diff --git a/Makefile b/Makefile
index 5d83efc..e4fc907 100644
--- a/Makefile
+++ b/Makefile
@@ -20,7 +20,7 @@ EXPORT_UNPREFIXED := yes
 FREESTANDING := no
 
 ifeq ($(FREESTANDING),yes)
-	CFLAGS += -DFREESTANDING -isystem arch/${ARCH}/freestanding
+	CFLAGS += -DFREESTANDING
 	EXPORT_UNPREFIXED = no
 endif
 
@@ -31,7 +31,7 @@ endif
 LIBUCONTEXT_C_SRC = $(wildcard arch/${ARCH}/*.c)
 LIBUCONTEXT_S_SRC = $(wildcard arch/${ARCH}/*.S)
 
-LIBUCONTEXT_VERSION := 0.13.1
+LIBUCONTEXT_VERSION := $(shell head VERSION)
 LIBUCONTEXT_OBJ = ${LIBUCONTEXT_C_SRC:.c=.o} ${LIBUCONTEXT_S_SRC:.S=.o}
 LIBUCONTEXT_SOVERSION = 0
 LIBUCONTEXT_NAME = libucontext.so
@@ -105,7 +105,7 @@ docs: ${MANPAGES}
 	$(CC) -std=gnu99 -D_BSD_SOURCE -fPIC -DPIC ${CFLAGS} ${CPPFLAGS} -c -o $@ $<
 
 .S.o:
-	$(CC) -fPIC -DPIC -DLIBUCONTEXT_ASSEMBLY ${CFLAGS} ${CPPFLAGS} -c -o $@ $<
+	$(CC) -fPIC -DPIC ${CFLAGS} ${CPPFLAGS} -c -o $@ $<
 
 ${LIBUCONTEXT_NAME}_clean:
 	rm -f ${LIBUCONTEXT_NAME}
@@ -199,13 +199,13 @@ examples/cooperative_threading: examples/cooperative_threading.c ${LIBUCONTEXT_N
 
 ifeq ($(FREESTANDING),no)
 
-include/libucontext/bits.h: arch/common/bits.h
-	cp arch/common/bits.h $@
+include/libucontext/bits.h: arch/common/include/libucontext/bits.h
+	cp $< $@
 
 else
 
-include/libucontext/bits.h: arch/${ARCH}/freestanding/bits.h
-	cp arch/${ARCH}/freestanding/bits.h $@
+include/libucontext/bits.h: arch/${ARCH}/include/libucontext/bits.h
+	cp $< $@
 
 endif
 
diff --git a/VERSION b/VERSION
new file mode 100644
index 0000000..ed0d9e9
--- /dev/null
+++ b/VERSION
@@ -0,0 +1 @@
+0.13.1
\ No newline at end of file
diff --git a/arch/aarch64/defs.h b/arch/aarch64/defs.h
index 469e69a..ee9fc7f 100644
--- a/arch/aarch64/defs.h
+++ b/arch/aarch64/defs.h
@@ -19,7 +19,7 @@
 #endif
 
 #define FETCH_LINKPTR(dest) \
-	asm("mov	x0, %0" : "=rm" ((dest)))
+	asm("mov	%0, x19" : "=r" ((dest)))
 
 #include "common-defs.h"
 
diff --git a/arch/aarch64/getcontext.S b/arch/aarch64/getcontext.S
index 98b659e..1a4911b 100644
--- a/arch/aarch64/getcontext.S
+++ b/arch/aarch64/getcontext.S
@@ -14,7 +14,11 @@
 
 ALIAS(getcontext, libucontext_getcontext)
 
-FUNC(libucontext_getcontext)
+	.global PROC_NAME(libucontext_getcontext);
+	.align  2;
+	TYPE(libucontext_getcontext)
+	ENT(libucontext_getcontext)
+PROC_NAME(libucontext_getcontext):
 	str	xzr, [x0, #REG_OFFSET(0)]
 
