P0660R10 <stop_token> And jthread #32
Comments
|
This is another issue I'd be very willing to implement! This one is definitely more involved than the other issues I've addressed and will require more involved testing as well. I'm willing to help implement this and provide the tests if this issue is open to work on? |
I caution you not to examine or refer to Nico's implementation: the licensing is unclear. There is a clear license statement for the "code examples" - Nico doesn't want people stealing his book material! - but I can't determine what license applies to any non "code examples". (It's not terribly clear what is or is not a "code example.") |
|
Would you then by chance have any recommendations, or should I just build upon the current std::thread implementation? |
I'd start by deriving |
|
Hi guys, what is the current state of this issue? @NathanSWard are you still working on this? |
|
Hi @jbrezina |
|
Update: |
|
Thanks - yeah, we need to be careful about licensing. Code written from scratch, Apache 2.0 with LLVM exception, and Boost are all ok. For anything else, we need to look into it ahead of time, to avoid accumulating work that can't be merged. |
Implements P0660R10, P1869R1, resolves microsoftGH-32, and VSO-951574. <atomic> / <memory>: Extract a new type _Locked_pointer which implements the low order bit tricks for atomic<shared_ptr> as a reusable component. yvals_core.h / *_feature_test_macros/test.cpp: Add feature test macro __cpp_lib_jthread and note that the feature is implemented. CMakeLists.txt: Add stop_token header. This will require copy_crt / setup changes in msvc. `<stop_token>` : New header. FIXME This needs a writeup after review. _Stop_callback_base forms a doubly-linked list of callbacks to invoke when stop is requested. It uses a function pointer rather than a virtual function for reduced RTTI codegen costs and to save an indirection of a vtbl access; the compiler being able to devirtualize this case is unlikely. (Using a function pointer rather than virtual was suggested by @lewissbaker ) _Stop_state is the shared state acting as the 'pipe' between stop sources and stop tokens. It has separate reference counts for sources and tokens, and all sources share 1 token count in the same way that all shared_ptrs share one weak count. The stop requested bit is merged into the sources counter to make atomically answering stop_possible possible without tricks. The counters are separate uint32_ts instead of one uint64_t to make operations manipulating them more efficient on 32 bit platforms, even though that means we might have to do a 1-2 more atomic ops on 64 bit platforms in some cases. The tricky part here is primarily how we implement the dtor effects for stop_callback. We can't use a straightforward bool in the stop_callback indicating whether the callback has been called, because once we invoke the callback the memory on which the stop_callback is stored might be gone. To resolve this, we have the thread processing request_stop publish the callback it is currently retiring, so a wait_callback destructor, upon seeing that another thread is executing the callback, can wait until the currently being executed callback is something else. `<thread>`: Extract the thread startup machinery from thread's ctor; notably, this introduces another layer of forwards in the constructing thread, but doesn't introduce any additional stack frames in the thread procedure, preserving the debugger experience. Implement jthread with a member thread and stop_source. Note that the standard currently makes the move assignment operator join on self move assign which is surprising; I asked L(E)WG if that is intentional. `<condition_variable>`: Implement the new interruptable waits; this code is effectively copied from the standard. I filed a defect report because [thread.condvarany.intwait]/7 depicts a call to a nonexistent variable `cv`.
Implements P0660R10, P1869R1, resolves microsoftGH-32, and VSO-951574. <atomic> / <memory>: Extract a new type _Locked_pointer which implements the low order bit tricks for atomic<shared_ptr> as a reusable component. yvals_core.h / *_feature_test_macros/test.cpp: Add feature test macro __cpp_lib_jthread and note that the feature is implemented. CMakeLists.txt: Add stop_token header. This will require copy_crt / setup changes in msvc. `<stop_token>` : New header. FIXME This needs a writeup after review. _Stop_callback_base forms a doubly-linked list of callbacks to invoke when stop is requested. It uses a function pointer rather than a virtual function for reduced RTTI codegen costs and to save an indirection of a vtbl access; the compiler being able to devirtualize this case is unlikely. (Using a function pointer rather than virtual was suggested by @lewissbaker ) _Stop_state is the shared state acting as the 'pipe' between stop sources and stop tokens. It has separate reference counts for sources and tokens, and all sources share 1 token count in the same way that all shared_ptrs share one weak count. The stop requested bit is merged into the sources counter to make atomically answering stop_possible possible without tricks. The counters are separate uint32_ts instead of one uint64_t to make operations manipulating them more efficient on 32 bit platforms, even though that means we might have to do a 1-2 more atomic ops on 64 bit platforms in some cases. The tricky part here is primarily how we implement the dtor effects for stop_callback. We can't use a straightforward bool in the stop_callback indicating whether the callback has been called, because once we invoke the callback the memory on which the stop_callback is stored might be gone. To resolve this, we have the thread processing request_stop publish the callback it is currently retiring, so a wait_callback destructor, upon seeing that another thread is executing the callback, can wait until the currently being executed callback is something else. `<thread>`: Extract the thread startup machinery from thread's ctor; notably, this introduces another layer of forwards in the constructing thread, but doesn't introduce any additional stack frames in the thread procedure, preserving the debugger experience. Implement jthread with a member thread and stop_source. Note that the standard currently makes the move assignment operator join on self move assign which is surprising; I asked L(E)WG if that is intentional. `<condition_variable>`: Implement the new interruptable waits; this code is effectively copied from the standard. I filed a defect report because [thread.condvarany.intwait]/7 depicts a call to a nonexistent variable `cv`.
StephanTLavavej
added
fixed
P0660R10
<stop_token>AndjthreadP1869R1 Renaming
condition_variable_any's Interruptible Wait MethodsLWG-3254 Strike
stop_token'soperator!=LWG-3362 Strike
stop_source'soperator!=The IDE's list of extensionless headers has been updated for VS 2019 16.5.
Feature-test macro as of WG21-N4842:
#define __cpp_lib_jthread 201911LThe text was updated successfully, but these errors were encountered: