[linux-2.6-block.git] / arch / mips / lib / csum_partial.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.
 *
 * Quick'n'dirty IP checksum ...
 *
 * Copyright (C) 1998, 1999 Ralf Baechle
 * Copyright (C) 1999 Silicon Graphics, Inc.
 * Copyright (C) 2007  Maciej W. Rozycki
 * Copyright (C) 2014 Imagination Technologies Ltd.
 */
#include <linux/errno.h>
#include <asm/asm.h>
#include <asm/asm-offsets.h>
#include <asm/regdef.h>

#ifdef CONFIG_64BIT
/*
 * As we are sharing code base with the mips32 tree (which use the o32 ABI
 * register definitions). We need to redefine the register definitions from
 * the n64 ABI register naming to the o32 ABI register naming.
 */
#undef t0
#undef t1
#undef t2
#undef t3
#define t0	$8
#define t1	$9
#define t2	$10
#define t3	$11
#define t4	$12
#define t5	$13
#define t6	$14
#define t7	$15

#define USE_DOUBLE
#endif

#ifdef USE_DOUBLE

#define LOAD   ld
#define LOAD32 lwu
#define ADD    daddu
#define NBYTES 8

#else

#define LOAD   lw
#define LOAD32 lw
#define ADD    addu
#define NBYTES 4

#endif /* USE_DOUBLE */

#define UNIT(unit)  ((unit)*NBYTES)

#define ADDC(sum,reg)						\
	.set	push;						\
	.set	noat;						\
	ADD	sum, reg;					\
	sltu	v1, sum, reg;					\
	ADD	sum, v1;					\
	.set	pop

#define ADDC32(sum,reg)						\
	.set	push;						\
	.set	noat;						\
	addu	sum, reg;					\
	sltu	v1, sum, reg;					\
	addu	sum, v1;					\
	.set	pop

#define CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3)	\
	LOAD	_t0, (offset + UNIT(0))(src);			\
	LOAD	_t1, (offset + UNIT(1))(src);			\
	LOAD	_t2, (offset + UNIT(2))(src);			\
	LOAD	_t3, (offset + UNIT(3))(src);			\
	ADDC(_t0, _t1);						\
	ADDC(_t2, _t3);						\
	ADDC(sum, _t0);						\
	ADDC(sum, _t2)

#ifdef USE_DOUBLE
#define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3)	\
	CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3)
#else
#define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3)	\
	CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3);	\
	CSUM_BIGCHUNK1(src, offset + 0x10, sum, _t0, _t1, _t2, _t3)
#endif

/*
 * a0: source address
 * a1: length of the area to checksum
 * a2: partial checksum
 */

#define src a0
#define sum v0

	.text
	.set	noreorder
	.align	5
LEAF(csum_partial)
	move	sum, zero
	move	t7, zero

	sltiu	t8, a1, 0x8
	bnez	t8, .Lsmall_csumcpy		/* < 8 bytes to copy */
	 move	t2, a1

	andi	t7, src, 0x1			/* odd buffer? */

.Lhword_align:
	beqz	t7, .Lword_align
	 andi	t8, src, 0x2

	lbu	t0, (src)
	LONG_SUBU	a1, a1, 0x1
#ifdef __MIPSEL__
	sll	t0, t0, 8
#endif
	ADDC(sum, t0)
	PTR_ADDU	src, src, 0x1
	andi	t8, src, 0x2

.Lword_align:
	beqz	t8, .Ldword_align
	 sltiu	t8, a1, 56

	lhu	t0, (src)
	LONG_SUBU	a1, a1, 0x2
	ADDC(sum, t0)
	sltiu	t8, a1, 56
	PTR_ADDU	src, src, 0x2

.Ldword_align:
	bnez	t8, .Ldo_end_words
	 move	t8, a1

	andi	t8, src, 0x4
	beqz	t8, .Lqword_align
	 andi	t8, src, 0x8

	LOAD32	t0, 0x00(src)
	LONG_SUBU	a1, a1, 0x4
	ADDC(sum, t0)
	PTR_ADDU	src, src, 0x4
	andi	t8, src, 0x8

