X-Git-Url: https://git.distorted.org.uk/~mdw/mLib/blobdiff_plain/18c831dcd0ae4d660c70ccac69d27ed2a97851be..4d4fb2913be1bb83a02e64fbb4dd0b4a6355354a:/utils/bits.h

diff --git a/utils/bits.h b/utils/bits.h
index f3fe08f..656eee5 100644
--- a/utils/bits.h
+++ b/utils/bits.h
@@ -40,6 +40,10 @@
 #  include <stdint.h>
 #endif
 
+#ifndef MLIB_COMPILER_H
+#  include "compiler.h"
+#endif
+
 /*----- Decide on some types ----------------------------------------------*/
 
 /* --- Make GNU C shut up --- */
@@ -314,117 +318,283 @@ typedef unsigned char octet, uint8;
    } while (0)
 #endif
 
+/* --- Endianness swapping --- */
+
+#if GCC_VERSION_P(4, 8)
+#  define ENDSWAP16(x) ((uint16)__builtin_bswap16(x))
+#endif
+#if GCC_VERSION_P(4, 3)
+#  define ENDSWAP32(x) ((uint32)__builtin_bswap32(x))
+#endif
+#if GCC_VERSION_P(4, 3) && defined(HAVE_UINT64)
+#  define ENDSWAP64(x) ((uint64)__builtin_bswap64(x))
+#endif
+
+#ifndef ENDSWAP8
+#  define ENDSWAP8(x) U8(x)
+#endif
+#ifndef ENDSWAP16
+#  define ENDSWAP16(x)							\
+	((((uint16)(x) >> 8)&0xff) |					\
+	 (((uint16)(x)&0xff) << 8))
+#endif
+#ifndef ENDSWAP24
+#  define ENDSWAP24(x)							\
+	((((uint24)(x) >> 16)&0xff) |					\
+	 ((uint24)(x)&0xff00) |						\
+	 ((uint24)((x)&0xff) << 16))
+#endif
+#ifndef ENDSWAP32
+#  define ENDSWAP32(x)							\
+	(ENDSWAP16(((uint32)(x) >> 16)&0xffff) |			\
+	 ((uint32)ENDSWAP16((x)&0xffff) << 16))
+#endif
+#if defined(HAVE_UINT64) && !defined(ENDSWAP64)
+#  define ENDSWAP64(x)							\
+	(ENDSWAP32(((uint64)(x) >> 32)&0xffffffff) |			\
+	 ((uint64)ENDSWAP32((x)&0xffffffff) << 32))
+#endif
+#ifdef HAVE_UINT64
+#  define ENDSWAP64_(z, x)						\
+	((z).i = ENDSWAP64((x).i))
+#else
+#  define ENDSWAP64_(z, x)						\
+	((z).lo = ENDSWAP32((x).hi),					\
+	 (z).hi = ENDSWAP32((x).lo))
+#endif
+
+#define MLIB_LITTLE_ENDIAN 1234
+#define MLIB_BIG_ENDIAN 4321
+#if defined(__ORDER_LITTLE_ENDIAN__) &&					\
+    __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#  define MLIB_BYTE_ORDER MLIB_LITTLE_ENDIAN
+#elif defined(__ORDER_BIG_ENDIAN__) &&					\
+      __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+#  define MLIB_BYTE_ORDER MLIB_BIG_ENDIAN
+#endif
+
+#if MLIB_BYTE_ORDER == MLIB_LITTLE_ENDIAN
+#  define HTOL16(x) (x)
+#  define LTOH16(x) (x)
+#  define HTOB16(x) ENDSWAP16(x)
+#  define BTOH16(x) ENDSWAP16(x)
+#  define HTOL24(x) (x)
+#  define LTOH24(x) (x)
+#  define HTOB24(x) ENDSWAP24(x)
+#  define BTOH24(x) ENDSWAP24(x)
+#  define HTOL32(x) (x)
+#  define LTOH32(x) (x)
+#  define HTOB32(x) ENDSWAP32(x)
+#  define BTOH32(x) ENDSWAP32(x)
+#  ifdef HAVE_UINT64
+#    define HTOL64(x) (x)
+#    define LTOH64(x) (x)
+#    define HTOB64(x) ENDSWAP64(x)
+#    define BTOH64(x) ENDSWAP64(x)
+#  endif
+#  define HTOL64_(z, x) ASSIGN64(z, x)
+#  define LTOH64_(z, x) ASSIGN64(z, x)
+#  define HTOB64_(z, x) ENDSWAP64_(z, x)
+#  define BTOH64_(z, x) ENDSWAP64_(z, x)
+#elif MLIB_BYTE_ORDER == MLIB_BIG_ENDIAN
+#  define HTOL16(x) ENDSWAP16(x)
+#  define LTOH16(x) ENDSWAP16(x)
+#  define HTOB16(x) (x)
+#  define BTOH16(x) (x)
+#  define HTOL24(x) ENDSWAP24(x)
+#  define LTOH24(x) ENDSWAP24(x)
+#  define HTOB24(x) (x)
+#  define BTOH24(x) (x)
+#  define HTOL32(x) ENDSWAP32(x)
+#  define LTOH32(x) ENDSWAP32(x)
+#  define HTOB32(x) (x)
+#  define BTOH32(x) (x)
+#  ifdef HAVE_UINT64
+#    define HTOL64(x) ENDSWAP64(x)
+#    define LTOH64(x) ENDSWAP64(x)
+#    define HTOB64(x) (x)
+#    define BTOH64(x) (x)
+#    define HTOL64_(z, x) ENDSWAP64_(z, x)
+#    define LTOH64_(z, x) ENDSWAP64_(z, x)
+#    define HTOB64_(z, x) ((z).i = (x).i)
+#    define BTOH64_(z, x) ((z).i = (x).i)
+#  endif
+#  define HTOL64_(z, x) ENDSWAP64_(z, x)
+#  define LTOH64_(z, x) ENDSWAP64_(z, x)
+#  define HTOB64_(z, x) ASSIGN64(z, x)
+#  define BTOH64_(z, x) ASSIGN64(z, x)
+#endif
+
+/* --- Unaligned access (GCC-specific) --- */
+
+#if GCC_VERSION_P(3, 3) && CHAR_BIT == 8
+#  define MLIB_MISALIGNED __attribute__((aligned(1), may_alias))
+#  if __SIZEOF_SHORT__ == 2
+     typedef MLIB_MISALIGNED unsigned short misaligned_uint16;
+#    define RAW16(p) (*(misaligned_uint16 *)(p))
+#  endif
+#  if __SIZEOF_INT__ == 4
+     typedef MLIB_MISALIGNED unsigned int misaligned_uint32;
+#    define RAW32(p) (*(misaligned_uint32 *)(p))
+#  elif __SIZEOF_LONG__ == 4
+     typedef MLIB_MISALIGNED unsigned long misaligned_uint32;
+#    define RAW32(p) (*(misaligned_uint32 *)(p))
+#  endif
+#  if __SIZEOF_LONG__ == 8
+     typedef MLIB_MISALIGNED unsigned long misaligned_uint64;
+#    define RAW64(p) (*(misaligned_uint64 *)(p))
+#  elif __SIZEOF_LONG_LONG__ == 8
+     typedef MLIB_MISALIGNED unsigned long long misaligned_uint64;
+#    define RAW64(p) (*(misaligned_uint64 *)(p))
+#  endif
+#endif
+
 /* --- Storage and retrieval --- */
 
