atmel_pdc.rar_The Next

#ifndef _VIDEO_ATAFB_UTILS_H #define _VIDEO_ATAFB_UTILS_H /* ================================================================= */ /* Utility Assembler Functions */ /* ================================================================= */ /* ====================================================================== */ /* Those of a delicate disposition might like to skip the next couple of * pages. * * These functions are drop in replacements for memmove and * memset(_, 0, _). However their five instances add at least a kilobyte * to the object file. You have been warned. * * Not a great fan of assembler for the sake of it, but I think * that these routines are at least 10 times faster than their C * equivalents for large blits, and that's important to the lowest level of * a graphics driver. Question is whether some scheme with the blitter * would be faster. I suspect not for simple text system - not much * asynchrony. * * Code is very simple, just gruesome expansion. Basic strategy is to * increase data moved/cleared at each step to 16 bytes to reduce * instruction per data move overhead. movem might be faster still * For more than 15 bytes, we try to align the write direction on a * longword boundary to get maximum speed. This is even more gruesome. * Unaligned read/write used requires 68020+ - think this is a problem? * * Sorry! */ /* ++roman: I've optimized Robert's original versions in some minor * aspects, e.g. moveq instead of movel, let gcc choose the registers, * use movem in some places... * For other modes than 1 plane, lots of more such assembler functions * were needed (e.g. the ones using movep or expanding color values). */ /* ++andreas: more optimizations: subl #65536,d0 replaced by clrw d0; subql #1,d0 for dbcc addal is faster than addaw movep is rather expensive compared to ordinary move's some functions rewritten in C for clarity, no speed loss */ static inline void *fb_memclear_small(void *s, size_t count) { if (!count) return 0; asm volatile ("\n" " lsr.l #1,%1 ; jcc 1f ; move.b %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.w %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.l %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.l %2,-(%0) ; move.l %2,-(%0)\n" "1:" : "=a" (s), "=d" (count) : "d" (0), "0" ((char *)s + count), "1" (count)); asm volatile ("\n" " subq.l #1,%1\n" " jcs 3f\n" " move.l %2,%%d4; move.l %2,%%d5; move.l %2,%%d6\n" "2: movem.l %2/%%d4/%%d5/%%d6,-(%0)\n" " dbra %1,2b\n" "3:" : "=a" (s), "=d" (count) : "d" (0), "0" (s), "1" (count) : "d4", "d5", "d6" ); return 0; } static inline void *fb_memclear(void *s, size_t count) { if (!count) return 0; if (count < 16) { asm volatile ("\n" " lsr.l #1,%1 ; jcc 1f ; clr.b (%0)+\n" "1: lsr.l #1,%1 ; jcc 1f ; clr.w (%0)+\n" "1: lsr.l #1,%1 ; jcc 1f ; clr.l (%0)+\n" "1: lsr.l #1,%1 ; jcc 1f ; clr.l (%0)+ ; clr.l (%0)+\n" "1:" : "=a" (s), "=d" (count) : "0" (s), "1" (count)); } else { long tmp; asm volatile ("\n" " move.l %1,%2\n" " lsr.l #1,%2 ; jcc 1f ; clr.b (%0)+ ; subq.w #1,%1\n" " lsr.l #1,%2 ; jcs 2f\n" /* %0 increased=>bit 2 switched*/ " clr.w (%0)+ ; subq.w #2,%1 ; jra 2f\n" "1: lsr.l #1,%2 ; jcc 2f\n" " clr.w (%0)+ ; subq.w #2,%1\n" "2: move.w %1,%2; lsr.l #2,%1 ; jeq 6f\n" " lsr.l #1,%1 ; jcc 3f ; clr.l (%0)+\n" "3: lsr.l #1,%1 ; jcc 4f ; clr.l (%0)+ ; clr.l (%0)+\n" "4: subq.l #1,%1 ; jcs 6f\n" "5: clr.l (%0)+; clr.l (%0)+ ; clr.l (%0)+ ; clr.l (%0)+\n" " dbra %1,5b ; clr.w %1; subq.l #1,%1; jcc 5b\n" "6: move.w %2,%1; btst #1,%1 ; jeq 7f ; clr.w (%0)+\n" "7: btst #0,%1 ; jeq 8f ; clr.b (%0)+\n" "8:" : "=a" (s), "=d" (count), "=d" (tmp) : "0" (s), "1" (count)); } return 0; } static inline void *fb_memset255(void *s, size_t count) { if (!count) return 0; asm volatile ("\n" " lsr.l #1,%1 ; jcc 1f ; move.b %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.w %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.l %2,-(%0)\n" "1: lsr.l #1,%1 ; jcc 1f ; move.l %2,-(%0) ; move.l %2,-(%0)\n" "1:" : "=a" (s), "=d" (count) : "d" (-1), "0" ((char *)s+count), "1" (count)); asm volatile ("\n" " subq.l #1,%1 ; jcs 3f\n" " move.l %2,%%d4; move.l %2,%%d5; move.l %2,%%d6\n" "2: movem.l %2/%%d4/%%d5/%%d6,-(%0)\n" " dbra %1,2b\n" "3:" : "=a" (s), "=d" (count) : "d" (-1), "0" (s), "1" (count) : "d4", "d5", "d6"); return 0; } static inline void *fb_memmove(void *d, const void *s, size_t count) { if (d < s) { if (count < 16) { asm volatile ("\n" " lsr.l #1,%2 ; jcc 1f ; move.b (%1)+,(%0)+\n" "1: lsr.l #1,%2 ; jcc 1f ; move.w (%1)+,(%0)+\n" "1: lsr.l #1,%2 ; jcc 1f ; move.l (%1)+,(%0)+\n" "1: lsr.l #1,%2 ; jcc 1f ; move.l (%1)+,(%0)+ ; move.l (%1)+,(%0)+\n" "1:" : "=a" (d), "=a" (s), "=d" (count) : "0" (d), "1" (s), "2" (count)); } else { long tmp; asm volatile ("\n" " move.l %0,%3\n" " lsr.l #1,%3 ; jcc 1f ; move.b (%1)+,(%0)+ ; subqw #1,%2\n" " lsr.l #1,%3 ; jcs 2f\n" /* %0 increased=>bit 2 switched*/ " move.w (%1)+,(%0)+ ; subqw #2,%2 ; jra 2f\n" "1: lsr.l #1,%3 ; jcc 2f\n" " move.w (%1)+,(%0)+ ; subqw #2,%2\n" "2: move.w %2,%-; lsr.l #2,%2 ; jeq 6f\n" " lsr.l #1,%2 ; jcc 3f ; move.l (%1)+,(%0)+\n" "3: lsr.l #1,%2 ; jcc 4f ; move.l (%1)+,(%0)+ ; move.l (%1)+,(%0)+\n" "4: subq.l #1,%2 ; jcs 6f\n" "5: move.l (%1)+,(%0)+; move.l (%1)+,(%0)+\n" " move.l (%1)+,(%0)+; move.l (%1)+,(%0)+\n" " dbra %2,5b ; clr.w %2; subq.l #1,%2; jcc 5b\n" "6: move.w %+,%2; btst #1,%2 ; jeq 7f ; move.w (%1)+,(%0)+\n" "7: btst #0,%2 ; jeq 8f ; move.b (%1)+,(%0)+\n" "8:" : "=a" (d), "=a" (s), "=d" (count), "=d" (tmp) : "0" (d), "1" (s), "2" (count)); } } else { if (count < 16) { asm volatile ("\n" " lsr.l #1,%2 ; jcc 1f ; move.b -(%1),-(%0)\n" "1: lsr.l #1,%2 ; jcc 1f ; move.w -(%1),-(%0)\n" "1: lsr.l #1,%2 ; jcc 1f ; move.l -(%1),-(%0)\n" "1: lsr.l #1,%2 ; jcc 1f ; move.l -(%1),-(%0) ; move.l -(%1),-(%0)\n" "1:" : "=a" (d), "=a" (s), "=d" (count) : "0" ((char *) d + count), "1" ((char *) s + count), "2" (count)); } else { long tmp; asm volatile ("\n" " move.l %0,%3\n" " lsr.l #1,%3 ; jcc 1f ; move.b -(%1),-(%0) ; subqw #1,%2\n" " lsr.l #1,%3 ; jcs 2f\n" /* %0 increased=>bit 2 switched*/ " move.w -(%1),-(%0) ; subqw #2,%2 ; jra 2f\n" "1: lsr.l #1,%3 ; jcc 2f\n" " move.w -(%1),-(%0) ; subqw #2,%2\n" "2: move.w %2,%-; lsr.l #2,%2 ; jeq 6f\n" " lsr.l #1,%2 ; jcc 3f ; move.l -(%1),-(%0)\n" "3: lsr.l #1,%2 ; jcc 4f ; move.l -(%1),-(%0) ; move.l -(%1),-(%0)\n" "4: subq.l #1,%2 ; jcs 6f\n" "5: move.l -(%1),-(%0); move.l -(%1),-(%0)\n" " move.l -(%1),-(%0); move.l -(%1),-(%0)\n" " dbra %2,5b ; clr.w %2; subq.l #1,%2; jcc 5b\n" "6: move.w %+,%2; btst #1,%2 ; jeq 7f ; move.w -(%1),-(%0)\n" "7: btst #0,%2 ; jeq 8f ; move.b -(%1),-(%0)\n" "8:" : "=a" (d), "=a" (s), "=d" (count), "=d" (tmp) : "0" ((char *) d + count), "1" ((char *) s + count), "2" (count)); } } return 0; } /* ++andreas: Simple and fast version of memmove, assumes size is divisible by 16, suitable for moving the whole screen bitplane */ static inline void fast_memmove(char *dst, const char *src, size_t size) { if (!size) return; if (dst < src) asm volatile ("\n" "1: movem.l (%0)+,%%d0/%%d1/%%a0/%%a1\n" " movem.l %%d0/%%d1/%%a0/%%a1,%1@\n" " addq.l #8,%1; addq.l #8,%1\n" " dbra %2,1b\n" " clr.w %2; subq.l #1,%2\n" " jcc 1b" : "=a" (src), "=a" (dst), "=d" (size) : "0" (src), "1" (dst), "2" (size / 16 - 1) : "d0", "d1", "a0", "a1", "memory"); else asm volatile ("\n" "1: subq.l #8,%0; subq.l #8,%0\n" " movem.l %0@,%%d0/%%d1/%%a0/%%a1\n" " movem.l %%d0/%%d1/%%a0/%%a1,-(%1)\n" " dbra %2,1b\n" " clr.w %2; subq.l #1,%2\n" " jcc 1b" : "=a" (sr