.Lqword_align:
	beqz	t8, .Loword_align
	 andi	t8, src, 0x10

#ifdef USE_DOUBLE
	ld	t0, 0x00(src)
	LONG_SUBU	a1, a1, 0x8
	ADDC(sum, t0)
#else
	lw	t0, 0x00(src)
	lw	t1, 0x04(src)
	LONG_SUBU	a1, a1, 0x8
	ADDC(sum, t0)
	ADDC(sum, t1)
#endif
	PTR_ADDU	src, src, 0x8
	andi	t8, src, 0x10

.Loword_align:
	beqz	t8, .Lbegin_movement
	 LONG_SRL	t8, a1, 0x7

#ifdef USE_DOUBLE
	ld	t0, 0x00(src)
	ld	t1, 0x08(src)
	ADDC(sum, t0)
	ADDC(sum, t1)
#else
	CSUM_BIGCHUNK1(src, 0x00, sum, t0, t1, t3, t4)
#endif
	LONG_SUBU	a1, a1, 0x10
	PTR_ADDU	src, src, 0x10
	LONG_SRL	t8, a1, 0x7

.Lbegin_movement:
	beqz	t8, 1f
	 andi	t2, a1, 0x40

.Lmove_128bytes:
	CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
	CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
	CSUM_BIGCHUNK(src, 0x40, sum, t0, t1, t3, t4)
	CSUM_BIGCHUNK(src, 0x60, sum, t0, t1, t3, t4)
	LONG_SUBU	t8, t8, 0x01
	.set	reorder				/* DADDI_WAR */
	PTR_ADDU	src, src, 0x80
	bnez	t8, .Lmove_128bytes
	.set	noreorder

1:
	beqz	t2, 1f
	 andi	t2, a1, 0x20

.Lmove_64bytes:
	CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
	CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
	PTR_ADDU	src, src, 0x40

1:
	beqz	t2, .Ldo_end_words
	 andi	t8, a1, 0x1c

.Lmove_32bytes:
	CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
	andi	t8, a1, 0x1c
	PTR_ADDU	src, src, 0x20

.Ldo_end_words:
	beqz	t8, .Lsmall_csumcpy
	 andi	t2, a1, 0x3
	LONG_SRL	t8, t8, 0x2

.Lend_words:
	LOAD32	t0, (src)
	LONG_SUBU	t8, t8, 0x1
	ADDC(sum, t0)
	.set	reorder				/* DADDI_WAR */
	PTR_ADDU	src, src, 0x4
	bnez	t8, .Lend_words
	.set	noreorder

/* unknown src alignment and < 8 bytes to go  */
.Lsmall_csumcpy:
	move	a1, t2

	andi	t0, a1, 4
	beqz	t0, 1f
	 andi	t0, a1, 2

	/* Still a full word to go  */
	ulw	t1, (src)
	PTR_ADDIU	src, 4
#ifdef USE_DOUBLE
	dsll	t1, t1, 32			/* clear lower 32bit */
#endif
	ADDC(sum, t1)

1:	move	t1, zero
	beqz	t0, 1f
	 andi	t0, a1, 1

	/* Still a halfword to go  */
	ulhu	t1, (src)
	PTR_ADDIU	src, 2

1:	beqz	t0, 1f
	 sll	t1, t1, 16

	lbu	t2, (src)
	 nop

#ifdef __MIPSEB__
	sll	t2, t2, 8
#endif
	or	t1, t2

1:	ADDC(sum, t1)

	/* fold checksum */
#ifdef USE_DOUBLE
	dsll32	v1, sum, 0
	daddu	sum, v1
	sltu	v1, sum, v1
	dsra32	sum, sum, 0
	addu	sum, v1
#endif

	/* odd buffer alignment? */
#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
	.set	push
	.set	arch=mips32r2
	wsbh	v1, sum
	movn	sum, v1, t7
	.set	pop
#else
	beqz	t7, 1f			/* odd buffer alignment? */
	 lui	v1, 0x00ff
	addu	v1, 0x00ff
	and	t0, sum, v1
	sll	t0, t0, 8
	srl	sum, sum, 8
	and	sum, sum, v1
	or	sum, sum, t0
1:
#endif
	.set	reorder
	/* Add the passed partial csum.	 */
	ADDC32(sum, a2)
	jr	ra
	.set	noreorder
	END(csum_partial)


/*
 * checksum and copy routines based on memcpy.S
 *
 *	csum_partial_copy_nocheck(src, dst, len, sum)
 *	__csum_partial_copy_kernel(src, dst, len, sum, errp)
 *
 * See "Spec" in memcpy.S for details.	Unlike __copy_user, all
 * function in this file use the standard calling convention.
 */