+#if defined(RAW16) && defined(LTOH16)
+#  define LOAD16_L(p) LTOH16(RAW16(p))
+#endif
+#if defined(RAW16) && defined(HTOL16)
+#  define STORE16_L(p, x) (RAW16(p) = HTOL16(x))
+#endif
+#if defined(RAW16) && defined(BTOH16)
+#  define LOAD16_B(p) BTOH16(RAW16(p))
+#endif
+#if defined(RAW16) && defined(HTOB16)
+#  define STORE16_B(p, x) (RAW16(p) = HTOB16(x))
+#endif
+
+#if defined(RAW32) && defined(LTOH32)
+#  define LOAD32_L(p) LTOH32(RAW32(p))
+#endif
+#if defined(RAW32) && defined(HTOL32)
+#  define STORE32_L(p, x) (RAW32(p) = HTOL32(x))
+#endif
+#if defined(RAW32) && defined(BTOH32)
+#  define LOAD32_B(p) BTOH32(RAW32(p))
+#endif
+#if defined(RAW32) && defined(HTOB32)
+#  define STORE32_B(p, x) (RAW32(p) = HTOB32(x))
+#endif
+
+#if defined(RAW64) && defined(LTOH64)
+#  define LOAD64_L(p) LTOH64(RAW64(p))
+#endif
+#if defined(RAW64) && defined(HTOL64)
+#  define STORE64_L(p, x) (RAW64(p) = HTOL64(x))
+#endif
+#if defined(RAW64) && defined(BTOH64)
+#  define LOAD64_B(p) BTOH64(RAW64(p))
+#endif
+#if defined(RAW64) && defined(HTOB64)
+#  define STORE64_B(p, x) (RAW64(p) = HTOB64(x))
+#endif
+
 #define GETBYTE(p, o) (((octet *)(p))[o] & MASK8)
 #define PUTBYTE(p, o, v) (((octet *)(p))[o] = U8((v)))
 
