On Tue, May 19, 2026 at 1:37 PM Branko Čibej <[email protected]> wrote:
> On 19. 5. 26 19:10, Timofei Zhakov wrote: > > On Tue, May 19, 2026 at 5:02 PM Branko Čibej <[email protected]> wrote: > >> On 19. 5. 26 15:19, Timofei Zhakov wrote: >> >> On Mon, May 18, 2026 at 11:20 PM Branko Čibej <[email protected]> wrote: >> >>> On 18. 5. 26 22:27, Ivan Zhakov wrote: >>> >>> On Mon, 18 May 2026 at 21:24, Timofei Zhakov <[email protected]> wrote: >>> >>>> On Mon, May 18, 2026 at 5:22 PM Branko Čibej <[email protected]> wrote: >>>> >>>>> On 18. 5. 26 17:04, Ivan Zhakov wrote: >>>>> >>>>> >>>> [..] >>> >>>> Yes, especially because the Unix and Win32 versions share a lot of >>>> things in common but sometimes need to do it in a slightly different way. >>>> >>>> Also what do you guys think about the fact that Win32 apr.h includes >>>> windows.h? It seems odd to me. APR promises to eliminate platform >>>> dependence by wrapping everything into POSIX-ish style API >>>> >>> >>> Cross-platform API. I'm not sure what you mean by POSIX-ish, but APR >>> isn't that. >>> >>> >> I mean that the API is similar. For example, apr_file_open() kind of >> works almost like fopen(). Not exactly but in my head I have this >> connection. Okay their signatures are very different and even naming, but I >> think there is some pattern of doing the same thing but in a slightly more >> modern way with a bit more control. >> >> >>> without extra unneeded junk and then all high level code essentially has >>>> GetLastError in the scope. I think if one wants to explicitly use win32 >>>> API, they should include those headers themselves. >>>> >>> >>> It's a good question. I agree that apr.h should not depend on system >>> headers like Windows.h. But one reason that Windows error codes are used in >>> apr_errno.h. >>> >>> >>> apr.h is a generated, *system-specific* header. Of course it can and >>> should depend on system headers if it needs them. It also includes >>> sys/types.h and sys/socket.h and sys/wait.h and os2.h and so on, depending >>> on the target. That makes perfect sense. >>> >>> >> The question I have is whether we really want it or not. >> >> In reality it's always included in every source file. Take svn_wc.h for >> example: >> >> [[[ >> #include <apr.h> >> #include <apr_pools.h> >> #include <apr_tables.h> >> #include <apr_hash.h> >> #include <apr_time.h> >> #include <apr_file_io.h> >> >> #include "svn_types.h" >> #include "svn_string.h" >> #include "svn_checksum.h" >> #include "svn_io.h" >> #include "svn_delta.h" /* for svn_stream_t */ >> #include "svn_opt.h" >> #include "svn_ra.h" /* for svn_ra_reporter_t type */ >> ]]] >> >> >> >> You'll have to be more specific: what exactly is wrong with that? apr.h >> uses symbols defined in windows.h, what else is it supposed to do? I really >> don't understand your objections. >> >> > It shouldn't use any symbols from it. That's what I'm saying. > > However, there are a few places that currently potentially do use it: > > [[[ > #if APR_HAS_UNICODE_FS > /* An arbitrary size that is digestable. True max is a bit less than 32000 > */ > #define APR_PATH_MAX 8192 > #else /* !APR_HAS_UNICODE_FS */ > #define APR_PATH_MAX MAX_PATH > #endif > ]]] > > and > > [[[ > /* Appears in later flavors, not the originals. */ > #ifndef in_addr6 > #define in6_addr in_addr6 > #endif > > #ifndef WS2TCPIP_INLINE > #define IN6_IS_ADDR_V4MAPPED(a) \ > ( (*(const apr_uint64_t *)(const void *)(&(a)->s6_addr[0]) == 0) \ > && (*(const apr_uint32_t *)(const void *)(&(a)->s6_addr[8]) == > ntohl(0x0000ffff))) > #endif > ]]] > > (MAX_PATH, ntohl, in_addr6 are the ones affected) > > I really think those should be somewhere in private headers or in the > sources. MAX_PATH is fine. But ipv6 magic - I don't like it. > > > I really don't know what to say (that isn't sarcastic or condescending). > Please take some time to research the consequences of what you're > suggesting and while you're doing that, think wider than just Windows. I'll > give you a hint: the key words are "APR" and "API". > > I guess I didn't quite reach the !(sarcastic || condescending) goal. My > apologies. > > Then we basically just have a header that nobody uses. It also brings a > lot of stuff into the autocomplete. > > > That's just about the worst argument for a wrong technical proposal that > I've ever seen. Get a better editor. > > -- Brane > At $dayjob we have a portability library of sorts and, like APR, it includes various system headers. Which ones depends on the system. Just to use this one example, MAX_PATH depends on the system so whatever program you build with the library will need to know what it is, for array sizing and other purposes. If you really want to eliminate all system headers, you could consider a "null system" which would include no system headers at all. It probably wouldn't build or run, but it might be useful for things like static analysis where you don't want the details of any particular system to influence the analysis. Any functions that must be provided by the underlying system would need to be NOPs or perhaps simulated in some way. Cheers, Nathan PS, This is a separate topic than CMake, unless there's a connection I missed?
