Don't fail to run DT_INIT and DT_INIT_ARRAY constructors if a shared library has DT_PREINIT_ARRAY constructors.

The GNU dynamic linker silently ignores a DT_PREINIT_ARRAY section in a shared library. We had ineffectual code that tried to report an error, which I tried to fix but got wrong --- my version still wouldn't report the error to the caller, but would prevent us from continuing to call constructors. Bug: 8825226 (cherry picked from commit 8147d3c284) Change-Id: Iad8e7e3743d1a5e6bd1d44ef40b1ab3499cb64f2
Adjust PC value in ARM stack trace.
2013-05-09 15:53:30 -07:00 · 2013-05-08 11:21:53 -07:00 · 2013-04-23 18:22:31 -07:00 · 2013-04-09 09:37:44 -07:00 · 2013-04-05 15:57:53 -07:00 · 2013-04-03 11:14:20 -07:00
893 changed files with 12989 additions and 13780 deletions
--- a/ABI-bugs.txt
+++ b/ABI-bugs.txt
@@ -0,0 +1,15 @@
+KNOWN ABI BUGS
+--------------
+
+  time_t is 32-bit. http://b/5819737
+
+  off_t is 32-bit. There is off64_t, but no _FILE_OFFSET_BITS support.
+
+  sigset_t is too small on ARM and x86 (but correct on MIPS), so support
+  for real-time signals is broken. http://b/5828899
+
+  Too few TLS slots mean we can't allocate 128 pthread_key_t instances,
+  which POSIX says should be the minimum.
+
+  atexit(3) handlers registered by a shared library aren't called on
+  dlclose(3); this only affects ARM. http://b/4998315
--- a/libc/Android.mk
+++ b/libc/Android.mk
@@ -11,7 +11,6 @@ libc_common_src_files := \
 	unistd/exec.c \
 	unistd/fnmatch.c \
 	unistd/getopt_long.c \
-	unistd/popen.c \
 	unistd/syslog.c \
 	unistd/system.c \
 	unistd/time.c \
@@ -111,35 +110,8 @@ libc_common_src_files := \
 	string/strspn.c \
 	string/strstr.c \
 	string/strtok.c \
-	wchar/wcpcpy.c \
-	wchar/wcpncpy.c \
-	wchar/wcscasecmp.c \
-	wchar/wcscat.c \
-	wchar/wcschr.c \
-	wchar/wcscmp.c \
-	wchar/wcscpy.c \
-	wchar/wcscspn.c \
-	wchar/wcsdup.c \
-	wchar/wcslcat.c \
-	wchar/wcslcpy.c \
-	wchar/wcslen.c \
-	wchar/wcsncasecmp.c \
-	wchar/wcsncat.c \
-	wchar/wcsncmp.c \
-	wchar/wcsncpy.c \
-	wchar/wcsnlen.c \
-	wchar/wcspbrk.c \
-	wchar/wcsrchr.c \
-	wchar/wcsspn.c \
-	wchar/wcsstr.c \
-	wchar/wcstok.c \
 	wchar/wcswidth.c \
 	wchar/wcsxfrm.c \
-	wchar/wmemchr.c \
-	wchar/wmemcmp.c \
-	wchar/wmemcpy.c \
-	wchar/wmemmove.c \
-	wchar/wmemset.c \
 	tzcode/asctime.c \
 	tzcode/difftime.c \
 	tzcode/localtime.c \
@@ -177,12 +149,10 @@ libc_common_src_files := \
 	bionic/isatty.c \
 	bionic/issetugid.c \
 	bionic/ldexp.c \
-	bionic/logd_write.c \
 	bionic/lseek64.c \
 	bionic/md5.c \
 	bionic/memchr.c \
 	bionic/memmem.c \
-	bionic/memmove_words.c \
 	bionic/memrchr.c \
 	bionic/memswap.c \
 	bionic/mmap.c \
@@ -196,7 +166,6 @@ libc_common_src_files := \
 	bionic/ptsname_r.c \
 	bionic/pututline.c \
 	bionic/pwrite.c \
-	bionic/realpath.c \
 	bionic/reboot.c \
 	bionic/recv.c \
 	bionic/sched_cpualloc.c \
@@ -234,7 +203,6 @@ libc_common_src_files := \
 	bionic/unlockpt.c \
 	bionic/usleep.c \
 	bionic/utmp.c \
-	bionic/wait.c \
 	bionic/wcscoll.c \
 	netbsd/gethnamaddr.c \
 	netbsd/inet/nsap_addr.c \
@@ -268,20 +236,34 @@ libc_common_src_files := \
 libc_bionic_src_files := \
    bionic/assert.cpp \
    bionic/brk.cpp \
-    bionic/debug_format.cpp \
    bionic/dirent.cpp \
-    bionic/eventfd.cpp \
+    bionic/__errno.c \
+    bionic/eventfd_read.cpp \
+    bionic/eventfd_write.cpp \
    bionic/__fgets_chk.cpp \
    bionic/getauxval.cpp \
    bionic/getcwd.cpp \
    bionic/libc_init_common.cpp \
+    bionic/libc_logging.cpp \
    bionic/libgen.cpp \
    bionic/__memcpy_chk.cpp \
    bionic/__memmove_chk.cpp \
    bionic/__memset_chk.cpp \
+    bionic/pthread_attr.cpp \
+    bionic/pthread_detach.cpp \
+    bionic/pthread_equal.cpp \
+    bionic/pthread_getcpuclockid.cpp \
+    bionic/pthread_getschedparam.cpp \
+    bionic/pthread_internals.cpp \
+    bionic/pthread_join.cpp \
+    bionic/pthread_kill.cpp \
+    bionic/pthread_self.cpp \
+    bionic/pthread_setname_np.cpp \
+    bionic/pthread_setschedparam.cpp \
    bionic/pthread_sigmask.cpp \
    bionic/raise.cpp \
    bionic/sbrk.cpp \
+    bionic/scandir.cpp \
    bionic/__set_errno.cpp \
    bionic/setlocale.cpp \
    bionic/signalfd.cpp \
@@ -297,13 +279,45 @@ libc_bionic_src_files := \
    bionic/__strncpy_chk.cpp \
    bionic/strsignal.cpp \
    bionic/stubs.cpp \
+    bionic/sysconf.cpp \
    bionic/tdestroy.cpp \
    bionic/tmpfile.cpp \
    bionic/__umask_chk.cpp \
    bionic/__vsnprintf_chk.cpp \
    bionic/__vsprintf_chk.cpp \
+    bionic/wait.cpp \
    bionic/wchar.cpp \

+libc_upstream_freebsd_src_files := \
+    upstream-freebsd/lib/libc/stdlib/realpath.c \
+    upstream-freebsd/lib/libc/string/wcpcpy.c \
+    upstream-freebsd/lib/libc/string/wcpncpy.c \
+    upstream-freebsd/lib/libc/string/wcscasecmp.c \
+    upstream-freebsd/lib/libc/string/wcscat.c \
+    upstream-freebsd/lib/libc/string/wcschr.c \
+    upstream-freebsd/lib/libc/string/wcscmp.c \
+    upstream-freebsd/lib/libc/string/wcscpy.c \
+    upstream-freebsd/lib/libc/string/wcscspn.c \
+    upstream-freebsd/lib/libc/string/wcsdup.c \
+    upstream-freebsd/lib/libc/string/wcslcat.c \
+    upstream-freebsd/lib/libc/string/wcslcpy.c \
+    upstream-freebsd/lib/libc/string/wcslen.c \
+    upstream-freebsd/lib/libc/string/wcsncasecmp.c \
+    upstream-freebsd/lib/libc/string/wcsncat.c \
+    upstream-freebsd/lib/libc/string/wcsncmp.c \
+    upstream-freebsd/lib/libc/string/wcsncpy.c \
+    upstream-freebsd/lib/libc/string/wcsnlen.c \
+    upstream-freebsd/lib/libc/string/wcspbrk.c \
+    upstream-freebsd/lib/libc/string/wcsrchr.c \
+    upstream-freebsd/lib/libc/string/wcsspn.c \
+    upstream-freebsd/lib/libc/string/wcsstr.c \
+    upstream-freebsd/lib/libc/string/wcstok.c \
+    upstream-freebsd/lib/libc/string/wmemchr.c \
+    upstream-freebsd/lib/libc/string/wmemcmp.c \
+    upstream-freebsd/lib/libc/string/wmemcpy.c \
+    upstream-freebsd/lib/libc/string/wmemmove.c \
+    upstream-freebsd/lib/libc/string/wmemset.c \
+
 libc_upstream_netbsd_src_files := \
    upstream-netbsd/common/lib/libc/hash/sha1/sha1.c \
    upstream-netbsd/common/lib/libc/inet/inet_addr.c \
@@ -311,6 +325,7 @@ libc_upstream_netbsd_src_files := \
    upstream-netbsd/libc/gen/ftw.c \
    upstream-netbsd/libc/gen/nftw.c \
    upstream-netbsd/libc/gen/nice.c \
+    upstream-netbsd/libc/gen/popen.c \
    upstream-netbsd/libc/gen/psignal.c \
    upstream-netbsd/libc/gen/setjmperr.c \
    upstream-netbsd/libc/gen/utime.c \
@@ -347,52 +362,11 @@ libc_upstream_netbsd_src_files := \
    upstream-netbsd/libc/string/strxfrm.c \
    upstream-netbsd/libc/unistd/killpg.c \

-# The following files are common, but must be compiled
-# with different C flags when building a static C library.
-#
-# The reason for this is the implementation of __get_tls()
-# that will differ between the shared and static versions
-# of the library.
-#
-# See comments in private/bionic_tls.h for more details.
-#
-# NOTE: bionic/pthread.c is added later to this list
-#       because it needs special handling on ARM, see
-#       below.
-#
-libc_static_common_src_files := \
-        bionic/__errno.c \
-        bionic/sysconf.cpp \
-
 # Architecture specific source files go here
 # =========================================================
 ifeq ($(TARGET_ARCH),arm)
 libc_common_src_files += \
-	arch-arm/bionic/abort_arm.S \
-	arch-arm/bionic/atomics_arm.c \
-	arch-arm/bionic/clone.S \
-	arch-arm/bionic/eabi.c \
-	arch-arm/bionic/_exit_with_stack_teardown.S \
-	arch-arm/bionic/ffs.S \
-	arch-arm/bionic/futex_arm.S \
-	arch-arm/bionic/__get_sp.S \
-	arch-arm/bionic/kill.S \
-	arch-arm/bionic/libgcc_compat.c \
-	arch-arm/bionic/memcmp16.S \
-	arch-arm/bionic/memcmp.S \
-	arch-arm/bionic/memcpy.S \
-	arch-arm/bionic/memset.S \
-	arch-arm/bionic/_setjmp.S \
-	arch-arm/bionic/setjmp.S \
-	arch-arm/bionic/sigsetjmp.S \
-	arch-arm/bionic/strcmp.S \
-	arch-arm/bionic/strcpy.S \
-	arch-arm/bionic/strlen.c.arm \
-	arch-arm/bionic/syscall.S \
-	arch-arm/bionic/tgkill.S \
-	arch-arm/bionic/tkill.S \
 	bionic/memmove.c.arm \
-	bionic/socketcalls.c \
 	string/bcopy.c \
 	string/strncmp.c \

@@ -407,79 +381,27 @@ libc_common_src_files += \

 libc_static_common_src_files += \
    bionic/pthread.c.arm \
+    bionic/pthread_create.cpp.arm \
    bionic/pthread_key.cpp.arm \

-# these are used by the static and dynamic versions of the libc
-# respectively
-libc_arch_static_src_files := \
-	arch-arm/bionic/exidx_static.c
-
-libc_arch_dynamic_src_files := \
-	arch-arm/bionic/exidx_dynamic.c
 endif # arm

 ifeq ($(TARGET_ARCH),x86)
 libc_common_src_files += \
-	arch-x86/bionic/__get_sp.S \
-	arch-x86/bionic/__get_tls.c \
-	arch-x86/bionic/__set_tls.c \
-	arch-x86/bionic/clone.S \
-	arch-x86/bionic/_exit_with_stack_teardown.S \
-	arch-x86/bionic/futex_x86.S \
-	arch-x86/bionic/setjmp.S \
-	arch-x86/bionic/_setjmp.S \
-	arch-x86/bionic/sigsetjmp.S \
-	arch-x86/bionic/vfork.S \
-	arch-x86/bionic/syscall.S \
-	arch-x86/string/bcopy_wrapper.S \
-	arch-x86/string/memcpy_wrapper.S \
-	arch-x86/string/memmove_wrapper.S \
-	arch-x86/string/bzero_wrapper.S \
-	arch-x86/string/memcmp_wrapper.S \
-	arch-x86/string/memset_wrapper.S \
-	arch-x86/string/strcmp_wrapper.S \
-	arch-x86/string/strncmp_wrapper.S \
-	arch-x86/string/strlen_wrapper.S \
-	string/strcpy.c \
-	bionic/pthread-atfork.c \
-	bionic/pthread-rwlocks.c \
-	bionic/pthread-timers.c \
-	bionic/ptrace.c
+    bionic/pthread-atfork.c \
+    bionic/pthread-rwlocks.c \
+    bionic/pthread-timers.c \
+    bionic/ptrace.c \
+    string/strcpy.c \

 libc_static_common_src_files += \
    bionic/pthread.c \
+    bionic/pthread_create.cpp \
    bionic/pthread_key.cpp \

-libc_arch_static_src_files := \
-	bionic/dl_iterate_phdr_static.c
-
-libc_arch_dynamic_src_files :=
 endif # x86

 ifeq ($(TARGET_ARCH),mips)
-libc_common_src_files += \
-	arch-mips/bionic/__get_sp.S \
-	arch-mips/bionic/__get_tls.c \
-	arch-mips/bionic/__set_tls.c \
-	arch-mips/bionic/_exit_with_stack_teardown.S \
-	arch-mips/bionic/_setjmp.S \
-	arch-mips/bionic/futex_mips.S \
-	arch-mips/bionic/bzero.S \
-	arch-mips/bionic/cacheflush.c \
-	arch-mips/bionic/clone.S \
-	arch-mips/bionic/ffs.S \
-	arch-mips/bionic/memcmp16.S \
-	arch-mips/bionic/memmove.c \
-	arch-mips/bionic/pipe.S \
-	arch-mips/bionic/setjmp.S \
-	arch-mips/bionic/sigsetjmp.S \
-	arch-mips/bionic/vfork.S
-
-libc_common_src_files += \
-	arch-mips/string/memset.S \
-	arch-mips/string/memcpy.S \
-	arch-mips/string/mips_strlen.c
-
 libc_common_src_files += \
 	bionic/memcmp.c \
 	string/bcopy.c \
@@ -495,14 +417,45 @@ libc_common_src_files += \

 libc_static_common_src_files += \
    bionic/pthread.c \
+    bionic/pthread_create.cpp \
    bionic/pthread_key.cpp \

-libc_arch_static_src_files := \
-	bionic/dl_iterate_phdr_static.c
-
-libc_arch_dynamic_src_files :=
 endif # mips

+ifeq ($(strip $(TARGET_CPU_VARIANT)),)
+$(warning TARGET_CPU_VARIANT is not defined)
+endif
+
+###########################################################
+## Add cpu specific source files.
+##
+## This can be called multiple times, but it will only add
+## the first source file for each unique $(1).
+## $(1): Unique identifier to identify the cpu variant
+##       implementation.
+## $(2): Cpu specific source file.
+###########################################################
+
+define libc-add-cpu-variant-src
+$(if $(filter true,$(_LIBC_ARCH_CPU_VARIANT_HAS_$(1))), \
+	, \
+     $(eval _LIBC_ARCH_CPU_VARIANT_HAS_$(1) := true) \
+     $(eval _LIBC_ARCH_CPU_VARIANT_SRC_FILE.$(1) := $(2)) \
+     $(eval _LIBC_ARCH_CPU_VARIANT_SRC_FILES += $(2)) \
+)
+endef
+
+_LIBC_ARCH_COMMON_SRC_FILES :=
+_LIBC_ARCH_CPU_VARIANT_SRC_FILES :=
+_LIBC_ARCH_STATIC_SRC_FILES :=
+_LIBC_ARCH_DYNAMIC_SRC_FILES :=
+include bionic/libc/arch-$(TARGET_ARCH)/$(TARGET_ARCH).mk
+
+libc_common_src_files += $(_LIBC_ARCH_COMMON_SRC_FILES)
+libc_common_src_files += $(_LIBC_ARCH_CPU_VARIANT_SRC_FILES)
+libc_arch_static_src_files := $(_LIBC_ARCH_STATIC_SRC_FILES)
+libc_arch_dynamic_src_files := $(_LIBC_ARCH_DYNAMIC_SRC_FILES)
+
 # Define some common cflags
 # ========================================================
 libc_common_cflags := \
@@ -539,27 +492,6 @@ ifeq ($(TARGET_ARCH),arm)
  libc_common_cflags += -DSOFTFLOAT
  libc_common_cflags += -fstrict-aliasing
  libc_crt_target_cflags := -mthumb-interwork
-  #
-  # Define HAVE_ARM_TLS_REGISTER macro to indicate to the C library
-  # that it should access the hardware TLS register directly in
-  # private/bionic_tls.h
-  #
-  # The value must match your kernel configuration
-  #
-  ifeq ($(ARCH_ARM_HAVE_TLS_REGISTER),true)
-    libc_common_cflags += -DHAVE_ARM_TLS_REGISTER
-  endif
-  #
-  # Define HAVE_32_BYTE_CACHE_LINES to indicate to C
-  # library it should use to 32-byte version of memcpy, and not
-  # the 64-byte version.
-  #
-  ifeq ($(ARCH_ARM_HAVE_32_BYTE_CACHE_LINES),true)
-    libc_common_cflags += -DHAVE_32_BYTE_CACHE_LINE
-  endif
-  ifeq ($(ARCH_ARM_USE_NON_NEON_MEMCPY),true)
-    libc_common_cflags += -DARCH_ARM_USE_NON_NEON_MEMCPY
-  endif
 endif # !arm

 ifeq ($(TARGET_ARCH),x86)
@@ -744,8 +676,8 @@ WITH_MALLOC_CHECK_LIBC_A := $(strip $(WITH_MALLOC_CHECK_LIBC_A))

 include $(CLEAR_VARS)

-LOCAL_SRC_FILES := bionic/ssp.cpp
-LOCAL_CFLAGS := $(libc_common_cflags) -fno-stack-protector
+LOCAL_SRC_FILES := bionic/__stack_chk_fail.cpp
+LOCAL_CFLAGS := $(libc_common_cflags) -fno-stack-protector -Werror
 LOCAL_C_INCLUDES := $(libc_common_c_includes)
 LOCAL_MODULE := libbionic_ssp
 LOCAL_ADDITIONAL_DEPENDENCIES := $(LOCAL_PATH)/Android.mk
@@ -754,6 +686,29 @@ LOCAL_SYSTEM_SHARED_LIBRARIES :=
 include $(BUILD_STATIC_LIBRARY)


+# ========================================================
+# libc_freebsd.a - upstream FreeBSD C library code
+# ========================================================
+#
+# These files are built with the freebsd-compat.h header file
+# automatically included.
+
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := $(libc_upstream_freebsd_src_files)
+LOCAL_CFLAGS := \
+    $(libc_common_cflags) \
+    -I$(LOCAL_PATH)/upstream-freebsd \
+    -I$(LOCAL_PATH)/upstream-freebsd/libc/include \
+    -include upstream-freebsd/freebsd-compat.h
+LOCAL_C_INCLUDES := $(libc_common_c_includes)
+LOCAL_MODULE := libc_freebsd
+LOCAL_ADDITIONAL_DEPENDENCIES := $(LOCAL_PATH)/Android.mk
+LOCAL_SYSTEM_SHARED_LIBRARIES :=
+
+include $(BUILD_STATIC_LIBRARY)
+
+
 # ========================================================
 # libc_netbsd.a - upstream NetBSD C library code
 # ========================================================
@@ -806,9 +761,12 @@ LOCAL_CFLAGS := $(libc_common_cflags) \
 LOCAL_C_INCLUDES := $(libc_common_c_includes)
 LOCAL_MODULE := libc_common
 LOCAL_ADDITIONAL_DEPENDENCIES := $(LOCAL_PATH)/Android.mk
-LOCAL_WHOLE_STATIC_LIBRARIES := libbionic_ssp libc_bionic libc_netbsd
+LOCAL_WHOLE_STATIC_LIBRARIES := libbionic_ssp libc_bionic libc_freebsd libc_netbsd
 LOCAL_SYSTEM_SHARED_LIBRARIES :=

+# TODO: split out the asflags.
+LOCAL_ASFLAGS := $(LOCAL_CFLAGS)
+
 include $(BUILD_STATIC_LIBRARY)


@@ -817,10 +775,10 @@ include $(BUILD_STATIC_LIBRARY)
 # ========================================================
 #
 # This is a version of the static C library that does not
-# include malloc. It's useful in situations when calling
-# the user wants to provide their own malloc implementation,
-# or wants to explicitly disallow the use of the use of malloc,
-# like the dynamic loader.
+# include malloc. It's useful in situations when the user wants
+# to provide their own malloc implementation, or wants to
+# explicitly disallow the use of the use of malloc,
+# such as in the dynamic loader.

 include $(CLEAR_VARS)

--- a/libc/NOTICE
+++ b/libc/NOTICE
@@ -4030,6 +4030,34 @@ SUCH DAMAGE.

 -------------------------------------------------------------------

+Copyright (c) 2013 ARM Ltd
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+3. The name of the company may not be used to endorse or promote
+   products derived from this software without specific prior written
+   permission.
+
+THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+-------------------------------------------------------------------
+
 Copyright (c)1999 Citrus Project,
 All rights reserved.

--- a/libc/README
+++ b/libc/README
@@ -1,54 +0,0 @@
-Welcome to Bionic, Android's small and custom C library for the Android
-platform.
-
-Bionic is mainly a port of the BSD C library to our Linux kernel with the
-following additions/changes:
-
- no support for locales
- no support for wide chars (i.e. multi-byte characters)
- its own smallish implementation of pthreads based on Linux futexes
- support for x86, ARM and ARM thumb CPU instruction sets and kernel interfaces
-
-Bionic is released under the standard 3-clause BSD License
-
-Bionic doesn't want to implement all features of a traditional C library, we only
-add features to it as we need them, and we try to keep things as simple and small
-as possible. Our goal is not to support scaling to thousands of concurrent threads
-on multi-processors machines; we're running this on cell-phones, damnit !!
-
-Note that Bionic doesn't provide a libthread_db or a libm implementation.
-
-
-Adding new syscalls:
-====================
-
-Bionic provides the gensyscalls.py Python script to automatically generate syscall
-stubs from the list defined in the file SYSCALLS.TXT. You can thus add a new syscall
-by doing the following:
-
- edit SYSCALLS.TXT
- add a new line describing your syscall, it should look like:
-
-   return_type  syscall_name(parameters)    syscall_number
-
- in the event where you want to differentiate the syscall function from its entry name,
-  use the alternate:
-
-   return_type  funcname:syscall_name(parameters)  syscall_number
-
- additionally, if the syscall number is different between ARM and x86, use:
-
-   return_type  funcname[:syscall_name](parameters)   arm_number,x86_number
-
- a syscall number can be -1 to indicate that the syscall is not implemented on
-  a given platform, for example:
-
-   void   __set_tls(void*)   arm_number,-1
-
-
-the comments in SYSCALLS.TXT contain more information about the line format
-
-You can also use the 'checksyscalls.py' script to check that all the syscall
-numbers you entered are correct. It does so by looking at the values defined in
-your Linux kernel headers. The script indicates where the values are incorrect
-and what is expected instead.
--- a/libc/SYSCALLS.TXT
+++ b/libc/SYSCALLS.TXT
@@ -1,13 +1,11 @@
-# this file is used to list all the syscalls that will be supported by
-# the Bionic C library. It is used to automatically generate the syscall
-# stubs, the list of syscall constants (__NR_xxxx) and the content of <linux/_unistd.h>
+# This file is used to automatically generate bionic's the system calls stubs.
 #
-# each non comment line has the following format:
+# Each non comment line has the following format:
 #
-# return_type    func_name[:syscall_name[:call_id]]([parameter_list])  (syscall_number|"stub")
+# return_type    func_name[:syscall_name[:call_id]]([parameter_list])  (1|-1|"stub")
 #
-# note that:
-#      - syscall_name correspond to the name of the syscall, which may differ from
+# Note that:
+#      - syscall_name corresponds to the name of the syscall, which may differ from
 #        the exported function name (example: the exit syscall is implemented by the _exit()
 #        function, which is not the same as the standard C exit() function which calls it)
 #        The call_id parameter, given that func_name and syscall_name have
@@ -18,299 +16,306 @@
 #      - each parameter type is assumed to be stored on 32 bits, there is no plan to support
 #        64-bit architectures at the moment
 #
-#      - it there is "stub" instead of a syscall number, the tool will not generate any
-#        assembler template for the syscall; it's up to the bionic implementation to provide
-#        a relevant C stub
+#      - the final field can be "1", meaning: generate a stub for each architecture,
+#        taking the constants from the kernel header files.
 #
-#      - additionally, if the syscall number is different amoung ARM, and x86, MIPS use:
-#        return_type funcname[:syscall_name](parameters) arm_number,x86_number,mips_number
+#      - the final field can be "stub" meaning: do not generate any stubs ---
+#        in this case, a hand-written custom stub must be provided.
+#        TODO: replace this with something like "custom" or "none", or remove
+#        it entirely.
 #
-# the file is processed by a python script named gensyscalls.py
+#      - the final field can be a three-element list of 1s and -1 meaning:
+#        this system call is only available on some of the architectures (1),
+#        and no stub should be generated for those architectures marked with -1.
+#        the order is arm,x86,mips.
+#        TODO: replace this with something more readable like "-arm,-mips" (meaning x86 only).
+#
+# This file is processed by a python script named gensyscalls.py.
 #

 # process management
-void    _exit:exit_group (int)      248,252,246
+void    _exit:exit_group (int)      1
 void    _exit_thread:exit (int)     1
-pid_t   __fork:fork (void)           2
-pid_t   _waitpid:waitpid (pid_t, int*, int, struct rusage*)   -1,7,7
-int     __waitid:waitid(int, pid_t, struct siginfo_t*, int,void*)          280,284,278
+pid_t   __fork:fork (void)           1
+pid_t   _waitpid:waitpid (pid_t, int*, int, struct rusage*)   -1,1,1
+int     __waitid:waitid(int, pid_t, struct siginfo_t*, int,void*)          1
+pid_t   wait4(pid_t pid, int *status, int options, struct rusage *rusage)   1

 # NOTE: this system call is never called directly, but we list it there
 #       to have __NR_clone properly defined.
 #
-pid_t   __sys_clone:clone (int, void*, int*, void*, int*) 120
+pid_t   __sys_clone:clone (int, void*, int*, void*, int*) 1

-int     execve (const char*, char* const*, char* const*)  11
+int     execve (const char*, char* const*, char* const*)  1

