diff --git a/FAQ b/FAQ
index 4712a232..5627a670 100644
--- a/FAQ
+++ b/FAQ
@@ -246,7 +246,7 @@ A: UPDATE: the defaults have been changed as of 2005-04-23.
    code.  Unfortunately, these options have been renamed on gcc 3.  You
    can fix this by creating a file ~/.diet/gcc containing this line:
 
-     -Os -fomit-frame-pointer -falign-jumps=1 -falign-loops=1 -mpreferred-stack-boundary=2
+     -Os -fomit-frame-pointer -falign-jumps=1 -falign-loops=1 -mpreferred-stack-boundary=4
 
    If you get this options not for diet -Os gcc but for diet -Os
    i386-linux-gcc, put this in ~/.diet/i386-linux-gcc instead.
diff --git a/Makefile b/Makefile
index 4f93cd30..167a304f 100644
--- a/Makefile
+++ b/Makefile
@@ -10,7 +10,7 @@ MAN1DIR=${prefix}/man/man1
 
 EXTRACFLAGS=
 
-MYARCH:=$(shell uname -m | sed -e 's/i[4-9]86/i386/' -e 's/armv[3-6]t\?e\?[lb]/arm/')
+MYARCH:=$(shell uname -m | sed -e 's/i[4-9]86/i386/' -e 's/armv[3-7]t\?e\?[lb]/arm/')
 
 # This extra-ugly cruft is here so make will not run uname and sed each
 # time it looks at $(OBJDIR).  This alone sped up running make when
@@ -107,6 +107,8 @@ CFLAGS=$(DEFAULTCFLAGS)
 CROSS=
 
 CC=gcc
+CCC=$(CROSS)$(CC)
+STRIP=$(COMMENT) $(CROSS)strip
 INC=-I. -isystem include
 
 VPATH=lib:libstdio:libugly:libcruft:libcrypt:libshell:liblatin1:libcompat:libdl:librpc:libregex:libm:profiling
@@ -141,7 +143,7 @@ endif
 
 ifneq ($(DEBUG),)
 CFLAGS = -g
-COMMENT = :
+STRIP = :
 endif
 CFLAGS += -W -Wall -Wextra -Wchar-subscripts -Wmissing-prototypes -Wmissing-declarations -Wno-switch -Wno-unused -Wredundant-decls -Wshadow
 
@@ -163,31 +165,31 @@ $(OBJDIR) $(PICODIR):
 
 % :: %,v
 
-$(OBJDIR)/%: $(OBJDIR)
-
 ifeq ($(CC),tcc)
-$(OBJDIR)/%.o: %.S $(ARCH)/syscalls.h
+$(OBJDIR)/%.o: %.S $(ARCH)/syscalls.h | $(OBJDIR)
 	$(CROSS)cpp $(INC) $< | $(CROSS)as --noexecstack -o $@
 
-$(OBJDIR)/%.o: %.c
+$(OBJDIR)/%.o: %.c | $(OBJDIR)
 	tcc -I. -Iinclude -c $< -o $@
-	$(COMMENT) -$(CROSS)strip -x -R .comment -R .note $@
+	-$(STRIP) -x -R .comment -R .note $@
 else
-$(OBJDIR)/pstart.o: start.S
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -DPROFILING -c $< $(ASM_CFLAGS) -o $@
+$(OBJDIR)/pstart.o: start.S | $(OBJDIR)
+	$(CCC) $(INC) $(CFLAGS) -DPROFILING -c $< $(ASM_CFLAGS) -o $@
 
-$(OBJDIR)/%.o: %.S $(ARCH)/syscalls.h
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -c $< $(ASM_CFLAGS) -o $@
+$(OBJDIR)/%.o: %.S $(ARCH)/syscalls.h | $(OBJDIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRAFLAGS) -c $< $(ASM_CFLAGS) -o $@
 
-$(OBJDIR)/pthread_%.o: libpthread/pthread_%.c
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -c $< -o $@
-	$(COMMENT) -$(CROSS)strip -x -R .comment -R .note $@
+$(OBJDIR)/pthread_%.o: libpthread/pthread_%.c | $(OBJDIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRAFLAGS) -c $< -o $@
+	-$(STRIP) -x -R .comment -R .note $@
 
-$(OBJDIR)/%.o: %.c
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -c $< -o $@ -D__dietlibc__
-	$(COMMENT) -$(CROSS)strip -x -R .comment -R .note $@
+$(OBJDIR)/%.o: %.c | $(OBJDIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRAFLAGS) -c $< -o $@ -D__dietlibc__
+	-$(STRIP) -x -R .comment -R .note $@
 endif
 
+
+
 ifeq ($(shell $(CC) -v 2>&1 | grep "gcc version"),gcc version 4.0.0)
 SAFE_CFLAGS=$(shell echo $(CFLAGS)|sed 's/-Os/-O2/')
 SAFER_CFLAGS=$(shell echo $(CFLAGS)|sed 's/-Os/-O/')
@@ -198,8 +200,8 @@ endif
 
 CC+=-D__dietlibc__
 
-$(OBJDIR)/crypt.o: libcrypt/crypt.c
-	$(CROSS)$(CC) $(INC) $(SAFER_CFLAGS) -c $< -o $@
+$(OBJDIR)/crypt.o: libcrypt/crypt.c | $(OBJDIR)
+	$(CCC) $(INC) $(SAFER_CFLAGS) -c $< -o $@
 
 DIETLIBC_OBJ = $(OBJDIR)/unified.o \
 $(SYSCALLOBJ) $(LIBOBJ) $(LIBSTDIOOBJ) $(LIBUGLYOBJ) \
@@ -213,9 +215,9 @@ $(OBJDIR)/dietlibc.a: $(DIETLIBC_OBJ) $(OBJDIR)/start.o
 $(OBJDIR)/librpc.a: $(LIBRPCOBJ)
 	$(CROSS)ar cru $@ $(LIBRPCOBJ)
 
-$(OBJDIR)/libcrypt.a:
+$(OBJDIR)/libcrypt.a: | $(OBJDIR)
 	touch dummy.c
-	$(CROSS)$(CC) -c dummy.c
+	$(CCC) -c dummy.c
 	$(CROSS)ar cru $@ dummy.o
 	rm -f dummy.c dummy.o
 
@@ -250,23 +252,23 @@ dyn_lib: $(PICODIR) $(PICODIR)/libc.so $(PICODIR)/dstart.o \
 	$(PICODIR)/libpthread.so $(PICODIR)/libdl.so $(PICODIR)/libcompat.so \
 	$(PICODIR)/libm.so $(PICODIR)/diet-dyn $(PICODIR)/diet-dyn-i
 
-$(PICODIR)/%.o: %.S $(ARCH)/syscalls.h
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -fPIC -D__DYN_LIB $(ASM_CFLAGS) -c $< -o $@
+$(PICODIR)/%.o: %.S $(ARCH)/syscalls.h | $(PICODIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRACFLAGS) -fPIC -D__DYN_LIB $(ASM_CFLAGS) -c $< -o $@
 
-$(PICODIR)/pthread_%.o: libpthread/pthread_%.c
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -fPIC -D__DYN_LIB -c $< -o $@
-	$(COMMENT) $(CROSS)strip -x -R .comment -R .note $@
+$(PICODIR)/pthread_%.o: libpthread/pthread_%.c | $(PICODIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRACFLAGS) -fPIC -D__DYN_LIB -c $< -o $@
+	$(STRIP) -x -R .comment -R .note $@
 
-$(PICODIR)/%.o: %.c
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -fPIC -D__DYN_LIB -c $< -o $@
-	$(COMMENT) $(CROSS)strip -x -R .comment -R .note $@
+$(PICODIR)/%.o: %.c | $(PICODIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRACFLAGS) -fPIC -D__DYN_LIB -c $< -o $@
+	$(STRIP) -x -R .comment -R .note $@
 
-$(PICODIR)/dstart.o: start.S
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -fPIC -D__DYN_LIB $(ASM_CFLAGS) -c $< -o $@
+$(PICODIR)/dstart.o: start.S | $(PICODIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRACFLAGS) -fPIC -D__DYN_LIB $(ASM_CFLAGS) -c $< -o $@
 
-$(PICODIR)/dyn_so_start.o: dyn_start.c
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -fPIC -D__DYN_LIB -D__DYN_LIB_SHARED -c $< -o $@
-	$(COMMENT) $(CROSS)strip -x -R .comment -R .note $@
+$(PICODIR)/dyn_so_start.o: dyn_start.c | $(PICODIR)
+	$(CCC) $(INC) $(CFLAGS) $(EXTRACFLAGS) -fPIC -D__DYN_LIB -D__DYN_LIB_SHARED -c $< -o $@
+	$(STRIP) -x -R .comment -R .note $@
 
 DYN_LIBC_PIC = $(LIBOBJ) $(LIBSTDIOOBJ) $(LIBUGLYOBJ) \
 $(LIBCRUFTOBJ) $(LIBCRYPTOBJ) $(LIBSHELLOBJ) $(LIBREGEXOBJ)
@@ -283,54 +285,54 @@ DYN_LIBCOMPAT_OBJS = $(patsubst $(OBJDIR)/%.o,$(PICODIR)/%.o,$(LIBCOMPATOBJ))
 DYN_LIBMATH_OBJS = $(patsubst $(OBJDIR)/%.o,$(PICODIR)/%.o,$(LIBMATHOBJ))
 
 $(PICODIR)/libc.so: $(PICODIR) $(DYN_LIBC_OBJ)
-	$(LD_UNSET) $(CROSS)$(CC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBC_OBJ) -lgcc -Wl,-soname=libc.so
+	$(LD_UNSET) $(CCC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBC_OBJ) -lgcc -Wl,-soname=libc.so
 
