thread_local is usually considered the hack to make unthreaded code littered with static variables useable from multiple thread contexts. It has overhead and reduces the compiler’s ability to optimize the code as compared to when parameters are used.
Also, until very recently, a lot compilers/platforms were unable to handle thread_local variables larger than a pointer size making it difficult to retrofit a lot of old code.
It's worth noting that `thread_local` does reduce register pressure. Unfortunately, almost no languages actually natively support the scoping that sane use of this requires.
I'm not aware of one, but it probably exists in some obscure language. I can describe how it should work though:
* ideally, thread-local variables should not be directly exposed at all
* variables intended to be dynamically scoped should be explicitly declared using the `dynamic` keyword at top level, using a thread-local variable (possibly of a different type) under the hood (contrary to languages that use dynamic scoping for all variables, this is opt-in, explicit, and more efficient. Note also that many languages do not have real thread-locals, they just pass the thread ID to a map, thus have atrocious performance.)
* reading a dynamically-scoped variable just accesses the thread-local, after unwrapping the container if needed. If the top level declaration didn't have an initializer, maybe this should throw an error?
* mutations to the dynamically-scoped variable require an explicit keyword, and involve saving the old value as if in a block-scoped local variable, and restoring it on unwind. If there's more than one mutation in a single block this can be optimized out.
* note that "assign a value wholesale" is only the most basic mutation; sometimes a stack is wanted (in which case the saved local value is just the depth) or a set (in which case the saved local value is the key to remove; this needs to be conditional)
* there should be escape hatches to manually do a save, restore, or write-without-saving. These should be used only to implement nontrivial mutations in the library or to port code using unsafe thread-locals.
It's possible to write a reasonable library implementation of this using macros in C++, GNU C, or maybe standard C23.
(an alternate approach would be to write code as if it were passing arguments to functions, but internally change them to writes to thread-local variables (only if not passed from a parameter of the same name) purely for register-pressure-optimization reasons)
Certain linker operations can be multi-threaded (not sure if this is specifically true for LLD). Particularly LTO in the GNU toolchain, but also there's been a lot of effort recently to make linking faster by actually having it use all available cores.
> Thread-local is way too magical for me. I wouldn't want to debug a system that made use of it.
There's a perfectly cromulent register just begging to be used; the circuitry has already been paid for, generating heat whether you like it or not, what magic are you afraid of here?
> Also, if you pass a param, then it can be shared.
Maybe, but if you design for sharing you'll never use your program might be bigger and slower as a result. Sometimes that matters.
> There's a perfectly cromulent register just begging to be used; [...] what magic are you afraid of here?
Most of the magic is not when using the thread-local variable, but when allocating it. When you declare a "static __thread char *p", how do you know that for instance this is located at the 123th word of the per-thread area? What if that declaration is on a dynamic library, which was loaded late (dlopen) into the process? What about threads which were started before that dynamic library was loaded, and therefore did not have enough space in their per-thread area for that thread-local variable, when they call into code which references it? What happens if the thread-local variable has an initializer?
The documentation at https://gcc.gnu.org/onlinedocs/gcc/Thread-Local.html links to a 81-page document describing four TLS access models, and that's just for Unix-style ELF; Windows platforms have their own complexities (which IIRC includes a per-process maximum of 64 or 1088 TLS slots, with slots above the first 64 being handled in a slightly different way).
When you declare a `static char *p;', how do you even know in which address of memory it is going to end up ?? How do you know what will happen if another compilation unit declares another variable of the same name? Another static library? Another dynamic library? What about initialization, what about other constructors that may read memory before main() runs? What about injected threads that are started before that? Madness, I tell you, absolute and utter madness.
> When you declare a "static __thread char p", how do you know that for instance this is located at the 123th word of the per-thread area?
Believe it or not, it's exactly the same way it knows that "static char p" is located at the 123rd word of the data section: The linker does it!
> What if that declaration is on a dynamic library, which was loaded late (dlopen) into the process?
Same thing, except the linker is dynamic.
> What about threads which were started before that dynamic library was loaded, and therefore did not have enough space in their per-thread area for that thread-local variable, when they call into code which references it? What happens if the thread-local variable has an initializer?
Seriously? I mean, you probably do have enough space because address space is huge, but dlopen should move TLS.
> links to a 81-page document describing four TLS access models, and that's just for Unix-style ELF
Just because someone can write 81-pages about something doesn't mean that it takes 81-pages to understand something.
Different people have different appetites for magic (and different definitions of what magic is).
For my first magic trick, I'd like to make something not equal to itself:
foo("bar") == foo("bar")
> false // Magic!
This is easy enough. You can make foo(..) do different things by giving it an implicit dependency on the state of the world. I'll notate implicit dependencies with [].
If you like magic, there's at least two fixes in Java land. 1: You can use some kind of mocking-framework magic to override what your actual clock does (in a way which is not visible in the non-test source code). 2: You can inject a fake clock service into your bean-wiring-framework so that you can control the clock during testing. Others seem to like these two, but they make me barf.
The way I fix it is to just make the implicit explicit instead, by turning them into parameters.
> Different people have different appetites for magic (and different definitions of what magic is).
Magic just means "I don't understand this"
And whilst I don't think it's that complicated, I also don't think you really need to understand how TLS works to use it any more than you need to understand auto mechanics to drive a car. It's faster than walking.
> I feel like thread-local takes this magic to the next level:
- TLS is indexed off of a base address, just like data.
- The thread "o" is going to be the same in each expression, so if thread[o] can live in a special-purpose register like %fs then it won't take a general-purpose register. If data can live in a special-purpose register...
Perhaps with a better picture you too will consider TLS is basically the same as non-TLS, except it's sometimes faster because, well, special circuitry.
If anything accessing via "data" is the weird one because even though it looks like you could just set data=0, you'd find almost everyone does it indirectly via %rip, and a variable "foo" is going to be %rip(-5000) in one place and %rip(-4000) in another, but would always be %fs:-26(0) if it were TLS.
I use thread_local a lot, but until recently, on Windows a delay-loaded dll with thread_local would've not worked, and the fix that is in place today is costly, okay that may not be the typical case, but it shows that support for such feature can create a lot of cost elsewhere.
Another pitfall with these is with thread-stealing concurrent schedulers - e.g. your worker thread now waits on something, and the scheduler decides to reuse the current thread for another worker - what is the meaning of thread_local there?
Another one would be coroutines (though haven't used them a lot in C/C++).
