summaryrefslogtreecommitdiffstats
path: root/firmware/target/coldfire/memcpy-coldfire.S
diff options
context:
space:
mode:
authorLinus Nielsen Feltzing <linus@haxx.se>2006-07-25 11:16:03 +0000
committerLinus Nielsen Feltzing <linus@haxx.se>2006-07-25 11:16:03 +0000
commit5a0cb68e43eacc98d275baef08296f154f3ae6dd (patch)
tree07e93ecd803fc76b2605d0610a5c19d584d13623 /firmware/target/coldfire/memcpy-coldfire.S
parent9051afa8a061152941ed4cf99e0ceb9bc4a3caf1 (diff)
downloadrockbox-5a0cb68e43eacc98d275baef08296f154f3ae6dd.tar.gz
rockbox-5a0cb68e43eacc98d275baef08296f154f3ae6dd.zip
Moved some assembly optimizations to the target tree
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@10318 a1c6a512-1295-4272-9138-f99709370657
Diffstat (limited to 'firmware/target/coldfire/memcpy-coldfire.S')
-rwxr-xr-xfirmware/target/coldfire/memcpy-coldfire.S680
1 files changed, 680 insertions, 0 deletions
diff --git a/firmware/target/coldfire/memcpy-coldfire.S b/firmware/target/coldfire/memcpy-coldfire.S
new file mode 100755
index 0000000000..523e1f5ed9
--- /dev/null
+++ b/firmware/target/coldfire/memcpy-coldfire.S
@@ -0,0 +1,680 @@
+/***************************************************************************
+ * __________ __ ___.
+ * Open \______ \ ____ ____ | | _\_ |__ _______ ___
+ * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ /
+ * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < <
+ * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \
+ * \/ \/ \/ \/ \/
+ * $Id$
+ *
+ * Copyright (C) 2004-2005 by Jens Arnold
+ *
+ * All files in this archive are subject to the GNU General Public License.
+ * See the file COPYING in the source tree root for full license agreement.
+ *
+ * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY
+ * KIND, either express or implied.
+ *
+ ****************************************************************************/
+#include "config.h"
+
+ .section .icode,"ax",@progbits
+
+#define FULLSPEED /* use burst writing for word aligned destinations */
+ .align 2
+ .global memcpy
+ .global __memcpy_fwd_entry
+ .type memcpy,@function
+
+/* Copies <length> bytes of data in memory from <source> to <dest>
+ * This version is optimized for speed
+ *
+ * arguments:
+ * (4,%sp) - destination address
+ * (8,%sp) - source address
+ * (12,%sp) - length
+ *
+ * return value:
+ * %d0 - destination address (like ANSI version)
+ *
+ * register usage:
+ * %a0 - current source address
+ * %a1 - current dest address
+ * %a2 - source end address (in line-copy loops)
+ * %d0 - data / scratch
+ * %d1 - source end address (byte and longword copy) / data / scratch
+ * %d2 - data / scratch
+ * %d3..%d7 - data
+ *
+ * For maximum speed this routine reads and writes whole lines using burst
+ * move (movem.l) where possible. For byte aligned destinations (long+1 and
+ * long+3) it writes longwords only. Same goes for word aligned destinations
+ * if FULLSPEED is undefined.
