Windows TLS destructors don't run in a statically linked binary

[nagisa]

On -msvc TLS destructors do not appear to be run at all.

On -gnu TLS destructors appear to run when a non-main thread exits (as expected) or std::process:exit is called (es expected), but not run when main thread is terminated normally (by returning from fn main; not expected).

Various variations of this/this program has been used for testing.

cc @retep998 who did the testing

[nagisa]

// The linker must not discard p_thread_callback_base. (We force a reference to this variable with a linker /INCLUDE:symbol pragma to ensure that.) If this variable is discarded, the OnThreadExit function will never be called.

and

//.CRT section is merged with .rdata on x64 so it must be constant data.
#pragma const_seg(".CRT$XLB")
// When defining a const variable, it must have external linkage to be sure the linker doesn't discard it.

are comments from the chromium’s TLS impl pointing towards something that might be missing from our implementation: we do not make sure data is constant on x64 and we do not make sure our symbol is \INCLUDEd.

[alexcrichton]

The bug here for threads in general specifically has to do with linkage, and I'm about to fix that in #28130, and for the main thread I've opened #28129 as it's not specific to Windows unfortunately.

For now I'm gonna recategorize this as "TLS destructors don't run in a statically linked binary on Windows"