1From 0bea89508c6be2b1a76650108d759d389e99bbf3 Mon Sep 17 00:00:00 2001
2From: Jeffy Chen <jeffy.chen@rock-chips.com>
3Date: Wed, 13 Apr 2022 16:32:02 +0800
4Subject: [PATCH] libc: string: arm: Support using glibc-neon version of APIs
5
6Set USE_GLIBC_NEON to enable.
7
8Signed-off-by: Jeffy Chen <jeffy.chen@rock-chips.com>
9---
10 Makerules                                     |   1 +
11 libc/string/Makefile.in                       |  10 +-
12 libc/string/arm/Makefile.in                   |  13 +
13 libc/string/arm/glibc-neon/README             |   4 +
14 libc/string/arm/glibc-neon/arm-features.h     |  59 ++
15 libc/string/arm/glibc-neon/bcopy.c            |  24 +
16 libc/string/arm/glibc-neon/bzero.c            |  28 +
17 libc/string/arm/glibc-neon/glibc_wrap.h       |  29 +
18 libc/string/arm/glibc-neon/memcmp.S           |   1 +
19 libc/string/arm/glibc-neon/memcpy.S           | 728 ++++++++++++++++++
20 libc/string/arm/glibc-neon/memmove.S          | 332 ++++++++
21 libc/string/arm/glibc-neon/memset.S           |  68 ++
22 libc/string/arm/glibc-neon/strcmp.S           | 504 ++++++++++++
23 libc/string/arm/glibc-neon/strlen.S           |  76 ++
24 libc/string/arm/glibc-neon/sysdep.h           | 339 ++++++++
25 .../arm/glibc-neon/sysdeps/generic/dwarf2.h   | 590 ++++++++++++++
26 .../arm/glibc-neon/sysdeps/generic/sysdep.h   |  97 +++
27 17 files changed, 2901 insertions(+), 2 deletions(-)
28 create mode 100644 libc/string/arm/Makefile.in
29 create mode 100644 libc/string/arm/glibc-neon/README
30 create mode 100644 libc/string/arm/glibc-neon/arm-features.h
31 create mode 100644 libc/string/arm/glibc-neon/bcopy.c
32 create mode 100644 libc/string/arm/glibc-neon/bzero.c
33 create mode 100644 libc/string/arm/glibc-neon/glibc_wrap.h
34 create mode 120000 libc/string/arm/glibc-neon/memcmp.S
35 create mode 100644 libc/string/arm/glibc-neon/memcpy.S
36 create mode 100644 libc/string/arm/glibc-neon/memmove.S
37 create mode 100644 libc/string/arm/glibc-neon/memset.S
38 create mode 100644 libc/string/arm/glibc-neon/strcmp.S
39 create mode 100644 libc/string/arm/glibc-neon/strlen.S
40 create mode 100644 libc/string/arm/glibc-neon/sysdep.h
41 create mode 100644 libc/string/arm/glibc-neon/sysdeps/generic/dwarf2.h
42 create mode 100644 libc/string/arm/glibc-neon/sysdeps/generic/sysdep.h
43
44diff --git a/Makerules b/Makerules
45index fd40e6c..2013ba9 100644
46--- a/Makerules
47+++ b/Makerules
48@@ -255,6 +255,7 @@ CFLAGS_gen.dep = -MT $@ -MD -MP -MF $(dir $@).$(notdir $@).dep
49
50 cmd_compile.c = $(CC) -c $< -o $@ \
51     $(filter-out $(CFLAGS-OMIT-$(notdir $<)), \
52+        $(CFLAGS-extra-$(subst $(top_srcdir),,$(<D))) \
53         $(CFLAGS) \
54         $(CFLAGS-for-library-members) \
55 	$(CFLAGS-$(suffix $@)) \
56diff --git a/libc/string/Makefile.in b/libc/string/Makefile.in
57index e7f2ccd..975b395 100644
58--- a/libc/string/Makefile.in
59+++ b/libc/string/Makefile.in
60@@ -5,6 +5,12 @@
61 # Licensed under the LGPL v2.1, see the file COPYING.LIB in this tarball.
62 #
63
64+ifeq ($(TARGET_ARCH),arm)
65+ifneq ($(USE_GLIBC_NEON),)
66+include $(top_srcdir)libc/string/$(TARGET_ARCH)/Makefile.in
67+endif
68+endif
69+
70 subdirs += libc/string/$(TARGET_ARCH) libc/string/generic
71
72 #
73@@ -25,8 +31,8 @@ STRING_SUBARCH_OBJS := $(STRING_SUBARCH_SOBJ) $(STRING_SUBARCH_COBJ)
74 endif
75
76 # Collect the arch specific implementation (asm, c files)
77-STRING_ARCH_DIR := $(top_srcdir)libc/string/$(TARGET_ARCH)
78-STRING_ARCH_OUT := $(top_builddir)libc/string/$(TARGET_ARCH)
79+STRING_ARCH_DIR ?= $(top_srcdir)libc/string/$(TARGET_ARCH)
80+STRING_ARCH_OUT ?= $(top_builddir)libc/string/$(TARGET_ARCH)
81
82 STRING_ARCH_SRC := $(wildcard $(STRING_ARCH_DIR)/*.c)
83 STRING_ARCH_OBJ := $(patsubst $(STRING_ARCH_DIR)/%.c,$(STRING_ARCH_OUT)/%.o,$(STRING_ARCH_SRC))
84diff --git a/libc/string/arm/Makefile.in b/libc/string/arm/Makefile.in
85new file mode 100644
86index 0000000..4cd7e24
87--- /dev/null
88+++ b/libc/string/arm/Makefile.in
89@@ -0,0 +1,13 @@
90+# Makefile for uClibc
91+#
92+# Copyright (C) 2022 Jeffy Chen <jeffy.chen@rock-chips.com>
93+#
94+# Licensed under the LGPL v2.1, see the file COPYING.LIB in this tarball.
95+#
96+
97+SUBDIR := libc/string/arm/glibc-neon
98+
99+STRING_ARCH_DIR := $(top_srcdir)$(SUBDIR)
100+STRING_ARCH_OUT := $(top_builddir)$(SUBDIR)
101+
102+CFLAGS-extra-$(SUBDIR) := -include glibc_wrap.h -I$(top_srcdir)$(SUBDIR)
103diff --git a/libc/string/arm/glibc-neon/README b/libc/string/arm/glibc-neon/README
104new file mode 100644
105index 0000000..ad72f05
106--- /dev/null
107+++ b/libc/string/arm/glibc-neon/README
108@@ -0,0 +1,4 @@
109+NOTE:
110+1/ Ported from glibc-2.34.9.
111+2/ glibc doesn't have an arm arch version of memcmp.
112+3/ uClibc is using strcmp as strcoll when locale disabled.
113diff --git a/libc/string/arm/glibc-neon/arm-features.h b/libc/string/arm/glibc-neon/arm-features.h
114new file mode 100644
115index 0000000..4a86e00
116--- /dev/null
117+++ b/libc/string/arm/glibc-neon/arm-features.h
118@@ -0,0 +1,59 @@
119+/* Macros to test for CPU features on ARM.  Generic ARM version.
120+   Copyright (C) 2012-2021 Free Software Foundation, Inc.
121+   This file is part of the GNU C Library.
122+
123+   The GNU C Library is free software; you can redistribute it and/or
124+   modify it under the terms of the GNU Lesser General Public
125+   License as published by the Free Software Foundation; either
126+   version 2.1 of the License, or (at your option) any later version.
127+
128+   The GNU C Library is distributed in the hope that it will be useful,
129+   but WITHOUT ANY WARRANTY; without even the implied warranty of
130+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
131+   Lesser General Public License for more details.
132+
133+   You should have received a copy of the GNU Lesser General Public
134+   License along with the GNU C Library.  If not, see
135+   <https://www.gnu.org/licenses/>.  */
136+
137+#ifndef _ARM_ARM_FEATURES_H
138+#define _ARM_ARM_FEATURES_H 1
139+
140+/* An OS-specific arm-features.h file should define ARM_HAVE_VFP to
141+   an appropriate expression for testing at runtime whether the VFP
142+   hardware is present.  We'll then redefine it to a constant if we
143+   know at compile time that we can assume VFP.  */
144+
145+#ifndef __SOFTFP__
146+/* The compiler is generating VFP instructions, so we're already
147+   assuming the hardware exists.  */
148+# undef ARM_HAVE_VFP
149+# define ARM_HAVE_VFP	1
150+#endif
151+
152+/* An OS-specific arm-features.h file may define ARM_ASSUME_NO_IWMMXT
153+   to indicate at compile time that iWMMXt hardware is never present
154+   at runtime (or that we never care about its state) and so need not
155+   be checked for.  */
156+
157+/* A more-specific arm-features.h file may define ARM_ALWAYS_BX to indicate
158+   that instructions using pc as a destination register must never be used,
159+   so a "bx" (or "blx") instruction is always required.  */
160+
161+/* The log2 of the minimum alignment required for an address that
162+   is the target of a computed branch (i.e. a "bx" instruction).
163+   A more-specific arm-features.h file may define this to set a more
164+   stringent requirement.
165+
166+   Using this only makes sense for code in ARM mode (where instructions
167+   always have a fixed size of four bytes), or for Thumb-mode code that is
168+   specifically aligning all the related branch targets to match (since
169+   Thumb instructions might be either two or four bytes).  */
170+#ifndef ARM_BX_ALIGN_LOG2
171+# define ARM_BX_ALIGN_LOG2	2
172+#endif
173+
174+/* An OS-specific arm-features.h file may define ARM_NO_INDEX_REGISTER to
175+   indicate that the two-register addressing modes must never be used.  */
176+
177+#endif  /* arm-features.h */
178diff --git a/libc/string/arm/glibc-neon/bcopy.c b/libc/string/arm/glibc-neon/bcopy.c
179new file mode 100644
180index 0000000..302bbbd
181--- /dev/null
182+++ b/libc/string/arm/glibc-neon/bcopy.c
183@@ -0,0 +1,24 @@
184+/* Copyright (C) 1991-2021 Free Software Foundation, Inc.
185+   This file is part of the GNU C Library.
186+
187+   The GNU C Library is free software; you can redistribute it and/or
188+   modify it under the terms of the GNU Lesser General Public
189+   License as published by the Free Software Foundation; either
190+   version 2.1 of the License, or (at your option) any later version.
191+
192+   The GNU C Library is distributed in the hope that it will be useful,
193+   but WITHOUT ANY WARRANTY; without even the implied warranty of
194+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
195+   Lesser General Public License for more details.
196+
197+   You should have received a copy of the GNU Lesser General Public
198+   License along with the GNU C Library; if not, see
199+   <https://www.gnu.org/licenses/>.  */
200+
201+#include <string.h>
202+
203+void
204+bcopy (const void *src, void *dest, size_t len)
205+{
206+  memmove (dest, src, len);
207+}
208diff --git a/libc/string/arm/glibc-neon/bzero.c b/libc/string/arm/glibc-neon/bzero.c
209new file mode 100644
210index 0000000..4391dad
211--- /dev/null
212+++ b/libc/string/arm/glibc-neon/bzero.c
213@@ -0,0 +1,28 @@
214+/* Copyright (C) 1991-2021 Free Software Foundation, Inc.
215+   This file is part of the GNU C Library.
216+
217+   The GNU C Library is free software; you can redistribute it and/or
218+   modify it under the terms of the GNU Lesser General Public
219+   License as published by the Free Software Foundation; either
220+   version 2.1 of the License, or (at your option) any later version.
221+
222+   The GNU C Library is distributed in the hope that it will be useful,
223+   but WITHOUT ANY WARRANTY; without even the implied warranty of
224+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
225+   Lesser General Public License for more details.
226+
227+   You should have received a copy of the GNU Lesser General Public
228+   License along with the GNU C Library; if not, see
229+   <https://www.gnu.org/licenses/>.  */
230+
231+#include <string.h>
232+
233+#undef __bzero
234+
235+/* Set N bytes of S to 0.  */
236+void
237+__bzero (void *s, size_t len)
238+{
239+  memset (s, '\0', len);
240+}
241+weak_alias (__bzero, bzero)
242diff --git a/libc/string/arm/glibc-neon/glibc_wrap.h b/libc/string/arm/glibc-neon/glibc_wrap.h
243new file mode 100644
244index 0000000..f00d50d
245--- /dev/null
246+++ b/libc/string/arm/glibc-neon/glibc_wrap.h
247@@ -0,0 +1,29 @@
248+/* Macros to adapt glibc version of string APIs.
249+   Copyright (C) 2022 Jeffy Chen <jeffy.chen@rock-chips.com>
250+   This file is part of the GNU C Library.
251+
252+   The GNU C Library is free software; you can redistribute it and/or
253+   modify it under the terms of the GNU Lesser General Public
254+   License as published by the Free Software Foundation; either
255+   version 2.1 of the License, or (at your option) any later version.
256+
257+   The GNU C Library is distributed in the hope that it will be useful,
258+   but WITHOUT ANY WARRANTY; without even the implied warranty of
259+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
260+   Lesser General Public License for more details.
261+
262+   You should have received a copy of the GNU Lesser General Public
263+   License along with the GNU C Library.  If not, see
264+   <https://www.gnu.org/licenses/>.  */
265+
266+#ifndef GLIBC_WRAP_H
267+#define GLIBC_WRAP_H
268+
269+#define IS_IN(lib) 1
270+
271+#define libc_hidden_builtin_def(fn) libc_hidden_def(fn)
272+
273+#define C_SYMBOL_NAME(name) name
274+#define MEMCPY_NEON
275+
276+#endif // GLIBC_WRAP_H
277diff --git a/libc/string/arm/glibc-neon/memcmp.S b/libc/string/arm/glibc-neon/memcmp.S
278new file mode 120000
279index 0000000..7c28a09
280--- /dev/null
281+++ b/libc/string/arm/glibc-neon/memcmp.S
282@@ -0,0 +1 @@
283+../memcmp.S
284\ No newline at end of file
285diff --git a/libc/string/arm/glibc-neon/memcpy.S b/libc/string/arm/glibc-neon/memcpy.S
286new file mode 100644
287index 0000000..ee562e8
288--- /dev/null
289+++ b/libc/string/arm/glibc-neon/memcpy.S
290@@ -0,0 +1,728 @@
291+/* NEON/VFP/ARM version of memcpy optimized for Cortex-A15.
292+   Copyright (C) 2013-2021 Free Software Foundation, Inc.
293+   This file is part of the GNU C Library.
294+
295+   The GNU C Library is free software; you can redistribute it and/or
296+   modify it under the terms of the GNU Lesser General Public
297+   License as published by the Free Software Foundation; either
298+   version 2.1 of the License, or (at your option) any later version.
299+
300+   The GNU C Library is distributed in the hope that it will be useful,
301+   but WITHOUT ANY WARRANTY; without even the implied warranty of
302+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
303+   Lesser General Public License for more details.
304+
305+   You should have received a copy of the GNU Lesser General Public
306+   License along with the GNU C Library; if not, see
307+   <https://www.gnu.org/licenses/>.
308+
309+   This memcpy routine is optimised for Cortex-A15 cores and takes advantage
310+   of VFP or NEON when built with the appropriate flags.
311+
312+   Assumptions:
313+
314+    ARMv6 (ARMv7-a if using Neon)
315+    ARM state
316+    Unaligned accesses
317+
318+ */
319+
320+/* Thumb cannot encode negative immediate offsets in memory operations.  */
321+#ifndef NO_THUMB
322+#define NO_THUMB
323+#endif
324+#include <sysdep.h>
325+#include <arm-features.h>
326+
327+	.syntax unified
328+	/* This implementation requires ARM state.  */
329+	.arm
330+
331+#ifdef MEMCPY_NEON
332+
333+	.fpu	neon
334+	.arch	armv7-a
335+# define FRAME_SIZE	4
336+# define USE_VFP
337+# define USE_NEON
338+
339+#elif defined (MEMCPY_VFP)
340+
341+	.arch	armv6
342+	.fpu	vfpv2
343+# define FRAME_SIZE	32
344+# define USE_VFP
345+
346+#else
347+	.arch	armv6
348+# define FRAME_SIZE    32
349+
350+#endif
351+
352+#define ALIGN(addr, align) addr:align
353+
354+#define INSN_SIZE	4
355+
356+/* Call parameters.  */
357+#define dstin	r0
358+#define src	r1
359+#define count	r2
360+
361+/* Locals.  */
362+#define tmp1	r3
363+#define dst	ip
364+#define tmp2	r8
365+
366+/* These two macros both work by repeated invocation of the macro
367+   dispatch_step (not defined here).  That macro performs one "step",
368+   doing one load instruction and one store instruction to copy one
369+   "unit".  On entry, TMP1 contains the number of bytes to be copied,
370+   a multiple of the unit size.  The macro clobbers TMP1 in the
371+   process of doing a computed jump to the tail containing the
372+   appropriate number of steps.
373+
374+   In dispatch_7_dword, dispatch_step is invoked seven times, with an
375+   argument that is 7 for the first and 1 for the last.  Units are
376+   double-words (8 bytes).  TMP1 is at most 56.
377+
378+   In dispatch_15_word, dispatch_step is invoked fifteen times,
379+   with an argument that is 15 for the first and 1 for the last.