+ */
+memcpy:
+ move.l (4,%sp),%a1 /* Destination */
+ move.l (8,%sp),%a0 /* Source */
+ move.l (12,%sp),%d1 /* Length */
+
+__memcpy_fwd_entry:
+ add.l %a0,%d1 /* %d1 = source end */
+
+ move.l %a0,%d0
+ addq.l #7,%d0
+ and.l #0xFFFFFFFC,%d0 /* %d0 = first source long bound + 4 */
+ cmp.l %d0,%d1 /* at least one aligned longword to copy? */
+ blo.w .bytes2_start /* no, jump directly to trailing byte loop */
+
+ subq.l #4,%d0 /* %d0 = first source long bound */
+ cmp.l %a0,%d0 /* any bytes to copy? */
+ jls .bytes1_end /* no: skip byte loop */
+
+ /* leading byte loop: copies 0..3 bytes */
+.bytes1_loop:
+ move.b (%a0)+,(%a1)+ /* copy byte */
+ cmp.l %a0,%d0 /* runs %a0 up to first long bound */
+ jhi .bytes1_loop
+
+.bytes1_end:
+ moveq.l #31,%d0
+ add.l %a0,%d0
+ and.l #0xFFFFFFF0,%d0 /* %d0 = first source line bound + 16 */
+ cmp.l %d0,%d1 /* at least one aligned line to copy? */
+ blo.w .long_start /* no: jump to longword copy loop */
+
+ lea.l (-28,%sp),%sp /* free up some registers */
+ movem.l %d2-%d7/%a2,(%sp)
+
+ moveq.l #16,%d2
+ sub.l %d2,%d0 /* %d0 = first source line bound */
+ move.l %d1,%a2 /* %a2 = end address */
+ lea.l (-15,%a2),%a2 /* adjust end address for loops doing 16 bytes/ pass */
+ move.l %a1,%d1
+ moveq.l #3,%d2 /* mask */
+ and.l %d2,%d1
+ jmp.l (2,%pc,%d1.l*4) /* switch (dest_addr & 3) */
+ bra.w .lines_do0_start
+ bra.w .lines_do1_start
+ bra.w .lines_do2_start
+ /* bra.w .lines_do3_start implicit */
+
+ /* byte aligned destination (long + 3): use line burst reads in main loop */
+.lines_do3_start:
+ moveq.l #24,%d1 /* shift count for shifting by 3 bytes */
+ cmp.l %a0,%d0 /* any leading longwords? */
+ jhi .lines_do3_head_start /* yes: leading longword copy */
+
+ movem.l (%a0),%d4-%d7 /* load first line */
+ lea.l (16,%a0),%a0
+ move.l %d4,%d2
+ lsr.l %d1,%d2 /* get high byte of first longword */
+ move.b %d2,(%a1)+ /* store byte */
+ jra .lines_do3_entry /* jump into main loop */
+
+.lines_do3_head_start:
+ move.l (%a0)+,%d7 /* load first longword */
+ move.l %d7,%d2
+ lsr.l %d1,%d2 /* get high byte */
+ move.b %d2,(%a1)+ /* store byte */
+ jra .lines_do3_head_entry /* jump into leading longword loop */
+
+.lines_do3_head_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ move.l %d7,%d2
+ lsr.l %d1,%d2 /* get high byte */
+ or.l %d2,%d6 /* combine with old lower 3 bytes */
+ move.l %d6,(%a1)+ /* store longword */
+.lines_do3_head_entry:
+ lsl.l #8,%d7 /* shift up lower 3 bytes */
+ cmp.l %a0,%d0 /* runs %a0 up to first line bound */
+ jhi .lines_do3_head_loop
+
+.lines_do3_loop:
+ move.l %d7,%d3 /* move last longword of old line away */
+ movem.l (%a0),%d4-%d7 /* load new line */
+ lea.l (16,%a0),%a0
+ move.l %d4,%d2
+ lsr.l %d1,%d2 /* get high byte of 1st longword */
+ or.l %d2,%d3 /* combine with old lower 3 bytes */
+ move.l %d3,(%a1)+ /* store longword */
+.lines_do3_entry:
+ lsl.l #8,%d4 /* shift up lower 3 bytes */
+ move.l %d5,%d2
+ lsr.l %d1,%d2 /* get high byte of 2nd longword */
+ or.l %d2,%d4 /* combine with 1st lower 3 bytes */
+ move.l %d4,(%a1)+ /* store longword */
+ lsl.l #8,%d5 /* shift up lower 3 bytes */
+ move.l %d6,%d2
+ lsr.l %d1,%d2 /* get high byte of 3rd longword */
+ or.l %d2,%d5 /* combine with 2nd lower 3 bytes */