 #define LOAD8(p) (GETBYTE((p), 0))
 #define STORE8(p, v) (PUTBYTE((p), 0, (v)))
 
-#define LOAD16_B(p)							\
-  (((uint16)GETBYTE((p), 0) << 8) |					\
-   ((uint16)GETBYTE((p), 1) << 0))
-#define LOAD16_L(p)							\
-  (((uint16)GETBYTE((p), 0) << 0) |					\
-   ((uint16)GETBYTE((p), 1) << 8))
+#ifndef LOAD16_B
+#  define LOAD16_B(p)
+	(((uint16)GETBYTE((p), 0) << 8) |				\
+	 ((uint16)GETBYTE((p), 1) << 0))
+#endif
+#ifndef LOAD16_L
+#  define LOAD16_L(p)							\
+	(((uint16)GETBYTE((p), 0) << 0) |				\
+	 ((uint16)GETBYTE((p), 1) << 8))
+#endif
 #define LOAD16(p) LOAD16_B((p))
 
-#define STORE16_B(p, v)							\
-  (PUTBYTE((p), 0, (uint16)(v) >> 8),					\
-   PUTBYTE((p), 1, (uint16)(v) >> 0))
-#define STORE16_L(p, v)							\
-  (PUTBYTE((p), 0, (uint16)(v) >> 0),					\
-   PUTBYTE((p), 1, (uint16)(v) >> 8))
+#ifndef STORE16_B
+#  define STORE16_B(p, v)						\
+	(PUTBYTE((p), 0, (uint16)(v) >> 8),				\
+	 PUTBYTE((p), 1, (uint16)(v) >> 0))
+#endif
+#ifndef STORE16_L
+#  define STORE16_L(p, v)						\
+	(PUTBYTE((p), 0, (uint16)(v) >> 0),				\
+	 PUTBYTE((p), 1, (uint16)(v) >> 8))
+#endif
 #define STORE16(p, v) STORE16_B((p), (v))
 
-#define LOAD24_B(p)							\
-  (((uint24)GETBYTE((p), 0) << 16) |					\
-   ((uint24)GETBYTE((p), 1) <<	8) |					\
-   ((uint24)GETBYTE((p), 2) <<	0))
-#define LOAD24_L(p)							\
-  (((uint24)GETBYTE((p), 0) <<	0) |					\
-   ((uint24)GETBYTE((p), 1) <<	8) |					\
-   ((uint24)GETBYTE((p), 2) << 16))
+#ifndef LOAD24_B
+#  define LOAD24_B(p)							\
+	(((uint24)GETBYTE((p), 0)	     << 16) |			\
+	 ((uint24)LOAD16_B((octet *)(p) + 1) <<  0))
+#endif
+#ifndef LOAD24_L
+#  define LOAD24_L(p)							\
+	(((uint24)LOAD16_L((octet *)(p) + 0) <<  0) |			\
+	 ((uint24)GETBYTE((p), 2)	     << 16))
+#endif
 #define LOAD24(p) LOAD24_B((p))
 
-#define STORE24_B(p, v)							\
-  (PUTBYTE((p), 0, (uint24)(v) >> 16),					\
-   PUTBYTE((p), 1, (uint24)(v) >>  8),					\
-   PUTBYTE((p), 2, (uint24)(v) >>  0))
-#define STORE24_L(p, v)							\
-  (PUTBYTE((p), 0, (uint24)(v) >>  0),					\
-   PUTBYTE((p), 1, (uint24)(v) >>  8),					\
-   PUTBYTE((p), 2, (uint24)(v) >> 16))
+#ifndef STORE24_B
+#  define STORE24_B(p, v)						\
+	(PUTBYTE((p), 0,	     (uint24)(v) >> 16),		\
+	 STORE16_B((octet *)(p) + 1, (uint24)(v) >>  0))
+#endif
+#ifndef STORE24_L
+#  define STORE24_L(p, v)						\
+	(STORE16_L((octet *)(p) + 0, (uint24)(v) >>  0),		\
+	 PUTBYTE((p), 2,	     (uint24)(v) >> 16))
+#endif
 #define STORE24(p, v) STORE24_B((p), (v))
 
