[DRAFT] One idea for the deduction guides

[Quuxplusone]

Posting as a PR just so that the CI will run (I hope). This all works on my local machine.

I see now the problem with not-constraining the less specific deduction guide.

template <typename Alloc, typename Value>
polymorphic(std::allocator_arg_t, Alloc, Value) -> polymorphic<
    Value, typename std::allocator_traits<Alloc>::template rebind_alloc<Value>>;

Before we can do overload resolution between this deduction guide and the more specific one, we must fully evaluate this one's type... which requires evaluating std::allocator_traits<Alloc>::template rebind_alloc<Value>, which explodes because the deduced type Alloc is not actually an allocator type. So, even though the other deduction guide would have been the better match, this one explodes before we even get to the overload-resolution step.

Here's a reduced example of what I'm talking about: https://godbolt.org/z/jWzn5PGn6

It looks to me as if the STL currently doesn't try to rebind allocators in deduction guides at all, with the sole exception of flat_set/flat_map, which rebind specifically in their from_range_t constructors and nowhere else. The from_range_t constructors (unlike the allocator-extended copy constructor) cannot deduce the allocator type in any other way, so it would be a hard error to pass in a non-allocator type there anyway.

So I think there would be good precedent for adopting an approach like the one in this PR. But I do recognize that this reduces the surface area of the library a little bit: I had to change some unit tests to avoid testing the pattern that will no longer work after this patch. However, notice that the analogous tests wouldn't pass for std::vector or std::flat_set, either!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.60%. Comparing base (840a406) to head (4554ce7).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #558   +/-   ##
=======================================
  Coverage   99.60%   99.60%           
=======================================
  Files          11       11           
  Lines         750      750           
  Branches       76       76           
=======================================
  Hits          747      747           
  Misses          3        3           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jbcoe]

Thanks @Quuxplusone we don't want to be doing things differently to vector so this looks like a sound direction.

I've been recovering from a cold so will be slow to properly respond to this.

On a first look, a minor nit is that we try to make indirect.h and polymorphic.h work as stand-alone single-include headers (for godbolt). I think your PR lightly couples them to feature_check.h. It would, however, be easy to clean up post-merge.