Use _NODISCARD_FRIEND everywhere

[StephanTLavavej]

Thanks to @MattStephanson in #2208, we now have _NODISCARD_FRIEND, a centralized workaround for a bug that affects CUDA.

In theory, we could restrict usage of _NODISCARD_FRIEND to C++14 (and possibly C++17 - I'm not sure if the newer CUDA versions that accept C++17 are affected by this bug). However, we've already accumulated some usage in C++20 code:

STL/stl/inc/__msvc_int128.hpp

Line 298 in 2787718

_NODISCARD_FRIEND constexpr bool operator<(const _Base128& _Left, const _Base128& _Right) noexcept {

As the macro is exactly the same number of characters, it seems simpler to use it everywhere.

This should fix DevCom-1555597 / VSO-1426765 "VS2019 standard exception headers: warning C5240: 'nodiscard': attribute is ignored in this syntactic position".

[AlexGuteniev]

I don't like that nodiscard is combined with friend. There's nodiscard message PR where nodiscard is replaced with more specific. Having nodiscard friend macro may require more message-specific macros

[CaseyCarter]

Having nodiscard friend macro may require more message-specific macros

_NODISCARD(message) and _NODISCARD_FRIEND(message) is two macros - doesn't seem too terrible to me.

[CaseyCarter]

I validated that there's now a single match for the regex _NODISCARD\s*friend in the STL, the expansion for the macro _NODISCARD_FRIEND.

[StephanTLavavej]

I don't like that nodiscard is combined with friend.

I'm also not delighted that this workaround is necessary, but CUDA support is an important scenario, and (as Casey mentioned) while there will be merge conflicts with nodiscard-message, I don't believe this represents a true conflict - notably, there should be no need for two kinds of messages (friends vs. non-friends).

At some point in the future, we may be able to drop support for CUDA 10.1 Update 2, but that's a decision for our bosses and boss-like entities (my vague understanding is that users in the CUDA ecosystem upgrade a bit faster than one might expect, due to the need to support new GPUs, but I don't know any concrete percentages for usage).

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.