380+   Units are words (4 bytes).  TMP1 is at most 60.  */
381+
382+#ifndef ARM_ALWAYS_BX
383+# if ARM_BX_ALIGN_LOG2 != 2
384+#  error case not handled
385+# endif
386+	.macro dispatch_7_dword
387+	rsb	tmp1, tmp1, #((7 * 8) - PC_OFS + INSN_SIZE)
388+	add	pc, pc, tmp1
389+	dispatch_step 7
390+	dispatch_step 6
391+	dispatch_step 5
392+	dispatch_step 4
393+	dispatch_step 3
394+	dispatch_step 2
395+	dispatch_step 1
396+	.purgem dispatch_step
397+	.endm
398+
399+	.macro dispatch_15_word
400+	rsb	tmp1, tmp1, #((15 * 4) - PC_OFS/2 + INSN_SIZE/2)
401+	add	pc, pc, tmp1, lsl #1
402+	dispatch_step 15
403+	dispatch_step 14
404+	dispatch_step 13
405+	dispatch_step 12
406+	dispatch_step 11
407+	dispatch_step 10
408+	dispatch_step 9
409+	dispatch_step 8
410+	dispatch_step 7
411+	dispatch_step 6
412+	dispatch_step 5
413+	dispatch_step 4
414+	dispatch_step 3
415+	dispatch_step 2
416+	dispatch_step 1
417+	.purgem dispatch_step
418+	.endm
419+#else
420+# if ARM_BX_ALIGN_LOG2 < 3
421+#  error case not handled
422+# endif
423+	.macro dispatch_helper steps, log2_bytes_per_step
424+	/* TMP1 gets (max_bytes - bytes_to_copy), where max_bytes is
425+	   (STEPS << LOG2_BYTES_PER_STEP).
426+	   So this is (steps_to_skip << LOG2_BYTES_PER_STEP).
427+	   Then it needs further adjustment to compensate for the
428+	   distance between the PC value taken below (0f + PC_OFS)
429+	   and the first step's instructions (1f).  */
430+	rsb	tmp1, tmp1, #((\steps << \log2_bytes_per_step) \
431+			      + ((1f - PC_OFS - 0f) \
432+				 >> (ARM_BX_ALIGN_LOG2 - \log2_bytes_per_step)))
433+	/* Shifting down LOG2_BYTES_PER_STEP gives us the number of
434+	   steps to skip, then shifting up ARM_BX_ALIGN_LOG2 gives us
435+	   the (byte) distance to add to the PC.  */
436+0:	add	tmp1, pc, tmp1, lsl #(ARM_BX_ALIGN_LOG2 - \log2_bytes_per_step)
437+	bx	tmp1
438+	.p2align ARM_BX_ALIGN_LOG2
439+1:
440+	.endm
441+
442+	.macro dispatch_7_dword
443+	dispatch_helper 7, 3
444+	.p2align ARM_BX_ALIGN_LOG2
445+	dispatch_step 7
446+	.p2align ARM_BX_ALIGN_LOG2
447+	dispatch_step 6
448+	.p2align ARM_BX_ALIGN_LOG2
449+	dispatch_step 5
450+	.p2align ARM_BX_ALIGN_LOG2
451+	dispatch_step 4
452+	.p2align ARM_BX_ALIGN_LOG2
453+	dispatch_step 3
454+	.p2align ARM_BX_ALIGN_LOG2
455+	dispatch_step 2
456+	.p2align ARM_BX_ALIGN_LOG2
457+	dispatch_step 1
458+	.p2align ARM_BX_ALIGN_LOG2
459+	.purgem dispatch_step
460+	.endm
461+
462+	.macro dispatch_15_word
463+	dispatch_helper 15, 2
464+	dispatch_step 15
465+	.p2align ARM_BX_ALIGN_LOG2
466+	dispatch_step 14
467+	.p2align ARM_BX_ALIGN_LOG2
468+	dispatch_step 13
469+	.p2align ARM_BX_ALIGN_LOG2
470+	dispatch_step 12
471+	.p2align ARM_BX_ALIGN_LOG2
472+	dispatch_step 11
473+	.p2align ARM_BX_ALIGN_LOG2
474+	dispatch_step 10
475+	.p2align ARM_BX_ALIGN_LOG2
476+	dispatch_step 9
477+	.p2align ARM_BX_ALIGN_LOG2
478+	dispatch_step 8
479+	.p2align ARM_BX_ALIGN_LOG2
480+	dispatch_step 7
481+	.p2align ARM_BX_ALIGN_LOG2
482+	dispatch_step 6
483+	.p2align ARM_BX_ALIGN_LOG2
484+	dispatch_step 5
485+	.p2align ARM_BX_ALIGN_LOG2
486+	dispatch_step 4
487+	.p2align ARM_BX_ALIGN_LOG2
488+	dispatch_step 3
489+	.p2align ARM_BX_ALIGN_LOG2
490+	dispatch_step 2
491+	.p2align ARM_BX_ALIGN_LOG2
492+	dispatch_step 1
493+	.p2align ARM_BX_ALIGN_LOG2
494+	.purgem dispatch_step
495+	.endm
496+
497+#endif
498+
499+#ifndef USE_NEON
500+/* For bulk copies using GP registers.  */
501+#define	A_l	r2		/* Call-clobbered.  */
502+#define	A_h	r3		/* Call-clobbered.  */
503+#define	B_l	r4
504+#define	B_h	r5
505+#define	C_l	r6
506+#define	C_h	r7
507+/* Don't use the pair r8,r9 because in some EABI variants r9 is reserved.  */
508+#define	D_l	r10
509+#define	D_h	r11
510+#endif
511+
512+/* Number of lines ahead to pre-fetch data.  If you change this the code
513+   below will need adjustment to compensate.  */
514+
515+#define prefetch_lines	5
516+
517+#ifdef USE_VFP
518+	.macro	cpy_line_vfp vreg, base
519+	vstr	\vreg, [dst, #\base]
520+	vldr	\vreg, [src, #\base]
521+	vstr	d0, [dst, #\base + 8]
522+	vldr	d0, [src, #\base + 8]
523+	vstr	d1, [dst, #\base + 16]
524+	vldr	d1, [src, #\base + 16]
525+	vstr	d2, [dst, #\base + 24]
526+	vldr	d2, [src, #\base + 24]
527+	vstr	\vreg, [dst, #\base + 32]
528+	vldr	\vreg, [src, #\base + prefetch_lines * 64 - 32]
529+	vstr	d0, [dst, #\base + 40]
530+	vldr	d0, [src, #\base + 40]
531+	vstr	d1, [dst, #\base + 48]
532+	vldr	d1, [src, #\base + 48]
533+	vstr	d2, [dst, #\base + 56]
534+	vldr	d2, [src, #\base + 56]
535+	.endm
536+
537+	.macro	cpy_tail_vfp vreg, base
538+	vstr	\vreg, [dst, #\base]
539+	vldr	\vreg, [src, #\base]
540+	vstr	d0, [dst, #\base + 8]
541+	vldr	d0, [src, #\base + 8]
542+	vstr	d1, [dst, #\base + 16]
543+	vldr	d1, [src, #\base + 16]
544+	vstr	d2, [dst, #\base + 24]
545+	vldr	d2, [src, #\base + 24]
546+	vstr	\vreg, [dst, #\base + 32]
547+	vstr	d0, [dst, #\base + 40]
548+	vldr	d0, [src, #\base + 40]
549+	vstr	d1, [dst, #\base + 48]
550+	vldr	d1, [src, #\base + 48]
551+	vstr	d2, [dst, #\base + 56]
552+	vldr	d2, [src, #\base + 56]
553+	.endm
554+#endif
555+
556+	.p2align 6
557+ENTRY(memcpy)
558+
559+	mov	dst, dstin	/* Preserve dstin, we need to return it.  */
560+	cmp	count, #64
561+	bhs	.Lcpy_not_short
562+	/* Deal with small copies quickly by dropping straight into the
563+	   exit block.  */
564+
565+.Ltail63unaligned:
566+#ifdef USE_NEON
567+	/* These need an extra layer of macro just to work around a
568+	   bug in the assembler's parser when an operand starts with
569+	   a {...}.  https://sourceware.org/bugzilla/show_bug.cgi?id=15647
570+	   tracks that bug; it was not fixed as of binutils-2.23.2.  */
571+	.macro neon_load_d0 reg
572+	vld1.8	{d0}, [\reg]!
573+	.endm
574+	.macro neon_store_d0 reg
575+	vst1.8	{d0}, [\reg]!
576+	.endm
577+
578+	and	tmp1, count, #0x38
579+	.macro dispatch_step i
580+	neon_load_d0 src
581+	neon_store_d0 dst
582+	.endm
583+	dispatch_7_dword
584+
585+	tst	count, #4
586+	ldrne	tmp1, [src], #4
587+	strne	tmp1, [dst], #4
588+#else
589+	/* Copy up to 15 full words of data.  May not be aligned.  */
590+	/* Cannot use VFP for unaligned data.  */
591+	and	tmp1, count, #0x3c
592+	add	dst, dst, tmp1
593+	add	src, src, tmp1
594+	/* Jump directly into the sequence below at the correct offset.  */
595+	.macro dispatch_step i
596+	ldr	tmp1, [src, #-(\i * 4)]
597+	str	tmp1, [dst, #-(\i * 4)]
598+	.endm
599+	dispatch_15_word
600+#endif
601+
602+	lsls	count, count, #31
603+	ldrhcs	tmp1, [src], #2
604+	ldrbne	src, [src]		/* Src is dead, use as a scratch.  */
605+	strhcs	tmp1, [dst], #2
606+	strbne	src, [dst]
607+	bx	lr
608+
609+.Lcpy_not_short:
610+	/* At least 64 bytes to copy, but don't know the alignment yet.  */
611+	str	tmp2, [sp, #-FRAME_SIZE]!
612+	cfi_adjust_cfa_offset (FRAME_SIZE)
613+	cfi_rel_offset (tmp2, 0)
614+	cfi_remember_state
615+	and	tmp2, src, #7
616+	and	tmp1, dst, #7
617+	cmp	tmp1, tmp2
618+	bne	.Lcpy_notaligned
619+
620+#ifdef USE_VFP
621+	/* Magic dust alert!  Force VFP on Cortex-A9.  Experiments show
622+	   that the FP pipeline is much better at streaming loads and
623+	   stores.  This is outside the critical loop.  */
624+	vmov.f32	s0, s0
625+#endif
626+
627+	/* SRC and DST have the same mutual 64-bit alignment, but we may
628+	   still need to pre-copy some bytes to get to natural alignment.
629+	   We bring SRC and DST into full 64-bit alignment.  */
630+	lsls	tmp2, dst, #29
631+	beq	1f
632+	rsbs	tmp2, tmp2, #0
633+	sub	count, count, tmp2, lsr #29
634+	ldrmi	tmp1, [src], #4
635+	strmi	tmp1, [dst], #4
636+	lsls	tmp2, tmp2, #2
637+	ldrhcs	tmp1, [src], #2
638+	ldrbne	tmp2, [src], #1
639+	strhcs	tmp1, [dst], #2
640+	strbne	tmp2, [dst], #1
641+
642+1:
643+	subs	tmp2, count, #64	/* Use tmp2 for count.  */
644+	blo	.Ltail63aligned
645+
646+	cmp	tmp2, #512
647+	bhs	.Lcpy_body_long
648+
649+.Lcpy_body_medium:			/* Count in tmp2.  */
650+#ifdef USE_VFP
651+1:
652+	vldr	d0, [src, #0]
653+	subs	tmp2, tmp2, #64
654+	vldr	d1, [src, #8]
655+	vstr	d0, [dst, #0]
656+	vldr	d0, [src, #16]
657+	vstr	d1, [dst, #8]
658+	vldr	d1, [src, #24]
659+	vstr	d0, [dst, #16]
660+	vldr	d0, [src, #32]
661+	vstr	d1, [dst, #24]
662+	vldr	d1, [src, #40]
663+	vstr	d0, [dst, #32]
664+	vldr	d0, [src, #48]
665+	vstr	d1, [dst, #40]
666+	vldr	d1, [src, #56]
667+	vstr	d0, [dst, #48]
668+	add	src, src, #64
669+	vstr	d1, [dst, #56]
670+	add	dst, dst, #64
671+	bhs	1b
672+	tst	tmp2, #0x3f
673+	beq	.Ldone
674+
675+.Ltail63aligned:			/* Count in tmp2.  */
676+	and	tmp1, tmp2, #0x38
677+	add	dst, dst, tmp1
678+	add	src, src, tmp1
679+	.macro dispatch_step i
680+	vldr	d0, [src, #-(\i * 8)]
681+	vstr	d0, [dst, #-(\i * 8)]
682+	.endm
683+	dispatch_7_dword
684+#else
685+	sub	src, src, #8
686+	sub	dst, dst, #8
687+1:
688+	ldrd	A_l, A_h, [src, #8]
689+	strd	A_l, A_h, [dst, #8]
690+	ldrd	A_l, A_h, [src, #16]
691+	strd	A_l, A_h, [dst, #16]
692+	ldrd	A_l, A_h, [src, #24]
693+	strd	A_l, A_h, [dst, #24]
694+	ldrd	A_l, A_h, [src, #32]
695+	strd	A_l, A_h, [dst, #32]
696+	ldrd	A_l, A_h, [src, #40]
697+	strd	A_l, A_h, [dst, #40]
698+	ldrd	A_l, A_h, [src, #48]
699+	strd	A_l, A_h, [dst, #48]
700+	ldrd	A_l, A_h, [src, #56]
701+	strd	A_l, A_h, [dst, #56]
702+	ldrd	A_l, A_h, [src, #64]!
703+	strd	A_l, A_h, [dst, #64]!
704+	subs	tmp2, tmp2, #64
705+	bhs	1b
706+	tst	tmp2, #0x3f
707+	bne	1f
708+	ldr	tmp2,[sp], #FRAME_SIZE
709+	cfi_adjust_cfa_offset (-FRAME_SIZE)
710+	cfi_restore (tmp2)
711+	bx	lr
712+
713+	cfi_restore_state
714+	cfi_remember_state
715+1:
716+	add	src, src, #8
717+	add	dst, dst, #8
718+
719+.Ltail63aligned:			/* Count in tmp2.  */
720+	/* Copy up to 7 d-words of data.  Similar to Ltail63unaligned, but
721+	   we know that the src and dest are 64-bit aligned so we can use
722+	   LDRD/STRD to improve efficiency.  */
723+	/* TMP2 is now negative, but we don't care about that.  The bottom
724+	   six bits still tell us how many bytes are left to copy.  */
725+
726+	and	tmp1, tmp2, #0x38
727+	add	dst, dst, tmp1
728+	add	src, src, tmp1
729+	.macro dispatch_step i
730+	ldrd	A_l, A_h, [src, #-(\i * 8)]
731+	strd	A_l, A_h, [dst, #-(\i * 8)]
732+	.endm
733+	dispatch_7_dword
734+#endif
735+
736+	tst	tmp2, #4
737+	ldrne	tmp1, [src], #4
738+	strne	tmp1, [dst], #4
739+	lsls	tmp2, tmp2, #31		/* Count (tmp2) now dead. */
740+	ldrhcs	tmp1, [src], #2
741+	ldrbne	tmp2, [src]
742+	strhcs	tmp1, [dst], #2
743+	strbne	tmp2, [dst]
744+
745+.Ldone:
746+	ldr	tmp2, [sp], #FRAME_SIZE
747+	cfi_adjust_cfa_offset (-FRAME_SIZE)
748+	cfi_restore (tmp2)
749+	bx	lr
750+
751+	cfi_restore_state
752+	cfi_remember_state
753+
754+.Lcpy_body_long:			/* Count in tmp2.  */
755+
756+	/* Long copy.  We know that there's at least (prefetch_lines * 64)
757+	   bytes to go.  */
758+#ifdef USE_VFP
759+	/* Don't use PLD.  Instead, read some data in advance of the current
760+	   copy position into a register.  This should act like a PLD
761+	   operation but we won't have to repeat the transfer.  */
762+
763+	vldr	d3, [src, #0]
764+	vldr	d4, [src, #64]
765+	vldr	d5, [src, #128]
766+	vldr	d6, [src, #192]
767+	vldr	d7, [src, #256]
768+
769+	vldr	d0, [src, #8]
770+	vldr	d1, [src, #16]
771+	vldr	d2, [src, #24]
772+	add	src, src, #32
773+
774+	subs	tmp2, tmp2, #prefetch_lines * 64 * 2
775+	blo	2f
776+1:
777+	cpy_line_vfp	d3, 0
778+	cpy_line_vfp	d4, 64
779+	cpy_line_vfp	d5, 128
780+	add	dst, dst, #3 * 64
781+	add	src, src, #3 * 64
782+	cpy_line_vfp	d6, 0
783+	cpy_line_vfp	d7, 64
784+	add	dst, dst, #2 * 64
785+	add	src, src, #2 * 64
786+	subs	tmp2, tmp2, #prefetch_lines * 64
787+	bhs	1b
788+
789+2:
790+	cpy_tail_vfp	d3, 0
791+	cpy_tail_vfp	d4, 64
792+	cpy_tail_vfp	d5, 128
793+	add	src, src, #3 * 64
794+	add	dst, dst, #3 * 64
795+	cpy_tail_vfp	d6, 0
796+	vstr	d7, [dst, #64]
797+	vldr	d7, [src, #64]
798+	vstr	d0, [dst, #64 + 8]
799+	vldr	d0, [src, #64 + 8]
800+	vstr	d1, [dst, #64 + 16]
801+	vldr	d1, [src, #64 + 16]
802+	vstr	d2, [dst, #64 + 24]
803+	vldr	d2, [src, #64 + 24]
804+	vstr	d7, [dst, #64 + 32]
805+	add	src, src, #96
806+	vstr	d0, [dst, #64 + 40]
807+	vstr	d1, [dst, #64 + 48]
808+	vstr	d2, [dst, #64 + 56]
809+	add	dst, dst, #128
810+	add	tmp2, tmp2, #prefetch_lines * 64
811+	b	.Lcpy_body_medium
812+#else
813+	/* Long copy.  Use an SMS style loop to maximize the I/O
814+	   bandwidth of the core.  We don't have enough spare registers
815+	   to synthesise prefetching, so use PLD operations.  */
816+	/* Pre-bias src and dst.  */
817+	sub	src, src, #8
818+	sub	dst, dst, #8
819+	pld	[src, #8]
820+	pld	[src, #72]
821+	subs	tmp2, tmp2, #64
822+	pld	[src, #136]
823+	ldrd	A_l, A_h, [src, #8]
824+	strd	B_l, B_h, [sp, #8]
825+	cfi_rel_offset (B_l, 8)
826+	cfi_rel_offset (B_h, 12)
827+	ldrd	B_l, B_h, [src, #16]
828+	strd	C_l, C_h, [sp, #16]
829+	cfi_rel_offset (C_l, 16)
830+	cfi_rel_offset (C_h, 20)
831+	ldrd	C_l, C_h, [src, #24]
832+	strd	D_l, D_h, [sp, #24]
833+	cfi_rel_offset (D_l, 24)
834+	cfi_rel_offset (D_h, 28)
835+	pld	[src, #200]
836+	ldrd	D_l, D_h, [src, #32]!