#define src a0
#define dst a1
#define len a2
#define psum a3
#define sum v0
#define odd t8
#define errptr t9

/*
 * The exception handler for loads requires that:
 *  1- AT contain the address of the byte just past the end of the source
 *     of the copy,
 *  2- src_entry <= src < AT, and
 *  3- (dst - src) == (dst_entry - src_entry),
 * The _entry suffix denotes values when __copy_user was called.
 *
 * (1) is set up up by __csum_partial_copy_from_user and maintained by
 *	not writing AT in __csum_partial_copy
 * (2) is met by incrementing src by the number of bytes copied
 * (3) is met by not doing loads between a pair of increments of dst and src
 *
 * The exception handlers for stores stores -EFAULT to errptr and return.
 * These handlers do not need to overwrite any data.
 */

/* Instruction type */
#define LD_INSN 1
#define ST_INSN 2
#define LEGACY_MODE 1
#define EVA_MODE    2
#define USEROP   1
#define KERNELOP 2

/*
 * Wrapper to add an entry in the exception table
 * in case the insn causes a memory exception.
 * Arguments:
 * insn    : Load/store instruction
 * type    : Instruction type
 * reg     : Register
 * addr    : Address
 * handler : Exception handler
 */
#define EXC(insn, type, reg, addr, handler)	\
	.if \mode == LEGACY_MODE;		\
9:		insn reg, addr;			\
		.section __ex_table,"a";	\
		PTR	9b, handler;		\
		.previous;			\
	/* This is enabled in EVA mode */	\
	.else;					\
		/* If loading from user or storing to user */	\
		.if ((\from == USEROP) && (type == LD_INSN)) || \
		    ((\to == USEROP) && (type == ST_INSN));	\
9:			__BUILD_EVA_INSN(insn##e, reg, addr);	\
			.section __ex_table,"a";		\
			PTR	9b, handler;			\
			.previous;				\
		.else;						\
			/* EVA without exception */		\
			insn reg, addr;				\
		.endif;						\
	.endif

#undef LOAD

#ifdef USE_DOUBLE

#define LOADK	ld /* No exception */
#define LOAD(reg, addr, handler)	EXC(ld, LD_INSN, reg, addr, handler)
#define LOADBU(reg, addr, handler)	EXC(lbu, LD_INSN, reg, addr, handler)
#define LOADL(reg, addr, handler)	EXC(ldl, LD_INSN, reg, addr, handler)
#define LOADR(reg, addr, handler)	EXC(ldr, LD_INSN, reg, addr, handler)
#define STOREB(reg, addr, handler)	EXC(sb, ST_INSN, reg, addr, handler)
#define STOREL(reg, addr, handler)	EXC(sdl, ST_INSN, reg, addr, handler)
#define STORER(reg, addr, handler)	EXC(sdr, ST_INSN, reg, addr, handler)
#define STORE(reg, addr, handler)	EXC(sd, ST_INSN, reg, addr, handler)
#define ADD    daddu
#define SUB    dsubu
#define SRL    dsrl
#define SLL    dsll
#define SLLV   dsllv
#define SRLV   dsrlv
#define NBYTES 8
#define LOG_NBYTES 3

#else

#define LOADK	lw /* No exception */
#define LOAD(reg, addr, handler)	EXC(lw, LD_INSN, reg, addr, handler)
#define LOADBU(reg, addr, handler)	EXC(lbu, LD_INSN, reg, addr, handler)
#define LOADL(reg, addr, handler)	EXC(lwl, LD_INSN, reg, addr, handler)
#define LOADR(reg, addr, handler)	EXC(lwr, LD_INSN, reg, addr, handler)
#define STOREB(reg, addr, handler)	EXC(sb, ST_INSN, reg, addr, handler)
#define STOREL(reg, addr, handler)	EXC(swl, ST_INSN, reg, addr, handler)
#define STORER(reg, addr, handler)	EXC(swr, ST_INSN, reg, addr, handler)
#define STORE(reg, addr, handler)	EXC(sw, ST_INSN, reg, addr, handler)
#define ADD    addu
#define SUB    subu
#define SRL    srl
#define SLL    sll
#define SLLV   sllv
#define SRLV   srlv
#define NBYTES 4
#define LOG_NBYTES 2