+ move.l %d5,(%a1)+ /* store longword */
+ lsl.l #8,%d6 /* shift up lower 3 bytes */
+ move.l %d7,%d2
+ lsr.l %d1,%d2 /* get high byte of 4th longword */
+ or.l %d2,%d6 /* combine with 3rd lower 3 bytes */
+ move.l %d6,(%a1)+ /* store longword */
+ lsl.l #8,%d7 /* shift up lower 3 bytes */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_do3_loop
+
+ lea.l (12,%a2),%a2 /* readjust end address for doing longwords */
+ cmp.l %a0,%a2 /* any trailing longwords? */
+ jls .lines_do3_tail_end /* no: just store last lower 3 bytes */
+
+.lines_do3_tail_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ move.l %d7,%d2
+ lsr.l %d1,%d2 /* get high byte */
+ or.l %d2,%d6 /* combine with old lower 3 bytes */
+ move.l %d6,(%a1)+ /* store longword */
+ lsl.l #8,%d7 /* shift up lower 3 bytes */
+ cmp.l %a0,%a2 /* runs %a0 up to last long bound */
+ jhi .lines_do3_tail_loop
+
+.lines_do3_tail_end:
+ swap %d7 /* get high word */
+ move.w %d7,(%a1)+ /* store word */
+ lsr.l %d1,%d7 /* get moved-up low byte */
+ move.b %d7,(%a1)+ /* store byte */
+ jra .lines_end
+
+ /* byte aligned destination (long + 1): use line burst reads in main loop */
+.lines_do1_start:
+ moveq.l #24,%d1 /* shift count for shifting by 3 bytes */
+ cmp.l %a0,%d0 /* any leading longwords? */
+ jhi .lines_do1_head_start /* yes: leading longword copy */
+
+ movem.l (%a0),%d4-%d7 /* load first line */
+ lea.l (16,%a0),%a0
+ move.l %d4,%d2 /* first longword, bytes 3210 */
+ lsr.l #8,%d2 /* first longword, bytes .321 */
+ swap %d2 /* first longword, bytes 21.3 */
+ move.b %d2,(%a1)+ /* store byte */
+ swap %d2 /* first longword, bytes .321 */
+ move.w %d2,(%a1)+ /* store word */
+ jra .lines_do1_entry
+
+.lines_do1_head_start:
+ move.l (%a0)+,%d7 /* load first longword */
+ move.l %d7,%d2 /* first longword, bytes 3210 */
+ lsr.l #8,%d2 /* first longword, bytes .321 */
+ swap %d2 /* first longword, bytes 21.3 */
+ move.b %d2,(%a1)+ /* store byte */
+ swap %d2 /* first longword, bytes .321 */
+ move.w %d2,(%a1)+ /* store word */
+ jra .lines_do1_head_entry
+
+.lines_do1_head_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ move.l %d7,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes */
+ or.l %d2,%d6 /* combine with old low byte */
+ move.l %d6,(%a1)+ /* store longword */
+.lines_do1_head_entry:
+ lsl.l %d1,%d7 /* shift up low byte */
+ cmp.l %a0,%d0 /* runs %a0 up to first line bound */
+ jhi .lines_do1_head_loop
+
+.lines_do1_loop:
+ move.l %d7,%d3 /* move last longword of old line away */
+ movem.l (%a0),%d4-%d7 /* load new line */
+ lea.l (16,%a0),%a0
+ move.l %d4,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes of 1st longword */
+ or.l %d2,%d3 /* combine with low byte of old longword */
+ move.l %d3,(%a1)+ /* store longword */
+.lines_do1_entry:
+ lsl.l %d1,%d4 /* shift up low byte */
+ move.l %d5,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes of 2nd longword */
+ or.l %d2,%d4 /* combine with low byte of 1st longword */
+ move.l %d4,(%a1)+ /* store longword */
+ lsl.l %d1,%d5 /* shift up low byte */
+ move.l %d6,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes of 3rd longword */
+ or.l %d2,%d5 /* combine with low byte of 2nd longword */
+ move.l %d5,(%a1)+ /* store longword */
+ lsl.l %d1,%d6 /* shift up low byte */
