improve support for c++ threading and synchronization

[cfriedt]

Overview

This change allows Zephyr developers to use most or all of the C++ facilities in gcc that depend on gthread support.

Those facilities include (but are likely not limited to)

• std::thread
• std::this_thread
• std::mutex
• std::recursive_mutex
• std::shared_mutex
• std::condition_variable
• std::binary_semaphore
• std::counting_semaphore
• std::lock_guard
• std::unique_lock
• std::scoped_lock
• std::shared_lock
• std::syncstream
• std::future
• std::async

Associated Zephyr SDK PRs

• libgcc: implement gthreads in terms of C11 threads gcc#27
• libc: add option for c11 threads crosstool-ng#27
• newlib: libc: create a generic machine/_threads.h newlib-cygwin#7

Associated Zephyr PRs

• posix: sysconf: add a "small" implementation option zephyr#69882
• arch: riscv: implement frame-pointer based stack unwinding zephyr#69912
• cpp: thread: support for std::thread, std::mutex, std::condition_variable, etc zephyr#43729

Zephyr SDK Toolchain Snapshots

• aarch64-zephyr-elf
• arc64-zephyr-elf
• arc-zephyr-elf
• arm-zephyr-elf
• microblazeel-zephyr-elf
• mips-zephyr-elf
• nios2-zephyr-elf
• riscv64-zephyr-elf*
• x86_64-zephyr-elf
• xtensa*-zephyr-elf

* The riscv64-zephyr-elf toolchain stack unwinding support is not used in Zephyr. This should be fixed by zephyrproject/zephyr#69912.

Testing branch

Test are in review at zephyrproject-rtos/zephyr#43729 .

The test suites can be run with e.g.

west build -p auto -b qemu_riscv64 -t run tests/lib/cpp/std/thread \
  -- \
  -DCONFIG_PICOLIBC=n \
  -DCONFIG_NEWLIB_LIBC=y

They should run via picolibc eventually as well, but IIRC, there is an error about a missing .spec file.

Help Wanted

Updating Picolibc

The few changes required to newlib might also need to be applied to picolibc in order to expose similar functionality.

Additional Testsuites

It would be helpful to add test suites for the following

• std::semaphore

• std::future

• std::coroutine

Additional testsuites can be added in parallel. They should not be considered blockers by any means, since they do not specifically depend on anything specific to Zephyr.

Technically, not dependent on threading, coroutine support in Zephyr might Just Work (TM). However, coroutines have the potential to be extremely impactful in Zephyr as they are far more memory efficient than threads (important for resource-constrained devices).

Building, Testing, and Uploading Additional SDK Toolchains

"Zephyr SDK Toolchain Snapshots" lists a number of unchecked boxes and unlinked toolchains without links. To check the box and populate the link, the following steps should be taken:

1. Build the toolchain

sh ./sdk-ng/contrib/linux_build_toolchain.sh -o /tmp/sdk-build -c arc64-zephyr-elf

2. Back up the existing toolchain (adjust your Zephyr SDK prefix as necessary)

rsync -avr /opt/zephyr/zephyr-sdk-0.16.5/arc64-zephyr-elf{,_orig}/

3. Overwrite the old toolchain with the new

rsync -avr --delete {/tmp/sdk-build/output,/opt/zephyr/zephyr-sdk-0.16.5}/arc64-zephyr-elf/

4. Ensure the following testsuites build, run, and pass

west build -p auto -b qemu_arc/qemu_arc_hs6x -t run tests/lib/cpp/std/condition_variable/ -- -DCONFIG_NEWLIB_LIBC=y -DCONFIG_PICOLIBC=n
west build -p auto -b qemu_arc/qemu_arc_hs6x -t run tests/lib/cpp/std/mutex/ -- -DCONFIG_NEWLIB_LIBC=y -DCONFIG_PICOLIBC=n
west build -p auto -b qemu_arc/qemu_arc_hs6x -t run tests/lib/cpp/std/thread/ -- -DCONFIG_NEWLIB_LIBC=y -DCONFIG_PICOLIBC=n

5. Compress the toolchain

cd /opt/zephyr/zephyr-sdk-0.16.5
tar cpJf ~/arc64-zephyr-elf.tar.xz arc64-zephyr-elf

6. Upload file to Google Drive
7. Update Link

LLVM Toolchain Integration

Currently, all of this work is being done against Zephyr's GCC-based SDK toolchains. It's possible that LLVM toolchain integration will be much easier.

Our testsuites should run successfully with both GCC-based toolchains as well as with LLVM.

[cfriedt]

This was considerably easier with the contrib/linux_build_toolchain.sh script 👍

[dkalowsk]

Neat!

[cfriedt]

• added support for thrd_sleep() (which is ISO C11) instead of nanosleep() (which is POSIX) in gcc/libstdc++-v3

[cfriedt]

This can probably be opened for review at this point. Likely I should make PRs for submodules too.

[cfriedt]

Two additional patches needed
Defines to get the required visibility and library behaviour.

diff --git a/libgcc/gthr-c11.h b/libgcc/gthr-c11.h
index 3cadb110d96..db2cc7a46f7 100644
--- a/libgcc/gthr-c11.h
+++ b/libgcc/gthr-c11.h
@@ -45,6 +45,10 @@ see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
 #define _GTHREAD_USE_COND_INIT_FUNC 1
 #define __GTHREAD_COND_INIT_FUNCTION __gthread_cond_init_function
 
+#define _GLIBCXX_THREAD_ABI_COMPAT 1
+#define _GLIBCXX_HAS_GTHREADS 1
+#define _GLIBCXX_USE_THRD_SLEEP 1
+
 #include <threads.h>
 #include <time.h>
 

Include the header for shared_ptr, which is used in one area in this file (if the header is not included there is a compile error)

diff --git a/libstdc++-v3/include/bits/std_thread.h b/libstdc++-v3/include/bits/std_thread.h
index dd625de3bc3..413ad4c9e09 100644
--- a/libstdc++-v3/include/bits/std_thread.h
+++ b/libstdc++-v3/include/bits/std_thread.h
@@ -41,6 +41,7 @@
 #include <bits/invoke.h>       // std::__invoke
 #include <bits/refwrap.h>       // not required, but helpful to users
 #include <bits/unique_ptr.h>   // std::unique_ptr
+#include <bits/shared_ptr.h>   // std::shared_ptr
 
 #ifdef _GLIBCXX_HAS_GTHREADS
 # include <bits/gthr.h>

[dkalowsk]

haven't been able to review the submodule changes but the configs changes look good

[cfriedt]

haven't been able to review the submodule changes but the configs changes look good

@dkalowsk - I still need to make an aarch64-zephyr-elf toolchain for you.

[cfriedt]

[cfriedt]

cc @XenulsWatching

[cfriedt]

The one oddity I found with using the C11 threads API to provide the C++11 threads API was that there is no equivalent for hardware_concurrency() - i.e. how many hardware threads are available.

[stephanosio]

LGTM

DNM until the submodule refs are updated.

[stephanosio]

@cfriedt Could you please rebase and update the submodule modules to point to the merged commits?

[cfriedt]

@stephanosio - should be updated now.