#endif /* USE_DOUBLE */

#ifdef CONFIG_CPU_LITTLE_ENDIAN
#define LDFIRST LOADR
#define LDREST	LOADL
#define STFIRST STORER
#define STREST	STOREL
#define SHIFT_DISCARD SLLV
#define SHIFT_DISCARD_REVERT SRLV
#else
#define LDFIRST LOADL
#define LDREST	LOADR
#define STFIRST STOREL
#define STREST	STORER
#define SHIFT_DISCARD SRLV
#define SHIFT_DISCARD_REVERT SLLV
#endif

#define FIRST(unit) ((unit)*NBYTES)
#define REST(unit)  (FIRST(unit)+NBYTES-1)

#define ADDRMASK (NBYTES-1)

#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
	.set	noat
#else
	.set	at=v1
#endif

	.macro __BUILD_CSUM_PARTIAL_COPY_USER mode, from, to, __nocheck

	PTR_ADDU	AT, src, len	/* See (1) above. */
	/* initialize __nocheck if this the first time we execute this
	 * macro
	 */
#ifdef CONFIG_64BIT
	move	errptr, a4
#else
	lw	errptr, 16(sp)
#endif
	.if \__nocheck == 1
	FEXPORT(csum_partial_copy_nocheck)
	.endif
	move	sum, zero
	move	odd, zero
	/*
	 * Note: dst & src may be unaligned, len may be 0
	 * Temps
	 */
	/*
	 * The "issue break"s below are very approximate.
	 * Issue delays for dcache fills will perturb the schedule, as will
	 * load queue full replay traps, etc.
	 *
	 * If len < NBYTES use byte operations.
	 */
	sltu	t2, len, NBYTES
	and	t1, dst, ADDRMASK
	bnez	t2, .Lcopy_bytes_checklen\@
	 and	t0, src, ADDRMASK
	andi	odd, dst, 0x1			/* odd buffer? */
	bnez	t1, .Ldst_unaligned\@
	 nop
	bnez	t0, .Lsrc_unaligned_dst_aligned\@
	/*
	 * use delay slot for fall-through
	 * src and dst are aligned; need to compute rem
	 */
.Lboth_aligned\@:
	 SRL	t0, len, LOG_NBYTES+3	 # +3 for 8 units/iter
	beqz	t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
	 nop
	SUB	len, 8*NBYTES		# subtract here for bgez loop
	.align	4
1:
	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
	LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
	LOAD(t5, UNIT(5)(src), .Ll_exc_copy\@)
	LOAD(t6, UNIT(6)(src), .Ll_exc_copy\@)
	LOAD(t7, UNIT(7)(src), .Ll_exc_copy\@)
	SUB	len, len, 8*NBYTES
	ADD	src, src, 8*NBYTES
	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
	ADDC(t0, t1)
	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
	ADDC(sum, t0)
	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
	ADDC(t2, t3)
	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
	ADDC(sum, t2)
	STORE(t4, UNIT(4)(dst),	.Ls_exc\@)
	ADDC(t4, t5)
	STORE(t5, UNIT(5)(dst),	.Ls_exc\@)
	ADDC(sum, t4)
	STORE(t6, UNIT(6)(dst),	.Ls_exc\@)
	ADDC(t6, t7)
	STORE(t7, UNIT(7)(dst),	.Ls_exc\@)
	ADDC(sum, t6)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, 8*NBYTES
	bgez	len, 1b
	.set	noreorder
	ADD	len, 8*NBYTES		# revert len (see above)

	/*
	 * len == the number of bytes left to copy < 8*NBYTES
	 */
.Lcleanup_both_aligned\@:
#define rem t7
	beqz	len, .Ldone\@
	 sltu	t0, len, 4*NBYTES
	bnez	t0, .Lless_than_4units\@
	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
	/*
	 * len >= 4*NBYTES
	 */
	LOAD(t0, UNIT(0)(src), .Ll_exc\@)
	LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
	LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
	LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
	SUB	len, len, 4*NBYTES
	ADD	src, src, 4*NBYTES
	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
	ADDC(t0, t1)
	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
	ADDC(sum, t0)
	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
	ADDC(t2, t3)
	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
	ADDC(sum, t2)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, 4*NBYTES
	beqz	len, .Ldone\@
	.set	noreorder
.Lless_than_4units\@:
	/*
	 * rem = len % NBYTES
	 */
	beq	rem, len, .Lcopy_bytes\@
	 nop
1:
	LOAD(t0, 0(src), .Ll_exc\@)
	ADD	src, src, NBYTES
	SUB	len, len, NBYTES
	STORE(t0, 0(dst), .Ls_exc\@)
	ADDC(sum, t0)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, NBYTES
	bne	rem, len, 1b
	.set	noreorder