 	/* save GPRs */
diff --git a/arch/aarch64/freestanding/bits.h b/arch/aarch64/include/libucontext/bits.h
similarity index 58%
rename from arch/aarch64/freestanding/bits.h
rename to arch/aarch64/include/libucontext/bits.h
index 3d6e8d7..8eb04f5 100644
--- a/arch/aarch64/freestanding/bits.h
+++ b/arch/aarch64/include/libucontext/bits.h
@@ -1,20 +1,20 @@
 #ifndef LIBUCONTEXT_BITS_H
 #define LIBUCONTEXT_BITS_H
 
-typedef unsigned long libucontext_greg_t;
-typedef unsigned long libucontext_gregset_t[34];
+typedef unsigned long long int libucontext_greg_t;
+typedef unsigned long int libucontext_sigset_t;
 
 typedef struct {
 	__uint128_t vregs[32];
-	unsigned int fpsr;
-	unsigned int fpcr;
+	libucontext_greg_t fpsr;
+	libucontext_greg_t fpcr;
 } libucontext_fpregset_t;
 
-typedef struct sigcontext {
-	unsigned long fault_address;
-	unsigned long regs[31];
-	unsigned long sp, pc, pstate;
-	long double __reserved[256];
+typedef struct {
+	libucontext_greg_t fault_address;
+	libucontext_greg_t regs[31];
+	libucontext_greg_t sp, pc, pstate;
+	double __reserved[256];
 } libucontext_mcontext_t;
 
 typedef struct {
@@ -27,6 +27,7 @@ typedef struct libucontext_ucontext {
 	unsigned long uc_flags;
 	struct libucontext_ucontext *uc_link;
 	libucontext_stack_t uc_stack;
+	libucontext_sigset_t uc_sigmask;
 	unsigned char __pad[128];
 	libucontext_mcontext_t uc_mcontext;
 } libucontext_ucontext_t;
diff --git a/arch/aarch64/makecontext.c b/arch/aarch64/makecontext.c
index fb291c5..de666ee 100644
--- a/arch/aarch64/makecontext.c
+++ b/arch/aarch64/makecontext.c
@@ -10,29 +10,38 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 #include <stdio.h>
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
 
+_Static_assert(offsetof(libucontext_ucontext_t, uc_mcontext.regs[0]) == R0_OFFSET, "R0_OFFSET is invalid");
+_Static_assert(offsetof(libucontext_ucontext_t, uc_mcontext.sp) == SP_OFFSET, "SP_OFFSET is invalid");
+_Static_assert(offsetof(libucontext_ucontext_t, uc_mcontext.pc) == PC_OFFSET, "PC_OFFSET is invalid");
+_Static_assert(offsetof(libucontext_ucontext_t, uc_mcontext.pstate) == PSTATE_OFFSET, "PSTATE_OFFSET is invalid");
 
 void
 libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int argc, ...)
 {
-	unsigned long *sp;
-	unsigned long *regp;
+	// specs require only 'int' arguments so stack arguments are int type
+	// while register argument take the full width of libucontext_greg_t
+	int *sp;
+	libucontext_greg_t *regp;
 	va_list va;
 	int i;
 
-	sp = (unsigned long *) ((uintptr_t) ucp->uc_stack.ss_sp + ucp->uc_stack.ss_size);
+	sp = (int *) ((uintptr_t) ucp->uc_stack.ss_sp + ucp->uc_stack.ss_size);
 	sp -= argc < 8 ? 0 : argc - 8;
-	sp = (unsigned long *) (((uintptr_t) sp & -16L));
+	sp = (int *) (((uintptr_t) sp & -16L));
 
 	ucp->uc_mcontext.sp = (uintptr_t) sp;
 	ucp->uc_mcontext.pc = (uintptr_t) func;
@@ -44,13 +53,14 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	regp = &(ucp->uc_mcontext.regs[0]);
 
 	for (i = 0; (i < argc && i < 8); i++)
-		*regp++ = va_arg (va, unsigned long);
+		*regp++ = va_arg (va, int);
 
 	for (; i < argc; i++)
-		*sp++ = va_arg (va, unsigned long);
+		*sp++ = va_arg (va, int);
 