-int     __setuid:setuid32 (uid_t)    213,213,-1
-int     __setuid:setuid (uid_t)   -1,-1,23
-uid_t   getuid:getuid32 ()         199,199,-1
-uid_t   getuid:getuid ()           -1,-1,24
-gid_t   getgid:getgid32 ()         200,200,-1
-gid_t   getgid:getgid ()           -1,-1,47
-uid_t   geteuid:geteuid32 ()       201,201,-1
-uid_t   geteuid:geteuid ()         -1,-1,49
-gid_t   getegid:getegid32 ()       202,202,-1
-gid_t   getegid:getegid ()         -1,-1,50
-uid_t   getresuid:getresuid32 (uid_t *ruid, uid_t *euid, uid_t *suid)   209,209,-1
-uid_t   getresuid:getresuid (uid_t *ruid, uid_t *euid, uid_t *suid)     -1,-1,186
-gid_t   getresgid:getresgid32 (gid_t *rgid, gid_t *egid, gid_t *sgid)   211,211,-1
-gid_t   getresgid:getresgid (gid_t *rgid, gid_t *egid, gid_t *sgid)     -1,-1,191
-pid_t   gettid()                   224,224,222
-ssize_t readahead(int, off64_t, size_t)     225,225,223
-int     getgroups:getgroups32(int, gid_t *)    205,205,-1
-int     getgroups:getgroups(int, gid_t *)      -1,-1,80
-pid_t   getpgid(pid_t)             132
-pid_t   getppid()                  64
-pid_t   getsid(pid_t)              147,147,151
-pid_t   setsid()                   66
-int     setgid:setgid32(gid_t)     214,214,-1
-int     setgid:setgid(gid_t)       -1,-1,46
+int     __setuid:setuid32 (uid_t)    1,1,-1
+int     __setuid:setuid (uid_t)   -1,-1,1
+uid_t   getuid:getuid32 ()         1,1,-1
+uid_t   getuid:getuid ()           -1,-1,1
+gid_t   getgid:getgid32 ()         1,1,-1
+gid_t   getgid:getgid ()           -1,-1,1
+uid_t   geteuid:geteuid32 ()       1,1,-1
+uid_t   geteuid:geteuid ()         -1,-1,1
+gid_t   getegid:getegid32 ()       1,1,-1
+gid_t   getegid:getegid ()         -1,-1,1
+uid_t   getresuid:getresuid32 (uid_t *ruid, uid_t *euid, uid_t *suid)   1,1,-1
+uid_t   getresuid:getresuid (uid_t *ruid, uid_t *euid, uid_t *suid)     -1,-1,1
+gid_t   getresgid:getresgid32 (gid_t *rgid, gid_t *egid, gid_t *sgid)   1,1,-1
+gid_t   getresgid:getresgid (gid_t *rgid, gid_t *egid, gid_t *sgid)     -1,-1,1
+pid_t   gettid()                   1
+ssize_t readahead(int, off64_t, size_t)     1
+int     getgroups:getgroups32(int, gid_t *)    1,1,-1
+int     getgroups:getgroups(int, gid_t *)      -1,-1,1
+pid_t   getpgid(pid_t)             1
+pid_t   getppid()                  1
+pid_t   getsid(pid_t)              1
+pid_t   setsid()                   1
+int     setgid:setgid32(gid_t)     1,1,-1
+int     setgid:setgid(gid_t)       -1,-1,1
 int     seteuid:seteuid32(uid_t)   stub
-int     __setreuid:setreuid32(uid_t, uid_t)   203,203,-1
-int     __setreuid:setreuid(uid_t, uid_t)     -1,-1,70
-int     __setresuid:setresuid32(uid_t, uid_t, uid_t)   208,208,-1
-int     __setresuid:setresuid(uid_t, uid_t, uid_t)     -1,-1,185
-int     setresgid:setresgid32(gid_t, gid_t, gid_t)   210,210,-1
-int     setresgid:setresgid(gid_t, gid_t, gid_t)     -1,-1,190
-void*   __brk:brk(void*)           45
+int     __setreuid:setreuid32(uid_t, uid_t)   1,1,-1
+int     __setreuid:setreuid(uid_t, uid_t)     -1,-1,1
+int     __setresuid:setresuid32(uid_t, uid_t, uid_t)   1,1,-1
+int     __setresuid:setresuid(uid_t, uid_t, uid_t)     -1,-1,1
+int     setresgid:setresgid32(gid_t, gid_t, gid_t)   1,1,-1
+int     setresgid:setresgid(gid_t, gid_t, gid_t)     -1,-1,1
+void*   __brk:brk(void*)           1
 # see comments in arch-arm/bionic/kill.S to understand why we don't generate an ARM stub for kill/tkill
-int     kill(pid_t, int)           -1,37,37
-int     tkill(pid_t tid, int sig)  -1,238,236
-int     tgkill(pid_t tgid, pid_t tid, int sig)  -1,270,266
-int     __ptrace:ptrace(int request, int pid, void* addr, void* data)  26
-int     __set_thread_area:set_thread_area(void*  user_desc)  -1,243,283
-int     __getpriority:getpriority(int, int)  96
-int     setpriority(int, int, int)   97
-int     setrlimit(int resource, const struct rlimit *rlp)  75
-int     getrlimit:ugetrlimit(int resource, struct rlimit *rlp)  191,191,-1
-int     getrlimit:getrlimit(int resource, struct rlimit *rlp)  -1,-1,76
-int     getrusage(int who, struct rusage*  r_usage)  77
-int     setgroups:setgroups32(int, const gid_t *)   206,206,-1
-int     setgroups:setgroups(int, const gid_t *)     -1,-1,81
+int     kill(pid_t, int)           -1,1,1
+int     tkill(pid_t tid, int sig)  -1,1,1
+int     tgkill(pid_t tgid, pid_t tid, int sig)  -1,1,1
+int     __ptrace:ptrace(int request, int pid, void* addr, void* data)  1
+int     __set_thread_area:set_thread_area(void*  user_desc)  -1,1,1
+int     __getpriority:getpriority(int, int)  1
+int     setpriority(int, int, int)   1
+int     setrlimit(int resource, const struct rlimit *rlp)  1
+int     getrlimit:ugetrlimit(int resource, struct rlimit *rlp)  1,1,-1
+int     getrlimit:getrlimit(int resource, struct rlimit *rlp)  -1,-1,1
+int     getrusage(int who, struct rusage*  r_usage)  1
+int     setgroups:setgroups32(int, const gid_t *)   1,1,-1
+int     setgroups:setgroups(int, const gid_t *)     -1,-1,1
 pid_t   getpgrp(void)  stub
-int     setpgid(pid_t, pid_t)  57
-pid_t   vfork(void)  190,-1,-1
-int     setregid:setregid32(gid_t, gid_t)  204,204,-1
-int     setregid:setregid(gid_t, gid_t)    -1,-1,71
-int     chroot(const char *)  61
+int     setpgid(pid_t, pid_t)  1
+pid_t   vfork(void)  1,-1,-1
+int     setregid:setregid32(gid_t, gid_t)  1,1,-1
+int     setregid:setregid(gid_t, gid_t)    -1,-1,1
+int     chroot(const char *)  1
 # IMPORTANT: Even though <sys/prctl.h> declares prctl(int,...), the syscall stub must take 6 arguments
 #            to match the kernel implementation.
-int     prctl(int option, unsigned int arg2, unsigned int arg3, unsigned int arg4, unsigned int arg5)  172,172,192
-int     capget(cap_user_header_t header, cap_user_data_t data) 184,184,204
-int     capset(cap_user_header_t header, const cap_user_data_t data) 185,185,205
-int     sigaltstack(const stack_t*, stack_t*) 186,186,206
-int     acct(const char*  filepath)  51
+int     prctl(int option, unsigned int arg2, unsigned int arg3, unsigned int arg4, unsigned int arg5) 1
+int     capget(cap_user_header_t header, cap_user_data_t data) 1
+int     capset(cap_user_header_t header, const cap_user_data_t data) 1
+int     sigaltstack(const stack_t*, stack_t*) 1
+int     acct(const char*  filepath)  1

 # file descriptors
-ssize_t     read (int, void*, size_t)        3
-ssize_t     write (int, const void*, size_t)       4
-ssize_t     pread64 (int, void *, size_t, off64_t) 180,180,200
-ssize_t     pwrite64 (int, void *, size_t, off64_t) 181,181,201
-int         __open:open (const char*, int, mode_t)  5
-int         __openat:openat (int, const char*, int, mode_t)  322,295,288
-int         close (int)                      6
+ssize_t     read (int, void*, size_t)        1
+ssize_t     write (int, const void*, size_t)       1
+ssize_t     pread64 (int, void *, size_t, off64_t) 1
+ssize_t     pwrite64 (int, void *, size_t, off64_t) 1
+int         __open:open (const char*, int, mode_t)  1
+int         __openat:openat (int, const char*, int, mode_t) 1
+int         close (int)                      1
 int         creat(const char*, mode_t)       stub
-off_t       lseek(int, off_t, int)           19
-int         __llseek:_llseek (int, unsigned long, unsigned long, loff_t*, int)  140
-pid_t       getpid ()    20
+off_t       lseek(int, off_t, int)           1
+int         __llseek:_llseek (int, unsigned long, unsigned long, loff_t*, int)  1
+pid_t       getpid ()    1
 void *      mmap(void *, size_t, int, int, int, long)  stub
-void *      __mmap2:mmap2(void*, size_t, int, int, int, long)   192,192,210
-int         munmap(void *, size_t)  91
-void *      mremap(void *, size_t, size_t, unsigned long)  163,163,167
-int         msync(const void *, size_t, int)    144
-int         mprotect(const void *, size_t, int)  125
-int         madvise(const void *, size_t, int)  220,219,218
-int         mlock(const void *addr, size_t len)    150,150,154
-int         munlock(const void *addr, size_t len)   151,151,155
-int         mlockall(int flags)   152,152,156
-int         munlockall()   153,153,157
-int         mincore(void*  start, size_t  length, unsigned char*  vec)   219,218,217
-int         __ioctl:ioctl(int, int, void *)  54
-int         readv(int, const struct iovec *, int)   145
-int         writev(int, const struct iovec *, int)  146
-int         __fcntl:fcntl(int, int, void*)  55
-int         flock(int, int)   143
-int         fchmod(int, mode_t)  94
-int         dup(int)  41
-int         pipe(int *)  42,42,-1
-int         pipe2(int *, int) 359,331,328
-int         dup2(int, int)   63
-int         select:_newselect(int, struct fd_set *, struct fd_set *, struct fd_set *, struct timeval *)  142
-int         ftruncate(int, off_t)  93
-int         ftruncate64(int, off64_t) 194,194,212
-int         getdents:getdents64(unsigned int, struct dirent *, unsigned int)   217,220,219
-int         fsync(int)  118
-int         fdatasync(int) 148,148,152
-int         fchown:fchown32(int, uid_t, gid_t)  207,207,-1
-int         fchown:fchown(int, uid_t, gid_t)    -1,-1,95
-void        sync(void)  36
-int         __fcntl64:fcntl64(int, int, void *)  221,221,220
-int         __fstatfs64:fstatfs64(int, size_t, struct statfs *)  267,269,256
-ssize_t     sendfile(int out_fd, int in_fd, off_t *offset, size_t count)  187,187,207
-int         fstatat:fstatat64(int dirfd, const char *path, struct stat *buf, int flags)   327,300,293
-int         mkdirat(int dirfd, const char *pathname, mode_t mode)  323,296,289
-int         fchownat(int dirfd, const char *path, uid_t owner, gid_t group, int flags)  325,298,291
-int         fchmodat(int dirfd, const char *path, mode_t mode, int flags)  333,306,299
-int         renameat(int olddirfd, const char *oldpath, int newdirfd, const char *newpath)  329,302,295
-int         fsetxattr(int, const char *, const void *, size_t, int) 228,228,226
-ssize_t     fgetxattr(int, const char *, void *, size_t) 231,231,229
-ssize_t     flistxattr(int, char *, size_t) 234,234,232
-int         fremovexattr(int, const char *) 237,237,235
+void *      __mmap2:mmap2(void*, size_t, int, int, int, long)   1
+int         munmap(void *, size_t)  1
+void *      mremap(void *, size_t, size_t, unsigned long)  1
+int         msync(const void *, size_t, int)    1
+int         mprotect(const void *, size_t, int)  1
+int         madvise(const void *, size_t, int)  1
+int         mlock(const void *addr, size_t len)    1
+int         munlock(const void *addr, size_t len)   1
+int         mlockall(int flags)   1
+int         munlockall()   1
+int         mincore(void*  start, size_t  length, unsigned char*  vec)   1
+int         __ioctl:ioctl(int, int, void *)  1
+int         readv(int, const struct iovec *, int)   1
+int         writev(int, const struct iovec *, int)  1
+int         __fcntl:fcntl(int, int, void*)  1
+int         flock(int, int)   1
+int         fchmod(int, mode_t)  1
+int         dup(int)  1
+int         pipe(int *)  1,1,-1
+int         pipe2(int *, int) 1
+int         dup2(int, int)   1
+int         select:_newselect(int, struct fd_set *, struct fd_set *, struct fd_set *, struct timeval *)  1
+int         ftruncate(int, off_t)  1
+int         ftruncate64(int, off64_t) 1
+int         getdents:getdents64(unsigned int, struct dirent *, unsigned int)   1
+int         fsync(int)  1
+int         fdatasync(int) 1
+int         fchown:fchown32(int, uid_t, gid_t)  1,1,-1
+int         fchown:fchown(int, uid_t, gid_t)    -1,-1,1
+void        sync(void)  1
+int         __fcntl64:fcntl64(int, int, void *)  1
+int         __fstatfs64:fstatfs64(int, size_t, struct statfs *)  1
+ssize_t     sendfile(int out_fd, int in_fd, off_t *offset, size_t count)  1
+int         fstatat:fstatat64(int dirfd, const char *path, struct stat *buf, int flags)   1
+int         mkdirat(int dirfd, const char *pathname, mode_t mode)  1
+int         fchownat(int dirfd, const char *path, uid_t owner, gid_t group, int flags)  1
+int         fchmodat(int dirfd, const char *path, mode_t mode, int flags)  1
+int         renameat(int olddirfd, const char *oldpath, int newdirfd, const char *newpath)  1
+int         fsetxattr(int, const char *, const void *, size_t, int) 1
+ssize_t     fgetxattr(int, const char *, void *, size_t) 1
+ssize_t     flistxattr(int, char *, size_t) 1
+int         fremovexattr(int, const char *) 1

 # file system
-int     link (const char*, const char*)  9
-int     unlink (const char*)             10
-int     unlinkat (int, const char *, int)   328,301,294
-int     chdir (const char*)              12
-int     mknod (const char*, mode_t, dev_t)  14
-int     chmod (const char*,mode_t)          15
-int     chown:chown32(const char *, uid_t, gid_t)  212,212,-1
-int     chown:chown(const char *, uid_t, gid_t)    -1,-1,202
-int     lchown:lchown32 (const char*, uid_t, gid_t)  198,198,-1
-int     lchown:lchown (const char*, uid_t, gid_t)  -1,-1,16
-int     mount (const char*, const char*, const char*, unsigned long, const void*)  21
+int     link (const char*, const char*)  1
+int     unlink (const char*)             1
+int     unlinkat (int, const char *, int)   1
+int     chdir (const char*)              1
+int     mknod (const char*, mode_t, dev_t)  1
+int     chmod (const char*,mode_t)          1
+int     chown:chown32(const char *, uid_t, gid_t)  1,1,-1
+int     chown:chown(const char *, uid_t, gid_t)    -1,-1,1
+int     lchown:lchown32 (const char*, uid_t, gid_t)  1,1,-1
+int     lchown:lchown (const char*, uid_t, gid_t)  -1,-1,1
+int     mount (const char*, const char*, const char*, unsigned long, const void*)  1
 int     umount(const char*)  stub
-int     umount2 (const char*, int)  52
-int     fstat:fstat64(int, struct stat*)    197,197,215
-int     stat:stat64(const char *, struct stat *)  195,195,213
-int     lstat:lstat64(const char *, struct stat *)  196,196,214
-int     mkdir(const char *, mode_t) 39
-int     readlink(const char *, char *, size_t)  85
-int     rmdir(const char *)  40
-int     rename(const char *, const char *)  38
-int     __getcwd:getcwd(char * buf, size_t size)  183,183,203
-int     access(const char *, int)  33
-int     faccessat(int, const char *, int, int)  334,307,300
-int     symlink(const char *, const char *)  83
-int     fchdir(int)    133
-int     truncate(const char*, off_t)    92
-int     setxattr(const char *, const char *, const void *, size_t, int) 226,226,224
-int     lsetxattr(const char *, const char *, const void *, size_t, int) 227,227,225
-ssize_t getxattr(const char *, const char *, void *, size_t) 229,229,227
-ssize_t lgetxattr(const char *, const char *, void *, size_t) 230,230,228
-ssize_t listxattr(const char *, char *, size_t) 232,232,230
-ssize_t llistxattr(const char *, char *, size_t) 233,233,231
-int     removexattr(const char *, const char *) 235,235,233
-int     lremovexattr(const char *, const char *) 236,236,234
-int     __statfs64:statfs64(const char *, size_t, struct statfs *)  266,268,255
-long    unshare(unsigned long)  337,310,303
+int     umount2 (const char*, int)  1
+int     fstat:fstat64(int, struct stat*)    1
+int     stat:stat64(const char *, struct stat *)  1
+int     lstat:lstat64(const char *, struct stat *)  1
+int     mkdir(const char *, mode_t) 1
+int     readlink(const char *, char *, size_t)  1
+int     rmdir(const char *)  1
+int     rename(const char *, const char *)  1
+int     __getcwd:getcwd(char * buf, size_t size)  1
+int     access(const char *, int)  1
+int     faccessat(int, const char *, int, int)  1
+int     symlink(const char *, const char *)  1
+int     fchdir(int)    1
+int     truncate(const char*, off_t)    1
+int     setxattr(const char *, const char *, const void *, size_t, int) 1
+int     lsetxattr(const char *, const char *, const void *, size_t, int) 1
+ssize_t getxattr(const char *, const char *, void *, size_t) 1
+ssize_t lgetxattr(const char *, const char *, void *, size_t) 1
+ssize_t listxattr(const char *, char *, size_t) 1
+ssize_t llistxattr(const char *, char *, size_t) 1
+int     removexattr(const char *, const char *) 1
+int     lremovexattr(const char *, const char *) 1
+int     __statfs64:statfs64(const char *, size_t, struct statfs *)  1
+long    unshare(unsigned long)  1

 # time
-int           pause ()                       29
-int           gettimeofday(struct timeval*, struct timezone*)       78
-int           settimeofday(const struct timeval*, const struct timezone*)   79
-clock_t       times(struct tms *)       43
-int           nanosleep(const struct timespec *, struct timespec *)   162,162,166
-int           clock_gettime(clockid_t clk_id, struct timespec *tp)    263,265,263
-int           clock_settime(clockid_t clk_id, const struct timespec *tp)  262,264,262
-int           clock_getres(clockid_t clk_id, struct timespec *res)   264,266,264
-int           clock_nanosleep(clockid_t clock_id, int flags, const struct timespec *req, struct timespec *rem)  265,267,265
-int           getitimer(int, const struct itimerval *)   105
-int           setitimer(int, const struct itimerval *, struct itimerval *)  104
-int           __timer_create:timer_create(clockid_t clockid, struct sigevent *evp, timer_t *timerid)    257,259,257
-int           __timer_settime:timer_settime(timer_t, int, const struct itimerspec*, struct itimerspec*) 258,260,258
-int           __timer_gettime:timer_gettime(timer_t, struct itimerspec*)                                259,261,259
-int           __timer_getoverrun:timer_getoverrun(timer_t)                                              260,262,260
-int           __timer_delete:timer_delete(timer_t)                                                      261,263,261
-int           utimes(const char*, const struct timeval tvp[2])                          269,271,267
-int           utimensat(int, const char *, const struct timespec times[2], int)         348,320,316
+int           pause ()                       1
+int           gettimeofday(struct timeval*, struct timezone*)       1
+int           settimeofday(const struct timeval*, const struct timezone*)   1
+clock_t       times(struct tms *)       1
+int           nanosleep(const struct timespec *, struct timespec *)   1
+int           clock_gettime(clockid_t clk_id, struct timespec *tp)    1
+int           clock_settime(clockid_t clk_id, const struct timespec *tp)  1
+int           clock_getres(clockid_t clk_id, struct timespec *res)   1
+int           clock_nanosleep(clockid_t clock_id, int flags, const struct timespec *req, struct timespec *rem)  1
+int           getitimer(int, const struct itimerval *)   1
+int           setitimer(int, const struct itimerval *, struct itimerval *)  1
+int           __timer_create:timer_create(clockid_t clockid, struct sigevent *evp, timer_t *timerid)    1
+int           __timer_settime:timer_settime(timer_t, int, const struct itimerspec*, struct itimerspec*) 1
+int           __timer_gettime:timer_gettime(timer_t, struct itimerspec*)                                1
+int           __timer_getoverrun:timer_getoverrun(timer_t)                                              1
+int           __timer_delete:timer_delete(timer_t)                                                      1
+int           utimes(const char*, const struct timeval tvp[2])                          1
+int           utimensat(int, const char *, const struct timespec times[2], int)         1

 # signals
-int     sigaction(int, const struct sigaction *, struct sigaction *)  67
-int     sigprocmask(int, const sigset_t *, sigset_t *)  126
-int     __sigsuspend:sigsuspend(int unused1, int unused2, unsigned mask)  72,72,-1
-int     __sigsuspend:sigsuspend(const sigset_t *mask)  -1,-1,72
-int     __rt_sigaction:rt_sigaction (int sig, const struct sigaction *act, struct sigaction *oact, size_t sigsetsize)  174,174,194
-int     __rt_sigprocmask:rt_sigprocmask (int  how, const sigset_t *set, sigset_t *oset, size_t sigsetsize)  175,175,195
-int     __rt_sigtimedwait:rt_sigtimedwait(const sigset_t *set, struct siginfo_t  *info, struct timespec_t  *timeout, size_t  sigset_size)  177,177,197
-int     sigpending(sigset_t *)  73
-int     signalfd4(int fd, const sigset_t *mask, size_t sizemask, int flags)  355,327,324
+int     sigaction(int, const struct sigaction *, struct sigaction *)  1
+int     sigprocmask(int, const sigset_t *, sigset_t *)  1
+int     __sigsuspend:sigsuspend(int unused1, int unused2, unsigned mask)  1,1,-1
+int     __sigsuspend:sigsuspend(const sigset_t *mask)  -1,-1,1
+int     __rt_sigaction:rt_sigaction (int sig, const struct sigaction *act, struct sigaction *oact, size_t sigsetsize)  1
+int     __rt_sigprocmask:rt_sigprocmask (int  how, const sigset_t *set, sigset_t *oset, size_t sigsetsize)  1
+int     __rt_sigtimedwait:rt_sigtimedwait(const sigset_t *set, struct siginfo_t  *info, struct timespec_t  *timeout, size_t  sigset_size)  1
+int     sigpending(sigset_t *)  1
+int     signalfd4(int fd, const sigset_t *mask, size_t sizemask, int flags)  1

 # sockets
-int           socket(int, int, int)              281,-1,183
-int           socketpair(int, int, int, int*)    288,-1,184
-int           bind(int, struct sockaddr *, int)  282,-1,169
-int           connect(int, struct sockaddr *, socklen_t)   283,-1,170
-int           listen(int, int)                   284,-1,174
-int           accept(int, struct sockaddr *, socklen_t *)  285,-1,168
-int           getsockname(int, struct sockaddr *, socklen_t *)  286,-1,172
-int           getpeername(int, struct sockaddr *, socklen_t *)  287,-1,171
-int           sendto(int, const void *, size_t, int, const struct sockaddr *, socklen_t)  290,-1,180
-int           recvfrom(int, void *, size_t, unsigned int, struct sockaddr *, socklen_t *)  292,-1,176
-int           shutdown(int, int)  293,-1,182
-int           setsockopt(int, int, int, const void *, socklen_t)  294,-1,181
-int           getsockopt(int, int, int, void *, socklen_t *)    295,-1,173
-int           sendmsg(int, const struct msghdr *, unsigned int)  296,-1,179
-int           recvmsg(int, struct msghdr *, unsigned int)   297,-1,177
+int           socket(int, int, int)              1,-1,1
+int           socketpair(int, int, int, int*)    1,-1,1
+int           bind(int, struct sockaddr *, int)  1,-1,1
+int           connect(int, struct sockaddr *, socklen_t)   1,-1,1
+int           listen(int, int)                   1,-1,1
+int           accept(int, struct sockaddr *, socklen_t *)  1,-1,1
+int           getsockname(int, struct sockaddr *, socklen_t *)  1,-1,1
+int           getpeername(int, struct sockaddr *, socklen_t *)  1,-1,1
+int           sendto(int, const void *, size_t, int, const struct sockaddr *, socklen_t)  1,-1,1
+int           recvfrom(int, void *, size_t, unsigned int, struct sockaddr *, socklen_t *)  1,-1,1
+int           shutdown(int, int)  1,-1,1
+int           setsockopt(int, int, int, const void *, socklen_t)  1,-1,1
+int           getsockopt(int, int, int, void *, socklen_t *)    1,-1,1
+int           sendmsg(int, const struct msghdr *, unsigned int)  1,-1,1
+int           recvmsg(int, struct msghdr *, unsigned int)   1,-1,1

 # sockets for x86. These are done as an "indexed" call to socketcall syscall.
-int           socket:socketcall:1 (int, int, int) -1,102,-1
-int           bind:socketcall:2 (int, struct sockaddr *, int)  -1,102,-1
-int           connect:socketcall:3(int, struct sockaddr *, socklen_t)   -1,102,-1
-int           listen:socketcall:4(int, int)                   -1,102,-1
-int           accept:socketcall:5(int, struct sockaddr *, socklen_t *)  -1,102,-1
-int           getsockname:socketcall:6(int, struct sockaddr *, socklen_t *)  -1,102,-1
-int           getpeername:socketcall:7(int, struct sockaddr *, socklen_t *)  -1,102,-1
-int           socketpair:socketcall:8(int, int, int, int*)    -1,102,-1
-int           sendto:socketcall:11(int, const void *, size_t, int, const struct sockaddr *, socklen_t)  -1,102,-1
-int           recvfrom:socketcall:12(int, void *, size_t, unsigned int, struct sockaddr *, socklen_t *)  -1,102,-1
-int           shutdown:socketcall:13(int, int)  -1,102,-1
-int           setsockopt:socketcall:14(int, int, int, const void *, socklen_t)  -1,102,-1
-int           getsockopt:socketcall:15(int, int, int, void *, socklen_t *)    -1,102,-1
-int           sendmsg:socketcall:16(int, const struct msghdr *, unsigned int)  -1,102,-1
-int           recvmsg:socketcall:17(int, struct msghdr *, unsigned int)   -1,102,-1
+int           socket:socketcall:1 (int, int, int) -1,1,-1
+int           bind:socketcall:2 (int, struct sockaddr *, int)  -1,1,-1
+int           connect:socketcall:3(int, struct sockaddr *, socklen_t)   -1,1,-1
+int           listen:socketcall:4(int, int)                   -1,1,-1
+int           accept:socketcall:5(int, struct sockaddr *, socklen_t *)  -1,1,-1
+int           getsockname:socketcall:6(int, struct sockaddr *, socklen_t *)  -1,1,-1
+int           getpeername:socketcall:7(int, struct sockaddr *, socklen_t *)  -1,1,-1
+int           socketpair:socketcall:8(int, int, int, int*)    -1,1,-1
+int           sendto:socketcall:11(int, const void *, size_t, int, const struct sockaddr *, socklen_t)  -1,1,-1
+int           recvfrom:socketcall:12(int, void *, size_t, unsigned int, struct sockaddr *, socklen_t *)  -1,1,-1
+int           shutdown:socketcall:13(int, int)  -1,1,-1
+int           setsockopt:socketcall:14(int, int, int, const void *, socklen_t)  -1,1,-1
+int           getsockopt:socketcall:15(int, int, int, void *, socklen_t *)    -1,1,-1
+int           sendmsg:socketcall:16(int, const struct msghdr *, unsigned int)  -1,1,-1
+int           recvmsg:socketcall:17(int, struct msghdr *, unsigned int)   -1,1,-1

