[PATCH] xtensa: Architecture support for Tensilica Xtensa Part 4

The attached patches provides part 4 of an architecture implementation for the
Tensilica Xtensa CPU series.

Signed-off-by: Chris Zankel <chris@zankel.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

9 files changed, 1968 insertions, 0 deletions

diff --git a/arch/xtensa/lib/Makefile b/arch/xtensa/lib/Makefile
new file mode 100644
index 0000000..ed935b5
--- /dev/null
+++ b/arch/xtensa/lib/Makefile
@@ -0,0 +1,7 @@
+#
+# Makefile for Xtensa-specific library files.
+#
+
+lib-y	+= memcopy.o memset.o checksum.o strcasecmp.o \
+	   usercopy.o strncpy_user.o strnlen_user.o
+lib-$(CONFIG_PCI) += pci-auto.o
diff --git a/arch/xtensa/lib/checksum.S b/arch/xtensa/lib/checksum.S
new file mode 100644
index 0000000..e2d64df
--- /dev/null
+++ b/arch/xtensa/lib/checksum.S
@@ -0,0 +1,410 @@
+/*
+ * INET		An implementation of the TCP/IP protocol suite for the LINUX
+ *		operating system.  INET is implemented using the  BSD Socket
+ *		interface as the means of communication with the user level.
+ *
+ *		IP/TCP/UDP checksumming routines
+ *
+ * Xtensa version:  Copyright (C) 2001 Tensilica, Inc. by Kevin Chea
+ *                  Optimized by Joe Taylor
+ *
+ *		This program is free software; you can redistribute it and/or
+ *		modify it under the terms of the GNU General Public License
+ *		as published by the Free Software Foundation; either version
+ *		2 of the License, or (at your option) any later version.
+ */
+
+#include <asm/errno.h>
+#include <linux/linkage.h>
+#define _ASMLANGUAGE
+#include <xtensa/config/core.h>
+
+/*
+ * computes a partial checksum, e.g. for TCP/UDP fragments
+ */
+
+/*
+ * unsigned int csum_partial(const unsigned char *buf, int len,
+ *                           unsigned int sum);
+ *    a2 = buf
+ *    a3 = len
+ *    a4 = sum
+ *
+ * This function assumes 2- or 4-byte alignment.  Other alignments will fail!
+ */
+
+/* ONES_ADD converts twos-complement math to ones-complement. */
+#define ONES_ADD(sum, val)	  \
+	add	sum, sum, val	; \
+	bgeu	sum, val, 99f	; \
+	addi	sum, sum, 1	; \
+99:				;
+
+.text
+ENTRY(csum_partial)
+	  /*
+	   * Experiments with Ethernet and SLIP connections show that buf
+	   * is aligned on either a 2-byte or 4-byte boundary.
+	   */
+	entry	sp, 32
+	extui	a5, a2, 0, 2
+	bnez	a5, 8f		/* branch if 2-byte aligned */
+	/* Fall-through on common case, 4-byte alignment */
+1:
+	srli	a5, a3, 5	/* 32-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a5, 2f
+#else
+	beqz	a5, 2f
+	slli	a5, a5, 5
+	add	a5, a5, a2	/* a5 = end of last 32-byte chunk */
+.Loop1:
+#endif
+	l32i	a6, a2, 0
+	l32i	a7, a2, 4
+	ONES_ADD(a4, a6)
+	ONES_ADD(a4, a7)
+	l32i	a6, a2, 8
+	l32i	a7, a2, 12
+	ONES_ADD(a4, a6)
+	ONES_ADD(a4, a7)
+	l32i	a6, a2, 16
+	l32i	a7, a2, 20
+	ONES_ADD(a4, a6)
+	ONES_ADD(a4, a7)
+	l32i	a6, a2, 24
+	l32i	a7, a2, 28
+	ONES_ADD(a4, a6)
+	ONES_ADD(a4, a7)
+	addi	a2, a2, 4*8
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a5, .Loop1
+#endif
+2:
+	extui	a5, a3, 2, 3	/* remaining 4-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a5, 3f
+#else
+	beqz	a5, 3f
+	slli	a5, a5, 2
+	add	a5, a5, a2	/* a5 = end of last 4-byte chunk */
+.Loop2:
+#endif
+	l32i	a6, a2, 0
+	ONES_ADD(a4, a6)
+	addi	a2, a2, 4
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a5, .Loop2
+#endif
+3:
+	_bbci.l	a3, 1, 5f	/* remaining 2-byte chunk */
+	l16ui	a6, a2, 0
+	ONES_ADD(a4, a6)
+	addi	a2, a2, 2
+5:
+	_bbci.l	a3, 0, 7f	/* remaining 1-byte chunk */
+6:	l8ui	a6, a2, 0
+#ifdef __XTENSA_EB__
+	slli	a6, a6, 8	/* load byte into bits 8..15 */
+#endif
+	ONES_ADD(a4, a6)
+7:
+	mov	a2, a4
+	retw
+
+	/* uncommon case, buf is 2-byte aligned */
+8:
+	beqz	a3, 7b		/* branch if len == 0 */
+	beqi	a3, 1, 6b	/* branch if len == 1 */
+
+	extui	a5, a2, 0, 1
+	bnez	a5, 8f		/* branch if 1-byte aligned */
+
+	l16ui	a6, a2, 0	/* common case, len >= 2 */
+	ONES_ADD(a4, a6)
+	addi	a2, a2, 2	/* adjust buf */
+	addi	a3, a3, -2	/* adjust len */
+	j	1b		/* now buf is 4-byte aligned */
+
+	/* case: odd-byte aligned, len > 1
+	 * This case is dog slow, so don't give us an odd address.
+	 * (I don't think this ever happens, but just in case.)
+	 */
+8:
+	srli	a5, a3, 2	/* 4-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a5, 2f
+#else
+	beqz	a5, 2f
+	slli	a5, a5, 2
+	add	a5, a5, a2	/* a5 = end of last 4-byte chunk */
+.Loop3:
+#endif
+	l8ui	a6, a2, 0	/* bits 24..31 */
+	l16ui	a7, a2, 1	/* bits  8..23 */
+	l8ui	a8, a2, 3	/* bits  0.. 8 */
+#ifdef	__XTENSA_EB__
+	slli	a6, a6, 24
+#else
+	slli	a8, a8, 24
+#endif
+	slli	a7, a7, 8
+	or	a7, a7, a6
+	or	a7, a7, a8
+	ONES_ADD(a4, a7)
+	addi	a2, a2, 4
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a5, .Loop3
+#endif
+2:
+	_bbci.l	a3, 1, 3f	/* remaining 2-byte chunk, still odd addr */
+	l8ui	a6, a2, 0
+	l8ui	a7, a2, 1
+#ifdef	__XTENSA_EB__
+	slli	a6, a6, 8
+#else
+	slli	a7, a7, 8
+#endif
+	or	a7, a7, a6
+	ONES_ADD(a4, a7)
+	addi	a2, a2, 2
+3:
+	j	5b		/* branch to handle the remaining byte */
+
+
+
+/*
+ * Copy from ds while checksumming, otherwise like csum_partial
+ *
+ * The macros SRC and DST specify the type of access for the instruction.
+ * thus we can call a custom exception handler for each access type.
+ */
+
+#define SRC(y...)			\
+	9999: y;			\
+	.section __ex_table, "a";	\
+	.long 9999b, 6001f	;	\
+	.previous
+
+#define DST(y...)			\
+	9999: y;			\
+	.section __ex_table, "a";	\
+	.long 9999b, 6002f	;	\
+	.previous
+
+/*
+unsigned int csum_partial_copy_generic (const char *src, char *dst, int len,
+					int sum, int *src_err_ptr, int *dst_err_ptr)
+	a2  = src
+	a3  = dst
+	a4  = len
+	a5  = sum
+	a6  = src_err_ptr
+	a7  = dst_err_ptr
+	a8  = temp
+	a9  = temp
+	a10 = temp
+	a11 = original len for exception handling
+	a12 = original dst for exception handling
+
+    This function is optimized for 4-byte aligned addresses.  Other
+    alignments work, but not nearly as efficiently.
