Replace the bswap function used in kplib with B_SWAP32, and merge the 32-bit MSVC and GCC assembly into compat.h.

TODO: Review the value of the continued use of special cases of these swap functions, including the OS-based variants for BSD and OS X, and this assembly.

git-svn-id: https://svn.eduke32.com/eduke32@5173 1a8010ca-5511-0410-912e-c29ae57300e0
This commit is contained in:
hendricks266 2015-05-03 07:03:12 +00:00
parent f141cc71f2
commit 5c4622a8ab
2 changed files with 28 additions and 47 deletions

View file

@ -248,12 +248,12 @@
# define B_LITTLE_ENDIAN 0
# define B_BIG_ENDIAN 1
# endif
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#elif defined(GEKKO) || defined(__ANDROID__)
# define B_LITTLE_ENDIAN 0
# define B_BIG_ENDIAN 1
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#elif defined(__OpenBSD__)
# include <machine/endian.h>
@ -316,7 +316,7 @@
# define B_LITTLE_ENDIAN 0
# define B_BIG_ENDIAN 1
# endif
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#elif defined(__QNX__)
# if defined __LITTLEENDIAN__
@ -326,7 +326,7 @@
# define B_LITTLE_ENDIAN 0
# define B_BIG_ENDIAN 1
# endif
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#elif defined(__sun)
# if defined _LITTLE_ENDIAN
@ -336,12 +336,12 @@
# define B_LITTLE_ENDIAN 0
# define B_BIG_ENDIAN 1
# endif
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#elif defined(_WIN32) || defined(SKYOS) || defined(__SYLLABLE__)
# define B_LITTLE_ENDIAN 1
# define B_BIG_ENDIAN 0
# define B_ENDIAN_C_INLINE 1
# define B_USE_COMPAT_SWAP 1
#endif
#if !defined(B_LITTLE_ENDIAN) || !defined(B_BIG_ENDIAN)
@ -365,18 +365,30 @@ defined __x86_64__ || defined __amd64__ || defined _M_X64 || defined _M_IA64 ||
extern "C" {
#endif
#if defined B_ENDIAN_X86_INLINE
# if defined(_MSC_VER)
// inline asm using bswap/xchg
# elif defined(__GNUC__)
// inline asm using bswap/xchg
# endif
#elif defined B_ENDIAN_C_INLINE
#if defined B_USE_COMPAT_SWAP
FORCE_INLINE uint16_t B_SWAP16(uint16_t s) { return (s >> 8) | (s << 8); }
# if !defined NOASM && defined __i386__ && defined _MSC_VER
FORCE_INLINE uint32_t B_SWAP32(uint32_t a)
{
_asm
{
mov eax, a
bswap eax
}
}
# elif !defined NOASM && defined __i386__ && defined __GNUC__
FORCE_INLINE uint32_t B_SWAP32(uint32_t a)
{
__asm__ __volatile__("bswap %0" : "+r"(a) : : "cc");
return a;
}
# else
FORCE_INLINE uint32_t B_SWAP32(uint32_t l)
{
return ((l >> 8) & 0xff00) | ((l & 0xff00) << 8) | (l << 24) | (l >> 24);
}
# endif
FORCE_INLINE uint64_t B_SWAP64(uint64_t l)
{
return (l >> 56) | ((l >> 40) & 0xff00) | ((l >> 24) & 0xff0000) | ((l >> 8) & 0xff000000) |

View file

@ -72,11 +72,6 @@ static __inline int32_t _lrotl(int32_t i, int sh)
#define min(a,b) (((a) < (b)) ? (a) : (b))
#endif
#if defined __clang__ && __clang_major__==3 && __clang_minor__==1
// clang 3.1 SVN r149129, assertion failure with inline asm
# define NOASM 1
#endif
#if defined(__GNUC__)
#undef _inline
#define _inline inline
@ -177,17 +172,6 @@ static uint8_t qhufbit0[1<<LOGQHUFSIZ0], qhufbit1[1<<LOGQHUFSIZ1];
#if defined(_MSC_VER) && !defined(NOASM)
#if B_BIG_ENDIAN == 1
static _inline uint32_t bswap(uint32_t a)
{
_asm
{
mov eax, a
bswap eax
}
}
#endif
static _inline int32_t bitrev(int32_t b, int32_t c)
{
_asm
@ -204,14 +188,6 @@ static _inline int32_t bitrev(int32_t b, int32_t c)
#elif defined(__GNUC__) && defined(__i386__) && !defined(NOASM)
#if B_BIG_ENDIAN == 1
static inline uint32_t bswap(uint32_t a)
{
__asm__ __volatile__("bswap %0" : "+r"(a) : : "cc");
return a;
}
#endif
static inline int32_t bitrev(int32_t b, int32_t c)
{
int32_t a = 0;
@ -223,13 +199,6 @@ static inline int32_t bitrev(int32_t b, int32_t c)
#else
#if B_BIG_ENDIAN == 1
static inline uint32_t bswap(uint32_t a)
{
return(((a&0xff0000)>>8) + ((a&0xff00)<<8) + (a<<24) + (a>>24));
}
#endif
static inline int32_t bitrev(int32_t b, int32_t c)
{
int32_t i, j;
@ -1120,9 +1089,9 @@ static void initkpeg()
#if B_BIG_ENDIAN == 1
for (i=0; i<1024; i++)
{
colclip[i] = bswap(colclip[i]);
colclipup8[i] = bswap(colclipup8[i]);
colclipup16[i] = bswap(colclipup16[i]);
colclip[i] = B_SWAP32(colclip[i]);
colclipup8[i] = B_SWAP32(colclipup8[i]);
colclipup16[i] = B_SWAP32(colclipup16[i]);
}
#endif