patch-2.3.23 linux/arch/sh/lib/memcpy.S

• Lines: 347
• Date: Mon Oct 18 11:16:13 1999
• Orig file: v2.3.22/linux/arch/sh/lib/memcpy.S
• Orig date: Tue Aug 31 17:29:13 1999

diff -u --recursive --new-file v2.3.22/linux/arch/sh/lib/memcpy.S linux/arch/sh/lib/memcpy.S
@@ -1,131 +1,227 @@
-! Taken from newlib-1.8.0
+/* $Id: memcpy.S,v 1.3 1999/09/28 11:32:48 gniibe Exp $
+ *
+ * "memcpy" implementation of SuperH
+ *
+ * Copyright (C) 1999  Niibe Yutaka
+ *
+ */
 
-!
-! Fast SH memcpy
-!
-! by Toshiyasu Morita (tm@netcom.com)
-! hacked by J"orn Rernnecke (amylaar@cygnus.co.uk) ("o for o-umlaut)
-!
-! Entry: r4: destination pointer
-!        r5: source pointer
-!        r6: byte count
-!
-! Exit:  r0: destination pointer
-!        r1-r7: trashed
-!
-! Notes: Usually one wants to do small reads and write a longword, but
-!        unfortunately it is difficult in some cases to concatanate bytes
-!        into a longword on the SH, so this does a longword read and small
-!        writes.
-!
-! This implementation makes two assumptions about how it is called:
-!
-! 1.: If the byte count is nonzero, the address of the last byte to be
-!     copied is unsigned greater than the address of the first byte to
-!     be copied.  This could be easily swapped for a signed comparison,
-!     but the algorithm used needs some comparison.
-!
-! 2.: When there are two or three bytes in the last word of an 11-or-bore
-!     bytes memory chunk to b copied, the rest of the word can be read
-!     without size effects.
-!     This could be easily changed by increasing the minumum size of
-!     a fast memcpy and the amount subtracted from r7 before L_2l_loop be 2,
-!     however, this would cost a few extra cyles on average.
-!
+/*
+ * void *memcpy(void *dst, const void *src, size_t n);
+ * No overlap between the memory of DST and of SRC are assumed.
+ */
 
 #include <linux/linkage.h>
 ENTRY(memcpy)
-	! Big endian version copies with decreasing addresses.
-	mov	r4,r0
-	add	r6,r0
-	sub	r4,r5
-	mov	#11,r1
-	cmp/hs	r1,r6
-	bf/s	L_small
+	tst	r6,r6
+	bt/s	9f		! if n=0, do nothing
+	 mov	r4,r0
+	sub	r4,r5		! From here, r5 has the distance to r0
+	add	r6,r0		! From here, r0 points the end of copying point
+	mov	#12,r1
+	cmp/gt	r6,r1
+	bt/s	7f		! if it's too small, copy a byte at once
 	 add	#-1,r5
-	mov	r5,r3
-	add	r0,r3
-	shlr	r3
-	bt/s	L_even
-	 mov	r4,r7
-	mov.b	@(r0,r5),r2
-	add	#-1,r3
-	mov.b	r2,@-r0
-L_even:
-	tst	#1,r0
-	add	#-1,r5
-	bf/s	L_odddst
-	 add	#8,r7
-	tst	#2,r0
-	bt	L_al4dst
-	add	#-1,r3
-	mov.w	@(r0,r5),r1
-	mov.w	r1,@-r0
-L_al4dst:
-	shlr	r3
-	bt	L_al4both
-	mov.w	@(r0,r5),r1
-	swap.w	r1,r1
-	add	#4,r7
-	add	#-4,r5
-	.align	2
-L_2l_loop:
-	mov.l	@(r0,r5),r2
-	xtrct	r2,r1
-	mov.l	r1,@-r0
-	cmp/hs	r7,r0
-	mov.l	@(r0,r5),r1
-	xtrct	r1,r2
-	mov.l	r2,@-r0
-	bt	L_2l_loop
-	bra	L_cleanup
-	 add	#5,r5
+	add	#1,r5
+	!			From here, r6 is free
+	!
+	!      r4   -->  [ ...  ] DST             [ ...  ] SRC
+	!	         [ ...  ]                 [ ...  ]
+	!	           :                        :
+	!      r0   -->  [ ...  ]       r0+r5 --> [ ...  ]
+	!
+	!
+	mov	r5,r1
+	mov	#3,r2
+	and	r2,r1
+	shll2	r1
+	mov	r0,r3		! Save the value on R0 to R3
+	mova	jmptable,r0
+	add	r1,r0
+	mov.l	@r0,r1
+	jmp	@r1
+	 mov	r3,r0		! and back to R0
+	.balign	4
+jmptable:
+	.long	case0
+	.long	case1
+	.long	case2
+	.long	case3
 
-	nop ! avoid nop in executed code.
-L_al4both:
-	add	#-2,r5
-	.align	2
-L_al4both_loop:
-	mov.l	@(r0,r5),r1
-	cmp/hs	r7,r0
-	bt/s	L_al4both_loop
+	! copy a byte at once
+7:	mov	r4,r2
+	add	#1,r2
+8:
+	cmp/hi	r2,r0
+	mov.b	@(r0,r5),r1
+	bt/s	8b			! while (r0>r2)
+	 mov.b	r1,@-r0
+9:
+	rts
+	 nop
+
+case0:
+	!
+	!	GHIJ KLMN OPQR -->  GHIJ KLMN OPQR
+	!
+	! First, align to long word boundary
+	mov	r0,r3
+	and	r2,r3
+	tst	r3,r3
+	bt/s	2f
+	 add	#-4,r5
+	add	#3,r5
+1:	dt	r3
+	mov.b	@(r0,r5),r1
+	bf/s	1b
+	 mov.b	r1,@-r0
+	!
+	add	#-3,r5
+2:	! Second, copy a long word at once
+	mov	r4,r2
+	add	#7,r2
+3:	mov.l	@(r0,r5),r1
+	cmp/hi	r2,r0
+	bt/s	3b
 	 mov.l	r1,@-r0
-	bra	L_cleanup
+	!
+	! Third, copy a byte at once, if necessary
+	cmp/eq	r4,r0
+	bt/s	9b
 	 add	#3,r5
+	bra	8b
+	 add	#-6,r2
 
