[PATCH] CRIS update: new subarchitecture v32

New CRIS sub architecture named v32.

From: Dave Jones <davej@redhat.com>

	Fix swapped kmalloc args

Signed-off-by: Mikael Starvik <starvik@axis.com>
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

11 files changed, 1653 insertions, 0 deletions

diff --git a/arch/cris/arch-v32/lib/Makefile b/arch/cris/arch-v32/lib/Makefile
new file mode 100644
index 0000000..05b3ec6
--- /dev/null
+++ b/arch/cris/arch-v32/lib/Makefile
@@ -0,0 +1,6 @@
+#
+# Makefile for Etrax-specific library files..
+#
+
+lib-y  = checksum.o checksumcopy.o string.o usercopy.o memset.o csumcpfruser.o spinlock.o
+
diff --git a/arch/cris/arch-v32/lib/checksum.S b/arch/cris/arch-v32/lib/checksum.S
new file mode 100644
index 0000000..32e6618
--- /dev/null
+++ b/arch/cris/arch-v32/lib/checksum.S
@@ -0,0 +1,111 @@
+/*
+ * A fast checksum routine using movem
+ * Copyright (c) 1998-2001, 2003 Axis Communications AB
+ *
+ * csum_partial(const unsigned char * buff, int len, unsigned int sum)
+ */
+
+	.globl	csum_partial
+csum_partial:
+
+	;; r10 - src
+	;; r11 - length
+	;; r12 - checksum
+
+	;; check for breakeven length between movem and normal word looping versions
+	;; we also do _NOT_ want to compute a checksum over more than the
+	;; actual length when length < 40
+
+	cmpu.w	80,$r11
+	blo	_word_loop
+	nop
+
+	;; need to save the registers we use below in the movem loop
+	;; this overhead is why we have a check above for breakeven length
+	;; only r0 - r8 have to be saved, the other ones are clobber-able
+	;; according to the ABI
+
+	subq	9*4,$sp
+	subq	10*4,$r11	; update length for the first loop
+	movem	$r8,[$sp]
+
+	;; do a movem checksum
+
+_mloop:	movem	[$r10+],$r9	; read 10 longwords
+
+	;; perform dword checksumming on the 10 longwords
+
+	add.d	$r0,$r12
+	addc	$r1,$r12
+	addc	$r2,$r12
+	addc	$r3,$r12
+	addc	$r4,$r12
+	addc	$r5,$r12
+	addc	$r6,$r12
+	addc	$r7,$r12
+	addc	$r8,$r12
+	addc	$r9,$r12
+
+	;; fold the carry into the checksum, to avoid having to loop the carry
+	;; back into the top
+
+	addc	0,$r12
+	addc	0,$r12		; do it again, since we might have generated a carry
+
+	subq	10*4,$r11
+	bge	_mloop
+	nop
+
+	addq	10*4,$r11	; compensate for last loop underflowing length
+
+	movem	[$sp+],$r8	; restore regs
+
+_word_loop:
+	;; only fold if there is anything to fold.
+
+	cmpq	0,$r12
+	beq	_no_fold
+
+	;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
+	;; r9 and r13 can be used as temporaries.
+
+	moveq	-1,$r9		; put 0xffff in r9, faster than move.d 0xffff,r9
+	lsrq	16,$r9
+
+	move.d	$r12,$r13
+	lsrq	16,$r13		; r13 = checksum >> 16
+	and.d	$r9,$r12		; checksum = checksum & 0xffff
+	add.d	$r13,$r12		; checksum += r13
+	move.d	$r12,$r13		; do the same again, maybe we got a carry last add
+	lsrq	16,$r13
+	and.d	$r9,$r12
+	add.d	$r13,$r12
+
+_no_fold:
+	cmpq	2,$r11
+	blt	_no_words
+	nop
+
+	;; checksum the rest of the words
+
+	subq	2,$r11
+
+_wloop:	subq	2,$r11
+	bge	_wloop
+	addu.w	[$r10+],$r12
+
+	addq	2,$r11
+
+_no_words:
+	;; see if we have one odd byte more
+	cmpq	1,$r11
+	beq	_do_byte
+	nop
+	ret
+	move.d	$r12,$r10
+
+_do_byte:
+	;; copy and checksum the last byte
+	addu.b	[$r10],$r12
+	ret
+	move.d	$r12,$r10
diff --git a/arch/cris/arch-v32/lib/checksumcopy.S b/arch/cris/arch-v32/lib/checksumcopy.S
new file mode 100644
index 0000000..9303ccb
--- /dev/null
+++ b/arch/cris/arch-v32/lib/checksumcopy.S
@@ -0,0 +1,120 @@
+/*
+ * A fast checksum+copy routine using movem
+ * Copyright (c) 1998, 2001, 2003 Axis Communications AB
+ *
+ * Authors:	Bjorn Wesen
+ *
+ * csum_partial_copy_nocheck(const char *src, char *dst,
+ *		             int len, unsigned int sum)
+ */
+
+	.globl	csum_partial_copy_nocheck
+csum_partial_copy_nocheck:
+
+	;; r10 - src
+	;; r11 - dst
+	;; r12 - length
+	;; r13 - checksum
+
+	;; check for breakeven length between movem and normal word looping versions
+	;; we also do _NOT_ want to compute a checksum over more than the
+	;; actual length when length < 40
+
+	cmpu.w	80,$r12
+	blo	_word_loop
+	nop
+
+	;; need to save the registers we use below in the movem loop
+	;; this overhead is why we have a check above for breakeven length
+	;; only r0 - r8 have to be saved, the other ones are clobber-able
+	;; according to the ABI
+
+	subq	9*4,$sp
+	subq	10*4,$r12	; update length for the first loop
+	movem	$r8,[$sp]
+
+	;; do a movem copy and checksum
+
+1:	;; A failing userspace access (the read) will have this as PC.
+_mloop:	movem	[$r10+],$r9	; read 10 longwords
+	movem	$r9,[$r11+]	; write 10 longwords
+
+	;; perform dword checksumming on the 10 longwords
+
+	add.d	$r0,$r13
+	addc	$r1,$r13
+	addc	$r2,$r13
+	addc	$r3,$r13
+	addc	$r4,$r13
+	addc	$r5,$r13
+	addc	$r6,$r13
+	addc	$r7,$r13
+	addc	$r8,$r13
+	addc	$r9,$r13
+
+	;; fold the carry into the checksum, to avoid having to loop the carry
+	;; back into the top
+
+	addc	0,$r13
+	addc	0,$r13		; do it again, since we might have generated a carry
+
+	subq	10*4,$r12
+	bge	_mloop
+	nop
+
+	addq	10*4,$r12	; compensate for last loop underflowing length
+
+	movem	[$sp+],$r8	; restore regs
+
+_word_loop:
+	;; only fold if there is anything to fold.