+ */
+
+ENTRY(csum_partial_copy_generic)
+	entry	sp, 32
+	mov	a12, a3
+	mov	a11, a4
+	or	a10, a2, a3
+
+	/* We optimize the following alignment tests for the 4-byte
+	aligned case.  Two bbsi.l instructions might seem more optimal
+	(commented out below).  However, both labels 5: and 3: are out
+	of the imm8 range, so the assembler relaxes them into
+	equivalent bbci.l, j combinations, which is actually
+	slower. */
+
+	extui	a9, a10, 0, 2
+	beqz	a9, 1f		/* branch if both are 4-byte aligned */
+	bbsi.l	a10, 0, 5f	/* branch if one address is odd */
+	j	3f		/* one address is 2-byte aligned */
+
+/*	_bbsi.l	a10, 0, 5f */	/* branch if odd address */
+/*	_bbsi.l	a10, 1, 3f */	/* branch if 2-byte-aligned address */
+
+1:
+	/* src and dst are both 4-byte aligned */
+	srli	a10, a4, 5	/* 32-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a10, 2f
+#else
+	beqz	a10, 2f
+	slli	a10, a10, 5
+	add	a10, a10, a2	/* a10 = end of last 32-byte src chunk */
+.Loop5:
+#endif
+SRC(	l32i	a9, a2, 0	)
+SRC(	l32i	a8, a2, 4	)
+DST(	s32i	a9, a3, 0	)
+DST(	s32i	a8, a3, 4	)
+	ONES_ADD(a5, a9)
+	ONES_ADD(a5, a8)
+SRC(	l32i	a9, a2, 8	)
+SRC(	l32i	a8, a2, 12	)
+DST(	s32i	a9, a3, 8	)
+DST(	s32i	a8, a3, 12	)
+	ONES_ADD(a5, a9)
+	ONES_ADD(a5, a8)
+SRC(	l32i	a9, a2, 16	)
+SRC(	l32i	a8, a2, 20	)
+DST(	s32i	a9, a3, 16	)
+DST(	s32i	a8, a3, 20	)
+	ONES_ADD(a5, a9)
+	ONES_ADD(a5, a8)
+SRC(	l32i	a9, a2, 24	)
+SRC(	l32i	a8, a2, 28	)
+DST(	s32i	a9, a3, 24	)
+DST(	s32i	a8, a3, 28	)
+	ONES_ADD(a5, a9)
+	ONES_ADD(a5, a8)
+	addi	a2, a2, 32
+	addi	a3, a3, 32
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a10, .Loop5
+#endif
+2:
+	extui	a10, a4, 2, 3	/* remaining 4-byte chunks */
+	extui	a4, a4, 0, 2	/* reset len for general-case, 2-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a10, 3f
+#else
+	beqz	a10, 3f
+	slli	a10, a10, 2
+	add	a10, a10, a2	/* a10 = end of last 4-byte src chunk */
+.Loop6:
+#endif
+SRC(	l32i	a9, a2, 0	)
+DST(	s32i	a9, a3, 0	)
+	ONES_ADD(a5, a9)
+	addi	a2, a2, 4
+	addi	a3, a3, 4
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a10, .Loop6
+#endif
+3:
+	/*
+	Control comes to here in two cases: (1) It may fall through
+	to here from the 4-byte alignment case to process, at most,
+	one 2-byte chunk.  (2) It branches to here from above if
+	either src or dst is 2-byte aligned, and we process all bytes
+	here, except for perhaps a trailing odd byte.  It's
+	inefficient, so align your addresses to 4-byte boundaries.
+
+	a2 = src
+	a3 = dst
+	a4 = len
+	a5 = sum
+	*/
+	srli	a10, a4, 1	/* 2-byte chunks */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a10, 4f
+#else
+	beqz	a10, 4f
+	slli	a10, a10, 1
+	add	a10, a10, a2	/* a10 = end of last 2-byte src chunk */
+.Loop7:
+#endif
+SRC(	l16ui	a9, a2, 0	)
+DST(	s16i	a9, a3, 0	)
+	ONES_ADD(a5, a9)
+	addi	a2, a2, 2
+	addi	a3, a3, 2
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a10, .Loop7
+#endif
+4:
+	/* This section processes a possible trailing odd byte. */
+	_bbci.l	a4, 0, 8f	/* 1-byte chunk */
+SRC(	l8ui	a9, a2, 0	)
+DST(	s8i	a9, a3, 0	)
+#ifdef __XTENSA_EB__
+	slli	a9, a9, 8	/* shift byte to bits 8..15 */
+#endif
+	ONES_ADD(a5, a9)
+8:
+	mov	a2, a5
+	retw
+
+5:
+	/* Control branch to here when either src or dst is odd.  We
+	process all bytes using 8-bit accesses.  Grossly inefficient,
+	so don't feed us an odd address. */
+
+	srli	a10, a4, 1	/* handle in pairs for 16-bit csum */
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a10, 6f
+#else
+	beqz	a10, 6f
+	slli	a10, a10, 1
+	add	a10, a10, a2	/* a10 = end of last odd-aligned, 2-byte src chunk */
+.Loop8:
+#endif
+SRC(	l8ui	a9, a2, 0	)
+SRC(	l8ui	a8, a2, 1	)
+DST(	s8i	a9, a3, 0	)
+DST(	s8i	a8, a3, 1	)
+#ifdef __XTENSA_EB__
+	slli	a9, a9, 8	/* combine into a single 16-bit value */
+#else				/* for checksum computation */
+	slli	a8, a8, 8
+#endif
+	or	a9, a9, a8
+	ONES_ADD(a5, a9)
+	addi	a2, a2, 2
+	addi	a3, a3, 2
+#if !XCHAL_HAVE_LOOPS
+	blt	a2, a10, .Loop8
+#endif
+6:
+	j	4b		/* process the possible trailing odd byte */
+
+
+# Exception handler:
+.section .fixup, "ax"
+/*
+	a6  = src_err_ptr
+	a7  = dst_err_ptr
+	a11 = original len for exception handling
+	a12 = original dst for exception handling
+*/
+
+6001:
+	_movi	a2, -EFAULT
+	s32i	a2, a6, 0	/* src_err_ptr */
+
+	# clear the complete destination - computing the rest
+	# is too much work
+	movi	a2, 0
+#if XCHAL_HAVE_LOOPS
+	loopgtz	a11, 2f
+#else
+	beqz	a11, 2f
+	add	a11, a11, a12	/* a11 = ending address */
+.Leloop:
+#endif
+	s8i	a2, a12, 0
+	addi	a12, a12, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a12, a11, .Leloop
+#endif
+2:
+	retw
+
+6002:
+	movi	a2, -EFAULT
+	s32i	a2, a7, 0	/* dst_err_ptr */
+	movi	a2, 0
+	retw
+
+.previous
+
diff --git a/arch/xtensa/lib/memcopy.S b/arch/xtensa/lib/memcopy.S
new file mode 100644
index 0000000..e8f6d7e
--- /dev/null
+++ b/arch/xtensa/lib/memcopy.S
@@ -0,0 +1,315 @@
+/*
+ * arch/xtensa/lib/hal/memcopy.S -- Core HAL library functions
+ * xthal_memcpy and xthal_bcopy
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2002 - 2005 Tensilica Inc.
+ */
+
+#include <xtensa/coreasm.h>
+
+	.macro	src_b	r, w0, w1
+#ifdef __XTENSA_EB__
+	src	\r, \w0, \w1
+#else
+	src	\r, \w1, \w0
+#endif
+	.endm
+
+	.macro	ssa8	r
+#ifdef __XTENSA_EB__
+	ssa8b	\r
+#else
+	ssa8l	\r
+#endif
+	.endm
+
+
+/*
+ * void *memcpy(void *dst, const void *src, size_t len);
+ * void *memmove(void *dst, const void *src, size_t len);
+ * void *bcopy(const void *src, void *dst, size_t len);
+ *
+ * This function is intended to do the same thing as the standard
+ * library function memcpy() (or bcopy()) for most cases.
+ * However, where the source and/or destination references
+ * an instruction RAM or ROM or a data RAM or ROM, that
+ * source and/or destination will always be accessed with
+ * 32-bit load and store instructions (as required for these
+ * types of devices).
+ *
+ * !!!!!!!  XTFIXME:
+ * !!!!!!!  Handling of IRAM/IROM has not yet
+ * !!!!!!!  been implemented.
+ *
+ * The bcopy version is provided here to avoid the overhead
+ * of an extra call, for callers that require this convention.
+ *
+ * The (general case) algorithm is as follows:
+ *   If destination is unaligned, align it by conditionally
+ *     copying 1 and 2 bytes.
+ *   If source is aligned,
+ *     do 16 bytes with a loop, and then finish up with
+ *     8, 4, 2, and 1 byte copies conditional on the length;
+ *   else (if source is unaligned),
+ *     do the same, but use SRC to align the source data.
+ *   This code tries to use fall-through branches for the common
+ *     case of aligned source and destination and multiple
+ *     of 4 (or 8) length.
