[SYCL] Finally fix iostream_proxy.hpp

[aelovikov-intel]

When _MT and _DLL are both defined, the user has passed /MD or /MDd and
is dynamically linking in the runtimes. Because std::cout etc are data,
they need to be marked dllimport in that case or else the linker cannot
find them.

[aelovikov-intel]

Many thanks to @AaronBallman for helping with this!

[AaronBallman]

FWIW, these changes LGTM.

For reference, when _MT and _DLL are both defined, the user has passed /MD or /MDd and is dynamically linking in the runtimes. Because std::cout etc are data, they need to be marked dllimport in that case or else the linker cannot find them.

[pvchupin]

when _MT and _DLL are both defined, the user has passed /MD or /MDd and is dynamically linking in the runtimes. Because std::cout etc are data, they need to be marked dllimport in that case or else the linker cannot find them.

This should be in PR description I think.

[aelovikov-intel]

I believe

Failed Tests (2):
  SYCL :: Assert/assert_in_simultaneously_multiple_tus.cpp
  SYCL :: Assert/assert_in_simultaneously_multiple_tus_one_ndebug.cpp

are known and there was/were issue(s) submitted for that. @intel/llvm-gatekeepers , ready to merge?

[pvchupin]

I've restarted testing yesterday to check if it's really flaky and it didn't help. Pushed one more restart today... just for curiosity...
I don't think fail is related to the change because it's NFC on Linux it seems.

#6463 looks related.
There is also #6520 for HIP.

[aelovikov-intel]

Failing test is being disabled at intel/llvm-test-suite#1127.