+ move.l %d7,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes of 4th longword */
+ or.l %d2,%d6 /* combine with low byte of 4th longword */
+ move.l %d6,(%a1)+ /* store longword */
+ lsl.l %d1,%d7 /* shift up low byte */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_do1_loop
+
+ lea.l (12,%a2),%a2 /* readjust end address for doing longwords */
+ cmp.l %a0,%a2 /* any trailing longwords? */
+ jls .lines_do1_tail_end /* no: just store last low byte */
+
+.lines_do1_tail_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ move.l %d7,%d2
+ lsr.l #8,%d2 /* get upper 3 bytes */
+ or.l %d2,%d6 /* combine with old low byte */
+ move.l %d6,(%a1)+ /* store longword */
+ lsl.l %d1,%d7 /* shift up low byte */
+ cmp.l %a0,%a2 /* runs %a0 up to last long bound */
+ jhi .lines_do1_tail_loop
+
+.lines_do1_tail_end:
+ lsr.l %d1,%d7 /* get shifted-up low byte */
+ move.b %d7,(%a1)+ /* store byte */
+ jra .lines_end
+
+ /* long aligned destination (line + 0/4/8/12): head */
+.lines_do0_head_loop:
+ move.l (%a0)+,(%a1)+ /* copy longword */
+.lines_do0_start:
+ cmp.l %a0,%d0 /* runs %a0 up to first line bound */
+ jhi .lines_do0_head_loop
+
+.lines_do0_head_end:
+ move.l %a1,%d1
+ lsr.l #2,%d1
+ moveq.l #3,%d0 /* mask */
+ and.l %d0,%d1
+ moveq.l #16,%d0 /* address increment for one main loop pass */
+ jmp.l (2,%pc,%d1.l*2) /* switch ((dest_addr >> 2) & 3) */
+ bra.b .lines_lo0_start
+ bra.b .lines_lo4_start
+ bra.b .lines_lo8_start
+ /* bra.b .lines_lo12_start implicit */
+
+ /* long aligned destination (line + 12): use line bursts in the loop */
+.lines_lo12_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ move.l %d4,(%a1)+ /* store 1st longword */
+ cmp.l %a0,%a2 /* any full lines? */
+ jls .lines_lo12_end /* no: skip main loop */
+
+.lines_lo12_loop:
+ move.l %d5,%d1 /* move last 3 longwords of old line away */
+ move.l %d6,%d2
+ move.l %d7,%d3
+ movem.l (%a0),%d4-%d7 /* load new line */
+ add.l %d0,%a0
+ movem.l %d1-%d4,(%a1) /* store line (3 old + 1 new longwords) */
+ add.l %d0,%a1
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo12_loop
+
+ /* long aligned destination (line + 0/4/8/12): tail */
+.lines_lo12_end:
+ move.l %d5,(%a1)+ /* store 3rd last longword */
+.lines_lo8_end:
+ move.l %d6,(%a1)+ /* store 2nd last longword */
+.lines_lo4_end:
+ move.l %d7,(%a1)+ /* store last longword */
+.lines_lo0_end:
+ lea.l (12,%a2),%a2 /* readjust end address for doing longwords */
+ cmp.l %a0,%a2 /* any trailing longwords? */
+ jls .lines_end /* no: get outta here */
+
+.lines_do0_tail_loop:
+ move.l (%a0)+,(%a1)+ /* copy longword */
+ cmp.l %a0,%a2 /* runs %a0 up to last long bound */
+ jhi .lines_do0_tail_loop
+
+ jra .lines_end
+
+ /* line aligned destination: use line bursts in the loop */
+.lines_lo0_start:
+.lines_lo0_loop:
+ movem.l (%a0),%d4-%d7 /* load line */
+ add.l %d0,%a0
+ movem.l %d4-%d7,(%a1) /* store line */
+ add.l %d0,%a1
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo0_loop
+
+ jra .lines_lo0_end /* handle trailing longwords */
+
+ /* long aligned destination (line + 4): use line bursts in the loop */
+.lines_lo4_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ move.l %d4,(%a1)+ /* store 1st longword */
+ move.l %d5,(%a1)+ /* store 2nd longword */
