Dustin J. Mitchell wrote: > As you probably know, on HP/UX, accept, getpeername, et al. take (.., > int *addrlen), not (.., socklen_t *addrlen) unless _XOPEN_SOURCE and > _XOPEN_SOURCE_EXTENDED are defined. However, if these macros are not > defined, socklen_t is still defined.
There are a few more problems too. This is from memory, I can check later, but as I recall 64-bit HP-UX defines socklen_t incorrectly to be long. IIRC it was long in both 32-bit and 64-bit environments. It should have been int. This led to failures in the 64-bit compile. > This means that the test in socklen.m4 doesn't have any effect (as it > finds the type already defined). The result is code that compiles, > but because of the endianness of the 64-bit processor, fails at > runtime. As I recall it was a size mismatch and not an endian mismatch. Because the size was mismatched the data was in the half that wasn't being used by the subroutine. IIRC data was stuffed into a 64-bit variable but then used as a 32-bit variable internally. The portion used was the most significant portion which was all zero. (It has been a few years and I might have details wrong.) However I think we are in agreement that it didn't work. :-} > I think it would make sense to run the compile check even if socklen_t > is defined. What I don't know is what to do if it turns out that > socklen_t is the wrong type to pass to these functions? "#undef > socklen_t\n#define socklen_t int"? Or try to use a type with a > different name, e.g., socklen_t_equiv, throughout the source code? I have a snippet of code that does just that, or rather exactly this: #if defined(__hpux) #undef socklen_t #define socklen_t int #endif With a FIXME comment that I never got back to about the problem being 64-bit compiles using the native HP ANSI C/C++ compiler in 64-bit mode defining socklen_t to a long so that sizeof (socklen_t) != sizeof (int). > I'm trying the latter solution within Amanda, since I can replace all > uses of socklen_t without much trouble. Is there a way to solve this > within gnulib, though? That would be nice. Bob