 $(PICODIR)/libpthread.so: $(DYN_PTHREAD_OBJS) dietfeatures.h $(PICODIR)/libc.so
-	$(LD_UNSET) $(CROSS)$(CC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_PTHREAD_OBJS) -L$(PICODIR) -lc -Wl,-soname=libpthread.so
+	$(LD_UNSET) $(CCC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_PTHREAD_OBJS) -L$(PICODIR) -lc -Wl,-soname=libpthread.so
 
 $(PICODIR)/libdl.so: libdl/_dl_main.c dietfeatures.h $(PICODIR)/libc.so
-	$(LD_UNSET) $(CROSS)$(CC) -D__OD_CLEAN_ROOM -DNODIETREF -fPIC -nostdlib -shared -Bsymbolic -Wl,-Bsymbolic \
+	$(LD_UNSET) $(CCC) -D__OD_CLEAN_ROOM -DNODIETREF -fPIC -nostdlib -shared -Bsymbolic -Wl,-Bsymbolic \
 		-o $@ $(SAFE_CFLAGS) $(INC) libdl/_dl_main.c -Wl,-soname=libdl.so
 
 $(OBJDIR)/pthread_create.o $(PICODIR)/pthread_create.o: dietfeatures.h
 $(OBJDIR)/pthread_internal.o $(PICODIR)/pthread_internal.o: dietfeatures.h
 
 #$(PICODIR)/libdl.so: $(DYN_LIBDL_OBJS) dietfeatures.h
-#	$(CROSS)$(CC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBDL_OBJS) -L$(PICODIR) -ldietc -Wl,-soname=libdl.so
+#	$(CCC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBDL_OBJS) -L$(PICODIR) -ldietc -Wl,-soname=libdl.so
 
 $(PICODIR)/libcompat.so: $(DYN_LIBCOMPAT_OBJS) dietfeatures.h $(PICODIR)/libc.so
-	$(LD_UNSET) $(CROSS)$(CC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBCOMPAT_OBJS) -L$(PICODIR) -lc -Wl,-soname=libcompat.so
+	$(LD_UNSET) $(CCC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBCOMPAT_OBJS) -L$(PICODIR) -lc -Wl,-soname=libcompat.so
 
 $(PICODIR)/libm.so: $(DYN_LIBMATH_OBJS) dietfeatures.h $(PICODIR)/libc.so
-	$(LD_UNSET) $(CROSS)$(CC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBMATH_OBJS) -L$(PICODIR) -lc -Wl,-soname=libm.so
+	$(LD_UNSET) $(CCC) -nostdlib -shared -o $@ $(CFLAGS) -fPIC $(DYN_LIBMATH_OBJS) -L$(PICODIR) -lc -Wl,-soname=libm.so
 
 
 $(SYSCALLOBJ): syscalls.h
 
 $(OBJDIR)/elftrunc: $(OBJDIR)/diet contrib/elftrunc.c
-	bin-$(MYARCH)/diet $(CROSS)$(CC) $(CFLAGS) -o $@ contrib/elftrunc.c
+	bin-$(MYARCH)/diet $(CCC) $(CFLAGS) -o $@ contrib/elftrunc.c
 
 $(OBJDIR)/dnsd: $(OBJDIR)/diet contrib/dnsd.c
-	bin-$(MYARCH)/diet $(CROSS)$(CC) $(CFLAGS) -o $@ contrib/dnsd.c
+	bin-$(MYARCH)/diet $(CCC) $(CFLAGS) -o $@ contrib/dnsd.c
 
 VERSION=dietlibc-$(shell head -n 1 CHANGES|sed 's/://')
 CURNAME=$(notdir $(shell pwd))
 
 $(OBJDIR)/diet: $(OBJDIR)/start.o $(OBJDIR)/dyn_start.o diet.c $(OBJDIR)/dietlibc.a $(OBJDIR)/dyn_stop.o
-	$(CROSS)$(CC) -isystem include $(CFLAGS) -nostdlib -o $@ $^ -DDIETHOME=\"$(HOME)\" -DVERSION=\"$(VERSION)\" -lgcc
-	$(CROSS)strip -R .comment -R .note $@
+	$(CCC) -isystem include $(CFLAGS) -nostdlib -o $@ $^ -DDIETHOME=\"$(HOME)\" -DVERSION=\"$(VERSION)\" -lgcc
+	$(STRIP) -R .comment -R .note $@
 
 $(OBJDIR)/diet-i: $(OBJDIR)/start.o $(OBJDIR)/dyn_start.o diet.c $(OBJDIR)/dietlibc.a $(OBJDIR)/dyn_stop.o
-	$(CROSS)$(CC) -isystem include $(CFLAGS) -nostdlib -o $@ $^ -DDIETHOME=\"$(prefix)\" -DVERSION=\"$(VERSION)\" -DINSTALLVERSION -lgcc
-	$(CROSS)strip -R .comment -R .note $@
+	$(CCC) -isystem include $(CFLAGS) -nostdlib -o $@ $^ -DDIETHOME=\"$(prefix)\" -DVERSION=\"$(VERSION)\" -DINSTALLVERSION -lgcc
+	$(STRIP) -R .comment -R .note $@
 
 $(PICODIR)/diet-dyn: $(PICODIR)/start.o $(PICODIR)/dyn_start.o diet.c
-	$(LD_UNSET) $(CROSS)$(CC) -isystem include $(CFLAGS) -fPIC -nostdlib -o $@ $^ -DDIETHOME=\"$(HOME)\" -D__DYN_LIB -DVERSION=\"$(VERSION)\" -L$(PICODIR) -lc -lgcc $(PICODIR)/dyn_stop.o -Wl,-dynamic-linker=$(HOME)/$(PICODIR)/libdl.so
-	$(CROSS)strip -R .command -R .note $@
+	$(LD_UNSET) $(CCC) -isystem include $(CFLAGS) -fPIC -nostdlib -o $@ $^ -DDIETHOME=\"$(HOME)\" -D__DYN_LIB -DVERSION=\"$(VERSION)\" -L$(PICODIR) -lc -lgcc $(PICODIR)/dyn_stop.o -Wl,-dynamic-linker=$(HOME)/$(PICODIR)/libdl.so
+	$(STRIP) -R .command -R .note $@
 
 $(PICODIR)/diet-dyn-i: $(PICODIR)/start.o $(PICODIR)/dyn_start.o diet.c
-	$(LD_UNSET) $(CROSS)$(CC) -isystem include $(CFLAGS) -fPIC -nostdlib -o $@ $^ -DDIETHOME=\"$(prefix)\" -D__DYN_LIB -DVERSION=\"$(VERSION)\" -L$(PICODIR) -lc -lgcc $(PICODIR)/dyn_stop.o -Wl,-dynamic-linker=$(ILIBDIR)/libdl.so -DINSTALLVERSION
-	$(CROSS)strip -R .command -R .note $@
+	$(LD_UNSET) $(CCC) -isystem include $(CFLAGS) -fPIC -nostdlib -o $@ $^ -DDIETHOME=\"$(prefix)\" -D__DYN_LIB -DVERSION=\"$(VERSION)\" -L$(PICODIR) -lc -lgcc $(PICODIR)/dyn_stop.o -Wl,-dynamic-linker=$(ILIBDIR)/libdl.so -DINSTALLVERSION
+	$(STRIP) -R .command -R .note $@
 
 $(OBJDIR)/djb: $(OBJDIR)/compile $(OBJDIR)/load
 
@@ -360,14 +362,14 @@ $(OBJDIR)/exports: $(OBJDIR)/dietlibc.a
 
 .PHONY: t t1
 t:
-	$(CROSS)$(CC) -g $(CFLAGS) -fno-builtin -nostdlib -isystem include -o t t.c $(OBJDIR)/start.o $(OBJDIR)/dyn_start.o $(OBJDIR)/dietlibc.a -lgcc $(OBJDIR)/dyn_stop.o -Wl,-Map,mapfile
+	$(CCC) -g $(CFLAGS) -fno-builtin -nostdlib -isystem include -o t t.c $(OBJDIR)/start.o $(OBJDIR)/dyn_start.o $(OBJDIR)/dietlibc.a -lgcc $(OBJDIR)/dyn_stop.o -Wl,-Map,mapfile
 
 t1:
-	$(CROSS)$(CC) -g -o t1 t.c
+	$(CCC) -g -o t1 t.c
 
 install-bin: $(OBJDIR)/start.o $(OBJDIR)/dietlibc.a $(OBJDIR)/librpc.a $(OBJDIR)/liblatin1.a $(OBJDIR)/libcompat.a $(OBJDIR)/elftrunc $(OBJDIR)/diet-i
 	$(INSTALL) -d $(DESTDIR)$(ILIBDIR) $(DESTDIR)$(MAN1DIR) $(DESTDIR)$(BINDIR)
-	$(INSTALL) $(OBJDIR)/start.o $(DESTDIR)$(ILIBDIR)/start.o
+	$(INSTALL) -m 644 $(OBJDIR)/start.o $(DESTDIR)$(ILIBDIR)/start.o
 	$(INSTALL) -m 644 $(OBJDIR)/libm.a $(OBJDIR)/libpthread.a $(OBJDIR)/librpc.a \
 $(OBJDIR)/liblatin1.a $(OBJDIR)/libcompat.a $(OBJDIR)/libcrypt.a $(DESTDIR)$(ILIBDIR)
 	$(INSTALL) -m 644 $(OBJDIR)/dietlibc.a $(DESTDIR)$(ILIBDIR)/libc.a
@@ -549,9 +551,7 @@ $(OBJDIR)/fcntl64.o: dietfeatures.h
 
 # WANT_SSP
 # This facepalm brought to you by: Ubuntu!
-$(OBJDIR)/stackgap.o: lib/stackgap.c dietfeatures.h
-	$(CROSS)$(CC) $(INC) $(CFLAGS) -c lib/stackgap.c -o $@ -D__dietlibc__ -fno-stack-protector
-	$(COMMENT) -$(CROSS)strip -x -R .comment -R .note $@
+$(OBJDIR)/stackgap.o $(PICODIR)/stackgap.o: EXTRACFLAGS:=-fno-stack-protector
 
