| 30 |
|
# error "Undefined 32-bit word blit function" |
| 31 |
|
#endif |
| 32 |
|
|
| 33 |
+ |
#if !defined(FB_BLIT_4) |
| 34 |
+ |
# error "Undefined 64-bit word blit function" |
| 35 |
+ |
#endif |
| 36 |
+ |
|
| 37 |
|
static void FB_FUNC_NAME(uint8 * dest, const uint8 * source, uint32 length) |
| 38 |
|
{ |
| 35 |
– |
#define DEREF_LONG_PTR(ptr, ofs) (((uint32 *)(ptr))[(ofs)]) |
| 39 |
|
#define DEREF_WORD_PTR(ptr, ofs) (((uint16 *)(ptr))[(ofs)]) |
| 40 |
+ |
#define DEREF_LONG_PTR(ptr, ofs) (((uint32 *)(ptr))[(ofs)]) |
| 41 |
+ |
#define DEREF_QUAD_PTR(ptr, ofs) (((uint64 *)(ptr))[(ofs)]) |
| 42 |
|
|
| 43 |
|
#ifndef UNALIGNED_PROFITABLE |
| 44 |
|
#if FB_DEPTH <= 8 |
| 59 |
|
#endif |
| 60 |
|
#endif |
| 61 |
|
|
| 62 |
< |
// Blit 4-byte words |
| 63 |
< |
if (length >= 4) { |
| 64 |
< |
const int remainder = (length / 4) % 8; |
| 65 |
< |
source += remainder * 4; |
| 66 |
< |
dest += remainder * 4; |
| 62 |
> |
// Blit 8-byte words |
| 63 |
> |
if (length >= 8) { |
| 64 |
> |
const int remainder = (length / 8) % 8; |
| 65 |
> |
source += remainder * 8; |
| 66 |
> |
dest += remainder * 8; |
| 67 |
|
|
| 68 |
< |
int n = ((length / 4) + 7) / 8; |
| 68 |
> |
int n = ((length / 8) + 7) / 8; |
| 69 |
|
switch (remainder) { |
| 70 |
|
case 0: do { |
| 71 |
< |
dest += 32; source += 32; |
| 72 |
< |
FB_BLIT_2(DEREF_LONG_PTR(dest, -8), DEREF_LONG_PTR(source, -8)); |
| 73 |
< |
case 7: FB_BLIT_2(DEREF_LONG_PTR(dest, -7), DEREF_LONG_PTR(source, -7)); |
| 74 |
< |
case 6: FB_BLIT_2(DEREF_LONG_PTR(dest, -6), DEREF_LONG_PTR(source, -6)); |
| 75 |
< |
case 5: FB_BLIT_2(DEREF_LONG_PTR(dest, -5), DEREF_LONG_PTR(source, -5)); |
| 76 |
< |
case 4: FB_BLIT_2(DEREF_LONG_PTR(dest, -4), DEREF_LONG_PTR(source, -4)); |
| 77 |
< |
case 3: FB_BLIT_2(DEREF_LONG_PTR(dest, -3), DEREF_LONG_PTR(source, -3)); |
| 78 |
< |
case 2: FB_BLIT_2(DEREF_LONG_PTR(dest, -2), DEREF_LONG_PTR(source, -2)); |
| 79 |
< |
case 1: FB_BLIT_2(DEREF_LONG_PTR(dest, -1), DEREF_LONG_PTR(source, -1)); |
| 71 |
> |
dest += 64; source += 64; |
| 72 |
> |
FB_BLIT_4(DEREF_QUAD_PTR(dest, -8), DEREF_QUAD_PTR(source, -8)); |
| 73 |
> |
case 7: FB_BLIT_4(DEREF_QUAD_PTR(dest, -7), DEREF_QUAD_PTR(source, -7)); |
| 74 |
> |
case 6: FB_BLIT_4(DEREF_QUAD_PTR(dest, -6), DEREF_QUAD_PTR(source, -6)); |
| 75 |
> |
case 5: FB_BLIT_4(DEREF_QUAD_PTR(dest, -5), DEREF_QUAD_PTR(source, -5)); |
| 76 |
> |
case 4: FB_BLIT_4(DEREF_QUAD_PTR(dest, -4), DEREF_QUAD_PTR(source, -4)); |
| 77 |
> |
case 3: FB_BLIT_4(DEREF_QUAD_PTR(dest, -3), DEREF_QUAD_PTR(source, -3)); |
| 78 |
> |
case 2: FB_BLIT_4(DEREF_QUAD_PTR(dest, -2), DEREF_QUAD_PTR(source, -2)); |
| 79 |
> |
case 1: FB_BLIT_4(DEREF_QUAD_PTR(dest, -1), DEREF_QUAD_PTR(source, -1)); |
| 80 |
|
} while (--n > 0); |
| 81 |
|
} |
| 82 |
|
} |
| 83 |
|
|
| 84 |
+ |
// There could be one long left to blit |
| 85 |
+ |
if (length & 4) { |
| 86 |
+ |
FB_BLIT_2(DEREF_LONG_PTR(dest, 0), DEREF_LONG_PTR(source, 0)); |
| 87 |
+ |
#if FB_DEPTH <= 16 |
| 88 |
+ |
dest += 4; |
| 89 |
+ |
source += 4; |
| 90 |
+ |
#endif |
| 91 |
+ |
} |
| 92 |
+ |
|
| 93 |
|
#if FB_DEPTH <= 16 |
| 94 |
|
// There could be one word left to blit |
| 95 |
|
if (length & 2) { |
| 121 |
|
#undef FB_BLIT_2 |
| 122 |
|
#endif |
| 123 |
|
|
| 124 |
+ |
#ifdef FB_BLIT_4 |
| 125 |
+ |
#undef FB_BLIT_4 |
| 126 |
+ |
#endif |
| 127 |
+ |
|
| 128 |
|
#ifdef FB_DEPTH |
| 129 |
|
#undef FB_DEPTH |
| 130 |
|
#endif |