	/*
	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
	 * A loop would do only a byte at a time with possible branch
	 * mispredicts.	 Can't do an explicit LOAD dst,mask,or,STORE
	 * because can't assume read-access to dst.  Instead, use
	 * STREST dst, which doesn't require read access to dst.
	 *
	 * This code should perform better than a simple loop on modern,
	 * wide-issue mips processors because the code has fewer branches and
	 * more instruction-level parallelism.
	 */
#define bits t2
	beqz	len, .Ldone\@
	 ADD	t1, dst, len	# t1 is just past last byte of dst
	li	bits, 8*NBYTES
	SLL	rem, len, 3	# rem = number of bits to keep
	LOAD(t0, 0(src), .Ll_exc\@)
	SUB	bits, bits, rem # bits = number of bits to discard
	SHIFT_DISCARD t0, t0, bits
	STREST(t0, -1(t1), .Ls_exc\@)
	SHIFT_DISCARD_REVERT t0, t0, bits
	.set reorder
	ADDC(sum, t0)
	b	.Ldone\@
	.set noreorder
.Ldst_unaligned\@:
	/*
	 * dst is unaligned
	 * t0 = src & ADDRMASK
	 * t1 = dst & ADDRMASK; T1 > 0
	 * len >= NBYTES
	 *
	 * Copy enough bytes to align dst
	 * Set match = (src and dst have same alignment)
	 */
#define match rem
	LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
	ADD	t2, zero, NBYTES
	LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
	SUB	t2, t2, t1	# t2 = number of bytes copied
	xor	match, t0, t1
	STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
	SLL	t4, t1, 3		# t4 = number of bits to discard
	SHIFT_DISCARD t3, t3, t4
	/* no SHIFT_DISCARD_REVERT to handle odd buffer properly */
	ADDC(sum, t3)
	beq	len, t2, .Ldone\@
	 SUB	len, len, t2
	ADD	dst, dst, t2
	beqz	match, .Lboth_aligned\@
	 ADD	src, src, t2

.Lsrc_unaligned_dst_aligned\@:
	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
	beqz	t0, .Lcleanup_src_unaligned\@
	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
1:
/*
 * Avoid consecutive LD*'s to the same register since some mips
 * implementations can't issue them in the same cycle.
 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
 * are to the same unit (unless src is aligned, but it's not).
 */
	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
	LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
	SUB	len, len, 4*NBYTES
	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
	LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
	LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
	LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
	LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
	LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
	ADD	src, src, 4*NBYTES
#ifdef CONFIG_CPU_SB1
	nop				# improves slotting
#endif
	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
	ADDC(t0, t1)
	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
	ADDC(sum, t0)
	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
	ADDC(t2, t3)
	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
	ADDC(sum, t2)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, 4*NBYTES
	bne	len, rem, 1b
	.set	noreorder

.Lcleanup_src_unaligned\@:
	beqz	len, .Ldone\@
	 and	rem, len, NBYTES-1  # rem = len % NBYTES
	beq	rem, len, .Lcopy_bytes\@
	 nop
1:
	LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
	LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
	ADD	src, src, NBYTES
	SUB	len, len, NBYTES
	STORE(t0, 0(dst), .Ls_exc\@)
	ADDC(sum, t0)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, NBYTES
	bne	len, rem, 1b
	.set	noreorder

.Lcopy_bytes_checklen\@:
	beqz	len, .Ldone\@
	 nop
.Lcopy_bytes\@:
	/* 0 < len < NBYTES  */
#ifdef CONFIG_CPU_LITTLE_ENDIAN
#define SHIFT_START 0
#define SHIFT_INC 8
#else
#define SHIFT_START 8*(NBYTES-1)
#define SHIFT_INC -8
#endif
	move	t2, zero	# partial word
	li	t3, SHIFT_START # shift
/* use .Ll_exc_copy here to return correct sum on fault */
#define COPY_BYTE(N)			\
	LOADBU(t0, N(src), .Ll_exc_copy\@);	\
	SUB	len, len, 1;		\
	STOREB(t0, N(dst), .Ls_exc\@);	\
	SLLV	t0, t0, t3;		\
	addu	t3, SHIFT_INC;		\
	beqz	len, .Lcopy_bytes_done\@; \
	 or	t2, t0