 # scheduler & real-time
-int sched_setscheduler(pid_t pid, int policy, const struct sched_param *param)  156,156,160
-int sched_getscheduler(pid_t pid)  157,157,161
-int sched_yield(void)  158,158,162
-int sched_setparam(pid_t pid, const struct sched_param *param)  154,154,158
-int sched_getparam(pid_t pid, struct sched_param *param)  155,155,159
-int sched_get_priority_max(int policy)  159,159,163
-int sched_get_priority_min(int policy)  160,160,164
-int sched_rr_get_interval(pid_t pid, struct timespec *interval)  161,161,165
-int sched_setaffinity(pid_t pid, size_t setsize, const cpu_set_t* set) 241,241,239
-int __sched_getaffinity:sched_getaffinity(pid_t pid, size_t setsize, cpu_set_t* set)  242,242,240
-int __getcpu:getcpu(unsigned *cpu, unsigned *node, void *unused) 345,318,312
+int sched_setscheduler(pid_t pid, int policy, const struct sched_param *param)  1
+int sched_getscheduler(pid_t pid)  1
+int sched_yield(void)  1
+int sched_setparam(pid_t pid, const struct sched_param *param)  1
+int sched_getparam(pid_t pid, struct sched_param *param)  1
+int sched_get_priority_max(int policy)  1
+int sched_get_priority_min(int policy)  1
+int sched_rr_get_interval(pid_t pid, struct timespec *interval)  1
+int sched_setaffinity(pid_t pid, size_t setsize, const cpu_set_t* set) 1
+int __sched_getaffinity:sched_getaffinity(pid_t pid, size_t setsize, cpu_set_t* set)  1
+int __getcpu:getcpu(unsigned *cpu, unsigned *node, void *unused) 1

 # io priorities
-int ioprio_set(int which, int who, int ioprio) 314,289,314
-int ioprio_get(int which, int who) 315,290,315
+int ioprio_set(int which, int who, int ioprio) 1
+int ioprio_get(int which, int who) 1

 # other
-int     uname(struct utsname *)  122
-pid_t   __wait4:wait4(pid_t pid, int *status, int options, struct rusage *rusage)   114
-mode_t  umask(mode_t)  60
-int      __reboot:reboot(int, int, int, void *)  88
-int     __syslog:syslog(int, char *, int)  103
-int     init_module(void *, unsigned long, const char *)  128
-int     delete_module(const char*, unsigned int)   129
-int     klogctl:syslog(int, char *, int)   103
-int     sysinfo(struct sysinfo *)  116
-int     personality(unsigned long)  136
-long    perf_event_open(struct perf_event_attr *attr_uptr, pid_t pid, int cpu, int group_fd, unsigned long flags) 364,336,333
+int     uname(struct utsname *)  1
+mode_t  umask(mode_t)  1
+int      __reboot:reboot(int, int, int, void *)  1
+int     __syslog:syslog(int, char *, int)  1
+int     init_module(void *, unsigned long, const char *)  1
+int     delete_module(const char*, unsigned int)   1
+int     klogctl:syslog(int, char *, int)   1
+int     sysinfo(struct sysinfo *)  1
+int     personality(unsigned long)  1
+long    perf_event_open(struct perf_event_attr *attr_uptr, pid_t pid, int cpu, int group_fd, unsigned long flags) 1

 # futex
-int	futex(void *, int, int, void *, void *, int) 240,240,238
+int	futex(void *, int, int, void *, void *, int) 1

 # epoll
-int     epoll_create(int size)     250,254,248
-int     epoll_ctl(int epfd, int op, int fd, struct epoll_event *event)    251,255,249
-int     epoll_wait(int epfd, struct epoll_event *events, int max, int timeout)   252,256,250
+int     epoll_create(int size)     1
+int     epoll_ctl(int epfd, int op, int fd, struct epoll_event *event)    1
+int     epoll_wait(int epfd, struct epoll_event *events, int max, int timeout)   1

-int     inotify_init(void)      316,291,284
-int     inotify_add_watch(int, const char *, unsigned int)  317,292,285
-int     inotify_rm_watch(int, unsigned int)  318,293,286
+int     inotify_init(void)      1
+int     inotify_add_watch(int, const char *, unsigned int)  1
+int     inotify_rm_watch(int, unsigned int)  1

-int     poll(struct pollfd *, unsigned int, long)  168,168,188
+int     poll(struct pollfd *, unsigned int, long)  1

-int     eventfd:eventfd2(unsigned int, int)  356,328,325
+int     eventfd:eventfd2(unsigned int, int)  1

 # ARM-specific ARM_NR_BASE == 0x0f0000 == 983040
-int     __set_tls:ARM_set_tls(void*)                                 983045,-1,-1
-int     cacheflush:ARM_cacheflush(long start, long end, long flags)  983042,-1,-1
+int     __set_tls:__ARM_NR_set_tls(void*)                                 1,-1,-1
+int     cacheflush:__ARM_NR_cacheflush(long start, long end, long flags)  1,-1,-1

 # MIPS-specific
-int	_flush_cache:cacheflush(char *addr, const int nbytes, const int op)	-1,-1,147
-int	syscall(int number,...) -1,-1,0
+int     _flush_cache:cacheflush(char *addr, const int nbytes, const int op) -1,-1,1
+int     syscall(int number,...) -1,-1,1
--- a/libc/arch-arm/arm.mk
+++ b/libc/arch-arm/arm.mk
@@ -0,0 +1,35 @@
+_LIBC_ARCH_COMMON_SRC_FILES := \
+    arch-arm/bionic/abort_arm.S \
+    arch-arm/bionic/atomics_arm.c \
+    arch-arm/bionic/clone.S \
+    arch-arm/bionic/eabi.c \
+    arch-arm/bionic/_exit_with_stack_teardown.S \
+    arch-arm/bionic/ffs.S \
+    arch-arm/bionic/futex_arm.S \
+    arch-arm/bionic/__get_sp.S \
+    arch-arm/bionic/kill.S \
+    arch-arm/bionic/libgcc_compat.c \
+    arch-arm/bionic/memcmp16.S \
+    arch-arm/bionic/memcmp.S \
+    arch-arm/bionic/_setjmp.S \
+    arch-arm/bionic/setjmp.S \
+    arch-arm/bionic/sigsetjmp.S \
+    arch-arm/bionic/strcpy.S \
+    arch-arm/bionic/strlen.c.arm \
+    arch-arm/bionic/syscall.S \
+    arch-arm/bionic/tgkill.S \
+    arch-arm/bionic/tkill.S \
+
+# These are used by the static and dynamic versions of the libc
+# respectively.
+_LIBC_ARCH_STATIC_SRC_FILES := \
+    arch-arm/bionic/exidx_static.c
+
+_LIBC_ARCH_DYNAMIC_SRC_FILES := \
+    arch-arm/bionic/exidx_dynamic.c
+
+ifeq ($(strip $(wildcard bionic/libc/arch-arm/$(TARGET_CPU_VARIANT)/$(TARGET_CPU_VARIANT).mk)),)
+$(error "TARGET_CPU_VARIANT not set or set to an unknown value. Possible values are cortex-a9, cortex-a15, krait. Use generic for devices that do not have a CPU similar to any of the supported cpu variants.")
+endif
+
+include bionic/libc/arch-arm/$(TARGET_CPU_VARIANT)/$(TARGET_CPU_VARIANT).mk
--- a/libc/arch-arm/bionic/_exit_with_stack_teardown.S
+++ b/libc/arch-arm/bionic/_exit_with_stack_teardown.S
@@ -27,7 +27,7 @@
 */

 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>
+#include <asm/unistd.h>

 // void _exit_with_stack_teardown(void* stackBase, int stackSize, int retCode)
 ENTRY(_exit_with_stack_teardown)
--- a/libc/arch-arm/bionic/atexit_legacy.c
+++ b/libc/arch-arm/bionic/atexit_legacy.c
@@ -27,7 +27,7 @@
 */

 #include <sys/types.h>
-#include <private/logd.h>
+#include <private/libc_logging.h>
 #include <stdio.h>

 /*
@@ -52,7 +52,7 @@ atexit(void (*func)(void))
     */
    static char const warning[] = "WARNING: generic atexit() called from legacy shared library\n";

-    __libc_android_log_write(ANDROID_LOG_WARN, "libc", warning);
+    __libc_format_log(ANDROID_LOG_WARN, "libc", warning);
    fprintf(stderr, warning);

    return (__cxa_atexit((void (*)(void *))func, NULL, NULL));
--- a/libc/arch-arm/bionic/clone.S
+++ b/libc/arch-arm/bionic/clone.S
@@ -26,13 +26,14 @@
 * SUCH DAMAGE.
 */

+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>
+#include <asm/unistd.h>

-// int  __pthread_clone(int (*fn)(void*), void* child_stack, int flags, void* arg);
+// int  __pthread_clone(void* (*fn)(void*), void* child_stack, int flags, void* arg);
 ENTRY(__pthread_clone)
    # Copy the args onto the new stack.
-    stmdb r1!, {r0, r3}
+    stmdb   r1!, {r0, r3}

    # The sys_clone system call only takes two arguments: 'flags' and 'child_stack'.
    # 'child_stack' is already in r1, but we need to move 'flags' into position.
@@ -45,10 +46,12 @@ ENTRY(__pthread_clone)
    movs    r0, r0
    beq     1f

-    # In parent, reload saved registers then either exit or set errno.
+    # In parent, reload saved registers then either return or set errno.
    ldmfd   sp!, {r4, r7}
-    bxne    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno

 1:  # The child.
    # pick the function arg and call address off the stack and jump
@@ -94,10 +97,12 @@ ENTRY(__bionic_clone)
    movs    r0, r0
    beq     1f

-    # In the parent, reload saved registers then either exit or set errno.
+    # In the parent, reload saved registers then either return or set errno.
    ldmfd   sp!, {r4, r5, r6, r7}
-    bxne    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno

 1:  # The child.
    ldr    r0, [sp, #-4]
--- a/libc/arch-arm/bionic/ffs.S
+++ b/libc/arch-arm/bionic/ffs.S
@@ -31,55 +31,11 @@
 #include <machine/asm.h>
 #include <machine/cpu-features.h>

-/*
- * ffs - find first set bit, this algorithm isolates the first set
- * bit, then multiplies the number by 0x0450fbaf which leaves the top
- * 6 bits as an index into the table.  This algorithm should be a win
- * over the checking each bit in turn as per the C compiled version.
- *
- * Some newer ARM architectures have an instruction named
- * CLZ (count leading Zero's) that is used
- *
- * This is the ffs algorithm devised by d.seal and posted to comp.sys.arm on
- * 16 Feb 1994.
- */
-
 ENTRY(ffs)
 	/* Standard trick to isolate bottom bit in r0 or 0 if r0 = 0 on entry */
 	rsb     r1, r0, #0
 	ands    r0, r0, r1
-#ifndef __ARM_HAVE_CLZ
-	/*
-	 * now r0 has at most one set bit, call this X
-	 * if X = 0, all further instructions are skipped
-	 */
-	adrne   r2, .L_ffs_table
-	orrne   r0, r0, r0, lsl #4  /* r0 = X * 0x11 */ 
-	orrne   r0, r0, r0, lsl #6  /* r0 = X * 0x451 */
-	rsbne   r0, r0, r0, lsl #16 /* r0 = X * 0x0450fbaf */
-              
-	/* now lookup in table indexed on top 6 bits of r0 */
-	ldrneb  r0, [ r2, r0, lsr #26 ]
-
-	bx		lr
-END(ffs)
-
-.text;
-.type .L_ffs_table, _ASM_TYPE_OBJECT;
-.L_ffs_table:
-/*               0   1   2   3   4   5   6   7           */
-	.byte	 0,  1,  2, 13,  3,  7,  0, 14  /*  0- 7 */
-	.byte	 4,  0,  8,  0,  0,  0,  0, 15  /*  8-15 */
-	.byte	11,  5,  0,  0,  9,  0,  0, 26  /* 16-23 */
-	.byte	 0,  0,  0,  0,  0, 22, 28, 16  /* 24-31 */
-	.byte	32, 12,  6,  0,  0,  0,  0,  0	/* 32-39 */
-	.byte	10,  0,  0, 25,  0,  0, 21, 27  /* 40-47 */
-	.byte	31,  0,  0,  0,  0, 24,  0, 20  /* 48-55 */
-	.byte   30,  0, 23, 19, 29, 18, 17,  0  /* 56-63 */
-#else /* !defined(__ARM_HAVE_CLZ) */
 	clzne	r0, r0
 	rsbne	r0, r0, #32
 	bx		lr
 END(ffs)
-#endif /* !defined(__ARM_HAVE_CLZ) */
-
--- a/libc/arch-arm/bionic/futex_arm.S
+++ b/libc/arch-arm/bionic/futex_arm.S
@@ -26,7 +26,7 @@
 * SUCH DAMAGE.
 */

-#include <sys/linux-syscalls.h>
+#include <asm/unistd.h>
 #include <machine/asm.h>

 #define FUTEX_WAIT 0
--- a/libc/arch-arm/bionic/kill.S
+++ b/libc/arch-arm/bionic/kill.S
@@ -25,6 +25,11 @@
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */
+
+#include <linux/err.h>
+#include <asm/unistd.h>
+#include <machine/asm.h>
+
 /* unlike our auto-generated syscall stubs, this code saves lr
   on the stack, as well as a few other registers. this makes
   our stack unwinder happy, when we generate debug stack
@@ -32,19 +37,14 @@
   abort due to a fatal runtime error (e.g. detection
   of a corrupted malloc heap).
 */
-#include <sys/linux-syscalls.h>
-#include <machine/asm.h>
-
-#ifndef __NR_kill
-#define __NR_kill   37
-#endif

 ENTRY(kill)
    stmfd   sp!, {r4-r7, ip, lr}
    ldr     r7, =__NR_kill
    swi     #0
    ldmfd   sp!, {r4-r7, ip, lr}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(kill)
--- a/libc/arch-arm/bionic/memcmp.S
+++ b/libc/arch-arm/bionic/memcmp.S
@@ -115,7 +115,7 @@ ENTRY(memcmp)
         * pointer somewhere else
         */
         mov        r4, r0
-        
+
        /* align first pointer to word boundary
         * offset = -src & 3
         */
@@ -151,7 +151,7 @@ ENTRY(memcmp)
        ldr         ip, [r1]
        subs        r2, r2, #(32 + 4)
        bmi         1f
-        
+
 0:      pld         [r4, #(CACHE_LINE_SIZE * 2)]
        pld         [r1, #(CACHE_LINE_SIZE * 2)]
        ldr         r0, [r4], #4
@@ -178,14 +178,14 @@ ENTRY(memcmp)
        ldreq       r0, [r4], #4
        ldreq       ip, [r1, #4]!
        eoreqs      r0, r0, lr
-        bne         2f        
+        bne         2f
        subs        r2, r2, #32
        bhs         0b

        /* do we have at least 4 bytes left? */
 1:      adds        r2, r2, #(32 - 4 + 4)
        bmi         4f
-        
+
        /* finish off 4 bytes at a time */
 3:      ldr         r0, [r4], #4
        ldr         ip, [r1], #4
@@ -233,17 +233,14 @@ ENTRY(memcmp)
        subs        r2, r2, #1
        bne         11b
        bx          lr
-END(memcmp)
-
-

 5:      /*************** non-congruent case ***************/
-        and         r0, r1, #3      
+        and         r0, r1, #3
        cmp         r0, #2
        bne         4f

        /* here, offset is 2 (16-bits aligned, special cased) */
-        
+
        /* make sure we have at least 16 bytes to process */
        subs        r2, r2, #16
        addmi       r2, r2, #16
@@ -341,3 +338,4 @@ END(memcmp)
        mov         r2, #4
 		ldmfd		sp!, {r5, r6, r7}
        b           8b
+END(memcmp)
--- a/libc/arch-arm/bionic/memcpy.a15.S
+++ b/libc/arch-arm/bionic/memcpy.a15.S
@@ -0,0 +1,406 @@
+/*
+ * Copyright (c) 2013 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+    /* Prototype: void *memcpy (void *dst, const void *src, size_t count).  */
+
+        /* Use the version of memcpy implemented using LDRD and STRD.
+           This version is tuned for Cortex-A15.
+           This might not be the best for other ARMv7-A CPUs,
+           but there is no predefine to distinguish between
+           different CPUs in the same architecture,
+           and this version is better than the plain memcpy provided in newlib.
+
+           Therefore, we use this version for all ARMv7-A CPUS.  */
+
+        /* To make the same code compile for both ARM and Thumb instruction
+       sets, switch to unified syntax at the beginning of this function.
+           However, by using the same code, we may be missing optimization
+       opportunities.  For instance, in LDRD/STRD instructions, the first
+       destination register must be even and the second consecutive in
+       ARM state, but not in Thumb state.  */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+        .syntax         unified
+
+ENTRY(memcpy)
+
+       /* Assumes that n >= 0, and dst, src are valid pointers.
+          If there is at least 8 bytes to copy, use LDRD/STRD.
+          If src and dst are misaligned with different offsets,
+          first copy byte by byte until dst is aligned,
+          and then copy using LDRD/STRD and shift if needed.
+          When less than 8 left, copy a word and then byte by byte.  */
+
+       /* Save registers (r0 holds the return value):
+          optimized push {r0, r4, r5, r6, r7, lr}.
+          To try and improve performance, stack layout changed,
+          i.e., not keeping the stack looking like users expect
+          (highest numbered register at highest address).  */
+        .save   {r0, lr}
+        push    {r0, lr}
+        .save   {r4, r5}
+        strd    r4, r5, [sp, #-8]!
+        .save   {r6, r7}
+        strd    r6, r7, [sp, #-8]!
+
+       /* TODO: Add debug frame directives.
+          We don't need exception unwind directives, because the code below
+      does not throw any exceptions and does not call any other functions.
+          Generally, newlib functions like this lack debug information for
+      assembler source.  */
+
+        /* Get copying of tiny blocks out of the way first.  */
+        /* Is there at least 4 bytes to copy?  */
+        subs    r2, r2, #4
+        blt     copy_less_than_4                 /* If n < 4.  */
+
+        /* Check word alignment.  */
+        ands    ip, r0, #3                       /* ip = last 2 bits of dst.  */
+        bne     dst_not_word_aligned             /* If dst is not word-aligned.  */
+
+        /* Get here if dst is word-aligned.  */
+        ands    ip, r1, #3                      /* ip = last 2 bits of src.  */
+        bne     src_not_word_aligned            /* If src is not word-aligned.  */
+word_aligned:
+        /* Get here if source and dst both are word-aligned.
+           The number of bytes remaining to copy is r2+4.  */
+
+        /* Is there is at least 64 bytes to copy?  */
+        subs    r2, r2, #60
+        blt     copy_less_than_64                /* If r2 + 4 < 64.  */
+
+        /* First, align the destination buffer to 8-bytes,
+           to make sure double loads and stores don't cross cache line boundary,
+           as they are then more expensive even if the data is in the cache
+           (require two load/store issue cycles instead of one).
+           If only one of the buffers is not 8-bytes aligned,
+           then it's more important to align dst than src,
+           because there is more penalty for stores
+           than loads that cross cacheline boundary.
+           This check and realignment are only worth doing
+           if there is a lot to copy.  */
+
+        /* Get here if dst is word aligned,
+           i.e., the 2 least significant bits are 0.
+           If dst is not 2w aligned (i.e., the 3rd bit is not set in dst),
+           then copy 1 word (4 bytes).  */
+        ands    r3, r0, #4
+        beq     11f                  /* If dst already two-word aligned.  */
+        ldr     r3, [r1], #4
+        str     r3, [r0], #4
+        subs    r2, r2, #4
+        blt     copy_less_than_64
+
+11:
+        /* TODO: Align to cacheline (useful for PLD optimization).  */
+
+        /* Every loop iteration copies 64 bytes.  */
+1:
+        .irp    offset, #0, #8, #16, #24, #32, #40, #48, #56
+        ldrd    r4, r5, [r1, \offset]
+        strd    r4, r5, [r0, \offset]
+        .endr
+
+        add     r0, r0, #64
+        add     r1, r1, #64
+        subs    r2, r2, #64
+        bge     1b                            /* If there is more to copy.  */
+
+copy_less_than_64:
+
+        /* Get here if less than 64 bytes to copy, -64 <= r2 < 0.
+           Restore the count if there is more than 7 bytes to copy.  */
+        adds    r2, r2, #56
+        blt     copy_less_than_8
+
+        /* Copy 8 bytes at a time.  */
+2:
+        ldrd    r4, r5, [r1], #8
+        strd    r4, r5, [r0], #8
+        subs    r2, r2, #8
+        bge     2b                            /* If there is more to copy.  */
+
+copy_less_than_8:
+
+        /* Get here if less than 8 bytes to copy, -8 <= r2 < 0.
+           Check if there is more to copy.  */
+        cmn     r2, #8
+        beq     return                          /* If r2 + 8 == 0.  */
+
+        /* Restore the count if there is more than 3 bytes to copy.  */
+        adds    r2, r2, #4
+        blt     copy_less_than_4
+
+        /* Copy 4 bytes.  */
+        ldr     r3, [r1], #4
+        str     r3, [r0], #4
+
+copy_less_than_4:
+        /* Get here if less than 4 bytes to copy, -4 <= r2 < 0.  */
+
+        /* Restore the count, check if there is more to copy.  */
+        adds    r2, r2, #4
+        beq     return                          /* If r2 == 0.  */
+
+        /* Get here with r2 is in {1,2,3}={01,10,11}.  */
+        /* Logical shift left r2, insert 0s, update flags.  */
+        lsls    r2, r2, #31
+
+        /* Copy byte by byte.
+           Condition ne means the last bit of r2 is 0.
+           Condition cs means the second to last bit of r2 is set,
+           i.e., r2 is 1 or 3.  */
+        itt     ne
+        ldrbne  r3, [r1], #1
+        strbne  r3, [r0], #1
+
+        itttt   cs
+        ldrbcs  r4, [r1], #1
+        ldrbcs  r5, [r1]
+        strbcs  r4, [r0], #1
+        strbcs  r5, [r0]
+
+return:
+        /* Restore registers: optimized pop {r0, r4, r5, r6, r7, pc}   */
+        /* This is the only return point of memcpy.  */
+        ldrd r6, r7, [sp], #8
+        ldrd r4, r5, [sp], #8
+        pop {r0, pc}
+
+#ifndef __ARM_FEATURE_UNALIGNED
+
+       /* The following assembly macro implements misaligned copy in software.
+          Assumes that dst is word aligned, src is at offset "pull" bits from
+      word, push = 32 - pull, and the number of bytes that remain to copy
+      is r2 + 4, r2 >= 0.  */
+
+       /* In the code below, r2 is the number of bytes that remain to be
+      written.  The number of bytes read is always larger, because we have
+      partial words in the shift queue.  */
+
+        .macro  miscopy pull push shiftleft shiftright
+
+        /* Align src to the previous word boundary.  */
+        bic     r1, r1, #3
+
+        /* Initialize the shift queue.  */
+        ldr     r5, [r1], #4                   /* Load a word from source.  */
+
+        subs    r2, r2, #4
+        blt     6f          /* Go to misaligned copy of less than 8 bytes.  */
+
+       /* Get here if there is more than 8 bytes to copy.
+          The number of bytes to copy is r2+8, r2 >= 0.  */
+
+       subs     r2, r2, #56
+       blt      4f         /* Go to misaligned copy of less than 64 bytes.  */
+
+3:
+       /* Get here if there is more than 64 bytes to copy.
+          The number of bytes to copy is r2+64, r2 >= 0.  */
+
+       /* Copy 64 bytes in every iteration.
+          Use a partial word from the shift queue.  */
+        .irp    offset, #0, #8, #16, #24, #32, #40, #48, #56
+        mov     r6, r5, \shiftleft #\pull
+        ldrd    r4, r5, [r1, \offset]
+        orr     r6, r6, r4, \shiftright #\push
+        mov     r7, r4, \shiftleft #\pull
+        orr     r7, r7, r5, \shiftright #\push
+        strd    r6, r7, [r0, \offset]
+        .endr
+
+        add     r1, r1, #64
+        add     r0, r0, #64
+        subs    r2, r2, #64
+        bge     3b
+
+4:
+       /* Get here if there is less than 64 bytes to copy (-64 <= r2 < 0)
+      and they are misaligned.  */
+
+       /* Restore the count if there is more than 7 bytes to copy.  */
+        adds    r2, r2, #56
+
+        blt     6f          /* Go to misaligned copy of less than 8 bytes.  */
+
+5:
+        /* Copy 8 bytes at a time.
+           Use a partial word from the shift queue.  */
+        mov     r6, r5, \shiftleft #\pull
+        ldrd    r4, r5, [r1], #8
+        orr     r6, r6, r4, \shiftright #\push
+        mov     r7, r4, \shiftleft #\pull
+        orr     r7, r7, r5, \shiftright #\push
+        strd    r6, r7, [r0], #8
+
+        subs    r2, r2, #8
+        bge     5b                        /* If there is more to copy.  */
+
+6:
+        /* Get here if there less than 8 bytes to copy (-8 <= r2 < 0)
+           and they are misaligned.  */
+
+        /* Check if there is more to copy.  */
+        cmn     r2, #8
+        beq     return
+
+        /* Check if there is less than 4 bytes to copy.  */
+        cmn     r2, #4
+
+        itt     lt
+        /* Restore src offset from word-align.  */
+        sublt   r1, r1, #(\push / 8)
+        blt     copy_less_than_4
+
+        /* Use a partial word from the shift queue.  */
+        mov     r3, r5, \shiftleft #\pull
+        /* Load a word from src, but without writeback
+           (this word is not fully written to dst).  */
+        ldr     r5, [r1]
+
+        /* Restore src offset from word-align.  */
+        add     r1, r1, #(\pull / 8)
+
+        /* Shift bytes to create one dst word and store it.  */
+        orr     r3, r3, r5, \shiftright #\push
+        str     r3, [r0], #4
+
+        /* Use single byte copying of the remaining bytes.  */
+        b       copy_less_than_4
+
+        .endm
+
+#endif /* not __ARM_FEATURE_UNALIGNED  */
+
+dst_not_word_aligned:
+
+       /* Get here when dst is not aligned and ip has the last 2 bits of dst,
+          i.e., ip is the offset of dst from word.
+          The number of bytes that remains to copy is r2 + 4,
+          i.e., there are at least 4 bytes to copy.
+          Write a partial word (0 to 3 bytes), such that dst becomes
+      word-aligned.  */
+
+       /* If dst is at ip bytes offset from a word (with 0 < ip < 4),
+          then there are (4 - ip) bytes to fill up to align dst to the next
+      word.  */
+        rsb     ip, ip, #4                        /* ip = #4 - ip.  */
+        cmp     ip, #2
+
+       /* Copy byte by byte with conditionals.  */
+        itt     gt
+        ldrbgt  r3, [r1], #1
+        strbgt  r3, [r0], #1
+
+        itt     ge
+        ldrbge  r4, [r1], #1
+        strbge  r4, [r0], #1
+
+        ldrb    lr, [r1], #1
+        strb    lr, [r0], #1
+
+       /* Update the count.
+          ip holds the number of bytes we have just copied.  */
+        subs    r2, r2, ip                        /* r2 = r2 - ip.  */
+        blt     copy_less_than_4                  /* If r2 < ip.  */
+
+       /* Get here if there are more than 4 bytes to copy.
+          Check if src is aligned.  If beforehand src and dst were not word
+      aligned but congruent (same offset), then now they are both
+      word-aligned, and we can copy the rest efficiently (without
+      shifting).  */
+        ands    ip, r1, #3                    /* ip = last 2 bits of src.  */
+        beq     word_aligned                  /* If r1 is word-aligned.  */
+
+src_not_word_aligned:
+       /* Get here when src is not word-aligned, but dst is word-aligned.
+          The number of bytes that remains to copy is r2+4.  */
+
+#ifdef __ARM_FEATURE_UNALIGNED
+       /* Copy word by word using LDR when alignment can be done in hardware,
+          i.e., SCTLR.A is set, supporting unaligned access in LDR and STR.  */
+        subs    r2, r2, #60
+        blt     8f
+
+7:
+        /* Copy 64 bytes in every loop iteration.  */
+        .irp    offset, #0, #4, #8, #12, #16, #20, #24, #28, #32, #36, #40, #44, #48, #52, #56, #60
+        ldr     r3, [r1, \offset]
+        str     r3, [r0, \offset]
+        .endr
+
+        add     r0, r0, #64
+        add     r1, r1, #64
+        subs    r2, r2, #64
+        bge     7b
+
+8:
+        /* Get here if less than 64 bytes to copy, -64 <= r2 < 0.
+           Check if there is more than 3 bytes to copy.  */
+        adds    r2, r2, #60
+        blt     copy_less_than_4
+
+9:
+       /* Get here if there is less than 64 but at least 4 bytes to copy,
+          where the number of bytes to copy is r2+4.  */
+        ldr     r3, [r1], #4
+        str     r3, [r0], #4
+        subs    r2, r2, #4
+        bge     9b
+
+        b       copy_less_than_4
+
+#else /* not __ARM_FEATURE_UNALIGNED  */
+
+       /* ip has last 2 bits of src,
+          i.e., ip is the offset of src from word, and ip > 0.
+          Compute shifts needed to copy from src to dst.  */
+        cmp     ip, #2
+        beq     miscopy_16_16             /* If ip == 2.  */
+        bge     miscopy_24_8              /* If ip == 3.  */
+
+        /* Get here if ip == 1.  */
+
+        /* Endian independent macros for shifting bytes within registers.  */
+
+#ifndef __ARMEB__
+miscopy_8_24:   miscopy pull=8 push=24 shiftleft=lsr shiftright=lsl
+miscopy_16_16:  miscopy pull=16 push=16 shiftleft=lsr shiftright=lsl
+miscopy_24_8:   miscopy pull=24 push=8 shiftleft=lsr shiftright=lsl
+#else  /* not __ARMEB__ */
+miscopy_8_24:   miscopy pull=8 push=24 shiftleft=lsl shiftright=lsr
+miscopy_16_16:  miscopy pull=16 push=16 shiftleft=lsl shiftright=lsr
+miscopy_24_8:   miscopy pull=24 push=8 shiftleft=lsl shiftright=lsr
+#endif  /* not __ARMEB__ */
+
+#endif  /* not __ARM_FEATURE_UNALIGNED  */
+
+END(memcpy)
--- a/libc/arch-arm/bionic/memset.S
+++ b/libc/arch-arm/bionic/memset.S
@@ -42,6 +42,7 @@
 ENTRY(bzero)
        mov     r2, r1
        mov     r1, #0
+        // Fall through to memset...
 END(bzero)