+
+	cmpq	0,$r13
+	beq	_no_fold
+
+	;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below
+	;; r9 can be used as temporary.
+
+	move.d	$r13,$r9
+	lsrq	16,$r9		; r0 = checksum >> 16
+	and.d	0xffff,$r13	; checksum = checksum & 0xffff
+	add.d	$r9,$r13	; checksum += r0
+	move.d	$r13,$r9	; do the same again, maybe we got a carry last add
+	lsrq	16,$r9
+	and.d	0xffff,$r13
+	add.d	$r9,$r13
+
+_no_fold:
+	cmpq	2,$r12
+	blt	_no_words
+	nop
+
+	;; copy and checksum the rest of the words
+
+	subq	2,$r12
+
+2:	;; A failing userspace access for the read below will have this as PC.
+_wloop:	move.w	[$r10+],$r9
+	addu.w	$r9,$r13
+	subq	2,$r12
+	bge	_wloop
+	move.w	$r9,[$r11+]
+
+	addq	2,$r12
+
+_no_words:
+	;; see if we have one odd byte more
+	cmpq	1,$r12
+	beq	_do_byte
+	nop
+	ret
+	move.d	$r13,$r10
+
+_do_byte:
+	;; copy and checksum the last byte
+3:	;; A failing userspace access for the read below will have this as PC.
+	move.b	[$r10],$r9
+	addu.b	$r9,$r13
+	move.b	$r9,[$r11]
+	ret
+	move.d	$r13,$r10
diff --git a/arch/cris/arch-v32/lib/csumcpfruser.S b/arch/cris/arch-v32/lib/csumcpfruser.S
new file mode 100644
index 0000000..600ec16
--- /dev/null
+++ b/arch/cris/arch-v32/lib/csumcpfruser.S
@@ -0,0 +1,69 @@
+/*
+ * Add-on to transform csum_partial_copy_nocheck in checksumcopy.S into
+ * csum_partial_copy_from_user by adding exception records.
+ *
+ * Copyright (C) 2001, 2003 Axis Communications AB.
+ *
+ * Author: Hans-Peter Nilsson.
+ */
+
+#include <asm/errno.h>
+
+/* Same function body, but a different name.  If we just added exception
+   records to _csum_partial_copy_nocheck and made it generic, we wouldn't
+   know a user fault from a kernel fault and we would have overhead in
+   each kernel caller for the error-pointer argument.
+
+   unsigned int csum_partial_copy_from_user
+     (const char *src, char *dst, int len, unsigned int sum, int *errptr);
+
+   Note that the errptr argument is only set if we encounter an error.
+   It is conveniently located on the stack, so the normal function body
+   does not have to handle it.  */
+
+#define csum_partial_copy_nocheck csum_partial_copy_from_user
+
+/* There are local labels numbered 1, 2 and 3 present to mark the
+   different from-user accesses.  */
+#include "checksumcopy.S"
+
+	.section .fixup,"ax"
+
+;; Here from the movem loop; restore stack.
+4:
+	movem	[$sp+],$r8
+;; r12 is already decremented.  Add back chunk_size-2.
+	addq	40-2,$r12
+
+;; Here from the word loop; r12 is off by 2; add it back.
+5:
+	addq	2,$r12
+
+;; Here from a failing single byte.
+6:
+
+;; Signal in *errptr that we had a failing access.
+	move.d	[$sp],$acr
+	moveq	-EFAULT,$r9
+	subq	4,$sp
+	move.d	$r9,[$acr]
+
+;; Clear the rest of the destination area using memset.  Preserve the
+;; checksum for the readable bytes.
+	move.d	$r13,[$sp]
+	subq	4,$sp
+	move.d	$r11,$r10
+	move	$srp,[$sp]
+	jsr	memset
+	clear.d	$r11
+
+	move	[$sp+],$srp
+	ret
+	move.d	[$sp+],$r10
+
+	.previous
+	.section __ex_table,"a"
+	.dword 1b,4b
+	.dword 2b,5b
+	.dword 3b,6b
+	.previous
diff --git a/arch/cris/arch-v32/lib/dram_init.S b/arch/cris/arch-v32/lib/dram_init.S
new file mode 100644
index 0000000..47b6cf5
--- /dev/null
+++ b/arch/cris/arch-v32/lib/dram_init.S
@@ -0,0 +1,120 @@
+/* $Id: dram_init.S,v 1.4 2005/04/24 18:48:32 starvik Exp $
+ *
+ * DRAM/SDRAM initialization - alter with care
+ * This file is intended to be included from other assembler files
+ *
+ * Note: This file may not modify r8 or r9 because they are used to
+ * carry information from the decompresser to the kernel
+ *
+ * Copyright (C) 2000-2003 Axis Communications AB
+ *
+ * Authors:  Mikael Starvik (starvik@axis.com)
+ */
+
+/* Just to be certain the config file is included, we include it here
+ * explicitely instead of depending on it being included in the file that
+ * uses this code.
+ */
+
+#include <linux/config.h>
+#include <asm/arch/hwregs/asm/reg_map_asm.h>
+#include <asm/arch/hwregs/asm/bif_core_defs_asm.h>
+
+	;; WARNING! The registers r8 and r9 are used as parameters carrying
+	;; information from the decompressor (if the kernel was compressed).
+	;; They should not be used in the code below.
+
+	; Refer to BIF MDS for a description of SDRAM initialization
+
+	; Bank configuration
+	move.d   REG_ADDR(bif_core, regi_bif_core, rw_sdram_cfg_grp0), $r0
+	move.d   CONFIG_ETRAX_SDRAM_GRP0_CONFIG, $r1
+	move.d   $r1, [$r0]
+	move.d   REG_ADDR(bif_core, regi_bif_core, rw_sdram_cfg_grp1), $r0
+	move.d   CONFIG_ETRAX_SDRAM_GRP1_CONFIG, $r1
+	move.d   $r1, [$r0]
+
+	; Calculate value of mrs_data
+	; CAS latency = 2 && bus_width = 32 => 0x40
+	; CAS latency = 3 && bus_width = 32 => 0x60
+	; CAS latency = 2 && bus_width = 16 => 0x20
+	; CAS latency = 3 && bus_width = 16 => 0x30
+
+	; Check if value is already supplied in kernel config
+	move.d   CONFIG_ETRAX_SDRAM_COMMAND, $r2
+	bne	 _set_timing
+	nop
+
+	move.d   0x40, $r4       ; Assume 32 bits and CAS latency = 2
+	move.d   CONFIG_ETRAX_SDRAM_TIMING, $r1
+ 	and.d    0x07, $r1       ; Get CAS latency
+	cmpq	 2, $r1		 ; CL = 2 ?