+ *
+ * Register use:
+ *	a0/ return address
+ *	a1/ stack pointer
+ *	a2/ return value
+ *	a3/ src
+ *	a4/ length
+ *	a5/ dst
+ *	a6/ tmp
+ *	a7/ tmp
+ *	a8/ tmp
+ *	a9/ tmp
+ *	a10/ tmp
+ *	a11/ tmp
+ */
+
+	.text
+	.align	4
+	.global	bcopy
+	.type   bcopy,@function
+bcopy:
+	entry	sp, 16		# minimal stack frame
+	# a2=src, a3=dst, a4=len
+	mov	a5, a3		# copy dst so that a2 is return value
+	mov	a3, a2
+	mov	a2, a5
+	j	.Lcommon	# go to common code for memcpy+bcopy
+
+
+/*
+ * Byte by byte copy
+ */
+	.align	4
+	.byte	0		# 1 mod 4 alignment for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lbytecopy:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, .Lbytecopydone
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a4, .Lbytecopydone
+	add	a7, a3, a4	# a7 = end address for source
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lnextbyte:
+	l8ui	a6, a3, 0
+	addi	a3, a3, 1
+	s8i	a6, a5, 0
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a7, .Lnextbyte
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lbytecopydone:
+	retw
+
+/*
+ * Destination is unaligned
+ */
+
+	.align	4
+.Ldst1mod2:	# dst is only byte aligned
+	_bltui	a4, 7, .Lbytecopy	# do short copies byte by byte
+
+	# copy 1 byte
+	l8ui	a6, a3,  0
+	addi	a3, a3,  1
+	addi	a4, a4, -1
+	s8i	a6, a5,  0
+	addi	a5, a5,  1
+	_bbci.l	a5, 1, .Ldstaligned	# if dst is now aligned, then
+					# return to main algorithm
+.Ldst2mod4:	# dst 16-bit aligned
+	# copy 2 bytes
+	_bltui	a4, 6, .Lbytecopy	# do short copies byte by byte
+	l8ui	a6, a3,  0
+	l8ui	a7, a3,  1
+	addi	a3, a3,  2
+	addi	a4, a4, -2
+	s8i	a6, a5,  0
+	s8i	a7, a5,  1
+	addi	a5, a5,  2
+	j	.Ldstaligned	# dst is now aligned, return to main algorithm
+
+	.align	4
+	.global	memcpy
+	.type   memcpy,@function
+memcpy:
+	.global	memmove
+	.type   memmove,@function
+memmove:
+
+	entry	sp, 16		# minimal stack frame
+	# a2/ dst, a3/ src, a4/ len
+	mov	a5, a2		# copy dst so that a2 is return value
+.Lcommon:
+	_bbsi.l	a2, 0, .Ldst1mod2	# if dst is 1 mod 2
+	_bbsi.l	a2, 1, .Ldst2mod4	# if dst is 2 mod 4
+.Ldstaligned:	# return here from .Ldst?mod? once dst is aligned
+	srli	a7, a4, 4	# number of loop iterations with 16B
+				# per iteration
+	movi	a8, 3		# if source is not aligned,
+	_bany	a3, a8, .Lsrcunaligned	# then use shifting copy
+	/*
+	 * Destination and source are word-aligned, use word copy.
+	 */
+	# copy 16 bytes per iteration for word-aligned dst and word-aligned src
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop1done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop1done
+	slli	a8, a7, 4
+	add	a8, a8, a3	# a8 = end of last 16B source chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1:
+	l32i	a6, a3,  0
+	l32i	a7, a3,  4
+	s32i	a6, a5,  0
+	l32i	a6, a3,  8
+	s32i	a7, a5,  4
+	l32i	a7, a3, 12
+	s32i	a6, a5,  8
+	addi	a3, a3, 16
+	s32i	a7, a5, 12
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a8, .Loop1
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1done:
+	bbci.l	a4, 3, .L2
+	# copy 8 bytes
+	l32i	a6, a3,  0
+	l32i	a7, a3,  4
+	addi	a3, a3,  8
+	s32i	a6, a5,  0
+	s32i	a7, a5,  4
+	addi	a5, a5,  8
+.L2:
+	bbsi.l	a4, 2, .L3
+	bbsi.l	a4, 1, .L4
+	bbsi.l	a4, 0, .L5
+	retw
+.L3:
+	# copy 4 bytes
+	l32i	a6, a3,  0
+	addi	a3, a3,  4
+	s32i	a6, a5,  0
+	addi	a5, a5,  4
+	bbsi.l	a4, 1, .L4
+	bbsi.l	a4, 0, .L5
+	retw
+.L4:
+	# copy 2 bytes
+	l16ui	a6, a3,  0
+	addi	a3, a3,  2
+	s16i	a6, a5,  0
+	addi	a5, a5,  2
+	bbsi.l	a4, 0, .L5
+	retw
+.L5:
+	# copy 1 byte
+	l8ui	a6, a3,  0
+	s8i	a6, a5,  0
+	retw
+
+/*
+ * Destination is aligned, Source is unaligned
+ */
+
+	.align	4
+.Lsrcunaligned:
+	_beqz	a4, .Ldone	# avoid loading anything for zero-length copies
+	# copy 16 bytes per iteration for word-aligned dst and unaligned src
+	ssa8	a3		# set shift amount from byte offset
+#define SIM_CHECKS_ALIGNMENT	1	/* set to 1 when running on ISS (simulator) with the
+					   lint or ferret client, or 0 to save a few cycles */
+#if XCHAL_UNALIGNED_LOAD_EXCEPTION || SIM_CHECKS_ALIGNMENT
+	and	a11, a3, a8	# save unalignment offset for below
+	sub	a3, a3, a11	# align a3
+#endif
+	l32i	a6, a3, 0	# load first word
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop2done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop2done
+	slli	a10, a7, 4
+	add	a10, a10, a3	# a10 = end of last 16B source chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2:
+	l32i	a7, a3,  4
+	l32i	a8, a3,  8
+	src_b	a6, a6, a7
+	s32i	a6, a5,  0
+	l32i	a9, a3, 12
+	src_b	a7, a7, a8
+	s32i	a7, a5,  4
+	l32i	a6, a3, 16
+	src_b	a8, a8, a9
+	s32i	a8, a5,  8
+	addi	a3, a3, 16
+	src_b	a9, a9, a6
+	s32i	a9, a5, 12
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a10, .Loop2
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2done:
+	bbci.l	a4, 3, .L12
+	# copy 8 bytes
+	l32i	a7, a3,  4
+	l32i	a8, a3,  8
+	src_b	a6, a6, a7
+	s32i	a6, a5,  0
+	addi	a3, a3,  8
+	src_b	a7, a7, a8
+	s32i	a7, a5,  4
+	addi	a5, a5,  8
+	mov	a6, a8
+.L12:
+	bbci.l	a4, 2, .L13
+	# copy 4 bytes
+	l32i	a7, a3,  4
+	addi	a3, a3,  4
+	src_b	a6, a6, a7
+	s32i	a6, a5,  0
+	addi	a5, a5,  4
+	mov	a6, a7
+.L13:
+#if XCHAL_UNALIGNED_LOAD_EXCEPTION || SIM_CHECKS_ALIGNMENT
+	add	a3, a3, a11	# readjust a3 with correct misalignment
+#endif
+	bbsi.l	a4, 1, .L14
+	bbsi.l	a4, 0, .L15
+.Ldone:	retw
+.L14:
+	# copy 2 bytes
+	l8ui	a6, a3,  0
+	l8ui	a7, a3,  1
+	addi	a3, a3,  2
+	s8i	a6, a5,  0
+	s8i	a7, a5,  1
+	addi	a5, a5,  2
+	bbsi.l	a4, 0, .L15
+	retw
+.L15:
+	# copy 1 byte
+	l8ui	a6, a3,  0
+	s8i	a6, a5,  0
+	retw
+
+/*
+ * Local Variables:
+ * mode:fundamental
+ * comment-start: "# "
+ * comment-start-skip: "# *"
+ * End:
+ */
diff --git a/arch/xtensa/lib/memset.S b/arch/xtensa/lib/memset.S
new file mode 100644
index 0000000..4de2513
--- /dev/null
+++ b/arch/xtensa/lib/memset.S
@@ -0,0 +1,160 @@
+/*
+ *  arch/xtensa/lib/memset.S
+ *
+ *  ANSI C standard library function memset
+ *  (Well, almost.  .fixup code might return zero.)
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+#include <xtensa/coreasm.h>
+
+/*
+ * void *memset(void *dst, int c, size_t length)
+ *
+ * The algorithm is as follows:
+ *   Create a word with c in all byte positions
+ *   If the destination is aligned,
+ *     do 16B chucks with a loop, and then finish up with
+ *     8B, 4B, 2B, and 1B stores conditional on the length.
+ *   If destination is unaligned, align it by conditionally
+ *     setting 1B and 2B and then go to aligned case.
+ *   This code tries to use fall-through branches for the common
+ *     case of an aligned destination (except for the branches to
+ *     the alignment labels).
+ */
+
+/* Load or store instructions that may cause exceptions use the EX macro. */
+
+#define EX(insn,reg1,reg2,offset,handler)	\
+9:	insn	reg1, reg2, offset;		\
+	.section __ex_table, "a";		\
+	.word	9b, handler;			\
+	.previous
+
+
+.text
+.align	4
+.global	memset
+.type	memset,@function
+memset:
+	entry	sp, 16		# minimal stack frame
+	# a2/ dst, a3/ c, a4/ length
+	extui	a3, a3, 0, 8	# mask to just 8 bits
+	slli	a7, a3, 8	# duplicate character in all bytes of word
+	or	a3, a3, a7	# ...