	COPY_BYTE(0)
	COPY_BYTE(1)
#ifdef USE_DOUBLE
	COPY_BYTE(2)
	COPY_BYTE(3)
	COPY_BYTE(4)
	COPY_BYTE(5)
#endif
	LOADBU(t0, NBYTES-2(src), .Ll_exc_copy\@)
	SUB	len, len, 1
	STOREB(t0, NBYTES-2(dst), .Ls_exc\@)
	SLLV	t0, t0, t3
	or	t2, t0
.Lcopy_bytes_done\@:
	ADDC(sum, t2)
.Ldone\@:
	/* fold checksum */
	.set	push
	.set	noat
#ifdef USE_DOUBLE
	dsll32	v1, sum, 0
	daddu	sum, v1
	sltu	v1, sum, v1
	dsra32	sum, sum, 0
	addu	sum, v1
#endif

#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
	.set	push
	.set	arch=mips32r2
	wsbh	v1, sum
	movn	sum, v1, odd
	.set	pop
#else
	beqz	odd, 1f			/* odd buffer alignment? */
	 lui	v1, 0x00ff
	addu	v1, 0x00ff
	and	t0, sum, v1
	sll	t0, t0, 8
	srl	sum, sum, 8
	and	sum, sum, v1
	or	sum, sum, t0
1:
#endif
	.set	pop
	.set reorder
	ADDC32(sum, psum)
	jr	ra
	.set noreorder

.Ll_exc_copy\@:
	/*
	 * Copy bytes from src until faulting load address (or until a
	 * lb faults)
	 *
	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
	 * may be more than a byte beyond the last address.
	 * Hence, the lb below may get an exception.
	 *
	 * Assumes src < THREAD_BUADDR($28)
	 */
	LOADK	t0, TI_TASK($28)
	 li	t2, SHIFT_START
	LOADK	t0, THREAD_BUADDR(t0)
1:
	LOADBU(t1, 0(src), .Ll_exc\@)
	ADD	src, src, 1
	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
	SLLV	t1, t1, t2
	addu	t2, SHIFT_INC
	ADDC(sum, t1)
	.set	reorder				/* DADDI_WAR */
	ADD	dst, dst, 1
	bne	src, t0, 1b
	.set	noreorder
.Ll_exc\@:
	LOADK	t0, TI_TASK($28)
	 nop
	LOADK	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
	 nop
	SUB	len, AT, t0		# len number of uncopied bytes
	/*
	 * Here's where we rely on src and dst being incremented in tandem,
	 *   See (3) above.
	 * dst += (fault addr - src) to put dst at first byte to clear
	 */
	ADD	dst, t0			# compute start address in a1
	SUB	dst, src
	/*
	 * Clear len bytes starting at dst.  Can't call __bzero because it
	 * might modify len.  An inefficient loop for these rare times...
	 */
	.set	reorder				/* DADDI_WAR */
	SUB	src, len, 1
	beqz	len, .Ldone\@
	.set	noreorder
1:	sb	zero, 0(dst)
	ADD	dst, dst, 1
	.set	push
	.set	noat
#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
	bnez	src, 1b
	 SUB	src, src, 1
#else
	li	v1, 1
	bnez	src, 1b
	 SUB	src, src, v1
#endif
	li	v1, -EFAULT
	b	.Ldone\@
	 sw	v1, (errptr)

.Ls_exc\@:
	li	v0, -1 /* invalid checksum */
	li	v1, -EFAULT
	jr	ra
	 sw	v1, (errptr)
	.set	pop
	.endm

LEAF(__csum_partial_copy_kernel)
#ifndef CONFIG_EVA
FEXPORT(__csum_partial_copy_to_user)
FEXPORT(__csum_partial_copy_from_user)
#endif
__BUILD_CSUM_PARTIAL_COPY_USER LEGACY_MODE USEROP USEROP 1
END(__csum_partial_copy_kernel)

#ifdef CONFIG_EVA
LEAF(__csum_partial_copy_to_user)
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE KERNELOP USEROP 0
END(__csum_partial_copy_to_user)

LEAF(__csum_partial_copy_from_user)
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE USEROP KERNELOP 0
END(__csum_partial_copy_from_user)
#endif