 # WANT_MALLOC_ZERO
 $(OBJDIR)/strndup.o: dietfeatures.h
diff --git a/arm/Makefile.add b/arm/Makefile.add
index 361e98b3..9c9a67c3 100644
--- a/arm/Makefile.add
+++ b/arm/Makefile.add
@@ -1,6 +1,7 @@
 
 LIBOBJ+=$(OBJDIR)/md5asm.o $(OBJDIR)/__aeabi_unwind_cpp.o
 LIBOBJ+=$(OBJDIR)/__fadvise.o $(OBJDIR)/arm_fadvise.o
+LIBOBJ+=$(OBJDIR)/__aeabi_read_tp.o $(OBJDIR)/arm_set_tls.o
 CFLAGS+=-Os -fomit-frame-pointer -fstrict-aliasing
 #ifdef __ARM_EABI__
 #CFLAGS+=-mabi=aapcs-linux -mno-thumb-interwork
diff --git a/arm/__aeabi_read_tp.S b/arm/__aeabi_read_tp.S
new file mode 100644
index 00000000..65fcf5f6
--- /dev/null
+++ b/arm/__aeabi_read_tp.S
@@ -0,0 +1,13 @@
+#include "arm-features.h"
+
+FUNC_START __aeabi_read_tp
+
+#if __ARM_ARCH__ < 6
+	mvn     r0, #0xf000
+	sub     pc, r0, #31
+#else
+	mrc     15, 0, r0, cr13, cr0, 3
+	bx      lr
+#endif
+
+FUNC_END __aeabi_read_tp
diff --git a/arm/__aeabi_unwind_cpp.S b/arm/__aeabi_unwind_cpp.S
index ca631bcb..9334eee8 100644
--- a/arm/__aeabi_unwind_cpp.S
+++ b/arm/__aeabi_unwind_cpp.S
@@ -1,21 +1,14 @@
-.text
-.global __aeabi_unwind_cpp_pr0
-.hidden __aeabi_unwind_cpp_pr0
-.type __aeabi_unwind_cpp_pr0, %function
+#include "arm-features.h"
 
-.global __aeabi_unwind_cpp_pr1
+FUNC_START __aeabi_unwind_cpp_pr0
+FUNC_START __aeabi_unwind_cpp_pr1
+FUNC_START __aeabi_unwind_cpp_pr2
+.hidden __aeabi_unwind_cpp_pr0
 .hidden __aeabi_unwind_cpp_pr1
-.type __aeabi_unwind_cpp_pr1, %function
-
-.global __aeabi_unwind_cpp_pr2
 .hidden __aeabi_unwind_cpp_pr2
-.type __aeabi_unwind_cpp_pr2, %function
 
-__aeabi_unwind_cpp_pr0:
-__aeabi_unwind_cpp_pr1:
-__aeabi_unwind_cpp_pr2:
-	mov	pc, lr	@ return from subroutine
+	RET
 
-.size __aeabi_unwind_cpp_pr0,.-__aeabi_unwind_cpp_pr0
-.size __aeabi_unwind_cpp_pr1,.-__aeabi_unwind_cpp_pr1
-.size __aeabi_unwind_cpp_pr2,.-__aeabi_unwind_cpp_pr2
+FUNC_END  __aeabi_unwind_cpp_pr2
+FUNC_END  __aeabi_unwind_cpp_pr1
+FUNC_END  __aeabi_unwind_cpp_pr0
diff --git a/arm/__fadvise.c b/arm/__fadvise.c
index 0aa1246d..aee78fc0 100644
--- a/arm/__fadvise.c
+++ b/arm/__fadvise.c
@@ -1,5 +1,7 @@
-#include <fcntl.h>
 #include "syscalls.h"
+#define _LINUX_SOURCE
+#include <fcntl.h>
+#include <unistd.h>
 
 #ifndef __NR_fadvise64
 long fadvise64_64(int fd, off64_t offset, off64_t len, int advice)
diff --git a/arm/__guard.S b/arm/__guard.S
index 7218d131..a4e53aa6 100644
--- a/arm/__guard.S
+++ b/arm/__guard.S
@@ -1,4 +1,5 @@
 .data
+.align 2
 .type __guard,#object
 .global __guard
 .type __stack_chk_guard,#object
@@ -7,3 +8,5 @@ __guard:
 __stack_chk_guard:
 .long 0xaff00
 
+.size __guard, . - __guard
+.size __stack_chk_guard, . - __stack_chk_guard
diff --git a/arm/__longjmp.S b/arm/__longjmp.S
index 31307bd2..3a8203fe 100644
--- a/arm/__longjmp.S
+++ b/arm/__longjmp.S
@@ -1,11 +1,23 @@
-.text
-.global __longjmp
-.type __longjmp,function
-__longjmp:
+#include "arm-features.h"
+
+FUNC_START	__longjmp
 	mov	ip, r0
 	movs	r0, r1
 	moveq	r0, #1
-#ifndef	__SOFTFP__
-	lfm	f4, 4, [ip], #48
+#ifndef __SOFTFP__
+# if __ARM_ARCH__ == 7
+	vldm	ip!, {d0-d15}
+#   ifdef __ARM_NEON__
+	vldm	ip!, {d16-d31}
+#   endif
+# else
+	lfm	f4, 4, [ip]!
+# endif
 #endif
-	ldmia   ip, {r4-r11, sp, pc}
+
+#ifdef __IWMMXT__
+#  warning "sigjmp will not restore iwmmxt coprocessor registers"
+#endif
+
+	ldmia   ip!, {r4-r11, sp, pc}
+FUNC_END	__longjmp
diff --git a/arm/__testandset.S b/arm/__testandset.S
index d9c57648..3b62c516 100644
--- a/arm/__testandset.S
+++ b/arm/__testandset.S
@@ -1,7 +1,15 @@
-.text
-.global __testandset
-__testandset:
+#include "arm-features.h"
+
+FUNC_START	__testandset
 	mov	r2, r0
 	mov	r1, #1
+# if __ARM_ARCH__ < 6
 	swp	r0, r1, [r2]
-	mov	pc, lr
+# else
+1:	ldrex	r0, [r2]
+	strex	r3, r1, [r2]
+	cmp	r3, #0
+	bne	1b
+# endif
+	RET
+FUNC_END	__testandset
diff --git a/arm/arm-features.h b/arm/arm-features.h
new file mode 100644
index 00000000..fa5d0451
--- /dev/null
+++ b/arm/arm-features.h
@@ -0,0 +1,110 @@
+/* --*- asm -*-- */
+
+#ifndef H_DIETLIBC_ARM_FEATURES_H
+#define H_DIETLIBC_ARM_FEATURES_H
+
+/* Stolen from gcc (gcc/config/arm/lib1funcs.asm) */
+#if defined(__ARM_ARCH_2__)
+# define __ARM_ARCH__ 2
+#endif
+
+#if defined(__ARM_ARCH_3__)
+# define __ARM_ARCH__ 3
+#endif
+
+#if defined(__ARM_ARCH_3M__) || defined(__ARM_ARCH_4__) \
+        || defined(__ARM_ARCH_4T__)
+# define __ARM_ARCH__ 4
+#endif
+
+#if defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_5T__) \
+        || defined(__ARM_ARCH_5E__) || defined(__ARM_ARCH_5TE__) \
+        || defined(__ARM_ARCH_5TEJ__)
+# define __ARM_ARCH__ 5
+#endif
+
+#if defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \
+        || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \
+        || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) \
+        || defined(__ARM_ARCH_6M__)
+# define __ARM_ARCH__ 6
+#endif
+
+#if defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \
+        || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__)
+# define __ARM_ARCH__ 7
+#endif
+
+#ifndef __ARM_ARCH__
+#error Unable to determine architecture.
+#endif
+
+#define DIET_JMPBUFSZ_REGS_REGULAR	10*32/8	/* r4-r11, sp, pc */
+#define DIET_JMPBUFSZ_REGS_FPv4		16*64/8 /* d0-d15 */
+
+#if !defined(__SOFTFP__) || defined(__IWMMXT__)
+#  define DIET_HAVE_COPROC_REGS	1
+#else
+#  undef  DIET_HAVE_COPROC_REGS
+#endif
+
+#ifdef __ASSEMBLER__
+
+.macro	FUNC_START name
+	.text
+	.align	0
+	.global	\name
+	.type	\name, %function
+\name:
+.endm
+
+.macro	FUNC_START_WEAK name
+	.text
+	.align	0
+	.weak	\name
+	.type	\name, %function
+\name:
+.endm
+
+.macro	FUNC_END name
+	.size	\name, . - \name
+.endm
+
+.macro	RET
+#if (__ARM_ARCH__ > 4) || defined(__ARM_ARCH_4T__)
+	bx	lr
+#else
+	mov	pc, lr
+#endif
+.endm
+
+.macro	SWI_UNIFIED	name
+#ifdef __ARM_EABI__
+	b	__unified_syscall_swi
+#else
+	swi	\name
+	b	__unified_syscall
+#endif
+.endm
+
+.macro	SWI_UNIFIED4
+#ifdef __ARM_EABI__
+	b	__unified_syscall_swi
+#else
+	swi	\name
+	b	__unified_syscall4
+#endif
+.endm
+
+.macro	LOAD_ARG4_5
+#if defined(__ARM_ARCH_4__) || defined(__ARM_ARCH_4T__) || defined(__ARM_ARCH_5T__)
+	ldr	r4, [sp,#16]
+	ldr	r5, [sp,#20]
+#else
+	ldrd	r4, [sp,#16]
+#endif
+.endm
+
+#endif	/* __ASSEMBLER__ */
+
+#endif	/* H_DIETLIBC_ARM_FEATURES_H */
diff --git a/arm/arm_set_tls.S b/arm/arm_set_tls.S
new file mode 100644
index 00000000..94c4ba45
--- /dev/null
+++ b/arm/arm_set_tls.S
@@ -0,0 +1,27 @@
+#include "syscalls.h"
+#include "arm-features.h"
+
+#ifdef __ARM_EABI__
+
+@ Calling convention:
+@   r0 ... thread data
+@ Returns:
+@   r0 ... low-level error code
+FUNC_START	__arm_set_tls
+	mov	ip, r7
+	ldr	r7, =__ARM_NR_set_tls
+	swi	0
+	mov	r7, ip
+
+	@ we do not have much options here to handle errors because
+	@ errno might not be available yet. Return error code as-is
+	@ and let it handle by the caller
+
+	RET
+FUNC_END	__arm_set_tls
+
+#else
+
+#warning "setup-tls only implemented for ARM_EABI!"
+
+#endif
diff --git a/arm/clone.S b/arm/clone.S
index 4a4b2f4a..7b43d66a 100644
--- a/arm/clone.S
+++ b/arm/clone.S
@@ -1,11 +1,7 @@
-
 #include <errno.h>
 #include "syscalls.h"
+#include "arm-features.h"
 