837+	b	1f
838+	.p2align	6
839+2:
840+	pld	[src, #232]
841+	strd	A_l, A_h, [dst, #40]
842+	ldrd	A_l, A_h, [src, #40]
843+	strd	B_l, B_h, [dst, #48]
844+	ldrd	B_l, B_h, [src, #48]
845+	strd	C_l, C_h, [dst, #56]
846+	ldrd	C_l, C_h, [src, #56]
847+	strd	D_l, D_h, [dst, #64]!
848+	ldrd	D_l, D_h, [src, #64]!
849+	subs	tmp2, tmp2, #64
850+1:
851+	strd	A_l, A_h, [dst, #8]
852+	ldrd	A_l, A_h, [src, #8]
853+	strd	B_l, B_h, [dst, #16]
854+	ldrd	B_l, B_h, [src, #16]
855+	strd	C_l, C_h, [dst, #24]
856+	ldrd	C_l, C_h, [src, #24]
857+	strd	D_l, D_h, [dst, #32]
858+	ldrd	D_l, D_h, [src, #32]
859+	bcs	2b
860+	/* Save the remaining bytes and restore the callee-saved regs.  */
861+	strd	A_l, A_h, [dst, #40]
862+	add	src, src, #40
863+	strd	B_l, B_h, [dst, #48]
864+	ldrd	B_l, B_h, [sp, #8]
865+	cfi_restore (B_l)
866+	cfi_restore (B_h)
867+	strd	C_l, C_h, [dst, #56]
868+	ldrd	C_l, C_h, [sp, #16]
869+	cfi_restore (C_l)
870+	cfi_restore (C_h)
871+	strd	D_l, D_h, [dst, #64]
872+	ldrd	D_l, D_h, [sp, #24]
873+	cfi_restore (D_l)
874+	cfi_restore (D_h)
875+	add	dst, dst, #72
876+	tst	tmp2, #0x3f
877+	bne	.Ltail63aligned
878+	ldr	tmp2, [sp], #FRAME_SIZE
879+	cfi_adjust_cfa_offset (-FRAME_SIZE)
880+	cfi_restore (tmp2)
881+	bx	lr
882+#endif
883+
884+	cfi_restore_state
885+	cfi_remember_state
886+
887+.Lcpy_notaligned:
888+	pld	[src, #0]
889+	pld	[src, #64]
890+	/* There's at least 64 bytes to copy, but there is no mutual
891+	   alignment.  */
892+	/* Bring DST to 64-bit alignment.  */
893+	lsls	tmp2, dst, #29
894+	pld	[src, #(2 * 64)]
895+	beq	1f
896+	rsbs	tmp2, tmp2, #0
897+	sub	count, count, tmp2, lsr #29
898+	ldrmi	tmp1, [src], #4
899+	strmi	tmp1, [dst], #4
900+	lsls	tmp2, tmp2, #2
901+	ldrbne	tmp1, [src], #1
902+	ldrhcs	tmp2, [src], #2
903+	strbne	tmp1, [dst], #1
904+	strhcs	tmp2, [dst], #2
905+1:
906+	pld	[src, #(3 * 64)]
907+	subs	count, count, #64
908+	ldrlo	tmp2, [sp], #FRAME_SIZE
909+	blo	.Ltail63unaligned
910+	pld	[src, #(4 * 64)]
911+
912+#ifdef USE_NEON
913+	/* These need an extra layer of macro just to work around a
914+	   bug in the assembler's parser when an operand starts with
915+	   a {...}.  */
916+	.macro neon_load_multi reglist, basereg
917+	vld1.8	{\reglist}, [\basereg]!
918+	.endm
919+	.macro neon_store_multi reglist, basereg
920+	vst1.8	{\reglist}, [ALIGN (\basereg, 64)]!
921+	.endm
922+
923+	neon_load_multi d0-d3, src
924+	neon_load_multi d4-d7, src
925+	subs	count, count, #64
926+	blo	2f
927+1:
928+	pld	[src, #(4 * 64)]
929+	neon_store_multi d0-d3, dst
930+	neon_load_multi d0-d3, src
931+	neon_store_multi d4-d7, dst
932+	neon_load_multi d4-d7, src
933+	subs	count, count, #64
934+	bhs	1b
935+2:
936+	neon_store_multi d0-d3, dst
937+	neon_store_multi d4-d7, dst
938+	ands	count, count, #0x3f
939+#else
940+	/* Use an SMS style loop to maximize the I/O bandwidth.  */
941+	sub	src, src, #4
942+	sub	dst, dst, #8
943+	subs	tmp2, count, #64	/* Use tmp2 for count.  */
944+	ldr	A_l, [src, #4]
945+	ldr	A_h, [src, #8]
946+	strd	B_l, B_h, [sp, #8]
947+	cfi_rel_offset (B_l, 8)
948+	cfi_rel_offset (B_h, 12)
949+	ldr	B_l, [src, #12]
950+	ldr	B_h, [src, #16]
951+	strd	C_l, C_h, [sp, #16]
952+	cfi_rel_offset (C_l, 16)
953+	cfi_rel_offset (C_h, 20)
954+	ldr	C_l, [src, #20]
955+	ldr	C_h, [src, #24]
956+	strd	D_l, D_h, [sp, #24]
957+	cfi_rel_offset (D_l, 24)
958+	cfi_rel_offset (D_h, 28)
959+	ldr	D_l, [src, #28]
960+	ldr	D_h, [src, #32]!
961+	b	1f
962+	.p2align	6
963+2:
964+	pld	[src, #(5 * 64) - (32 - 4)]
965+	strd	A_l, A_h, [dst, #40]
966+	ldr	A_l, [src, #36]
967+	ldr	A_h, [src, #40]
968+	strd	B_l, B_h, [dst, #48]
969+	ldr	B_l, [src, #44]
970+	ldr	B_h, [src, #48]
971+	strd	C_l, C_h, [dst, #56]
972+	ldr	C_l, [src, #52]
973+	ldr	C_h, [src, #56]
974+	strd	D_l, D_h, [dst, #64]!
975+	ldr	D_l, [src, #60]
976+	ldr	D_h, [src, #64]!
977+	subs	tmp2, tmp2, #64
978+1:
979+	strd	A_l, A_h, [dst, #8]
980+	ldr	A_l, [src, #4]
981+	ldr	A_h, [src, #8]
982+	strd	B_l, B_h, [dst, #16]
983+	ldr	B_l, [src, #12]
984+	ldr	B_h, [src, #16]
985+	strd	C_l, C_h, [dst, #24]
986+	ldr	C_l, [src, #20]
987+	ldr	C_h, [src, #24]
988+	strd	D_l, D_h, [dst, #32]
989+	ldr	D_l, [src, #28]
990+	ldr	D_h, [src, #32]
991+	bcs	2b
992+
993+	/* Save the remaining bytes and restore the callee-saved regs.  */
994+	strd	A_l, A_h, [dst, #40]
995+	add	src, src, #36
996+	strd	B_l, B_h, [dst, #48]
997+	ldrd	B_l, B_h, [sp, #8]
998+	cfi_restore (B_l)
999+	cfi_restore (B_h)
1000+	strd	C_l, C_h, [dst, #56]
1001+	ldrd	C_l, C_h, [sp, #16]
1002+	cfi_restore (C_l)
1003+	cfi_restore (C_h)
1004+	strd	D_l, D_h, [dst, #64]
1005+	ldrd	D_l, D_h, [sp, #24]
1006+	cfi_restore (D_l)
1007+	cfi_restore (D_h)
1008+	add	dst, dst, #72
1009+	ands	count, tmp2, #0x3f
1010+#endif
1011+	ldr	tmp2, [sp], #FRAME_SIZE
1012+	cfi_adjust_cfa_offset (-FRAME_SIZE)
1013+	cfi_restore (tmp2)
1014+	bne	.Ltail63unaligned
1015+	bx	lr
1016+
1017+END(memcpy)
1018+libc_hidden_builtin_def (memcpy)
1019diff --git a/libc/string/arm/glibc-neon/memmove.S b/libc/string/arm/glibc-neon/memmove.S
1020new file mode 100644
1021index 0000000..b01164a
1022--- /dev/null
1023+++ b/libc/string/arm/glibc-neon/memmove.S
1024@@ -0,0 +1,332 @@
1025+/* Copyright (C) 2006-2021 Free Software Foundation, Inc.
1026+   This file is part of the GNU C Library.
1027+
1028+   Contributed by MontaVista Software, Inc. (written by Nicolas Pitre)
1029+
1030+   The GNU C Library is free software; you can redistribute it and/or
1031+   modify it under the terms of the GNU Lesser General Public
1032+   License as published by the Free Software Foundation; either
1033+   version 2.1 of the License, or (at your option) any later version.
1034+
1035+   The GNU C Library is distributed in the hope that it will be useful,
1036+   but WITHOUT ANY WARRANTY; without even the implied warranty of
1037+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
1038+   Lesser General Public License for more details.
1039+
1040+   You should have received a copy of the GNU Lesser General Public
1041+   License along with the GNU C Library.  If not, see
1042+   <https://www.gnu.org/licenses/>.  */
1043+
1044+/* Thumb requires excessive IT insns here.  */
1045+#define NO_THUMB
1046+#include <sysdep.h>
1047+#include <arm-features.h>
1048+
1049+/*
1050+ * Data preload for architectures that support it (ARM V5TE and above)
1051+ */
1052+#if (!defined (__ARM_ARCH_2__) && !defined (__ARM_ARCH_3__) \
1053+     && !defined (__ARM_ARCH_3M__) && !defined (__ARM_ARCH_4__) \
1054+     && !defined (__ARM_ARCH_4T__) && !defined (__ARM_ARCH_5__) \
1055+     && !defined (__ARM_ARCH_5T__))
1056+#define PLD(code...)    code
1057+#else
1058+#define PLD(code...)
1059+#endif
1060+
1061+/*
1062+ * This can be used to enable code to cacheline align the source pointer.
1063+ * Experiments on tested architectures (StrongARM and XScale) didn't show
1064+ * this a worthwhile thing to do.  That might be different in the future.
1065+ */
1066+//#define CALGN(code...)        code
1067+#define CALGN(code...)
1068+
1069+/*
1070+ * Endian independent macros for shifting bytes within registers.
1071+ */
1072+#ifndef __ARMEB__
1073+#define PULL            lsr
1074+#define PUSH            lsl
1075+#else
1076+#define PULL            lsl
1077+#define PUSH            lsr
1078+#endif
1079+
1080+		.text
1081+		.syntax unified
1082+
1083+/*
1084+ * Prototype: void *memmove(void *dest, const void *src, size_t n);
1085+ *
1086+ * Note:
1087+ *
1088+ * If the memory regions don't overlap, we simply branch to memcpy which is
1089+ * normally a bit faster. Otherwise the copy is done going downwards.
1090+ */
1091+
1092+ENTRY(memmove)
1093+
1094+		subs	ip, r0, r1
1095+		cmphi	r2, ip
1096+#if !IS_IN (libc)
1097+		bls	memcpy
1098+#else
1099+		bls	HIDDEN_JUMPTARGET(memcpy)
1100+#endif
1101+
1102+		push	{r0, r4, lr}
1103+		cfi_adjust_cfa_offset (12)
1104+		cfi_rel_offset (r4, 4)
1105+		cfi_rel_offset (lr, 8)
1106+
1107+		cfi_remember_state
1108+
1109+		add	r1, r1, r2
1110+		add	r0, r0, r2
1111+		subs	r2, r2, #4
1112+		blo	8f
1113+		ands	ip, r0, #3
1114+	PLD(	pld	[r1, #-4]		)
1115+		bne	9f
1116+		ands	ip, r1, #3
1117+		bne	10f
1118+
1119+1:		subs	r2, r2, #(28)
1120+		push	{r5 - r8}
1121+		cfi_adjust_cfa_offset (16)
1122+		cfi_rel_offset (r5, 0)
1123+		cfi_rel_offset (r6, 4)
1124+		cfi_rel_offset (r7, 8)
1125+		cfi_rel_offset (r8, 12)
1126+		blo     5f
1127+
1128+	CALGN(	ands	ip, r1, #31		)
1129+	CALGN(	sbcsne	r4, ip, r2		)  @ C is always set here
1130+	CALGN(	bcs	2f			)
1131+	CALGN(	adr	r4, 6f			)
1132+	CALGN(	subs	r2, r2, ip		)  @ C is set here
1133+#ifndef ARM_ALWAYS_BX
1134+	CALGN(	add	pc, r4, ip, lsl	#(ARM_BX_ALIGN_LOG2 - 2))
1135+#else
1136+	CALGN(	add	r4, r4, ip, lsl	#(ARM_BX_ALIGN_LOG2 - 2))
1137+	CALGN(	bx	r4			)
1138+#endif
1139+
1140+	PLD(	pld	[r1, #-4]		)
1141+2:	PLD(	cmp	r2, #96			)
1142+	PLD(	pld	[r1, #-32]		)
1143+	PLD(    blo     4f                      )
1144+	PLD(	pld	[r1, #-64]		)
1145+	PLD(	pld	[r1, #-96]		)
1146+
1147+3:	PLD(	pld	[r1, #-128]		)
1148+4:		ldmdb	r1!, {r3, r4, r5, r6, r7, r8, ip, lr}
1149+		subs	r2, r2, #32
1150+		stmdb	r0!, {r3, r4, r5, r6, r7, r8, ip, lr}
1151+		bhs     3b
1152+
1153+5:		ands	ip, r2, #28
1154+		rsb	ip, ip, #32
1155+#ifndef ARM_ALWAYS_BX
1156+		/* C is always clear here.  */
1157+		addne	pc, pc, ip, lsl #(ARM_BX_ALIGN_LOG2 - 2)
1158+		b	7f
1159+#else
1160+		beq	7f
1161+		push	{r10}
1162+		cfi_adjust_cfa_offset (4)
1163+		cfi_rel_offset (r10, 0)
1164+0:		add	r10, pc, ip, lsl #(ARM_BX_ALIGN_LOG2 - 2)
1165+		/* If alignment is not perfect, then there will be some
1166+		   padding (nop) instructions between this BX and label 6.
1167+		   The computation above assumed that two instructions
1168+		   later is exactly the right spot.  */
1169+		add	r10, #(6f - (0b + PC_OFS))
1170+		bx	r10
1171+#endif
1172+		.p2align ARM_BX_ALIGN_LOG2
1173+6:		nop
1174+		.p2align ARM_BX_ALIGN_LOG2
1175+		ldr	r3, [r1, #-4]!
1176+		.p2align ARM_BX_ALIGN_LOG2
1177+		ldr	r4, [r1, #-4]!
1178+		.p2align ARM_BX_ALIGN_LOG2
1179+		ldr	r5, [r1, #-4]!
1180+		.p2align ARM_BX_ALIGN_LOG2
1181+		ldr	r6, [r1, #-4]!
1182+		.p2align ARM_BX_ALIGN_LOG2
1183+		ldr	r7, [r1, #-4]!
1184+		.p2align ARM_BX_ALIGN_LOG2
1185+		ldr	r8, [r1, #-4]!
1186+		.p2align ARM_BX_ALIGN_LOG2
1187+		ldr	lr, [r1, #-4]!
1188+
1189+#ifndef ARM_ALWAYS_BX
1190+		add	pc, pc, ip, lsl #(ARM_BX_ALIGN_LOG2 - 2)
1191+		nop
1192+#else
1193+0:		add	r10, pc, ip, lsl #(ARM_BX_ALIGN_LOG2 - 2)
1194+		/* If alignment is not perfect, then there will be some
1195+		   padding (nop) instructions between this BX and label 66.
