~jscott/which

Remove unnecessary include of <limits.h>

Thanks to deheader for catching this!
Initialize pathname_found.

clang-tidy pointed out that if the addition overflow
checks were to turn out true, the loop could stop and
reading pathname_found would be undefined.
Add an assertion that a loop runs at least once.

This should demonstrate that pathname_found isn't
used uninitialized.
Suppress a couple GCC warnings.
Use open_memstream() to dynamically allocate the path list.

Trying to count the number of paths in the string beforehand
was error-prone, as is trying to reallocate the memory ourselves.
Make the error handling more robust.

In particular, update the value of threads_joined so
we don't accidentally join with the same thread twice.
Use faccessat() to determine executable permissions.

Be portable to POSIX-conforming systems. These changes
too were drawn from the Code Review Stack Exchange.
Fix attempts to join with a thread multiple times.
Stop using thread cancellation.

It was pointed out to me at the Code Review Stack Exchange
at https://codereview.stackexchange.com/q/274956 that
none of the child threads make calls to blocking functions,
so we may as well not use thread cancellation. This simplifies
things substantially.
Add the restrict qualifier to our comparison functions.

Since the pointers point to const-qualified objects,
here the restrict qualifier merely indicates that
the pointed-to objects are not modified.
Don't deadlock waiting for children to finish if there are none.
Suppress Cppcheck warning about our usage of strtok().
Fix the build time XSI check.
Add some tests and assertions to the build system.
Add a missing call to free().

I found this by noticing that a call to free() from
the "goto ladder" wasn't included in our resource cleanup
on success.
Handle pthread_create() returning EAGAIN more gracefully.

Previously, we'd enter a busy loop with calling sched_yield()
and trying to create the thread again. This is suboptimal: a
far cleaner approach which is less likely to cause failed
thread creation attempts is to wait for a thread to terminate.

The best way to do this is with a condition variable which
gets raised when a thread terminates, and so that's what we
do here.
Install which using Meson.
Bump the copyright year in meson.build.

This is for the previous commit.
Remove the global sequential thread ID.

Using a global variable for a thread-specific parameter
was hideous, as was the semaphore necessary to guard it.
Although potentially less portable in principle, we now
pass a thread's sequential identifier (which serves as an
index into various arrays) as the sole argument into
each of our threads via type punning.

This is cleaner and requires fewer resources. I don't
know of any system on which this won't work, but a
check has been added in the build system to ensure
that convering the seq. thread ID to a void* is valid.

The way errors are handled in each of the threads was
tweaked because apparently _XOPEN_SOURCE (but not
_POSIX_C_SOURCE) causes the GNU C Library to expose
NL_TEXTMAX, and as of this writing it has the value
INT_MAX which would cause a stack overflow.
Next