sun4u-opl/common/memcpy.s

77 	prefetch [%o1], N_READS_STRONG
79 	cmp	%o1, %o0	! if from address is >= to use forward copy
81 	sub	%o0, %o1, %o4	! get difference of two addresses
91 	add     %o1, %o2, %o1           ! get to end of source space
101 	ldub	[%o1-1], %o3		! load last byte
103 	sub	%o1, 4, %o1
104 	ldub	[%o1+2], %o3		! load 2nd from last byte
107 	ldub	[%o1+1], %o3		! load 3rd from last byte
110 	ldub	[%o1], %o3		! load 4th from last byte
117 	dec	%o1			! decrement src address
118 	ldub	[%o1], %o3		! read a byte
129 	prefetch [%o1 - (4 * BLOCK_SIZE)], #one_read
135 	dec	%o1			! decrement src address
136 	ldub	[%o1], %o3		! read a byte
144 	andcc	%o1, 7, %g0		! chk src long word alignment
162 	ldx	[%o1-8], %o4		! load
163 	prefetch [%o1 - (8 * BLOCK_SIZE)], #one_read
167 	ldx	[%o1-16], %o3		! a block of 32 bytes
168 	sub	%o1, 32, %o1		! decrease src ptr by 32
170 	ldx	[%o1+8], %o4
173 	ldx	[%o1], %o3
180 	ldx	[%o1-8], %o4		! load and store 16 bytes
181 	sub	%o1, 16, %o1		! decrease src ptr by 16
184 	ldx	[%o1], %o3		!
194 	ldx	[%o1-8], %o4		! load 8 bytes
195 	sub	%o1, 8, %o1		! decrease src ptr by 8
215 	alignaddr %o1, %g0, %o5		! align src
220 	sub	%o1, %o4, %o1		!
303 	ldub	[%o1-1], %o3		! load last byte
305 	sub	%o1, 4, %o1
306 	ldub	[%o1+2], %o3		! load 2nd from last byte
309 	ldub	[%o1+1], %o3		! load 3rd from last byte
312 	ldub	[%o1], %o3		! load 4th from last byte
316 	dec	%o1			! decrement src address
317 	ldub	[%o1], %o3		! read a byte
334 	prefetch [%o1], N_READS_STRONG
341 	or	%o0, %o1, %o3		! prepare alignment check
346 	ldub	[%o1], %o3		! read byte
349 	ldub	[%o1+1], %o3		! repeat for a total of 4 bytes
350 	add	%o1, 4, %o1		! advance SRC by 4
352 	ldub	[%o1-2], %o3
355 	ldub	[%o1-1], %o3
364 	ldub	[%o1], %o3		! load one byte
368 	ldub	[%o1+1], %o3		! load second byte
372 	ldub	[%o1+2], %o3		! load third byte
380 	lduw	[%o1], %o3		! read word
384 	add	%o1, 8, %o1		! update SRC
385 	lduw	[%o1-4], %o3		! read word
395 	lduw	[%o1], %o3
396 	add	%o1, 4, %o1
408 	lduw	[%o1], %o3		! read word
413 	ldub	[%o1+4], %o3		! load one byte
416 	ldub	[%o1+5], %o3		! load second byte
420 	ldub	[%o1+6], %o3		! load third byte
427 	prefetch [%o1 + (4 * BLOCK_SIZE)], #one_read
430 	neg	%o1, %o3
441 	ldub	[%o1], %o4
443 	inc	%o1
448 	! Now DST is 8-byte aligned.  o0, o1, o2 are current.
451 	andcc	%o1, 0x3, %g0		! test alignment
452 	prefetch [%o1 + (1 * BLOCK_SIZE)], #one_read
464 	andcc	%o1, 0x7, %g0		! test word alignment
465 	prefetch [%o1 + (2 * BLOCK_SIZE)], #one_read
470 	prefetch [%o1 + (3 * BLOCK_SIZE)], #one_read
477 	prefetch [%o1 + (4 * BLOCK_SIZE)], #one_read
478 	ld	[%o1], %o4		! load
482 	ld	[%o1+4], %o3		! a block of 16 bytes
483 	add	%o1, 16, %o1		! increase src ptr by 16
485 	ld	[%o1-8], %o4
488 	ld	[%o1-4], %o3
498 	ld	[%o1], %o4		! load 4 bytes
501 	add	%o1, 8, %o1		! increase src ptr by 8
502 	ld	[%o1-4], %o3		! load 4 bytes
511 	ld	[%o1], %o4		! load 4 bytes
513 	add	%o1, 4, %o1		! increase src ptr by 4
521 	ldub	[%o1], %o3		! load one byte
524 	ldub	[%o1+1], %o3		! load second byte
528 	ldub	[%o1+2], %o3		! load third byte
546 	prefetch [%o1 + (3 * BLOCK_SIZE)], #one_read
553 	prefetch [%o1 + (4 * BLOCK_SIZE)], #one_read
554 	ldx	[%o1], %o4		! load
558 	ldx	[%o1+8], %o3		! a block of 32 bytes
559 	add	%o1, 32, %o1		! increase src ptr by 32
561 	ldx	[%o1-16], %o4
564 	ldx	[%o1-8], %o3
571 	ldx	[%o1], %o4		! load and store 16 bytes
572 	add	%o1, 16, %o1		! increase src ptr by 16
575 	ldx	[%o1-8], %o3		!