+	beq	 _bw_check
+	nop
+	move.d   0x60, $r4
+
+_bw_check:
+	; Assume that group 0 width is equal to group 1. This assumption
+	; is wrong for a group 1 only hardware (such as the grand old
+	; StorPoint+).
+	move.d   CONFIG_ETRAX_SDRAM_GRP0_CONFIG, $r1
+	and.d    0x200, $r1	; DRAM width is bit 9
+	beq      _set_timing
+	lslq	 2, $r4		;  mrs_data starts at bit 2
+	lsrq     1, $r4		;  16 bits. Shift down value.
+
+	; Set timing parameters (refresh off to avoid Guinness TR 83)
+_set_timing:
+	move.d   CONFIG_ETRAX_SDRAM_TIMING, $r1
+	and.d    ~(3 << reg_bif_core_rw_sdram_timing___ref___lsb), $r1
+	move.d   REG_ADDR(bif_core, regi_bif_core, rw_sdram_timing), $r0
+	move.d   $r1, [$r0]
+
+	; Issue NOP command
+	move.d REG_ADDR(bif_core, regi_bif_core, rw_sdram_cmd), $r5
+	moveq regk_bif_core_nop, $r1
+	move.d $r1, [$r5]
+
+	; Wait 200us
+	move.d   10000, $r2
+1:	bne      1b
+	subq     1, $r2
+
+	; Issue initialization command sequence
+	move.d   _sdram_commands_start, $r2
+	and.d    0x000fffff, $r2 ; Make sure commands are read from flash
+	move.d   _sdram_commands_end,  $r3
+	and.d    0x000fffff, $r3
+1:	clear.d  $r6
+	move.b   [$r2+], $r6 	; Load command
+	or.d     $r4, $r6	; Add calculated mrs
+	move.d   $r6, [$r5]	; Write rw_sdram_cmd
+	; Wait 80 ns between each command
+	move.d	 4000, $r7
+2:	bne	 2b
+	subq	 1, $r7
+	cmp.d    $r2, $r3	; Last command?
+	bne      1b
+	nop
+
+	; Start refresh
+	move.d   CONFIG_ETRAX_SDRAM_TIMING, $r1
+	move.d   REG_ADDR(bif_core, regi_bif_core, rw_sdram_timing), $r0
+	move.d   $r1, [$r0]
+
+	; Initialization finished
+	ba       _sdram_commands_end
+	nop
+
+_sdram_commands_start:
+	.byte   regk_bif_core_pre ; Precharge
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_ref ; refresh
+	.byte   regk_bif_core_mrs ; mrs
+_sdram_commands_end:
diff --git a/arch/cris/arch-v32/lib/hw_settings.S b/arch/cris/arch-v32/lib/hw_settings.S
new file mode 100644
index 0000000..5182e8c
--- /dev/null
+++ b/arch/cris/arch-v32/lib/hw_settings.S
@@ -0,0 +1,73 @@
+/*
+ * $Id: hw_settings.S,v 1.3 2005/04/24 18:36:57 starvik Exp $
+ *
+ * This table is used by some tools to extract hardware parameters.
+ * The table should be included in the kernel and the decompressor.
+ * Don't forget to update the tools if you change this table.
+ *
+ * Copyright (C) 2001 Axis Communications AB
+ *
+ * Authors:  Mikael Starvik (starvik@axis.com)
+ */
+
+#include <linux/config.h>
+#include <asm/arch/hwregs/asm/reg_map_asm.h>
+#include <asm/arch/hwregs/asm/bif_core_defs_asm.h>
+#include <asm/arch/hwregs/asm/gio_defs_asm.h>
+
+	.ascii "HW_PARAM_MAGIC" ; Magic number
+	.dword 0xc0004000	; Kernel start address
+
+	; Debug port
+#ifdef CONFIG_ETRAX_DEBUG_PORT0
+	.dword 0
+#elif defined(CONFIG_ETRAX_DEBUG_PORT1)
+	.dword 1
+#elif defined(CONFIG_ETRAX_DEBUG_PORT2)
+	.dword 2
+#elif defined(CONFIG_ETRAX_DEBUG_PORT3)
+	.dword 3
+#else
+	.dword 4 ; No debug
+#endif
+
+	; Register values
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_grp1_cfg)
+	.dword CONFIG_ETRAX_MEM_GRP1_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_grp2_cfg)
+	.dword CONFIG_ETRAX_MEM_GRP2_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_grp3_cfg)
+	.dword CONFIG_ETRAX_MEM_GRP3_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_grp4_cfg)
+	.dword CONFIG_ETRAX_MEM_GRP4_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_sdram_cfg_grp0)
+	.dword CONFIG_ETRAX_SDRAM_GRP0_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_sdram_cfg_grp1)
+	.dword CONFIG_ETRAX_SDRAM_GRP1_CONFIG
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_sdram_timing)
+	.dword CONFIG_ETRAX_SDRAM_TIMING
+	.dword REG_ADDR(bif_core, regi_bif_core, rw_sdram_cmd)
+	.dword CONFIG_ETRAX_SDRAM_COMMAND
+
+	.dword REG_ADDR(gio, regi_gio, rw_pa_dout)
+	.dword CONFIG_ETRAX_DEF_GIO_PA_OUT
+	.dword REG_ADDR(gio, regi_gio, rw_pa_oe)
+	.dword CONFIG_ETRAX_DEF_GIO_PA_OE
+	.dword REG_ADDR(gio, regi_gio, rw_pb_dout)
+	.dword CONFIG_ETRAX_DEF_GIO_PB_OUT
+	.dword REG_ADDR(gio, regi_gio, rw_pb_oe)
+	.dword CONFIG_ETRAX_DEF_GIO_PB_OE
+	.dword REG_ADDR(gio, regi_gio, rw_pc_dout)
+	.dword CONFIG_ETRAX_DEF_GIO_PC_OUT
+	.dword REG_ADDR(gio, regi_gio, rw_pc_oe)
+	.dword CONFIG_ETRAX_DEF_GIO_PC_OE
+	.dword REG_ADDR(gio, regi_gio, rw_pd_dout)
+	.dword CONFIG_ETRAX_DEF_GIO_PD_OUT
+	.dword REG_ADDR(gio, regi_gio, rw_pd_oe)
+	.dword CONFIG_ETRAX_DEF_GIO_PD_OE
+	.dword REG_ADDR(gio, regi_gio, rw_pe_dout)
+	.dword CONFIG_ETRAX_DEF_GIO_PE_OUT
+	.dword REG_ADDR(gio, regi_gio, rw_pe_oe)
+	.dword CONFIG_ETRAX_DEF_GIO_PE_OE
+
+	.dword 0 ; No more register values
diff --git a/arch/cris/arch-v32/lib/memset.c b/arch/cris/arch-v32/lib/memset.c
new file mode 100644
index 0000000..ffca121
--- /dev/null
+++ b/arch/cris/arch-v32/lib/memset.c
@@ -0,0 +1,253 @@
+/*#************************************************************************#*/
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# FUNCTION NAME: memset()                                                 */
+/*#                                                                         */
+/*# PARAMETERS:  void* dst;   Destination address.                          */
+/*#              int     c;   Value of byte to write.                       */
+/*#              int   len;   Number of bytes to write.                     */
+/*#                                                                         */
+/*# RETURNS:     dst.                                                       */
+/*#                                                                         */
+/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */
+/*#              Framework taken from memcpy.  This routine is              */
+/*#              very sensitive to compiler changes in register allocation. */
+/*#              Should really be rewritten to avoid this problem.          */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# HISTORY                                                                 */
+/*#                                                                         */
+/*# DATE      NAME            CHANGES                                       */
+/*# ----      ----            -------                                       */
+/*# 990713    HP              Tired of watching this function (or           */
+/*#                           really, the nonoptimized generic              */
+/*#                           implementation) take up 90% of simulator      */
+/*#                           output.  Measurements needed.                 */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+
+#include <linux/types.h>
+
+/* No, there's no macro saying 12*4, since it is "hard" to get it into
+   the asm in a good way.  Thus better to expose the problem everywhere.
+   */
+
+/* Assuming 1 cycle per dword written or read (ok, not really true), and
+   one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1)
+   so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */
+
+#define ZERO_BLOCK_SIZE (1*12*4)
+
+void *memset(void *pdst,
+             int c,
+             size_t plen)
+{
+  /* Ok.  Now we want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this. */
+
+  register char *return_dst __asm__ ("r10") = pdst;
+  register int n __asm__ ("r12") = plen;
+  register int lc __asm__ ("r11") = c;
+
+  /* Most apps use memset sanely.  Only those memsetting about 3..4
+     bytes or less get penalized compared to the generic implementation
+     - and that's not really sane use. */
+
+  /* Ugh.  This is fragile at best.  Check with newer GCC releases, if
+     they compile cascaded "x |= x << 8" sanely! */
+  __asm__("movu.b %0,$r13	\n\
+           lslq 8,$r13		\n\
+	   move.b %0,$r13	\n\
+	   move.d $r13,%0	\n\
+	   lslq 16,$r13		\n\
+	   or.d $r13,%0"
+          : "=r" (lc) : "0" (lc) : "r13");
+
+  {
+    register char *dst __asm__ ("r13") = pdst;
+
+  /* This is NONPORTABLE, but since this whole routine is     */
+  /* grossly nonportable that doesn't matter.                 */
+
+  if (((unsigned long) pdst & 3) != 0
+     /* Oops! n=0 must be a legal call, regardless of alignment. */
+      && n >= 3)
+  {
+    if ((unsigned long)dst & 1)
+    {
+      *dst = (char) lc;
+      n--;
+      dst++;
+    }
+
+    if ((unsigned long)dst & 2)
+    {
+      *(short *)dst = lc;
+      n -= 2;
+      dst += 2;
+    }
+  }
+
+  /* Now the fun part.  For the threshold value of this, check the equation
+     above. */
+  /* Decide which copying method to use. */
+  if (n >= ZERO_BLOCK_SIZE)
+  {
+    /* For large copies we use 'movem' */
+
+  /* It is not optimal to tell the compiler about clobbering any
+     registers; that will move the saving/restoring of those registers
+     to the function prologue/epilogue, and make non-movem sizes
+     suboptimal.
+
+      This method is not foolproof; it assumes that the "asm reg"
+     declarations at the beginning of the function really are used
+     here (beware: they may be moved to temporary registers).
+      This way, we do not have to save/move the registers around into
+     temporaries; we can safely use them straight away.
+
+      If you want to check that the allocation was right; then
+      check the equalities in the first comment.  It should say
+      "r13=r13, r12=r12, r11=r11" */
+    __asm__ volatile ("							\n\
+        ;; Check that the register asm declaration got right.		\n\
+        ;; The GCC manual says it will work, but there *has* been bugs.	\n\
+	.ifnc %0-%1-%4,$r13-$r12-$r11					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll clobber in the movem process	\n\
+	;; on the stack.  Don't mention them to gcc, it will only be	\n\
+	;; upset.							\n\
+	subq 	11*4,$sp						\n\
+        movem   $r10,[$sp]						\n\
+									\n\
+        move.d  $r11,$r0						\n\
+        move.d  $r11,$r1						\n\
+        move.d  $r11,$r2						\n\
+        move.d  $r11,$r3						\n\
+        move.d  $r11,$r4						\n\
+        move.d  $r11,$r5						\n\
+        move.d  $r11,$r6						\n\
+        move.d  $r11,$r7						\n\
+        move.d  $r11,$r8						\n\
+        move.d  $r11,$r9						\n\
+        move.d  $r11,$r10						\n\
+									\n\
+        ;; Now we've got this:						\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+        ;; Update n for the first loop					\n\
+        subq    12*4,$r12						\n\
+0:									\n\
+        subq   12*4,$r12						\n\
+        bge     0b							\n\
+	movem	$r11,[$r13+]						\n\
+									\n\
+        addq   12*4,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+        movem [$sp+],$r10"
+
+     /* Outputs */ : "=r" (dst), "=r" (n)
+     /* Inputs */ : "0" (dst), "1" (n), "r" (lc));
+  }
+
+    /* Either we directly starts copying, using dword copying
+       in a loop, or we copy as much as possible with 'movem'
+       and then the last block (<44 bytes) is copied here.
+       This will work since 'movem' will have updated src,dst,n. */
+
+    while ( n >= 16 )
+    {
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      n -= 16;
+    }
+
+    /* A switch() is definitely the fastest although it takes a LOT of code.
+     * Particularly if you inline code this.