-#define LOAD32_B(p)							\
-  (((uint32)GETBYTE((p), 0) << 24) |					\
-   ((uint32)GETBYTE((p), 1) << 16) |					\
-   ((uint32)GETBYTE((p), 2) <<	8) |					\
-   ((uint32)GETBYTE((p), 3) <<	0))
-#define LOAD32_L(p)							\
-  (((uint32)GETBYTE((p), 0) <<	0) |					\
-   ((uint32)GETBYTE((p), 1) <<	8) |					\
-   ((uint32)GETBYTE((p), 2) << 16) |					\
-   ((uint32)GETBYTE((p), 3) << 24))
+#ifndef LOAD32_B
+#  define LOAD32_B(p)							\
+	(((uint32)LOAD16_B((octet *)(p) + 0) << 16) |			\
+	 ((uint32)LOAD16_B((octet *)(p) + 2) <<  0))
+#endif
+#ifndef LOAD32_L
+#  define LOAD32_L(p)							\
+	(((uint32)LOAD16_L((octet *)(p) + 0) <<  0) |			\
+	 ((uint32)LOAD16_L((octet *)(p) + 2) << 16))
+#endif
 #define LOAD32(p) LOAD32_B((p))
 
-#define STORE32_B(p, v)							\
-  (PUTBYTE((p), 0, (uint32)(v) >> 24),					\
-   PUTBYTE((p), 1, (uint32)(v) >> 16),					\
-   PUTBYTE((p), 2, (uint32)(v) >>  8),					\
-   PUTBYTE((p), 3, (uint32)(v) >>  0))
-#define STORE32_L(p, v)							\
-  (PUTBYTE((p), 0, (uint32)(v) >>  0),					\
-   PUTBYTE((p), 1, (uint32)(v) >>  8),					\
-   PUTBYTE((p), 2, (uint32)(v) >> 16),					\
-   PUTBYTE((p), 3, (uint32)(v) >> 24))
+#ifndef STORE32_B
+#  define STORE32_B(p, v)						\
+	(STORE16_B((octet *)(p) + 0, (uint32)(v) >> 16),		\
+	 STORE16_B((octet *)(p) + 2, (uint32)(v) >>  0))
+#endif
+#ifndef STORE32_L
+#  define STORE32_L(p, v)						\
+	(STORE16_L((octet *)(p) + 0, (uint32)(v) >>  0),		\
+	 STORE16_L((octet *)(p) + 2, (uint32)(v) >> 16))
+#endif
 #define STORE32(p, v) STORE32_B((p), (v))
 
 #ifdef HAVE_UINT64
 
-#  define LOAD64_B(p)							\
-    (((uint64)GETBYTE((p), 0) << 56) |					\
-     ((uint64)GETBYTE((p), 1) << 48) |					\
-     ((uint64)GETBYTE((p), 2) << 40) |					\
-     ((uint64)GETBYTE((p), 3) << 32) |					\
-     ((uint64)GETBYTE((p), 4) << 24) |					\
-     ((uint64)GETBYTE((p), 5) << 16) |					\
-     ((uint64)GETBYTE((p), 6) <<  8) |					\
-     ((uint64)GETBYTE((p), 7) <<  0))
-#  define LOAD64_L(p)							\
-    (((uint64)GETBYTE((p), 0) <<  0) |					\
-     ((uint64)GETBYTE((p), 1) <<  8) |					\
-     ((uint64)GETBYTE((p), 2) << 16) |					\
-     ((uint64)GETBYTE((p), 3) << 24) |					\
-     ((uint64)GETBYTE((p), 4) << 32) |					\
-     ((uint64)GETBYTE((p), 5) << 40) |					\
-     ((uint64)GETBYTE((p), 6) << 48) |					\
-     ((uint64)GETBYTE((p), 7) << 56))
+#  ifndef LOAD64_B
+#    define LOAD64_B(p)							\
+	(((uint64)LOAD32_B((octet *)(p) + 0) << 32) |			\
+	 ((uint64)LOAD32_B((octet *)(p) + 4) <<  0))
+#  endif
+#  ifndef LOAD64_L
+#    define LOAD64_L(p)							\
+	(((uint64)LOAD32_L((octet *)(p) + 0) <<  0) |			\
+	 ((uint64)LOAD32_L((octet *)(p) + 4) << 32))
+#  endif
 #  define LOAD64(p) LOAD64_B((p))
 #  define LOAD64_B_(d, p) ((d).i = LOAD64_B((p)))
 #  define LOAD64_L_(d, p) ((d).i = LOAD64_L((p)))
 #  define LOAD64_(d, p) LOAD64_B_((d), (p))
 