+	slli	a7, a3, 16	# ...
+	or	a3, a3, a7	# ...
+	mov	a5, a2		# copy dst so that a2 is return value
+	movi	a6, 3		# for alignment tests
+	bany	a2, a6, .Ldstunaligned # if dst is unaligned
+.L0:	# return here from .Ldstunaligned when dst is aligned
+	srli	a7, a4, 4	# number of loop iterations with 16B
+				# per iteration
+	bnez	a4, .Laligned
+	retw
+
+/*
+ * Destination is word-aligned.
+ */
+	# set 16 bytes per iteration for word-aligned dst
+	.align	4		# 1 mod 4 alignment for LOOPNEZ
+	.byte	0		# (0 mod 4 alignment for LBEG)
+.Laligned:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop1done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop1done
+	slli	a6, a7, 4
+	add	a6, a6, a5	# a6 = end of last 16B chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1:
+	EX(s32i, a3, a5,  0, memset_fixup)
+	EX(s32i, a3, a5,  4, memset_fixup)
+	EX(s32i, a3, a5,  8, memset_fixup)
+	EX(s32i, a3, a5, 12, memset_fixup)
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a5, a6, .Loop1
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1done:
+	bbci.l	a4, 3, .L2
+	# set 8 bytes
+	EX(s32i, a3, a5,  0, memset_fixup)
+	EX(s32i, a3, a5,  4, memset_fixup)
+	addi	a5, a5,  8
+.L2:
+	bbci.l	a4, 2, .L3
+	# set 4 bytes
+	EX(s32i, a3, a5,  0, memset_fixup)
+	addi	a5, a5,  4
+.L3:
+	bbci.l	a4, 1, .L4
+	# set 2 bytes
+	EX(s16i, a3, a5,  0, memset_fixup)
+	addi	a5, a5,  2
+.L4:
+	bbci.l	a4, 0, .L5
+	# set 1 byte
+	EX(s8i, a3, a5,  0, memset_fixup)
+.L5:
+.Lret1:
+	retw
+
+/*
+ * Destination is unaligned
+ */
+
+.Ldstunaligned:
+	bltui	a4, 8, .Lbyteset	# do short copies byte by byte
+	bbci.l	a5, 0, .L20		# branch if dst alignment half-aligned
+	# dst is only byte aligned
+	# set 1 byte
+	EX(s8i, a3, a5,  0, memset_fixup)
+	addi	a5, a5,  1
+	addi	a4, a4, -1
+	# now retest if dst aligned
+	bbci.l	a5, 1, .L0	# if now aligned, return to main algorithm
+.L20:
+	# dst half-aligned
+	# set 2 bytes
+	EX(s16i, a3, a5,  0, memset_fixup)
+	addi	a5, a5,  2
+	addi	a4, a4, -2
+	j	.L0		# dst is now aligned, return to main algorithm
+
+/*
+ * Byte by byte set
+ */
+	.align	4
+	.byte	0		# 1 mod 4 alignment for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lbyteset:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, .Lbytesetdone
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a4, .Lbytesetdone
+	add	a6, a5, a4	# a6 = ending address
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lbyteloop:
+	EX(s8i, a3, a5, 0, memset_fixup)
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a5, a6, .Lbyteloop
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lbytesetdone:
+	retw
+
+
+	.section .fixup, "ax"
+	.align	4
+
+/* We return zero if a failure occurred. */
+
+memset_fixup:
+	movi	a2, 0
+	retw
diff --git a/arch/xtensa/lib/pci-auto.c b/arch/xtensa/lib/pci-auto.c
new file mode 100644
index 0000000..90c790f
--- /dev/null
+++ b/arch/xtensa/lib/pci-auto.c
@@ -0,0 +1,352 @@
+/*
+ * arch/xtensa/kernel/pci-auto.c
+ *
+ * PCI autoconfiguration library
+ *
+ * Copyright (C) 2001 - 2005 Tensilica Inc.
+ *
+ * Chris Zankel <zankel@tensilica.com, cez@zankel.net>
+ *
+ * Based on work from Matt Porter <mporter@mvista.com>
+ *
+ * This program is free software; you can redistribute  it and/or modify it
+ * under  the terms of  the GNU General  Public License as published by the
+ * Free Software Foundation;  either version 2 of the  License, or (at your
+ * option) any later version.
+ */
+
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/pci.h>
+
+#include <asm/pci-bridge.h>
+
+
+/*
+ *
+ * Setting up a PCI
+ *
+ * pci_ctrl->first_busno = <first bus number (0)>
+ * pci_ctrl->last_busno = <last bus number (0xff)>
+ * pci_ctrl->ops = <PCI config operations>
+ * pci_ctrl->map_irq = <function to return the interrupt number for a device>
+ *
+ * pci_ctrl->io_space.start = <IO space start address (PCI view)>
+ * pci_ctrl->io_space.end = <IO space end address (PCI view)>
+ * pci_ctrl->io_space.base = <IO space offset: address 0 from CPU space>
+ * pci_ctrl->mem_space.start = <MEM space start address (PCI view)>
+ * pci_ctrl->mem_space.end = <MEM space end address (PCI view)>
+ * pci_ctrl->mem_space.base = <MEM space offset: address 0 from CPU space>
+ *
+ * pcibios_init_resource(&pci_ctrl->io_resource, <IO space start>,
+ * 			 <IO space end>, IORESOURCE_IO, "PCI host bridge");
+ * pcibios_init_resource(&pci_ctrl->mem_resources[0], <MEM space start>,
+ * 			 <MEM space end>, IORESOURCE_MEM, "PCI host bridge");
+ *
+ * pci_ctrl->last_busno = pciauto_bus_scan(pci_ctrl,pci_ctrl->first_busno);
+ *
+ * int __init pciauto_bus_scan(struct pci_controller *pci_ctrl, int current_bus)
+ *
+ */
+
+
+/* define DEBUG to print some debugging messages. */
+
+#undef DEBUG
+
+#ifdef DEBUG
+# define DBG(x...) printk(x)
+#else
+# define DBG(x...)
+#endif
+
+static int pciauto_upper_iospc;
+static int pciauto_upper_memspc;
+
+static struct pci_dev pciauto_dev;
+static struct pci_bus pciauto_bus;
+
+/*
+ * Helper functions
+ */
+
+/* Initialize the bars of a PCI device.  */
+
+static void __init
+pciauto_setup_bars(struct pci_dev *dev, int bar_limit)
+{
+	int bar_size;
+	int bar, bar_nr;
+	int *upper_limit;
+	int found_mem64 = 0;
+
+	for (bar = PCI_BASE_ADDRESS_0, bar_nr = 0;
+	     bar <= bar_limit;
+	     bar+=4, bar_nr++)
+	{
+		/* Tickle the BAR and get the size */
+		pci_write_config_dword(dev, bar, 0xffffffff);
+		pci_read_config_dword(dev, bar, &bar_size);
+
+		/* If BAR is not implemented go to the next BAR */
+		if (!bar_size)
+			continue;
+
+		/* Check the BAR type and set our address mask */
+		if (bar_size & PCI_BASE_ADDRESS_SPACE_IO)
+		{
+			bar_size &= PCI_BASE_ADDRESS_IO_MASK;
+			upper_limit = &pciauto_upper_iospc;
+			DBG("PCI Autoconfig: BAR %d, I/O, ", bar_nr);
+		}
+		else
+		{
+			if ((bar_size & PCI_BASE_ADDRESS_MEM_TYPE_MASK) ==
+			    PCI_BASE_ADDRESS_MEM_TYPE_64)
+				found_mem64 = 1;
+
+			bar_size &= PCI_BASE_ADDRESS_MEM_MASK;
+			upper_limit = &pciauto_upper_memspc;
+			DBG("PCI Autoconfig: BAR %d, Mem, ", bar_nr);
+		}
+
+		/* Allocate a base address (bar_size is negative!) */
+		*upper_limit = (*upper_limit + bar_size) & bar_size;
+
+		/* Write it out and update our limit */
+		pci_write_config_dword(dev, bar, *upper_limit);
+
+		/*
+		 * If we are a 64-bit decoder then increment to the
+		 * upper 32 bits of the bar and force it to locate
+		 * in the lower 4GB of memory.
