<crtdefs.h>: Header is missing from this repo

[BillyONeal]

Describe the bug
In #593 @AlexGuteniev indicated that our __dllimport macros are coming from crtdefs.h which is not presently in the repo; that is part of "vcruntime".

We've resolved half of this by implementing new satellite DLLs with a .def file and not using any import/export macro. This is the approach that we chose (much later after this issue was originally filed) for the latest satellite DLL atomic_wait:

STL/stl/src/msvcp_atomic_wait.src

Line 4 in c2527c4

; atomic wait satellite DLL definition

The remaining work here is that we could move vcruntime into github (we already made our directory structure allow for that). We're hopefully planning to do this in the future, but no ETA yet.

[AlexGuteniev]

• We could move vcruntime into github (we already made our directory structure allow for that).

That sounds interesting as it might help in other places.

I've asked why is _YIELD_PROCESSOR() defined as yield instruction on ARM:

STL/stl/inc/xatomic.h

Line 40 in 285187b

#define _YIELD_PROCESSOR() __yield()

but not as pause on x86 / x86-64:

STL/stl/inc/xatomic.h

Line 31 in 285187b

#define _YIELD_PROCESSOR()

Assuming there's no particular reason, and putting pause would be useful, I'll need _mm_pause in <intrin0.h>. But again, <intrin0.h> is not here on Github.

[BillyONeal]

The answer I've heard from folks that supposedly know better is "never _mm_pause ever" -- that macro was defined long before my time but it was defined as empty for x86/x64 and I don't think we would want to change that without extremely strong profiling evidence that it was a good idea.

[AlexGuteniev]

Ok, then YieldProcessor() in SDK headers is probably wrong.
Anyway, I guess I need to seek answer in Intel / AMD manuals.

[AlexGuteniev]

(But hey, Intel TBB uses _mm_pause! They are supposed to know the best)

[AlexGuteniev]

Even WaitOnAddress itself uses pause for its internal spinning (I don't have access to Windows source, but can step into using a debugger).

I'm not insisting _YIELD_PROCESSOR() should be redefined, but I think the full rule of _mm_pause usage is not as simple as "never _mm_pause ever"

[jonwil]

It would definitely be nice to see the source code to the VC Runtime stuff (vcruntime140.dll, vcruntime.lib, msvcrt.lib and the like) made open source.

I assume crtdefs.h is needed by things other than the STL and moving it into the STL repo isn't an option?

[StephanTLavavej]

Yeah, that would be a layering issue. We're looking into open-sourcing almost all of vcruntime/vcstartup but there are many gears that need to turn before that can happen.

[StephanTLavavej]

Variant of "We could do all new satellite DLLs with a different header than crtdefs.h that lives in this repo." - we could simply move the relevant macros into yvals.h instead of introducing a whole header. (After solving logistics issues related to the move, since MSVC has to ship Previews before we can consume its headers.)

[AlexGuteniev]

I think I have fourth idea, which looks better.

Can do this, like .def without .def, in implementation .cpp:

#ifdef _BUILDING_SATELLITE_ATOMIC_WAIT

#pragma comment(linker, "/export:__std_atomic_set_api_level")
#pragma comment(linker, "/export:__std_atomic_get_spin_count")
#pragma comment(linker, "/export:__std_atomic_wait_get_deadline")
#pragma comment(linker, "/export:__std_atomic_wait_direct")
...
#endif // _BUILDING_SATELLITE_ATOMIC_WAIT

Then don't need to decorate names anyhow, but still using preprocessor condition passed from make file

The best part of it is that I can conditionally use forwarders (#763):

#ifdef _ATOMIC_WAIT_STATICALLY_AVAILABLE_TO_IMPL

#pragma comment(linker, "/export:__std_atomic_notify_one_direct=api-ms-win-core-synch-l1-2-0.WakeByAddressSingle")
#pragma comment(linker, "/export:__std_atomic_notify_all_direct=api-ms-win-core-synch-l1-2-0.WakeByAddressAll")

#else // ^^^ statically available / not statically available vvv

#pragma comment(linker, "/export:__std_atomic_notify_one_direct")
#pragma comment(linker, "/export:__std_atomic_notify_all_direct")

#endif // _ATOMIC_WAIT_STATICALLY_AVAILABLE_TO_IMPL

[AlexGuteniev]

Syntax with #pragma comment(linker, "/export...") proved to be not a good idea on x86, when there are name decorations involved dependent on calling conventions. .def file somehow bypasses the decoration problem, and #pragma comment somehow fails.
Manually specifying decorated name, which would work for #pragma comment, looks like a bad option from maintenance viewpoint, and just looks ugly.
Still #pragma comment works well for frowarders -- if there would be a decision to use them.