 ENTRY(memset)
--- a/libc/arch-arm/bionic/strcmp.a15.S
+++ b/libc/arch-arm/bionic/strcmp.a15.S
@@ -0,0 +1,787 @@
+/*
+ * Copyright (c) 2013 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "arm_asm.h"
+
+#ifdef __ARMEB__
+#define S2LOMEM lsl
+#define S2LOMEMEQ lsleq
+#define S2HIMEM lsr
+#define MSB 0x000000ff
+#define LSB 0xff000000
+#define BYTE0_OFFSET 24
+#define BYTE1_OFFSET 16
+#define BYTE2_OFFSET 8
+#define BYTE3_OFFSET 0
+#else /* not  __ARMEB__ */
+#define S2LOMEM lsr
+#define S2LOMEMEQ lsreq
+#define S2HIMEM lsl
+#define BYTE0_OFFSET 0
+#define BYTE1_OFFSET 8
+#define BYTE2_OFFSET 16
+#define BYTE3_OFFSET 24
+#define MSB 0xff000000
+#define LSB 0x000000ff
+#endif /* not  __ARMEB__ */
+
+.syntax         unified
+
+#if defined (__thumb__)
+        .thumb
+        .thumb_func
+#endif
+        .global strcmp
+        .type   strcmp, %function
+strcmp:
+
+#if (defined (__thumb__) && !defined (__thumb2__))
+1:
+        ldrb    r2, [r0]
+        ldrb    r3, [r1]
+        adds    r0, r0, #1
+        adds    r1, r1, #1
+        cmp     r2, #0
+        beq     2f
+        cmp     r2, r3
+        beq     1b
+2:
+        subs    r0, r2, r3
+        bx      lr
+#elif (defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED))
+1:
+        ldrb    r2, [r0], #1
+        ldrb    r3, [r1], #1
+        cmp     r2, #1
+        it      cs
+        cmpcs   r2, r3
+        beq     1b
+        subs    r0, r2, r3
+        RETURN
+
+
+#elif (defined (_ISA_THUMB_2) || defined (_ISA_ARM_6))
+      /* Use LDRD whenever possible.  */
+
+/* The main thing to look out for when comparing large blocks is that
+   the loads do not cross a page boundary when loading past the index
+   of the byte with the first difference or the first string-terminator.
+
+   For example, if the strings are identical and the string-terminator
+   is at index k, byte by byte comparison will not load beyond address
+   s1+k and s2+k; word by word comparison may load up to 3 bytes beyond
+   k; double word - up to 7 bytes.  If the load of these bytes crosses
+   a page boundary, it might cause a memory fault (if the page is not mapped)
+   that would not have happened in byte by byte comparison.
+
+   If an address is (double) word aligned, then a load of a (double) word
+   from that address will not cross a page boundary.
+   Therefore, the algorithm below considers word and double-word alignment
+   of strings separately.  */
+
+/* High-level description of the algorithm.
+
+   * The fast path: if both strings are double-word aligned,
+     use LDRD to load two words from each string in every loop iteration.
+   * If the strings have the same offset from a word boundary,
+     use LDRB to load and compare byte by byte until
+     the first string is aligned to a word boundary (at most 3 bytes).
+     This is optimized for quick return on short unaligned strings.
+   * If the strings have the same offset from a double-word boundary,
+     use LDRD to load two words from each string in every loop iteration, as in the fast path.
+   * If the strings do not have the same offset from a double-word boundary,
+     load a word from the second string before the loop to initialize the queue.
+     Use LDRD to load two words from every string in every loop iteration.
+     Inside the loop, load the second word from the second string only after comparing
+     the first word, using the queued value, to guarantee safety across page boundaries.
+   * If the strings do not have the same offset from a word boundary,
+     use LDR and a shift queue. Order of loads and comparisons matters,
+     similarly to the previous case.
+
+   * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value.
+   * The only difference between ARM and Thumb modes is the use of CBZ instruction.
+   * The only difference between big and little endian is the use of REV in little endian
+     to compute the return value, instead of MOV.
+   * No preload. [TODO.]
+*/
+
+        .macro m_cbz reg label
+#ifdef __thumb2__
+        cbz     \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        beq     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbz */
+
+        .macro m_cbnz reg label
+#ifdef __thumb2__
+        cbnz    \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        bne     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbnz */
+
+        .macro  init
+        /* Macro to save temporary registers and prepare magic values.  */
+        subs    sp, sp, #16
+        strd    r4, r5, [sp, #8]
+        strd    r6, r7, [sp]
+        mvn     r6, #0  /* all F */
+        mov     r7, #0  /* all 0 */
+        .endm   /* init */
+
+        .macro  magic_compare_and_branch w1 w2 label
+        /* Macro to compare registers w1 and w2 and conditionally branch to label.  */
+        cmp     \w1, \w2        /* Are w1 and w2 the same?  */
+        magic_find_zero_bytes \w1
+        it      eq
+        cmpeq   ip, #0          /* Is there a zero byte in w1?  */
+        bne     \label
+        .endm /* magic_compare_and_branch */
+
+        .macro  magic_find_zero_bytes w1
+        /* Macro to find all-zero bytes in w1, result is in ip.  */
+#if (defined (__ARM_FEATURE_DSP))
+        uadd8   ip, \w1, r6
+        sel     ip, r7, r6
+#else /* not defined (__ARM_FEATURE_DSP) */
+        /* __ARM_FEATURE_DSP is not defined for some Cortex-M processors.
+        Coincidently, these processors only have Thumb-2 mode, where we can use the
+        the (large) magic constant available directly as an immediate in instructions.
+        Note that we cannot use the magic constant in ARM mode, where we need
+        to create the constant in a register.  */
+        sub     ip, \w1, #0x01010101
+        bic     ip, ip, \w1
+        and     ip, ip, #0x80808080
+#endif /* not defined (__ARM_FEATURE_DSP) */
+        .endm /* magic_find_zero_bytes */
+
+        .macro  setup_return w1 w2
+#ifdef __ARMEB__
+        mov     r1, \w1
+        mov     r2, \w2
+#else /* not  __ARMEB__ */
+        rev     r1, \w1
+        rev     r2, \w2
+#endif /* not  __ARMEB__ */
+        .endm /* setup_return */
+
+        /*
+        optpld r0, #0
+        optpld r1, #0
+        */
+
+        /* Are both strings double-word aligned?  */
+        orr     ip, r0, r1
+        tst     ip, #7
+        bne     do_align
+
+        /* Fast path.  */
+        init
+
+doubleword_aligned:
+
+        /* Get here when the strings to compare are double-word aligned.  */
+        /* Compare two words in every iteration.  */
+        .p2align        2
+2:
+        /*
+        optpld r0, #16
+        optpld r1, #16
+        */
+
+        /* Load the next double-word from each string.  */
+        ldrd    r2, r3, [r0], #8
+        ldrd    r4, r5, [r1], #8
+
+        magic_compare_and_branch w1=r2, w2=r4, label=return_24
+        magic_compare_and_branch w1=r3, w2=r5, label=return_35
+        b       2b
+
+do_align:
+        /* Is the first string word-aligned?  */
+        ands    ip, r0, #3
+        beq     word_aligned_r0
+
+        /* Fast compare byte by byte until the first string is word-aligned.  */
+        /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes
+        to read until the next word boudnary is 4-ip.  */
+        bic     r0, r0, #3
+        ldr     r2, [r0], #4
+        lsls    ip, ip, #31
+        beq     byte2
+        bcs     byte3
+
+byte1:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE1_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte2:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE2_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte3:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE3_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbnz  reg=r3, label=word_aligned_r0
+
+fast_return:
+        mov     r0, ip
+        bx      lr
+
+word_aligned_r0:
+        init
+        /* The first string is word-aligned.  */
+        /* Is the second string word-aligned?  */
+        ands    ip, r1, #3
+        bne     strcmp_unaligned
+
+word_aligned:
+        /* The strings are word-aligned. */
+        /* Is the first string double-word aligned?  */
+        tst     r0, #4
+        beq     doubleword_aligned_r0
+
+        /* If r0 is not double-word aligned yet, align it by loading
+        and comparing the next word from each string.  */
+        ldr     r2, [r0], #4
+        ldr     r4, [r1], #4
+        magic_compare_and_branch w1=r2 w2=r4 label=return_24
+
+doubleword_aligned_r0:
+        /* Get here when r0 is double-word aligned.  */
+        /* Is r1 doubleword_aligned?  */
+        tst     r1, #4
+        beq     doubleword_aligned
+
+        /* Get here when the strings to compare are word-aligned,
+        r0 is double-word aligned, but r1 is not double-word aligned.  */
+
+        /* Initialize the queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare two words in every iteration.  */
+        .p2align        2
+3:
+        /*
+        optpld r0, #16
+        optpld r1, #16
+        */
+
+        /* Load the next double-word from each string and compare.  */
+        ldrd    r2, r3, [r0], #8
+        magic_compare_and_branch w1=r2 w2=r5 label=return_25
+        ldrd    r4, r5, [r1], #8
+        magic_compare_and_branch w1=r3 w2=r4 label=return_34
+        b       3b
+
+        .macro miscmp_word offsetlo offsethi
+        /* Macro to compare misaligned strings.  */
+        /* r0, r1 are word-aligned, and at least one of the strings
+        is not double-word aligned.  */
+        /* Compare one word in every loop iteration.  */
+        /* OFFSETLO is the original bit-offset of r1 from a word-boundary,
+        OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word).  */
+
+        /* Initialize the shift queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare one word from each string in every loop iteration.  */
+        .p2align        2
+7:
+        ldr     r3, [r0], #4
+        S2LOMEM r5, r5, #\offsetlo
+        magic_find_zero_bytes w1=r3
+        cmp     r7, ip, S2HIMEM #\offsetlo
+        and     r2, r3, r6, S2LOMEM #\offsetlo
+        it      eq
+        cmpeq   r2, r5
+        bne     return_25
+        ldr     r5, [r1], #4
+        cmp     ip, #0
+        eor r3, r2, r3
+        S2HIMEM r2, r5, #\offsethi
+        it      eq
+        cmpeq   r3, r2
+        bne     return_32
+        b       7b
+        .endm /* miscmp_word */
+
+strcmp_unaligned:
+        /* r0 is word-aligned, r1 is at offset ip from a word.  */
+        /* Align r1 to the (previous) word-boundary.  */
+        bic     r1, r1, #3
+
+        /* Unaligned comparison word by word using LDRs. */
+        cmp     ip, #2
+        beq     miscmp_word_16                    /* If ip == 2.  */
+        bge     miscmp_word_24                    /* If ip == 3.  */
+        miscmp_word offsetlo=8 offsethi=24        /* If ip == 1.  */
+miscmp_word_16:  miscmp_word offsetlo=16 offsethi=16
+miscmp_word_24:  miscmp_word offsetlo=24 offsethi=8
+
+
+return_32:
+        setup_return w1=r3, w2=r2
+        b       do_return
+return_34:
+        setup_return w1=r3, w2=r4
+        b       do_return
+return_25:
+        setup_return w1=r2, w2=r5
+        b       do_return
+return_35:
+        setup_return w1=r3, w2=r5
+        b       do_return
+return_24:
+        setup_return w1=r2, w2=r4
+
+do_return:
+
+#ifdef __ARMEB__
+        mov     r0, ip
+#else /* not  __ARMEB__ */
+        rev     r0, ip
+#endif /* not  __ARMEB__ */
+
+        /* Restore temporaries early, before computing the return value.  */
+        ldrd    r6, r7, [sp]
+        ldrd    r4, r5, [sp, #8]
+        adds    sp, sp, #16
+
+        /* There is a zero or a different byte between r1 and r2.  */
+        /* r0 contains a mask of all-zero bytes in r1.  */
+        /* Using r0 and not ip here because cbz requires low register.  */
+        m_cbz   reg=r0, label=compute_return_value
+        clz     r0, r0
+        /* r0 contains the number of bits on the left of the first all-zero byte in r1.  */
+        rsb     r0, r0, #24
+        /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1.  */
+        lsr     r1, r1, r0
+        lsr     r2, r2, r0
+
+compute_return_value:
+        movs    r0, #1
+        cmp     r1, r2
+        /* The return value is computed as follows.
+        If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return.
+        If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0,
+        which means r0:=r0-r0-1 and r0 is #-1 at return.
+        If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1,
+        which means r0:=r0-r0 and r0 is #0 at return.
+        (C==0 and Z==1) cannot happen because the carry bit is "not borrow".  */
+        it      ls
+        sbcls   r0, r0, r0
+        bx      lr
+
+
+#else   /* !(defined (_ISA_THUMB_2) || defined (_ISA_ARM_6)
+             defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) ||
+             (defined (__thumb__) && !defined (__thumb2__))) */
+
+        /* Use LDR whenever possible. */
+
+#ifdef __thumb2__
+#define magic1(REG) 0x01010101
+#define magic2(REG) 0x80808080
+#else
+#define magic1(REG) REG
+#define magic2(REG) REG, lsl #7
+#endif
+
+        optpld  r0
+        optpld  r1
+        eor     r2, r0, r1
+        tst     r2, #3
+        /* Strings not at same byte offset from a word boundary.  */
+        bne     strcmp_unaligned
+        ands    r2, r0, #3
+        bic     r0, r0, #3
+        bic     r1, r1, #3
+        ldr     ip, [r0], #4
+        it      eq
+        ldreq   r3, [r1], #4
+        beq     1f
+        /* Although s1 and s2 have identical initial alignment, they are
+        not currently word aligned.  Rather than comparing bytes,
+    make sure that any bytes fetched from before the addressed
+    bytes are forced to 0xff.  Then they will always compare
+    equal.  */
+        eor     r2, r2, #3
+        lsl     r2, r2, #3
+        mvn     r3, MSB
+        S2LOMEM        r2, r3, r2
+        ldr     r3, [r1], #4
+        orr     ip, ip, r2
+        orr     r3, r3, r2
+1:
+#ifndef __thumb2__
+              /* Load the 'magic' constant 0x01010101.  */
+        str     r4, [sp, #-4]!
+        mov     r4, #1
+        orr     r4, r4, r4, lsl #8
+        orr     r4, r4, r4, lsl #16
+#endif
+        .p2align        2
+4:
+        optpld  r0, #8
+        optpld  r1, #8
+        sub     r2, ip, magic1(r4)
+        cmp     ip, r3
+        itttt   eq
+        /* check for any zero bytes in first word */
+        biceq   r2, r2, ip
+        tsteq   r2, magic2(r4)
+        ldreq   ip, [r0], #4
+        ldreq   r3, [r1], #4
+        beq     4b
+2:
+        /* There's a zero or a different byte in the word */
+        S2HIMEM  r0, ip, #24
+        S2LOMEM  ip, ip, #8
+        cmp     r0, #1
+        it      cs
+        cmpcs   r0, r3, S2HIMEM #24
+        it      eq
+        S2LOMEMEQ r3, r3, #8
+        beq     2b
+        /* On a big-endian machine, r0 contains the desired byte in bits
+        0-7; on a little-endian machine they are in bits 24-31.  In
+        both cases the other bits in r0 are all zero.  For r3 the
+        interesting byte is at the other end of the word, but the
+        other bits are not necessarily zero.  We need a signed result
+        representing the differnece in the unsigned bytes, so for the
+        little-endian case we can't just shift the interesting bits
+        up.  */
+#ifdef __ARMEB__
+        sub     r0, r0, r3, lsr #24
+#else
+        and     r3, r3, #255
+#ifdef __thumb2__
+        /* No RSB instruction in Thumb2 */
+        lsr     r0, r0, #24
+        sub     r0, r0, r3
+#else
+        rsb     r0, r3, r0, lsr #24
+#endif
+#endif
+#ifndef __thumb2__
+        ldr     r4, [sp], #4
+#endif
+        RETURN
+
+
+strcmp_unaligned:
+
+#if 0
+        /* The assembly code below is based on the following alogrithm.  */
+#ifdef __ARMEB__
+#define RSHIFT <<
+#define LSHIFT >>
+#else
+#define RSHIFT >>
+#define LSHIFT <<
+#endif
+
+#define body(shift)                         \
+  mask = 0xffffffffU RSHIFT shift;                  \
+  w1 = *wp1++;                              \
+  w2 = *wp2++;                              \
+  do                                    \
+    {                                   \
+      t1 = w1 & mask;                           \
+      if (__builtin_expect(t1 != w2 RSHIFT shift, 0))           \
+    {                               \
+      w2 RSHIFT= shift;                     \
+      break;                            \
+    }                               \
+      if (__builtin_expect(((w1 - b1) & ~w1) & (b1 << 7), 0))       \
+    {                               \
+      /* See comment in assembler below re syndrome on big-endian */\
+      if ((((w1 - b1) & ~w1) & (b1 << 7)) & mask)           \
+        w2 RSHIFT= shift;                       \
+      else                              \
+        {                               \
+          w2 = *wp2;                        \
+          t1 = w1 RSHIFT (32 - shift);              \
+          w2 = (w2 LSHIFT (32 - shift)) RSHIFT (32 - shift);    \
+        }                               \
+      break;                            \
+    }                               \
+      w2 = *wp2++;                          \
+      t1 ^= w1;                             \
+      if (__builtin_expect(t1 != w2 LSHIFT (32 - shift), 0))        \
+    {                               \
+      t1 = w1 >> (32 - shift);                  \
+      w2 = (w2 << (32 - shift)) RSHIFT (32 - shift);        \
+      break;                            \
+    }                               \
+      w1 = *wp1++;                          \
+    } while (1)
+
+  const unsigned* wp1;
+  const unsigned* wp2;
+  unsigned w1, w2;
+  unsigned mask;
+  unsigned shift;
+  unsigned b1 = 0x01010101;
+  char c1, c2;
+  unsigned t1;
+
+  while (((unsigned) s1) & 3)
+    {
+      c1 = *s1++;
+      c2 = *s2++;
+      if (c1 == 0 || c1 != c2)
+    return c1 - (int)c2;
+    }
+  wp1 = (unsigned*) (((unsigned)s1) & ~3);
+  wp2 = (unsigned*) (((unsigned)s2) & ~3);
+  t1 = ((unsigned) s2) & 3;
+  if (t1 == 1)
+    {
+      body(8);
+    }
+  else if (t1 == 2)
+    {
+      body(16);
+    }
+  else
+    {
+      body (24);
+    }
+
+  do
+    {
+#ifdef __ARMEB__
+      c1 = (char) t1 >> 24;
+      c2 = (char) w2 >> 24;
+#else /* not  __ARMEB__ */
+      c1 = (char) t1;
+      c2 = (char) w2;
+#endif /* not  __ARMEB__ */
+      t1 RSHIFT= 8;
+      w2 RSHIFT= 8;
+    } while (c1 != 0 && c1 == c2);
+  return c1 - c2;
+#endif /* 0 */
+
+
+        wp1 .req r0
+        wp2 .req r1
+        b1  .req r2
+        w1  .req r4
+        w2  .req r5
+        t1  .req ip
+        @ r3 is scratch
+
+        /* First of all, compare bytes until wp1(sp1) is word-aligned. */
+1:
+        tst     wp1, #3
+        beq     2f
+        ldrb    r2, [wp1], #1
+        ldrb    r3, [wp2], #1
+        cmp     r2, #1
+        it      cs
+        cmpcs   r2, r3
+        beq     1b
+        sub     r0, r2, r3
+        RETURN
+
+2:
+        str     r5, [sp, #-4]!
+        str     r4, [sp, #-4]!
+        //stmfd   sp!, {r4, r5}
+        mov     b1, #1
+        orr     b1, b1, b1, lsl #8
+        orr     b1, b1, b1, lsl #16
+
+        and     t1, wp2, #3
+        bic     wp2, wp2, #3
+        ldr     w1, [wp1], #4
+        ldr     w2, [wp2], #4
+        cmp     t1, #2
+        beq     2f
+        bhi     3f
+
+        /* Critical inner Loop: Block with 3 bytes initial overlap */
+        .p2align        2
+1:
+        bic     t1, w1, MSB
+        cmp     t1, w2, S2LOMEM #8
+        sub     r3, w1, b1
+        bic     r3, r3, w1
+        bne     4f
+        ands    r3, r3, b1, lsl #7
+        it      eq
+        ldreq   w2, [wp2], #4
+        bne     5f
+        eor     t1, t1, w1
+        cmp     t1, w2, S2HIMEM #24
+        bne     6f
+        ldr     w1, [wp1], #4
+        b       1b
+4:
+        S2LOMEM        w2, w2, #8
+        b       8f
+
+5:
+#ifdef __ARMEB__
+        /* The syndrome value may contain false ones if the string ends
+        with the bytes 0x01 0x00 */
+        tst     w1, #0xff000000
+        itt     ne
+        tstne   w1, #0x00ff0000
+        tstne   w1, #0x0000ff00
+        beq     7f
+#else
+        bics    r3, r3, #0xff000000
+        bne     7f
+#endif
+        ldrb    w2, [wp2]
+        S2LOMEM  t1, w1, #24
+#ifdef __ARMEB__
+        lsl     w2, w2, #24
+#endif
+        b       8f
+
+6:
+        S2LOMEM  t1, w1, #24
+        and     w2, w2, LSB
+        b       8f
+
+        /* Critical inner Loop: Block with 2 bytes initial overlap */
+        .p2align        2
+2:
+        S2HIMEM  t1, w1, #16
+        sub     r3, w1, b1
+        S2LOMEM  t1, t1, #16
+        bic     r3, r3, w1
+        cmp     t1, w2, S2LOMEM #16
+        bne     4f
+        ands    r3, r3, b1, lsl #7
+        it      eq
+        ldreq   w2, [wp2], #4
+        bne     5f
+        eor     t1, t1, w1
+        cmp     t1, w2, S2HIMEM #16
+        bne     6f
+        ldr     w1, [wp1], #4
+        b       2b
+
+5:
+#ifdef __ARMEB__
+        /* The syndrome value may contain false ones if the string ends
+        with the bytes 0x01 0x00 */
+        tst     w1, #0xff000000
+        it      ne
+        tstne   w1, #0x00ff0000
+        beq     7f
+#else
+        lsls    r3, r3, #16
+        bne     7f
+#endif
+        ldrh    w2, [wp2]
+        S2LOMEM  t1, w1, #16
+#ifdef __ARMEB__
+        lsl     w2, w2, #16
+#endif
+        b       8f
+
+6:
+        S2HIMEM  w2, w2, #16
+        S2LOMEM  t1, w1, #16
+4:
+        S2LOMEM  w2, w2, #16
+        b       8f
+
+        /* Critical inner Loop: Block with 1 byte initial overlap */
+        .p2align        2
+3:
+        and     t1, w1, LSB
+        cmp     t1, w2, S2LOMEM #24
+        sub     r3, w1, b1
+        bic     r3, r3, w1
+        bne     4f
+        ands    r3, r3, b1, lsl #7
+        it      eq
+        ldreq   w2, [wp2], #4
+        bne     5f
+        eor     t1, t1, w1
+        cmp     t1, w2, S2HIMEM #8
+        bne     6f
+        ldr     w1, [wp1], #4
+        b       3b
+4:
+        S2LOMEM  w2, w2, #24
+        b       8f
+5:
+        /* The syndrome value may contain false ones if the string ends
+        with the bytes 0x01 0x00 */
+        tst     w1, LSB
+        beq     7f
+        ldr     w2, [wp2], #4
+6:
+        S2LOMEM  t1, w1, #8
+        bic     w2, w2, MSB
+        b       8f
+7:
+        mov     r0, #0
+        //ldmfd   sp!, {r4, r5}
+        ldr     r4, [sp], #4
+        ldr     r5, [sp], #4
+        RETURN
+8:
+        and     r2, t1, LSB
+        and     r0, w2, LSB
+        cmp     r0, #1
+        it      cs
+        cmpcs   r0, r2
+        itt     eq
+        S2LOMEMEQ        t1, t1, #8
+        S2LOMEMEQ        w2, w2, #8
+        beq     8b
+        sub     r0, r2, r0
+        //ldmfd   sp!, {r4, r5}
+        ldr     r4, [sp], #4
+        ldr     r5, [sp], #4
+        RETURN
+
+#endif /* !(defined (_ISA_THUMB_2) || defined (_ISA_ARM_6)
+            defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) ||
+            (defined (__thumb__) && !defined (__thumb2__))) */
--- a/libc/arch-arm/bionic/syscall.S
+++ b/libc/arch-arm/bionic/syscall.S
@@ -26,8 +26,8 @@
 * SUCH DAMAGE.
 */

