[dmd-cxx] Fix Issue 19481: mutex.d(95) Error: pthread_mutex_init failed

[ibuclaw]

Committing the backport before it's in master... should give you a couple months to fix failing mecca in #2411 before gdc switches to ddmd/master when stage1 opens for gcc-10.

[dlang-bot]

Thanks for your pull request, @ibuclaw!

Bugzilla references

Auto-close	Bugzilla	Severity	Description
✓	19481	major	Aborting from local/libphobos/libdruntime/core/sync/mutex.d(95) Error: pthread_mutex_init failed.

Testing this PR locally

If you don't have a local development environment setup, you can use Digger to test this PR:

dub fetch digger
dub run digger -- build "dmd-cxx + druntime#2534"

[rainers]

Maybe mutexAlign could be a property of Mutex itself so explicit alignment could be used by overwriting mutexAlign with a constant.

[ibuclaw]

There's nothing wrong with the alignment of Mutex. If you're lucky, it's the size of N in void[N][2] that's the issue, as only the first index is suitably aligned. (You're unlucky if the alignment of pointers is smaller than the alignment of the OS mutex type).

There is no one constant alignment either, what type of and what alignment of the mutex is varies between platforms.

[ibuclaw]

Committing, if an amendment is needed before May, then so be it.

[rainers]

There's nothing wrong with the alignment of Mutex.

Mutex.alignof might be unusable here because it is the alignment of the class reference (the unlucky case):

class C
{
	double x;
}

struct S
{
	align(16) int x;
}

// for -m32
pragma(msg, Object.alignof); // 4
pragma(msg, C.alignof);      // 4
pragma(msg, C.x.alignof);    // 8
pragma(msg, S.alignof);      // 16
pragma(msg, S.x.alignof);    // 4

Agreed, the aligned size of the object is another issue.

There is no one constant alignment either, what type of and what alignment of the mutex is varies between platforms.

I meant something like

version(mySpecialOS)
    enum mutexAlign = 16; // due to explicit alignment on field of mutex
else
    enum mutexAlign = classInstanceAlignment!Mutex;

so this declaration doesn't have to be repeated everytime something similar is done.

But probably the better solution would be for alignof(C.x) to return the explicit alignment for symbols, not the alignment of the type.

[ibuclaw]

Mutex.alignof might be unusable here because it is the alignment of the class reference (the unlucky case):

Yes, you can't use Mutex.alignof, which why it is being changed from. The underlying record is suitably aligned, hence I re-iterate there is no problem with Mutex (i.e: when used as new Mutex()).

I meant something like

I don't see a reason for that, as you almost certainly mean OS_mutex_type.alignof (Currently: CRITICAL_SECTION on Windows, pthread_mutex_t on Posix, etc...) rather than a hard-coded number.

I had considered something similar to core.sync.semaphore.

public:
    /// Aliases the operating-system-specific mutex type.
    version (Windows)        alias Handle = CRITICAL_SECTION;
    /// ditto
    else version (Posix)     alias Handle = pthread_mutex_t;

private:
    /// Handle to the system-specific mutex.
    Handle m_hndl;

But ultimately went for this instead over align(Mutex.Handle.alignof) xxx, as I'd rather not assume that the OS mutex has the greater alignment requirement of the class.