+ move.l %d6,(%a1)+ /* store 3rd longword */
+ cmp.l %a0,%a2 /* any full lines? */
+ jls .lines_lo4_end /* no: skip main loop */
+
+.lines_lo4_loop:
+ move.l %d7,%d3 /* move last longword of old line away */
+ movem.l (%a0),%d4-%d7 /* load new line */
+ add.l %d0,%a0
+ movem.l %d3-%d6,(%a1) /* store line (1 old + 3 new longwords) */
+ add.l %d0,%a1
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo4_loop
+
+ jra .lines_lo4_end /* handle trailing longwords */
+
+ /* long aligned destination (line + 8): use line bursts in the loop */
+.lines_lo8_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ move.l %d4,(%a1)+ /* store 1st longword */
+ move.l %d5,(%a1)+ /* store 2nd longword */
+ cmp.l %a0,%a2
+ jls .lines_lo8_end
+
+.lines_lo8_loop:
+ move.l %d6,%d2 /* move last 2 longwords of old line away */
+ move.l %d7,%d3
+ movem.l (%a0),%d4-%d7 /* load new line */
+ add.l %d0,%a0
+ movem.l %d2-%d5,(%a1) /* store line (2 old + 2 new longwords) */
+ add.l %d0,%a1
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo8_loop
+
+ jra .lines_lo8_end /* handle trailing longwords */
+
+#ifdef FULLSPEED
+
+ /* word aligned destination (line + 2/6/10/14): head */
+.lines_do2_start:
+ cmp.l %a0,%d0 /* any leading longwords? */
+ jls .lines_do2_selector /* no: jump to mainloop selector */
+
+ move.l (%a0)+,%d7 /* load first longword */
+ swap %d7 /* swap words */
+ move.w %d7,(%a1)+ /* store high word */
+ cmp.l %a0,%d0 /* any more longword? */
+ jls .lines_do2_head_end /* no: skip head loop */
+
+.lines_do2_head_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ swap %d7 /* swap words */
+ move.w %d7,%d6 /* combine high word with old low word */
+ move.l %d6,(%a1)+ /* store longword */
+ cmp.l %a0,%d0 /* runs %a0 up to first line bound */
+ jhi .lines_do2_head_loop
+
+.lines_do2_head_end:
+ swap %d7 /* undo swap */
+ move.w %d7,(%a1)+ /* store word */
+
+.lines_do2_selector:
+ move.l %a1,%d1
+ lsr.l #2,%d1
+ moveq.l #3,%d0 /* mask */
+ and.l %d0,%d1
+ moveq.l #16,%d0 /* address increment for one main loop pass */
+ jmp.l (2,%pc,%d1.l*4) /* switch ((dest_addr >> 2) & 3) */
+ bra.w .lines_lo2_start
+ bra.w .lines_lo6_start
+ bra.w .lines_lo10_start
+ /* bra.w .lines_lo14_start implicit */
+
+ /* word aligned destination (line + 14): use line bursts in the loop */
+.lines_lo14_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,(%a1)+ /* store word */
+ jra .lines_lo14_entry /* jump into main loop */
+
+.lines_lo14_loop:
+ move.l %d4,%d0 /* move old line away */
+ move.l %d5,%d1
+ move.l %d6,%d2
+ move.l %d7,%d3
+ movem.l (%a0),%d4-%d7 /* load new line */
+ lea.l (16,%a0),%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,%d3 /* combine 1st high word with old low word */
+ movem.l %d0-%d3,(%a1) /* store line */
+ lea.l (16,%a1),%a1
+.lines_lo14_entry:
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo14_loop
+
+ /* word aligned destination (line + 2/6/10/14): tail */
+.lines_lo14_end:
+ move.l %d4,(%a1)+ /* store third last longword */
+.lines_lo10_end:
+ move.l %d5,(%a1)+ /* store second last longword */
+.lines_lo6_end:
+ move.l %d6,(%a1)+ /* store last longword */
+.lines_lo2_end:
+ lea.l (12,%a2),%a2 /* readjust end address for doing longwords */