+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(syscall)
    mov     ip, sp
@@ -39,7 +39,8 @@ ENTRY(syscall)
    ldmfd   ip, {r3, r4, r5, r6}
    swi     #0
    ldmfd   sp!, {r4, r5, r6, r7}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(syscall)
--- a/libc/arch-arm/bionic/tgkill.S
+++ b/libc/arch-arm/bionic/tgkill.S
@@ -25,6 +25,11 @@
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */
+
+#include <linux/err.h>
+#include <asm/unistd.h>
+#include <machine/asm.h>
+
 /* unlike our auto-generated syscall stubs, this code saves lr
   on the stack, as well as a few other registers. this makes
   our stack unwinder happy, when we generate debug stack
@@ -33,19 +38,13 @@
   of a corrupted malloc heap).
 */

-#include <sys/linux-syscalls.h>
-#include <machine/asm.h>
-
-#ifndef __NR_tgkill
-#define __NR_tgkill  268
-#endif
-
 ENTRY(tgkill)
    stmfd   sp!, {r4-r7, ip, lr}
    ldr     r7, =__NR_tgkill
    swi     #0
    ldmfd   sp!, {r4-r7, ip, lr}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(tgkill)
--- a/libc/arch-arm/bionic/tkill.S
+++ b/libc/arch-arm/bionic/tkill.S
@@ -25,6 +25,11 @@
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 */
+
+#include <linux/err.h>
+#include <asm/unistd.h>
+#include <machine/asm.h>
+
 /* unlike our auto-generated syscall stubs, this code saves lr
   on the stack, as well as a few other registers. this makes
   our stack unwinder happy, when we generate debug stack
@@ -33,19 +38,13 @@
   of a corrupted malloc heap).
 */