+     */
+    switch (n)
+    {
+      case 0:
+        break;
+      case 1:
+        *(char*)dst = (char) lc;
+        break;
+      case 2:
+        *(short*)dst = (short) lc;
+        break;
+      case 3:
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 4:
+        *((long*)dst)++ = lc;
+        break;
+      case 5:
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 6:
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 7:
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 8:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        break;
+      case 9:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 10:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 11:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 12:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        break;
+      case 13:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 14:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 15:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+    }
+  }
+
+  return return_dst; /* destination pointer. */
+} /* memset() */
diff --git a/arch/cris/arch-v32/lib/nand_init.S b/arch/cris/arch-v32/lib/nand_init.S
new file mode 100644
index 0000000..aba5c75
--- /dev/null
+++ b/arch/cris/arch-v32/lib/nand_init.S
@@ -0,0 +1,179 @@
+##=============================================================================
+##
+##      nand_init.S
+##
+##      The bootrom copies data from the NAND flash to the internal RAM but
+##      due to a bug/feature we can only trust the 256 first bytes. So this
+##      code copies more data from NAND flash to internal RAM. Obvioulsy this
+##      code must fit in the first 256 bytes so alter with care.
+##
+##	Some notes about the bug/feature for future reference:
+##        The bootrom copies the first 127 KB from NAND flash to internal
+##        memory. The problem is that it does a bytewise copy. NAND flashes
+##        does autoincrement on the address so for a 16-bite device each
+##        read/write increases the address by two. So the copy loop in the
+##        bootrom will discard every second byte. This is solved by inserting
+##        zeroes in every second byte in the first erase block.
+##
+##        The bootrom also incorrectly assumes that it can read the flash
+##        linear with only one read command but the flash will actually
+##        switch between normal area and spare area if you do that so we
+##        can't trust more than the first 256 bytes.
+##
+##=============================================================================
+
+#include <asm/arch/hwregs/asm/reg_map_asm.h>
+#include <asm/arch/hwregs/asm/gio_defs_asm.h>
+#include <asm/arch/hwregs/asm/pinmux_defs_asm.h>
+#include <asm/arch/hwregs/asm/bif_core_defs_asm.h>
+#include <asm/arch/hwregs/asm/config_defs_asm.h>
+#include <linux/config.h>
+
+;; There are 8-bit NAND flashes and 16-bit NAND flashes.
+;; We need to treat them slightly different.
+#if CONFIG_ETRAX_FLASH_BUSWIDTH==2
+#define PAGE_SIZE 256
+#else
+#error 2
+#define PAGE_SIZE 512
+#endif
+#define ERASE_BLOCK 16384
+
+;; GPIO pins connected to NAND flash
+#define CE 4
+#define CLE 5
+#define ALE 6
+#define BY 7
+
+;; Address space for NAND flash
+#define NAND_RD_ADDR 0x90000000
+#define NAND_WR_ADDR 0x94000000
+
+#define READ_CMD 0x00
+
+;; Readability macros
+#define CSP_MASK \
+	REG_MASK(bif_core, rw_grp3_cfg, gated_csp0) | \
+	REG_MASK(bif_core, rw_grp3_cfg, gated_csp1)
+#define CSP_VAL \
+	REG_STATE(bif_core, rw_grp3_cfg, gated_csp0, rd) | \
+	REG_STATE(bif_core, rw_grp3_cfg, gated_csp1, wr)
+
+;;----------------------------------------------------------------------------
+;; Macros to set/clear GPIO bits
+
+.macro SET x
+	or.b   (1<<\x),$r9
+	move.d $r9, [$r2]
+.endm
+
+.macro CLR x
+	and.b  ~(1<<\x),$r9
+	move.d $r9, [$r2]
+.endm
+
+;;----------------------------------------------------------------------------
+
+nand_boot:
+	;; Check if nand boot was selected
+	move.d REG_ADDR(config, regi_config, r_bootsel), $r0
+	move.d [$r0], $r0
+	and.d  REG_MASK(config, r_bootsel, boot_mode), $r0
+	cmp.d  REG_STATE(config, r_bootsel, boot_mode, nand), $r0
+	bne normal_boot ; No NAND boot
+	nop
+
+copy_nand_to_ram:
+	;; copy_nand_to_ram
+	;; Arguments
+	;;   r10 - destination
+	;;   r11 - source offset
+	;;   r12 - size
+	;;   r13 - Address to jump to after completion
+	;; Note : r10-r12 are clobbered on return
+	;; Registers used:
+	;;   r0 - NAND_RD_ADDR
+	;;   r1 - NAND_WR_ADDR
+	;;   r2 - reg_gio_rw_pa_dout
+	;;   r3 - reg_gio_r_pa_din
+	;;   r4 - tmp
+	;;   r5 - byte counter within a page
+	;;   r6 - reg_pinmux_rw_pa
+	;;   r7 - reg_gio_rw_pa_oe
+	;;   r8 - reg_bif_core_rw_grp3_cfg
+	;;   r9 - reg_gio_rw_pa_dout shadow
+	move.d 0x90000000, $r0
+	move.d 0x94000000, $r1
+	move.d REG_ADDR(gio, regi_gio, rw_pa_dout), $r2
+	move.d REG_ADDR(gio, regi_gio, r_pa_din), $r3
+	move.d REG_ADDR(pinmux, regi_pinmux, rw_pa), $r6
+	move.d REG_ADDR(gio, regi_gio, rw_pa_oe), $r7
+	move.d REG_ADDR(bif_core, regi_bif_core, rw_grp3_cfg), $r8
+
+#if CONFIG_ETRAX_FLASH_BUSWIDTH==2
+	lsrq	1, $r11
+#endif
+	;; Set up GPIO
+	move.d [$r2], $r9
+	move.d [$r7], $r4
+	or.b (1<<ALE) | (1 << CLE) | (1<<CE), $r4
+	move.d $r4, [$r7]
+
+	;; Set up bif
+	move.d [$r8], $r4
+	and.d CSP_MASK, $r4
+	or.d CSP_VAL, $r4
+	move.d $r4, [$r8]
+
+1:	;; Copy one page
+	CLR CE
+	SET CLE
+	moveq	READ_CMD, $r4
+	move.b	$r4, [$r1]
+	moveq	20, $r4
+2:	bne	2b
+	subq	1, $r4
+	CLR CLE
+	SET ALE
+	clear.w [$r1] 		; Column address = 0
+	move.d	$r11, $r4
+	lsrq	8, $r4
+	move.b	$r4, [$r1]	; Row address
+	lsrq	8, $r4
+	move.b	$r4, [$r1]	; Row adddress
+	moveq	20, $r4
+2:	bne	2b
+	subq	1, $r4
+	CLR ALE
+2:	move.d	[$r3], $r4
+	and.d	1 << BY, $r4
+	beq 2b
+	movu.w  PAGE_SIZE, $r5
+2:	; Copy one byte/word
+#if CONFIG_ETRAX_FLASH_BUSWIDTH==2
+	move.w  [$r0], $r4
+#else
+	move.b  [$r0], $r4
+#endif
+	subq	1, $r5
+	bne	2b
+#if CONFIG_ETRAX_FLASH_BUSWIDTH==2
+	move.w  $r4, [$r10+]
+	subu.w	PAGE_SIZE*2, $r12
+#else
+	move.b  $r4, [$r10+]
+	subu.w	PAGE_SIZE, $r12
+#endif
+	bpl	1b
+	addu.w	PAGE_SIZE, $r11
+
+	;; End of copy
+	jump	$r13
+	nop
+
+	;; This will warn if the code above is too large. If you consider
+	;; to remove this you don't understand the bug/feature.