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/aarch64/setcontext.S b/arch/aarch64/setcontext.S
index d7c8d8c..70b9573 100644
--- a/arch/aarch64/setcontext.S
+++ b/arch/aarch64/setcontext.S
@@ -14,7 +14,11 @@
 
 ALIAS(setcontext, libucontext_setcontext)
 
-FUNC(libucontext_setcontext)
+	.global PROC_NAME(libucontext_setcontext);
+	.align  2;
+	TYPE(libucontext_setcontext)
+	ENT(libucontext_setcontext)
+PROC_NAME(libucontext_setcontext):
 	/* restore GPRs */
 	ldp	x18, x19, [x0, #REG_OFFSET(18)]
 	ldp	x20, x21, [x0, #REG_OFFSET(20)]
diff --git a/arch/aarch64/swapcontext.S b/arch/aarch64/swapcontext.S
index 17a0b99..2af525e 100644
--- a/arch/aarch64/swapcontext.S
+++ b/arch/aarch64/swapcontext.S
@@ -14,7 +14,11 @@
 
 ALIAS(swapcontext, libucontext_swapcontext)
 
-FUNC(libucontext_swapcontext)
+	.global PROC_NAME(libucontext_swapcontext);
+	.align  2;
+	TYPE(libucontext_swapcontext)
+	ENT(libucontext_swapcontext)
+PROC_NAME(libucontext_swapcontext):
 	str	xzr, [x0, #REG_OFFSET(0)]
 
 	/* save GPRs */
@@ -50,7 +54,7 @@ FUNC(libucontext_swapcontext)
 
 	/* move x1 to x0 and call setcontext */
 	mov	x0, x1
-	bl	libucontext_setcontext
+	bl	PROC_NAME(libucontext_setcontext)
 
 	/* hmm, we came back here try to return */
 	mov	x30, x28
diff --git a/arch/aarch64/trampoline.c b/arch/aarch64/trampoline.c
index 83eb453..699a050 100644
--- a/arch/aarch64/trampoline.c
+++ b/arch/aarch64/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/arm/defs.h b/arch/arm/defs.h
index 0834d9e..3b6b9ba 100644
--- a/arch/arm/defs.h
+++ b/arch/arm/defs.h
@@ -6,7 +6,7 @@
 #define TYPE(__proc)
 
 #define FETCH_LINKPTR(dest) \
-	asm("movs    r0, %0" : "=rm" ((dest)))
+	asm("movs    %0, r4" : "=r" ((dest)))
 
 #include "common-defs.h"
 
diff --git a/arch/arm/freestanding/bits.h b/arch/arm/include/libucontext/bits.h
similarity index 100%
rename from arch/arm/freestanding/bits.h
rename to arch/arm/include/libucontext/bits.h
diff --git a/arch/arm/makecontext.c b/arch/arm/makecontext.c
index 9d1ba8d..fd46df1 100644
--- a/arch/arm/makecontext.c
+++ b/arch/arm/makecontext.c
@@ -10,7 +10,9 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
@@ -19,6 +21,7 @@
 
 
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -56,5 +59,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/arm/trampoline.c b/arch/arm/trampoline.c
index 83eb453..699a050 100644
--- a/arch/arm/trampoline.c
+++ b/arch/arm/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/common/common-defs.h b/arch/common/common-defs.h
index 8a5f12d..8ba97bf 100644
--- a/arch/common/common-defs.h
+++ b/arch/common/common-defs.h
@@ -18,15 +18,27 @@
 #endif
 
 #ifndef TYPE
-# define TYPE(__proc)	.type	__proc, @function;
+# ifdef __clang__
+#  define TYPE(__proc) // .type not supported
+# else
+#  define TYPE(__proc)	.type	__proc, @function;
+#endif
+#endif
+
+#ifndef PROC_NAME
+# ifdef __MACH__
+#  define PROC_NAME(__proc) _ ## __proc
+# else
+#  define PROC_NAME(__proc) __proc
+# endif
 #endif
 
