/* Optimized strcmp for Xtensa.
   Copyright (C) 2001, 2007 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, see
   <http://www.gnu.org/licenses/>.  */

#include <sysdep.h>
#include <bits/xtensa-config.h>
#include <features.h>

#ifdef __XTENSA_EB__
#define	MASK0 0xff000000
#define	MASK1 0x00ff0000
#define	MASK2 0x0000ff00
#define	MASK3 0x000000ff
#else
#define	MASK0 0x000000ff
#define	MASK1 0x0000ff00
#define	MASK2 0x00ff0000
#define	MASK3 0xff000000
#endif

#define MASK4 0x40404040

	.text
	.align 4
	.literal_position
	.literal .Lmask0, MASK0
	.literal .Lmask1, MASK1
	.literal .Lmask2, MASK2
	.literal .Lmask3, MASK3
	.literal .Lmask4, MASK4
ENTRY (strcmp)
	/* a2 = s1, a3 = s2 */

	l8ui	a8, a2, 0	/* byte 0 from s1 */
	l8ui	a9, a3, 0	/* byte 0 from s2 */
	movi	a10, 3		/* mask */
	bne	a8, a9, .Lretdiff

	or	a11, a2, a3
	bnone	a11, a10, .Laligned

	xor	a11, a2, a3	/* compare low two bits of s1 and s2 */
	bany	a11, a10, .Lunaligned	/* if they have different alignment */

	/* s1/s2 are not word-aligned.  */
	addi	a2, a2, 1	/* advance s1 */
	beqz	a8, .Leq	/* bytes equal, if zero, strings are equal */
	addi	a3, a3, 1	/* advance s2 */
	bnone	a2, a10, .Laligned /* if s1/s2 now aligned */
	l8ui	a8, a2, 0	/* byte 1 from s1 */
	l8ui	a9, a3, 0	/* byte 1 from s2 */
	addi	a2, a2, 1	/* advance s1 */
	bne	a8, a9, .Lretdiff /* if different, return difference */
	beqz	a8, .Leq	/* bytes equal, if zero, strings are equal */
	addi	a3, a3, 1	/* advance s2 */
	bnone	a2, a10, .Laligned /* if s1/s2 now aligned */
	l8ui	a8, a2, 0	/* byte 2 from s1 */
	l8ui	a9, a3, 0	/* byte 2 from s2 */
	addi	a2, a2, 1	/* advance s1 */
	bne	a8, a9, .Lretdiff /* if different, return difference */
	beqz	a8, .Leq	/* bytes equal, if zero, strings are equal */
	addi	a3, a3, 1	/* advance s2 */
	j	.Laligned

/* s1 and s2 have different alignment.

   If the zero-overhead loop option is available, use an (almost)
   infinite zero-overhead loop with conditional exits so we only pay
   for taken branches when exiting the loop.

   Note: It is important for this unaligned case to come before the
   code for aligned strings, because otherwise some of the branches
   above cannot reach and have to be transformed to branches around
   jumps.  The unaligned code is smaller and the branches can reach
   over it.  */

	.align	4
	/* (2 mod 4) alignment for loop instruction */
.Lunaligned:
#if XCHAL_HAVE_LOOPS
	_movi.n	a8, 0		/* set up for the maximum loop count */
	loop	a8, .Lretdiff	/* loop forever (almost anyway) */
#endif
.Lnextbyte:
	l8ui	a8, a2, 0
	l8ui	a9, a3, 0
	addi	a2, a2, 1
	bne	a8, a9, .Lretdiff
	addi	a3, a3, 1
#if XCHAL_HAVE_LOOPS
	beqz	a8, .Lretdiff
#else
	bnez	a8, .Lnextbyte
#endif
.Lretdiff:
	sub	a2, a8, a9
	retw

/* s1 is word-aligned; s2 is word-aligned.

   If the zero-overhead loop option is available, use an (almost)
   infinite zero-overhead loop with conditional exits so we only pay
   for taken branches when exiting the loop.  */

/* New algorithm, relying on the fact that all normal ASCII is between
   32 and 127.