1196+		   The computation above assumed that two instructions
1197+		   later is exactly the right spot.  */
1198+		add	r10, #(66f - (0b + PC_OFS))
1199+		bx	r10
1200+#endif
1201+		.p2align ARM_BX_ALIGN_LOG2
1202+66:		nop
1203+		.p2align ARM_BX_ALIGN_LOG2
1204+		str	r3, [r0, #-4]!
1205+		.p2align ARM_BX_ALIGN_LOG2
1206+		str	r4, [r0, #-4]!
1207+		.p2align ARM_BX_ALIGN_LOG2
1208+		str	r5, [r0, #-4]!
1209+		.p2align ARM_BX_ALIGN_LOG2
1210+		str	r6, [r0, #-4]!
1211+		.p2align ARM_BX_ALIGN_LOG2
1212+		str	r7, [r0, #-4]!
1213+		.p2align ARM_BX_ALIGN_LOG2
1214+		str	r8, [r0, #-4]!
1215+		.p2align ARM_BX_ALIGN_LOG2
1216+		str	lr, [r0, #-4]!
1217+
1218+#ifdef ARM_ALWAYS_BX
1219+		pop	{r10}
1220+		cfi_adjust_cfa_offset (-4)
1221+		cfi_restore (r10)
1222+#endif
1223+
1224+	CALGN(	bcs	2b			)
1225+
1226+7:		pop	{r5 - r8}
1227+		cfi_adjust_cfa_offset (-16)
1228+		cfi_restore (r5)
1229+		cfi_restore (r6)
1230+		cfi_restore (r7)
1231+		cfi_restore (r8)
1232+
1233+8:		movs	r2, r2, lsl #31
1234+		ldrbne	r3, [r1, #-1]!
1235+		ldrbcs	r4, [r1, #-1]!
1236+		ldrbcs	ip, [r1, #-1]
1237+		strbne	r3, [r0, #-1]!
1238+		strbcs	r4, [r0, #-1]!
1239+		strbcs	ip, [r0, #-1]
1240+
1241+#if ((defined (__ARM_ARCH_4T__) && defined (__THUMB_INTERWORK__)) \
1242+     || defined (ARM_ALWAYS_BX))
1243+		pop	{r0, r4, lr}
1244+		cfi_adjust_cfa_offset (-12)
1245+		cfi_restore (r4)
1246+		cfi_restore (lr)
1247+		bx      lr
1248+#else
1249+		pop	{r0, r4, pc}
1250+#endif
1251+
1252+		cfi_restore_state
1253+
1254+9:		cmp	ip, #2
1255+		ldrbgt	r3, [r1, #-1]!
1256+		ldrbge	r4, [r1, #-1]!
1257+		ldrb	lr, [r1, #-1]!
1258+		strbgt	r3, [r0, #-1]!
1259+		strbge	r4, [r0, #-1]!
1260+		subs	r2, r2, ip
1261+		strb	lr, [r0, #-1]!
1262+		blo	8b
1263+		ands	ip, r1, #3
1264+		beq	1b
1265+
1266+10:		bic	r1, r1, #3
1267+		cmp	ip, #2
1268+		ldr	r3, [r1, #0]
1269+		beq	17f
1270+		blt	18f
1271+
1272+
1273+		.macro	backward_copy_shift push pull
1274+
1275+		subs	r2, r2, #28
1276+		blo	14f
1277+
1278+	CALGN(	ands	ip, r1, #31		)
1279+	CALGN(	rsb	ip, ip, #32		)
1280+	CALGN(	sbcsne	r4, ip, r2		)  @ C is always set here
1281+	CALGN(	subcc	r2, r2, ip		)
1282+	CALGN(	bcc	15f			)
1283+
1284+11:		push	{r5 - r8, r10}
1285+		cfi_adjust_cfa_offset (20)
1286+		cfi_rel_offset (r5, 0)
1287+		cfi_rel_offset (r6, 4)
1288+		cfi_rel_offset (r7, 8)
1289+		cfi_rel_offset (r8, 12)
1290+		cfi_rel_offset (r10, 16)
1291+
1292+	PLD(	pld	[r1, #-4]		)
1293+	PLD(	cmp	r2, #96			)
1294+	PLD(	pld	[r1, #-32]		)
1295+	PLD(	blo	13f			)
1296+	PLD(	pld	[r1, #-64]		)
1297+	PLD(	pld	[r1, #-96]		)
1298+
1299+12:	PLD(	pld	[r1, #-128]		)
1300+13:		ldmdb   r1!, {r7, r8, r10, ip}
1301+		mov     lr, r3, PUSH #\push
1302+		subs    r2, r2, #32
1303+		ldmdb   r1!, {r3, r4, r5, r6}
1304+		orr     lr, lr, ip, PULL #\pull
1305+		mov     ip, ip, PUSH #\push
1306+		orr     ip, ip, r10, PULL #\pull
1307+		mov     r10, r10, PUSH #\push
1308+		orr     r10, r10, r8, PULL #\pull
1309+		mov     r8, r8, PUSH #\push
1310+		orr     r8, r8, r7, PULL #\pull
1311+		mov     r7, r7, PUSH #\push
1312+		orr     r7, r7, r6, PULL #\pull
1313+		mov     r6, r6, PUSH #\push
1314+		orr     r6, r6, r5, PULL #\pull
1315+		mov     r5, r5, PUSH #\push
1316+		orr     r5, r5, r4, PULL #\pull
1317+		mov     r4, r4, PUSH #\push
1318+		orr     r4, r4, r3, PULL #\pull
1319+		stmdb   r0!, {r4 - r8, r10, ip, lr}
1320+		bhs	12b
1321+
1322+		pop	{r5 - r8, r10}
1323+		cfi_adjust_cfa_offset (-20)
1324+		cfi_restore (r5)
1325+		cfi_restore (r6)
1326+		cfi_restore (r7)
1327+		cfi_restore (r8)
1328+		cfi_restore (r10)
1329+
1330+14:		ands	ip, r2, #28
1331+		beq	16f
1332+
1333+15:		mov     lr, r3, PUSH #\push
1334+		ldr	r3, [r1, #-4]!
1335+		subs	ip, ip, #4
1336+		orr	lr, lr, r3, PULL #\pull
1337+		str	lr, [r0, #-4]!
1338+		bgt	15b
1339+	CALGN(	cmp	r2, #0			)
1340+	CALGN(	bge	11b			)
1341+
1342+16:		add	r1, r1, #(\pull / 8)
1343+		b	8b
1344+
1345+		.endm
1346+
1347+
1348+		backward_copy_shift	push=8	pull=24
1349+
1350+17:		backward_copy_shift	push=16	pull=16
1351+
1352+18:		backward_copy_shift	push=24	pull=8
1353+
1354+
1355+END(memmove)
1356+libc_hidden_builtin_def (memmove)
1357diff --git a/libc/string/arm/glibc-neon/memset.S b/libc/string/arm/glibc-neon/memset.S
1358new file mode 100644
1359index 0000000..dc89ca7
1360--- /dev/null
1361+++ b/libc/string/arm/glibc-neon/memset.S
1362@@ -0,0 +1,68 @@
1363+/* Copyright (C) 1998-2021 Free Software Foundation, Inc.
1364+   This file is part of the GNU C Library.
1365+
1366+   The GNU C Library is free software; you can redistribute it and/or
1367+   modify it under the terms of the GNU Lesser General Public
1368+   License as published by the Free Software Foundation; either
1369+   version 2.1 of the License, or (at your option) any later version.
1370+
1371+   The GNU C Library is distributed in the hope that it will be useful,
1372+   but WITHOUT ANY WARRANTY; without even the implied warranty of
1373+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
1374+   Lesser General Public License for more details.
1375+
1376+   You should have received a copy of the GNU Lesser General Public
1377+   License along with the GNU C Library.  If not, see
1378+   <https://www.gnu.org/licenses/>.  */
1379+
1380+/* Thumb requires excessive IT insns here.  */
1381+#define NO_THUMB
1382+#include <sysdep.h>
1383+
1384+	.text
1385+	.syntax unified
1386+
1387+/* void *memset (dstpp, c, len) */
1388+
1389+ENTRY(memset)
1390+	mov	r3, r0
1391+	cmp	r2, #8
1392+	bcc	2f		@ less than 8 bytes to move
1393+
1394+1:
1395+	tst	r3, #3		@ aligned yet?
1396+	strbne	r1, [r3], #1
1397+	subne	r2, r2, #1
1398+	bne	1b
1399+
1400+	and	r1, r1, #255	@ clear any sign bits
1401+	orr	r1, r1, r1, lsl $8
1402+	orr	r1, r1, r1, lsl $16
1403+	mov	ip, r1
1404+
1405+1:
1406+	subs	r2, r2, #8
1407+	stmiacs	r3!, {r1, ip}	@ store up to 32 bytes per loop iteration
1408+	subscs	r2, r2, #8
1409+	stmiacs	r3!, {r1, ip}
1410+	subscs	r2, r2, #8
1411+	stmiacs	r3!, {r1, ip}
1412+	subscs	r2, r2, #8
1413+	stmiacs	r3!, {r1, ip}
1414+	bcs	1b
1415+
1416+	and	r2, r2, #7
1417+2:
1418+	subs	r2, r2, #1	@ store up to 4 bytes per loop iteration
1419+	strbcs	r1, [r3], #1
1420+	subscs	r2, r2, #1
1421+	strbcs	r1, [r3], #1
1422+	subscs	r2, r2, #1
1423+	strbcs	r1, [r3], #1
1424+	subscs	r2, r2, #1
1425+	strbcs	r1, [r3], #1
1426+	bcs	2b
1427+
1428+	DO_RET(lr)
1429+END(memset)
1430+libc_hidden_builtin_def (memset)
1431diff --git a/libc/string/arm/glibc-neon/strcmp.S b/libc/string/arm/glibc-neon/strcmp.S
1432new file mode 100644
1433index 0000000..98a3c6c
1434--- /dev/null
1435+++ b/libc/string/arm/glibc-neon/strcmp.S
1436@@ -0,0 +1,504 @@
1437+/* strcmp implementation for ARMv7-A, optimized for Cortex-A15.
1438+   Copyright (C) 2012-2021 Free Software Foundation, Inc.
1439+   This file is part of the GNU C Library.
1440+
1441+   The GNU C Library is free software; you can redistribute it and/or
1442+   modify it under the terms of the GNU Lesser General Public
1443+   License as published by the Free Software Foundation; either
1444+   version 2.1 of the License, or (at your option) any later version.
1445+
1446+   The GNU C Library is distributed in the hope that it will be useful,
1447+   but WITHOUT ANY WARRANTY; without even the implied warranty of
1448+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
1449+   Lesser General Public License for more details.
1450+
1451+   You should have received a copy of the GNU Lesser General Public
1452+   License along with the GNU C Library.  If not, see
1453+   <https://www.gnu.org/licenses/>.  */
1454+
1455+#include <arm-features.h>
1456+#include <sysdep.h>
1457+
1458+/* Implementation of strcmp for ARMv7 when DSP instructions are
1459+   available.  Use ldrd to support wider loads, provided the data
1460+   is sufficiently aligned.  Use saturating arithmetic to optimize
1461+   the compares.  */
1462+
1463+/* Build Options:
1464+   STRCMP_PRECHECK: Run a quick pre-check of the first byte in the
1465+   string.  If comparing completely random strings the pre-check will
1466+   save time, since there is a very high probability of a mismatch in
1467+   the first character: we save significant overhead if this is the
1468+   common case.  However, if strings are likely to be identical (e.g.
1469+   because we're verifying a hit in a hash table), then this check
1470+   is largely redundant.  */
1471+
1472+#define STRCMP_PRECHECK	1
1473+
1474+	.syntax unified
1475+
1476+#ifdef __ARM_BIG_ENDIAN
1477+# define S2LO lsl
1478+# define S2LOEQ lsleq
1479+# define S2HI lsr
1480+# define MSB 0x000000ff
1481+# define LSB 0xff000000
1482+# define BYTE0_OFFSET 24
1483+# define BYTE1_OFFSET 16
1484+# define BYTE2_OFFSET 8
1485+# define BYTE3_OFFSET 0
1486+#else /* not  __ARM_BIG_ENDIAN */
1487+# define S2LO lsr
1488+# define S2LOEQ lsreq
1489+# define S2HI lsl
1490+# define BYTE0_OFFSET 0
1491+# define BYTE1_OFFSET 8
1492+# define BYTE2_OFFSET 16
1493+# define BYTE3_OFFSET 24
1494+# define MSB 0xff000000
1495+# define LSB 0x000000ff
1496+#endif /* not  __ARM_BIG_ENDIAN */
1497+
1498+/* Parameters and result.  */
1499+#define src1		r0
1500+#define src2		r1
1501+#define result		r0	/* Overlaps src1.  */
1502+
1503+/* Internal variables.  */
1504+#define tmp1		r4
1505+#define tmp2		r5
1506+#define const_m1	r12
1507+
1508+/* Additional internal variables for 64-bit aligned data.  */
1509+#define data1a		r2
1510+#define data1b		r3
1511+#define data2a		r6
1512+#define data2b		r7
1513+#define syndrome_a	tmp1
1514+#define syndrome_b	tmp2
1515+
1516+/* Additional internal variables for 32-bit aligned data.  */
1517+#define data1		r2
1518+#define data2		r3
1519+#define syndrome	tmp2
1520+
1521+
1522+	.thumb
1523+
1524+/* In Thumb code we can't use MVN with a register shift, but we do have ORN.  */
1525+.macro prepare_mask mask_reg, nbits_reg
1526+	S2HI \mask_reg, const_m1, \nbits_reg
1527+.endm
1528+.macro apply_mask data_reg, mask_reg
1529+	orn \data_reg, \data_reg, \mask_reg
1530+.endm
1531+
1532+	/* Macro to compute and return the result value for word-aligned
1533+	   cases.  */
1534+	.macro strcmp_epilogue_aligned synd d1 d2 restore_r6
1535+#ifdef __ARM_BIG_ENDIAN
1536+	/* If data1 contains a zero byte, then syndrome will contain a 1 in
1537+	   bit 7 of that byte.  Otherwise, the highest set bit in the
1538+	   syndrome will highlight the first different bit.  It is therefore
1539+	   sufficient to extract the eight bits starting with the syndrome
1540+	   bit.  */
1541+	clz	tmp1, \synd
1542+	lsl	r1, \d2, tmp1
1543+	.if \restore_r6
1544+	ldrd	r6, r7, [sp, #8]
1545+	.endif
1546+	lsl	\d1, \d1, tmp1
1547+	lsr	result, \d1, #24
1548+	ldrd	r4, r5, [sp], #16
1549+	cfi_remember_state
1550+	cfi_def_cfa_offset (0)
1551+	cfi_restore (r4)
1552+	cfi_restore (r5)
1553+	cfi_restore (r6)
1554+	cfi_restore (r7)
1555+	sub	result, result, r1, lsr #24
1556+	bx	lr
1557+#else
1558+	/* To use the big-endian trick we'd have to reverse all three words.
1559+	   that's slower than this approach.  */
1560+	rev	\synd, \synd
1561+	clz	tmp1, \synd
1562+	bic	tmp1, tmp1, #7
1563+	lsr	r1, \d2, tmp1
1564+	.if \restore_r6
1565+	ldrd	r6, r7, [sp, #8]
1566+	.endif
1567+	lsr	\d1, \d1, tmp1
1568+	and	result, \d1, #255
1569+	and	r1, r1, #255
1570+	ldrd	r4, r5, [sp], #16
1571+	cfi_remember_state
1572+	cfi_def_cfa_offset (0)
1573+	cfi_restore (r4)
1574+	cfi_restore (r5)
1575+	cfi_restore (r6)
1576+	cfi_restore (r7)
1577+	sub	result, result, r1
1578+
1579+	bx	lr
1580+#endif
1581+	.endm
1582+
1583+	.text
1584+	.p2align	5
1585+.Lstrcmp_start_addr:
1586+#if STRCMP_PRECHECK == 1
1587+.Lfastpath_exit:
1588+	sub	r0, r2, r3
1589+	bx	lr
1590+	nop
1591+#endif
1592+ENTRY (strcmp)
1593+#if STRCMP_PRECHECK == 1
1594+	ldrb	r2, [src1]
1595+	ldrb	r3, [src2]
1596+	cmp	r2, #1
1597+	it	cs
1598+	cmpcs	r2, r3
1599+	bne	.Lfastpath_exit
1600+#endif
1601+	strd	r4, r5, [sp, #-16]!
