add support for CMAKE_OSX_ARCHITECTURES when detecting SIZEOF variables

[JonasVautherin]

When cross-compiling for iOS, toolchains set CMAKE_OSX_ARCHITECTURES. As per the doc:

The variable “${VARIABLE}” may be “0” when CMAKE_OSX_ARCHITECTURES has multiple architectures for building OS X universal binaries. This indicates that the type size varies across architectures.

In this case, e.g. ${SIZEOF_LONG} gets set to "0", and therefore #cmakedefine SIZEOF_LONG ${SIZEOF_LONG} becomes "/* #undef SIZEOF_LONG */" in curl_config.h, resulting in a broken build.

However, and still described in the docs, when ${SIZEOF_LONG} is "0", CMake also defines other values, including ${SIZEOF_LONG_CODE}, which gives, e.g. for CMAKE_OSX_ARCHITECTURES=armv7;armv7s;arm64:

#if defined(__ARM_ARCH_7A__)
# define SIZEOF_LONG 4
#elif defined(__ARM_ARCH_7S__)
# define SIZEOF_LONG 4
#elif defined(__aarch64__)
# define SIZEOF_LONG 8
#else
# error SIZEOF_LONG unknown
#endif

Instead of using #cmakedefine and ${SIZEOF_LONG}, I believe it always works to use ${SIZEOF_LONG_CODE} directly, while enabling support for multiple architectures in CMAKE_OSX_ARCHITECTURES.

It is working for me on Linux, and on macos when cross-compiling for iOS (armv7;armv7s;arm64).

[JonasVautherin]

I don't get how my changes could reduce the coverage by 8%. And the other error seems to be a timeout in the CI.

[bagder]

@JonasVautherin no, you're right. The coveralls service acts up like this every now and then, so in this particular case we'll just ignore it. And the cirrus-ci freebsd test sometimes times out and that's another one to ignore...

[snikulov]

@JonasVautherin pardon me... But why #cmakedefine directive was removed?
Was it some syntax change in CMake?
Could you please explain how it works now according to documentation?

[snikulov]

@JonasVautherin After some investigation, I've discovered your point.
While it seems to be working, it's not clear for understanding.
No any direct references from configure_file to CheckTypeSize provided.

[jzakrzewski]

@snikulov surprised me also but CheckTypeSize documentation makes it clear now. Nevertheless, it feels pretty disconnected but I don't know if there's a way to make it easier to read.

[snikulov]

@jzakrzewski perhaps @bradking could shed some light on this.
@bradking Is it correct to use ${SIZEOF_LONG_CODE} instead of #cmakedefine?

[jzakrzewski]

Judging from the documentation I''d say it is correct. But there's this "Where the heck is that coming from?" moment. I guess for an expert it may be obvious. I haven't written (yet) enough CMake to know it from the first glance.
Maybe a reminder-comment somewhere would suffice? Because this seems quite elegant once you know why it works.

[bagder]

Maybe an added comment with the link to the documentation for it is enough?

[JonasVautherin]

I updated with a description before the SIZEOF_ section, I hope it makes it clearer. There is again a timeout in the CI (Cirrus), and I don't get why appveyor fails.

[jzakrzewski]

For me personally that'd be enough.

[snikulov]

LGTM.

[bagder]

Thanks!