Chris Zankel writes:

The following patches add support for the Xtensa processor architecture
to uClibc. They are based on a recent SVN checkout (12/05/2007).

The first patch (attached to this post) adds Xtensa support to various
shared configuration and make files. The following patches then include
the Xtensa specific files and directories.

I welcome any feedback and would appreciate it if you could include the
patches into the mainline tree. I am certainly committed to maintain the port. 

Bob Wilson was kind enough to review the patches.

Some notes about the architecture: Xtensa is a configurable and
extensible processor architecture developed by Tensilica. For more
information, please visit: www.linux-xtensa.org.

49 files changed, 4274 insertions, 0 deletions

diff --git a/libc/string/xtensa/Makefile b/libc/string/xtensa/Makefile
new file mode 100644
index 000000000..0a95346fd
--- /dev/null
+++ b/libc/string/xtensa/Makefile
@@ -0,0 +1,13 @@
+# Makefile for uClibc
+#
+# Copyright (C) 2000-2005 Erik Andersen <andersen@uclibc.org>
+#
+# Licensed under the LGPL v2.1, see the file COPYING.LIB in this tarball.
+#
+
+top_srcdir:=../../../
+top_builddir:=../../../
+all: objs
+include $(top_builddir)Rules.mak
+include ../Makefile.in
+include $(top_srcdir)Makerules
diff --git a/libc/string/xtensa/memcpy.S b/libc/string/xtensa/memcpy.S
new file mode 100644
index 000000000..19f3a6818
--- /dev/null
+++ b/libc/string/xtensa/memcpy.S
@@ -0,0 +1,297 @@
+/* Optimized memcpy 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, write to the Free
+   Software Foundation, Inc., 51 Franklin Street - Fifth Floor,
+   Boston, MA 02110-1301, USA.  */
+
+#include "../../sysdeps/linux/xtensa/sysdep.h"
+#include <bits/xtensa-config.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
+
+/* If the Xtensa Unaligned Load Exception option is not used, this
+   code can run a few cycles faster by relying on the low address bits
+   being ignored.  However, if the code is then run with an Xtensa ISS
+   client that checks for unaligned accesses, it will produce a lot of
+   warning messages.  Set this flag to disable the use of unaligned
+   accesses and keep the ISS happy.  */
+
+#define UNALIGNED_ADDRESSES_CHECKED 1
+
+/* Do not use .literal_position in the ENTRY macro.  */
+#undef LITERAL_POSITION
+#define LITERAL_POSITION
+
+
+/* void *memcpy (void *dst, const void *src, size_t len)
+
+   The algorithm is as follows:
+
+   If the destination is unaligned, align it by conditionally
+   copying 1- and/or 2-byte pieces.
+
+   If the source is aligned, copy 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.  */
+
+
+/* Byte by byte copy.  */
+
+	.text
+	.align	4
+	.literal_position
+__memcpy_aux:
+
+	/* Skip a byte to get 1 mod 4 alignment for LOOPNEZ
+	   (0 mod 4 alignment for LBEG).  */
+	.byte	0
+
+.Lbytecopy:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, 2f
+#else
+	beqz	a4, 2f
+	add	a7, a3, a4	// a7 = end address for source
+#endif
+1:	l8ui	a6, a3, 0
+	addi	a3, a3, 1
+	s8i	a6, a5, 0
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a7, 1b
+#endif
+2:	retw
+
+
+/* Destination is unaligned.  */
+
+	.align	4
+.Ldst1mod2: // dst is only byte aligned
+
+	/* Do short copies byte-by-byte.  */
+	_bltui	a4, 7, .Lbytecopy