+	.org 256
+	.org ERASE_BLOCK
+
+normal_boot:
diff --git a/arch/cris/arch-v32/lib/spinlock.S b/arch/cris/arch-v32/lib/spinlock.S
new file mode 100644
index 0000000..2437ae7
--- /dev/null
+++ b/arch/cris/arch-v32/lib/spinlock.S
@@ -0,0 +1,33 @@
+;; Core of the spinlock implementation
+;;
+;; Copyright (C) 2004 Axis Communications AB.
+;;
+;; Author: Mikael Starvik
+
+
+	.global cris_spin_lock
+	.global cris_spin_trylock
+
+	.text
+
+cris_spin_lock:
+	clearf	p
+1:	test.d	[$r10]
+	beq	1b
+	clearf	p
+	ax
+	clear.d [$r10]
+	bcs     1b
+	clearf	p
+	ret
+	nop
+
+cris_spin_trylock:
+	clearf	p
+1:	move.d	[$r10], $r11
+	ax
+	clear.d [$r10]
+        bcs	1b
+        clearf	p
+	ret
+	move.d	$r11,$r10
diff --git a/arch/cris/arch-v32/lib/string.c b/arch/cris/arch-v32/lib/string.c
new file mode 100644
index 0000000..98e282a
--- /dev/null
+++ b/arch/cris/arch-v32/lib/string.c
@@ -0,0 +1,219 @@
+/*#************************************************************************#*/
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# FUNCTION NAME: memcpy()                                                 */
+/*#                                                                         */
+/*# PARAMETERS:  void* dst;   Destination address.                          */
+/*#              void* src;   Source address.                               */
+/*#              int   len;   Number of bytes to copy.                      */
+/*#                                                                         */
+/*# RETURNS:     dst.                                                       */
+/*#                                                                         */
+/*# DESCRIPTION: Copies len bytes of memory from src to dst.  No guarantees */
+/*#              about copying of overlapping memory areas. This routine is */
+/*#              very sensitive to compiler changes in register allocation. */
+/*#              Should really be rewritten to avoid this problem.          */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# HISTORY                                                                 */
+/*#                                                                         */
+/*# DATE      NAME            CHANGES                                       */
+/*# ----      ----            -------                                       */
+/*# 941007    Kenny R         Creation                                      */
+/*# 941011    Kenny R         Lots of optimizations and inlining.           */
+/*# 941129    Ulf A           Adapted for use in libc.                      */
+/*# 950216    HP              N==0 forgotten if non-aligned src/dst.        */
+/*#                           Added some optimizations.                     */
+/*# 001025    HP              Make src and dst char *.  Align dst to	    */
+/*#			      dword, not just word-if-both-src-and-dst-	    */
+/*#			      are-misaligned.				    */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+
+#include <linux/types.h>
+
+void *memcpy(void *pdst,
+             const void *psrc,
+             size_t pn)
+{
+  /* Ok.  Now we want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+  register void *return_dst __asm__ ("r10") = pdst;
+  register char *dst __asm__ ("r13") = pdst;
+  register const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+
+
+  /* When src is aligned but not dst, this makes a few extra needless
+     cycles.  I believe it would take as many to check that the
+     re-alignment was unnecessary.  */
+  if (((unsigned long) dst & 3) != 0
+      /* Don't align if we wouldn't copy more than a few bytes; so we
+	 don't have to check further for overflows.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      n--;
+      *(char*)dst = *(char*)src;
+      src++;
+      dst++;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      n -= 2;
+      *(short*)dst = *(short*)src;
+      src += 2;
+      dst += 2;
+    }
+  }
+
+  /* Decide which copying method to use.  Movem is dirt cheap, so the
+     overheap is low enough to always use the minimum block size as the
+     threshold.  */
+  if (n >= 44)
+  {
+    /* For large copies we use 'movem' */
+
+  /* It is not optimal to tell the compiler about clobbering any
+     registers; that will move the saving/restoring of those registers
+     to the function prologue/epilogue, and make non-movem sizes
+     suboptimal.  */
+    __asm__ volatile ("							\n\
+        ;; Check that the register asm declaration got right.		\n\
+        ;; The GCC manual explicitly says TRT will happen.		\n\
+	.ifnc %0-%1-%2,$r13-$r11-$r12					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+									\n\
+	;; on the stack.						\n\
+	subq 	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+        ;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+        ;; Update n for the first loop					\n\
+        subq    44,$r12							\n\
+0:									\n\
+	movem	[$r11+],$r10						\n\
+        subq   44,$r12							\n\
+        bge     0b							\n\
+	movem	$r10,[$r13+]						\n\
+									\n\
+        addq   44,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+        movem [$sp+],$r10"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n)
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n));
+
+  }
+
+  /* Either we directly starts copying, using dword copying
+     in a loop, or we copy as much as possible with 'movem'
+     and then the last block (<44 bytes) is copied here.
+     This will work since 'movem' will have updated src,dst,n. */
+
+  while ( n >= 16 )
+  {
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    n -= 16;
+  }
+
+  /* A switch() is definitely the fastest although it takes a LOT of code.
+   * Particularly if you inline code this.