-	.text
-	.weak 	clone
-	.global __clone
-	
 @
 @ Some slightly tricky stuff here... edit with care :-)
 @
@@ -15,13 +11,12 @@
 
 #define CLONE_VM      0x00000100
 #define CLONE_THREAD  0x00010000
-	
+
 @ ;  don't do this yet
 @#define RESET_PID
 
-		
-clone:
-__clone:
+FUNC_START_WEAK	clone
+FUNC_START	__clone
 	@ ; start with a sanity check
 	cmp	r0, #0
 	cmpne	r1, #0
@@ -31,13 +26,13 @@ __clone:
 	@ ; insert args into stack
 	str	r3, [r1, #-4]!
 	str	r0, [r1, #-4]!
-	
+
 	@ ; do the system call
 	@ ; get the flags
 	mov	r0, r2
-#ifdef RESET_PID	
+#ifdef RESET_PID
 	mov	ip, r2
-#endif	
+#endif
 	@ ; child sp is already in r1
 	stmfd	sp!, {r4, r7}
 	@stmdb	sp!, {r4, r7}
@@ -46,14 +41,14 @@ __clone:
 	ldr	r4, [sp, #16]
         ldr     r7, =__NR_clone
 	swi	0
-	
+
 	cmp	r0, #0
 	beq	1f
-	ldmfd	sp!, {r4, r7}	
+	ldmfd	sp!, {r4, r7}
 	blt	__unified_syscall	@ (return code < 0): handle as an error
-	bx	lr
-1:	
-#ifdef RESET_PID		
+	RET
+1:
+#ifdef RESET_PID
 	tst	ip, #CLONE_THREAD
 	bne	3f
 
@@ -67,21 +62,22 @@ __clone:
 	swi	0
 	str	r0, [r1, #PID_OFFSET] @	; not defined yet ??
 	str	r0, [r1, #TID_OFFSET] @	; not defined yet ??
-3:	
+3:
 #endif
 	@ ; pick the function arg and call address off the stack and execute
 	ldr	r0, [sp, #4]
 	mov	lr, pc
 	ldr	pc, [sp], #8
-	
+
 	@ ; and we're done, passing return value through r0
 	b	_exit			@ branch to _exit (PIC safe)
+FUNC_END	__clone
+FUNC_END	clone
 
-	
 #else
-		
-clone:
-__clone:
+
+FUNC_START_WEAK	clone
+FUNC_START	__clone
 	movs	r12, r0			@ check function pointer
 	cmpne	r1, #0			@ if function check for stack pointer
 	moveq	r0, #-EINVAL		@ if one is not available set errno value
@@ -101,5 +97,7 @@ __clone:
 	ldmia	sp!, { r0, pc } 	@ load function param and jump to thread function
 
 1:	b	_exit			@ branch to _exit (PIC safe)
+FUNC_END	__clone
+FUNC_END	clone
 
 #endif
diff --git a/arm/dyn_syscalls.S b/arm/dyn_syscalls.S
index a4baf28f..56ec38a5 100644
--- a/arm/dyn_syscalls.S
+++ b/arm/dyn_syscalls.S
@@ -3,6 +3,8 @@
  * Olaf Dreesen
  */
 
+#include "arm-features.h"
+
 #include "../syscalls.s/environ.S"
 #include "../syscalls.s/errno.S"
 
@@ -25,7 +27,7 @@ __unified_syscall:
 
 	mvn	r0, #0
 #include "dietuglyweaks.h"
-	mov	pc, lr
+	RET
 
 /* ok now include all syscalls.s (*.S) and sysdep *.S */
 #include "mmap.S"
@@ -104,7 +106,6 @@ __unified_syscall:
 #include "../syscalls.s/n_sigprocmask.S"
 #include "../syscalls.s/n_sigsuspend.S"
 #include "../syscalls.s/nanosleep.S"
-#include "../syscalls.s/nice.S"
 #include "../syscalls.s/open.S"
 #include "../syscalls.s/pause.S"
 #include "../syscalls.s/personality.S"
diff --git a/arm/md5asm.S b/arm/md5asm.S
index 370b5039..feb0926a 100644
--- a/arm/md5asm.S
+++ b/arm/md5asm.S
@@ -1,3 +1,5 @@
+#include "arm-features.h"
+
 /*****************************************************************************
 * Copyright (C) 2002 Andre McCurdy  <armccurdy@yahoo.co.uk>
 *
@@ -75,8 +77,9 @@ MD5Init:
     mov     r1, #0
     str     r1, [r0, #0x10]                 @ initial count[0] = 0
     str     r1, [r0, #0x14]                 @ initial count[1] = 0
-    mov     pc, lr                          @ return
+    RET
 
+   .align 3
 1: .word    0x67452301                      @ initial MD5 context->state[0]
    .word    0xefcdab89                      @ initial MD5 context->state[1]
    .word    0x98badcfe                      @ initial MD5 context->state[2]
@@ -128,6 +131,7 @@ MD5Update:
     @ static void __MD5Transform (uint32_t *buf, const uint32_t *in, int repeat);
     @ --
 
+.align 3
 MD5MagicData:
 
 1: .word     0xd76aa478, 0xe8c7b756, 0x242070db, 0xc1bdceee
@@ -148,6 +152,7 @@ MD5MagicData:
    .word     0xf7537e82, 0xbd3af235, 0x2ad7d2bb, 0xeb86d391, (17f-19f-4)
    .word     0x6e4120A9, 0x20657264, 0x7543634d, 0x00796472, (19f-19f-4)
 
+.align 2
 __MD5Transform:
 
     cmp     r2, #0
diff --git a/arm/mmap.S b/arm/mmap.S
index d3ea1317..26824a59 100644
--- a/arm/mmap.S
+++ b/arm/mmap.S
@@ -1,3 +1,5 @@
+#include "arm-features.h"
+
 #include <errno.h>
 #include "syscalls.h"
 
@@ -28,7 +30,7 @@ mmap:
 	ldr     r4, [sp], #4
 	ldr     r5, [sp], #4
 	cmn	r0, #4096
-	mov	pc, lr			@ return
+	RET
 
 #else	
 		
diff --git a/arm/setjmp.S b/arm/setjmp.S
index 6b850d40..afe563dc 100644
--- a/arm/setjmp.S
+++ b/arm/setjmp.S
@@ -1,17 +1,41 @@
-.text
-.weak setjmp
-setjmp:
-.global __setjmp
-__setjmp:
+#include "arm-features.h"
+
+FUNC_START_WEAK	setjmp
+FUNC_START	__setjmp
 	mov	r1, #0
-.global __sigsetjmp
-__sigsetjmp:
-.weak sigsetjmp
-sigsetjmp:
-#ifndef	__SOFTFP__
-	sfm	f4, 4, [r0], #48
+FUNC_END	__setjmp
+FUNC_END	setjmp
+
+FUNC_START_WEAK	sigsetjmp
+FUNC_START	__sigsetjmp
+
+#ifdef DIET_HAVE_COPROC_REGS
+	/* we have to work on a copy of 'r0' (jmpbuf *) */
+	mov	ip, r0
+#endif
+
+#ifndef __SOFTFP__
+# if __ARM_ARCH__ >= 7
+	vstmia	ip!, {d0-d15}
+#   ifdef __ARM_NEON__
+	vstmia	ip!, {d16-d31}
+#   endif
+# else
+	sfm	f4, 4, [ip]!
+# endif
+#endif
+
+#ifdef __IWMMXT__
+#  warning "setjmp will not save iwmmxt coprocessor registers"
 #endif
+
+#ifndef DIET_HAVE_COPROC_REGS
 	stmia	r0, {r4-r11, sp, lr}
-	sub	r0, r0, #48
-	b	__sigjmp_save
 
+#else
+	stmia	ip!, {r4-r11, sp, lr}
+#endif
+
+	b	__sigjmp_save
+FUNC_END	__sigsetjmp
+FUNC_END	sigsetjmp
diff --git a/arm/start.S b/arm/start.S
index d68d49d6..1e5e0090 100644
--- a/arm/start.S
+++ b/arm/start.S
@@ -1,23 +1,10 @@
-
 #include "dietfeatures.h"
 #include "syscalls.h"
+#include "arm-features.h"
 
-	.text
 #ifdef __ARM_EABI__
-	.align 4
-#else
-	.align 2
-#endif
-	
-	.global _start
-	.weak 	exit
-	.global _exit
-
-
-#ifdef __ARM_EABI__
-
-_start:
 
+FUNC_START	_start
 	mov	fp, #0			@ clear the frame pointer
 	ldr	a1, [sp]		@ argc
 	add	a2, sp, #4		@ argv
@@ -25,27 +12,29 @@ _start:
 	add	a3, a2, a1, lsl #2	@ &argv[argc]
 	add	a3, a3, #4		@ envp	
 	str	a3, [ip, #0]		@ environ = envp
-	bl	main
-		
+	bl	CALL_IN_STARTCODE
+
 @
 @ The exit status from main() is already in r0.
 @ We need to branch to 'exit' in case we have linked with 'atexit'.
 @
 	bl	exit
+FUNC_END	_start
 
-exit:
-_exit:
-
+FUNC_START	_exit
+FUNC_START_WEAK	exit
 	mov	r7, #__NR_exit		
 	swi	0			@ never returns.
 