-#include <sys/linux-syscalls.h>
-#include <machine/asm.h>
-
-#ifndef __NR_tkill
-#define __NR_tkill  238
-#endif
-
 ENTRY(tkill)
    stmfd   sp!, {r4-r7, ip, lr}
    ldr     r7, =__NR_tkill
    swi     #0
    ldmfd   sp!, {r4-r7, ip, lr}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(tkill)
--- a/libc/arch-arm/cortex-a15/bionic/memcpy.S
+++ b/libc/arch-arm/cortex-a15/bionic/memcpy.S
@@ -0,0 +1,146 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+/* Assumes neon instructions and a cache line size of 64 bytes. */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+/*
+ * This code assumes it is running on a processor that supports all arm v7
+ * instructions, that supports neon instructions, and that has a 64 byte
+ * cache line.
+ */
+
+        .text
+        .fpu    neon
+
+#define CACHE_LINE_SIZE     64
+
+ENTRY(memcpy)
+        .save       {r0, lr}
+        /* start preloading as early as possible */
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        stmfd       sp!, {r0, lr}
+        pld         [r1, #(CACHE_LINE_SIZE*1)]
+
+        /* do we have at least 16-bytes to copy (needed for alignment below) */
+        cmp         r2, #16
+        blo         5f
+
+        /* align destination to cache-line for the write-buffer */
+        rsb         r3, r0, #0
+        ands        r3, r3, #0xF
+        beq         0f
+
+        /* copy up to 15-bytes (count in r3) */
+        sub         r2, r2, r3
+        movs        ip, r3, lsl #31
+        ldrmib      lr, [r1], #1
+        strmib      lr, [r0], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+        movs        ip, r3, lsl #29
+        bge         1f
+        // copies 4 bytes, destination 32-bits aligned
+        vld4.8      {d0[0], d1[0], d2[0], d3[0]}, [r1]!
+        vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r0, :32]!
+1:      bcc         2f
+        // copies 8 bytes, destination 64-bits aligned
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0, :64]!
+2:
+
+0:      /* preload immediately the next cache line, which we may need */
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        pld         [r1, #(CACHE_LINE_SIZE*1)]
+
+        /* make sure we have at least 64 bytes to copy */
+        subs        r2, r2, #64
+        blo         2f
+
+        /* Preload all the cache lines we need.
+         * NOTE: The number of pld below depends on CACHE_LINE_SIZE,
+         * ideally we would increase the distance in the main loop to
+         * avoid the goofy code below. In practice this doesn't seem to make
+         * a big difference.
+         * NOTE: The value CACHE_LINE_SIZE * 4 was chosen through
+         * experimentation.
+         */
+        pld         [r1, #(CACHE_LINE_SIZE*2)]
+        pld         [r1, #(CACHE_LINE_SIZE*3)]
+        pld         [r1, #(CACHE_LINE_SIZE*4)]
+
+1:      /* The main loop copies 64 bytes at a time */
+        vld1.8      {d0  - d3},   [r1]!
+        vld1.8      {d4  - d7},   [r1]!
+        pld         [r1, #(CACHE_LINE_SIZE*4)]
+        subs        r2, r2, #64
+        vst1.8      {d0  - d3},   [r0, :128]!
+        vst1.8      {d4  - d7},   [r0, :128]!
+        bhs         1b
+
+2:      /* fix-up the remaining count and make sure we have >= 32 bytes left */
+        add         r2, r2, #64
+        subs        r2, r2, #32
+        blo         4f
+
+3:      /* 32 bytes at a time. These cache lines were already preloaded */
+        vld1.8      {d0 - d3},  [r1]!
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3},  [r0, :128]!
+        bhs         3b
+4:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         5f
+        // copies 16 bytes, 128-bits aligned
+        vld1.8      {d0, d1}, [r1]!
+        vst1.8      {d0, d1}, [r0, :128]!
+
+5:      /* copy up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vld4.8      {d0[0], d1[0], d2[0], d3[0]}, [r1]!
+        vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        ldrmib      r3, [r1], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strmib      r3, [r0], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+
+        ldmfd       sp!, {r0, lr}
+        bx          lr
+END(memcpy)
--- a/libc/arch-arm/cortex-a15/bionic/memset.S
+++ b/libc/arch-arm/cortex-a15/bionic/memset.S
@@ -0,0 +1,106 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+		/*
+		 * Optimized memset() for ARM.
+         *
+         * memset() returns its first argument.
+		 */
+
+    .fpu    neon
+
+ENTRY(bzero)
+        mov     r2, r1
+        mov     r1, #0
+        // Fall through to memset...
+END(bzero)
+
+ENTRY(memset)
+        .save       {r0}
+        stmfd       sp!, {r0}
+
+        vdup.8      q0, r1
+
+        /* do we have at least 16-bytes to write (needed for alignment below) */
+        cmp         r2, #16
+        blo         3f
+
+        /* align destination to 16 bytes for the write-buffer */
+        rsb         r3, r0, #0
+        ands        r3, r3, #0xF
+        beq         2f
+
+        /* write up to 15-bytes (count in r3) */
+        sub         r2, r2, r3
+        movs        ip, r3, lsl #31
+        strmib      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        movs        ip, r3, lsl #29
+        bge         1f
+
+        // writes 4 bytes, 32-bits aligned
+        vst1.32     {d0[0]}, [r0, :32]!
+1:      bcc         2f
+
+        // writes 8 bytes, 64-bits aligned
+        vst1.8      {d0}, [r0, :64]!
+2:
+        /* make sure we have at least 32 bytes to write */
+        subs        r2, r2, #32
+        blo         2f
+        vmov        q1, q0
+
+1:      /* The main loop writes 32 bytes at a time */
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3}, [r0, :128]!
+        bhs         1b
+
+2:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         3f
+
+        // writes 16 bytes, 128-bits aligned
+        vst1.8      {d0, d1}, [r0, :128]!
+3:      /* write up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vst1.32     {d0[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        strmib      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        ldmfd       sp!, {r0}
+        bx          lr
+END(memset)
--- a/libc/arch-arm/cortex-a15/bionic/strcmp.S
+++ b/libc/arch-arm/cortex-a15/bionic/strcmp.S
@@ -0,0 +1,377 @@
+/*
+ * Copyright (c) 2013 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+#ifdef __ARMEB__
+#define S2LOMEM lsl
+#define S2LOMEMEQ lsleq
+#define S2HIMEM lsr
+#define MSB 0x000000ff
+#define LSB 0xff000000
+#define BYTE0_OFFSET 24
+#define BYTE1_OFFSET 16
+#define BYTE2_OFFSET 8
+#define BYTE3_OFFSET 0
+#else /* not  __ARMEB__ */
+#define S2LOMEM lsr
+#define S2LOMEMEQ lsreq
+#define S2HIMEM lsl
+#define BYTE0_OFFSET 0
+#define BYTE1_OFFSET 8
+#define BYTE2_OFFSET 16
+#define BYTE3_OFFSET 24
+#define MSB 0xff000000
+#define LSB 0x000000ff
+#endif /* not  __ARMEB__ */
+
+.syntax         unified
+
+#if defined (__thumb__)
+        .thumb
+        .thumb_func
+#endif
+
+ENTRY(strcmp)
+      /* Use LDRD whenever possible.  */
+
+/* The main thing to look out for when comparing large blocks is that
+   the loads do not cross a page boundary when loading past the index
+   of the byte with the first difference or the first string-terminator.
+
+   For example, if the strings are identical and the string-terminator
+   is at index k, byte by byte comparison will not load beyond address
+   s1+k and s2+k; word by word comparison may load up to 3 bytes beyond
+   k; double word - up to 7 bytes.  If the load of these bytes crosses
+   a page boundary, it might cause a memory fault (if the page is not mapped)
+   that would not have happened in byte by byte comparison.
+
+   If an address is (double) word aligned, then a load of a (double) word
+   from that address will not cross a page boundary.
+   Therefore, the algorithm below considers word and double-word alignment
+   of strings separately.  */
+
+/* High-level description of the algorithm.
+
+   * The fast path: if both strings are double-word aligned,
+     use LDRD to load two words from each string in every loop iteration.
+   * If the strings have the same offset from a word boundary,
+     use LDRB to load and compare byte by byte until
+     the first string is aligned to a word boundary (at most 3 bytes).
+     This is optimized for quick return on short unaligned strings.
+   * If the strings have the same offset from a double-word boundary,
+     use LDRD to load two words from each string in every loop iteration, as in the fast path.
+   * If the strings do not have the same offset from a double-word boundary,
+     load a word from the second string before the loop to initialize the queue.
+     Use LDRD to load two words from every string in every loop iteration.
+     Inside the loop, load the second word from the second string only after comparing
+     the first word, using the queued value, to guarantee safety across page boundaries.
+   * If the strings do not have the same offset from a word boundary,
+     use LDR and a shift queue. Order of loads and comparisons matters,
+     similarly to the previous case.
+
+   * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value.
+   * The only difference between ARM and Thumb modes is the use of CBZ instruction.
+   * The only difference between big and little endian is the use of REV in little endian
+     to compute the return value, instead of MOV.
+*/
+
+        .macro m_cbz reg label
+#ifdef __thumb2__
+        cbz     \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        beq     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbz */
+
+        .macro m_cbnz reg label
+#ifdef __thumb2__
+        cbnz    \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        bne     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbnz */
+
+        .macro  init
+        /* Macro to save temporary registers and prepare magic values.  */
+        subs    sp, sp, #16
+        strd    r4, r5, [sp, #8]
+        strd    r6, r7, [sp]
+        mvn     r6, #0  /* all F */
+        mov     r7, #0  /* all 0 */
+        .endm   /* init */
+
+        .macro  magic_compare_and_branch w1 w2 label
+        /* Macro to compare registers w1 and w2 and conditionally branch to label.  */
+        cmp     \w1, \w2        /* Are w1 and w2 the same?  */
+        magic_find_zero_bytes \w1
+        it      eq
+        cmpeq   ip, #0          /* Is there a zero byte in w1?  */
+        bne     \label
+        .endm /* magic_compare_and_branch */
+
+        .macro  magic_find_zero_bytes w1
+        /* Macro to find all-zero bytes in w1, result is in ip.  */
+#if (defined (__ARM_FEATURE_DSP))
+        uadd8   ip, \w1, r6
+        sel     ip, r7, r6
+#else /* not defined (__ARM_FEATURE_DSP) */
+        /* __ARM_FEATURE_DSP is not defined for some Cortex-M processors.
+        Coincidently, these processors only have Thumb-2 mode, where we can use the
+        the (large) magic constant available directly as an immediate in instructions.
+        Note that we cannot use the magic constant in ARM mode, where we need
+        to create the constant in a register.  */
+        sub     ip, \w1, #0x01010101
+        bic     ip, ip, \w1
+        and     ip, ip, #0x80808080
+#endif /* not defined (__ARM_FEATURE_DSP) */
+        .endm /* magic_find_zero_bytes */
+
+        .macro  setup_return w1 w2
+#ifdef __ARMEB__
+        mov     r1, \w1
+        mov     r2, \w2
+#else /* not  __ARMEB__ */
+        rev     r1, \w1
+        rev     r2, \w2
+#endif /* not  __ARMEB__ */
+        .endm /* setup_return */
+
+        pld [r0, #0]
+        pld [r1, #0]
+
+        /* Are both strings double-word aligned?  */
+        orr     ip, r0, r1
+        tst     ip, #7
+        bne     do_align
+
+        /* Fast path.  */
+        init
+
+doubleword_aligned:
+
+        /* Get here when the strings to compare are double-word aligned.  */
+        /* Compare two words in every iteration.  */
+        .p2align        2
+2:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string.  */
+        ldrd    r2, r3, [r0], #8
+        ldrd    r4, r5, [r1], #8
+
+        magic_compare_and_branch w1=r2, w2=r4, label=return_24
+        magic_compare_and_branch w1=r3, w2=r5, label=return_35
+        b       2b
+
+do_align:
+        /* Is the first string word-aligned?  */
+        ands    ip, r0, #3
+        beq     word_aligned_r0
+
+        /* Fast compare byte by byte until the first string is word-aligned.  */
+        /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes
+        to read until the next word boundary is 4-ip.  */
+        bic     r0, r0, #3
+        ldr     r2, [r0], #4
+        lsls    ip, ip, #31
+        beq     byte2
+        bcs     byte3
+
+byte1:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE1_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte2:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE2_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte3:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE3_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbnz  reg=r3, label=word_aligned_r0
+
+fast_return:
+        mov     r0, ip
+        bx      lr
+
+word_aligned_r0:
+        init
+        /* The first string is word-aligned.  */
+        /* Is the second string word-aligned?  */
+        ands    ip, r1, #3
+        bne     strcmp_unaligned
+
+word_aligned:
+        /* The strings are word-aligned. */
+        /* Is the first string double-word aligned?  */
+        tst     r0, #4
+        beq     doubleword_aligned_r0
+
+        /* If r0 is not double-word aligned yet, align it by loading
+        and comparing the next word from each string.  */
+        ldr     r2, [r0], #4
+        ldr     r4, [r1], #4
+        magic_compare_and_branch w1=r2 w2=r4 label=return_24
+
+doubleword_aligned_r0:
+        /* Get here when r0 is double-word aligned.  */
+        /* Is r1 doubleword_aligned?  */
+        tst     r1, #4
+        beq     doubleword_aligned
+
+        /* Get here when the strings to compare are word-aligned,
+        r0 is double-word aligned, but r1 is not double-word aligned.  */
+
+        /* Initialize the queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare two words in every iteration.  */
+        .p2align        2
+3:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string and compare.  */
+        ldrd    r2, r3, [r0], #8
+        magic_compare_and_branch w1=r2 w2=r5 label=return_25
+        ldrd    r4, r5, [r1], #8
+        magic_compare_and_branch w1=r3 w2=r4 label=return_34
+        b       3b
+
+        .macro miscmp_word offsetlo offsethi
+        /* Macro to compare misaligned strings.  */
+        /* r0, r1 are word-aligned, and at least one of the strings
+        is not double-word aligned.  */
+        /* Compare one word in every loop iteration.  */
+        /* OFFSETLO is the original bit-offset of r1 from a word-boundary,
+        OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word).  */
+
+        /* Initialize the shift queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare one word from each string in every loop iteration.  */
+        .p2align        2
+7:
+        ldr     r3, [r0], #4
+        S2LOMEM r5, r5, #\offsetlo
+        magic_find_zero_bytes w1=r3
+        cmp     r7, ip, S2HIMEM #\offsetlo
+        and     r2, r3, r6, S2LOMEM #\offsetlo
+        it      eq
+        cmpeq   r2, r5
+        bne     return_25
+        ldr     r5, [r1], #4
+        cmp     ip, #0
+        eor r3, r2, r3
+        S2HIMEM r2, r5, #\offsethi
+        it      eq
+        cmpeq   r3, r2
+        bne     return_32
+        b       7b
+        .endm /* miscmp_word */
+
+strcmp_unaligned:
+        /* r0 is word-aligned, r1 is at offset ip from a word.  */
+        /* Align r1 to the (previous) word-boundary.  */
+        bic     r1, r1, #3
+
+        /* Unaligned comparison word by word using LDRs. */
+        cmp     ip, #2
+        beq     miscmp_word_16                    /* If ip == 2.  */
+        bge     miscmp_word_24                    /* If ip == 3.  */
+        miscmp_word offsetlo=8 offsethi=24        /* If ip == 1.  */
+miscmp_word_16:  miscmp_word offsetlo=16 offsethi=16
+miscmp_word_24:  miscmp_word offsetlo=24 offsethi=8
+
+
+return_32:
+        setup_return w1=r3, w2=r2
+        b       do_return
+return_34:
+        setup_return w1=r3, w2=r4
+        b       do_return
+return_25:
+        setup_return w1=r2, w2=r5
+        b       do_return
+return_35:
+        setup_return w1=r3, w2=r5
+        b       do_return
+return_24:
+        setup_return w1=r2, w2=r4
+
+do_return:
+
+#ifdef __ARMEB__
+        mov     r0, ip
+#else /* not  __ARMEB__ */
+        rev     r0, ip
+#endif /* not  __ARMEB__ */
+
+        /* Restore temporaries early, before computing the return value.  */
+        ldrd    r6, r7, [sp]
+        ldrd    r4, r5, [sp, #8]
+        adds    sp, sp, #16
+
+        /* There is a zero or a different byte between r1 and r2.  */
+        /* r0 contains a mask of all-zero bytes in r1.  */
+        /* Using r0 and not ip here because cbz requires low register.  */
+        m_cbz   reg=r0, label=compute_return_value
+        clz     r0, r0
+        /* r0 contains the number of bits on the left of the first all-zero byte in r1.  */
+        rsb     r0, r0, #24
+        /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1.  */
+        lsr     r1, r1, r0
+        lsr     r2, r2, r0
+
+compute_return_value:
+        movs    r0, #1
+        cmp     r1, r2
+        /* The return value is computed as follows.
+        If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return.
+        If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0,
+        which means r0:=r0-r0-1 and r0 is #-1 at return.
+        If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1,
+        which means r0:=r0-r0 and r0 is #0 at return.
+        (C==0 and Z==1) cannot happen because the carry bit is "not borrow".  */
+        it      ls
+        sbcls   r0, r0, r0
+        bx      lr
+END(strcmp)
--- a/libc/arch-arm/cortex-a15/cortex-a15.mk
+++ b/libc/arch-arm/cortex-a15/cortex-a15.mk
@@ -0,0 +1,5 @@
+$(call libc-add-cpu-variant-src,MEMCPY,arch-arm/cortex-a15/bionic/memcpy.S)
+$(call libc-add-cpu-variant-src,MEMSET,arch-arm/cortex-a15/bionic/memset.S)
+$(call libc-add-cpu-variant-src,STRCMP,arch-arm/cortex-a15/bionic/strcmp.S)
+
+include bionic/libc/arch-arm/generic/generic.mk
--- a/libc/arch-arm/cortex-a7/cortex-a7.mk
+++ b/libc/arch-arm/cortex-a7/cortex-a7.mk
@@ -0,0 +1 @@
+include bionic/libc/arch-arm/cortex-a15/cortex-a15.mk
--- a/libc/arch-arm/cortex-a9/bionic/memcpy.S
+++ b/libc/arch-arm/cortex-a9/bionic/memcpy.S
@@ -0,0 +1,211 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+/*
+ * This code assumes it is running on a processor that supports all arm v7
+ * instructions, that supports neon instructions, and that has a 32 byte
+ * cache line.
+ */
+
+        .text
+        .fpu    neon
+
+#define CACHE_LINE_SIZE     32
+
+ENTRY(memcpy)
+        .save       {r0, lr}
+        /* start preloading as early as possible */
+        pld         [r1, #(CACHE_LINE_SIZE * 0)]
+        stmfd       sp!, {r0, lr}
+        pld         [r1, #(CACHE_LINE_SIZE * 2)]
+
+        // Check so divider is at least 16 bytes, needed for alignment code.
+        cmp         r2, #16
+        blo         5f
+
+
+        /* check if buffers are aligned. If so, run arm-only version */
+        eor         r3, r0, r1
+        ands        r3, r3, #0x3
+        beq         11f
+
+        /* Check the upper size limit for Neon unaligned memory access in memcpy */
+        cmp         r2, #224
+        blo         3f
+
+        /* align destination to 16 bytes for the write-buffer */
+        rsb         r3, r0, #0
+        ands        r3, r3, #0xF
+        beq         3f
+
+        /* copy up to 15-bytes (count in r3) */
+        sub         r2, r2, r3
+        movs        ip, r3, lsl #31
+        ldrmib      lr, [r1], #1
+        strmib      lr, [r0], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+        movs        ip, r3, lsl #29
+        bge         1f
+        // copies 4 bytes, destination 32-bits aligned
+        vld1.32     {d0[0]}, [r1]!
+        vst1.32     {d0[0]}, [r0, :32]!
+1:      bcc         2f
+        // copies 8 bytes, destination 64-bits aligned
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0, :64]!
+2:
+        /* preload immediately the next cache line, which we may need */
+        pld         [r1, #(CACHE_LINE_SIZE * 0)]
+        pld         [r1, #(CACHE_LINE_SIZE * 2)]
+3:
+        /* make sure we have at least 64 bytes to copy */
+        subs        r2, r2, #64
+        blo         2f
+
+        /* preload all the cache lines we need */
+        pld         [r1, #(CACHE_LINE_SIZE * 4)]
+        pld         [r1, #(CACHE_LINE_SIZE * 6)]
+
+1:      /* The main loop copies 64 bytes at a time */
+        vld1.8      {d0 - d3}, [r1]!
+        vld1.8      {d4 - d7}, [r1]!
+        pld         [r1, #(CACHE_LINE_SIZE * 6)]
+        subs        r2, r2, #64
+        vst1.8      {d0 - d3}, [r0]!
+        vst1.8      {d4 - d7}, [r0]!
+        bhs         1b
+
+2:      /* fix-up the remaining count and make sure we have >= 32 bytes left */
+        add         r2, r2, #64
+        subs        r2, r2, #32
+        blo         4f
+
+3:      /* 32 bytes at a time. These cache lines were already preloaded */
+        vld1.8      {d0 - d3}, [r1]!
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3}, [r0]!
+        bhs         3b
+
+4:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         5f
+        // copies 16 bytes, 128-bits aligned
+        vld1.8      {d0, d1}, [r1]!
+        vst1.8      {d0, d1}, [r0]!
+5:      /* copy up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vld1.32     {d0[0]}, [r1]!
+        vst1.32     {d0[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        ldrmib      r3, [r1], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strmib      r3, [r0], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+
+        ldmfd       sp!, {r0, lr}
+        bx          lr
+11:
+        /* Simple arm-only copy loop to handle aligned copy operations */
+        stmfd       sp!, {r4, r5, r6, r7, r8}
+        pld         [r1, #(CACHE_LINE_SIZE * 4)]
+
+        /* Check alignment */
+        rsb         r3, r1, #0
+        ands        r3, #3
+        beq         2f
+
+        /* align source to 32 bits. We need to insert 2 instructions between
+         * a ldr[b|h] and str[b|h] because byte and half-word instructions
+         * stall 2 cycles.
+         */
+        movs        r12, r3, lsl #31
+        sub         r2, r2, r3      /* we know that r3 <= r2 because r2 >= 4 */
+        ldrmib      r3, [r1], #1
+        ldrcsb      r4, [r1], #1
+        ldrcsb      r5, [r1], #1
+        strmib      r3, [r0], #1
+        strcsb      r4, [r0], #1
+        strcsb      r5, [r0], #1
+
+2:
+        subs        r2, r2, #64
+        blt         4f
+
+3:      /* Main copy loop, copying 64 bytes at a time */
+        pld         [r1, #(CACHE_LINE_SIZE * 8)]
+        ldmia       r1!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        stmia       r0!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        ldmia       r1!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        stmia       r0!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        subs        r2, r2, #64
+        bge         3b
+
+4:      /* Check if there are > 32 bytes left */
+        adds        r2, r2, #64
+        subs        r2, r2, #32
+        blt         5f
+
+        /* Copy 32 bytes */
+        ldmia       r1!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        stmia       r0!, {r3, r4, r5, r6, r7, r8, r12, lr}
+        subs        r2, #32
+
+5:      /* Handle any remaining bytes */
+        adds        r2, #32
+        beq         6f
+
+        movs        r12, r2, lsl #28
+        ldmcsia     r1!, {r3, r4, r5, r6}   /* 16 bytes */
+        ldmmiia     r1!, {r7, r8}           /*  8 bytes */
+        stmcsia     r0!, {r3, r4, r5, r6}
+        stmmiia     r0!, {r7, r8}
+        movs        r12, r2, lsl #30
+        ldrcs       r3, [r1], #4            /*  4 bytes */
+        ldrmih      r4, [r1], #2            /*  2 bytes */
+        strcs       r3, [r0], #4
+        strmih      r4, [r0], #2
+        tst         r2, #0x1
+        ldrneb      r3, [r1]                /*  last byte  */
+        strneb      r3, [r0]
+6:
+        ldmfd       sp!, {r4, r5, r6, r7, r8}
+        ldmfd       sp!, {r0, pc}
+END(memcpy)
--- a/libc/arch-arm/cortex-a9/bionic/memset.S
+++ b/libc/arch-arm/cortex-a9/bionic/memset.S
@@ -0,0 +1,152 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+/*
+ * This code assumes it is running on a processor that supports all arm v7
+ * instructions and that supports neon instructions.
+ */
+
+    .fpu    neon
+
+ENTRY(bzero)
+        mov     r2, r1
+        mov     r1, #0
+END(bzero)
+
+/* memset() returns its first argument.  */
+ENTRY(memset)
+        # The neon memset only wins for less than 132.
+        cmp         r2, #132
+        bhi         11f
+
+        .save       {r0}
+        stmfd       sp!, {r0}
+
+        vdup.8      q0, r1
+
+        /* make sure we have at least 32 bytes to write */
+        subs        r2, r2, #32
+        blo         2f
+        vmov        q1, q0
+
+1:      /* The main loop writes 32 bytes at a time */
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3}, [r0]!
+        bhs         1b
+
+2:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         3f
+
+        // writes 16 bytes, 128-bits aligned
+        vst1.8      {d0, d1}, [r0]!
+3:      /* write up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vst1.32     {d0[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        strmib      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        ldmfd       sp!, {r0}
+        bx          lr
+11:
+        /* compute the offset to align the destination
+         * offset = (4-(src&3))&3 = -src & 3
+         */
+
+        .save       {r0, r4-r7, lr}
+        stmfd       sp!, {r0, r4-r7, lr}
+        rsb         r3, r0, #0
+        ands        r3, r3, #3
+        cmp         r3, r2
+        movhi       r3, r2
+
+        /* splat r1 */
+        mov         r1, r1, lsl #24
+        orr         r1, r1, r1, lsr #8
+        orr         r1, r1, r1, lsr #16
+
+        movs        r12, r3, lsl #31
+        strcsb      r1, [r0], #1    /* can't use strh (alignment unknown) */
+        strcsb      r1, [r0], #1
+        strmib      r1, [r0], #1
+        subs        r2, r2, r3
+        ldmlsfd     sp!, {r0, r4-r7, lr}   /* return */
+        bxls        lr
+
+        /* align the destination to a cache-line */
+        mov         r12, r1
+        mov         lr, r1
+        mov         r4, r1
+        mov         r5, r1
+        mov         r6, r1
+        mov         r7, r1
+
+        rsb         r3, r0, #0
+        ands        r3, r3, #0x1C
+        beq         3f
+        cmp         r3, r2
+        andhi       r3, r2, #0x1C
+        sub         r2, r2, r3
+
+        /* conditionally writes 0 to 7 words (length in r3) */
+        movs        r3, r3, lsl #28
+        stmcsia     r0!, {r1, lr}
+        stmcsia     r0!, {r1, lr}
+        stmmiia     r0!, {r1, lr}
+        movs        r3, r3, lsl #2
+        strcs       r1, [r0], #4
+
+3:
+        subs        r2, r2, #32
+        mov         r3, r1
+        bmi         2f
+1:      subs        r2, r2, #32
+        stmia       r0!, {r1,r3,r4,r5,r6,r7,r12,lr}
+        bhs         1b
+2:      add         r2, r2, #32
+
+        /* conditionally stores 0 to 31 bytes */
+        movs        r2, r2, lsl #28
+        stmcsia     r0!, {r1,r3,r12,lr}
+        stmmiia     r0!, {r1, lr}
+        movs        r2, r2, lsl #2
+        strcs       r1, [r0], #4
+        strmih      r1, [r0], #2
+        movs        r2, r2, lsl #2
+        strcsb      r1, [r0]
+        ldmfd       sp!, {r0, r4-r7, lr}
+        bx          lr
+END(memset)
--- a/libc/arch-arm/cortex-a9/bionic/strcmp.S
+++ b/libc/arch-arm/cortex-a9/bionic/strcmp.S
@@ -0,0 +1,544 @@
+/*
+ * Copyright (c) 2013 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+#ifdef __ARMEB__
+#define S2LOMEM lsl
+#define S2LOMEMEQ lsleq
+#define S2HIMEM lsr
+#define MSB 0x000000ff
+#define LSB 0xff000000
+#define BYTE0_OFFSET 24
+#define BYTE1_OFFSET 16
+#define BYTE2_OFFSET 8
+#define BYTE3_OFFSET 0
+#else /* not  __ARMEB__ */
+#define S2LOMEM lsr
+#define S2LOMEMEQ lsreq
+#define S2HIMEM lsl
+#define BYTE0_OFFSET 0
+#define BYTE1_OFFSET 8
+#define BYTE2_OFFSET 16
+#define BYTE3_OFFSET 24
+#define MSB 0xff000000
+#define LSB 0x000000ff
+#endif /* not  __ARMEB__ */
+
+.syntax         unified
+
+#if defined (__thumb__)
+        .thumb
+        .thumb_func
+#endif
+
+ENTRY(strcmp)
+      /* Use LDRD whenever possible.  */
+
+/* The main thing to look out for when comparing large blocks is that
+   the loads do not cross a page boundary when loading past the index
+   of the byte with the first difference or the first string-terminator.
+
+   For example, if the strings are identical and the string-terminator
+   is at index k, byte by byte comparison will not load beyond address
+   s1+k and s2+k; word by word comparison may load up to 3 bytes beyond
+   k; double word - up to 7 bytes.  If the load of these bytes crosses
+   a page boundary, it might cause a memory fault (if the page is not mapped)
+   that would not have happened in byte by byte comparison.
+
+   If an address is (double) word aligned, then a load of a (double) word
+   from that address will not cross a page boundary.
+   Therefore, the algorithm below considers word and double-word alignment
+   of strings separately.  */
+
+/* High-level description of the algorithm.
+
+   * The fast path: if both strings are double-word aligned,
+     use LDRD to load two words from each string in every loop iteration.
+   * If the strings have the same offset from a word boundary,
+     use LDRB to load and compare byte by byte until
+     the first string is aligned to a word boundary (at most 3 bytes).
+     This is optimized for quick return on short unaligned strings.
+   * If the strings have the same offset from a double-word boundary,
+     use LDRD to load two words from each string in every loop iteration, as in the fast path.
+   * If the strings do not have the same offset from a double-word boundary,
+     load a word from the second string before the loop to initialize the queue.
+     Use LDRD to load two words from every string in every loop iteration.
+     Inside the loop, load the second word from the second string only after comparing
+     the first word, using the queued value, to guarantee safety across page boundaries.
+   * If the strings do not have the same offset from a word boundary,
+     use LDR and a shift queue. Order of loads and comparisons matters,
+     similarly to the previous case.
+
+   * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value.
+   * The only difference between ARM and Thumb modes is the use of CBZ instruction.
+   * The only difference between big and little endian is the use of REV in little endian
+     to compute the return value, instead of MOV.
+*/
+
+        .macro m_cbz reg label
+#ifdef __thumb2__
+        cbz     \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        beq     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbz */
+
+        .macro m_cbnz reg label
+#ifdef __thumb2__
+        cbnz    \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        bne     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbnz */
+
+        .macro  init
+        /* Macro to save temporary registers and prepare magic values.  */
+        subs    sp, sp, #16
+        strd    r4, r5, [sp, #8]
+        strd    r6, r7, [sp]
+        mvn     r6, #0  /* all F */
+        mov     r7, #0  /* all 0 */
+        .endm   /* init */
+
+        .macro  magic_compare_and_branch w1 w2 label
+        /* Macro to compare registers w1 and w2 and conditionally branch to label.  */
+        cmp     \w1, \w2        /* Are w1 and w2 the same?  */
+        magic_find_zero_bytes \w1
+        it      eq
+        cmpeq   ip, #0          /* Is there a zero byte in w1?  */
+        bne     \label
+        .endm /* magic_compare_and_branch */
+
+        .macro  magic_find_zero_bytes w1
+        /* Macro to find all-zero bytes in w1, result is in ip.  */
+#if (defined (__ARM_FEATURE_DSP))
+        uadd8   ip, \w1, r6
+        sel     ip, r7, r6
+#else /* not defined (__ARM_FEATURE_DSP) */
+        /* __ARM_FEATURE_DSP is not defined for some Cortex-M processors.
+        Coincidently, these processors only have Thumb-2 mode, where we can use the
+        the (large) magic constant available directly as an immediate in instructions.
+        Note that we cannot use the magic constant in ARM mode, where we need
+        to create the constant in a register.  */
+        sub     ip, \w1, #0x01010101
+        bic     ip, ip, \w1
+        and     ip, ip, #0x80808080
+#endif /* not defined (__ARM_FEATURE_DSP) */
+        .endm /* magic_find_zero_bytes */
+
+        .macro  setup_return w1 w2
+#ifdef __ARMEB__
+        mov     r1, \w1
+        mov     r2, \w2
+#else /* not  __ARMEB__ */
+        rev     r1, \w1
+        rev     r2, \w2
+#endif /* not  __ARMEB__ */
+        .endm /* setup_return */
+
+        pld [r0, #0]
+        pld [r1, #0]
+
+        /* Are both strings double-word aligned?  */
+        orr     ip, r0, r1
+        tst     ip, #7
+        bne     do_align
+
+        /* Fast path.  */
+        init
+
+doubleword_aligned:
+
+        /* Get here when the strings to compare are double-word aligned.  */
+        /* Compare two words in every iteration.  */
+        .p2align        2
