Don't memcmp(x, y, -1) from ranges::lexicographical_compare

[CaseyCarter]

... even with two potentially-infinite ranges.

Drive by: Pull non-dependent test cases out of the instantiator::call template magic in tests/P0896R4_ranges_alg_lexicographical_compare.

Fixes VSO-1854238 / AB#1854238

[strega-nil-ms]

I think we should maybe reconsider calling memcmp(x, y, -1) in this case entirely, at least under __SANITIZE_ADDRESS__, since, while memcmp is UB when you don't call it with fully valid buffers¹, calling lexicographical_compare(i1, unreachable_sentinel, i2, unreachable_sentinel) is valid (at least, that's my understanding; I may fully be wrong about that and this is a test of an additional guarantee we provide).

Footnotes

1. https://trust-in-soft.com/blog/2015/12/21/memcmp-requires-pointers-to-fully-valid-buffers/ ↩

[CaseyCarter]

I think we should maybe reconsider calling memcmp(x, y, -1) in this case entirely, at least under __SANITIZE_ADDRESS__, since, while memcmp is UB when you don't call it with fully valid buffers1, calling lexicographical_compare(i1, unreachable_sentinel, i2, unreachable_sentinel) is valid (at least, that's my understanding; I may fully be wrong about that and this is a test of an additional guarantee we provide).

Footnotes

1. https://trust-in-soft.com/blog/2015/12/21/memcmp-requires-pointers-to-fully-valid-buffers/ ↩

Yeah, it's hard to argue in support of memcmp(x, y, -1). The initial motivation was "it works, and doing so means we don't need a special-case check for the case where both ranges are 'infinite'". But given how questionable this is in the first place and how quickly we've found a case where it doesn't "work", I should just add the special-case check and be done with it.

[CaseyCarter]

Okay, fixed. Thanks for reminding me how icky this was in the first place.

[CaseyCarter]

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