 #define FUNC(__proc)					\
-	.global __proc;					\
+	.global PROC_NAME(__proc);			\
 	.align  2;					\
 	TYPE(__proc)					\
 	ENT(__proc)					\
-__proc:							\
+PROC_NAME(__proc):					\
 	SETUP_FRAME(__proc)
 #ifdef __clang__
 #define END(__proc)
@@ -46,10 +58,4 @@ __proc:							\
 
 #define REG_OFFSET(__reg)       (MCONTEXT_GREGS + ((__reg) * REG_SZ))
 
-#ifndef LIBUCONTEXT_ASSEMBLY
-
-#include <libucontext/libucontext.h>
-
-#endif
-
 #endif
diff --git a/arch/common/common-trampoline.c b/arch/common/common-trampoline.c
index f182f99..dc5f2d3 100644
--- a/arch/common/common-trampoline.c
+++ b/arch/common/common-trampoline.c
@@ -19,6 +19,7 @@ libucontext_trampoline(void)
 {
 	libucontext_ucontext_t *uc_link;
 
+    // FIXME: there's no guarantee that input is not clobbered!
 	FETCH_LINKPTR(uc_link);
 
 	if (uc_link == NULL)
diff --git a/arch/common/bits.h b/arch/common/include/libucontext/bits.h
similarity index 100%
rename from arch/common/bits.h
rename to arch/common/include/libucontext/bits.h
diff --git a/arch/m68k/defs.h b/arch/m68k/defs.h
index 6a00651..fc19252 100644
--- a/arch/m68k/defs.h
+++ b/arch/m68k/defs.h
@@ -27,7 +27,7 @@
 #define PC_OFFSET	REG_OFFSET(REG_PC)
 
 #define FETCH_LINKPTR(dest) \
-	asm("mov.l (%%sp, %%d7.l * 4), %0" : "=rm" ((dest)))
+	asm("mov.l (%%sp, %%d7.l * 4), %0" :: "r" ((dest)))
 
 #include "common-defs.h"
 
diff --git a/arch/m68k/freestanding/bits.h b/arch/m68k/include/libucontext/bits.h
similarity index 100%
rename from arch/m68k/freestanding/bits.h
rename to arch/m68k/include/libucontext/bits.h
diff --git a/arch/m68k/makecontext.c b/arch/m68k/makecontext.c
index 2cbc0f7..fb380f1 100644
--- a/arch/m68k/makecontext.c
+++ b/arch/m68k/makecontext.c
@@ -10,13 +10,16 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 #include <stdio.h>
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -55,5 +58,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	*sp++ = (libucontext_greg_t) ucp->uc_link;
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/m68k/trampoline.c b/arch/m68k/trampoline.c
index 83eb453..699a050 100644
--- a/arch/m68k/trampoline.c
+++ b/arch/m68k/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/mips/freestanding/bits.h b/arch/mips/include/libucontext/bits.h
similarity index 100%
rename from arch/mips/freestanding/bits.h
rename to arch/mips/include/libucontext/bits.h
diff --git a/arch/mips64/freestanding/bits.h b/arch/mips64/include/libucontext/bits.h
similarity index 100%
rename from arch/mips64/freestanding/bits.h
rename to arch/mips64/include/libucontext/bits.h
diff --git a/arch/ppc/makecontext.c b/arch/ppc/makecontext.c
index cc2a882..f8ed384 100644
--- a/arch/ppc/makecontext.c
+++ b/arch/ppc/makecontext.c
@@ -11,12 +11,15 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stdarg.h>
 #include <stdint.h>
 
 
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -55,5 +58,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(), int argc, .
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/ppc64/makecontext.c b/arch/ppc64/makecontext.c
index 9a9d8b2..e42d66a 100644
--- a/arch/ppc64/makecontext.c
+++ b/arch/ppc64/makecontext.c
@@ -11,12 +11,15 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stdarg.h>
 #include <stdint.h>
 