1602+	cfi_def_cfa_offset (16)
1603+	cfi_offset (r4, -16)
1604+	cfi_offset (r5, -12)
1605+	orr	tmp1, src1, src2
1606+	strd	r6, r7, [sp, #8]
1607+	cfi_offset (r6, -8)
1608+	cfi_offset (r7, -4)
1609+	mvn	const_m1, #0
1610+	lsl	r2, tmp1, #29
1611+	cbz	r2, .Lloop_aligned8
1612+
1613+.Lnot_aligned:
1614+	eor	tmp1, src1, src2
1615+	tst	tmp1, #7
1616+	bne	.Lmisaligned8
1617+
1618+	/* Deal with mutual misalignment by aligning downwards and then
1619+	   masking off the unwanted loaded data to prevent a difference.  */
1620+	and	tmp1, src1, #7
1621+	bic	src1, src1, #7
1622+	and	tmp2, tmp1, #3
1623+	bic	src2, src2, #7
1624+	lsl	tmp2, tmp2, #3	/* Bytes -> bits.  */
1625+	ldrd	data1a, data1b, [src1], #16
1626+	tst	tmp1, #4
1627+	ldrd	data2a, data2b, [src2], #16
1628+	prepare_mask tmp1, tmp2
1629+	apply_mask data1a, tmp1
1630+	apply_mask data2a, tmp1
1631+	beq	.Lstart_realigned8
1632+	apply_mask data1b, tmp1
1633+	mov	data1a, const_m1
1634+	apply_mask data2b, tmp1
1635+	mov	data2a, const_m1
1636+	b	.Lstart_realigned8
1637+
1638+	/* Unwind the inner loop by a factor of 2, giving 16 bytes per
1639+	   pass.  */
1640+	.p2align 5,,12  /* Don't start in the tail bytes of a cache line.  */
1641+	.p2align 2	/* Always word aligned.  */
1642+.Lloop_aligned8:
1643+	ldrd	data1a, data1b, [src1], #16
1644+	ldrd	data2a, data2b, [src2], #16
1645+.Lstart_realigned8:
1646+	uadd8	syndrome_b, data1a, const_m1	/* Only want GE bits,  */
1647+	eor	syndrome_a, data1a, data2a
1648+	sel	syndrome_a, syndrome_a, const_m1
1649+	cbnz	syndrome_a, .Ldiff_in_a
1650+	uadd8	syndrome_b, data1b, const_m1	/* Only want GE bits.  */
1651+	eor	syndrome_b, data1b, data2b
1652+	sel	syndrome_b, syndrome_b, const_m1
1653+	cbnz	syndrome_b, .Ldiff_in_b
1654+
1655+	ldrd	data1a, data1b, [src1, #-8]
1656+	ldrd	data2a, data2b, [src2, #-8]
1657+	uadd8	syndrome_b, data1a, const_m1	/* Only want GE bits,  */
1658+	eor	syndrome_a, data1a, data2a
1659+	sel	syndrome_a, syndrome_a, const_m1
1660+	uadd8	syndrome_b, data1b, const_m1	/* Only want GE bits.  */
1661+	eor	syndrome_b, data1b, data2b
1662+	sel	syndrome_b, syndrome_b, const_m1
1663+	/* Can't use CBZ for backwards branch.  */
1664+	orrs	syndrome_b, syndrome_b, syndrome_a /* Only need if s_a == 0 */
1665+	beq	.Lloop_aligned8
1666+
1667+.Ldiff_found:
1668+	cbnz	syndrome_a, .Ldiff_in_a
1669+
1670+.Ldiff_in_b:
1671+	strcmp_epilogue_aligned syndrome_b, data1b, data2b 1
1672+
1673+.Ldiff_in_a:
1674+	cfi_restore_state
1675+	strcmp_epilogue_aligned syndrome_a, data1a, data2a 1
1676+
1677+	cfi_restore_state
1678+.Lmisaligned8:
1679+	tst	tmp1, #3
1680+	bne	.Lmisaligned4
1681+	ands	tmp1, src1, #3
1682+	bne	.Lmutual_align4
1683+
1684+	/* Unrolled by a factor of 2, to reduce the number of post-increment
1685+	   operations.  */
1686+.Lloop_aligned4:
1687+	ldr	data1, [src1], #8
1688+	ldr	data2, [src2], #8
1689+.Lstart_realigned4:
1690+	uadd8	syndrome, data1, const_m1	/* Only need GE bits.  */
1691+	eor	syndrome, data1, data2
1692+	sel	syndrome, syndrome, const_m1
1693+	cbnz	syndrome, .Laligned4_done
1694+	ldr	data1, [src1, #-4]
1695+	ldr	data2, [src2, #-4]
1696+	uadd8	syndrome, data1, const_m1
1697+	eor	syndrome, data1, data2
1698+	sel	syndrome, syndrome, const_m1
1699+	cmp	syndrome, #0
1700+	beq	.Lloop_aligned4
1701+
1702+.Laligned4_done:
1703+	strcmp_epilogue_aligned syndrome, data1, data2, 0
1704+
1705+.Lmutual_align4:
1706+	cfi_restore_state
1707+	/* Deal with mutual misalignment by aligning downwards and then
1708+	   masking off the unwanted loaded data to prevent a difference.  */
1709+	lsl	tmp1, tmp1, #3	/* Bytes -> bits.  */
1710+	bic	src1, src1, #3
1711+	ldr	data1, [src1], #8
1712+	bic	src2, src2, #3
1713+	ldr	data2, [src2], #8
1714+
1715+	prepare_mask tmp1, tmp1
1716+	apply_mask data1, tmp1
1717+	apply_mask data2, tmp1
1718+	b	.Lstart_realigned4
1719+
1720+.Lmisaligned4:
1721+	ands	tmp1, src1, #3
1722+	beq	.Lsrc1_aligned
1723+	sub	src2, src2, tmp1
1724+	bic	src1, src1, #3
1725+	lsls	tmp1, tmp1, #31
1726+	ldr	data1, [src1], #4
1727+	beq	.Laligned_m2
1728+	bcs	.Laligned_m1
1729+
1730+#if STRCMP_PRECHECK == 0
1731+	ldrb	data2, [src2, #1]
1732+	uxtb	tmp1, data1, ror #BYTE1_OFFSET
1733+	subs	tmp1, tmp1, data2
1734+	bne	.Lmisaligned_exit
1735+	cbz	data2, .Lmisaligned_exit
1736+
1737+.Laligned_m2:
1738+	ldrb	data2, [src2, #2]
1739+	uxtb	tmp1, data1, ror #BYTE2_OFFSET
1740+	subs	tmp1, tmp1, data2
1741+	bne	.Lmisaligned_exit
1742+	cbz	data2, .Lmisaligned_exit
1743+
1744+.Laligned_m1:
1745+	ldrb	data2, [src2, #3]
1746+	uxtb	tmp1, data1, ror #BYTE3_OFFSET
1747+	subs	tmp1, tmp1, data2
1748+	bne	.Lmisaligned_exit
1749+	add	src2, src2, #4
1750+	cbnz	data2, .Lsrc1_aligned
1751+#else  /* STRCMP_PRECHECK */
1752+	/* If we've done the pre-check, then we don't need to check the
1753+	   first byte again here.  */
1754+	ldrb	data2, [src2, #2]
1755+	uxtb	tmp1, data1, ror #BYTE2_OFFSET
1756+	subs	tmp1, tmp1, data2
1757+	bne	.Lmisaligned_exit
1758+	cbz	data2, .Lmisaligned_exit
1759+
1760+.Laligned_m2:
1761+	ldrb	data2, [src2, #3]
1762+	uxtb	tmp1, data1, ror #BYTE3_OFFSET
1763+	subs	tmp1, tmp1, data2
1764+	bne	.Lmisaligned_exit
1765+	cbnz	data2, .Laligned_m1
1766+#endif
1767+
1768+.Lmisaligned_exit:
1769+	mov	result, tmp1
1770+	ldr	r4, [sp], #16
1771+	cfi_remember_state
1772+	cfi_def_cfa_offset (0)
1773+	cfi_restore (r4)
1774+	cfi_restore (r5)
1775+	cfi_restore (r6)
1776+	cfi_restore (r7)
1777+	bx	lr
1778+
1779+#if STRCMP_PRECHECK == 1
1780+.Laligned_m1:
1781+	add	src2, src2, #4
1782+#endif
1783+.Lsrc1_aligned:
1784+	cfi_restore_state
1785+	/* src1 is word aligned, but src2 has no common alignment
1786+	   with it.  */
1787+	ldr	data1, [src1], #4
1788+	lsls	tmp1, src2, #31		/* C=src2[1], Z=src2[0].  */
1789+
1790+	bic	src2, src2, #3
1791+	ldr	data2, [src2], #4
1792+	bhi	.Loverlap1		/* C=1, Z=0 => src2[1:0] = 0b11.  */
1793+	bcs	.Loverlap2		/* C=1, Z=1 => src2[1:0] = 0b10.  */
1794+
1795+	/* (overlap3) C=0, Z=0 => src2[1:0] = 0b01.  */
1796+.Loverlap3:
1797+	bic	tmp1, data1, #MSB
1798+	uadd8	syndrome, data1, const_m1
1799+	eors	syndrome, tmp1, data2, S2LO #8
1800+	sel	syndrome, syndrome, const_m1
1801+	bne	4f
1802+	cbnz	syndrome, 5f
1803+	ldr	data2, [src2], #4
1804+	eor	tmp1, tmp1, data1
1805+	cmp	tmp1, data2, S2HI #24
1806+	bne	6f
1807+	ldr	data1, [src1], #4
1808+	b	.Loverlap3
1809+4:
1810+	S2LO	data2, data2, #8
1811+	b	.Lstrcmp_tail
1812+
1813+5:
1814+	bics	syndrome, syndrome, #MSB
1815+	bne	.Lstrcmp_done_equal
1816+
1817+	/* We can only get here if the MSB of data1 contains 0, so
1818+	   fast-path the exit.  */
1819+	ldrb	result, [src2]
1820+	ldrd	r4, r5, [sp], #16
1821+	cfi_remember_state
1822+	cfi_def_cfa_offset (0)
1823+	cfi_restore (r4)
1824+	cfi_restore (r5)
1825+	/* R6/7 Not used in this sequence.  */
1826+	cfi_restore (r6)
1827+	cfi_restore (r7)
1828+	neg	result, result
1829+	bx	lr
1830+
1831+6:
1832+	cfi_restore_state
1833+	S2LO	data1, data1, #24
1834+	and	data2, data2, #LSB
1835+	b	.Lstrcmp_tail
1836+
1837+	.p2align 5,,12	/* Ensure at least 3 instructions in cache line.  */
1838+.Loverlap2:
1839+	and	tmp1, data1, const_m1, S2LO #16
1840+	uadd8	syndrome, data1, const_m1
1841+	eors	syndrome, tmp1, data2, S2LO #16
1842+	sel	syndrome, syndrome, const_m1
1843+	bne	4f
1844+	cbnz	syndrome, 5f
1845+	ldr	data2, [src2], #4
1846+	eor	tmp1, tmp1, data1
1847+	cmp	tmp1, data2, S2HI #16
1848+	bne	6f
1849+	ldr	data1, [src1], #4
1850+	b	.Loverlap2
1851+4:
1852+	S2LO	data2, data2, #16
1853+	b	.Lstrcmp_tail
1854+5:
1855+	ands	syndrome, syndrome, const_m1, S2LO #16
1856+	bne	.Lstrcmp_done_equal
1857+
1858+	ldrh	data2, [src2]
1859+	S2LO	data1, data1, #16
1860+#ifdef __ARM_BIG_ENDIAN
1861+	lsl	data2, data2, #16
1862+#endif
1863+	b	.Lstrcmp_tail
1864+
1865+6:
1866+	S2LO	data1, data1, #16
1867+	and	data2, data2, const_m1, S2LO #16
1868+	b	.Lstrcmp_tail
1869+
1870+	.p2align 5,,12	/* Ensure at least 3 instructions in cache line.  */
1871+.Loverlap1:
1872+	and	tmp1, data1, #LSB
1873+	uadd8	syndrome, data1, const_m1
1874+	eors	syndrome, tmp1, data2, S2LO #24
1875+	sel	syndrome, syndrome, const_m1
1876+	bne	4f
1877+	cbnz	syndrome, 5f
1878+	ldr	data2, [src2], #4
1879+	eor	tmp1, tmp1, data1
1880+	cmp	tmp1, data2, S2HI #8
1881+	bne	6f
1882+	ldr	data1, [src1], #4
1883+	b	.Loverlap1
1884+4:
1885+	S2LO	data2, data2, #24
1886+	b	.Lstrcmp_tail
1887+5:
1888+	tst	syndrome, #LSB
1889+	bne	.Lstrcmp_done_equal
1890+	ldr	data2, [src2]
1891+6:
1892+	S2LO	data1, data1, #8
1893+	bic	data2, data2, #MSB
1894+	b	.Lstrcmp_tail
1895+
1896+.Lstrcmp_done_equal:
1897+	mov	result, #0
1898+	ldrd	r4, r5, [sp], #16
1899+	cfi_remember_state
1900+	cfi_def_cfa_offset (0)
1901+	cfi_restore (r4)
1902+	cfi_restore (r5)
1903+	/* R6/7 not used in this sequence.  */
1904+	cfi_restore (r6)
1905+	cfi_restore (r7)
1906+	bx	lr
1907+
1908+.Lstrcmp_tail:
1909+	cfi_restore_state
1910+#ifndef __ARM_BIG_ENDIAN
1911+	rev	data1, data1
1912+	rev	data2, data2
1913+	/* Now everything looks big-endian...  */
1914+#endif
1915+	uadd8	tmp1, data1, const_m1
1916+	eor	tmp1, data1, data2
1917+	sel	syndrome, tmp1, const_m1
1918+	clz	tmp1, syndrome
1919+	lsl	data1, data1, tmp1
1920+	lsl	data2, data2, tmp1
1921+	lsr	result, data1, #24
1922+	ldrd	r4, r5, [sp], #16
1923+	cfi_def_cfa_offset (0)
1924+	cfi_restore (r4)
1925+	cfi_restore (r5)
1926+	/* R6/7 not used in this sequence.  */
1927+	cfi_restore (r6)
1928+	cfi_restore (r7)
1929+	sub	result, result, data2, lsr #24
1930+	bx	lr
1931+END (strcmp)
1932+libc_hidden_builtin_def (strcmp)
1933+
1934+# From ../strcmp.S
1935+#include <features.h>
1936+
1937+#ifndef __UCLIBC_HAS_LOCALE__
1938+strong_alias(strcmp,strcoll)
1939+libc_hidden_def(strcoll)
1940+#endif
1941diff --git a/libc/string/arm/glibc-neon/strlen.S b/libc/string/arm/glibc-neon/strlen.S
1942new file mode 100644
1943index 0000000..dbb6344
1944--- /dev/null
1945+++ b/libc/string/arm/glibc-neon/strlen.S
1946@@ -0,0 +1,76 @@
1947+/* Copyright (C) 1998-2021 Free Software Foundation, Inc.
1948+   This file is part of the GNU C Library.
1949+
1950+   The GNU C Library is free software; you can redistribute it and/or
1951+   modify it under the terms of the GNU Lesser General Public
1952+   License as published by the Free Software Foundation; either
1953+   version 2.1 of the License, or (at your option) any later version.
1954+
1955+   The GNU C Library is distributed in the hope that it will be useful,
1956+   but WITHOUT ANY WARRANTY; without even the implied warranty of
1957+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
1958+   Lesser General Public License for more details.
1959+
1960+   You should have received a copy of the GNU Lesser General Public
1961+   License along with the GNU C Library.  If not, see
1962+   <https://www.gnu.org/licenses/>.  */
1963+
1964+/* Thumb requires excessive IT insns here.  */
1965+#define NO_THUMB
1966+#include <sysdep.h>
1967+
1968+/* size_t strlen(const char *S)
1969+ * entry: r0 -> string
1970+ * exit: r0 = len
1971+ */
1972+
1973+	.syntax unified
1974+	.text
1975+
1976+ENTRY(strlen)
1977+	bic     r1, r0, $3              @ addr of word containing first byte
1978+	ldr     r2, [r1], $4            @ get the first word
1979+	ands    r3, r0, $3              @ how many bytes are duff?
1980+	rsb     r0, r3, $0              @ get - that number into counter.
1981+	beq     Laligned                @ skip into main check routine if no
1982+					@ more
1983+#ifdef __ARMEB__
1984+	orr     r2, r2, $0xff000000     @ set this byte to non-zero
1985+	subs    r3, r3, $1              @ any more to do?
1986+	orrgt   r2, r2, $0x00ff0000     @ if so, set this byte
1987+	subs    r3, r3, $1              @ more?
1988+	orrgt   r2, r2, $0x0000ff00     @ then set.
1989+#else
1990+	orr     r2, r2, $0x000000ff     @ set this byte to non-zero
1991+	subs    r3, r3, $1              @ any more to do?
1992+	orrgt   r2, r2, $0x0000ff00     @ if so, set this byte
1993+	subs    r3, r3, $1              @ more?
1994+	orrgt   r2, r2, $0x00ff0000     @ then set.
1995+#endif
1996+Laligned:				@ here, we have a word in r2.  Does it
1997+	tst     r2, $0x000000ff         @ contain any zeroes?
