mirror of
https://github.com/nzp-team/fteqw.git
synced 2024-11-25 05:11:35 +00:00
221d4032af
git-svn-id: https://svn.code.sf.net/p/fteqw/code/trunk@6001 fc73d0e0-1445-4013-8a0c-d673dee63da5
71 lines
1.9 KiB
C
71 lines
1.9 KiB
C
#ifndef QCLIB_PROGTYPE_H
|
|
#define QCLIB_PROGTYPE_H
|
|
|
|
#if _MSC_VER >= 1300
|
|
#define QC_ALIGN(a) __declspec(align(a))
|
|
#elif (__GNUC__ >= 3) || defined(__clang__)
|
|
#define QC_ALIGN(a) __attribute__((aligned(a)))
|
|
#else
|
|
#define QC_ALIGN(a) //I hope misaligned accesses are okay...
|
|
#endif
|
|
|
|
#if 0
|
|
//64bit primitives allows for:
|
|
// greater precision timers (so maps can last longer without getting restarted)
|
|
// planet-sized maps (with the engine's vec_t types changed too, and with some sort of magic for the gpu's precision).
|
|
//TODO: for this to work, someone'll have to go through the code to somehow deal with the vec_t/pvec_t/float differences.
|
|
#warning FTE isnt ready for this.
|
|
#include <stdint.h>
|
|
typedef double pvec_t;
|
|
typedef int64_t pint_t;
|
|
typedef uint64_t puint_t;
|
|
|
|
#include <inttypes.h>
|
|
#define pPRId PRId64
|
|
#define pPRIi PRIi64
|
|
#define pPRIu PRIu64
|
|
#define pPRIx PRIx64
|
|
#define QCVM_64
|
|
#else
|
|
//use 32bit types, for sanity.
|
|
typedef float pvec_t;
|
|
typedef int pint_t;
|
|
typedef unsigned int puint_t;
|
|
#ifdef _MSC_VER
|
|
typedef QC_ALIGN(4) __int64 pint64_t;
|
|
typedef QC_ALIGN(4) unsigned __int64 puint64_t;
|
|
|
|
#define pPRId "d"
|
|
#define pPRIi "i"
|
|
#define pPRIu "u"
|
|
#define pPRIx "x"
|
|
#define pPRIi64 "I64i"
|
|
#define pPRIu64 "I64u"
|
|
#define pPRIx64 "I64x"
|
|
#else
|
|
#include <inttypes.h>
|
|
typedef int64_t pint64_t QC_ALIGN(4);
|
|
typedef uint64_t puint64_t QC_ALIGN(4);
|
|
|
|
#define pPRId PRId32
|
|
#define pPRIi PRIi32
|
|
#define pPRIu PRIu32
|
|
#define pPRIx PRIx32
|
|
#define pPRIi64 PRIi64
|
|
#define pPRIu64 PRIu64
|
|
#define pPRIx64 PRIx64
|
|
#endif
|
|
#define QCVM_32
|
|
#endif
|
|
|
|
typedef QC_ALIGN(4) double pdouble_t; //the qcvm uses vectors and stuff, so any 64bit types are only 4-byte aligned. we don't do atomics so this is fine so long as the compiler handles it for us.
|
|
typedef unsigned int pbool;
|
|
typedef pvec_t pvec3_t[3];
|
|
typedef pint_t progsnum_t;
|
|
typedef puint_t func_t;
|
|
typedef puint_t string_t;
|
|
|
|
extern pvec3_t pvec3_origin;
|
|
|
|
#endif /* QCLIB_PROGTYPE_H */
|
|
|