+ cmp.l %a0,%a2 /* any trailing longwords? */
+ jls .lines_do2_tail_end /* no: skip tail loop */
+
+.lines_do2_tail_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ swap %d7 /* swap words */
+ move.w %d7,%d6 /* combine high word with old low word */
+ move.l %d6,(%a1)+ /* store longword */
+ cmp.l %a0,%a2 /* runs %a0 up to last long bound */
+ jhi .lines_do2_tail_loop
+
+.lines_do2_tail_end:
+ swap %d7 /* undo swap */
+ move.w %d7,(%a1)+ /* store last word */
+ jra .lines_end
+
+ /* word aligned destination (line + 2): use line bursts in the loop */
+.lines_lo2_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,(%a1)+ /* store high word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3nd high word with 2nd low word */
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ move.l %d4,(%a1)+ /* store 1st longword */
+ move.l %d5,(%a1)+ /* store 2nd longword */
+ move.l %d6,(%a1)+ /* store 3rd longword */
+ cmp.l %a0,%a2 /* any full lines? */
+ jls .lines_lo2_end /* no: skip main loop */
+
+.lines_lo2_loop:
+ move.l %d7,%d3 /* move last longword of old line away */
+ movem.l (%a0),%d4-%d7 /* load line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,%d3 /* combine 1st high word with old low word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ movem.l %d3-%d6,(%a1) /* store line */
+ add.l %d0,%a1
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo2_loop
+
+ jra .lines_lo2_end /* handle trailing longwords */
+
+ /* word aligned destination (line + 6): use line bursts in the loop */
+.lines_lo6_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,(%a1)+ /* store high word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ move.l %d4,(%a1)+ /* store 1st longword */
+ move.l %d5,(%a1)+ /* store 2nd longword */
+ jra .lines_lo6_entry /* jump into main loop */
+
+.lines_lo6_loop:
+ move.l %d6,%d2 /* move last 2 longwords of old line away */
+ move.l %d7,%d3
+ movem.l (%a0),%d4-%d7 /* load line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,%d3 /* combine 1st high word with old low word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ movem.l %d2-%d5,(%a1) /* store line */
+ add.l %d0,%a1
+.lines_lo6_entry:
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo6_loop
+
+ jra .lines_lo6_end /* handle trailing longwords */
+
+ /* word aligned destination (line + 10): use line bursts in the loop */
+.lines_lo10_start:
+ movem.l (%a0),%d4-%d7 /* load first line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,(%a1)+ /* store high word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ move.l %d4,(%a1)+ /* store 1st longword */
+ jra .lines_lo10_entry /* jump into main loop */
+
+.lines_lo10_loop:
+ move.l %d5,%d1 /* move last 3 longwords of old line away */
+ move.l %d6,%d2
+ move.l %d7,%d3
+ movem.l (%a0),%d4-%d7 /* load line */
+ add.l %d0,%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,%d3 /* combine 1st high word with old low word */
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ movem.l %d1-%d4,(%a1) /* store line */
+ add.l %d0,%a1
+.lines_lo10_entry:
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_lo10_loop
+
+ jra .lines_lo10_end /* handle trailing longwords */
+
+#else /* !FULLSPEED */
+
+ /* word aligned destination (long + 2): use line burst reads in the loop */