1998+	tstne   r2, $0x0000ff00         @
1999+	tstne   r2, $0x00ff0000         @
2000+	tstne   r2, $0xff000000         @
2001+	addne   r0, r0, $4              @ if not, the string is 4 bytes longer
2002+	ldrne   r2, [r1], $4            @ and we continue to the next word
2003+	bne     Laligned                @
2004+Llastword:				@ drop through to here once we find a
2005+#ifdef __ARMEB__
2006+	tst     r2, $0xff000000         @ word that has a zero byte in it
2007+	addne   r0, r0, $1              @
2008+	tstne   r2, $0x00ff0000         @ and add up to 3 bytes on to it
2009+	addne   r0, r0, $1              @
2010+	tstne   r2, $0x0000ff00         @ (if first three all non-zero, 4th
2011+	addne   r0, r0, $1              @  must be zero)
2012+#else
2013+	tst     r2, $0x000000ff         @ word that has a zero byte in it
2014+	addne   r0, r0, $1              @
2015+	tstne   r2, $0x0000ff00         @ and add up to 3 bytes on to it
2016+	addne   r0, r0, $1              @
2017+	tstne   r2, $0x00ff0000         @ (if first three all non-zero, 4th
2018+	addne   r0, r0, $1              @  must be zero)
2019+#endif
2020+	DO_RET(lr)
2021+END(strlen)
2022+libc_hidden_builtin_def (strlen)
2023diff --git a/libc/string/arm/glibc-neon/sysdep.h b/libc/string/arm/glibc-neon/sysdep.h
2024new file mode 100644
2025index 0000000..cceb4a9
2026--- /dev/null
2027+++ b/libc/string/arm/glibc-neon/sysdep.h
2028@@ -0,0 +1,339 @@
2029+/* Assembler macros for ARM.
2030+   Copyright (C) 1997-2021 Free Software Foundation, Inc.
2031+   This file is part of the GNU C Library.
2032+
2033+   The GNU C Library is free software; you can redistribute it and/or
2034+   modify it under the terms of the GNU Lesser General Public
2035+   License as published by the Free Software Foundation; either
2036+   version 2.1 of the License, or (at your option) any later version.
2037+
2038+   The GNU C Library is distributed in the hope that it will be useful,
2039+   but WITHOUT ANY WARRANTY; without even the implied warranty of
2040+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
2041+   Lesser General Public License for more details.
2042+
2043+   You should have received a copy of the GNU Lesser General Public
2044+   License along with the GNU C Library.  If not, see
2045+   <https://www.gnu.org/licenses/>.  */
2046+
2047+#include <sysdeps/generic/sysdep.h>
2048+#include <features.h>
2049+
2050+#ifndef __ASSEMBLER__
2051+# include <stdint.h>
2052+#else
2053+# include <arm-features.h>
2054+#endif
2055+
2056+/* The __ARM_ARCH define is provided by gcc 4.8.  Construct it otherwise.  */
2057+#ifndef __ARM_ARCH
2058+# ifdef __ARM_ARCH_2__
2059+#  define __ARM_ARCH 2
2060+# elif defined (__ARM_ARCH_3__) || defined (__ARM_ARCH_3M__)
2061+#  define __ARM_ARCH 3
2062+# elif defined (__ARM_ARCH_4__) || defined (__ARM_ARCH_4T__)
2063+#  define __ARM_ARCH 4
2064+# elif defined (__ARM_ARCH_5__) || defined (__ARM_ARCH_5E__) \
2065+       || defined(__ARM_ARCH_5T__) || defined(__ARM_ARCH_5TE__) \
2066+       || defined(__ARM_ARCH_5TEJ__)
2067+#  define __ARM_ARCH 5
2068+# elif defined (__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \
2069+       || defined (__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) \
2070+       || defined (__ARM_ARCH_6K__) || defined(__ARM_ARCH_6T2__)
2071+#  define __ARM_ARCH 6
2072+# elif defined (__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \
2073+       || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) \
2074+       || defined(__ARM_ARCH_7EM__)
2075+#  define __ARM_ARCH 7
2076+# else
2077+#  error unknown arm architecture
2078+# endif
2079+#endif
2080+
2081+#if __ARM_ARCH > 4 || defined (__ARM_ARCH_4T__)
2082+# define ARCH_HAS_BX
2083+#endif
2084+#if __ARM_ARCH > 4
2085+# define ARCH_HAS_BLX
2086+#endif
2087+#if __ARM_ARCH > 6 || defined (__ARM_ARCH_6K__) || defined (__ARM_ARCH_6ZK__)
2088+# define ARCH_HAS_HARD_TP
2089+#endif
2090+#if __ARM_ARCH > 6 || defined (__ARM_ARCH_6T2__)
2091+# define ARCH_HAS_T2
2092+#endif
2093+
2094+#ifdef	__ASSEMBLER__
2095+
2096+/* Syntactic details of assembler.  */
2097+
2098+#define ALIGNARG(log2) log2
2099+#define ASM_SIZE_DIRECTIVE(name) .size name,.-name
2100+
2101+#define PLTJMP(_x)	_x##(PLT)
2102+
2103+#ifdef ARCH_HAS_BX
2104+# define BX(R)		bx	R
2105+# define BXC(C, R)	bx##C	R
2106+# ifdef ARCH_HAS_BLX
2107+#  define BLX(R)	blx	R
2108+# else
2109+#  define BLX(R)	mov	lr, pc; bx R
2110+# endif
2111+#else
2112+# define BX(R)		mov	pc, R
2113+# define BXC(C, R)	mov##C	pc, R
2114+# define BLX(R)		mov	lr, pc; mov pc, R
2115+#endif
2116+
2117+#define DO_RET(R)	BX(R)
2118+#define RETINSTR(C, R)	BXC(C, R)
2119+
2120+/* Define an entry point visible from C.  */
2121+#define	ENTRY(name)					\
2122+	.globl	C_SYMBOL_NAME(name);			\
2123+	.type	C_SYMBOL_NAME(name),%function;		\
2124+	.align	ALIGNARG(4);				\
2125+  C_LABEL(name)						\
2126+	CFI_SECTIONS;					\
2127+	cfi_startproc;					\
2128+	CALL_MCOUNT
2129+
2130+#define CFI_SECTIONS					\
2131+	.cfi_sections .debug_frame
2132+
2133+#undef	END
2134+#define END(name)					\
2135+	cfi_endproc;					\
2136+	ASM_SIZE_DIRECTIVE(name)
2137+
2138+/* If compiled for profiling, call `mcount' at the start of each function.  */
2139+#ifdef	PROF
2140+/* Call __gnu_mcount_nc (GCC >= 4.4).  */
2141+#define CALL_MCOUNT					\
2142+	push	{lr};					\
2143+	cfi_adjust_cfa_offset (4);			\
2144+	cfi_rel_offset (lr, 0);				\
2145+	bl	PLTJMP(mcount);				\
2146+	cfi_adjust_cfa_offset (-4);			\
2147+	cfi_restore (lr)
2148+#else
2149+#define CALL_MCOUNT		/* Do nothing.  */
2150+#endif
2151+
2152+/* Since C identifiers are not normally prefixed with an underscore
2153+   on this system, the asm identifier `syscall_error' intrudes on the
2154+   C name space.  Make sure we use an innocuous name.  */
2155+#define	syscall_error	__syscall_error
2156+#define mcount		__gnu_mcount_nc
2157+
2158+/* Tag_ABI_align8_preserved: This code preserves 8-byte
2159+   alignment in any callee.  */
2160+	.eabi_attribute 25, 1
2161+/* Tag_ABI_align8_needed: This code may require 8-byte alignment from
2162+   the caller.  */
2163+	.eabi_attribute 24, 1
2164+
2165+/* The thumb2 encoding is reasonably complete.  Unless suppressed, use it.  */
2166+	.syntax unified
2167+# if defined(__thumb2__) && !defined(NO_THUMB)
2168+	.thumb
2169+#else
2170+#  undef __thumb__
2171+#  undef __thumb2__
2172+	.arm
2173+# endif
2174+
2175+/* Load or store to/from address X + Y into/from R, (maybe) using T.
2176+   X or Y can use T freely; T can be R if OP is a load.  The first
2177+   version eschews the two-register addressing mode, while the
2178+   second version uses it.  */
2179+# define LDST_INDEXED_NOINDEX(OP, R, T, X, Y)		\
2180+	add	T, X, Y;				\
2181+	OP	R, [T]
2182+# define LDST_INDEXED_INDEX(OP, R, X, Y)		\
2183+	OP	R, [X, Y]
2184+
2185+# ifdef ARM_NO_INDEX_REGISTER
2186+/* We're never using the two-register addressing mode, so this
2187+   always uses an intermediate add.  */
2188+#  define LDST_INDEXED(OP, R, T, X, Y)	LDST_INDEXED_NOINDEX (OP, R, T, X, Y)
2189+#  define LDST_PC_INDEXED(OP, R, T, X)	LDST_INDEXED_NOINDEX (OP, R, T, pc, X)
2190+# else
2191+/* The two-register addressing mode is OK, except on Thumb with pc.  */
2192+#  define LDST_INDEXED(OP, R, T, X, Y)	LDST_INDEXED_INDEX (OP, R, X, Y)
2193+#  ifdef __thumb2__
2194+#   define LDST_PC_INDEXED(OP, R, T, X)	LDST_INDEXED_NOINDEX (OP, R, T, pc, X)
2195+#  else
2196+#   define LDST_PC_INDEXED(OP, R, T, X)	LDST_INDEXED_INDEX (OP, R, pc, X)
2197+#  endif
2198+# endif
2199+
2200+/* Load or store to/from a pc-relative EXPR into/from R, using T.  */
2201+# ifdef __thumb2__
2202+#  define LDST_PCREL(OP, R, T, EXPR) \
2203+	ldr	T, 98f;					\
2204+	.subsection 2;					\
2205+98:	.word	EXPR - 99f - PC_OFS;			\
2206+	.previous;					\
2207+99:	add	T, T, pc;				\
2208+	OP	R, [T]
2209+# elif defined (ARCH_HAS_T2) && ARM_PCREL_MOVW_OK
2210+#  define LDST_PCREL(OP, R, T, EXPR)			\
2211+	movw	T, #:lower16:EXPR - 99f - PC_OFS;	\
2212+	movt	T, #:upper16:EXPR - 99f - PC_OFS;	\
2213+99:	LDST_PC_INDEXED (OP, R, T, T)
2214+# else
2215+#  define LDST_PCREL(OP, R, T, EXPR) \
2216+	ldr	T, 98f;					\
2217+	.subsection 2;					\
2218+98:	.word	EXPR - 99f - PC_OFS;			\
2219+	.previous;					\
2220+99:	OP	R, [pc, T]
2221+# endif
2222+
2223+/* Load from a global SYMBOL + CONSTANT into R, using T.  */
2224+# if defined (ARCH_HAS_T2) && !defined (PIC)
2225+#  define LDR_GLOBAL(R, T, SYMBOL, CONSTANT)				\
2226+	movw	T, #:lower16:SYMBOL;					\
2227+	movt	T, #:upper16:SYMBOL;					\
2228+	ldr	R, [T, $CONSTANT]
2229+# elif defined (ARCH_HAS_T2) && defined (PIC) && ARM_PCREL_MOVW_OK
2230+#  define LDR_GLOBAL(R, T, SYMBOL, CONSTANT)				\
2231+	movw	R, #:lower16:_GLOBAL_OFFSET_TABLE_ - 97f - PC_OFS;	\
2232+	movw	T, #:lower16:99f - 98f - PC_OFS;			\
2233+	movt	R, #:upper16:_GLOBAL_OFFSET_TABLE_ - 97f - PC_OFS;	\
2234+	movt	T, #:upper16:99f - 98f - PC_OFS;			\
2235+	.pushsection .rodata.cst4, "aM", %progbits, 4;			\
2236+	.balign 4;							\
2237+99:	.word	SYMBOL##(GOT);						\
2238+	.popsection;							\
2239+97:	add	R, R, pc;						\
2240+98:	LDST_PC_INDEXED (ldr, T, T, T);					\
2241+	LDST_INDEXED (ldr, R, T, R, T);					\
2242+	ldr	R, [R, $CONSTANT]
2243+# else
2244+#  define LDR_GLOBAL(R, T, SYMBOL, CONSTANT)		\
2245+	ldr	T, 99f;					\
2246+	ldr	R, 100f;				\
2247+98:	add	T, T, pc;				\
2248+	ldr	T, [T, R];				\
2249+	.subsection 2;					\
2250+99:	.word	_GLOBAL_OFFSET_TABLE_ - 98b - PC_OFS;	\
2251+100:	.word	SYMBOL##(GOT);				\
2252+	.previous;					\
2253+	ldr	R, [T, $CONSTANT]
2254+# endif
2255+
2256+/* This is the same as LDR_GLOBAL, but for a SYMBOL that is known to
2257+   be in the same linked object (as for one with hidden visibility).
2258+   We can avoid the GOT indirection in the PIC case.  For the pure
2259+   static case, LDR_GLOBAL is already optimal.  */
2260+# ifdef PIC
2261+#  define LDR_HIDDEN(R, T, SYMBOL, CONSTANT) \
2262+  LDST_PCREL (ldr, R, T, SYMBOL + CONSTANT)
2263+# else
2264+#  define LDR_HIDDEN(R, T, SYMBOL, CONSTANT) \
2265+  LDR_GLOBAL (R, T, SYMBOL, CONSTANT)
2266+# endif
2267+
2268+/* Cope with negative memory offsets, which thumb can't encode.
2269+   Use NEGOFF_ADJ_BASE to (conditionally) alter the base register,
2270+   and then NEGOFF_OFF1 to use 0 for thumb and the offset for arm,
2271+   or NEGOFF_OFF2 to use A-B for thumb and A for arm.  */
2272+# ifdef __thumb2__
2273+#  define NEGOFF_ADJ_BASE(R, OFF)	add R, R, $OFF
2274+#  define NEGOFF_ADJ_BASE2(D, S, OFF)	add D, S, $OFF
2275+#  define NEGOFF_OFF1(R, OFF)		[R]
2276+#  define NEGOFF_OFF2(R, OFFA, OFFB)	[R, $((OFFA) - (OFFB))]
2277+# else
2278+#  define NEGOFF_ADJ_BASE(R, OFF)
2279+#  define NEGOFF_ADJ_BASE2(D, S, OFF)	mov D, S
2280+#  define NEGOFF_OFF1(R, OFF)		[R, $OFF]
2281+#  define NEGOFF_OFF2(R, OFFA, OFFB)	[R, $OFFA]
2282+# endif
2283+
2284+/* Helper to get the TLS base pointer.  The interface is that TMP is a
2285+   register that may be used to hold the LR, if necessary.  TMP may be
2286+   LR itself to indicate that LR need not be saved.  The base pointer
2287+   is returned in R0.  Only R0 and TMP are modified.  */
2288+
2289+# ifdef ARCH_HAS_HARD_TP
2290+/* If the cpu has cp15 available, use it.  */
2291+#  define GET_TLS(TMP)		mrc p15, 0, r0, c13, c0, 3
2292+# else
2293+/* At this generic level we have no tricks to pull.  Call the ABI routine.  */
2294+#  define GET_TLS(TMP)					\
2295+	push	{ r1, r2, r3, lr };			\
2296+	cfi_remember_state;				\
2297+	cfi_adjust_cfa_offset (16);			\
2298+	cfi_rel_offset (r1, 0);				\
2299+	cfi_rel_offset (r2, 4);				\
2300+	cfi_rel_offset (r3, 8);				\
2301+	cfi_rel_offset (lr, 12);			\
2302+	bl	__aeabi_read_tp;			\
2303+	pop	{ r1, r2, r3, lr };			\
2304+	cfi_restore_state
2305+# endif /* ARCH_HAS_HARD_TP */
2306+
2307+/* These are the directives used for EABI unwind info.
2308+   Wrap them in macros so another configuration's sysdep.h
2309+   file can define them away if it doesn't use EABI unwind info.  */
2310+# define eabi_fnstart		.fnstart
2311+# define eabi_fnend		.fnend
2312+# define eabi_save(...)		.save __VA_ARGS__
2313+# define eabi_cantunwind	.cantunwind
2314+# define eabi_pad(n)		.pad n
2315+
2316+#endif	/* __ASSEMBLER__ */
2317+
2318+/* This number is the offset from the pc at the current location.  */
2319+#ifdef __thumb__
2320+# define PC_OFS  4
2321+#else
2322+# define PC_OFS  8
2323+#endif
2324+
2325+/* Pointer mangling support.  */
2326+#if (IS_IN (rtld) \
2327+     || (!defined SHARED && (IS_IN (libc) || IS_IN (libpthread))))
2328+# ifdef __ASSEMBLER__
2329+#  define PTR_MANGLE_LOAD(guard, tmp)					\
2330+  LDR_HIDDEN (guard, tmp, C_SYMBOL_NAME(__pointer_chk_guard_local), 0)
2331+#  define PTR_MANGLE(dst, src, guard, tmp)				\
2332+  PTR_MANGLE_LOAD(guard, tmp);						\
2333+  PTR_MANGLE2(dst, src, guard)
2334+/* Use PTR_MANGLE2 for efficiency if guard is already loaded.  */
2335+#  define PTR_MANGLE2(dst, src, guard)		\
2336+  eor dst, src, guard
2337+#  define PTR_DEMANGLE(dst, src, guard, tmp)	\
2338+  PTR_MANGLE (dst, src, guard, tmp)
2339+#  define PTR_DEMANGLE2(dst, src, guard)	\
2340+  PTR_MANGLE2 (dst, src, guard)
2341+# else
2342+extern uintptr_t __pointer_chk_guard_local attribute_relro attribute_hidden;
2343+#  define PTR_MANGLE(var) \
2344+  (var) = (__typeof (var)) ((uintptr_t) (var) ^ __pointer_chk_guard_local)
2345+#  define PTR_DEMANGLE(var)     PTR_MANGLE (var)
2346+# endif
2347+#else
2348+# ifdef __ASSEMBLER__
2349+#  define PTR_MANGLE_LOAD(guard, tmp)					\
2350+  LDR_GLOBAL (guard, tmp, C_SYMBOL_NAME(__pointer_chk_guard), 0);
2351+#  define PTR_MANGLE(dst, src, guard, tmp)				\
2352+  PTR_MANGLE_LOAD(guard, tmp);						\
2353+  PTR_MANGLE2(dst, src, guard)
2354+/* Use PTR_MANGLE2 for efficiency if guard is already loaded.  */
2355+#  define PTR_MANGLE2(dst, src, guard)		\
2356+  eor dst, src, guard
2357+#  define PTR_DEMANGLE(dst, src, guard, tmp)	\
2358+  PTR_MANGLE (dst, src, guard, tmp)
2359+#  define PTR_DEMANGLE2(dst, src, guard)	\
2360+  PTR_MANGLE2 (dst, src, guard)
2361+# else
2362+extern uintptr_t __pointer_chk_guard attribute_relro;
2363+#  define PTR_MANGLE(var) \
2364+  (var) = (__typeof (var)) ((uintptr_t) (var) ^ __pointer_chk_guard)
2365+#  define PTR_DEMANGLE(var)     PTR_MANGLE (var)
2366+# endif
2367+#endif
2368diff --git a/libc/string/arm/glibc-neon/sysdeps/generic/dwarf2.h b/libc/string/arm/glibc-neon/sysdeps/generic/dwarf2.h
2369new file mode 100644
2370index 0000000..0d08da0
2371--- /dev/null
2372+++ b/libc/string/arm/glibc-neon/sysdeps/generic/dwarf2.h
2373@@ -0,0 +1,590 @@
2374+/* Declarations and definitions of codes relating to the DWARF2 symbolic
2375+   debugging information format.