+   */
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      *(char*)dst = *(char*)src;
+      break;
+    case 2:
+      *(short*)dst = *(short*)src;
+      break;
+    case 3:
+      *((short*)dst)++ = *((short*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 4:
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 5:
+      *((long*)dst)++ = *((long*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 6:
+      *((long*)dst)++ = *((long*)src)++;
+      *(short*)dst = *(short*)src;
+      break;
+    case 7:
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 8:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 9:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 10:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *(short*)dst = *(short*)src;
+      break;
+    case 11:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 12:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 13:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+    case 14:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *(short*)dst = *(short*)src;
+      break;
+    case 15:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *(char*)dst = *(char*)src;
+      break;
+  }
+
+  return return_dst; /* destination pointer. */
+} /* memcpy() */
diff --git a/arch/cris/arch-v32/lib/usercopy.c b/arch/cris/arch-v32/lib/usercopy.c
new file mode 100644
index 0000000..f0b0846
--- /dev/null
+++ b/arch/cris/arch-v32/lib/usercopy.c
@@ -0,0 +1,470 @@
+/*
+ * User address space access functions.
+ * The non-inlined parts of asm-cris/uaccess.h are here.
+ *
+ * Copyright (C) 2000, 2003 Axis Communications AB.
+ *
+ * Written by Hans-Peter Nilsson.
+ * Pieces used from memcpy, originally by Kenny Ranerup long time ago.
+ */
+
+#include <asm/uaccess.h>
+
+/* Asm:s have been tweaked (within the domain of correctness) to give
+   satisfactory results for "gcc version 3.2.1 Axis release R53/1.53-v32".
+
+   Check regularly...
+
+   Note that for CRISv32, the PC saved at a bus-fault is the address
+   *at* the faulting instruction, with a special case for instructions
+   in delay slots: then it's the address of the branch.  Note also that
+   in contrast to v10, a postincrement in the instruction is *not*
+   performed at a bus-fault; the register is seen having the original
+   value in fault handlers.  */
+
+
+/* Copy to userspace.  This is based on the memcpy used for
+   kernel-to-kernel copying; see "string.c".  */
+
+unsigned long
+__copy_user (void __user *pdst, const void *psrc, unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+  register char *dst __asm__ ("r13") = pdst;
+  register const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+
+  /* When src is aligned but not dst, this makes a few extra needless
+     cycles.  I believe it would take as many to check that the
+     re-alignment was unnecessary.  */
+  if (((unsigned long) dst & 3) != 0
+      /* Don't align if we wouldn't copy more than a few bytes; so we
+	 don't have to check further for overflows.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      __asm_copy_to_user_1 (dst, src, retn);
+      n--;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      __asm_copy_to_user_2 (dst, src, retn);
+      n -= 2;
+    }
+  }
+
+  /* Movem is dirt cheap.  The overheap is low enough to always use the
+     minimum possible block size as the threshold.  */
+  if (n >= 44)
+  {
+    /* For large copies we use 'movem'.  */
+
+    /* It is not optimal to tell the compiler about clobbering any
+       registers; that will move the saving/restoring of those registers
+       to the function prologue/epilogue, and make non-movem sizes
+       suboptimal.  */
+    __asm__ volatile ("\
+        ;; Check that the register asm declaration got right.		\n\
+        ;; The GCC manual explicitly says TRT will happen.		\n\
+	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+	;; on the stack.						\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	44,$r12							\n\
+0:									\n\
+	movem	[$r11+],$r10						\n\
+	subq   44,$r12							\n\
+1:	bge	0b							\n\
+	movem	$r10,[$r13+]						\n\
+3:									\n\
+	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+2:									\n\
+	.section .fixup,\"ax\"						\n\
+4:									\n\
+; When failing on any of the 1..44 bytes in a chunk, we adjust back the	\n\
+; source pointer and just drop through	to the by-16 and by-4 loops to	\n\
+; get the correct number of failing bytes.  This necessarily means a	\n\
+; few extra exceptions, but invalid user pointers shouldn't happen in	\n\
+; time-critical code anyway.						\n\
+	jump 3b								\n\
+	subq 44,$r11							\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 1b,4b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
+
+  }
+
+  while (n >= 16)
+  {
+    __asm_copy_to_user_16 (dst, src, retn);
+    n -= 16;
+  }
+
+  /* Having a separate by-four loops cuts down on cache footprint.
+     FIXME:  Test with and without; increasing switch to be 0..15.  */
+  while (n >= 4)
+  {
+    __asm_copy_to_user_4 (dst, src, retn);
+    n -= 4;
+  }
+
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      __asm_copy_to_user_1 (dst, src, retn);
+      break;
+    case 2:
+      __asm_copy_to_user_2 (dst, src, retn);
+      break;
+    case 3:
+      __asm_copy_to_user_3 (dst, src, retn);
+      break;
+  }
+
+  return retn;
+}
+
+/* Copy from user to kernel, zeroing the bytes that were inaccessible in
+   userland.  The return-value is the number of bytes that were
+   inaccessible.  */
+
+unsigned long
+__copy_user_zeroing (void __user *pdst, const void *psrc, unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+     As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on.  */
+
+  register char *dst __asm__ ("r13") = pdst;
+  register const char *src __asm__ ("r11") = psrc;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+  /* The best reason to align src is that we then know that a read-fault
+     was for aligned bytes; there's no 1..3 remaining good bytes to
+     pickle.  */
+  if (((unsigned long) src & 3) != 0)
+  {
+    if (((unsigned long) src & 1) && n != 0)
+    {
+      __asm_copy_from_user_1 (dst, src, retn);
+      n--;
+    }
+
+    if (((unsigned long) src & 2) && n >= 2)
+    {
+      __asm_copy_from_user_2 (dst, src, retn);
+      n -= 2;
+    }
+
+    /* We only need one check after the unalignment-adjustments, because
+       if both adjustments were done, either both or neither reference
+       had an exception.  */
+    if (retn != 0)
+      goto copy_exception_bytes;
+  }
+
+  /* Movem is dirt cheap.  The overheap is low enough to always use the
+     minimum possible block size as the threshold.  */
+  if (n >= 44)
+  {
+    /* It is not optimal to tell the compiler about clobbering any
+       registers; that will move the saving/restoring of those registers
+       to the function prologue/epilogue, and make non-movem sizes
+       suboptimal.  */
+    __asm__ volatile ("\
+	.ifnc %0%1%2%3,$r13$r11$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+	;; on the stack.						\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	44,$r12							\n\
+0:									\n\
+	movem	[$r11+],$r10						\n\
+									\n\
+	subq   44,$r12							\n\
+	bge	0b							\n\
+	movem	$r10,[$r13+]						\n\
+									\n\
+4:									\n\
+	addq   44,$r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+	.section .fixup,\"ax\"						\n\
+									\n\
+;; Do not jump back into the loop if we fail.  For some uses, we get a	\n\
+;; page fault somewhere on the line.  Without checking for page limits,	\n\
+;; we don't know where, but we need to copy accurately and keep an	\n\
+;; accurate count; not just clear the whole line.  To do that, we fall	\n\
+;; down in the code below, proceeding with smaller amounts.  It should	\n\
+;; be kept in mind that we have to cater to code like what at one time	\n\
+;; was in fs/super.c:							\n\
+;;  i = size - copy_from_user((void *)page, data, size);		\n\
+;; which would cause repeated faults while clearing the remainder of	\n\
+;; the SIZE bytes at PAGE after the first fault.			\n\
+;; A caveat here is that we must not fall through from a failing page	\n\
+;; to a valid page.							\n\
+									\n\
+3:									\n\
+	jump	4b ;; Fall through, pretending the fault didn't happen.	\n\
+	nop								\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 0b,3b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn));
+  }
+
+  /* Either we directly start copying here, using dword copying in a loop,
+     or we copy as much as possible with 'movem' and then the last block
+     (<44 bytes) is copied here.  This will work since 'movem' will have
+     updated src, dst and n.  (Except with failing src.)