+2:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string.  */
+        ldrd    r2, r3, [r0], #8
+        ldrd    r4, r5, [r1], #8
+
+        magic_compare_and_branch w1=r2, w2=r4, label=return_24
+        magic_compare_and_branch w1=r3, w2=r5, label=return_35
+        b       2b
+
+do_align:
+        /* Is the first string word-aligned?  */
+        ands    ip, r0, #3
+        beq     word_aligned_r0
+
+        /* Fast compare byte by byte until the first string is word-aligned.  */
+        /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes
+        to read until the next word boundary is 4-ip.  */
+        bic     r0, r0, #3
+        ldr     r2, [r0], #4
+        lsls    ip, ip, #31
+        beq     byte2
+        bcs     byte3
+
+byte1:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE1_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte2:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE2_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte3:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE3_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbnz  reg=r3, label=word_aligned_r0
+
+fast_return:
+        mov     r0, ip
+        bx      lr
+
+word_aligned_r0:
+        init
+        /* The first string is word-aligned.  */
+        /* Is the second string word-aligned?  */
+        ands    ip, r1, #3
+        bne     strcmp_unaligned
+
+word_aligned:
+        /* The strings are word-aligned. */
+        /* Is the first string double-word aligned?  */
+        tst     r0, #4
+        beq     doubleword_aligned_r0
+
+        /* If r0 is not double-word aligned yet, align it by loading
+        and comparing the next word from each string.  */
+        ldr     r2, [r0], #4
+        ldr     r4, [r1], #4
+        magic_compare_and_branch w1=r2 w2=r4 label=return_24
+
+doubleword_aligned_r0:
+        /* Get here when r0 is double-word aligned.  */
+        /* Is r1 doubleword_aligned?  */
+        tst     r1, #4
+        beq     doubleword_aligned
+
+        /* Get here when the strings to compare are word-aligned,
+        r0 is double-word aligned, but r1 is not double-word aligned.  */
+
+        /* Initialize the queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare two words in every iteration.  */
+        .p2align        2
+3:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string and compare.  */
+        ldrd    r2, r3, [r0], #8
+        magic_compare_and_branch w1=r2 w2=r5 label=return_25
+        ldrd    r4, r5, [r1], #8
+        magic_compare_and_branch w1=r3 w2=r4 label=return_34
+        b       3b
+
+        .macro miscmp_word offsetlo offsethi
+        /* Macro to compare misaligned strings.  */
+        /* r0, r1 are word-aligned, and at least one of the strings
+        is not double-word aligned.  */
+        /* Compare one word in every loop iteration.  */
+        /* OFFSETLO is the original bit-offset of r1 from a word-boundary,
+        OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word).  */
+
+        /* Initialize the shift queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare one word from each string in every loop iteration.  */
+        .p2align        2
+7:
+        ldr     r3, [r0], #4
+        S2LOMEM r5, r5, #\offsetlo
+        magic_find_zero_bytes w1=r3
+        cmp     r7, ip, S2HIMEM #\offsetlo
+        and     r2, r3, r6, S2LOMEM #\offsetlo
+        it      eq
+        cmpeq   r2, r5
+        bne     return_25
+        ldr     r5, [r1], #4
+        cmp     ip, #0
+        eor r3, r2, r3
+        S2HIMEM r2, r5, #\offsethi
+        it      eq
+        cmpeq   r3, r2
+        bne     return_32
+        b       7b
+        .endm /* miscmp_word */
+
+return_32:
+        setup_return w1=r3, w2=r2
+        b       do_return
+return_34:
+        setup_return w1=r3, w2=r4
+        b       do_return
+return_25:
+        setup_return w1=r2, w2=r5
+        b       do_return
+return_35:
+        setup_return w1=r3, w2=r5
+        b       do_return
+return_24:
+        setup_return w1=r2, w2=r4
+
+do_return:
+
+#ifdef __ARMEB__
+        mov     r0, ip
+#else /* not  __ARMEB__ */
+        rev     r0, ip
+#endif /* not  __ARMEB__ */
+
+        /* Restore temporaries early, before computing the return value.  */
+        ldrd    r6, r7, [sp]
+        ldrd    r4, r5, [sp, #8]
+        adds    sp, sp, #16
+
+        /* There is a zero or a different byte between r1 and r2.  */
+        /* r0 contains a mask of all-zero bytes in r1.  */
+        /* Using r0 and not ip here because cbz requires low register.  */
+        m_cbz   reg=r0, label=compute_return_value
+        clz     r0, r0
+        /* r0 contains the number of bits on the left of the first all-zero byte in r1.  */
+        rsb     r0, r0, #24
+        /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1.  */
+        lsr     r1, r1, r0
+        lsr     r2, r2, r0
+
+compute_return_value:
+        movs    r0, #1
+        cmp     r1, r2
+        /* The return value is computed as follows.
+        If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return.
+        If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0,
+        which means r0:=r0-r0-1 and r0 is #-1 at return.
+        If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1,
+        which means r0:=r0-r0 and r0 is #0 at return.
+        (C==0 and Z==1) cannot happen because the carry bit is "not borrow".  */
+        it      ls
+        sbcls   r0, r0, r0
+        bx      lr
+
+    /* The code from the previous version of strcmp.S handles all of the
+     * cases where the first string and seconds string cannot both be
+     * aligned to a word boundary faster than the new algorithm. See
+     * bionic/libc/arch-arm/cortex-a15/bionic/strcmp.S for the unedited
+     * version of the code.
+     */
+strcmp_unaligned:
+	wp1 .req r0
+	wp2 .req r1
+	b1  .req r2
+	w1  .req r4
+	w2  .req r5
+	t1  .req ip
+	@ r3 is scratch
+
+2:
+	mov	b1, #1
+	orr	b1, b1, b1, lsl #8
+	orr	b1, b1, b1, lsl #16
+
+	and	t1, wp2, #3
+	bic	wp2, wp2, #3
+	ldr	w1, [wp1], #4
+	ldr	w2, [wp2], #4
+	cmp	t1, #2
+	beq	2f
+	bhi	3f
+
+	/* Critical inner Loop: Block with 3 bytes initial overlap */
+	.p2align	2
+1:
+	bic	t1, w1, #MSB
+	cmp	t1, w2, S2LOMEM #8
+	sub	r3, w1, b1
+	bic	r3, r3, w1
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, S2HIMEM #24
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	1b
+4:
+	S2LOMEM	w2, w2, #8
+	b	8f
+
+5:
+#ifdef __ARMEB__
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #0xff000000
+	itt	ne
+	tstne	w1, #0x00ff0000
+	tstne	w1, #0x0000ff00
+	beq	7f
+#else
+	bics	r3, r3, #0xff000000
+	bne	7f
+#endif
+	ldrb	w2, [wp2]
+	S2LOMEM	t1, w1, #24
+#ifdef __ARMEB__
+	lsl	w2, w2, #24
+#endif
+	b	8f
+
+6:
+	S2LOMEM	t1, w1, #24
+	and	w2, w2, #LSB
+	b	8f
+
+	/* Critical inner Loop: Block with 2 bytes initial overlap */
+	.p2align	2
+2:
+	S2HIMEM	t1, w1, #16
+	sub	r3, w1, b1
+	S2LOMEM	t1, t1, #16
+	bic	r3, r3, w1
+	cmp	t1, w2, S2LOMEM #16
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, S2HIMEM #16
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	2b
+
+5:
+#ifdef __ARMEB__
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #0xff000000
+	it	ne
+	tstne	w1, #0x00ff0000
+	beq	7f
+#else
+	lsls	r3, r3, #16
+	bne	7f
+#endif
+	ldrh	w2, [wp2]
+	S2LOMEM	t1, w1, #16
+#ifdef __ARMEB__
+	lsl	w2, w2, #16
+#endif
+	b	8f
+
+6:
+	S2HIMEM	w2, w2, #16
+	S2LOMEM	t1, w1, #16
+4:
+	S2LOMEM	w2, w2, #16
+	b	8f
+
+	/* Critical inner Loop: Block with 1 byte initial overlap */
+	.p2align	2
+3:
+	and	t1, w1, #LSB
+	cmp	t1, w2, S2LOMEM #24
+	sub	r3, w1, b1
+	bic	r3, r3, w1
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, S2HIMEM #8
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	3b
+4:
+	S2LOMEM	w2, w2, #24
+	b	8f
+5:
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #LSB
+	beq	7f
+	ldr	w2, [wp2], #4
+6:
+	S2LOMEM	t1, w1, #8
+	bic	w2, w2, #MSB
+	b	8f
+7:
+	mov	r0, #0
+
+    /* Restore registers and stack. */
+    ldrd    r6, r7, [sp]
+    ldrd    r4, r5, [sp, #8]
+    adds    sp, sp, #16
+
+	bx	lr
+
+8:
+	and	r2, t1, #LSB
+	and	r0, w2, #LSB
+	cmp	r0, #1
+	it	cs
+	cmpcs	r0, r2
+	itt	eq
+	S2LOMEMEQ	t1, t1, #8
+	S2LOMEMEQ	w2, w2, #8
+	beq	8b
+	sub	r0, r2, r0
+
+    /* Restore registers and stack. */
+    ldrd    r6, r7, [sp]
+    ldrd    r4, r5, [sp, #8]
+    adds    sp, sp, #16
+
+	bx	lr
+END(strcmp)
--- a/libc/arch-arm/cortex-a9/cortex-a9.mk
+++ b/libc/arch-arm/cortex-a9/cortex-a9.mk
@@ -0,0 +1,5 @@
+$(call libc-add-cpu-variant-src,MEMCPY,arch-arm/cortex-a9/bionic/memcpy.S)
+$(call libc-add-cpu-variant-src,MEMSET,arch-arm/cortex-a9/bionic/memset.S)
+$(call libc-add-cpu-variant-src,STRCMP,arch-arm/cortex-a9/bionic/strcmp.S)
+
+include bionic/libc/arch-arm/generic/generic.mk
--- a/libc/arch-arm/generic/bionic/memcpy.S
+++ b/libc/arch-arm/generic/bionic/memcpy.S
@@ -0,0 +1,380 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+        /*
+         * Optimized memcpy() for ARM.
+         *
+         * note that memcpy() always returns the destination pointer,
+         * so we have to preserve R0.
+         */
+
+ENTRY(memcpy)
+        /* The stack must always be 64-bits aligned to be compliant with the
+         * ARM ABI. Since we have to save R0, we might as well save R4
+         * which we can use for better pipelining of the reads below
+         */
+        .save       {r0, r4, lr}
+        stmfd       sp!, {r0, r4, lr}
+        /* Making room for r5-r11 which will be spilled later */
+        .pad        #28
+        sub         sp, sp, #28
+
+        // preload the destination because we'll align it to a cache line
+        // with small writes. Also start the source "pump".
+        PLD         (r0, #0)
+        PLD         (r1, #0)
+        PLD         (r1, #32)
+
+        /* it simplifies things to take care of len<4 early */
+        cmp         r2, #4
+        blo         copy_last_3_and_return
+
+        /* compute the offset to align the source
+         * offset = (4-(src&3))&3 = -src & 3
+         */
+        rsb         r3, r1, #0
+        ands        r3, r3, #3
+        beq         src_aligned
+
+        /* align source to 32 bits. We need to insert 2 instructions between
+         * a ldr[b|h] and str[b|h] because byte and half-word instructions
+         * stall 2 cycles.
+         */
+        movs        r12, r3, lsl #31
+        sub         r2, r2, r3      /* we know that r3 <= r2 because r2 >= 4 */
+        ldrmib      r3, [r1], #1
+        ldrcsb      r4, [r1], #1
+        ldrcsb      r12,[r1], #1
+        strmib      r3, [r0], #1
+        strcsb      r4, [r0], #1
+        strcsb      r12,[r0], #1
+
+src_aligned:
+
+        /* see if src and dst are aligned together (congruent) */
+        eor         r12, r0, r1
+        tst         r12, #3
+        bne         non_congruent
+
+        /* Use post-incriment mode for stm to spill r5-r11 to reserved stack
+         * frame. Don't update sp.
+         */
+        stmea       sp, {r5-r11}
+
+        /* align the destination to a cache-line */
+        rsb         r3, r0, #0
+        ands        r3, r3, #0x1C
+        beq         congruent_aligned32
+        cmp         r3, r2
+        andhi       r3, r2, #0x1C
+
+        /* conditionally copies 0 to 7 words (length in r3) */
+        movs        r12, r3, lsl #28
+        ldmcsia     r1!, {r4, r5, r6, r7}   /* 16 bytes */
+        ldmmiia     r1!, {r8, r9}           /*  8 bytes */
+        stmcsia     r0!, {r4, r5, r6, r7}
+        stmmiia     r0!, {r8, r9}
+        tst         r3, #0x4
+        ldrne       r10,[r1], #4            /*  4 bytes */
+        strne       r10,[r0], #4
+        sub         r2, r2, r3
+
+congruent_aligned32:
+        /*
+         * here source is aligned to 32 bytes.
+         */
+
+cached_aligned32:
+        subs        r2, r2, #32
+        blo         less_than_32_left
+
+        /*
+         * We preload a cache-line up to 64 bytes ahead. On the 926, this will
+         * stall only until the requested world is fetched, but the linefill
+         * continues in the the background.
+         * While the linefill is going, we write our previous cache-line
+         * into the write-buffer (which should have some free space).
+         * When the linefill is done, the writebuffer will
+         * start dumping its content into memory
+         *
+         * While all this is going, we then load a full cache line into
+         * 8 registers, this cache line should be in the cache by now
+         * (or partly in the cache).
+         *
+         * This code should work well regardless of the source/dest alignment.
+         *
+         */
+
+        // Align the preload register to a cache-line because the cpu does
+        // "critical word first" (the first word requested is loaded first).
+        bic         r12, r1, #0x1F
+        add         r12, r12, #64
+
+1:      ldmia       r1!, { r4-r11 }
+        PLD         (r12, #64)
+        subs        r2, r2, #32
+
+        // NOTE: if r12 is more than 64 ahead of r1, the following ldrhi
+        // for ARM9 preload will not be safely guarded by the preceding subs.
+        // When it is safely guarded the only possibility to have SIGSEGV here
+        // is because the caller overstates the length.
+        ldrhi       r3, [r12], #32      /* cheap ARM9 preload */
+        stmia       r0!, { r4-r11 }
+        bhs         1b
+
+        add         r2, r2, #32
+
+
+
+
+less_than_32_left:
+        /*
+         * less than 32 bytes left at this point (length in r2)
+         */
+
+        /* skip all this if there is nothing to do, which should
+         * be a common case (if not executed the code below takes
+         * about 16 cycles)
+         */
+        tst         r2, #0x1F
+        beq         1f
+
+        /* conditionnaly copies 0 to 31 bytes */
+        movs        r12, r2, lsl #28
+        ldmcsia     r1!, {r4, r5, r6, r7}   /* 16 bytes */
+        ldmmiia     r1!, {r8, r9}           /*  8 bytes */
+        stmcsia     r0!, {r4, r5, r6, r7}
+        stmmiia     r0!, {r8, r9}
+        movs        r12, r2, lsl #30
+        ldrcs       r3, [r1], #4            /*  4 bytes */
+        ldrmih      r4, [r1], #2            /*  2 bytes */
+        strcs       r3, [r0], #4
+        strmih      r4, [r0], #2
+        tst         r2, #0x1
+        ldrneb      r3, [r1]                /*  last byte  */
+        strneb      r3, [r0]
+
+        /* we're done! restore everything and return */
+1:      ldmfd       sp!, {r5-r11}
+        ldmfd       sp!, {r0, r4, lr}
+        bx          lr
+
+        /********************************************************************/
+
+non_congruent:
+        /*
+         * here source is aligned to 4 bytes
+         * but destination is not.
+         *
+         * in the code below r2 is the number of bytes read
+         * (the number of bytes written is always smaller, because we have
+         * partial words in the shift queue)
+         */
+        cmp         r2, #4
+        blo         copy_last_3_and_return
+
+        /* Use post-incriment mode for stm to spill r5-r11 to reserved stack
+         * frame. Don't update sp.
+         */
+        stmea       sp, {r5-r11}
+
+        /* compute shifts needed to align src to dest */
+        rsb         r5, r0, #0
+        and         r5, r5, #3          /* r5 = # bytes in partial words */
+        mov         r12, r5, lsl #3     /* r12 = right */
+        rsb         lr, r12, #32        /* lr = left  */
+
+        /* read the first word */
+        ldr         r3, [r1], #4
+        sub         r2, r2, #4
+
+        /* write a partial word (0 to 3 bytes), such that destination
+         * becomes aligned to 32 bits (r5 = nb of words to copy for alignment)
+         */
+        movs        r5, r5, lsl #31
+        strmib      r3, [r0], #1
+        movmi       r3, r3, lsr #8
+        strcsb      r3, [r0], #1
+        movcs       r3, r3, lsr #8
+        strcsb      r3, [r0], #1
+        movcs       r3, r3, lsr #8
+
+        cmp         r2, #4
+        blo         partial_word_tail
+
+        /* Align destination to 32 bytes (cache line boundary) */
+1:      tst         r0, #0x1c
+        beq         2f
+        ldr         r5, [r1], #4
+        sub         r2, r2, #4
+        orr         r4, r3, r5,     lsl lr
+        mov         r3, r5,         lsr r12
+        str         r4, [r0], #4
+        cmp         r2, #4
+        bhs         1b
+        blo         partial_word_tail
+
+        /* copy 32 bytes at a time */
+2:      subs        r2, r2, #32
+        blo         less_than_thirtytwo
+
+        /* Use immediate mode for the shifts, because there is an extra cycle
+         * for register shifts, which could account for up to 50% of
+         * performance hit.
+         */
+
+        cmp         r12, #24
+        beq         loop24
+        cmp         r12, #8
+        beq         loop8
+
+loop16:
+        ldr         r12, [r1], #4
+1:      mov         r4, r12
+        ldmia       r1!, {   r5,r6,r7,  r8,r9,r10,r11}
+        PLD         (r1, #64)
+        subs        r2, r2, #32
+        ldrhs       r12, [r1], #4
+        orr         r3, r3, r4,     lsl #16
+        mov         r4, r4,         lsr #16
+        orr         r4, r4, r5,     lsl #16
+        mov         r5, r5,         lsr #16
+        orr         r5, r5, r6,     lsl #16
+        mov         r6, r6,         lsr #16
+        orr         r6, r6, r7,     lsl #16
+        mov         r7, r7,         lsr #16
+        orr         r7, r7, r8,     lsl #16
+        mov         r8, r8,         lsr #16
+        orr         r8, r8, r9,     lsl #16
+        mov         r9, r9,         lsr #16
+        orr         r9, r9, r10,    lsl #16
+        mov         r10, r10,       lsr #16
+        orr         r10, r10, r11,  lsl #16
+        stmia       r0!, {r3,r4,r5,r6, r7,r8,r9,r10}
+        mov         r3, r11,        lsr #16
+        bhs         1b
+        b           less_than_thirtytwo
+
+loop8:
+        ldr         r12, [r1], #4
+1:      mov         r4, r12
+        ldmia       r1!, {   r5,r6,r7,  r8,r9,r10,r11}
+        PLD         (r1, #64)
+        subs        r2, r2, #32
+        ldrhs       r12, [r1], #4
+        orr         r3, r3, r4,     lsl #24
+        mov         r4, r4,         lsr #8
+        orr         r4, r4, r5,     lsl #24
+        mov         r5, r5,         lsr #8
+        orr         r5, r5, r6,     lsl #24
+        mov         r6, r6,         lsr #8
+        orr         r6, r6, r7,     lsl #24
+        mov         r7, r7,         lsr #8
+        orr         r7, r7, r8,     lsl #24
+        mov         r8, r8,         lsr #8
+        orr         r8, r8, r9,     lsl #24
+        mov         r9, r9,         lsr #8
+        orr         r9, r9, r10,    lsl #24
+        mov         r10, r10,       lsr #8
+        orr         r10, r10, r11,  lsl #24
+        stmia       r0!, {r3,r4,r5,r6, r7,r8,r9,r10}
+        mov         r3, r11,        lsr #8
+        bhs         1b
+        b           less_than_thirtytwo
+
+loop24:
+        ldr         r12, [r1], #4
+1:      mov         r4, r12
+        ldmia       r1!, {   r5,r6,r7,  r8,r9,r10,r11}
+        PLD         (r1, #64)
+        subs        r2, r2, #32
+        ldrhs       r12, [r1], #4
+        orr         r3, r3, r4,     lsl #8
+        mov         r4, r4,         lsr #24
+        orr         r4, r4, r5,     lsl #8
+        mov         r5, r5,         lsr #24
+        orr         r5, r5, r6,     lsl #8
+        mov         r6, r6,         lsr #24
+        orr         r6, r6, r7,     lsl #8
+        mov         r7, r7,         lsr #24
+        orr         r7, r7, r8,     lsl #8
+        mov         r8, r8,         lsr #24
+        orr         r8, r8, r9,     lsl #8
+        mov         r9, r9,         lsr #24
+        orr         r9, r9, r10,    lsl #8
+        mov         r10, r10,       lsr #24
+        orr         r10, r10, r11,  lsl #8
+        stmia       r0!, {r3,r4,r5,r6, r7,r8,r9,r10}
+        mov         r3, r11,        lsr #24
+        bhs         1b
+
+
+less_than_thirtytwo:
+        /* copy the last 0 to 31 bytes of the source */
+        rsb         r12, lr, #32        /* we corrupted r12, recompute it  */
+        add         r2, r2, #32
+        cmp         r2, #4
+        blo         partial_word_tail
+
+1:      ldr         r5, [r1], #4
+        sub         r2, r2, #4
+        orr         r4, r3, r5,     lsl lr
+        mov         r3, r5,         lsr r12
+        str         r4, [r0], #4
+        cmp         r2, #4
+        bhs         1b
+
+partial_word_tail:
+        /* we have a partial word in the input buffer */
+        movs        r5, lr, lsl #(31-3)
+        strmib      r3, [r0], #1
+        movmi       r3, r3, lsr #8
+        strcsb      r3, [r0], #1
+        movcs       r3, r3, lsr #8
+        strcsb      r3, [r0], #1
+
+        /* Refill spilled registers from the stack. Don't update sp. */
+        ldmfd       sp, {r5-r11}
+
+copy_last_3_and_return:
+        movs        r2, r2, lsl #31 /* copy remaining 0, 1, 2 or 3 bytes */
+        ldrmib      r2, [r1], #1
+        ldrcsb      r3, [r1], #1
+        ldrcsb      r12,[r1]
+        strmib      r2, [r0], #1
+        strcsb      r3, [r0], #1
+        strcsb      r12,[r0]
+
+        /* we're done! restore sp and spilled registers and return */
+        add         sp,  sp, #28
+        ldmfd       sp!, {r0, r4, lr}
+        bx          lr
+END(memcpy)
--- a/libc/arch-arm/generic/bionic/memset.S
+++ b/libc/arch-arm/generic/bionic/memset.S
@@ -0,0 +1,109 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/asm.h>
+
+        /*
+         * Optimized memset() for ARM.
+         *
+         * memset() returns its first argument.
+         */
+
+ENTRY(bzero)
+        mov     r2, r1
+        mov     r1, #0
+END(bzero)
+
+ENTRY(memset)
+        /* compute the offset to align the destination
+         * offset = (4-(src&3))&3 = -src & 3
+         */
+        .save       {r0, r4-r7, lr}
+        stmfd       sp!, {r0, r4-r7, lr}
+        rsb         r3, r0, #0
+        ands        r3, r3, #3
+        cmp         r3, r2
+        movhi       r3, r2
+
+        /* splat r1 */
+        mov         r1, r1, lsl #24
+        orr         r1, r1, r1, lsr #8
+        orr         r1, r1, r1, lsr #16
+
+        movs        r12, r3, lsl #31
+        strcsb      r1, [r0], #1    /* can't use strh (alignment unknown) */
+        strcsb      r1, [r0], #1
+        strmib      r1, [r0], #1
+        subs        r2, r2, r3
+        ldmlsfd     sp!, {r0, r4-r7, lr}    /* return */
+        bxls        lr
+
+        /* align the destination to a cache-line */
+        mov         r12, r1
+        mov         lr, r1
+        mov         r4, r1
+        mov         r5, r1
+        mov         r6, r1
+        mov         r7, r1
+
+        rsb         r3, r0, #0
+        ands        r3, r3, #0x1C
+        beq         3f
+        cmp         r3, r2
+        andhi       r3, r2, #0x1C
+        sub         r2, r2, r3
+
+        /* conditionally writes 0 to 7 words (length in r3) */
+        movs        r3, r3, lsl #28
+        stmcsia     r0!, {r1, lr}
+        stmcsia     r0!, {r1, lr}
+        stmmiia     r0!, {r1, lr}
+        movs        r3, r3, lsl #2
+        strcs       r1, [r0], #4
+
+3:
+        subs        r2, r2, #32
+        mov         r3, r1
+        bmi         2f
+1:      subs        r2, r2, #32
+        stmia       r0!, {r1,r3,r4,r5,r6,r7,r12,lr}
+        bhs         1b
+2:      add         r2, r2, #32
+
+        /* conditionally stores 0 to 31 bytes */
+        movs        r2, r2, lsl #28
+        stmcsia     r0!, {r1,r3,r12,lr}
+        stmmiia     r0!, {r1, lr}
+        movs        r2, r2, lsl #2
+        strcs       r1, [r0], #4
+        strmih      r1, [r0], #2
+        movs        r2, r2, lsl #2
+        strcsb      r1, [r0]
+        ldmfd       sp!, {r0, r4-r7, lr}
+        bx          lr
+END(memset)
--- a/libc/arch-arm/generic/bionic/strcmp.S
+++ b/libc/arch-arm/generic/bionic/strcmp.S
@@ -0,0 +1,317 @@
+/*
+ * Copyright (c) 2011 The Android Open Source Project
+ * Copyright (c) 2008 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+	.text
+
+#ifdef __ARMEB__
+#define SHFT2LSB lsl
+#define SHFT2LSBEQ lsleq
+#define SHFT2MSB lsr
+#define SHFT2MSBEQ lsreq
+#define MSB 0x000000ff
+#define LSB 0xff000000
+#else
+#define SHFT2LSB lsr
+#define SHFT2LSBEQ lsreq
+#define SHFT2MSB lsl
+#define SHFT2MSBEQ lsleq
+#define MSB 0xff000000
+#define LSB 0x000000ff
+#endif
+
+#define magic1(REG) REG
+#define magic2(REG) REG, lsl #7
+
+ENTRY(strcmp)
+	PLD(r0, #0)
+	PLD(r1, #0)
+	eor	r2, r0, r1
+	tst	r2, #3
+
+	/* Strings not at same byte offset from a word boundary.  */
+	bne	.Lstrcmp_unaligned
+	ands	r2, r0, #3
+	bic	r0, r0, #3
+	bic	r1, r1, #3
+	ldr	ip, [r0], #4
+	it	eq
+	ldreq	r3, [r1], #4
+	beq	1f
+
+	/* Although s1 and s2 have identical initial alignment, they are
+	 * not currently word aligned.  Rather than comparing bytes,
+	 * make sure that any bytes fetched from before the addressed
+	 * bytes are forced to 0xff.  Then they will always compare
+	 * equal.
+	 */
+	eor	r2, r2, #3
+	lsl	r2, r2, #3
+	mvn	r3, #MSB
+	SHFT2LSB	r2, r3, r2
+	ldr	r3, [r1], #4
+	orr	ip, ip, r2
+	orr	r3, r3, r2
+1:
+	/* Load the 'magic' constant 0x01010101. */
+	str	r4, [sp, #-4]!
+	mov	r4, #1
+	orr	r4, r4, r4, lsl #8
+	orr	r4, r4, r4, lsl #16
+	.p2align	2
+4:
+	PLD(r0, #8)
+	PLD(r1, #8)
+	sub	r2, ip, magic1(r4)
+	cmp	ip, r3
+	itttt	eq
+
+	/* check for any zero bytes in first word */
+	biceq	r2, r2, ip
+	tsteq	r2, magic2(r4)
+	ldreq	ip, [r0], #4
+	ldreq	r3, [r1], #4
+	beq	4b
+2:
+	/* There's a zero or a different byte in the word */
+	SHFT2MSB	r0, ip, #24
+	SHFT2LSB	ip, ip, #8
+	cmp	r0, #1
+	it	cs
+	cmpcs	r0, r3, SHFT2MSB #24
+	it	eq
+	SHFT2LSBEQ r3, r3, #8
+	beq	2b
+	/* On a big-endian machine, r0 contains the desired byte in bits
+	 * 0-7; on a little-endian machine they are in bits 24-31.  In
+	 * both cases the other bits in r0 are all zero.  For r3 the
+	 * interesting byte is at the other end of the word, but the
+	 * other bits are not necessarily zero.  We need a signed result
+	 * representing the differnece in the unsigned bytes, so for the
+	 * little-endian case we can't just shift the interesting bits up.
+	 */
+#ifdef __ARMEB__
+	sub	r0, r0, r3, lsr #24
+#else
+	and	r3, r3, #255
+	/* No RSB instruction in Thumb2 */
+#ifdef __thumb2__
+	lsr	r0, r0, #24
+	sub	r0, r0, r3
+#else
+	rsb	r0, r3, r0, lsr #24
+#endif
+#endif
+	ldr	r4, [sp], #4
+	bx	lr
+
+.Lstrcmp_unaligned:
+	wp1 .req r0
+	wp2 .req r1
+	b1  .req r2
+	w1  .req r4
+	w2  .req r5
+	t1  .req ip
+	@ r3 is scratch
+
+	/* First of all, compare bytes until wp1(sp1) is word-aligned. */
+1:
+	tst	wp1, #3
+	beq	2f
+	ldrb	r2, [wp1], #1
+	ldrb	r3, [wp2], #1
+	cmp	r2, #1
+	it	cs
+	cmpcs	r2, r3
+	beq	1b
+	sub	r0, r2, r3
+	bx	lr
+
+2:
+	str	r5, [sp, #-4]!
+	str	r4, [sp, #-4]!
+	mov	b1, #1
+	orr	b1, b1, b1, lsl #8
+	orr	b1, b1, b1, lsl #16
+
+	and	t1, wp2, #3
+	bic	wp2, wp2, #3
+	ldr	w1, [wp1], #4
+	ldr	w2, [wp2], #4
+	cmp	t1, #2
+	beq	2f
+	bhi	3f
+
+	/* Critical inner Loop: Block with 3 bytes initial overlap */
+	.p2align	2
+1:
+	bic	t1, w1, #MSB
+	cmp	t1, w2, SHFT2LSB #8
+	sub	r3, w1, b1
+	bic	r3, r3, w1
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, SHFT2MSB #24
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	1b
+4:
+	SHFT2LSB	w2, w2, #8
+	b	8f
+
+5:
+#ifdef __ARMEB__
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #0xff000000
+	itt	ne
+	tstne	w1, #0x00ff0000
+	tstne	w1, #0x0000ff00
+	beq	7f
+#else
+	bics	r3, r3, #0xff000000
+	bne	7f
+#endif
+	ldrb	w2, [wp2]
+	SHFT2LSB	t1, w1, #24
+#ifdef __ARMEB__
+	lsl	w2, w2, #24
+#endif
+	b	8f
+
+6:
+	SHFT2LSB	t1, w1, #24
+	and	w2, w2, #LSB
+	b	8f
+
+	/* Critical inner Loop: Block with 2 bytes initial overlap */
+	.p2align	2
+2:
+	SHFT2MSB	t1, w1, #16
+	sub	r3, w1, b1
+	SHFT2LSB	t1, t1, #16
+	bic	r3, r3, w1
+	cmp	t1, w2, SHFT2LSB #16
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, SHFT2MSB #16
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	2b
+
+5:
+#ifdef __ARMEB__
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #0xff000000
+	it	ne
+	tstne	w1, #0x00ff0000
+	beq	7f
+#else
+	lsls	r3, r3, #16
+	bne	7f
+#endif
+	ldrh	w2, [wp2]
+	SHFT2LSB	t1, w1, #16
+#ifdef __ARMEB__
+	lsl	w2, w2, #16
+#endif
+	b	8f
+
+6:
+	SHFT2MSB	w2, w2, #16
+	SHFT2LSB	t1, w1, #16
+4:
+	SHFT2LSB	w2, w2, #16
+	b	8f
+
+	/* Critical inner Loop: Block with 1 byte initial overlap */
+	.p2align	2
+3:
+	and	t1, w1, #LSB
+	cmp	t1, w2, SHFT2LSB #24
+	sub	r3, w1, b1
+	bic	r3, r3, w1
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, SHFT2MSB #8
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	3b
+4:
+	SHFT2LSB	w2, w2, #24
+	b	8f
+5:
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #LSB
+	beq	7f
+	ldr	w2, [wp2], #4
+6:
+	SHFT2LSB	t1, w1, #8
+	bic	w2, w2, #MSB
+	b	8f
+7:
+	mov	r0, #0
+	ldr	r4, [sp], #4
+	ldr	r5, [sp], #4
+	bx	lr
+
+8:
+	and	r2, t1, #LSB
+	and	r0, w2, #LSB
+	cmp	r0, #1
+	it	cs
+	cmpcs	r0, r2
+	itt	eq
+	SHFT2LSBEQ	t1, t1, #8
+	SHFT2LSBEQ	w2, w2, #8
+	beq	8b
+	sub	r0, r2, r0
+	ldr	r4, [sp], #4
+	ldr	r5, [sp], #4
+	bx	lr
+END(strcmp)
--- a/libc/arch-arm/generic/generic.mk
+++ b/libc/arch-arm/generic/generic.mk
@@ -0,0 +1,3 @@
+$(call libc-add-cpu-variant-src,MEMCPY,arch-arm/generic/bionic/memcpy.S)
+$(call libc-add-cpu-variant-src,MEMSET,arch-arm/generic/bionic/memset.S)
+$(call libc-add-cpu-variant-src,STRCMP,arch-arm/generic/bionic/strcmp.S)
--- a/libc/arch-arm/krait/bionic/memcpy.S
+++ b/libc/arch-arm/krait/bionic/memcpy.S
@@ -0,0 +1,146 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+/* Assumes neon instructions and a cache line size of 32 bytes. */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+/*
+ * This code assumes it is running on a processor that supports all arm v7
+ * instructions, that supports neon instructions, and that has a 32 byte
+ * cache line.
+ */
+
+        .text
+        .fpu    neon
+
+#define CACHE_LINE_SIZE     32
+
+ENTRY(memcpy)
+        .save       {r0, lr}
+        /* start preloading as early as possible */
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        stmfd       sp!, {r0, lr}
+        pld         [r1, #(CACHE_LINE_SIZE*2)]
+
+        /* do we have at least 16-bytes to copy (needed for alignment below) */
+        cmp         r2, #16
+        blo         5f
+
+        /* align destination to cache-line for the write-buffer */
+        rsb         r3, r0, #0
+        ands        r3, r3, #0xF
+        beq         0f
+
+        /* copy up to 15-bytes (count in r3) */
+        sub         r2, r2, r3
+        movs        ip, r3, lsl #31
+        ldrmib      lr, [r1], #1
+        strmib      lr, [r0], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+        movs        ip, r3, lsl #29
+        bge         1f
+        // copies 4 bytes, destination 32-bits aligned
+        vld4.8      {d0[0], d1[0], d2[0], d3[0]}, [r1]!
+        vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r0, :32]!
+1:      bcc         2f
+        // copies 8 bytes, destination 64-bits aligned
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0, :64]!
+2:
+
+0:      /* preload immediately the next cache line, which we may need */
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        pld         [r1, #(CACHE_LINE_SIZE*2)]
+
+        /* make sure we have at least 64 bytes to copy */
+        subs        r2, r2, #64
+        blo         2f
+
+        /* Preload all the cache lines we need.
+         * NOTE: The number of pld below depends on CACHE_LINE_SIZE,
+         * ideally we would increase the distance in the main loop to
+         * avoid the goofy code below. In practice this doesn't seem to make
+         * a big difference.
+         * NOTE: The value CACHE_LINE_SIZE * 8 was chosen through
+         * experimentation.
+         */
+        pld         [r1, #(CACHE_LINE_SIZE*4)]
+        pld         [r1, #(CACHE_LINE_SIZE*6)]
+        pld         [r1, #(CACHE_LINE_SIZE*8)]
+
+1:      /* The main loop copies 64 bytes at a time */
+        vld1.8      {d0  - d3},   [r1]!
+        vld1.8      {d4  - d7},   [r1]!
+        pld         [r1, #(CACHE_LINE_SIZE*8)]
+        subs        r2, r2, #64
+        vst1.8      {d0  - d3},   [r0, :128]!
+        vst1.8      {d4  - d7},   [r0, :128]!
+        bhs         1b
+
+2:      /* fix-up the remaining count and make sure we have >= 32 bytes left */
+        add         r2, r2, #64
+        subs        r2, r2, #32
+        blo         4f
+
+3:      /* 32 bytes at a time. These cache lines were already preloaded */
+        vld1.8      {d0 - d3},  [r1]!
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3},  [r0, :128]!
+        bhs         3b
+4:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         5f
+        // copies 16 bytes, 128-bits aligned
+        vld1.8      {d0, d1}, [r1]!
+        vst1.8      {d0, d1}, [r0, :128]!
+
+5:      /* copy up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vld1.8      {d0}, [r1]!
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vld4.8      {d0[0], d1[0], d2[0], d3[0]}, [r1]!
+        vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        ldrmib      r3, [r1], #1
+        ldrcsb      ip, [r1], #1
+        ldrcsb      lr, [r1], #1
+        strmib      r3, [r0], #1
+        strcsb      ip, [r0], #1
+        strcsb      lr, [r0], #1
+
+        ldmfd       sp!, {r0, lr}
+        bx          lr
+END(memcpy)
--- a/libc/arch-arm/krait/bionic/memset.S
+++ b/libc/arch-arm/krait/bionic/memset.S
@@ -0,0 +1,81 @@
+/*
+ * Copyright (C) 2008 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+/*
+ * This code assumes it is running on a processor that supports all arm v7
+ * instructions, that supports neon instructions, and that supports
+ * unaligned neon instruction accesses to memory.
+ */
+
+    .fpu    neon
+
+ENTRY(bzero)
+        mov     r2, r1
