llvm/llvm-project#79793 is still in flight, but the affected test is blocked by a suspected MSVC compiler bug with constexpr, so we don't need to wait.

It's probably the same bug as DevCom-10439137 "Discarded id-expression causes unnecessary reading".

Yep, I agree! Great catch, I didn't know about that bug report. I've updated the entry.

I also noted in VSO-1869865 (the internal mirror of DevCom-10439137) that the ICE I just reduced, VSO-1948221 "x86chk ICE with constexpr type_info: Assertion failed: isIndirection()", is likely related - it involved constexpr and a (void) cast, which I found curious at the time.

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

I had to push an additional commit because new tests appeared that require Unix headers, and the MSVC-internal test harness doesn't understand this.

Should UNSUPPORTED: !has-unix-headers be added to magic_comments.txt? I don't understand what this file does, but apparently it's used by the MSVC-internal test harness.

@cpplearner That's a good idea. magic_comments.txt is indeed used by the MSVC-internal test harness - it's an extremely primitive mechanism that I wrote a while ago (later adapted by Casey) where if a test mentions any of those "magic comments", it's skipped. This imitates the real test harness's implementation of "feature" detection.

We already have // REQUIRES: has-unix-headers there; arguably upstream should be changed to say that, avoiding the current double negation. (The comment matching we do is also whole-line, as I recall.)

Since the PR and its mirror have passed checks, I'll explore doing this the next time I update LLVM. Thanks!