+	.align	2
 .L3:	.word 	environ
 
-	
-#else	
+FUNC_END	exit
+FUNC_END	_exit
 
-	
-_start:
+#else
+
+FUNC_START	_start
 
 #ifdef WANT_DYNAMIC
 	mov	a4, a1			@ save dynamic ld.so _fini
@@ -83,18 +72,20 @@ _start:
 @ We need to branch to 'exit' in case we have linked with 'atexit'.
 @
 	bl	exit
+FUNC_END	_start
 
-exit:
-_exit:
-
+FUNC_START	_exit
+FUNC_START_WEAK	exit
 #ifdef PROFILING
 	mov	r4, r0			@ save a copy of exit status
 	bl	_stop_monitor
 	mov	r0, r4
 #endif
 	swi	$__NR_exit		@ never returns.
+FUNC_END	exit
+FUNC_END	_exit
 
-
+	.align	2
 #ifdef __DYN_LIB
 .L3:	.word 	environ(GOT)
 .L4:	.word 	_GLOBAL_OFFSET_TABLE_-(1b+8)
diff --git a/arm/strcpy.S b/arm/strcpy.S
index 20e1029d..7a865620 100644
--- a/arm/strcpy.S
+++ b/arm/strcpy.S
@@ -1,10 +1,7 @@
 #include "dietfeatures.h"
+#include "arm-features.h"
 
-.text
-	.align	2
-	.global	strcpy
-
-strcpy:
+FUNC_START	strcpy
 #ifndef WANT_SMALL_STRING_ROUTINES
 	mov	ip, r0
 	ands	r2, r1, #3
@@ -61,6 +58,5 @@ strcpy:
 	ldrneb	r2, [r1], #1
 #endif
 	bne	.Lloop
-	mov	pc, lr
-.Lfe1:
-	.size	 strcpy,.Lfe1-strcpy
+	RET
+FUNC_END	strcpy
diff --git a/arm/strlen.S b/arm/strlen.S
index 6b2b459a..a6af8f0d 100644
--- a/arm/strlen.S
+++ b/arm/strlen.S
@@ -1,12 +1,7 @@
 #include "dietfeatures.h"
+#include "arm-features.h"
 
-	.text
-	.align	2
-
-	.global	strlen
-
-strlen:
-
+FUNC_START	strlen
 #if 0
 	teq	a1, #0			@ is string pointer NULL ??
 	moveq	pc, lr			@ if so, return 0
@@ -61,12 +56,10 @@ strlen:
 	sub	a1, a1, a2
 #endif
 
-	mov	pc, lr
+	RET
 
 #ifndef WANT_SMALL_STRING_ROUTINES
 .Lmagic:
 	.word	0x01010101
 #endif
-
-.Lstrlen:
-	.size	strlen,.Lstrlen-strlen
+FUNC_END	strlen
diff --git a/arm/syscalls.h b/arm/syscalls.h
index 896fd5e2..bef48a52 100644
--- a/arm/syscalls.h
+++ b/arm/syscalls.h
@@ -727,9 +727,9 @@
 #define __ARGS_getpeername		0
 #define __ARGS_socketpair		0
 #define __ARGS_send			0
-#define __ARGS_sendto			0
+#define __ARGS_sendto			1
 #define __ARGS_recv			0
-#define __ARGS_recvfrom			0
+#define __ARGS_recvfrom			1
 #define __ARGS_shutdown			0
 #define __ARGS_setsockopt		0
 #define __ARGS_getsockopt		0
@@ -824,70 +824,30 @@
 
 #ifdef __ASSEMBLER__
 
-#ifdef __ARM_EABI__
+#include "arm-features.h"
 
 #define syscall_weak(name,wsym,sym) __syscall_weak __NR_##name, wsym, sym, __ARGS_##name
 .macro __syscall_weak name wsym sym typ
-.text
-.type \wsym,function
-.weak \wsym
-\wsym:
-.type \sym,function
-.global \sym
-\sym:
-        stmfd	sp!,{r4,r5,r7,lr}
-	ldr	r4, [sp,#16]
-	ldr	r5, [sp,#20]
-        ldr     r7, =\name
-	swi	0
-	b	__unified_syscall
+FUNC_START_WEAK	\wsym
+__syscall	\name, \sym, \typ
+FUNC_END	\wsym
 .endm
 
+#ifdef __ARM_EABI__
 
 #define syscall(name,sym) __syscall __NR_##name, sym, __ARGS_##name
 .macro __syscall name sym typ
-.text
-.type \sym,function
-.global \sym
-\sym:
-        stmfd	sp!,{r4,r5,r7,lr}
-	ldr	r4, [sp,#16]
-	ldr	r5, [sp,#20]
-        ldr     r7, =\name
-	swi	0
-	b	__unified_syscall
+FUNC_START	\sym
+        ldr     ip, =\name
+	b	__unified_syscall_swi
+FUNC_END	\sym
 .endm
 
 #else
 
-#define syscall_weak(name,wsym,sym) __syscall_weak $__NR_##name, wsym, sym, __ARGS_##name
-.macro __syscall_weak name wsym sym typ
-.text
-.type \wsym,function
-.weak \wsym
-\wsym:
-.type \sym,function
-.global \sym
-\sym:
-.ifgt \typ
-	mov	ip, sp
-	stmfd	sp!,{r4, r5, r6}
-	ldmia	ip, {r4, r5, r6}
-.endif
-	swi	\name
-.ifgt \typ
-	b	__unified_syscall4
-.else
-	b	__unified_syscall
-.endif
-.endm
-
 #define syscall(name,sym) __syscall $__NR_##name, sym, __ARGS_##name
 .macro __syscall name sym typ
-.text
-.type \sym,function
-.global \sym
-\sym:
+FUNC_START	\sym
 .ifgt \typ
 	mov	ip, sp
 	stmfd	sp!,{r4, r5, r6}
@@ -899,6 +859,7 @@
 .else
 	b	__unified_syscall
 .endif
+FUNC_END	\sym
 .endm
 
 #endif
diff --git a/arm/unified.S b/arm/unified.S
index e6ea3f66..7f6d5311 100644
--- a/arm/unified.S
+++ b/arm/unified.S
@@ -1,22 +1,29 @@
-
 #include <dietfeatures.h>
+#include "arm-features.h"
 
-	.text
 #ifdef __ARM_EABI__
-	.align 4
-#else
-	.align 2
-#endif 	
-	.global __unified_syscall
-	.global __unified_syscall4
 
+/* expects:
+ * r0-r3 ... syscall arguments 0-3
+ * ip    ... syscall number
+ */
+FUNC_START	__unified_syscall_swi
+	.hidden	__unified_syscall_swi
+        stmfd	sp!,{r4,r5,r7,lr}
+	mov	r7, ip
+	LOAD_ARG4_5
+	swi	0
+	/* fallthrough to __unified4_syscall */
+FUNC_END	__unified_syscall_swi
 
-#ifdef __ARM_EABI__
-
-__unified_syscall4:
-__unified_syscall:
-	
+/* expects:
+ * r0    ... syscall return value
+ * original r4-r7 + lr on stack
+ */
+FUNC_START	__unified_syscall
+	.hidden	__unified_syscall
         cmn     r0, #4096
+#ifndef WANT_THREAD_SAFE
         rsbcs   r2, r0, #0
         ldrcs   r3, 1f
         mvncs   r0, #0
@@ -25,17 +32,28 @@ __unified_syscall:
 	.balign 4
 1:
         .word   errno
-	
+#else
+	bcc	1f
+	rsb	r4, r0, #0
+	bl	__errno_location
+	str	r4, [r0]
+	mvn	r0, #0
+1:
+        ldmfd   sp!,{r4,r5,r7,pc}
+#endif
+FUNC_END	__unified_syscall
+
 /* here we go and "reuse" the return for weak-void functions */
 #include "dietuglyweaks.h"
 
-	mov	pc, lr			@ return
+	RET
 
-#else	
-	
-__unified_syscall4:
+#else
+
+FUNC_START	__unified_syscall4
 	ldmia	sp!, {r4, r5, r6}
-__unified_syscall:
+
+FUNC_START	__unified_syscall
 	cmn	r0, #4096
 	movcc	pc, lr			@ return value comes direct from kernel.
 
