shotgun/lib/memutil.h

   1 #include <stdint.h>
   2
   3 static inline void fast_memcpy32_x86(void *v_to, void *v_from, int words) {
   4   uint32_t *to = (uint32_t*)(v_to);
   5   uint32_t *from = (uint32_t*)(v_from);
   6   while(words--) { *to++ = *from++; }
   7 }
   8
   9 static inline void fast_memcpy64(void *v_to, void *v_from, int words) {
  10   uint64_t *to = (uint64_t*)(v_to);
  11   uint64_t *from = (uint64_t*)(v_from);
  12   while(words--) { *to++ = *from++; }
  13 }
  14
  15 /* Found this tip floating around the internet. PPC has a
  16    preincrement and store opcode, but not post. */
  17 static inline void fast_memcpy32_ppc(void *v_to, void *v_from, int words) {
  18   uint32_t *to = (uint32_t*)(v_to);
  19   uint32_t *from = (uint32_t*)(v_from);
  20   --to;
  21   --from;
  22
  23   while(words--) { *++to = *++from; }
  24 }
  25
  26 static inline void fast_memfill32_x86(void *v_to, uint32_t val, int words) {
  27   uint32_t *to = (uint32_t*)(v_to);
  28   while(words--) { *to++ = val; }
  29 }
  30
  31 static inline void fast_memfill64(void *v_to, uint64_t val, int words) {
  32   uint64_t *to = (uint64_t*)(v_to);
  33   while(words--) { *to++ = val; }
  34 }
  35
  36 static inline void fast_memfill32_ppc(void *v_to, uint32_t val, int words) {
  37   uint32_t *to = (uint32_t*)(v_to);
  38   --to;
  39   while(words--) { *++to = val; }
  40 }
  41
  42 static inline void fast_memfill32_s20(void *v_to, uint32_t val) {
  43   uint32_t *to = (uint32_t*)v_to;
  44   *(to) = val;
  45   *(to + 1) = val;
  46   *(to + 2) = val;
  47   *(to + 3) = val;
  48   *(to + 4) = val;
  49 }
  50
  51 static inline void fast_memfill64_s20(void *v_to, uint64_t val) {
  52   uint64_t *to = (uint64_t*)v_to;
  53   *(to) = val;
  54   *(to + 1) = val;
  55   *(to + 2) = val;
  56   *(to + 3) = val;
  57   *(to + 4) = val;
  58 }
  59
  60 #ifdef __POWERPC__
  61 #define fast_memcpy32(a,b,c) fast_memcpy32_ppc(a,b,c)
  62 #define fast_memfill32(a,b,c) fast_memfill32_ppc(a,b,c)
  63 #else
  64 #define fast_memcpy32(a,b,c) fast_memcpy32_x86(a,b,c)
  65 #define fast_memfill32(a,b,c) fast_memfill32_x86(a,b,c)
  66 #endif
  67
  68 #ifdef __x86_64__
  69 #define fast_memcpy(a,b,c) fast_memcpy64(a,b,c)
  70 #define fast_memfill(a,b,c) fast_memfill64(a,b,c)
  71 #define fast_memfill_s20(a,b) fast_memfill64_s20(a,b)
  72 #else
  73 #define fast_memcpy(a,b,c) fast_memcpy32(a,b,c)
  74 #define fast_memfill(a,b,c) fast_memfill32(a,b,c)
  75 #define fast_memfill_s20(a,b) fast_memfill32_s20(a,b)
  76 #endif