On 09/11/2023 14:47, Bruno Haible wrote:
That's a good question. Yes, sometimes configure tests can be formulated
in a generic way (i.e. "on all platforms where PTHREAD_RWLOCK_INITIALIZER
is not defined") or in a specific way (i.e. "on Mac OS X < 10.5") or in
an intermediate way (i.e. "on all macOS version which lack
PTHREAD_RWLOCK_INITIALIZER").
When autoconf was created, in the 1990ies, new OS versions were frequently
forked off from existing OS versions. So, if a bug was detected, say, in
Solaris, there were good chances that the same bug also existed in SINIX
or perhaps IRIX. And if a bug existed in SunOS 4, there were chances that
it existed also in Ultrix. In these times, configure tests should better
be generic.
Nowadays, the set of OSes has stabilized. Let's look, for example, at the
bugs in the <fenv.h> functions that I discovered in the last three weeks:
- 22 bugs are specific to a single platform.
- 11 bugs exist on two or more platforms.
Also, remember that a generic test
- takes longer to write than a specific test,
- may fail on other platforms for unrelated reasons.
Therefore, when 2/3 of all bugs are specific to a single platform, it
often is overkill to use a generic test. It just depends on the nature
of the bug.
Here, in this case, I chose to put in a workaround only for macOS,
because I hope that if the problem (PTHREAD_RWLOCK_INITIALIZER being
undefined) occurs on other platforms, it has better workarounds:
defining some platform-specific macro, using PTHREAD_RWLOCK_INITIALIZER_NP,
or some other obvious initializer).
Understood, thank you.
Sevan
