On 2025-04-06 Pali Rohár wrote: > On Saturday 29 March 2025 18:56:56 Lasse Collin wrote: > > On 2025-03-29 Pali Rohár wrote: > > > Problem is that you cannot define _USE_32BIT_TIME_T on UCRT > > > builds. There are header ifdef checks that combination of > > > _USE_32BIT_TIME_T and _UCRT throw preprocessor #error. But UCRT > > > has _stat32 and also _stat32i64 functions. So this would not > > > work. > > > > I might not fully understand how much this matters. > > > > - Currently one doesn't need to build any stat*.c files for UCRT. > > The .def file redirects statxxx to _statxxx. > > > > - The way I see it, having the fstat/stat/wstat symbols matters > > mostly for Autoconf checks. Normal applications rely on > > <sys/stat.h>, and then the default symbol doesn't matter. > > > > <ftw.h> uses "struct stat", not any underscore version. When apps > > cannot get 32-bit time_t variants of "struct stat" from > > <sys/stat.h>, > > Application can get it when building in msvcrt mode. Application > cannot get it when building in UCRT mode. Both clang and gcc allows > to chose build mode at application build time (not just at mingw-w64 > build time). gcc has -mcrtdll= switch and clang has -lmsvc* switches.
This is the thing I had missed. Thanks! > > With MSVCRT, stat64 comes from stat64.c, and if that is built with > > some -fsanitize, then the type mismatch between <sys/stat.h> and > > stat64.c would be detected. The problem would be stat() when it's > > redirected to stat64() because stat64.c expects "struct stat64 *" > > while stat() uses "struct stat *". > > When I'm looking at the ftw() implementation, it has already exactly > same problem. It is taking an fcb function parameter of type: > > int (*fcb) (const char *, const STRUCT_STAT *, int) > > But it calling it as a different signature: > > (*ctx->fcb) (buf, st, flag, ftw); > > It contains additional argument. > > So I think that the ftw code to work with clang already needs to > disable CFI checks. Yes. I commented this shortly on 2025-03-29. The callback part is easy to fix, for example, by having two function pointers in the ctx_t structure. But since it's not causing real-world problems (other than the -fsanitize usage), this doesn't need to be modified now. > Hence I would propose to fix the broken stat functions to work without > CFI and later if somebody comes with a suitable solution we can fix it > also for CFI-enabled compilers. I agree. The patches are a major improvement already. :-) If you like, feel free to include my stat patches 1-8 from 2025-03-22 when you submit yours. Alternatively, I will send mine separately after yours have been merged. Do whichever you prefer. Thanks! -- Lasse Collin _______________________________________________ Mingw-w64-public mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