@@ -53,10 +71,13 @@ __unified_syscall:
 /* here we go and "reuse" the return for weak-void functions */
 #include "dietuglyweaks.h"
 
-	mov	pc, lr			@ return
+	RET
 
 #ifndef WANT_THREAD_SAFE
+.align 2
 .L0:	.long 	errno
 #endif
+FUNC_END	__unified_syscall
+FUNC_END	__unified_syscall4
 
 #endif
diff --git a/arm/waitpid.S b/arm/waitpid.S
index 2c3a75b3..facaff13 100644
--- a/arm/waitpid.S
+++ b/arm/waitpid.S
@@ -1,15 +1,8 @@
-.text
-#ifdef __ARM_EABI__
-.align	4
-#else	
-.align	2
-#endif
-.weak	waitpid
-.type	waitpid, %function
-waitpid:
-.global	__libc_waitpid
-.type	__libc_waitpid, %function
-__libc_waitpid:
-	mov	r3, #0
-	b	wait4
-	.size	waitpid, .-waitpid
+#include "arm-features.h"
+
+FUNC_START_WEAK	waitpid
+FUNC_START	__libc_waitpid
+ 	mov	r3, #0
+ 	b	wait4
+FUNC_END	__libc_waitpid
+FUNC_END	waitpid
diff --git a/diet.c b/diet.c
index 5f3a2ba1..e0406381 100644
--- a/diet.c
+++ b/diet.c
@@ -25,7 +25,7 @@ static void error(const char *message) {
 }
 
 static const char* Os[] = {
-  "i386","-Os","-mpreferred-stack-boundary=2",
+  "i386","-Os","-mpreferred-stack-boundary=4",
 	 "-falign-functions=1","-falign-jumps=1",
 	 "-falign-loops=1","-fomit-frame-pointer",0,
   "x86_64","-Os",0,
diff --git a/findcflags.sh b/findcflags.sh
index 2d600d6a..5b07d970 100755
--- a/findcflags.sh
+++ b/findcflags.sh
@@ -3,6 +3,6 @@ version=`${1:-gcc} -v 2>&1 |sed -n '/gcc version/ {s/gcc version //;p;}'`
 case $version in
   2.9*) echo -march=i386 -Os -fomit-frame-pointer -malign-functions=1 -malign-jumps=1 -malign-loops=1 -mpreferred-stack-boundary=2 ;;
   3.0*) echo -march=i386 -Os -fomit-frame-pointer -malign-functions=1 -malign-jumps=1 -malign-loops=1 -mpreferred-stack-boundary=2 ;;
-  [34]*) echo -Os -fomit-frame-pointer -falign-functions=1 -falign-jumps=1 -falign-loops=1 -mpreferred-stack-boundary=2;;
+  [34]*) echo -Os -fomit-frame-pointer -falign-functions=1 -falign-jumps=1 -falign-loops=1 -mpreferred-stack-boundary=4;;
   *) echo -O2 -pipe -fomit-frame-pointer ;;
 esac
diff --git a/ia64/clone.S b/ia64/clone.S
index 0f879704..38bd1a64 100644
--- a/ia64/clone.S
+++ b/ia64/clone.S
@@ -1,4 +1,4 @@
-/* This is untested code which probably won´t work out of the box! */
+/* This is untested code which probably won't work out of the box! */
 
 #include "syscalls.h"
 #include <errno.h>
@@ -43,4 +43,5 @@ __clone:
 	br.call.dptk.few b0=__error_unified_syscall
 	br.ret.sptk.few b0
 .endp __clone
+.endp __clone2
 .size __clone, . - __clone
diff --git a/ia64/unified.S b/ia64/unified.S
index 548e0506..4158d9ec 100644
--- a/ia64/unified.S
+++ b/ia64/unified.S
@@ -16,9 +16,7 @@
 .text
 
 .globl __unified_syscall
-.proc  __unified_syscall
 .globl __error_unified_syscall
-.proc  __error_unified_syscall
 .globl _exit
 .proc  _exit
 
@@ -27,11 +25,16 @@ _exit:
 .endp _exit
 .size _exit, . - _exit
 
+.proc  __unified_syscall
 __unified_syscall:
 	break.i 0x100000
 	movl r2=errno
 	cmp.eq p6,p0=-1,r10
 	;;
+.endp __unified_syscall
+.size __unified_syscall, . - __unified_syscall
+
+.proc  __error_unified_syscall
 __error_unified_syscall:
 (p6)    st4 [r2]=r8
 (p6)    mov r8=-1
@@ -39,6 +42,6 @@ __error_unified_syscall:
 #include "dietuglyweaks.h"
 
 	br.ret.sptk.few rp
+.endp __error_unified_syscall
 
-.size __unified_syscall, __error_unified_syscall - __unified_syscall
 .size __error_unified_syscall, . - __error_unified_syscall
diff --git a/include/fcntl.h b/include/fcntl.h
index 24b64ba4..1c6c32b7 100644
--- a/include/fcntl.h
+++ b/include/fcntl.h
@@ -621,8 +621,10 @@ struct flock64
 extern int fcntl (int __fd, int __cmd, ...) __THROW;
 #ifndef __NO_STAT64
 extern int fcntl64 (int __fd, int __cmd, ...) __THROW;
+extern int fstatat64(int dirfd, const char *pathname, struct stat *buf, int flags) __THROW;
 #if defined _FILE_OFFSET_BITS && _FILE_OFFSET_BITS == 64
 #define fcntl fcntl64
+#define fstatat fstatat64
 #endif
 #endif
 
diff --git a/include/setjmp.h b/include/setjmp.h
index 10b0a07b..7fb03031 100644
--- a/include/setjmp.h
+++ b/include/setjmp.h
@@ -166,7 +166,7 @@ typedef int __jmp_buf[3];
 #ifdef __arm__
 #define __JMP_BUF_SP            8
 #ifndef __ASSEMBLER__
-typedef int __jmp_buf[24];
+typedef int __jmp_buf[10 + 16*2 + 16*2];
 #endif
 #endif
 
diff --git a/include/stdlib.h b/include/stdlib.h
index 4976f865..f6e760ab 100644
--- a/include/stdlib.h
+++ b/include/stdlib.h
@@ -29,7 +29,7 @@ long int strtol(const char *nptr, char **endptr, int base) __THROW;
 unsigned long int strtoul(const char *nptr, char **endptr, int base) __THROW;
 
 extern int __ltostr(char *s, unsigned int size, unsigned long i, unsigned int base, int UpCase) __THROW;
-extern int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned int prec2,int g) __THROW;
+extern int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned int prec2,int flags) __THROW;
 
 #if !defined(__STRICT_ANSI__) || __STDC_VERSION__ + 0 >= 199900L
 __extension__ long long int strtoll(const char *nptr, char **endptr, int base) __THROW;
@@ -43,7 +43,7 @@ double atof(const char *nptr) __THROW;
 __extension__ long long int atoll(const char *nptr);
 
 void exit(int status) __THROW __attribute__((__noreturn__));
-void abort(void) __THROW;
+void abort(void) __THROW __attribute__((__noreturn__));
 
 extern int rand(void) __THROW;
 extern int rand_r(unsigned int *seed) __THROW;
diff --git a/include/unistd.h b/include/unistd.h
index e1e4457d..56494c0c 100644
--- a/include/unistd.h
+++ b/include/unistd.h
@@ -39,6 +39,8 @@ loff_t lseek64(int fildes, loff_t offset, int whence) __THROW;
 #if defined _FILE_OFFSET_BITS && _FILE_OFFSET_BITS == 64
 #define lseek(fildes,offset,whence) lseek64(fildes,offset,whence)
 #endif
+#else
+#define lseek64(fildes,offset,whence) lseek(fildes,offset,whence)
 #endif
 
 int chdir(const char *path) __THROW;
diff --git a/lib/__dtostr.c b/lib/__dtostr.c
index 1d082e3b..bc612008 100644
--- a/lib/__dtostr.c
+++ b/lib/__dtostr.c
@@ -5,13 +5,15 @@
 
 static int copystring(char* buf,int maxlen, const char* s) {
   int i;
-  for (i=0; i<3&&i<maxlen; ++i)
+  for (i=0; i<maxlen; ++i) {
     buf[i]=s[i];
-  if (i<maxlen) { buf[i]=0; ++i; }
+    if (!s[i])
+      break;
+  }
   return i;
 }
 
-int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned int prec2,int g) {
+int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned int prec2,int flags) {
 #if 1
   union {
     unsigned long long l;
@@ -35,8 +37,12 @@ int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned i
   double tmp;
   char *oldbuf=buf;
 
-  if ((i=isinf(d))) return copystring(buf,maxlen,i>0?"inf":"-inf");
-  if (isnan(d)) return copystring(buf,maxlen,"nan");
+  if (isinf(d))
+    return copystring(buf,maxlen,
+		      (d<0)?
+		      (flags&0x02?"-INF":"-inf"):
+		      (flags&0x02?"INF":"inf"));
+  if (isnan(d)) return copystring(buf,maxlen,flags&0x02?"NAN":"nan");
   e10=1+(long)(e*0.30102999566398119802); /* log10(2) */
   /* Wir iterieren von Links bis wir bei 0 sind oder maxlen erreicht
    * ist.  Wenn maxlen erreicht ist, machen wir das nochmal in
@@ -126,7 +132,7 @@ int __dtostr(double d,char *buf,unsigned int maxlen,unsigned int prec,unsigned i
   if (prec2 || prec>(unsigned int)(buf-oldbuf)+1) {	/* more digits wanted */
     if (!maxlen) return 0; --maxlen;
     *buf='.'; ++buf;
-    if (g) {
+    if ((flags & 0x01)) {
       if (prec2) prec=prec2;
       prec-=buf-oldbuf-1;
     } else {
diff --git a/lib/__utime.c b/lib/__utime.c
index e0132653..2d0e4af6 100644
--- a/lib/__utime.c
+++ b/lib/__utime.c
@@ -1,5 +1,7 @@
-#include <utime.h>
 #include <syscalls.h>
+#define _BSD_SOURCE
+#include <utime.h>
+#include <sys/time.h>
 
 #ifndef __NR_utime
 int utime(const char *filename, const struct utimbuf *times)
diff --git a/lib/__v_printf.c b/lib/__v_printf.c
index 36202f59..964c0056 100644
--- a/lib/__v_printf.c
+++ b/lib/__v_printf.c
@@ -4,6 +4,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <errno.h>
+#include <math.h>
 #include "dietstdio.h"
 #include "dietwarning.h"
 
