| 32 |
|
|
| 33 |
|
static void FB_FUNC_NAME(uint8 * dest, const uint8 * source, uint32 length) |
| 34 |
|
{ |
| 35 |
+ |
union { |
| 36 |
+ |
uint8 * bp; |
| 37 |
+ |
uint16 * wp; |
| 38 |
+ |
uint32 * lp; |
| 39 |
+ |
} s, d; |
| 40 |
+ |
|
| 41 |
+ |
s.bp = (uint8 *)source; |
| 42 |
+ |
d.bp = dest; |
| 43 |
+ |
|
| 44 |
|
#if FB_DEPTH <= 8 |
| 45 |
|
// Align source and dest to 16-bit word boundaries |
| 46 |
|
if (FB_DEPTH <= 8 && ((unsigned long) source) & 1) { |
| 47 |
< |
*dest++ = *source++; |
| 47 |
> |
*d.bp++ = *s.bp++; |
| 48 |
|
length -= 1; |
| 49 |
|
} |
| 50 |
|
#endif |
| 51 |
|
|
| 43 |
– |
// source and dest are mutually aligned |
| 44 |
– |
uint16 * swp = ((uint16 *)source); |
| 45 |
– |
uint16 * dwp = ((uint16 *) dest ); |
| 46 |
– |
|
| 52 |
|
#if FB_DEPTH <= 8 |
| 53 |
|
if (length >= 2) { |
| 54 |
|
#endif |
| 56 |
|
#if FB_DEPTH <= 16 |
| 57 |
|
// Align source and dest to 32-bit word boundaries |
| 58 |
|
if (((unsigned long) source) & 2) { |
| 59 |
< |
const uint16 val = *swp++; |
| 60 |
< |
FB_BLIT_1(*dwp++, val); |
| 59 |
> |
const uint16 val = *s.wp++; |
| 60 |
> |
FB_BLIT_1(*d.wp++, val); |
| 61 |
|
length -= 2; |
| 62 |
|
} |
| 63 |
|
#endif |
| 65 |
|
// Blit 4-byte words |
| 66 |
|
if (length >= 4) { |
| 67 |
|
const int remainder = (length / 4) % 8; |
| 68 |
< |
uint32 * slp = (uint32 *)swp + remainder; |
| 69 |
< |
uint32 * dlp = (uint32 *)dwp + remainder; |
| 68 |
> |
s.lp += remainder; |
| 69 |
> |
d.lp += remainder; |
| 70 |
|
|
| 71 |
|
int n = ((length / 4) + 7) / 8; |
| 72 |
|
switch (remainder) { |
| 73 |
|
case 0: do { |
| 74 |
< |
slp += 8; dlp += 8; |
| 75 |
< |
FB_BLIT_2(dlp[-8], slp[-8]); |
| 76 |
< |
case 7: FB_BLIT_2(dlp[-7], slp[-7]); |
| 77 |
< |
case 6: FB_BLIT_2(dlp[-6], slp[-6]); |
| 78 |
< |
case 5: FB_BLIT_2(dlp[-5], slp[-5]); |
| 79 |
< |
case 4: FB_BLIT_2(dlp[-4], slp[-4]); |
| 80 |
< |
case 3: FB_BLIT_2(dlp[-3], slp[-3]); |
| 81 |
< |
case 2: FB_BLIT_2(dlp[-2], slp[-2]); |
| 82 |
< |
case 1: FB_BLIT_2(dlp[-1], slp[-1]); |
| 74 |
> |
s.lp += 8; d.lp += 8; |
| 75 |
> |
FB_BLIT_2(d.lp[-8], s.lp[-8]); |
| 76 |
> |
case 7: FB_BLIT_2(d.lp[-7], s.lp[-7]); |
| 77 |
> |
case 6: FB_BLIT_2(d.lp[-6], s.lp[-6]); |
| 78 |
> |
case 5: FB_BLIT_2(d.lp[-5], s.lp[-5]); |
| 79 |
> |
case 4: FB_BLIT_2(d.lp[-4], s.lp[-4]); |
| 80 |
> |
case 3: FB_BLIT_2(d.lp[-3], s.lp[-3]); |
| 81 |
> |
case 2: FB_BLIT_2(d.lp[-2], s.lp[-2]); |
| 82 |
> |
case 1: FB_BLIT_2(d.lp[-1], s.lp[-1]); |
| 83 |
|
} while (--n > 0); |
| 84 |
|
} |
| 85 |
|
} |
| 86 |
|
|
| 87 |
|
#if FB_DEPTH <= 16 |
| 88 |
< |
// There might remain one word to blit |
| 88 |
> |
// There might remain at least one word to blit |
| 89 |
|
if (length & 2) { |
| 90 |
< |
uint16 * const s = (uint16 *)(((uint8 *)swp) + length - 2); |
| 91 |
< |
uint16 * const d = (uint16 *)(((uint8 *)dwp) + length - 2); |
| 92 |
< |
FB_BLIT_1(*d, *s); |
| 90 |
> |
FB_BLIT_1(*d.wp, *s.wp); |
| 91 |
> |
#if FB_DEPTH <= 8 |
| 92 |
> |
d.wp++; |
| 93 |
> |
s.wp++; |
| 94 |
> |
#endif |
| 95 |
|
} |
| 96 |
|
#endif |
| 97 |
|
|
| 98 |
|
#if FB_DEPTH <= 8 |
| 99 |
|
} |
| 100 |
+ |
|
| 101 |
+ |
// There might remain one byte to blit |
| 102 |
+ |
if (length & 1) |
| 103 |
+ |
*d.bp = *s.bp; |
| 104 |
|
#endif |
| 105 |
|
} |
| 106 |
|
|