2376+   Copyright (C) 1992-2021 Free Software Foundation, Inc.
2377+   Contributed by Gary Funck (gary@intrepid.com).  Derived from the
2378+   DWARF 1 implementation written by Ron Guilmette (rfg@monkeys.com).
2379+
2380+   This file is part of the GNU C Library.
2381+
2382+   The GNU C Library is free software; you can redistribute it and/or
2383+   modify it under the terms of the GNU Lesser General Public
2384+   License as published by the Free Software Foundation; either
2385+   version 2.1 of the License, or (at your option) any later version.
2386+
2387+   The GNU C Library is distributed in the hope that it will be useful,
2388+   but WITHOUT ANY WARRANTY; without even the implied warranty of
2389+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
2390+   Lesser General Public License for more details.
2391+
2392+   You should have received a copy of the GNU Lesser General Public
2393+   License along with the GNU C Library; if not, see
2394+   <https://www.gnu.org/licenses/>.  */
2395+
2396+#ifndef _DWARF2_H
2397+#define _DWARF2_H	1
2398+
2399+/* This file is derived from the DWARF specification (a public document)
2400+   Revision 2.0.0 (July 27, 1993) developed by the UNIX International
2401+   Programming Languages Special Interest Group (UI/PLSIG) and distributed
2402+   by UNIX International.  Copies of this specification are available from
2403+   UNIX International, 20 Waterview Boulevard, Parsippany, NJ, 07054.  */
2404+
2405+/* This file is shared between GCC and GDB, and should not contain
2406+   prototypes.  */
2407+
2408+#ifndef __ASSEMBLER__
2409+/* Tag names and codes.  */
2410+
2411+enum dwarf_tag
2412+  {
2413+    DW_TAG_padding = 0x00,
2414+    DW_TAG_array_type = 0x01,
2415+    DW_TAG_class_type = 0x02,
2416+    DW_TAG_entry_point = 0x03,
2417+    DW_TAG_enumeration_type = 0x04,
2418+    DW_TAG_formal_parameter = 0x05,
2419+    DW_TAG_imported_declaration = 0x08,
2420+    DW_TAG_label = 0x0a,
2421+    DW_TAG_lexical_block = 0x0b,
2422+    DW_TAG_member = 0x0d,
2423+    DW_TAG_pointer_type = 0x0f,
2424+    DW_TAG_reference_type = 0x10,
2425+    DW_TAG_compile_unit = 0x11,
2426+    DW_TAG_string_type = 0x12,
2427+    DW_TAG_structure_type = 0x13,
2428+    DW_TAG_subroutine_type = 0x15,
2429+    DW_TAG_typedef = 0x16,
2430+    DW_TAG_union_type = 0x17,
2431+    DW_TAG_unspecified_parameters = 0x18,
2432+    DW_TAG_variant = 0x19,
2433+    DW_TAG_common_block = 0x1a,
2434+    DW_TAG_common_inclusion = 0x1b,
2435+    DW_TAG_inheritance = 0x1c,
2436+    DW_TAG_inlined_subroutine = 0x1d,
2437+    DW_TAG_module = 0x1e,
2438+    DW_TAG_ptr_to_member_type = 0x1f,
2439+    DW_TAG_set_type = 0x20,
2440+    DW_TAG_subrange_type = 0x21,
2441+    DW_TAG_with_stmt = 0x22,
2442+    DW_TAG_access_declaration = 0x23,
2443+    DW_TAG_base_type = 0x24,
2444+    DW_TAG_catch_block = 0x25,
2445+    DW_TAG_const_type = 0x26,
2446+    DW_TAG_constant = 0x27,
2447+    DW_TAG_enumerator = 0x28,
2448+    DW_TAG_file_type = 0x29,
2449+    DW_TAG_friend = 0x2a,
2450+    DW_TAG_namelist = 0x2b,
2451+    DW_TAG_namelist_item = 0x2c,
2452+    DW_TAG_packed_type = 0x2d,
2453+    DW_TAG_subprogram = 0x2e,
2454+    DW_TAG_template_type_param = 0x2f,
2455+    DW_TAG_template_value_param = 0x30,
2456+    DW_TAG_thrown_type = 0x31,
2457+    DW_TAG_try_block = 0x32,
2458+    DW_TAG_variant_part = 0x33,
2459+    DW_TAG_variable = 0x34,
2460+    DW_TAG_volatile_type = 0x35,
2461+    /* SGI/MIPS Extensions */
2462+    DW_TAG_MIPS_loop = 0x4081,
2463+    /* GNU extensions */
2464+    DW_TAG_format_label = 0x4101,	/* for FORTRAN 77 and Fortran 90 */
2465+    DW_TAG_function_template = 0x4102,	/* for C++ */
2466+    DW_TAG_class_template = 0x4103,	/* for C++ */
2467+    DW_TAG_GNU_BINCL = 0x4104,
2468+    DW_TAG_GNU_EINCL = 0x4105
2469+  };
2470+
2471+#define DW_TAG_lo_user	0x4080
2472+#define DW_TAG_hi_user	0xffff
2473+
2474+/* flag that tells whether entry has a child or not */
2475+#define DW_children_no   0
2476+#define	DW_children_yes  1
2477+
2478+/* Form names and codes.  */
2479+enum dwarf_form
2480+  {
2481+    DW_FORM_addr = 0x01,
2482+    DW_FORM_block2 = 0x03,
2483+    DW_FORM_block4 = 0x04,
2484+    DW_FORM_data2 = 0x05,
2485+    DW_FORM_data4 = 0x06,
2486+    DW_FORM_data8 = 0x07,
2487+    DW_FORM_string = 0x08,
2488+    DW_FORM_block = 0x09,
2489+    DW_FORM_block1 = 0x0a,
2490+    DW_FORM_data1 = 0x0b,
2491+    DW_FORM_flag = 0x0c,
2492+    DW_FORM_sdata = 0x0d,
2493+    DW_FORM_strp = 0x0e,
2494+    DW_FORM_udata = 0x0f,
2495+    DW_FORM_ref_addr = 0x10,
2496+    DW_FORM_ref1 = 0x11,
2497+    DW_FORM_ref2 = 0x12,
2498+    DW_FORM_ref4 = 0x13,
2499+    DW_FORM_ref8 = 0x14,
2500+    DW_FORM_ref_udata = 0x15,
2501+    DW_FORM_indirect = 0x16
2502+  };
2503+
2504+/* Attribute names and codes.  */
2505+
2506+enum dwarf_attribute
2507+  {
2508+    DW_AT_sibling = 0x01,
2509+    DW_AT_location = 0x02,
2510+    DW_AT_name = 0x03,
2511+    DW_AT_ordering = 0x09,
2512+    DW_AT_subscr_data = 0x0a,
2513+    DW_AT_byte_size = 0x0b,
2514+    DW_AT_bit_offset = 0x0c,
2515+    DW_AT_bit_size = 0x0d,
2516+    DW_AT_element_list = 0x0f,
2517+    DW_AT_stmt_list = 0x10,
2518+    DW_AT_low_pc = 0x11,
2519+    DW_AT_high_pc = 0x12,
2520+    DW_AT_language = 0x13,
2521+    DW_AT_member = 0x14,
2522+    DW_AT_discr = 0x15,
2523+    DW_AT_discr_value = 0x16,
2524+    DW_AT_visibility = 0x17,
2525+    DW_AT_import = 0x18,
2526+    DW_AT_string_length = 0x19,
2527+    DW_AT_common_reference = 0x1a,
2528+    DW_AT_comp_dir = 0x1b,
2529+    DW_AT_const_value = 0x1c,
2530+    DW_AT_containing_type = 0x1d,
2531+    DW_AT_default_value = 0x1e,
2532+    DW_AT_inline = 0x20,
2533+    DW_AT_is_optional = 0x21,
2534+    DW_AT_lower_bound = 0x22,
2535+    DW_AT_producer = 0x25,
2536+    DW_AT_prototyped = 0x27,
2537+    DW_AT_return_addr = 0x2a,
2538+    DW_AT_start_scope = 0x2c,
2539+    DW_AT_stride_size = 0x2e,
2540+    DW_AT_upper_bound = 0x2f,
2541+    DW_AT_abstract_origin = 0x31,
2542+    DW_AT_accessibility = 0x32,
2543+    DW_AT_address_class = 0x33,
2544+    DW_AT_artificial = 0x34,
2545+    DW_AT_base_types = 0x35,
2546+    DW_AT_calling_convention = 0x36,
2547+    DW_AT_count = 0x37,
2548+    DW_AT_data_member_location = 0x38,
2549+    DW_AT_decl_column = 0x39,
2550+    DW_AT_decl_file = 0x3a,
2551+    DW_AT_decl_line = 0x3b,
2552+    DW_AT_declaration = 0x3c,
2553+    DW_AT_discr_list = 0x3d,
2554+    DW_AT_encoding = 0x3e,
2555+    DW_AT_external = 0x3f,
2556+    DW_AT_frame_base = 0x40,
2557+    DW_AT_friend = 0x41,
2558+    DW_AT_identifier_case = 0x42,
2559+    DW_AT_macro_info = 0x43,
2560+    DW_AT_namelist_items = 0x44,
2561+    DW_AT_priority = 0x45,
2562+    DW_AT_segment = 0x46,
2563+    DW_AT_specification = 0x47,
2564+    DW_AT_static_link = 0x48,
2565+    DW_AT_type = 0x49,
2566+    DW_AT_use_location = 0x4a,
2567+    DW_AT_variable_parameter = 0x4b,
2568+    DW_AT_virtuality = 0x4c,
2569+    DW_AT_vtable_elem_location = 0x4d,
2570+    /* SGI/MIPS Extensions */
2571+    DW_AT_MIPS_fde = 0x2001,
2572+    DW_AT_MIPS_loop_begin = 0x2002,
2573+    DW_AT_MIPS_tail_loop_begin = 0x2003,
2574+    DW_AT_MIPS_epilog_begin = 0x2004,
2575+    DW_AT_MIPS_loop_unroll_factor = 0x2005,
2576+    DW_AT_MIPS_software_pipeline_depth = 0x2006,
2577+    DW_AT_MIPS_linkage_name = 0x2007,
2578+    DW_AT_MIPS_stride = 0x2008,
2579+    DW_AT_MIPS_abstract_name = 0x2009,
2580+    DW_AT_MIPS_clone_origin = 0x200a,
2581+    DW_AT_MIPS_has_inlines = 0x200b,
2582+    /* GNU extensions.  */
2583+    DW_AT_sf_names = 0x2101,
2584+    DW_AT_src_info = 0x2102,
2585+    DW_AT_mac_info = 0x2103,
2586+    DW_AT_src_coords = 0x2104,
2587+    DW_AT_body_begin = 0x2105,
2588+    DW_AT_body_end = 0x2106
2589+  };
2590+
2591+#define DW_AT_lo_user	0x2000	/* implementation-defined range start */
2592+#define DW_AT_hi_user	0x3ff0	/* implementation-defined range end */
2593+
2594+/* Location atom names and codes.  */
2595+
2596+enum dwarf_location_atom
2597+  {
2598+    DW_OP_addr = 0x03,
2599+    DW_OP_deref = 0x06,
2600+    DW_OP_const1u = 0x08,
2601+    DW_OP_const1s = 0x09,
2602+    DW_OP_const2u = 0x0a,
2603+    DW_OP_const2s = 0x0b,
2604+    DW_OP_const4u = 0x0c,
2605+    DW_OP_const4s = 0x0d,
2606+    DW_OP_const8u = 0x0e,
2607+    DW_OP_const8s = 0x0f,
2608+    DW_OP_constu = 0x10,
2609+    DW_OP_consts = 0x11,
2610+    DW_OP_dup = 0x12,
2611+    DW_OP_drop = 0x13,
2612+    DW_OP_over = 0x14,
2613+    DW_OP_pick = 0x15,
2614+    DW_OP_swap = 0x16,
2615+    DW_OP_rot = 0x17,
2616+    DW_OP_xderef = 0x18,
2617+    DW_OP_abs = 0x19,
2618+    DW_OP_and = 0x1a,
2619+    DW_OP_div = 0x1b,
2620+    DW_OP_minus = 0x1c,
2621+    DW_OP_mod = 0x1d,
2622+    DW_OP_mul = 0x1e,
2623+    DW_OP_neg = 0x1f,
2624+    DW_OP_not = 0x20,
2625+    DW_OP_or = 0x21,
2626+    DW_OP_plus = 0x22,
2627+    DW_OP_plus_uconst = 0x23,
2628+    DW_OP_shl = 0x24,
2629+    DW_OP_shr = 0x25,
2630+    DW_OP_shra = 0x26,
2631+    DW_OP_xor = 0x27,
2632+    DW_OP_bra = 0x28,
2633+    DW_OP_eq = 0x29,
2634+    DW_OP_ge = 0x2a,
2635+    DW_OP_gt = 0x2b,
2636+    DW_OP_le = 0x2c,
2637+    DW_OP_lt = 0x2d,
2638+    DW_OP_ne = 0x2e,
2639+    DW_OP_skip = 0x2f,
2640+    DW_OP_lit0 = 0x30,
2641+    DW_OP_lit1 = 0x31,
2642+    DW_OP_lit2 = 0x32,
2643+    DW_OP_lit3 = 0x33,
2644+    DW_OP_lit4 = 0x34,
2645+    DW_OP_lit5 = 0x35,
2646+    DW_OP_lit6 = 0x36,
2647+    DW_OP_lit7 = 0x37,
2648+    DW_OP_lit8 = 0x38,
2649+    DW_OP_lit9 = 0x39,
2650+    DW_OP_lit10 = 0x3a,
2651+    DW_OP_lit11 = 0x3b,
2652+    DW_OP_lit12 = 0x3c,
2653+    DW_OP_lit13 = 0x3d,
2654+    DW_OP_lit14 = 0x3e,
2655+    DW_OP_lit15 = 0x3f,
2656+    DW_OP_lit16 = 0x40,
2657+    DW_OP_lit17 = 0x41,
2658+    DW_OP_lit18 = 0x42,
2659+    DW_OP_lit19 = 0x43,
2660+    DW_OP_lit20 = 0x44,
2661+    DW_OP_lit21 = 0x45,
2662+    DW_OP_lit22 = 0x46,
2663+    DW_OP_lit23 = 0x47,
2664+    DW_OP_lit24 = 0x48,
2665+    DW_OP_lit25 = 0x49,
2666+    DW_OP_lit26 = 0x4a,
2667+    DW_OP_lit27 = 0x4b,
2668+    DW_OP_lit28 = 0x4c,
2669+    DW_OP_lit29 = 0x4d,
2670+    DW_OP_lit30 = 0x4e,
2671+    DW_OP_lit31 = 0x4f,
2672+    DW_OP_reg0 = 0x50,
2673+    DW_OP_reg1 = 0x51,
2674+    DW_OP_reg2 = 0x52,
2675+    DW_OP_reg3 = 0x53,
2676+    DW_OP_reg4 = 0x54,
2677+    DW_OP_reg5 = 0x55,
2678+    DW_OP_reg6 = 0x56,
2679+    DW_OP_reg7 = 0x57,
2680+    DW_OP_reg8 = 0x58,
2681+    DW_OP_reg9 = 0x59,
2682+    DW_OP_reg10 = 0x5a,
2683+    DW_OP_reg11 = 0x5b,
2684+    DW_OP_reg12 = 0x5c,
2685+    DW_OP_reg13 = 0x5d,
2686+    DW_OP_reg14 = 0x5e,
2687+    DW_OP_reg15 = 0x5f,
2688+    DW_OP_reg16 = 0x60,
2689+    DW_OP_reg17 = 0x61,
2690+    DW_OP_reg18 = 0x62,
2691+    DW_OP_reg19 = 0x63,
2692+    DW_OP_reg20 = 0x64,
2693+    DW_OP_reg21 = 0x65,
2694+    DW_OP_reg22 = 0x66,
2695+    DW_OP_reg23 = 0x67,
2696+    DW_OP_reg24 = 0x68,
2697+    DW_OP_reg25 = 0x69,
2698+    DW_OP_reg26 = 0x6a,
2699+    DW_OP_reg27 = 0x6b,
2700+    DW_OP_reg28 = 0x6c,
2701+    DW_OP_reg29 = 0x6d,
2702+    DW_OP_reg30 = 0x6e,
2703+    DW_OP_reg31 = 0x6f,
2704+    DW_OP_breg0 = 0x70,
2705+    DW_OP_breg1 = 0x71,
2706+    DW_OP_breg2 = 0x72,
2707+    DW_OP_breg3 = 0x73,
2708+    DW_OP_breg4 = 0x74,
2709+    DW_OP_breg5 = 0x75,
2710+    DW_OP_breg6 = 0x76,
2711+    DW_OP_breg7 = 0x77,
2712+    DW_OP_breg8 = 0x78,