@@ -346,45 +347,49 @@ int __v_printf(struct arg_printf* fn, const char *format, va_list arg_ptr)
 #ifdef WANT_FLOATING_POINT_IN_PRINTF
       /* print a floating point value */
       case 'f':
+      case 'F':
       case 'g':
+      case 'G':
 	{
-	  int g=(ch=='g');
+	  int flags=(((ch&0x5f)=='G') ? 0x01 : 0x00) | ((ch&0x20) ? 0x00 : 0x02);
 	  double d=va_arg(arg_ptr,double);
 	  s=buf+1;
 	  if (width==0) width=1;
 	  if (!flag_dot) preci=6;
 	  if (flag_sign || d < +0.0) flag_in_sign=1;
 
-	  sz=__dtostr(d,s,sizeof(buf)-1,width,preci,g);
-
-	  if (flag_dot) {
-	    char *tmp;
-	    if ((tmp=strchr(s,'.'))) {
-	      if (preci || flag_hash) ++tmp;
-	      while (preci>0 && *++tmp) --preci;
-	      *tmp=0;
-	    } else if (flag_hash) {
-	      s[sz]='.';
-	      s[++sz]='\0';
+	  sz=__dtostr(d,s,sizeof(buf)-1,width,preci,flags);
+
+	  if (!isnan(d) && !isinf(d)) {		/* skip NaN + INF values */
+	    if (flag_dot) {
+	      char *tmp;
+	      if ((tmp=strchr(s,'.'))) {
+		if (preci || flag_hash) ++tmp;
+		while (preci>0 && *++tmp) --preci;
+		*tmp=0;
+	      } else if (flag_hash) {
+		s[sz]='.';
+		s[++sz]='\0';
+	      }
 	    }
-	  }
 
-	  if (g) {
-	    char *tmp,*tmp1;	/* boy, is _this_ ugly! */
-	    if ((tmp=strchr(s,'.'))) {
-	      tmp1=strchr(tmp,'e');
-	      while (*tmp) ++tmp;
-	      if (tmp1) tmp=tmp1;
-	      while (*--tmp=='0') ;
-	      if (*tmp!='.') ++tmp;
-	      *tmp=0;
-	      if (tmp1) strcpy(tmp,tmp1);
+	    if ((flags&0x01)) {
+	      char *tmp,*tmp1;	/* boy, is _this_ ugly! */
+	      if ((tmp=strchr(s,'.'))) {
+		tmp1=strchr(tmp,'e');
+		while (*tmp) ++tmp;
+		if (tmp1) tmp=tmp1;
+		while (*--tmp=='0') ;
+		if (*tmp!='.') ++tmp;
+		*tmp=0;
+		if (tmp1) strcpy(tmp,tmp1);
+	      }
 	    }
-	  }
 	  
-	  if ((flag_sign || flag_space) && d>=0) {
-	    *(--s)=(flag_sign)?'+':' ';
-	    ++sz;
+	    if ((flag_sign || flag_space) && d>=0) {
+	      *(--s)=(flag_sign)?'+':' ';
+	      ++sz;
+	    }
 	  }
 	  
 	  sz=strlen(s);
diff --git a/lib/stackgap.c b/lib/stackgap.c
index 55842b14..5bf7bf98 100644
--- a/lib/stackgap.c
+++ b/lib/stackgap.c
@@ -126,6 +126,9 @@ void __setup_tls(tcbhead_t* mainthread) {
 #elif defined(__sparc__)
   register tcbhead_t* __thread_self __asm("%g7");
   __thread_self=mainthread;
+#elif defined(__arm__)
+  extern void __arm_set_tls(void*);
+  __arm_set_tls(mainthread);
 #else
 #warning "no idea how to enable TLS on this platform, edit lib/stackgap.c"
 #endif
diff --git a/libpthread/pthread_errno.c b/libpthread/pthread_errno.c
deleted file mode 100644
index d035544a..00000000
--- a/libpthread/pthread_errno.c
+++ /dev/null
@@ -1,5 +0,0 @@
-int errno;
-
-#include "dietwarning.h"
-
-link_warning("errno","\e[1;33;41m>>> your multithreaded code uses errno! <<<\e[0m");
diff --git a/libugly/asctime.c b/libugly/asctime.c
index e1fcc870..9d304e6b 100644
--- a/libugly/asctime.c
+++ b/libugly/asctime.c
@@ -1,6 +1,6 @@
 #include <time.h>
 
-static char buf[25];
+static char buf[26];
 
 char *asctime(const struct tm *timeptr) {
   return asctime_r(timeptr,buf);
diff --git a/libugly/asctime_r.c b/libugly/asctime_r.c
index 43c5e2cb..aae09ad0 100644
--- a/libugly/asctime_r.c
+++ b/libugly/asctime_r.c
@@ -25,5 +25,6 @@ char *asctime_r(const struct tm *t, char *buf) {
   num2str(buf+20,(t->tm_year+1900)/100);
   num2str(buf+22,(t->tm_year+1900)%100);
   buf[24]='\n';
+  buf[25]='\0';
   return buf;
 }
diff --git a/libugly/strftime.c b/libugly/strftime.c
index 2a1e8a8b..d327d793 100644
--- a/libugly/strftime.c
+++ b/libugly/strftime.c
@@ -45,38 +45,39 @@ size_t  strftime ( char* dst, size_t max, const char* format, const struct tm* t
 	    else
 again:
 	    switch (*format) {
-//          case '%': *p++ = '%';  				 break;			// reduce size of jump table
-	    case 'n': *p++ = '\n'; 				 break;
-	    case 't': *p++ = '\t'; 				 break;
+//          case '%': *p++ = '%';  				break;			// reduce size of jump table
+	    case 'n': *p++ = '\n'; 				break;
+	    case 't': *p++ = '\t'; 				break;
 	    case 'O': case 'E': ++format; goto again;
-	    case 'c': src = "%b %a %d %k:%M:%S %Z %Y";        	 goto _strf;
-	    case 'r': src = "%I:%M:%S %p";                    	 goto _strf;
-	    case 'R': src = "%H:%M";      			 goto _strf;
-	    case 'x': src = "%b %a %d";   			 goto _strf;
-	    case 'X': src = "%k:%M:%S";   			 goto _strf;
-	    case 'D': src = "%m/%d/%y";   			 goto _strf;
+	    case 'c': src = "%b %a %d %k:%M:%S %Z %Y";        	goto _strf;
+	    case 'r': src = "%I:%M:%S %p";                    	goto _strf;
+	    case 'R': src = "%H:%M";      			goto _strf;
+	    case 'x': src = "%b %a %d";   			goto _strf;
+	    case 'X': src = "%k:%M:%S";   			goto _strf;
+	    case 'D': src = "%m/%d/%y";   			goto _strf;
+	    case 'F': src = "%Y-%m-%d";				goto _strf;
 	    case 'T': src = "%H:%M:%S";
 	       _strf: p  += strftime (p, (size_t)(dst+max-p), src, tm); 	 break;
-	    case 'a': src = sweekdays [tm->tm_wday]; 		 goto _str;
-	    case 'A': src = weekdays  [tm->tm_wday]; 		 goto _str;
+	    case 'a': src = sweekdays [tm->tm_wday]; 		goto _str;
+	    case 'A': src = weekdays  [tm->tm_wday]; 		goto _str;
 	    case 'h':
-	    case 'b': src = smonths   [tm->tm_mon];  		 goto _str;
-	    case 'B': src = months    [tm->tm_mon];  		 goto _str;
+	    case 'b': src = smonths   [tm->tm_mon];  		goto _str;
+	    case 'B': src = months    [tm->tm_mon];  		goto _str;
 	    case 'p': src = ampm [tm->tm_hour > 12 ? 3 : 2]; goto _str;
 	    case 'P': src = ampm [tm->tm_hour > 12 ? 1 : 0]; goto _str;
-	    case 'C': no  = tm->tm_year/100 + 19; 		 goto _no;
-	    case 'd': no  = tm->tm_mday;          		 goto _no;
-	    case 'e': no  = tm->tm_mday;          		 goto _nos;
-	    case 'H': no  = tm->tm_hour;          		 goto _no;
-	    case 'I': no  = tm->tm_hour % 12;     		 goto _no;
-	    case 'j': no  = tm->tm_yday;          		 goto _no;
-	    case 'k': no  = tm->tm_hour;          		 goto _nos;
-	    case 'l': no  = tm->tm_hour % 12;     		 goto _nos;
-	    case 'm': no  = tm->tm_mon + 1;         		 goto _no;
-	    case 'M': no  = tm->tm_min;           		 goto _no;
-	    case 'S': no  = tm->tm_sec;           		 goto _no;
-	    case 'u': no  = tm->tm_wday ? tm->tm_wday : 7; 	 goto _no;
-	    case 'w': no  = tm->tm_wday;              		 goto _no;
+	    case 'C': no  = tm->tm_year/100 + 19; 		goto _no;
+	    case 'd': no  = tm->tm_mday;          		goto _no;
+	    case 'e': no  = tm->tm_mday;          		goto _nos;
+	    case 'H': no  = tm->tm_hour;          		goto _no;
+	    case 'I': no  = tm->tm_hour % 12;     		goto _no;
+	    case 'j': no  = tm->tm_yday;          		goto _no;
+	    case 'k': no  = tm->tm_hour;          		goto _nos;
+	    case 'l': no  = tm->tm_hour % 12;     		goto _nos;
+	    case 'm': no  = tm->tm_mon + 1;         		goto _no;
+	    case 'M': no  = tm->tm_min;           		goto _no;
+	    case 'S': no  = tm->tm_sec;           		goto _no;
+	    case 'u': no  = tm->tm_wday ? tm->tm_wday : 7; 	goto _no;
+	    case 'w': no  = tm->tm_wday;              		goto _no;
 	    case 'U': no  = (tm->tm_yday - tm->tm_wday + 7) / 7; goto _no;
 	    case 'W': no  = (tm->tm_yday - (tm->tm_wday - 1 + 7) % 7 + 7) / 7; goto _no;
 	    case 's': {
diff --git a/libugly/strptime.c b/libugly/strptime.c
index 9d7f530c..7babbb7d 100644
--- a/libugly/strptime.c
+++ b/libugly/strptime.c
@@ -119,10 +119,10 @@ char* strptime(const char* s,const char* format, struct tm* tm) {
 	++s;
 	break;
       case 'x':
-	s=strptime(s,"%b %a %d",tm);
+	s=strptime(s,"%m/%d/%y",tm);
 	break;
       case 'X':
-	s=strptime(s,"%k:%M:%S",tm);
+	s=strptime(s,"%H:%M:%S",tm);
 	break;
       case 'y':
 	i=getint(&s,2);
diff --git a/parisc/strcmp.S b/parisc/strcmp.S
index 9dddd59d..737cbdb1 100644
--- a/parisc/strcmp.S
+++ b/parisc/strcmp.S
@@ -1,6 +1,8 @@
 .text
 .globl strcmp
 