+.lines_do2_start:
+ cmp.l %a0,%d0 /* any leading longwords? */
+ jhi .lines_do2_head_start /* yes: leading longword copy */
+
+ movem.l (%a0),%d4-%d7 /* load first line */
+ lea.l (16,%a0),%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,(%a1)+ /* store high word */
+ jra .lines_do2_entry /* jump into main loop */
+
+.lines_do2_head_start:
+ move.l (%a0)+,%d7 /* load first longword */
+ swap %d7 /* swap words */
+ move.w %d7,(%a1)+ /* store high word */
+ cmp.l %a0,%d0 /* any full longword? */
+ jls .lines_do2_loop /* no: skip head loop */
+
+.lines_do2_head_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ swap %d7 /* swap words */
+ move.w %d7,%d6 /* combine high word with old low word */
+ move.l %d6,(%a1)+ /* store longword */
+ cmp.l %a0,%d0 /* runs %a0 up to first line bound */
+ jhi .lines_do2_head_loop
+
+.lines_do2_loop:
+ move.l %d7,%d3 /* move last longword of old line away */
+ movem.l (%a0),%d4-%d7 /* load line */
+ lea.l (16,%a0),%a0
+ swap %d4 /* swap words of 1st long */
+ move.w %d4,%d3 /* combine 1st high word with old low word */
+ move.l %d3,(%a1)+ /* store 1st longword */
+.lines_do2_entry:
+ swap %d5 /* swap words of 2nd long */
+ move.w %d5,%d4 /* combine 2nd high word with 1st low word */
+ move.l %d4,(%a1)+ /* store 2nd longword */
+ swap %d6 /* swap words of 3rd long */
+ move.w %d6,%d5 /* combine 3rd high word with 2nd low word */
+ move.l %d5,(%a1)+ /* store 3rd longword */
+ swap %d7 /* swap words of 4th long */
+ move.w %d7,%d6 /* combine 4th high word with 3rd low word */
+ move.l %d6,(%a1)+ /* store 4th longword */
+ cmp.l %a0,%a2 /* runs %a0 up to last line bound */
+ jhi .lines_do2_loop
+
+.lines_do2_end:
+ lea.l (12,%a2),%a2 /* readjust end address for doing longwords */
+ cmp.l %a0,%a2 /* any trailing longwords? */
+ jls .lines_do2_tail_end /* no: skip tail loop */
+
+.lines_do2_tail_loop:
+ move.l %d7,%d6 /* move old longword away */
+ move.l (%a0)+,%d7 /* load new longword */
+ swap %d7 /* swap words */
+ move.w %d7,%d6 /* combine high word with old low word */
+ move.l %d6,(%a1)+ /* store longword */
+ cmp.l %a0,%a2 /* runs %a0 up to last long bound */
+ jhi .lines_do2_tail_loop
+
+.lines_do2_tail_end:
+ swap %d7 /* undo swap */
+ move.w %d7,(%a1)+ /* store last word */
+ /* jra .lines_end implicit */
+
+#endif /* !FULLSPEED */
+
+.lines_end:
+ addq.l #3,%a2 /* readjust end address */
+ move.l %a2,%d1 /* end address in %d1 again */
+ movem.l (%sp),%d2-%d7/%a2 /* restore registers */
+ lea.l (28,%sp),%sp
+ jra .bytes2_start /* jump to trailing byte loop */
+
+.long_start:
+ subq.l #3,%d1 /* adjust end address for doing 4 bytes/ pass */
+
+ /* longword copy loop - no lines */
+.long_loop:
+ move.l (%a0)+,(%a1)+ /* copy longword (write can be unaligned) */
+ cmp.l %a0,%d1 /* runs %a0 up to last long bound */
+ jhi .long_loop
+
+ addq.l #3,%d1 /* readjust end address */
+ cmp.l %a0,%d1 /* any bytes left? */
+ jls .bytes2_end /* no: skip trailing byte loop */
+
+ /* trailing byte loop */
+.bytes2_loop:
+ move.b (%a0)+,(%a1)+ /* copy byte */
+.bytes2_start:
+ cmp.l %a0,%d1 /* runs %a0 up to end address */
+ jhi .bytes2_loop
+
+.bytes2_end:
+ move.l (4,%sp),%d0 /* return destination */
+ rts
+
+.end:
+ .size memcpy,.end-memcpy