-	nop ! avoid nop in executed code.
-L_odddst:
-	shlr	r3
-	bt	L_al4src
-	mov.w	@(r0,r5),r1
-	mov.b	r1,@-r0
-	shlr8	r1
-	mov.b	r1,@-r0
-L_al4src:
-	add	#-2,r5
-	.align	2
-L_odd_loop:
-	mov.l	@(r0,r5),r2
-	cmp/hs	r7,r0
-	mov.b	r2,@-r0
-	shlr8	r2
-	mov.w	r2,@-r0
-	shlr16	r2
-	mov.b	r2,@-r0
-	bt	L_odd_loop
-
-	add	#3,r5
-L_cleanup:
-L_small:
+case1:
+	!
+	!	GHIJ KLMN OPQR -->  ...G HIJK LMNO PQR.
+	!
+	! First, align to long word boundary
+	mov	r0,r3
+	and	r2,r3
+	tst	r3,r3
+	bt/s	2f
+	 add	#-1,r5
+1:	dt	r3
+	mov.b	@(r0,r5),r1
+	bf/s	1b
+	 mov.b	r1,@-r0
+	!
+2:	! Second, read a long word and write a long word at once
+	mov.l	@(r0,r5),r1
+	add	#-4,r5
+	mov	r4,r2
+	add	#7,r2
+	!
+#ifdef __LITTLE_ENDIAN__
+3:	mov	r1,r3		! RQPO
+	shll16	r3
+	shll8	r3		! Oxxx
+	mov.l	@(r0,r5),r1	! NMLK
+	mov	r1,r6
+	shlr8	r6		! xNML
+	or	r6,r3		! ONML
+	cmp/hi	r2,r0
+	bt/s	3b
+	 mov.l	r3,@-r0
+#else
+3:	mov	r1,r3		! OPQR
+	shlr16	r3
+	shlr8	r3		! xxxO
+	mov.l	@(r0,r5),r1	! KLMN
+	mov	r1,r6
+	shll8	r6		! LMNx
+	or	r6,r3		! LMNO
+	cmp/hi	r2,r0
+	bt/s	3b
+	 mov.l	r3,@-r0
+#endif
+	!
+	! Third, copy a byte at once, if necessary
 	cmp/eq	r4,r0
-	bt	L_ready
-	add	#1,r4
-	.align	2
-L_cleanup_loop:
-	mov.b	@(r0,r5),r2
+	bt/s	9b
+	 add	#4,r5
+	bra	8b
+	 add	#-6,r2
+
+case2:
+	!
+	!	GHIJ KLMN OPQR -->  ..GH IJKL MNOP QR..
+	!
+	! First, align to word boundary
+	tst	#1,r0
+	bt/s	2f
+	 add	#-1,r5
+	mov.b	@(r0,r5),r1
+	mov.b	r1,@-r0
+	!
+2:	! Second, read a word and write a word at once
+	add	#-1,r5
+	mov	r4,r2
+	add	#3,r2
+	!
+3:	mov.w	@(r0,r5),r1
+	cmp/hi	r2,r0
+	bt/s	3b
+	 mov.w	r1,@-r0
+	!
+	! Third, copy a byte at once, if necessary
 	cmp/eq	r4,r0
-	mov.b	r2,@-r0
-	bf	L_cleanup_loop
-L_ready:
+	bt/s	9b
+	 add	#1,r5
+	mov.b	@(r0,r5),r1
 	rts
-	 nop
+	 mov.b	r1,@-r0
+
+case3:
+	!
+	!	GHIJ KLMN OPQR -->  .GHI JKLM NOPQ R...
+	!
+	! First, align to long word boundary
+	mov	r0,r3
+	and	r2,r3
+	tst	r3,r3
+	bt/s	2f
+	 add	#-1,r5
+1:	dt	r3
+	mov.b	@(r0,r5),r1
+	bf/s	1b
+	 mov.b	r1,@-r0
+	!
+2:	! Second, read a long word and write a long word at once
+	add	#-2,r5
+	mov.l	@(r0,r5),r1
+	add	#-4,r5
+	mov	r4,r2
+	add	#7,r2
+	!
+#ifdef __LITTLE_ENDIAN__
+3:	mov	r1,r3		! RQPO
+	shll8	r3		! QPOx
+	mov.l	@(r0,r5),r1	! NMLK
+	mov	r1,r6
+	shlr16	r6
+	shlr8	r6		! xxxN
+	or	r6,r3		! QPON
+	cmp/hi	r2,r0
+	bt/s	3b
+	 mov.l	r3,@-r0
+#else
+3:	mov	r1,r3		! OPQR
+	shlr8	r3		! xOPQ
+	mov.l	@(r0,r5),r1	! KLMN
+	mov	r1,r6
+	shll16	r6
+	shll8	r6		! Nxxx
+	or	r6,r3		! NOPQ
+	cmp/hi	r2,r0
+	bt/s	3b
+	 mov.l	r3,@-r0
+#endif
+	!
+	! Third, copy a byte at once, if necessary
+	cmp/eq	r4,r0
+	bt/s	9b
+	 add	#6,r5
+	bra	8b
+	 add	#-6,r2