-#  define STORE64_B(p, v)						\
-    (PUTBYTE((p), 0, (uint64)(v) >> 56),				\
-     PUTBYTE((p), 1, (uint64)(v) >> 48),				\
-     PUTBYTE((p), 2, (uint64)(v) >> 40),				\
-     PUTBYTE((p), 3, (uint64)(v) >> 32),				\
-     PUTBYTE((p), 4, (uint64)(v) >> 24),				\
-     PUTBYTE((p), 5, (uint64)(v) >> 16),				\
-     PUTBYTE((p), 6, (uint64)(v) >>  8),				\
-     PUTBYTE((p), 7, (uint64)(v) >>  0))
-#  define STORE64_L(p, v)						\
-    (PUTBYTE((p), 0, (uint64)(v) >>  0),				\
-     PUTBYTE((p), 1, (uint64)(v) >>  8),				\
-     PUTBYTE((p), 2, (uint64)(v) >> 16),				\
-     PUTBYTE((p), 3, (uint64)(v) >> 24),				\
-     PUTBYTE((p), 4, (uint64)(v) >> 32),				\
-     PUTBYTE((p), 5, (uint64)(v) >> 40),				\
-     PUTBYTE((p), 6, (uint64)(v) >> 48),				\
-     PUTBYTE((p), 7, (uint64)(v) >> 56))
+#  ifndef STORE64_B
+#    define STORE64_B(p, v)						\
+	(STORE32_B((octet *)(p) + 0, (uint64)(v) >> 32),		\
+	 STORE32_B((octet *)(p) + 4, (uint64)(v) >>  0))
+#  endif
+#  ifndef STORE64_L
+#    define STORE64_L(p, v)						\
+	(STORE32_L((octet *)(p) + 0, (uint64)(v) >>  0),		\
+	 STORE32_L((octet *)(p) + 4, (uint64)(v) >> 32))
+#  endif
 #  define STORE64(p, v) STORE64_B((p), (v))
 #  define STORE64_B_(p, v) STORE64_B((p), (v).i)
 #  define STORE64_L_(p, v) STORE64_L((p), (v).i)
@@ -433,19 +603,19 @@ typedef unsigned char octet, uint8;
 #else
 
 #  define LOAD64_B_(d, p)						\
-    ((d).hi = LOAD32_B((octet *)(p) + 0),				\
-     (d).lo = LOAD32_B((octet *)(p) + 4))
+	((d).hi = LOAD32_B((octet *)(p) + 0),				\
+	 (d).lo = LOAD32_B((octet *)(p) + 4))
 #  define LOAD64_L_(d, p)						\
-    ((d).lo = LOAD32_L((octet *)(p) + 0),				\
-     (d).hi = LOAD32_L((octet *)(p) + 4))
+	((d).lo = LOAD32_L((octet *)(p) + 0),				\
+	 (d).hi = LOAD32_L((octet *)(p) + 4))
 #  define LOAD64_(d, p) LOAD64_B_((d), (p))
 
 #  define STORE64_B_(p, v)						\
-    (STORE32_B((octet *)(p) + 0, (v).hi),				\
-     STORE32_B((octet *)(p) + 4, (v).lo))
+	(STORE32_B((octet *)(p) + 0, (v).hi),				\
+	 STORE32_B((octet *)(p) + 4, (v).lo))
 #  define STORE64_L_(p, v)						\
-    (STORE32_L((octet *)(p) + 0, (v).lo),				\
-     STORE32_L((octet *)(p) + 4, (v).hi))
+	(STORE32_L((octet *)(p) + 0, (v).lo),				\
+	 STORE32_L((octet *)(p) + 4, (v).hi))
 #  define STORE64_(p, v) STORE64_B_((p), (v))
 
 #endif