Runtime: emulate TLS with a map for WASI target

[MaxDesiatov]

Runtime thread-local storage has to be emulated for WebAssembly/WASI as WebAssembly currently doesn't support threads.

Related to SR-9307.

(cc @compnerd @kateinoigakukun)

[compnerd]

Why cant you use -femulated-tls?

[compnerd]

CC: @mikeash

[MaxDesiatov]

I didn't know about -femulated-tls, need to try that 😄

[xwu]

Suggested change

	_stdlib_tls_map = new _stdlib_tls_map_t();
	_stdlib_tls_map = new _stdlib_tls_map_t();

Nit: if still pursuing this PR, would you mind running the changes through the clang-format tool?

[MaxDesiatov]

@compnerd I had another look at -femulated-tls and it's unclear to me how SWIFT_THREAD_GETSPECIFIC should be defined with -femulated-tls? We would still have to implement pthread_getspecific stubs, as wasi-sdk has no pthreads. Would that be substantially different from what is proposed in this PR?

[mikeash]

I think the best long-term plan would be to add a mode to ThreadLocalStorage.cpp that can use a thread-local variable, like we have in Exclusivity.cpp. I'm not sure why we only have that in one of those files, maybe they just evolved separately. That would allow you to use -femulated-tls. Not sure if you want to do that here or save it for another time, though.

[compnerd]

Using -femulated-tls means that you don't need to provide the mapping for the TLS access. Although it will be slower than real TLS, its shifting the responsibility to a well tested library implementation (compiler-rt). With the emulated TLS, you do not use the pthread local API but rather __thread annotations and let the compiler DTRT.

[filip-sakel]

Sorry to revive this, but is TLS still an issue with built-in concurrency?

[mikeash]

The runtime still needs thread-local storage for various things even if application code is using higher-level constructs. For example, retrieving a task-local variable in user code works by using TLS to fetch the current task running on the current thread, then retrieving the thread-local variable from that task.

[compnerd]

The map access needs locking, which I believe is missing or am I misreading the patch?

Can we get away with just using -femul-tls when building for WASI? That should let us use a tested implementation for the mapping from compiler-Ty.

[kateinoigakukun]

Closing in favor of cleaner SWIFT_THREADING_NONE way