   Rather than check all bytes for zero:
   Take one word (4 bytes).  Call it w1.
   Shift w1 left by one into w1'.
   Or w1 and w1'.  For all normal ASCII bit 6 will be 1; for zero it won't.
   Check that all 4 bit 6's (one for each byte) are one:
   If they are, we are definitely not done.
   If they are not, we are probably done, but need to check for zero.  */

	.align	4
#if XCHAL_HAVE_LOOPS
.Laligned:
	.begin	no-transform
	l32r	a4, .Lmask0	/* mask for byte 0 */
	l32r	a7, .Lmask4
	/* Loop forever.  (a4 is more than than the maximum number
	   of iterations) */
	loop	a4, .Laligned_done

	/* First unrolled loop body.  */
	l32i	a8, a2, 0	/* get word from s1 */
	l32i	a9, a3, 0	/* get word from s2 */
	slli	a5, a8, 1
	bne	a8, a9, .Lwne2
	or	a9, a8, a5
	bnall	a9, a7, .Lprobeq

	/* Second unrolled loop body.  */
	l32i	a8, a2, 4	/* get word from s1+4 */
	l32i	a9, a3, 4	/* get word from s2+4 */
	slli	a5, a8, 1
	bne	a8, a9, .Lwne2
	or	a9, a8, a5
	bnall	a9, a7, .Lprobeq2

	addi	a2, a2, 8	/* advance s1 pointer */
	addi	a3, a3, 8	/* advance s2 pointer */
.Laligned_done:
	or	a1, a1, a1	/* nop */

.Lprobeq2:
	/* Adjust pointers to account for the loop unrolling.  */
	addi	a2, a2, 4
	addi	a3, a3, 4

#else /* !XCHAL_HAVE_LOOPS */

.Laligned:
	movi	a4, MASK0	/* mask for byte 0 */
	movi	a7, MASK4
	j	.Lfirstword
.Lnextword:
	addi	a2, a2, 4	/* advance s1 pointer */
	addi	a3, a3, 4	/* advance s2 pointer */
.Lfirstword:
	l32i	a8, a2, 0	/* get word from s1 */
	l32i	a9, a3, 0	/* get word from s2 */
	slli	a5, a8, 1
	bne	a8, a9, .Lwne2
	or	a9, a8, a5
	ball	a9, a7, .Lnextword
#endif /* !XCHAL_HAVE_LOOPS */

	/* align (0 mod 4) */
.Lprobeq:
	/* Words are probably equal, but check for sure.
	   If not, loop over the rest of string using normal algorithm.  */

	bnone	a8, a4, .Leq	/* if byte 0 is zero */
	l32r	a5, .Lmask1	/* mask for byte 1 */
	l32r	a6, .Lmask2	/* mask for byte 2 */
	bnone	a8, a5, .Leq	/* if byte 1 is zero */
	l32r	a7, .Lmask3	/* mask for byte 3 */
	bnone	a8, a6, .Leq	/* if byte 2 is zero */
	bnone	a8, a7, .Leq	/* if byte 3 is zero */
	addi.n	a2, a2, 4	/* advance s1 pointer */
	addi.n	a3, a3, 4	/* advance s2 pointer */
#if XCHAL_HAVE_LOOPS

	/* align (1 mod 4) */
	loop	a4, .Leq	/* loop forever (a4 is bigger than max iters) */
	.end	no-transform

	l32i	a8, a2, 0	/* get word from s1 */
	l32i	a9, a3, 0	/* get word from s2 */
	addi	a2, a2, 4	/* advance s1 pointer */
	bne	a8, a9, .Lwne
	bnone	a8, a4, .Leq	/* if byte 0 is zero */
	bnone	a8, a5, .Leq	/* if byte 1 is zero */
	bnone	a8, a6, .Leq	/* if byte 2 is zero */
	bnone	a8, a7, .Leq	/* if byte 3 is zero */
	addi	a3, a3, 4	/* advance s2 pointer */