2713+    DW_OP_breg9 = 0x79,
2714+    DW_OP_breg10 = 0x7a,
2715+    DW_OP_breg11 = 0x7b,
2716+    DW_OP_breg12 = 0x7c,
2717+    DW_OP_breg13 = 0x7d,
2718+    DW_OP_breg14 = 0x7e,
2719+    DW_OP_breg15 = 0x7f,
2720+    DW_OP_breg16 = 0x80,
2721+    DW_OP_breg17 = 0x81,
2722+    DW_OP_breg18 = 0x82,
2723+    DW_OP_breg19 = 0x83,
2724+    DW_OP_breg20 = 0x84,
2725+    DW_OP_breg21 = 0x85,
2726+    DW_OP_breg22 = 0x86,
2727+    DW_OP_breg23 = 0x87,
2728+    DW_OP_breg24 = 0x88,
2729+    DW_OP_breg25 = 0x89,
2730+    DW_OP_breg26 = 0x8a,
2731+    DW_OP_breg27 = 0x8b,
2732+    DW_OP_breg28 = 0x8c,
2733+    DW_OP_breg29 = 0x8d,
2734+    DW_OP_breg30 = 0x8e,
2735+    DW_OP_breg31 = 0x8f,
2736+    DW_OP_regx = 0x90,
2737+    DW_OP_fbreg = 0x91,
2738+    DW_OP_bregx = 0x92,
2739+    DW_OP_piece = 0x93,
2740+    DW_OP_deref_size = 0x94,
2741+    DW_OP_xderef_size = 0x95,
2742+    DW_OP_nop = 0x96
2743+  };
2744+
2745+#define DW_OP_lo_user	0x80	/* implementation-defined range start */
2746+#define DW_OP_hi_user	0xff	/* implementation-defined range end */
2747+
2748+/* Type encodings.  */
2749+
2750+enum dwarf_type
2751+  {
2752+    DW_ATE_void = 0x0,
2753+    DW_ATE_address = 0x1,
2754+    DW_ATE_boolean = 0x2,
2755+    DW_ATE_complex_float = 0x3,
2756+    DW_ATE_float = 0x4,
2757+    DW_ATE_signed = 0x5,
2758+    DW_ATE_signed_char = 0x6,
2759+    DW_ATE_unsigned = 0x7,
2760+    DW_ATE_unsigned_char = 0x8
2761+  };
2762+
2763+#define	DW_ATE_lo_user 0x80
2764+#define	DW_ATE_hi_user 0xff
2765+
2766+/* Array ordering names and codes.  */
2767+enum dwarf_array_dim_ordering
2768+  {
2769+    DW_ORD_row_major = 0,
2770+    DW_ORD_col_major = 1
2771+  };
2772+
2773+/* access attribute */
2774+enum dwarf_access_attribute
2775+  {
2776+    DW_ACCESS_public = 1,
2777+    DW_ACCESS_protected = 2,
2778+    DW_ACCESS_private = 3
2779+  };
2780+
2781+/* visibility */
2782+enum dwarf_visibility_attribute
2783+  {
2784+    DW_VIS_local = 1,
2785+    DW_VIS_exported = 2,
2786+    DW_VIS_qualified = 3
2787+  };
2788+
2789+/* virtuality */
2790+enum dwarf_virtuality_attribute
2791+  {
2792+    DW_VIRTUALITY_none = 0,
2793+    DW_VIRTUALITY_virtual = 1,
2794+    DW_VIRTUALITY_pure_virtual = 2
2795+  };
2796+
2797+/* case sensitivity */
2798+enum dwarf_id_case
2799+  {
2800+    DW_ID_case_sensitive = 0,
2801+    DW_ID_up_case = 1,
2802+    DW_ID_down_case = 2,
2803+    DW_ID_case_insensitive = 3
2804+  };
2805+
2806+/* calling convention */
2807+enum dwarf_calling_convention
2808+  {
2809+    DW_CC_normal = 0x1,
2810+    DW_CC_program = 0x2,
2811+    DW_CC_nocall = 0x3
2812+  };
2813+
2814+#define DW_CC_lo_user 0x40
2815+#define DW_CC_hi_user 0xff
2816+
2817+/* inline attribute */
2818+enum dwarf_inline_attribute
2819+  {
2820+    DW_INL_not_inlined = 0,
2821+    DW_INL_inlined = 1,
2822+    DW_INL_declared_not_inlined = 2,
2823+    DW_INL_declared_inlined = 3
2824+  };
2825+
2826+/* discriminant lists */
2827+enum dwarf_discrim_list
2828+  {
2829+    DW_DSC_label = 0,
2830+    DW_DSC_range = 1
2831+  };
2832+
2833+/* line number opcodes */
2834+enum dwarf_line_number_ops
2835+  {
2836+    DW_LNS_extended_op = 0,
2837+    DW_LNS_copy = 1,
2838+    DW_LNS_advance_pc = 2,
2839+    DW_LNS_advance_line = 3,
2840+    DW_LNS_set_file = 4,
2841+    DW_LNS_set_column = 5,
2842+    DW_LNS_negate_stmt = 6,
2843+    DW_LNS_set_basic_block = 7,
2844+    DW_LNS_const_add_pc = 8,
2845+    DW_LNS_fixed_advance_pc = 9
2846+  };
2847+
2848+/* line number extended opcodes */
2849+enum dwarf_line_number_x_ops
2850+  {
2851+    DW_LNE_end_sequence = 1,
2852+    DW_LNE_set_address = 2,
2853+    DW_LNE_define_file = 3
2854+  };
2855+
2856+/* call frame information */
2857+enum dwarf_call_frame_info
2858+  {
2859+    DW_CFA_advance_loc = 0x40,
2860+    DW_CFA_offset = 0x80,
2861+    DW_CFA_restore = 0xc0,
2862+    DW_CFA_nop = 0x00,
2863+    DW_CFA_set_loc = 0x01,
2864+    DW_CFA_advance_loc1 = 0x02,
2865+    DW_CFA_advance_loc2 = 0x03,
2866+    DW_CFA_advance_loc4 = 0x04,
2867+    DW_CFA_offset_extended = 0x05,
2868+    DW_CFA_restore_extended = 0x06,
2869+    DW_CFA_undefined = 0x07,
2870+    DW_CFA_same_value = 0x08,
2871+    DW_CFA_register = 0x09,
2872+    DW_CFA_remember_state = 0x0a,
2873+    DW_CFA_restore_state = 0x0b,
2874+    DW_CFA_def_cfa = 0x0c,
2875+    DW_CFA_def_cfa_register = 0x0d,
2876+    DW_CFA_def_cfa_offset = 0x0e,
2877+    DW_CFA_def_cfa_expression = 0x0f,
2878+    DW_CFA_expression = 0x10,
2879+    /* Dwarf 2.1 */
2880+    DW_CFA_offset_extended_sf = 0x11,
2881+    DW_CFA_def_cfa_sf = 0x12,
2882+    DW_CFA_def_cfa_offset_sf = 0x13,
2883+
2884+    /* SGI/MIPS specific */
2885+    DW_CFA_MIPS_advance_loc8 = 0x1d,
2886+
2887+    /* GNU extensions */
2888+    DW_CFA_GNU_window_save = 0x2d,
2889+    DW_CFA_GNU_args_size = 0x2e,
2890+    DW_CFA_GNU_negative_offset_extended = 0x2f
2891+  };
2892+
2893+#define DW_CIE_ID	  0xffffffff
2894+#define DW_CIE_VERSION	  1
2895+
2896+#define DW_CFA_extended   0
2897+#define DW_CFA_low_user   0x1c
2898+#define DW_CFA_high_user  0x3f
2899+
2900+#define DW_CHILDREN_no		     0x00
2901+#define DW_CHILDREN_yes		     0x01
2902+
2903+#define DW_ADDR_none		0
2904+
2905+/* Source language names and codes.  */
2906+
2907+enum dwarf_source_language
2908+  {
2909+    DW_LANG_C89 = 0x0001,
2910+    DW_LANG_C = 0x0002,
2911+    DW_LANG_Ada83 = 0x0003,
2912+    DW_LANG_C_plus_plus = 0x0004,
2913+    DW_LANG_Cobol74 = 0x0005,
2914+    DW_LANG_Cobol85 = 0x0006,
2915+    DW_LANG_Fortran77 = 0x0007,
2916+    DW_LANG_Fortran90 = 0x0008,
2917+    DW_LANG_Pascal83 = 0x0009,
2918+    DW_LANG_Modula2 = 0x000a,
2919+    DW_LANG_Java = 0x000b,
2920+    DW_LANG_Mips_Assembler = 0x8001
2921+  };
2922+
2923+
2924+#define DW_LANG_lo_user 0x8000	/* implementation-defined range start */
2925+#define DW_LANG_hi_user 0xffff	/* implementation-defined range start */
2926+
2927+/* Names and codes for macro information.  */
2928+
2929+enum dwarf_macinfo_record_type
2930+  {
2931+    DW_MACINFO_define = 1,
2932+    DW_MACINFO_undef = 2,
2933+    DW_MACINFO_start_file = 3,
2934+    DW_MACINFO_end_file = 4,
2935+    DW_MACINFO_vendor_ext = 255
2936+  };
2937+
2938+#endif /* !ASSEMBLER */
2939+
2940+/* @@@ For use with GNU frame unwind information.  */
2941+
2942+#define DW_EH_PE_absptr		0x00
2943+#define DW_EH_PE_omit		0xff
2944+
2945+#define DW_EH_PE_uleb128	0x01
2946+#define DW_EH_PE_udata2		0x02
2947+#define DW_EH_PE_udata4		0x03
2948+#define DW_EH_PE_udata8		0x04
2949+#define DW_EH_PE_sleb128	0x09
2950+#define DW_EH_PE_sdata2		0x0A
2951+#define DW_EH_PE_sdata4		0x0B
2952+#define DW_EH_PE_sdata8		0x0C
2953+#define DW_EH_PE_signed		0x08
2954+
2955+#define DW_EH_PE_pcrel		0x10
2956+#define DW_EH_PE_textrel	0x20
2957+#define DW_EH_PE_datarel	0x30
2958+#define DW_EH_PE_funcrel	0x40
2959+#define DW_EH_PE_aligned	0x50
2960+
2961+#define DW_EH_PE_indirect	0x80
2962+
2963+#endif /* dwarf2.h */
2964diff --git a/libc/string/arm/glibc-neon/sysdeps/generic/sysdep.h b/libc/string/arm/glibc-neon/sysdeps/generic/sysdep.h
2965new file mode 100644
2966index 0000000..39cac91
2967--- /dev/null
2968+++ b/libc/string/arm/glibc-neon/sysdeps/generic/sysdep.h
2969@@ -0,0 +1,97 @@
2970+/* Generic asm macros used on many machines.
2971+   Copyright (C) 1991-2021 Free Software Foundation, Inc.
2972+   This file is part of the GNU C Library.
2973+
2974+   The GNU C Library is free software; you can redistribute it and/or
2975+   modify it under the terms of the GNU Lesser General Public
2976+   License as published by the Free Software Foundation; either
2977+   version 2.1 of the License, or (at your option) any later version.
2978+
2979+   The GNU C Library is distributed in the hope that it will be useful,
2980+   but WITHOUT ANY WARRANTY; without even the implied warranty of
2981+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
2982+   Lesser General Public License for more details.
2983+
2984+   You should have received a copy of the GNU Lesser General Public
2985+   License along with the GNU C Library; if not, see
2986+   <https://www.gnu.org/licenses/>.  */
2987+
2988+#ifndef C_LABEL
2989+
2990+/* Define a macro we can use to construct the asm name for a C symbol.  */
2991+# define C_LABEL(name)	name##:
2992+
2993+#endif
2994+
2995+#ifdef __ASSEMBLER__
2996+/* Mark the end of function named SYM.  This is used on some platforms
2997+   to generate correct debugging information.  */
2998+# ifndef END
2999+#  define END(sym)
3000+# endif
3001+
3002+# ifndef JUMPTARGET
3003+#  define JUMPTARGET(sym)	sym
3004+# endif
3005+#endif
3006+
3007+/* Makros to generate eh_frame unwind information.  */
3008+#ifdef __ASSEMBLER__
3009+# define cfi_startproc			.cfi_startproc
3010+# define cfi_endproc			.cfi_endproc
3011+# define cfi_def_cfa(reg, off)		.cfi_def_cfa reg, off
3012+# define cfi_def_cfa_register(reg)	.cfi_def_cfa_register reg
3013+# define cfi_def_cfa_offset(off)	.cfi_def_cfa_offset off
3014+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
3015+# define cfi_offset(reg, off)		.cfi_offset reg, off
3016+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
3017+# define cfi_register(r1, r2)		.cfi_register r1, r2
3018+# define cfi_return_column(reg)	.cfi_return_column reg
3019+# define cfi_restore(reg)		.cfi_restore reg
3020+# define cfi_same_value(reg)		.cfi_same_value reg
3021+# define cfi_undefined(reg)		.cfi_undefined reg
3022+# define cfi_remember_state		.cfi_remember_state
3023+# define cfi_restore_state		.cfi_restore_state
3024+# define cfi_window_save		.cfi_window_save
3025+# define cfi_personality(enc, exp)	.cfi_personality enc, exp
3026+# define cfi_lsda(enc, exp)		.cfi_lsda enc, exp
3027+
3028+#else /* ! ASSEMBLER */
3029+
3030+# define CFI_STRINGIFY(Name) CFI_STRINGIFY2 (Name)
3031+# define CFI_STRINGIFY2(Name) #Name
3032+# define CFI_STARTPROC	".cfi_startproc"
3033+# define CFI_ENDPROC	".cfi_endproc"
3034+# define CFI_DEF_CFA(reg, off)	\
3035+   ".cfi_def_cfa " CFI_STRINGIFY(reg) "," CFI_STRINGIFY(off)
3036+# define CFI_DEF_CFA_REGISTER(reg) \
3037+   ".cfi_def_cfa_register " CFI_STRINGIFY(reg)
3038+# define CFI_DEF_CFA_OFFSET(off) \
3039+   ".cfi_def_cfa_offset " CFI_STRINGIFY(off)
3040+# define CFI_ADJUST_CFA_OFFSET(off) \
3041+   ".cfi_adjust_cfa_offset " CFI_STRINGIFY(off)
3042+# define CFI_OFFSET(reg, off) \
3043+   ".cfi_offset " CFI_STRINGIFY(reg) "," CFI_STRINGIFY(off)
3044+# define CFI_REL_OFFSET(reg, off) \
3045+   ".cfi_rel_offset " CFI_STRINGIFY(reg) "," CFI_STRINGIFY(off)
3046+# define CFI_REGISTER(r1, r2) \
3047+   ".cfi_register " CFI_STRINGIFY(r1) "," CFI_STRINGIFY(r2)
3048+# define CFI_RETURN_COLUMN(reg) \
3049+   ".cfi_return_column " CFI_STRINGIFY(reg)
3050+# define CFI_RESTORE(reg) \
3051+   ".cfi_restore " CFI_STRINGIFY(reg)
3052+# define CFI_UNDEFINED(reg) \
3053+   ".cfi_undefined " CFI_STRINGIFY(reg)
3054+# define CFI_REMEMBER_STATE \
3055+   ".cfi_remember_state"
3056+# define CFI_RESTORE_STATE \
3057+   ".cfi_restore_state"
3058+# define CFI_WINDOW_SAVE \
3059+   ".cfi_window_save"
3060+# define CFI_PERSONALITY(enc, exp) \
3061+   ".cfi_personality " CFI_STRINGIFY(enc) "," CFI_STRINGIFY(exp)
3062+# define CFI_LSDA(enc, exp) \
3063+   ".cfi_lsda " CFI_STRINGIFY(enc) "," CFI_STRINGIFY(exp)
3064+#endif
3065+
3066+#include "dwarf2.h"
3067--
30682.20.1
3069
3070