+
+     Since we want to keep src accurate, we can't use
+     __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and
+     retn, but not src (by design; it's value is ignored elsewhere).  */
+
+  while (n >= 4)
+  {
+    __asm_copy_from_user_4 (dst, src, retn);
+    n -= 4;
+
+    if (retn)
+      goto copy_exception_bytes;
+  }
+
+  /* If we get here, there were no memory read faults.  */
+  switch (n)
+  {
+    /* These copies are at least "naturally aligned" (so we don't have
+       to check each byte), due to the src alignment code before the
+       movem loop.  The *_3 case *will* get the correct count for retn.  */
+    case 0:
+      /* This case deliberately left in (if you have doubts check the
+	 generated assembly code).  */
+      break;
+    case 1:
+      __asm_copy_from_user_1 (dst, src, retn);
+      break;
+    case 2:
+      __asm_copy_from_user_2 (dst, src, retn);
+      break;
+    case 3:
+      __asm_copy_from_user_3 (dst, src, retn);
+      break;
+  }
+
+  /* If we get here, retn correctly reflects the number of failing
+     bytes.  */
+  return retn;
+
+copy_exception_bytes:
+  /* We already have "retn" bytes cleared, and need to clear the
+     remaining "n" bytes.  A non-optimized simple byte-for-byte in-line
+     memset is preferred here, since this isn't speed-critical code and
+     we'd rather have this a leaf-function than calling memset.  */
+  {
+    char *endp;
+    for (endp = dst + n; dst < endp; dst++)
+      *dst = 0;
+  }
+
+  return retn + n;
+}
+
+/* Zero userspace.  */
+
+unsigned long
+__do_clear_user (void __user *pto, unsigned long pn)
+{
+  /* We want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     FIXME: Comment for old gcc version.  Check.
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+  register char *dst __asm__ ("r13") = pto;
+  register int n __asm__ ("r12") = pn;
+  register int retn __asm__ ("r10") = 0;
+
+
+  if (((unsigned long) dst & 3) != 0
+     /* Don't align if we wouldn't copy more than a few bytes.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      __asm_clear_1 (dst, retn);
+      n--;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      __asm_clear_2 (dst, retn);
+      n -= 2;
+    }
+  }
+
+  /* Decide which copying method to use.
+     FIXME: This number is from the "ordinary" kernel memset.  */
+  if (n >= 48)
+  {
+    /* For large clears we use 'movem' */
+
+    /* It is not optimal to tell the compiler about clobbering any
+       call-saved registers; that will move the saving/restoring of
+       those registers to the function prologue/epilogue, and make
+       non-movem sizes suboptimal.
+
+       This method is not foolproof; it assumes that the "asm reg"
+       declarations at the beginning of the function really are used
+       here (beware: they may be moved to temporary registers).
+       This way, we do not have to save/move the registers around into
+       temporaries; we can safely use them straight away.
+
+      If you want to check that the allocation was right; then
+      check the equalities in the first comment.  It should say
+      something like "r13=r13, r11=r11, r12=r12". */
+    __asm__ volatile ("\
+	.ifnc %0%1%2,$r13$r12$r10					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll clobber in the movem process	\n\
+	;; on the stack.  Don't mention them to gcc, it will only be	\n\
+	;; upset.							\n\
+	subq	11*4,$sp						\n\
+	movem	$r10,[$sp]						\n\
+									\n\
+	clear.d $r0							\n\
+	clear.d $r1							\n\
+	clear.d $r2							\n\
+	clear.d $r3							\n\
+	clear.d $r4							\n\
+	clear.d $r5							\n\
+	clear.d $r6							\n\
+	clear.d $r7							\n\
+	clear.d $r8							\n\
+	clear.d $r9							\n\
+	clear.d $r10							\n\
+	clear.d $r11							\n\
+									\n\
+	;; Now we've got this:						\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+	;; Update n for the first loop					\n\
+	subq	12*4,$r12						\n\
+0:									\n\
+	subq   12*4,$r12						\n\
+1:									\n\
+	bge	0b							\n\
+	movem	$r11,[$r13+]						\n\
+									\n\
+	addq   12*4,$r12 ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+	movem [$sp+],$r10						\n\
+2:									\n\
+	.section .fixup,\"ax\"						\n\
+3:									\n\
+	movem [$sp],$r10						\n\
+	addq 12*4,$r10							\n\
+	addq 12*4,$r13							\n\
+	movem $r10,[$sp]						\n\
+	jump 0b								\n\
+	clear.d $r10							\n\
+									\n\
+	.previous							\n\
+	.section __ex_table,\"a\"					\n\
+	.dword 1b,3b							\n\
+	.previous"
+
+     /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn)
+     /* Inputs */ : "0" (dst), "1" (n), "2" (retn)
+     /* Clobber */ : "r11");
+  }
+
+  while (n >= 16)
+  {
+    __asm_clear_16 (dst, retn);
+    n -= 16;
+  }
+
+  /* Having a separate by-four loops cuts down on cache footprint.
+     FIXME:  Test with and without; increasing switch to be 0..15.  */
+  while (n >= 4)
+  {
+    __asm_clear_4 (dst, retn);
+    n -= 4;
+  }
+
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      __asm_clear_1 (dst, retn);
+      break;
+    case 2:
+      __asm_clear_2 (dst, retn);
+      break;
+    case 3:
+      __asm_clear_3 (dst, retn);
+      break;
+  }
+
+  return retn;
+}