/* vms_compat.h - OpenVMS 8.4 Alpha (DEC C) compatibility shims. * * Pulled in by coab.h, before anything else, when the compiler predefines * __VMS. Nothing includes this file directly and nothing outside it tests * __VMS except the four places listed at the bottom, so the whole platform * divergence is visible from here. * * Modelled on the sibling OpenVMS ports' shims - [-.ROTT.ROTT]VMS_COMPAT.H and * [-.CHOCOLATE_DUKE.ENGINE.SRC]VMS_COMPAT.H - but much smaller, because this * tree was written as portable C99 rather than reverse-engineered 1994 DOS C: * there is no Watcom inline asm, no DPMI, no __int64, no far pointers, no * strcmpi/stricmp/filelength, and no 16-bit segment arithmetic to emulate. * * See PORTING-VMS.md for the whole story; the comments here cover only what a * reader of the source needs. */ #ifndef COAB_VMS_COMPAT_H #define COAB_VMS_COMPAT_H #ifdef __VMS /* -------------------------------------------------------------------------- * GCC function attributes. * * Seven declarations carry __attribute__((format(printf,N,M))) or * ((noreturn)): log.h's log_write and log_fatal, quit.h's game_print_and_exit * and game_print_and_exit_msg, ecl.h's ecl_vm_log, and savegame.c's two static * record loggers. These are GCC diagnostics with no DEC C equivalent, and DEC * C treats the unknown keyword as a syntax error at the declaration - i.e. the * build stops on the first file that includes log.h, which is all of them. * * Defining the keyword away is the standard shim and costs only printf-argument * checking, which the Linux and macOS builds already do (-Wall -Wextra, warning * free). Doing it here rather than editing seven declarations keeps the * checking on the platforms that can perform it. * * Note the double parentheses at the use sites: __attribute__((a,b)) passes * "a,b" as ONE macro argument because the comma sits inside the inner parens, * so a one-parameter macro is right even for ((format(printf,1,2),noreturn)). * * Losing ((noreturn)) on game_print_and_exit is harmless - it is a warning and * optimisation hint, and the function really does longjmp out - but it is why * MISSINGRETURN is in DESCRIP.MMS's disabled-diagnostics list: DEC C can now * see a fallthrough tail where GCC could not. * -------------------------------------------------------------------------- */ #ifndef __attribute__ #define __attribute__(x) #endif /* -------------------------------------------------------------------------- * The unprefixed VMS macro, which SDL 1.2's headers require. * * SDL's begin_code.h does * * #ifndef DECLSPEC * #ifdef VMS * #define DECLSPEC * #endif * #endif * * and has NO #else. So if VMS is not defined, DECLSPEC is never defined at * all, and every prototype in every SDL header - "extern DECLSPEC int SDLCALL * SDL_Init(Uint32)" - becomes a declaration beginning with an unknown * identifier. That is a syntax error on the first line of SDL_error.h and the * diagnostics that follow name DECLSPEC, not VMS, so it is not an obvious * failure to read. * * DEC C predefines the unprefixed VMS only in its relaxed modes; a strict * /STANDARD suppresses every unprefixed predefined macro (VMS, vax, alpha, * unix) and keeps only the __-prefixed ones, of which SDL knows nothing. * Defining it here rather than in DESCRIP.MMS keeps the two in step whatever * $(CSTD) turns out to be, and only ever adds a macro DEC C would have added * itself. * * Reached in time because platform_sdl1.c - the one file that includes * - includes "platform.h" first, and that includes coab.h, and that includes * this. Do not reorder those. * -------------------------------------------------------------------------- */ #ifndef VMS #define VMS 1 #endif /* -------------------------------------------------------------------------- * strcasecmp lives in on VMS, not . * * Three call sites - vfs.c's vfs_save_resolve and save_name_cmp, and * savegame.c's .sav suffix test. All three include only, which is * correct on Linux and macOS (glibc and libc++ both declare strcasecmp there * via _DEFAULT_SOURCE) and leaves it implicitly declared on VMS. An implicit * declaration returning int happens to work here, but %CC-I-IMPLICITFUNC on * three files in every build is noise, and DEC C is entitled to reject it. * * Included from here rather than added to those two files so that the include * lists stay as the reference build has them. * -------------------------------------------------------------------------- */ #include /* -------------------------------------------------------------------------- * Binary file access - THE ONE THING