+.weak strcoll
+strcoll:
 strcmp:
 	.PROC
 	.CALLINFO
diff --git a/s390/Makefile.add b/s390/Makefile.add
index b70b9715..65c20f00 100644
--- a/s390/Makefile.add
+++ b/s390/Makefile.add
@@ -1,4 +1,4 @@
 
 CFLAGS+=-Os # -fomit-frame-pointer -fstrict-aliasing
 VPATH:=s390:syscalls.s:$(VPATH)
-LIBOBJ+=$(OBJDIR)/__waitpid.o
+LIBOBJ+=$(OBJDIR)/__waitpid.o $(OBJDIR)/__fadvise.o
diff --git a/s390/__fadvise.c b/s390/__fadvise.c
new file mode 100644
index 00000000..f06b8a3c
--- /dev/null
+++ b/s390/__fadvise.c
@@ -0,0 +1,29 @@
+#include "syscalls.h"
+
+#define _LINUX_SOURCE
+#include <fcntl.h>
+#include <unistd.h>
+
+struct fadvise64_64_layout
+{
+  int fd;
+  off64_t offset;
+  off64_t len;
+  int advise;
+};
+
+long fadvise64_64(int fd, off64_t offset, off64_t len, int advise)
+{
+  struct fadvise64_64_layout parameters;
+  extern long __s390_fadvise64_64(struct fadvise64_64_layout *args);
+
+  parameters.fd = fd;
+  parameters.offset = offset;
+  parameters.len = len;
+  parameters.advise = advise;
+
+  return __s390_fadvise64_64(&parameters);
+}
+
+int posix_fadvise(int fd, off64_t offset, off64_t len, int advise)
+  __attribute__((__alias__("fadvise64_64")));
diff --git a/s390/setjmp.S b/s390/setjmp.S
index 7522373c..762f9e34 100644
--- a/s390/setjmp.S
+++ b/s390/setjmp.S
@@ -14,6 +14,7 @@ __setjmp:
 
 .global _setjmp
 .type	_setjmp,@function
+_setjmp:
 	lhi	%r3,0
 .size	_setjmp,.-_setjmp
 
@@ -26,15 +27,18 @@ __sigsetjmp:
 	std	%f6,48(%r2)
 
 	/* Make a tail call to __sigjmp_save; it takes the same args.  */
-	basr	%r1,%r0
+	basr	%r1,0
 #ifdef	PIC
-.L0:	la	%r1,.L1-.L0(0,%r1)
+.L0:	al	%r1,.L1-.L0(0,%r1)
 	l	%r1,__sigjmp_save@GOT12(0,%r1)
+#else
+.L0:	l	%r1,.L1-.L0(0,%r1)
+#endif
 	br	%r1
+	.p2align 3
+#ifdef	PIC
 .L1:	.long	_GLOBAL_OFFSET_TABLE_-.L0
 #else
-	l	%r1,.L1(0,%r1)
-	br	%r1
 .L1:	.long	__sigjmp_save
 #endif
 .size __sigsetjmp,.-__sigsetjmp;
diff --git a/s390/syscalls.h b/s390/syscalls.h
index a383547d..922a72bc 100644
--- a/s390/syscalls.h
+++ b/s390/syscalls.h
@@ -277,6 +277,7 @@
 #define __NR_mknodat		290
 #define __NR_fchownat		291
 #define __NR_futimesat		292
+#define __NR_fstatat64		293
 #define __NR_unlinkat		294
 #define __NR_renameat		295
 #define __NR_linkat		296
diff --git a/s390x/setjmp.S b/s390x/setjmp.S
index 7522373c..762f9e34 100644
--- a/s390x/setjmp.S
+++ b/s390x/setjmp.S
@@ -14,6 +14,7 @@ __setjmp:
 
 .global _setjmp
 .type	_setjmp,@function
+_setjmp:
 	lhi	%r3,0
 .size	_setjmp,.-_setjmp
 
@@ -26,15 +27,18 @@ __sigsetjmp:
 	std	%f6,48(%r2)
 
 	/* Make a tail call to __sigjmp_save; it takes the same args.  */
-	basr	%r1,%r0
+	basr	%r1,0
 #ifdef	PIC
-.L0:	la	%r1,.L1-.L0(0,%r1)
+.L0:	al	%r1,.L1-.L0(0,%r1)
 	l	%r1,__sigjmp_save@GOT12(0,%r1)
+#else
+.L0:	l	%r1,.L1-.L0(0,%r1)
+#endif
 	br	%r1
+	.p2align 3
+#ifdef	PIC
 .L1:	.long	_GLOBAL_OFFSET_TABLE_-.L0
 #else
-	l	%r1,.L1(0,%r1)
-	br	%r1
 .L1:	.long	__sigjmp_save
 #endif
 .size __sigsetjmp,.-__sigsetjmp;
diff --git a/sparc/strchr.S b/sparc/strchr.S
index e27e4541..8c2c78c6 100644
--- a/sparc/strchr.S
+++ b/sparc/strchr.S
@@ -11,6 +11,8 @@
 
 .text
 .globl strchr
+.weak index
+index:
 strchr:
 	ldub [%o0], %o2
 	cmp %o2, %o1
diff --git a/sparc/strcmp.S b/sparc/strcmp.S
index 50c82b5e..81379f20 100644
--- a/sparc/strcmp.S
+++ b/sparc/strcmp.S
@@ -11,6 +11,8 @@
 
 .text
 .globl strcmp
+.weak strcoll
+strcoll:
 strcmp:
 	clr %o4
 .Lloop:
diff --git a/syscalls.s/fadvise64.S b/syscalls.s/fadvise64.S
index 89a6eea6..23e01e46 100644
--- a/syscalls.s/fadvise64.S
+++ b/syscalls.s/fadvise64.S
@@ -3,15 +3,14 @@
 
 #include "syscalls.h"
 
+#ifdef __NR_fadvise64
+syscall(fadvise64,fadvise64)
+
 #ifndef __NR_fadvise64_64
+posix_fadvise = fadvise64
 .globl posix_fadvise
-.type posix_fadvise,@function
-posix_fadvise:
 #endif
 
-#ifdef __NR_fadvise64
-syscall(fadvise64,fadvise64)
-
 #endif
 
 #endif
diff --git a/syscalls.s/newfstatat.S b/syscalls.s/newfstatat.S
index daa27147..f10d087a 100644
--- a/syscalls.s/newfstatat.S
+++ b/syscalls.s/newfstatat.S
@@ -2,4 +2,12 @@
 
 #ifdef __NR_newfstatat
 syscall(newfstatat,newfstatat)
+
+#ifdef __NR_fstatat64
+#  error __NR_newfstatat and __NR_fstatat64 must not both be defined
+#endif
+
+fstatat = newfstatat
+.globl fstatat
+
 #endif
diff --git a/test/asprintf.c b/test/asprintf.c
index 996a5aa5..0d4f2eb3 100644
--- a/test/asprintf.c
+++ b/test/asprintf.c
@@ -13,7 +13,7 @@ int main(int argc, char **argv) {
  assert(strlen(path) == asprintlen);
 	     
  printf("%s\n", path);
- asprintlen=asprintf(&path, "/proc" "/%d/stat", strlen(argv[1]));
+ asprintlen=asprintf(&path, "/proc" "/%zu/stat", strlen(argv[1]));
  assert(strlen(path) == asprintlen);
  printf("%s\n", path);
 
diff --git a/test/atexit.c b/test/atexit.c
index 709a3e44..79ee392b 100644
--- a/test/atexit.c
+++ b/test/atexit.c
@@ -2,7 +2,7 @@
 #include <unistd.h>
 
 void blah(void) {
-  write(2,"atexit\n",7);
+  write(1,"atexit\n",7);
 }
 
 int main() {
diff --git a/test/cycles.c b/test/cycles.c
index 35547f8a..c7805cd3 100644
--- a/test/cycles.c
+++ b/test/cycles.c
@@ -12,6 +12,18 @@
   asm volatile ("rdtsc" : "=a" (l), "=d" (h)); \
   dst = (((uint64_t)h) << 32) | l;                             \
 } while (0)
+#elif defined (__powerpc64__)
+#define RDTSC(dst) asm volatile ("mftb %0" : "=r" (dst))
+#elif defined (__powerpc__)
+#define RDTSC(dst) do { \
+  uint32_t chk, tbl, tbu; \
+  /* The code below is as suggested in Motorola reference manual for 32 bits PPCs. */ \
+  __asm__ __volatile__ ("1: mftbu %0; mftb %1; mftbu %2; cmpw %2,%0; bne 1b" \
+    : "=r" (tbu), "=r" (tbl), "=r" (chk) ); \
+  dst = ((uint64_t)tbu << 32) | tbl; \
+} while (0)
+#elif defined (__sparcv9__)
+#define RDTSC(dst) asm volatile ("rd %%tick, %%0":"=r"(dst))
 #else
 #error "Unimplemented rdtsc"
 #endif