+		 */
+
+		if (found_mem64)
+			pci_write_config_dword(dev, (bar+=4), 0x00000000);
+
+		DBG("size=0x%x, address=0x%x\n", ~bar_size + 1, *upper_limit);
+	}
+}
+
+/* Initialize the interrupt number. */
+
+static void __init
+pciauto_setup_irq(struct pci_controller* pci_ctrl,struct pci_dev *dev,int devfn)
+{
+	u8 pin;
+	int irq = 0;
+
+	pci_read_config_byte(dev, PCI_INTERRUPT_PIN, &pin);
+
+	/* Fix illegal pin numbers. */
+
+	if (pin == 0 || pin > 4)
+		pin = 1;
+
+	if (pci_ctrl->map_irq)
+		irq = pci_ctrl->map_irq(dev, PCI_SLOT(devfn), pin);
+
+	if (irq == -1)
+		irq = 0;
+
+	DBG("PCI Autoconfig: Interrupt %d, pin %d\n", irq, pin);
+
+	pci_write_config_byte(dev, PCI_INTERRUPT_LINE, irq);
+}
+
+
+static void __init
+pciauto_prescan_setup_bridge(struct pci_dev *dev, int current_bus,
+			     int sub_bus, int *iosave, int *memsave)
+{
+	/* Configure bus number registers */
+	pci_write_config_byte(dev, PCI_PRIMARY_BUS, current_bus);
+	pci_write_config_byte(dev, PCI_SECONDARY_BUS, sub_bus + 1);
+	pci_write_config_byte(dev, PCI_SUBORDINATE_BUS,	0xff);
+
+	/* Round memory allocator to 1MB boundary */
+	pciauto_upper_memspc &= ~(0x100000 - 1);
+	*memsave = pciauto_upper_memspc;
+
+	/* Round I/O allocator to 4KB boundary */
+	pciauto_upper_iospc &= ~(0x1000 - 1);
+	*iosave = pciauto_upper_iospc;
+
+	/* Set up memory and I/O filter limits, assume 32-bit I/O space */
+	pci_write_config_word(dev, PCI_MEMORY_LIMIT,
+			      ((pciauto_upper_memspc - 1) & 0xfff00000) >> 16);
+	pci_write_config_byte(dev, PCI_IO_LIMIT,
+			      ((pciauto_upper_iospc - 1) & 0x0000f000) >> 8);
+	pci_write_config_word(dev, PCI_IO_LIMIT_UPPER16,
+			      ((pciauto_upper_iospc - 1) & 0xffff0000) >> 16);
+}
+
+static void __init
+pciauto_postscan_setup_bridge(struct pci_dev *dev, int current_bus, int sub_bus,
+			      int *iosave, int *memsave)
+{
+	int cmdstat;
+
+	/* Configure bus number registers */
+	pci_write_config_byte(dev, PCI_SUBORDINATE_BUS,	sub_bus);
+
+	/*
+	 * Round memory allocator to 1MB boundary.
+	 * If no space used, allocate minimum.
+	 */
+	pciauto_upper_memspc &= ~(0x100000 - 1);
+	if (*memsave == pciauto_upper_memspc)
+		pciauto_upper_memspc -= 0x00100000;
+
+	pci_write_config_word(dev, PCI_MEMORY_BASE, pciauto_upper_memspc >> 16);
+
+	/* Allocate 1MB for pre-fretch */
+	pci_write_config_word(dev, PCI_PREF_MEMORY_LIMIT,
+			      ((pciauto_upper_memspc - 1) & 0xfff00000) >> 16);
+
+	pciauto_upper_memspc -= 0x100000;
+
+	pci_write_config_word(dev, PCI_PREF_MEMORY_BASE,
+			      pciauto_upper_memspc >> 16);
+
+	/* Round I/O allocator to 4KB boundary */
+	pciauto_upper_iospc &= ~(0x1000 - 1);
+	if (*iosave == pciauto_upper_iospc)
+		pciauto_upper_iospc -= 0x1000;
+
+	pci_write_config_byte(dev, PCI_IO_BASE,
+			      (pciauto_upper_iospc & 0x0000f000) >> 8);
+	pci_write_config_word(dev, PCI_IO_BASE_UPPER16,
+			      pciauto_upper_iospc >> 16);
+
+	/* Enable memory and I/O accesses, enable bus master */
+	pci_read_config_dword(dev, PCI_COMMAND, &cmdstat);
+	pci_write_config_dword(dev, PCI_COMMAND,
+			       cmdstat |
+			       PCI_COMMAND_IO |
+			       PCI_COMMAND_MEMORY |
+			       PCI_COMMAND_MASTER);
+}
+
+/*
+ * Scan the current PCI bus.
+ */
+
+
+int __init pciauto_bus_scan(struct pci_controller *pci_ctrl, int current_bus)
+{
+	int sub_bus, pci_devfn, pci_class, cmdstat, found_multi=0;
+	unsigned short vid;
+	unsigned char header_type;
+	struct pci_dev *dev = &pciauto_dev;
+
+        pciauto_dev.bus = &pciauto_bus;
+        pciauto_dev.sysdata = pci_ctrl;
+	pciauto_bus.ops = pci_ctrl->ops;
+
+	/*
+	 * Fetch our I/O and memory space upper boundaries used
+	 * to allocated base addresses on this pci_controller.
+	 */
+
+	if (current_bus == pci_ctrl->first_busno)
+	{
+		pciauto_upper_iospc = pci_ctrl->io_resource.end + 1;
+		pciauto_upper_memspc = pci_ctrl->mem_resources[0].end + 1;
+	}
+
+	sub_bus = current_bus;
+
+	for (pci_devfn = 0; pci_devfn < 0xff; pci_devfn++)
+	{
+		/* Skip our host bridge */
+		if ((current_bus == pci_ctrl->first_busno) && (pci_devfn == 0))
+			continue;
+
+		if (PCI_FUNC(pci_devfn) && !found_multi)
+			continue;
+
+		pciauto_bus.number = current_bus;
+		pciauto_dev.devfn = pci_devfn;
+
+		/* If config space read fails from this device, move on */
+		if (pci_read_config_byte(dev, PCI_HEADER_TYPE, &header_type))
+			continue;
+
+		if (!PCI_FUNC(pci_devfn))
+			found_multi = header_type & 0x80;
+		pci_read_config_word(dev, PCI_VENDOR_ID, &vid);
+
+		if (vid == 0xffff || vid == 0x0000) {
+			found_multi = 0;
+			continue;
+		}
+
+		pci_read_config_dword(dev, PCI_CLASS_REVISION, &pci_class);
+
+		if ((pci_class >> 16) == PCI_CLASS_BRIDGE_PCI) {
+
+			int iosave, memsave;
+
+			DBG("PCI Autoconfig: Found P2P bridge, device %d\n",
+			    PCI_SLOT(pci_devfn));
+
+			/* Allocate PCI I/O and/or memory space */
+			pciauto_setup_bars(dev, PCI_BASE_ADDRESS_1);
+
+			pciauto_prescan_setup_bridge(dev, current_bus, sub_bus,
+					&iosave, &memsave);
+			sub_bus = pciauto_bus_scan(pci_ctrl, sub_bus+1);
+			pciauto_postscan_setup_bridge(dev, current_bus, sub_bus,
+					&iosave, &memsave);
+			pciauto_bus.number = current_bus;
+
+			continue;
+
+		}
+
+
+#if 0
+		/* Skip legacy mode IDE controller */
+
+		if ((pci_class >> 16) == PCI_CLASS_STORAGE_IDE) {
+
+			unsigned char prg_iface;
+			pci_read_config_byte(dev, PCI_CLASS_PROG, &prg_iface);
+
+			if (!(prg_iface & PCIAUTO_IDE_MODE_MASK)) {
+				DBG("PCI Autoconfig: Skipping legacy mode "
+				    "IDE controller\n");
+				continue;
+			}
+		}
+#endif
+
+		/*
+		 * Found a peripheral, enable some standard
+		 * settings
+		 */
+
+		pci_read_config_dword(dev, PCI_COMMAND,	&cmdstat);
+		pci_write_config_dword(dev, PCI_COMMAND,
+				cmdstat |
+					PCI_COMMAND_IO |
+					PCI_COMMAND_MEMORY |
+					PCI_COMMAND_MASTER);
+		pci_write_config_byte(dev, PCI_LATENCY_TIMER, 0x80);
+
+		/* Allocate PCI I/O and/or memory space */
+		DBG("PCI Autoconfig: Found Bus %d, Device %d, Function %d\n",
+		    current_bus, PCI_SLOT(pci_devfn), PCI_FUNC(pci_devfn) );
+
+		pciauto_setup_bars(dev, PCI_BASE_ADDRESS_5);
+		pciauto_setup_irq(pci_ctrl, dev, pci_devfn);
+	}
+	return sub_bus;
+}
+
+
+
+
+
diff --git a/arch/xtensa/lib/strcasecmp.c b/arch/xtensa/lib/strcasecmp.c
new file mode 100644
index 0000000..165b2d6
--- /dev/null
+++ b/arch/xtensa/lib/strcasecmp.c
@@ -0,0 +1,32 @@
+/*
+ *  linux/arch/xtensa/lib/strcasecmp.c
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+#include <linux/string.h>
+
+
+/* We handle nothing here except the C locale.  Since this is used in
+   only one place, on strings known to contain only 7 bit ASCII, this
+   is ok.  */
+
