113: Clang: fix AST representation of expanded template arguments.

Fri Aug 19 02:21:16 PDT 2022

davrec added a comment.

In D128113#3733764 <https://reviews.llvm.org/D128113#3733764>, @mizvekov wrote:

> In D128113#3733051 <https://reviews.llvm.org/D128113#3733051>, @joanahalili wrote:
>
>> We have a translation unit, on which we see an increase of compilation time and clang memory allocation from 11GB to 14GB. We are working on an isolated case.
>
> Thanks for looking into this!
>
> What I can imagine may be happening here is that if we instantiate a template with a very large argument pack consisting of mostly the same template arguments, we will create many more `SubstTemplateTypeParmType` nodes with this patch, as they won't unique so much anymore, because each will have a different pack index.

If this is indeed a major problem, and/or if performance is an issue for other of the patches in the stack, maybe the solution is along the lines of what I raised in https://discourse.llvm.org/t/rfc-improving-diagnostics-with-template-specialization-resugaring/64294/5: introduce an option that determines how much sugar we construct in the AST, and modify `ASTContext::get*Type(...)` accordingly.

For now it could have two possible values, 'medium' and 'maximum', and `getSubstTemplateTypeParmType(...)` could be modified so it always uses PackIndex=0 whenever it is only set to 'medium'.  And, for later patches, resugaring would only be enabled when it is set to 'maximum'.  (And maybe later a 'minimum' option could be experimented with which disables all sugar except that needed to keep tests passing.)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128113/new/

https://reviews.llvm.org/D128113