Split wchars tests from the normal variant

[Jonathan Wakely]

On 22/12/20 18:27 -0300, Alexandre Oliva wrote:
>
>This change extracts apart the wchar specific parts of character
>conversion tests to allow conditonalizating these parts on actual
>wchar support while applying the rest more generally.
>
>This turned out useful during our work on the libstdc++ support
>for VxWorks, to expose the problematic areas more precisely.
>
>Regstrapped on x86_64-linux-gnu, and tested with -x-arm-wrs-vxworks7r2.
>Ok to install?  (dg-requires-wchars is added by another patch by
>Corentin, that I posted a few minutes ago)
>
>While updating Corentin's patch for mainline, I brought over to the
>split-out test even the preprocessor conditional that is present in the
>current version of the test, but required/implied by dg-requires-wchars.
>Maybe that's excessive.  Maybe the whole patch is excessive, given that
>conditional, but I didn't want to just drop it without asking for
>others' thoughts.

I do think this is excessive. The point of the test is only to verify
that calling from_chars with wchar_t gives an error. I don't think we
need to make that conditional on whether wchar_t is supported or not.
Adding a whole new test and checking the dg-requires... condition adds
non-zero overhead to the testsuite.

Following the theme of my other replies, maybe _GLIBCXX_USE_WCHAR_T
isn't even the right thing to check here. We don't require any
support for wchar_t in this test, we only require the type to be
defined. Simple changing _GLIBCXX_USE_WCHAR_T to __SIZEOF_WCHAR_T__
seems like a better fix. That will mean that we use the type if it's
defined, and not otherwise. We don't care if the library actually
supports wchar_t specializations for std::char_traits etc. because we
are expecting to get an error anyway.