+        mov     r1, #0
+END(bzero)
+
+/* memset() returns its first argument.  */
+ENTRY(memset)
+        .save       {r0}
+        stmfd       sp!, {r0}
+
+        vdup.8      q0, r1
+
+        /* make sure we have at least 32 bytes to write */
+        subs        r2, r2, #32
+        blo         2f
+        vmov        q1, q0
+
+1:      /* The main loop writes 32 bytes at a time */
+        subs        r2, r2, #32
+        vst1.8      {d0 - d3}, [r0]!
+        bhs         1b
+
+2:      /* less than 32 left */
+        add         r2, r2, #32
+        tst         r2, #0x10
+        beq         3f
+
+        // writes 16 bytes, 128-bits aligned
+        vst1.8      {d0, d1}, [r0]!
+3:      /* write up to 15-bytes (count in r2) */
+        movs        ip, r2, lsl #29
+        bcc         1f
+        vst1.8      {d0}, [r0]!
+1:      bge         2f
+        vst1.32     {d0[0]}, [r0]!
+2:      movs        ip, r2, lsl #31
+        strmib      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        strcsb      r1, [r0], #1
+        ldmfd       sp!, {r0}
+        bx          lr
+END(memset)
--- a/libc/arch-arm/krait/bionic/strcmp.S
+++ b/libc/arch-arm/krait/bionic/strcmp.S
@@ -0,0 +1,477 @@
+/*
+ * Copyright (c) 2013 ARM Ltd
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. The name of the company may not be used to endorse or promote
+ *    products derived from this software without specific prior written
+ *    permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
+ * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
+ * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+ * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <machine/cpu-features.h>
+#include <machine/asm.h>
+
+#ifdef __ARMEB__
+#define S2LOMEM lsl
+#define S2LOMEMEQ lsleq
+#define S2HIMEM lsr
+#define MSB 0x000000ff
+#define LSB 0xff000000
+#define BYTE0_OFFSET 24
+#define BYTE1_OFFSET 16
+#define BYTE2_OFFSET 8
+#define BYTE3_OFFSET 0
+#else /* not  __ARMEB__ */
+#define S2LOMEM lsr
+#define S2LOMEMEQ lsreq
+#define S2HIMEM lsl
+#define BYTE0_OFFSET 0
+#define BYTE1_OFFSET 8
+#define BYTE2_OFFSET 16
+#define BYTE3_OFFSET 24
+#define MSB 0xff000000
+#define LSB 0x000000ff
+#endif /* not  __ARMEB__ */
+
+.syntax         unified
+
+#if defined (__thumb__)
+        .thumb
+        .thumb_func
+#endif
+
+ENTRY(strcmp)
+      /* Use LDRD whenever possible.  */
+
+/* The main thing to look out for when comparing large blocks is that
+   the loads do not cross a page boundary when loading past the index
+   of the byte with the first difference or the first string-terminator.
+
+   For example, if the strings are identical and the string-terminator
+   is at index k, byte by byte comparison will not load beyond address
+   s1+k and s2+k; word by word comparison may load up to 3 bytes beyond
+   k; double word - up to 7 bytes.  If the load of these bytes crosses
+   a page boundary, it might cause a memory fault (if the page is not mapped)
+   that would not have happened in byte by byte comparison.
+
+   If an address is (double) word aligned, then a load of a (double) word
+   from that address will not cross a page boundary.
+   Therefore, the algorithm below considers word and double-word alignment
+   of strings separately.  */
+
+/* High-level description of the algorithm.
+
+   * The fast path: if both strings are double-word aligned,
+     use LDRD to load two words from each string in every loop iteration.
+   * If the strings have the same offset from a word boundary,
+     use LDRB to load and compare byte by byte until
+     the first string is aligned to a word boundary (at most 3 bytes).
+     This is optimized for quick return on short unaligned strings.
+   * If the strings have the same offset from a double-word boundary,
+     use LDRD to load two words from each string in every loop iteration, as in the fast path.
+   * If the strings do not have the same offset from a double-word boundary,
+     load a word from the second string before the loop to initialize the queue.
+     Use LDRD to load two words from every string in every loop iteration.
+     Inside the loop, load the second word from the second string only after comparing
+     the first word, using the queued value, to guarantee safety across page boundaries.
+   * If the strings do not have the same offset from a word boundary,
+     use LDR and a shift queue. Order of loads and comparisons matters,
+     similarly to the previous case.
+
+   * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value.
+   * The only difference between ARM and Thumb modes is the use of CBZ instruction.
+   * The only difference between big and little endian is the use of REV in little endian
+     to compute the return value, instead of MOV.
+*/
+
+        .macro m_cbz reg label
+#ifdef __thumb2__
+        cbz     \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        beq     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbz */
+
+        .macro m_cbnz reg label
+#ifdef __thumb2__
+        cbnz    \reg, \label
+#else   /* not defined __thumb2__ */
+        cmp     \reg, #0
+        bne     \label
+#endif /* not defined __thumb2__ */
+        .endm /* m_cbnz */
+
+        .macro  init
+        /* Macro to save temporary registers and prepare magic values.  */
+        subs    sp, sp, #16
+        strd    r4, r5, [sp, #8]
+        strd    r6, r7, [sp]
+        mvn     r6, #0  /* all F */
+        mov     r7, #0  /* all 0 */
+        .endm   /* init */
+
+        .macro  magic_compare_and_branch w1 w2 label
+        /* Macro to compare registers w1 and w2 and conditionally branch to label.  */
+        cmp     \w1, \w2        /* Are w1 and w2 the same?  */
+        magic_find_zero_bytes \w1
+        it      eq
+        cmpeq   ip, #0          /* Is there a zero byte in w1?  */
+        bne     \label
+        .endm /* magic_compare_and_branch */
+
+        .macro  magic_find_zero_bytes w1
+        /* Macro to find all-zero bytes in w1, result is in ip.  */
+#if (defined (__ARM_FEATURE_DSP))
+        uadd8   ip, \w1, r6
+        sel     ip, r7, r6
+#else /* not defined (__ARM_FEATURE_DSP) */
+        /* __ARM_FEATURE_DSP is not defined for some Cortex-M processors.
+        Coincidently, these processors only have Thumb-2 mode, where we can use the
+        the (large) magic constant available directly as an immediate in instructions.
+        Note that we cannot use the magic constant in ARM mode, where we need
+        to create the constant in a register.  */
+        sub     ip, \w1, #0x01010101
+        bic     ip, ip, \w1
+        and     ip, ip, #0x80808080
+#endif /* not defined (__ARM_FEATURE_DSP) */
+        .endm /* magic_find_zero_bytes */
+
+        .macro  setup_return w1 w2
+#ifdef __ARMEB__
+        mov     r1, \w1
+        mov     r2, \w2
+#else /* not  __ARMEB__ */
+        rev     r1, \w1
+        rev     r2, \w2
+#endif /* not  __ARMEB__ */
+        .endm /* setup_return */
+
+        pld [r0, #0]
+        pld [r1, #0]
+
+        /* Are both strings double-word aligned?  */
+        orr     ip, r0, r1
+        tst     ip, #7
+        bne     do_align
+
+        /* Fast path.  */
+        init
+
+doubleword_aligned:
+
+        /* Get here when the strings to compare are double-word aligned.  */
+        /* Compare two words in every iteration.  */
+        .p2align        2
+2:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string.  */
+        ldrd    r2, r3, [r0], #8
+        ldrd    r4, r5, [r1], #8
+
+        magic_compare_and_branch w1=r2, w2=r4, label=return_24
+        magic_compare_and_branch w1=r3, w2=r5, label=return_35
+        b       2b
+
+do_align:
+        /* Is the first string word-aligned?  */
+        ands    ip, r0, #3
+        beq     word_aligned_r0
+
+        /* Fast compare byte by byte until the first string is word-aligned.  */
+        /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes
+        to read until the next word boundary is 4-ip.  */
+        bic     r0, r0, #3
+        ldr     r2, [r0], #4
+        lsls    ip, ip, #31
+        beq     byte2
+        bcs     byte3
+
+byte1:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE1_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte2:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE2_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbz   reg=r3, label=fast_return
+
+byte3:
+        ldrb    ip, [r1], #1
+        uxtb    r3, r2, ror #BYTE3_OFFSET
+        subs    ip, r3, ip
+        bne     fast_return
+        m_cbnz  reg=r3, label=word_aligned_r0
+
+fast_return:
+        mov     r0, ip
+        bx      lr
+
+word_aligned_r0:
+        init
+        /* The first string is word-aligned.  */
+        /* Is the second string word-aligned?  */
+        ands    ip, r1, #3
+        bne     strcmp_unaligned
+
+word_aligned:
+        /* The strings are word-aligned. */
+        /* Is the first string double-word aligned?  */
+        tst     r0, #4
+        beq     doubleword_aligned_r0
+
+        /* If r0 is not double-word aligned yet, align it by loading
+        and comparing the next word from each string.  */
+        ldr     r2, [r0], #4
+        ldr     r4, [r1], #4
+        magic_compare_and_branch w1=r2 w2=r4 label=return_24
+
+doubleword_aligned_r0:
+        /* Get here when r0 is double-word aligned.  */
+        /* Is r1 doubleword_aligned?  */
+        tst     r1, #4
+        beq     doubleword_aligned
+
+        /* Get here when the strings to compare are word-aligned,
+        r0 is double-word aligned, but r1 is not double-word aligned.  */
+
+        /* Initialize the queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare two words in every iteration.  */
+        .p2align        2
+3:
+        pld [r0, #16]
+        pld [r1, #16]
+
+        /* Load the next double-word from each string and compare.  */
+        ldrd    r2, r3, [r0], #8
+        magic_compare_and_branch w1=r2 w2=r5 label=return_25
+        ldrd    r4, r5, [r1], #8
+        magic_compare_and_branch w1=r3 w2=r4 label=return_34
+        b       3b
+
+        .macro miscmp_word offsetlo offsethi
+        /* Macro to compare misaligned strings.  */
+        /* r0, r1 are word-aligned, and at least one of the strings
+        is not double-word aligned.  */
+        /* Compare one word in every loop iteration.  */
+        /* OFFSETLO is the original bit-offset of r1 from a word-boundary,
+        OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word).  */
+
+        /* Initialize the shift queue.  */
+        ldr     r5, [r1], #4
+
+        /* Compare one word from each string in every loop iteration.  */
+        .p2align        2
+7:
+        ldr     r3, [r0], #4
+        S2LOMEM r5, r5, #\offsetlo
+        magic_find_zero_bytes w1=r3
+        cmp     r7, ip, S2HIMEM #\offsetlo
+        and     r2, r3, r6, S2LOMEM #\offsetlo
+        it      eq
+        cmpeq   r2, r5
+        bne     return_25
+        ldr     r5, [r1], #4
+        cmp     ip, #0
+        eor r3, r2, r3
+        S2HIMEM r2, r5, #\offsethi
+        it      eq
+        cmpeq   r3, r2
+        bne     return_32
+        b       7b
+        .endm /* miscmp_word */
+
+strcmp_unaligned:
+        /* r0 is word-aligned, r1 is at offset ip from a word.  */
+        /* Align r1 to the (previous) word-boundary.  */
+        bic     r1, r1, #3
+
+        /* Unaligned comparison word by word using LDRs. */
+        cmp     ip, #2
+        beq     miscmp_word_16                    /* If ip == 2.  */
+        bge     miscmp_word_24                    /* If ip == 3.  */
+        miscmp_word offsetlo=8 offsethi=24        /* If ip == 1.  */
+miscmp_word_24:  miscmp_word offsetlo=24 offsethi=8
+
+
+return_32:
+        setup_return w1=r3, w2=r2
+        b       do_return
+return_34:
+        setup_return w1=r3, w2=r4
+        b       do_return
+return_25:
+        setup_return w1=r2, w2=r5
+        b       do_return
+return_35:
+        setup_return w1=r3, w2=r5
+        b       do_return
+return_24:
+        setup_return w1=r2, w2=r4
+
+do_return:
+
+#ifdef __ARMEB__
+        mov     r0, ip
+#else /* not  __ARMEB__ */
+        rev     r0, ip
+#endif /* not  __ARMEB__ */
+
+        /* Restore temporaries early, before computing the return value.  */
+        ldrd    r6, r7, [sp]
+        ldrd    r4, r5, [sp, #8]
+        adds    sp, sp, #16
+
+        /* There is a zero or a different byte between r1 and r2.  */
+        /* r0 contains a mask of all-zero bytes in r1.  */
+        /* Using r0 and not ip here because cbz requires low register.  */
+        m_cbz   reg=r0, label=compute_return_value
+        clz     r0, r0
+        /* r0 contains the number of bits on the left of the first all-zero byte in r1.  */
+        rsb     r0, r0, #24
+        /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1.  */
+        lsr     r1, r1, r0
+        lsr     r2, r2, r0
+
+compute_return_value:
+        movs    r0, #1
+        cmp     r1, r2
+        /* The return value is computed as follows.
+        If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return.
+        If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0,
+        which means r0:=r0-r0-1 and r0 is #-1 at return.
+        If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1,
+        which means r0:=r0-r0 and r0 is #0 at return.
+        (C==0 and Z==1) cannot happen because the carry bit is "not borrow".  */
+        it      ls
+        sbcls   r0, r0, r0
+        bx      lr
+
+    /* The code from the previous version of strcmp.S handles this
+     * particular case (the second string is 2 bytes off a word alignment)
+     * faster than any current version. In this very specific case, use the
+     * previous version. See bionic/libc/arch-arm/cortex-a15/bionic/strcmp.S
+     * for the unedited version of this code.
+     */
+miscmp_word_16:
+	wp1 .req r0
+	wp2 .req r1
+	b1  .req r2
+	w1  .req r4
+	w2  .req r5
+	t1  .req ip
+	@ r3 is scratch
+
+    /* At this point, wp1 (r0) has already been word-aligned. */
+2:
+	mov	b1, #1
+	orr	b1, b1, b1, lsl #8
+	orr	b1, b1, b1, lsl #16
+
+	and	t1, wp2, #3
+	bic	wp2, wp2, #3
+	ldr	w1, [wp1], #4
+	ldr	w2, [wp2], #4
+
+	/* Critical inner Loop: Block with 2 bytes initial overlap */
+	.p2align	2
+2:
+	S2HIMEM	t1, w1, #16
+	sub	r3, w1, b1
+	S2LOMEM	t1, t1, #16
+	bic	r3, r3, w1
+	cmp	t1, w2, S2LOMEM #16
+	bne	4f
+	ands	r3, r3, b1, lsl #7
+	it	eq
+	ldreq	w2, [wp2], #4
+	bne	5f
+	eor	t1, t1, w1
+	cmp	t1, w2, S2HIMEM #16
+	bne	6f
+	ldr	w1, [wp1], #4
+	b	2b
+
+5:
+#ifdef __ARMEB__
+	/* The syndrome value may contain false ones if the string ends
+	 * with the bytes 0x01 0x00
+	 */
+	tst	w1, #0xff000000
+	it	ne
+	tstne	w1, #0x00ff0000
+	beq	7f
+#else
+	lsls	r3, r3, #16
+	bne	7f
+#endif
+	ldrh	w2, [wp2]
+	S2LOMEM	t1, w1, #16
+#ifdef __ARMEB__
+	lsl	w2, w2, #16
+#endif
+	b	8f
+
+6:
+	S2HIMEM	w2, w2, #16
+	S2LOMEM	t1, w1, #16
+4:
+	S2LOMEM	w2, w2, #16
+	b	8f
+
+7:
+	mov	r0, #0
+
+    /* Restore registers and stack. */
+    ldrd    r6, r7, [sp]
+    ldrd    r4, r5, [sp, #8]
+    adds    sp, sp, #16
+
+	bx	lr
+
+8:
+	and	r2, t1, #LSB
+	and	r0, w2, #LSB
+	cmp	r0, #1
+	it	cs
+	cmpcs	r0, r2
+	itt	eq
+	S2LOMEMEQ	t1, t1, #8
+	S2LOMEMEQ	w2, w2, #8
+	beq	8b
+	sub	r0, r2, r0
+
+    /* Restore registers and stack. */
+    ldrd    r6, r7, [sp]
+    ldrd    r4, r5, [sp, #8]
+    adds    sp, sp, #16
+
+	bx	lr
+END(strcmp)
--- a/libc/arch-arm/krait/krait.mk
+++ b/libc/arch-arm/krait/krait.mk
@@ -0,0 +1,5 @@
+$(call libc-add-cpu-variant-src,MEMCPY,arch-arm/krait/bionic/memcpy.S)
+$(call libc-add-cpu-variant-src,MEMSET,arch-arm/krait/bionic/memset.S)
+$(call libc-add-cpu-variant-src,STRCMP,arch-arm/krait/bionic/strcmp.S)
+
+include bionic/libc/arch-arm/generic/generic.mk
--- a/libc/arch-arm/syscalls.mk
+++ b/libc/arch-arm/syscalls.mk
@@ -4,6 +4,7 @@ syscall_src += arch-arm/syscalls/_exit.S
 syscall_src += arch-arm/syscalls/_exit_thread.S
 syscall_src += arch-arm/syscalls/__fork.S
 syscall_src += arch-arm/syscalls/__waitid.S
+syscall_src += arch-arm/syscalls/wait4.S
 syscall_src += arch-arm/syscalls/__sys_clone.S
 syscall_src += arch-arm/syscalls/execve.S
 syscall_src += arch-arm/syscalls/__setuid.S
@@ -180,7 +181,6 @@ syscall_src += arch-arm/syscalls/__getcpu.S
 syscall_src += arch-arm/syscalls/ioprio_set.S
 syscall_src += arch-arm/syscalls/ioprio_get.S
 syscall_src += arch-arm/syscalls/uname.S
-syscall_src += arch-arm/syscalls/__wait4.S
 syscall_src += arch-arm/syscalls/umask.S
 syscall_src += arch-arm/syscalls/__reboot.S
 syscall_src += arch-arm/syscalls/__syslog.S
--- a/libc/arch-arm/syscalls/__brk.S
+++ b/libc/arch-arm/syscalls/__brk.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__brk)
    mov     ip, r7
    ldr     r7, =__NR_brk
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__brk)
--- a/libc/arch-arm/syscalls/__fcntl.S
+++ b/libc/arch-arm/syscalls/__fcntl.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__fcntl)
    mov     ip, r7
    ldr     r7, =__NR_fcntl
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__fcntl)
--- a/libc/arch-arm/syscalls/__fcntl64.S
+++ b/libc/arch-arm/syscalls/__fcntl64.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__fcntl64)
    mov     ip, r7
    ldr     r7, =__NR_fcntl64
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__fcntl64)
--- a/libc/arch-arm/syscalls/__fork.S
+++ b/libc/arch-arm/syscalls/__fork.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__fork)
    mov     ip, r7
    ldr     r7, =__NR_fork
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__fork)
--- a/libc/arch-arm/syscalls/__fstatfs64.S
+++ b/libc/arch-arm/syscalls/__fstatfs64.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__fstatfs64)
    mov     ip, r7
    ldr     r7, =__NR_fstatfs64
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__fstatfs64)
--- a/libc/arch-arm/syscalls/__getcpu.S
+++ b/libc/arch-arm/syscalls/__getcpu.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__getcpu)
    mov     ip, r7
    ldr     r7, =__NR_getcpu
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__getcpu)
--- a/libc/arch-arm/syscalls/__getcwd.S
+++ b/libc/arch-arm/syscalls/__getcwd.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__getcwd)
    mov     ip, r7
    ldr     r7, =__NR_getcwd
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__getcwd)
--- a/libc/arch-arm/syscalls/__getpriority.S
+++ b/libc/arch-arm/syscalls/__getpriority.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__getpriority)
    mov     ip, r7
    ldr     r7, =__NR_getpriority
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__getpriority)
--- a/libc/arch-arm/syscalls/__ioctl.S
+++ b/libc/arch-arm/syscalls/__ioctl.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__ioctl)
    mov     ip, r7
    ldr     r7, =__NR_ioctl
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__ioctl)
--- a/libc/arch-arm/syscalls/__llseek.S
+++ b/libc/arch-arm/syscalls/__llseek.S
@@ -1,6 +1,7 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__llseek)
    mov     ip, sp
@@ -10,7 +11,8 @@ ENTRY(__llseek)
    ldr     r7, =__NR__llseek
    swi     #0
    ldmfd   sp!, {r4, r5, r6, r7}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__llseek)
--- a/libc/arch-arm/syscalls/__mmap2.S
+++ b/libc/arch-arm/syscalls/__mmap2.S
@@ -1,6 +1,7 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__mmap2)
    mov     ip, sp
@@ -10,7 +11,8 @@ ENTRY(__mmap2)
    ldr     r7, =__NR_mmap2
    swi     #0
    ldmfd   sp!, {r4, r5, r6, r7}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__mmap2)
--- a/libc/arch-arm/syscalls/__open.S
+++ b/libc/arch-arm/syscalls/__open.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__open)
    mov     ip, r7
    ldr     r7, =__NR_open
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__open)
--- a/libc/arch-arm/syscalls/__openat.S
+++ b/libc/arch-arm/syscalls/__openat.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__openat)
    mov     ip, r7
    ldr     r7, =__NR_openat
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__openat)
--- a/libc/arch-arm/syscalls/__ptrace.S
+++ b/libc/arch-arm/syscalls/__ptrace.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__ptrace)
    mov     ip, r7
    ldr     r7, =__NR_ptrace
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__ptrace)
--- a/libc/arch-arm/syscalls/__reboot.S
+++ b/libc/arch-arm/syscalls/__reboot.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__reboot)
    mov     ip, r7
    ldr     r7, =__NR_reboot
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__reboot)
--- a/libc/arch-arm/syscalls/__rt_sigaction.S
+++ b/libc/arch-arm/syscalls/__rt_sigaction.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__rt_sigaction)
    mov     ip, r7
    ldr     r7, =__NR_rt_sigaction
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__rt_sigaction)
--- a/libc/arch-arm/syscalls/__rt_sigprocmask.S
+++ b/libc/arch-arm/syscalls/__rt_sigprocmask.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__rt_sigprocmask)
    mov     ip, r7
    ldr     r7, =__NR_rt_sigprocmask
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__rt_sigprocmask)
--- a/libc/arch-arm/syscalls/__rt_sigtimedwait.S
+++ b/libc/arch-arm/syscalls/__rt_sigtimedwait.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__rt_sigtimedwait)
    mov     ip, r7
    ldr     r7, =__NR_rt_sigtimedwait
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__rt_sigtimedwait)
--- a/libc/arch-arm/syscalls/__sched_getaffinity.S
+++ b/libc/arch-arm/syscalls/__sched_getaffinity.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__sched_getaffinity)
    mov     ip, r7
    ldr     r7, =__NR_sched_getaffinity
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__sched_getaffinity)
--- a/libc/arch-arm/syscalls/__set_tls.S
+++ b/libc/arch-arm/syscalls/__set_tls.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__set_tls)
    mov     ip, r7
-    ldr     r7, =__NR_ARM_set_tls
+    ldr     r7, =__ARM_NR_set_tls
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__set_tls)
--- a/libc/arch-arm/syscalls/__setresuid.S
+++ b/libc/arch-arm/syscalls/__setresuid.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__setresuid)
    mov     ip, r7
    ldr     r7, =__NR_setresuid32
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__setresuid)
--- a/libc/arch-arm/syscalls/__setreuid.S
+++ b/libc/arch-arm/syscalls/__setreuid.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__setreuid)
    mov     ip, r7
    ldr     r7, =__NR_setreuid32
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__setreuid)
--- a/libc/arch-arm/syscalls/__setuid.S
+++ b/libc/arch-arm/syscalls/__setuid.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__setuid)
    mov     ip, r7
    ldr     r7, =__NR_setuid32
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__setuid)
--- a/libc/arch-arm/syscalls/__sigsuspend.S
+++ b/libc/arch-arm/syscalls/__sigsuspend.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__sigsuspend)
    mov     ip, r7
    ldr     r7, =__NR_sigsuspend
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__sigsuspend)
--- a/libc/arch-arm/syscalls/__statfs64.S
+++ b/libc/arch-arm/syscalls/__statfs64.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__statfs64)
    mov     ip, r7
    ldr     r7, =__NR_statfs64
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__statfs64)
--- a/libc/arch-arm/syscalls/__sys_clone.S
+++ b/libc/arch-arm/syscalls/__sys_clone.S
@@ -1,6 +1,7 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__sys_clone)
    mov     ip, sp
@@ -10,7 +11,8 @@ ENTRY(__sys_clone)
    ldr     r7, =__NR_clone
    swi     #0
    ldmfd   sp!, {r4, r5, r6, r7}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__sys_clone)
--- a/libc/arch-arm/syscalls/__syslog.S
+++ b/libc/arch-arm/syscalls/__syslog.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__syslog)
    mov     ip, r7
    ldr     r7, =__NR_syslog
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__syslog)
--- a/libc/arch-arm/syscalls/__timer_create.S
+++ b/libc/arch-arm/syscalls/__timer_create.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__timer_create)
    mov     ip, r7
    ldr     r7, =__NR_timer_create
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__timer_create)
--- a/libc/arch-arm/syscalls/__timer_delete.S
+++ b/libc/arch-arm/syscalls/__timer_delete.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__timer_delete)
    mov     ip, r7
    ldr     r7, =__NR_timer_delete
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__timer_delete)
--- a/libc/arch-arm/syscalls/__timer_getoverrun.S
+++ b/libc/arch-arm/syscalls/__timer_getoverrun.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__timer_getoverrun)
    mov     ip, r7
    ldr     r7, =__NR_timer_getoverrun
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__timer_getoverrun)
--- a/libc/arch-arm/syscalls/__timer_gettime.S
+++ b/libc/arch-arm/syscalls/__timer_gettime.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__timer_gettime)
    mov     ip, r7
    ldr     r7, =__NR_timer_gettime
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__timer_gettime)
--- a/libc/arch-arm/syscalls/__timer_settime.S
+++ b/libc/arch-arm/syscalls/__timer_settime.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__timer_settime)
    mov     ip, r7
    ldr     r7, =__NR_timer_settime
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__timer_settime)
--- a/libc/arch-arm/syscalls/__wait4.S
+++ b/libc/arch-arm/syscalls/__wait4.S
@@ -1,13 +0,0 @@
-/* autogenerated by gensyscalls.py */
-#include <machine/asm.h>
-#include <sys/linux-syscalls.h>
-
-ENTRY(__wait4)
-    mov     ip, r7
-    ldr     r7, =__NR_wait4
-    swi     #0
-    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
-END(__wait4)
--- a/libc/arch-arm/syscalls/__waitid.S
+++ b/libc/arch-arm/syscalls/__waitid.S
@@ -1,6 +1,7 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(__waitid)
    mov     ip, sp
@@ -10,7 +11,8 @@ ENTRY(__waitid)
    ldr     r7, =__NR_waitid
    swi     #0
    ldmfd   sp!, {r4, r5, r6, r7}
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(__waitid)
--- a/libc/arch-arm/syscalls/_exit.S
+++ b/libc/arch-arm/syscalls/_exit.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(_exit)
    mov     ip, r7
    ldr     r7, =__NR_exit_group
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(_exit)
--- a/libc/arch-arm/syscalls/_exit_thread.S
+++ b/libc/arch-arm/syscalls/_exit_thread.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(_exit_thread)
    mov     ip, r7
    ldr     r7, =__NR_exit
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(_exit_thread)
--- a/libc/arch-arm/syscalls/accept.S
+++ b/libc/arch-arm/syscalls/accept.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(accept)
    mov     ip, r7
    ldr     r7, =__NR_accept
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(accept)
--- a/libc/arch-arm/syscalls/access.S
+++ b/libc/arch-arm/syscalls/access.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(access)
    mov     ip, r7
    ldr     r7, =__NR_access
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(access)
--- a/libc/arch-arm/syscalls/acct.S
+++ b/libc/arch-arm/syscalls/acct.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(acct)
    mov     ip, r7
    ldr     r7, =__NR_acct
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(acct)
--- a/libc/arch-arm/syscalls/bind.S
+++ b/libc/arch-arm/syscalls/bind.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(bind)
    mov     ip, r7
    ldr     r7, =__NR_bind
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(bind)
--- a/libc/arch-arm/syscalls/cacheflush.S
+++ b/libc/arch-arm/syscalls/cacheflush.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(cacheflush)
    mov     ip, r7
-    ldr     r7, =__NR_ARM_cacheflush
+    ldr     r7, =__ARM_NR_cacheflush
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(cacheflush)
--- a/libc/arch-arm/syscalls/capget.S
+++ b/libc/arch-arm/syscalls/capget.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(capget)
    mov     ip, r7
    ldr     r7, =__NR_capget
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(capget)
--- a/libc/arch-arm/syscalls/capset.S
+++ b/libc/arch-arm/syscalls/capset.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(capset)
    mov     ip, r7
    ldr     r7, =__NR_capset
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(capset)
--- a/libc/arch-arm/syscalls/chdir.S
+++ b/libc/arch-arm/syscalls/chdir.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(chdir)
    mov     ip, r7
    ldr     r7, =__NR_chdir
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(chdir)
--- a/libc/arch-arm/syscalls/chmod.S
+++ b/libc/arch-arm/syscalls/chmod.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(chmod)
    mov     ip, r7
    ldr     r7, =__NR_chmod
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(chmod)
--- a/libc/arch-arm/syscalls/chown.S
+++ b/libc/arch-arm/syscalls/chown.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(chown)
    mov     ip, r7
    ldr     r7, =__NR_chown32
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(chown)
--- a/libc/arch-arm/syscalls/chroot.S
+++ b/libc/arch-arm/syscalls/chroot.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(chroot)
    mov     ip, r7
    ldr     r7, =__NR_chroot
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(chroot)
--- a/libc/arch-arm/syscalls/clock_getres.S
+++ b/libc/arch-arm/syscalls/clock_getres.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(clock_getres)
    mov     ip, r7
    ldr     r7, =__NR_clock_getres
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(clock_getres)
--- a/libc/arch-arm/syscalls/clock_gettime.S
+++ b/libc/arch-arm/syscalls/clock_gettime.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(clock_gettime)
    mov     ip, r7
    ldr     r7, =__NR_clock_gettime
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(clock_gettime)
--- a/libc/arch-arm/syscalls/clock_nanosleep.S
+++ b/libc/arch-arm/syscalls/clock_nanosleep.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(clock_nanosleep)
    mov     ip, r7
    ldr     r7, =__NR_clock_nanosleep
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(clock_nanosleep)
--- a/libc/arch-arm/syscalls/clock_settime.S
+++ b/libc/arch-arm/syscalls/clock_settime.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(clock_settime)
    mov     ip, r7
    ldr     r7, =__NR_clock_settime
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(clock_settime)
--- a/libc/arch-arm/syscalls/close.S
+++ b/libc/arch-arm/syscalls/close.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(close)
    mov     ip, r7
    ldr     r7, =__NR_close
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(close)
--- a/libc/arch-arm/syscalls/connect.S
+++ b/libc/arch-arm/syscalls/connect.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(connect)
    mov     ip, r7
    ldr     r7, =__NR_connect
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(connect)
--- a/libc/arch-arm/syscalls/delete_module.S
+++ b/libc/arch-arm/syscalls/delete_module.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(delete_module)
    mov     ip, r7
    ldr     r7, =__NR_delete_module
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(delete_module)
--- a/libc/arch-arm/syscalls/dup.S
+++ b/libc/arch-arm/syscalls/dup.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(dup)
    mov     ip, r7
    ldr     r7, =__NR_dup
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(dup)
--- a/libc/arch-arm/syscalls/dup2.S
+++ b/libc/arch-arm/syscalls/dup2.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(dup2)
    mov     ip, r7
    ldr     r7, =__NR_dup2
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(dup2)
--- a/libc/arch-arm/syscalls/epoll_create.S
+++ b/libc/arch-arm/syscalls/epoll_create.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(epoll_create)
    mov     ip, r7
    ldr     r7, =__NR_epoll_create
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(epoll_create)
--- a/libc/arch-arm/syscalls/epoll_ctl.S
+++ b/libc/arch-arm/syscalls/epoll_ctl.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(epoll_ctl)
    mov     ip, r7
    ldr     r7, =__NR_epoll_ctl
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(epoll_ctl)
--- a/libc/arch-arm/syscalls/epoll_wait.S
+++ b/libc/arch-arm/syscalls/epoll_wait.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(epoll_wait)
    mov     ip, r7
    ldr     r7, =__NR_epoll_wait
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(epoll_wait)
--- a/libc/arch-arm/syscalls/eventfd.S
+++ b/libc/arch-arm/syscalls/eventfd.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(eventfd)
    mov     ip, r7
    ldr     r7, =__NR_eventfd2
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(eventfd)
--- a/libc/arch-arm/syscalls/execve.S
+++ b/libc/arch-arm/syscalls/execve.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(execve)
    mov     ip, r7
    ldr     r7, =__NR_execve
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(execve)
--- a/libc/arch-arm/syscalls/faccessat.S
+++ b/libc/arch-arm/syscalls/faccessat.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(faccessat)
    mov     ip, r7
    ldr     r7, =__NR_faccessat
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(faccessat)
--- a/libc/arch-arm/syscalls/fchdir.S
+++ b/libc/arch-arm/syscalls/fchdir.S
@@ -1,13 +1,15 @@
 /* autogenerated by gensyscalls.py */
+#include <asm/unistd.h>
+#include <linux/err.h>
 #include <machine/asm.h>
-#include <sys/linux-syscalls.h>

 ENTRY(fchdir)
    mov     ip, r7
    ldr     r7, =__NR_fchdir
    swi     #0
    mov     r7, ip
-    movs    r0, r0
-    bxpl    lr
-    b       __set_syscall_errno
+    cmn     r0, #(MAX_ERRNO + 1)
+    bxls    lr
+    neg     r0, r0
+    b       __set_errno
 END(fchdir)
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`include bionic/libc/arch-arm/cortex-a15/cortex-a15.mk`