THAT SILENTLY CORRUPTS DATA ON VMS. * * "rb"/"wb" is accepted and does nothing: there is no text/binary distinction * in the CRTL, and what actually decides whether a stream of bytes survives is * the RMS record layer. A file opened without "ctx=stm" is read and written a * RECORD at a time, so record boundaries are inserted, stripped or padded and * the byte offsets that DAX block reading and savegame record writing depend on * no longer line up. It cannot be folded into the mode string - it is a * variadic argument to fopen - which is why this is a function and not a macro. * * Every binary open in the tree therefore goes through vfs_fopen() (vfs.c): * the DAX reads in vfs_read_file, the savegame and character-roster opens in * fileio.c's file_assign, the self-test's PPM dumps in display.c and its * synthetic save in selftest.c. log.c's fopen is deliberately NOT one of them * - the log is text, and leaving it in the default variable-record format is * what makes TYPE and EDIT work on it. * * The two macros below are what vfs_fopen expands to; they are here so that * "SEARCH [.SRC]*.* VMS_FOPEN" finds the moving parts from either end. * * VMS_FOPEN_CREATE additionally asks for stream-LF, the format every other * byte-stream file on a VMS system uses (it is what UNZIP and GZIP produce). * Without it a newly created file gets the CRTL's default variable-length * records, and although "ctx=stm" would still read it back correctly on this * system, nothing else on the machine could - and a savegame copied from a DOS * or Windows install, which is stream-LF, would not match the files this port * writes. RMS ignores rfm on an open of an EXISTING file, so it is only ever * applied at creation. * -------------------------------------------------------------------------- */ #define VMS_FOPEN_READ(path, mode) fopen((path), (mode), "ctx=stm") #define VMS_FOPEN_CREATE(path, mode) fopen((path), (mode), "ctx=stm", \ "rfm=stmlf") /* -------------------------------------------------------------------------- * What is NOT here, and why. * * No shim. Nothing in this tree calls backtrace(); the sibling * ROTT and Duke ports needed one because their DOS-era error handlers dumped * a stack. * * No snprintf fallback. All ~700 uses assume the C99 function. The CRTL on * OpenVMS 8.4 has it; CONFIGURE.COM probes for it anyway and says so, because * older CRTLs do not and the failure is otherwise a mystery. See * PORTING-VMS.md for the fallback to add if you meet one. * * No clock_gettime fallback. Two call sites, rnd.c's seeding and text.c's * time01, both asking only for CLOCK_REALTIME and neither on a hot path. * OpenVMS 8.4's CRTL has it. If yours does not, PORTING-VMS.md has the * gettimeofday() shim to paste in right here - it is four lines - and * CONFIGURE.COM tells you whether you need it. * * No / substitutes HERE - they are in [.VMSINC] instead, * which is a separate directory rather than more typedefs in this file for one * concrete reason: [.SRC]RND.C's first line is "#include ", before * "rnd.h" and so before this file is in scope at all. Nothing this header * defines can reach it; only the include path can. CONFIGURE.COM adds * [.VMSINC] to /INCLUDE only if the real headers are missing, which they are on * any DEC C before V7. See PORTING-VMS.md and the header comments in the two * files - the trap described there is a real bug that was hit, not a * hypothetical. * * No 64-bit pointer qualifier. DESCRIP.MMS leaves DEC C's Alpha default of * 32-bit pointers, as every other port in this tree does. This one is a * consistency choice rather than a correctness one: no pointer here is ever * stored in an integer (checked - the engine's u32 fields hold DAX offsets, * never addresses), so /POINTER_SIZE=64 would also work. * * The other four places that test __VMS, all in vfs.c and all about RMS or * ODS-2 rather than about the compiler: * * vfs_fopen the "ctx=stm" opens above. * dir_spec stat() and opendir() need "dir/" or "[dir]" for a * DIRECTORY; a bare "dir" or "./dir" names a FILE of that * name. Called by path_is_dir, index_dir and the two * savegame scans. * strip_version readdir() reports "NAME.EXT;1", version included. Called * by index_dir and the two savegame scans. * user_dir the save and log directories. $HOME/.local/share is * neither a legal ODS-2 name nor a path the CRTL can * translate, so both land in the current default directory, * exactly as the sibling ROTT port does it. * -------------------------------------------------------------------------- */ #endif /* __VMS */ #endif /* COAB_VMS_COMPAT_H */