+int strcasecmp(const char *a, const char *b)
+{
+	int ca, cb;
+
+	do {
+		ca = *a++ & 0xff;
+		cb = *b++ & 0xff;
+		if (ca >= 'A' && ca <= 'Z')
+			ca += 'a' - 'A';
+		if (cb >= 'A' && cb <= 'Z')
+			cb += 'a' - 'A';
+	} while (ca == cb && ca != '\0');
+
+	return ca - cb;
+}
diff --git a/arch/xtensa/lib/strncpy_user.S b/arch/xtensa/lib/strncpy_user.S
new file mode 100644
index 0000000..71d55df
--- /dev/null
+++ b/arch/xtensa/lib/strncpy_user.S
@@ -0,0 +1,224 @@
+/*
+ *  arch/xtensa/lib/strncpy_user.S
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Returns: -EFAULT if exception before terminator, N if the entire
+ *  buffer filled, else strlen.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+#include <xtensa/coreasm.h>
+#include <linux/errno.h>
+
+/* Load or store instructions that may cause exceptions use the EX macro. */
+
+#define EX(insn,reg1,reg2,offset,handler)	\
+9:	insn	reg1, reg2, offset;		\
+	.section __ex_table, "a";		\
+	.word	9b, handler;			\
+	.previous
+
+/*
+ * char *__strncpy_user(char *dst, const char *src, size_t len)
+ */
+.text
+.begin literal
+.align	4
+.Lmask0:
+	.byte	0xff, 0x00, 0x00, 0x00
+.Lmask1:
+	.byte	0x00, 0xff, 0x00, 0x00
+.Lmask2:
+	.byte	0x00, 0x00, 0xff, 0x00
+.Lmask3:
+	.byte	0x00, 0x00, 0x00, 0xff
+.end literal
+
+# Register use
+#   a0/ return address
+#   a1/ stack pointer
+#   a2/ return value
+#   a3/ src
+#   a4/ len
+#   a5/ mask0
+#   a6/ mask1
+#   a7/ mask2
+#   a8/ mask3
+#   a9/ tmp
+#   a10/ tmp
+#   a11/ dst
+#   a12/ tmp
+
+.align	4
+.global	__strncpy_user
+.type	__strncpy_user,@function
+__strncpy_user:
+	entry	sp, 16		# minimal stack frame
+	# a2/ dst, a3/ src, a4/ len
+	mov	a11, a2		# leave dst in return value register
+	beqz	a4, .Lret	# if len is zero
+	l32r	a5, .Lmask0	# mask for byte 0
+	l32r	a6, .Lmask1	# mask for byte 1
+	l32r	a7, .Lmask2	# mask for byte 2
+	l32r	a8, .Lmask3	# mask for byte 3
+	bbsi.l	a3, 0, .Lsrc1mod2 # if only  8-bit aligned
+	bbsi.l	a3, 1, .Lsrc2mod4 # if only 16-bit aligned
+.Lsrcaligned:	# return here when src is word-aligned
+	srli	a12, a4, 2	# number of loop iterations with 4B per loop
+	movi	a9, 3
+	bnone	a11, a9, .Laligned
+	j	.Ldstunaligned
+
+.Lsrc1mod2:	# src address is odd
+	EX(l8ui, a9, a3, 0, fixup_l)	# get byte 0
+	addi	a3, a3, 1		# advance src pointer
+	EX(s8i, a9, a11, 0, fixup_s)	# store byte 0
+	beqz	a9, .Lret		# if byte 0 is zero
+	addi	a11, a11, 1		# advance dst pointer
+	addi	a4, a4, -1		# decrement len
+	beqz	a4, .Lret		# if len is zero
+	bbci.l	a3, 1, .Lsrcaligned	# if src is now word-aligned
+
+.Lsrc2mod4:	# src address is 2 mod 4
+	EX(l8ui, a9, a3, 0, fixup_l)	# get byte 0
+	/* 1-cycle interlock */
+	EX(s8i, a9, a11, 0, fixup_s)	# store byte 0
+	beqz	a9, .Lret		# if byte 0 is zero
+	addi	a11, a11, 1		# advance dst pointer
+	addi	a4, a4, -1		# decrement len
+	beqz	a4, .Lret		# if len is zero
+	EX(l8ui, a9, a3, 1, fixup_l)	# get byte 0
+	addi	a3, a3, 2		# advance src pointer
+	EX(s8i, a9, a11, 0, fixup_s)	# store byte 0
+	beqz	a9, .Lret		# if byte 0 is zero
+	addi	a11, a11, 1		# advance dst pointer
+	addi	a4, a4, -1		# decrement len
+	bnez	a4, .Lsrcaligned	# if len is nonzero
+.Lret:
+	sub	a2, a11, a2		# compute strlen
+	retw
+
+/*
+ * dst is word-aligned, src is word-aligned
+ */
+	.align	4		# 1 mod 4 alignment for LOOPNEZ
+	.byte	0		# (0 mod 4 alignment for LBEG)
+.Laligned:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a12, .Loop1done
+#else
+	beqz	a12, .Loop1done
+	slli	a12, a12, 2
+	add	a12, a12, a11	# a12 = end of last 4B chunck
+#endif
+.Loop1:
+	EX(l32i, a9, a3, 0, fixup_l)	# get word from src
+	addi	a3, a3, 4		# advance src pointer
+	bnone	a9, a5, .Lz0		# if byte 0 is zero
+	bnone	a9, a6, .Lz1		# if byte 1 is zero
+	bnone	a9, a7, .Lz2		# if byte 2 is zero
+	EX(s32i, a9, a11, 0, fixup_s)	# store word to dst
+	bnone	a9, a8, .Lz3		# if byte 3 is zero
+	addi	a11, a11, 4		# advance dst pointer
+#if !XCHAL_HAVE_LOOPS
+	blt	a11, a12, .Loop1
+#endif
+
+.Loop1done:
+	bbci.l	a4, 1, .L100
+	# copy 2 bytes
+	EX(l16ui, a9, a3, 0, fixup_l)
+	addi	a3, a3, 2		# advance src pointer
+#ifdef __XTENSA_EB__
+	bnone	a9, a7, .Lz0		# if byte 2 is zero
+	bnone	a9, a8, .Lz1		# if byte 3 is zero
+#else
+	bnone	a9, a5, .Lz0		# if byte 0 is zero
+	bnone	a9, a6, .Lz1		# if byte 1 is zero
+#endif
+	EX(s16i, a9, a11, 0, fixup_s)
+	addi	a11, a11, 2		# advance dst pointer
+.L100:
+	bbci.l	a4, 0, .Lret
+	EX(l8ui, a9, a3, 0, fixup_l)
+	/* slot */
+	EX(s8i, a9, a11, 0, fixup_s)
+	beqz	a9, .Lret		# if byte is zero
+	addi	a11, a11, 1-3		# advance dst ptr 1, but also cancel
+					# the effect of adding 3 in .Lz3 code
+	/* fall thru to .Lz3 and "retw" */
+
+.Lz3:	# byte 3 is zero
+	addi	a11, a11, 3		# advance dst pointer
+	sub	a2, a11, a2		# compute strlen
+	retw
+.Lz0:	# byte 0 is zero
+#ifdef __XTENSA_EB__
+	movi	a9, 0
+#endif /* __XTENSA_EB__ */
+	EX(s8i, a9, a11, 0, fixup_s)
+	sub	a2, a11, a2		# compute strlen
+	retw
+.Lz1:	# byte 1 is zero
+#ifdef __XTENSA_EB__
+        extui   a9, a9, 16, 16
+#endif /* __XTENSA_EB__ */
+	EX(s16i, a9, a11, 0, fixup_s)
+	addi	a11, a11, 1		# advance dst pointer
+	sub	a2, a11, a2		# compute strlen
+	retw
+.Lz2:	# byte 2 is zero
+#ifdef __XTENSA_EB__
+        extui   a9, a9, 16, 16
+#endif /* __XTENSA_EB__ */
+	EX(s16i, a9, a11, 0, fixup_s)
+	movi	a9, 0
+	EX(s8i, a9, a11, 2, fixup_s)
+	addi	a11, a11, 2		# advance dst pointer
+	sub	a2, a11, a2		# compute strlen
+	retw
+
+	.align	4		# 1 mod 4 alignment for LOOPNEZ
+	.byte	0		# (0 mod 4 alignment for LBEG)
+.Ldstunaligned:
+/*
+ * for now just use byte copy loop
+ */
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, .Lunalignedend
+#else
+	beqz	a4, .Lunalignedend
+	add	a12, a11, a4		# a12 = ending address
+#endif /* XCHAL_HAVE_LOOPS */
+.Lnextbyte:
+	EX(l8ui, a9, a3, 0, fixup_l)
+	addi	a3, a3, 1
+	EX(s8i, a9, a11, 0, fixup_s)
+	beqz	a9, .Lunalignedend
+	addi	a11, a11, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a11, a12, .Lnextbyte
+#endif
+
+.Lunalignedend:
+	sub	a2, a11, a2		# compute strlen
+	retw
+
+
+	.section .fixup, "ax"
+	.align	4
+
+	/* For now, just return -EFAULT.  Future implementations might
+	 * like to clear remaining kernel space, like the fixup
+	 * implementation in memset().  Thus, we differentiate between
+	 * load/store fixups. */
+
+fixup_s:
+fixup_l:
+	movi	a2, -EFAULT
+	retw
+
diff --git a/arch/xtensa/lib/strnlen_user.S b/arch/xtensa/lib/strnlen_user.S
new file mode 100644
index 0000000..cdff4d6
--- /dev/null
+++ b/arch/xtensa/lib/strnlen_user.S
@@ -0,0 +1,147 @@
+/*
+ *  arch/xtensa/lib/strnlen_user.S
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Returns strnlen, including trailing zero terminator.