+
+	/* Copy 1 byte.  */
+	l8ui	a6, a3, 0
+	addi	a3, a3, 1
+	addi	a4, a4, -1
+	s8i	a6, a5, 0
+	addi	a5, a5, 1
+
+	/* Return to main algorithm if dst is now aligned.  */
+	_bbci.l	a5, 1, .Ldstaligned
+
+.Ldst2mod4: // dst has 16-bit alignment
+
+	/* Do short copies byte-by-byte.  */
+	_bltui	a4, 6, .Lbytecopy
+
+	/* Copy 2 bytes.  */
+	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
+
+	/* dst is now aligned; return to main algorithm.  */
+	j	.Ldstaligned
+
+
+ENTRY (memcpy)
+	/* a2 = dst, a3 = src, a4 = len */
+
+	mov	a5, a2		// copy dst so that a2 is return value
+	_bbsi.l	a2, 0, .Ldst1mod2
+	_bbsi.l	a2, 1, .Ldst2mod4
+.Ldstaligned:
+
+	/* Get number of loop iterations with 16B per iteration.  */
+	srli	a7, a4, 4
+
+	/* Check if source is aligned.  */
+	movi	a8, 3
+	_bany	a3, a8, .Lsrcunaligned
+
+	/* Destination and source are word-aligned, use word copy.  */
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, 2f
+#else
+	beqz	a7, 2f
+	slli	a8, a7, 4
+	add	a8, a8, a3	// a8 = end of last 16B source chunk
+#endif
+1:	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, 1b
+#endif
+
+	/* Copy any leftover pieces smaller than 16B.  */
+2:	bbci.l	a4, 3, 3f
+
+	/* 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
+
+3:	bbsi.l	a4, 2, 4f
+	bbsi.l	a4, 1, 5f
+	bbsi.l	a4, 0, 6f
+	retw
+
+	/* Copy 4 bytes.  */
+4:	l32i	a6, a3, 0
+	addi	a3, a3, 4
+	s32i	a6, a5, 0
+	addi	a5, a5, 4
+	bbsi.l	a4, 1, 5f
+	bbsi.l	a4, 0, 6f
+	retw
+
+	/* Copy 2 bytes.  */
+5:	l16ui	a6, a3, 0
+	addi	a3, a3, 2
+	s16i	a6, a5, 0
+	addi	a5, a5, 2
+	bbsi.l	a4, 0, 6f
+	retw
+
+	/* Copy 1 byte.  */
+6:	l8ui	a6, a3, 0
+	s8i	a6, a5, 0
+
+.Ldone:
+	retw
+
+
+/* Destination is aligned; source is unaligned.  */
+
+	.align	4
+.Lsrcunaligned:
+	/* Avoid loading anything for zero-length copies.  */
+	_beqz	a4, .Ldone
+
+	/* Copy 16 bytes per iteration for word-aligned dst and
+	   unaligned src.  */
+	ssa8	a3		// set shift amount from byte offset
+#if UNALIGNED_ADDRESSES_CHECKED
+	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, 2f
+#else
+	beqz	a7, 2f
+	slli	a10, a7, 4
+	add	a10, a10, a3	// a10 = end of last 16B source chunk
+#endif
+1:	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, 1b
+#endif
+
+2:	bbci.l	a4, 3, 3f
+
+	/* 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
+
+3:	bbci.l	a4, 2, 4f
+
+	/* 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
+4:
+#if UNALIGNED_ADDRESSES_CHECKED
+	add	a3, a3, a11	// readjust a3 with correct misalignment
+#endif
+	bbsi.l	a4, 1, 5f
+	bbsi.l	a4, 0, 6f
+	retw
+
+	/* Copy 2 bytes.  */
+5:	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, 6f
+	retw
+
+	/* Copy 1 byte.  */
+6:	l8ui	a6, a3, 0
+	s8i	a6, a5, 0
+	retw
+
+libc_hidden_def (memcpy)
diff --git a/libc/string/xtensa/memset.S b/libc/string/xtensa/memset.S
new file mode 100644
index 000000000..c0928825d
--- /dev/null
+++ b/libc/string/xtensa/memset.S
@@ -0,0 +1,165 @@
+/* Optimized memset 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, write to the Free
+   Software Foundation, Inc., 51 Franklin Street - Fifth Floor,
+   Boston, MA 02110-1301, USA.  */
+
+#include "../../sysdeps/linux/xtensa/sysdep.h"
+#include <bits/xtensa-config.h>
+
+/* Do not use .literal_position in the ENTRY macro.  */
+#undef LITERAL_POSITION
