build: fix gperftools build on OS X

[zuercher]

gperftools-2.6.1 doesn't build on OS X without adding -D_XOPEN_SOURCE=1 to enable some deprecated libc code.

[PiotrSikora]

1 is not valid _XOPEN_SOURCE value, please use 700.

[mattklein123]

Also I think you want CPPFLAGS="$CPPFLAGS -D_XOPEN_SOURCE=700"

[zuercher]

For _XOPEN_SOURCE=700 or _XOPEN_SOURCE=600, gperftools fails to compile because getpagesize is not defined in the version of _POSIX_C_SOURCE that those definitions imply.

getpagesize is defined when _XOPEN_SOURCE=500, but syscall isn't. However, defining _DARWIN_C_SOURCE brings syscall back.

[PiotrSikora]

Actually, looking into this a bit closer, <ucontext.h> is only included in the benchmark tool (file: benchmark/getcontext_light.cc), which we don't use.

The library correctly includes <sys/ucontext.h>, so it doesn't require extra magic on macOS:

#if defined(HAVE_SYS_UCONTEXT_H)
#include <sys/ucontext.h>
#elif defined(HAVE_UCONTEXT_H)
#include <ucontext.h>
#elif defined(HAVE_CYGWIN_SIGNAL_H)
#include <cygwin/signal.h>
typedef ucontext ucontext_t;
#else
typedef int ucontext_t;   // just to quiet the compiler, mostly
#endif

Perhaps we could just build the library itself? It's a bit hacky, but:

make V=1 tcmalloc_and_profiler_unittest

seems to work just fine.

[zuercher]

We use make install to get the headers into $THIRDPARTY_BUILD/include and that target depends on building everything (even though the test that uses the benchmark library is not installed).

The other alternative is probably to remove support for it in the Mac build altogether. I think that's basically disabling tcmalloc and skipping this build recipe.

[mattklein123]

I think this is fine. @PiotrSikora any particular reason to not just do this if it works?

[zuercher]

FWIW: gperftools/gperftools#910

[PiotrSikora]

@mattklein123 I'm just worried that Envoy and gperftools might pull different struct definitions, because of different _XOPEN_SOURCE and _DARWIN_C_SOURCE defines, but I'm not sure if that's a valid concern. I think it's fine to do it for now.

[mattklein123]

I really know nothing about this. It seems reasonable to me for now. @zuercher perhaps can you put in a comment in that script with a TODO which links to the gperftools bug so maybe we can clean this up at some point or let people know why this is being set and what it is working around?