+ *  Zero indicates error.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+#include <xtensa/coreasm.h>
+
+/* Load or store instructions that may cause exceptions use the EX macro. */
+
+#define EX(insn,reg1,reg2,offset,handler)	\
+9:	insn	reg1, reg2, offset;		\
+	.section __ex_table, "a";		\
+	.word	9b, handler;			\
+	.previous
+
+/*
+ * size_t __strnlen_user(const char *s, size_t len)
+ */
+.text
+.begin literal
+.align	4
+.Lmask0:
+	.byte	0xff, 0x00, 0x00, 0x00
+.Lmask1:
+	.byte	0x00, 0xff, 0x00, 0x00
+.Lmask2:
+	.byte	0x00, 0x00, 0xff, 0x00
+.Lmask3:
+	.byte	0x00, 0x00, 0x00, 0xff
+.end literal
+
+# Register use:
+#   a2/ src
+#   a3/ len
+#   a4/ tmp
+#   a5/ mask0
+#   a6/ mask1
+#   a7/ mask2
+#   a8/ mask3
+#   a9/ tmp
+#   a10/ tmp
+
+.align	4
+.global	__strnlen_user
+.type	__strnlen_user,@function
+__strnlen_user:
+	entry	sp, 16		# minimal stack frame
+	# a2/ s, a3/ len
+	addi	a4, a2, -4	# because we overincrement at the end;
+				# we compensate with load offsets of 4
+	l32r	a5, .Lmask0	# mask for byte 0
+	l32r	a6, .Lmask1	# mask for byte 1
+	l32r	a7, .Lmask2	# mask for byte 2
+	l32r	a8, .Lmask3	# mask for byte 3
+	bbsi.l	a2, 0, .L1mod2	# if only  8-bit aligned
+	bbsi.l	a2, 1, .L2mod4	# if only 16-bit aligned
+
+/*
+ * String is word-aligned.
+ */
+.Laligned:
+	srli	a10, a3, 2	# number of loop iterations with 4B per loop
+#if XCHAL_HAVE_LOOPS
+	loopnez	a10, .Ldone
+#else
+	beqz	a10, .Ldone
+	slli	a10, a10, 2
+	add	a10, a10, a4	# a10 = end of last 4B chunk
+#endif /* XCHAL_HAVE_LOOPS */
+.Loop:
+	EX(l32i, a9, a4, 4, lenfixup)	# get next word of string
+	addi	a4, a4, 4		# advance string pointer
+	bnone	a9, a5, .Lz0		# if byte 0 is zero
+	bnone	a9, a6, .Lz1		# if byte 1 is zero
+	bnone	a9, a7, .Lz2		# if byte 2 is zero
+	bnone	a9, a8, .Lz3		# if byte 3 is zero
+#if !XCHAL_HAVE_LOOPS
+	blt	a4, a10, .Loop
+#endif
+
+.Ldone:
+	EX(l32i, a9, a4, 4, lenfixup)	# load 4 bytes for remaining checks
+
+	bbci.l	a3, 1, .L100
+	# check two more bytes (bytes 0, 1 of word)
+	addi	a4, a4, 2	# advance string pointer
+	bnone	a9, a5, .Lz0	# if byte 0 is zero
+	bnone	a9, a6, .Lz1	# if byte 1 is zero
+.L100:
+	bbci.l	a3, 0, .L101
+	# check one more byte (byte 2 of word)
+	# Actually, we don't need to check.  Zero or nonzero, we'll add one.
+	# Do not add an extra one for the NULL terminator since we have
+	#  exhausted the original len parameter.
+	addi	a4, a4, 1	# advance string pointer
+.L101:
+	sub	a2, a4, a2	# compute length
+	retw
+
+# NOTE that in several places below, we point to the byte just after
+# the zero byte in order to include the NULL terminator in the count.
+
+.Lz3:	# byte 3 is zero
+	addi	a4, a4, 3	# point to zero byte
+.Lz0:	# byte 0 is zero
+	addi	a4, a4, 1	# point just beyond zero byte
+	sub	a2, a4, a2	# subtract to get length
+	retw
+.Lz1:	# byte 1 is zero
+	addi	a4, a4, 1+1	# point just beyond zero byte
+	sub	a2, a4, a2	# subtract to get length
+	retw
+.Lz2:	# byte 2 is zero
+	addi	a4, a4, 2+1	# point just beyond zero byte
+	sub	a2, a4, a2	# subtract to get length
+	retw
+
+.L1mod2:	# address is odd
+	EX(l8ui, a9, a4, 4, lenfixup)	# get byte 0
+	addi	a4, a4, 1		# advance string pointer
+	beqz	a9, .Lz3		# if byte 0 is zero
+	bbci.l	a4, 1, .Laligned	# if string pointer is now word-aligned
+
+.L2mod4:	# address is 2 mod 4
+	addi	a4, a4, 2	# advance ptr for aligned access
+	EX(l32i, a9, a4, 0, lenfixup)	# get word with first two bytes of string
+	bnone	a9, a7, .Lz2	# if byte 2 (of word, not string) is zero
+	bany	a9, a8, .Laligned # if byte 3 (of word, not string) is nonzero
+	# byte 3 is zero
+	addi	a4, a4, 3+1	# point just beyond zero byte
+	sub	a2, a4, a2	# subtract to get length
+	retw
+
+	.section .fixup, "ax"
+	.align	4
+lenfixup:
+	movi	a2, 0
+	retw
+
diff --git a/arch/xtensa/lib/usercopy.S b/arch/xtensa/lib/usercopy.S
new file mode 100644
index 0000000..265db26
--- /dev/null
+++ b/arch/xtensa/lib/usercopy.S
@@ -0,0 +1,321 @@
+/*
+ *  arch/xtensa/lib/usercopy.S
+ *
+ *  Copy to/from user space (derived from arch/xtensa/lib/hal/memcopy.S)
+ *
+ *  DO NOT COMBINE this function with <arch/xtensa/lib/hal/memcopy.S>.
+ *  It needs to remain separate and distinct.  The hal files are part
+ *  of the the Xtensa link-time HAL, and those files may differ per
+ *  processor configuration.  Patching the kernel for another
+ *  processor configuration includes replacing the hal files, and we
+ *  could loose the special functionality for accessing user-space
+ *  memory during such a patch.  We sacrifice a little code space here
+ *  in favor to simplify code maintenance.
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+
+/*
+ * size_t __xtensa_copy_user (void *dst, const void *src, size_t len);
+ *
+ * The returned value is the number of bytes not copied.  Implies zero
+ * is success.
+ *
+ * The general case algorithm is as follows:
+ *   If the destination and source are both aligned,
+ *     do 16B chunks with a loop, and then finish up with
+ *     8B, 4B, 2B, and 1B copies conditional on the length.
+ *   If destination is aligned and source unaligned,
+ *     do the same, but use SRC to align the source data.
+ *   If destination is unaligned, align it by conditionally
+ *     copying 1B and 2B and then retest.
+ *   This code tries to use fall-through braches for the common
+ *     case of aligned destinations (except for the branches to
+ *     the alignment label).