#else /* !XCHAL_HAVE_LOOPS */

	j	.Lfirstword2
.Lnextword2:
	addi	a3, a3, 4	/* advance s2 pointer */
.Lfirstword2:
	l32i	a8, a2, 0	/* get word from s1 */
	l32i	a9, a3, 0	/* get word from s2 */
	addi	a2, a2, 4	/* advance s1 pointer */
	bne	a8, a9, .Lwne
	bnone	a8, a4, .Leq	/* if byte 0 is zero */
	bnone	a8, a5, .Leq	/* if byte 1 is zero */
	bnone	a8, a6, .Leq	/* if byte 2 is zero */
	bany	a8, a7, .Lnextword2	/* if byte 3 is zero */
#endif /* !XCHAL_HAVE_LOOPS */

	/* Words are equal; some byte is zero.  */
.Leq:	movi	a2, 0		/* return equal */
	retw

.Lwne2:	/* Words are not equal.  On big-endian processors, if none of the
	   bytes are zero, the return value can be determined by a simple
	   comparison.  */
#ifdef __XTENSA_EB__
	or	a10, a8, a5
	bnall	a10, a7, .Lsomezero
	bgeu	a8, a9, .Lposreturn
	movi	a2, -1
	retw
.Lposreturn:
	movi	a2, 1
	retw
.Lsomezero:	/* There is probably some zero byte. */
#endif /* __XTENSA_EB__ */
.Lwne:	/* Words are not equal.  */
	xor	a2, a8, a9	/* get word with nonzero in byte that differs */
	bany	a2, a4, .Ldiff0	/* if byte 0 differs */
	movi	a5, MASK1	/* mask for byte 1 */
	bnone	a8, a4, .Leq	/* if byte 0 is zero */
	bany	a2, a5, .Ldiff1	/* if byte 1 differs */
	movi	a6, MASK2	/* mask for byte 2 */
	bnone	a8, a5, .Leq	/* if byte 1 is zero */
	bany	a2, a6, .Ldiff2	/* if byte 2 differs */
	bnone	a8, a6, .Leq	/* if byte 2 is zero */
#ifdef __XTENSA_EB__
.Ldiff3:
.Ldiff2:
.Ldiff1:
	/* Byte 0 is equal (at least) and there is a difference before a zero
	   byte.  Just subtract words to get the return value.
	   The high order equal bytes cancel, leaving room for the sign.  */
	sub	a2, a8, a9
	retw

.Ldiff0:
	/* Need to make room for the sign, so can't subtract whole words.  */
	extui	a10, a8, 24, 8
	extui	a11, a9, 24, 8
	sub	a2, a10, a11
	retw

#else /* !__XTENSA_EB__ */
	/* Little-endian is a little more difficult because can't subtract
	   whole words.  */
.Ldiff3:
	/* Bytes 0-2 are equal; byte 3 is different.
	   For little-endian need to have a sign bit for the difference.  */
	extui	a10, a8, 24, 8
	extui	a11, a9, 24, 8
	sub	a2, a10, a11
	retw

.Ldiff0:
	/* Byte 0 is different.  */
	extui	a10, a8, 0, 8
	extui	a11, a9, 0, 8
	sub	a2, a10, a11
	retw

.Ldiff1:
	/* Byte 0 is equal; byte 1 is different.  */
	extui	a10, a8, 8, 8
	extui	a11, a9, 8, 8
	sub	a2, a10, a11
	retw

.Ldiff2:
	/* Bytes 0-1 are equal; byte 2 is different.  */
	extui	a10, a8, 16, 8
	extui	a11, a9, 16, 8
	sub	a2, a10, a11
	retw

#endif /* !__XTENSA_EB */

libc_hidden_def (strcmp)

#ifndef __UCLIBC_HAS_LOCALE__
strong_alias (strcmp, strcoll)
libc_hidden_def (strcoll)
#endif