+#define LITERAL_POSITION
+
+/* 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, set 16B chunks with a loop, and then
+   finish up with 8B, 4B, 2B, and 1B stores conditional on the length.
+
+   If the destination is unaligned, align it by conditionally
+   setting 1B and/or 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).  */
+
+
+/* Byte-by-byte set.  */
+
+	.text
+	.align	4
+	.literal_position
+__memset_aux:
+
+	/* Skip a byte to get 1 mod 4 alignment for LOOPNEZ
+	   (0 mod 4 alignment for LBEG).  */
+	.byte	0
+
+.Lbyteset:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, 2f
+#else
+	beqz	a4, 2f
+	add	a6, a5, a4	// a6 = ending address
+#endif
+1:	s8i	a3, a5, 0
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a5, a6, 1b
+#endif
+2:	retw
+
+
+/* Destination is unaligned.  */
+
+	.align	4
+
+.Ldst1mod2: // dst is only byte aligned
+
+	/* Do short sizes byte-by-byte.  */
+	bltui	a4, 8, .Lbyteset
+
+	/* Set 1 byte.  */
+	s8i	a3, a5, 0
+	addi	a5, a5, 1
+	addi	a4, a4, -1
+
+	/* Now retest if dst is aligned.  */
+	_bbci.l	a5, 1, .Ldstaligned
+
+.Ldst2mod4: // dst has 16-bit alignment
+
+	/* Do short sizes byte-by-byte.  */
+	bltui	a4, 8, .Lbyteset
+
+	/* Set 2 bytes.  */
+	s16i	a3, a5, 0
+	addi	a5, a5, 2
+	addi	a4, a4, -2
+
+	/* dst is now aligned; return to main algorithm */
+	j	.Ldstaligned
+
+
+ENTRY (memset)
+	/* a2 = dst, a3 = c, a4 = length */
+
+	/* Duplicate character into all bytes of word.  */
+	extui	a3, a3, 0, 8
+	slli	a7, a3, 8
+	or	a3, a3, a7
+	slli	a7, a3, 16
+	or	a3, a3, a7
+
+	mov	a5, a2		// copy dst so that a2 is return value
+
+	/* Check if dst is unaligned.  */
+	_bbsi.l	a2, 0, .Ldst1mod2
+	_bbsi.l	a2, 1, .Ldst2mod4
+.Ldstaligned:
+
+	/* Get number of loop iterations with 16B per iteration.  */
+	srli	a7, a4, 4
+
+	/* Destination is word-aligned.  */
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, 2f
+#else
+	beqz	a7, 2f
+	slli	a6, a7, 4
+	add	a6, a6, a5	// a6 = end of last 16B chunk
+#endif
+	/* Set 16 bytes per iteration.  */
+1:	s32i	a3, a5, 0
+	s32i	a3, a5, 4
+	s32i	a3, a5, 8
+	s32i	a3, a5, 12
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a5, a6, 1b
+#endif
+
+	/* Set any leftover pieces smaller than 16B.  */
+2:	bbci.l	a4, 3, 3f
+
+	/* Set 8 bytes.  */
+	s32i	a3, a5, 0
+	s32i	a3, a5, 4
+	addi	a5, a5, 8
+
+3:	bbci.l	a4, 2, 4f
+
+	/* Set 4 bytes.  */
+	s32i	a3, a5, 0
+	addi	a5, a5, 4
+
+4:	bbci.l	a4, 1, 5f
+
+	/* Set 2 bytes.  */
+	s16i	a3, a5, 0
+	addi	a5, a5, 2
+
+5:	bbci.l	a4, 0, 6f
+
+	/* Set 1 byte.  */
+	s8i	a3, a5, 0
+6:	retw
+
+libc_hidden_def (memset)
diff --git a/libc/string/xtensa/strcmp.S b/libc/string/xtensa/strcmp.S
new file mode 100644
index 000000000..90c418d12
--- /dev/null
+++ b/libc/string/xtensa/strcmp.S
@@ -0,0 +1,313 @@
+/* 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, write to the Free
+   Software Foundation, Inc., 51 Franklin Street - Fifth Floor,
+   Boston, MA 02110-1301, USA.  */
+
+#include "../../sysdeps/linux/xtensa/sysdep.h"
+#include <bits/xtensa-config.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
+
+	.literal .Lmask0, MASK0
+	.literal .Lmask1, MASK1
+	.literal .Lmask2, MASK2
+	.literal .Lmask3, MASK3
+	.literal .Lmask4, MASK4
+
+	.text
+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