+ *
+ * Register use:
+ *	a0/ return address
+ *	a1/ stack pointer
+ *	a2/ return value
+ *	a3/ src
+ *	a4/ length
+ *	a5/ dst
+ *	a6/ tmp
+ *	a7/ tmp
+ *	a8/ tmp
+ *	a9/ tmp
+ *	a10/ tmp
+ *	a11/ original length
+ */
+
+#include <xtensa/coreasm.h>
+
+#ifdef __XTENSA_EB__
+#define ALIGN(R, W0, W1) src	R, W0, W1
+#define SSA8(R)	ssa8b R
+#else
+#define ALIGN(R, W0, W1) src	R, W1, W0
+#define SSA8(R)	ssa8l R
+#endif
+
+/* Load or store instructions that may cause exceptions use the EX macro. */
+
+#define EX(insn,reg1,reg2,offset,handler)	\
+9:	insn	reg1, reg2, offset;		\
+	.section __ex_table, "a";		\
+	.word	9b, handler;			\
+	.previous
+
+
+	.text
+	.align	4
+	.global	__xtensa_copy_user
+	.type	__xtensa_copy_user,@function
+__xtensa_copy_user:
+	entry	sp, 16		# minimal stack frame
+	# a2/ dst, a3/ src, a4/ len
+	mov	a5, a2		# copy dst so that a2 is return value
+	mov	a11, a4		# preserve original len for error case
+.Lcommon:
+	bbsi.l	a2, 0, .Ldst1mod2 # if dst is 1 mod 2
+	bbsi.l	a2, 1, .Ldst2mod4 # if dst is 2 mod 4
+.Ldstaligned:	# return here from .Ldstunaligned when dst is aligned
+	srli	a7, a4, 4	# number of loop iterations with 16B
+				# per iteration
+	movi	a8, 3		  # if source is also aligned,
+	bnone	a3, a8, .Laligned # then use word copy
+	SSA8(	a3)		# set shift amount from byte offset
+	bnez	a4, .Lsrcunaligned
+	movi	a2, 0		# return success for len==0
+	retw
+
+/*
+ * Destination is unaligned
+ */
+
+.Ldst1mod2:	# dst is only byte aligned
+	bltui	a4, 7, .Lbytecopy	# do short copies byte by byte
+
+	# copy 1 byte
+	EX(l8ui, a6, a3, 0, l_fixup)
+	addi	a3, a3,  1
+	EX(s8i, a6, a5,  0, s_fixup)
+	addi	a5, a5,  1
+	addi	a4, a4, -1
+	bbci.l	a5, 1, .Ldstaligned	# if dst is now aligned, then
+					# return to main algorithm
+.Ldst2mod4:	# dst 16-bit aligned
+	# copy 2 bytes
+	bltui	a4, 6, .Lbytecopy	# do short copies byte by byte
+	EX(l8ui, a6, a3, 0, l_fixup)
+	EX(l8ui, a7, a3, 1, l_fixup)
+	addi	a3, a3,  2
+	EX(s8i, a6, a5,  0, s_fixup)
+	EX(s8i, a7, a5,  1, s_fixup)
+	addi	a5, a5,  2
+	addi	a4, a4, -2
+	j	.Ldstaligned	# dst is now aligned, return to main algorithm
+
+/*
+ * Byte by byte copy
+ */
+	.align	4
+	.byte	0		# 1 mod 4 alignment for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lbytecopy:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, .Lbytecopydone
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a4, .Lbytecopydone
+	add	a7, a3, a4	# a7 = end address for source
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lnextbyte:
+	EX(l8ui, a6, a3, 0, l_fixup)
+	addi	a3, a3, 1
+	EX(s8i, a6, a5, 0, s_fixup)
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a7, .Lnextbyte
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lbytecopydone:
+	movi	a2, 0		# return success for len bytes copied
+	retw
+
+/*
+ * Destination and source are word-aligned.
+ */
+	# copy 16 bytes per iteration for word-aligned dst and word-aligned src
+	.align	4		# 1 mod 4 alignment for LOOPNEZ
+	.byte	0		# (0 mod 4 alignment for LBEG)
+.Laligned:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop1done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop1done
+	slli	a8, a7, 4
+	add	a8, a8, a3	# a8 = end of last 16B source chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1:
+	EX(l32i, a6, a3,  0, l_fixup)
+	EX(l32i, a7, a3,  4, l_fixup)
+	EX(s32i, a6, a5,  0, s_fixup)
+	EX(l32i, a6, a3,  8, l_fixup)
+	EX(s32i, a7, a5,  4, s_fixup)
+	EX(l32i, a7, a3, 12, l_fixup)
+	EX(s32i, a6, a5,  8, s_fixup)
+	addi	a3, a3, 16
+	EX(s32i, a7, a5, 12, s_fixup)
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a8, .Loop1
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1done:
+	bbci.l	a4, 3, .L2
+	# copy 8 bytes
+	EX(l32i, a6, a3,  0, l_fixup)
+	EX(l32i, a7, a3,  4, l_fixup)
+	addi	a3, a3,  8
+	EX(s32i, a6, a5,  0, s_fixup)
+	EX(s32i, a7, a5,  4, s_fixup)
+	addi	a5, a5,  8
+.L2:
+	bbci.l	a4, 2, .L3
+	# copy 4 bytes
+	EX(l32i, a6, a3,  0, l_fixup)
+	addi	a3, a3,  4
+	EX(s32i, a6, a5,  0, s_fixup)
+	addi	a5, a5,  4
+.L3:
+	bbci.l	a4, 1, .L4
+	# copy 2 bytes
+	EX(l16ui, a6, a3,  0, l_fixup)
+	addi	a3, a3,  2
+	EX(s16i,  a6, a5,  0, s_fixup)
+	addi	a5, a5,  2
+.L4:
+	bbci.l	a4, 0, .L5
+	# copy 1 byte
+	EX(l8ui, a6, a3,  0, l_fixup)
+	EX(s8i,  a6, a5,  0, s_fixup)
+.L5:
+	movi	a2, 0		# return success for len bytes copied
+	retw
+
+/*
+ * Destination is aligned, Source is unaligned
+ */
+
+	.align	4
+	.byte	0		# 1 mod 4 alignement for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lsrcunaligned:
+	# copy 16 bytes per iteration for word-aligned dst and unaligned src
+	and	a10, a3, a8	# save unalignment offset for below
+	sub	a3, a3, a10	# align a3 (to avoid sim warnings only; not needed for hardware)
+	EX(l32i, a6, a3, 0, l_fixup)	# load first word
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop2done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop2done
+	slli	a10, a7, 4
+	add	a10, a10, a3	# a10 = end of last 16B source chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2:
+	EX(l32i, a7, a3,  4, l_fixup)
+	EX(l32i, a8, a3,  8, l_fixup)
+	ALIGN(	a6, a6, a7)
+	EX(s32i, a6, a5,  0, s_fixup)
+	EX(l32i, a9, a3, 12, l_fixup)
+	ALIGN(	a7, a7, a8)
+	EX(s32i, a7, a5,  4, s_fixup)
+	EX(l32i, a6, a3, 16, l_fixup)
+	ALIGN(	a8, a8, a9)
+	EX(s32i, a8, a5,  8, s_fixup)
+	addi	a3, a3, 16
+	ALIGN(	a9, a9, a6)
+	EX(s32i, a9, a5, 12, s_fixup)
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a10, .Loop2
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2done:
+	bbci.l	a4, 3, .L12
+	# copy 8 bytes
+	EX(l32i, a7, a3,  4, l_fixup)
+	EX(l32i, a8, a3,  8, l_fixup)
+	ALIGN(	a6, a6, a7)
+	EX(s32i, a6, a5,  0, s_fixup)
+	addi	a3, a3,  8
+	ALIGN(	a7, a7, a8)
+	EX(s32i, a7, a5,  4, s_fixup)
+	addi	a5, a5,  8
+	mov	a6, a8
+.L12:
+	bbci.l	a4, 2, .L13
+	# copy 4 bytes
+	EX(l32i, a7, a3,  4, l_fixup)
+	addi	a3, a3,  4
+	ALIGN(	a6, a6, a7)
+	EX(s32i, a6, a5,  0, s_fixup)
+	addi	a5, a5,  4
+	mov	a6, a7
+.L13:
+	add	a3, a3, a10	# readjust a3 with correct misalignment
+	bbci.l	a4, 1, .L14
+	# copy 2 bytes
+	EX(l8ui, a6, a3,  0, l_fixup)
+	EX(l8ui, a7, a3,  1, l_fixup)
+	addi	a3, a3,  2
+	EX(s8i, a6, a5,  0, s_fixup)
+	EX(s8i, a7, a5,  1, s_fixup)
+	addi	a5, a5,  2
+.L14:
+	bbci.l	a4, 0, .L15
+	# copy 1 byte
+	EX(l8ui, a6, a3,  0, l_fixup)
+	EX(s8i,  a6, a5,  0, s_fixup)
+.L15:
+	movi	a2, 0		# return success for len bytes copied
+	retw
+
+
+	.section .fixup, "ax"
+	.align	4
+
+/* a2 = original dst; a5 = current dst; a11= original len
+ * bytes_copied = a5 - a2
+ * retval = bytes_not_copied = original len - bytes_copied
+ * retval = a11 - (a5 - a2)
+ *
+ * Clearing the remaining pieces of kernel memory plugs security
+ * holes.  This functionality is the equivalent of the *_zeroing
+ * functions that some architectures provide.
+ */
+
+.Lmemset:
+	.word	memset
+
+s_fixup:
+	sub	a2, a5, a2	/* a2 <-- bytes copied */
+	sub	a2, a11, a2	/* a2 <-- bytes not copied */
+	retw
+
+l_fixup:
+	sub	a2, a5, a2	/* a2 <-- bytes copied */
+	sub	a2, a11, a2	/* a2 <-- bytes not copied == return value */
+
+	/* void *memset(void *s, int c, size_t n); */
+	mov	a6, a5		/* s */
+	movi	a7, 0		/* c */
+	mov	a8, a2		/* n */
+	l32r	a4, .Lmemset
+	callx4	a4
+	/* Ignore memset return value in a6. */
+	/* a2 still contains bytes not copied. */
+	retw
+