585 	ldx	[%o1], %o4		! load 8 bytes
586 	add	%o1, 8, %o1		! increase src ptr by 8
599 	prefetch [%o1 + (2 * BLOCK_SIZE)], #one_read
600 	prefetch [%o1 + (3 * BLOCK_SIZE)], #one_read
604 	add	%o1, 8, %o1		! prepare to round SRC upward
633 	!                          o1
636 	!                                   o1
639 	!                                   o1
642 	alignaddr %o1, %g0, %o1		! set GSR.ALIGN and align o1
654 	ldd	[%o1-8], %d0
656 	add	%o1, %o3, %o1		! back up o1
658 	ldda	[%o1]ASI_FL8_P, %d2
659 	inc	%o1
660 	andcc	%o1, 7, %g0
671 	ldd	[%o1], %d2
674 	add	%o1, 8, %o1		! update SRC
678 	ldd	[%o1], %d0
680 	add	%o1, 16, %o1		! update SRC
684 	ldd	[%o1 - 8], %d2
699 	sub	%o1, 8, %o1
704 	! Currently, o1 is pointing to the next double-aligned byte in SRC
705 	! The 8 bytes starting at [o1-8] are available in d0
729 	ldda	[%o1]ASI_FL8_P, %d2
730 	inc	%o1
745 	ldub	[%o1], %o3
748 	inc	%o1
758 	add	%o1, %o3, %o1 		! Back up %o1
762 	ldda	[%o1]ASI_FL8_P, %d2
763 	inc	%o1
773 	ldub	[%o1], %o3
775 	inc	%o1
791 	! %o1 I/O 8-byte aligned (and we've set GSR.ALIGN)
792 	! %d0 I/O already loaded with SRC data from [%o1-8]
799 	! Recall that %o1 is further into SRC than %o0 is into DST
801 	prefetch [%o1 + (4 * BLOCK_SIZE)], #one_read
802 	prefetch [%o1 + (8 * BLOCK_SIZE)], #one_read
807 	prefetch [%o1 + (12 * BLOCK_SIZE)], #one_read
809 	ldd	[%o1], %f2
810 	ldd	[%o1 + 0x8], %f4
812 	ldd	[%o1 + 0x10], %f6
814 	ldd	[%o1 + 0x18], %f8
816 	ldd	[%o1 + 0x20], %f10
819 	prefetch [%o1 + (16 * BLOCK_SIZE)], #one_read
820 	ldd	[%o1 + 0x28], %f12
823 	ldd	[%o1 + 0x30], %f14
825 	ldd	[%o1 + 0x38], %f0
827 	add	%o1, BLOCK_SIZE, %o1	! update SRC
832 	ldd	[%o1], %f2
834 	ldd	[%o1 + 0x8], %f4
845 	prefetch [%o1 + (24 * BLOCK_SIZE) + BLOCK_SIZE], #one_read
847 	ldd	[%o1 + 0x10], %f6
849 	ldd	[%o1 + 0x18], %f8
851 	ldd	[%o1 + 0x20], %f10
853 	ldd	[%o1 + 0x28], %f12
855 	ldd	[%o1 + 0x30], %f14
857 	ldd	[%o1 + 0x38], %f0
859 	prefetch [%o1 + (18 * BLOCK_SIZE)], #one_read
863 	add	%o1, BLOCK_SIZE, %o1	! update SRC
871 	ldd	[%o1], %f2
885 	ldd	[%o1 + 0x8], %f4
886 	ldd	[%o1 + 0x10], %f6
888 	ldd	[%o1 + 0x18], %f8
890 	ldd	[%o1 + 0x20], %f10
892 	ldd	[%o1 + 0x28], %f12
894 	ldd	[%o1 + 0x30], %f14
913 	! %o1 I/O 8-byte aligned (and we've set GSR.ALIGN)
914 	! %d0 I/O already loaded with SRC data from [%o1-8]
921 	! Recall that %o1 is further into SRC than %o0 is into DST
923 	ldd	[%o1], %f2
924 	ldd	[%o1 + 0x8], %f4
926 	ldd	[%o1 + 0x10], %f6
928 	ldd	[%o1 + 0x18], %f8
930 	ldd	[%o1 + 0x20], %f10
933 	ldd	[%o1 + 0x28], %f12
935 	prefetch [%o1 + (16 * BLOCK_SIZE)], #one_read
937 	ldd	[%o1 + 0x30], %f14
939 	ldd	[%o1 + 0x38], %f0
940 	prefetch [%o1 + (17 * BLOCK_SIZE)], #one_read
942 	add	%o1, BLOCK_SIZE, %o1	! update SRC
951 	ldd	[%o1], %f2
953 	ldd	[%o1 + 0x8], %f4
957 	ldd	[%o1 + 0x10], %f6
959 	ldd	[%o1 + 0x18], %f8
961 	ldd	[%o1 + 0x20], %f10
963 	ldd	[%o1 + 0x28], %f12
965 	ldd	[%o1 + 0x30], %f14
966 	prefetch [%o1 + (2 * BLOCK_SIZE)], #n_reads
968 	ldd	[%o1 + 0x38], %f0
970 	prefetch [%o1 + (25 * BLOCK_SIZE)], #one_read
974 	prefetch [%o1 + (18 * BLOCK_SIZE)], #one_read
976 	add	%o1, BLOCK_SIZE, %o1		! update SRC
985 	ldd	[%o1], %f2
999 	ldd	[%o1 + 0x8], %f4
1000 	ldd	[%o1 + 0x10], %f6
1002 	ldd	[%o1 + 0x18], %f8
1004 	ldd	[%o1 + 0x20], %f10
1006 	ldd	[%o1 + 0x28], %f12
1008 	ldd	[%o1 + 0x30], %f14