 
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -56,5 +59,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(), int argc, .
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/riscv64/defs.h b/arch/riscv64/defs.h
index 191e8c2..d9b7647 100644
--- a/arch/riscv64/defs.h
+++ b/arch/riscv64/defs.h
@@ -48,7 +48,7 @@
 #define PC_OFFSET	REG_OFFSET(REG_PC)
 
 #define FETCH_LINKPTR(dest) \
-	asm("mv	%0, s1" : "=rm" ((dest)))
+	asm("mv	%0, s1" : "=r" ((dest)))
 
 #include "common-defs.h"
 
diff --git a/arch/riscv64/freestanding/bits.h b/arch/riscv64/include/libucontext/bits.h
similarity index 100%
rename from arch/riscv64/freestanding/bits.h
rename to arch/riscv64/include/libucontext/bits.h
diff --git a/arch/riscv64/makecontext.c b/arch/riscv64/makecontext.c
index d8609f8..aee671c 100644
--- a/arch/riscv64/makecontext.c
+++ b/arch/riscv64/makecontext.c
@@ -10,13 +10,16 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 #include <stdio.h>
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -56,5 +59,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/riscv64/trampoline.c b/arch/riscv64/trampoline.c
index 83eb453..699a050 100644
--- a/arch/riscv64/trampoline.c
+++ b/arch/riscv64/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/s390x/freestanding/bits.h b/arch/s390x/include/libucontext/bits.h
similarity index 100%
rename from arch/s390x/freestanding/bits.h
rename to arch/s390x/include/libucontext/bits.h
diff --git a/arch/s390x/makecontext.c b/arch/s390x/makecontext.c
index e74c6f3..f970da4 100644
--- a/arch/s390x/makecontext.c
+++ b/arch/s390x/makecontext.c
@@ -10,7 +10,9 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
@@ -18,6 +20,7 @@
 
 
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -62,5 +65,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	ucp->uc_mcontext.gregs[15] = (uintptr_t) sp;
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/sh/defs.h b/arch/sh/defs.h
index dd13d47..7cddaef 100644
--- a/arch/sh/defs.h
+++ b/arch/sh/defs.h
@@ -13,7 +13,7 @@
 #define REG_MACL		(21)
 
 #define FETCH_LINKPTR(dest)	\
-	asm("mov r8, %0" : "=rm" (dest));
+	asm("mov r8, %0" : "=r" (dest));
 
 #include "common-defs.h"
 
diff --git a/arch/sh/freestanding/bits.h b/arch/sh/include/libucontext/bits.h
similarity index 100%
rename from arch/sh/freestanding/bits.h
rename to arch/sh/include/libucontext/bits.h
diff --git a/arch/sh/makecontext.c b/arch/sh/makecontext.c
index ac0c2b7..9bc4782 100644
--- a/arch/sh/makecontext.c
+++ b/arch/sh/makecontext.c
@@ -10,13 +10,16 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 #include <stdio.h>
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -53,5 +56,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/sh/trampoline.c b/arch/sh/trampoline.c
index 83eb453..699a050 100644
--- a/arch/sh/trampoline.c
+++ b/arch/sh/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/x86/defs.h b/arch/x86/defs.h
index 2b0c68d..0bf73ba 100644
--- a/arch/x86/defs.h
+++ b/arch/x86/defs.h
@@ -60,7 +60,7 @@
 #define MCONTEXT_GREGS	(20)
 
 #define FETCH_LINKPTR(dest) \
-	asm("movl (%%esp, %%ebx, 4), %0" : "=rm" ((dest)));
+	asm("movl (%%esp, %%ebx, 4), %0" : "=r" ((dest)));
 
 #include "common-defs.h"
 
diff --git a/arch/x86/freestanding/bits.h b/arch/x86/include/libucontext/bits.h
similarity index 100%
rename from arch/x86/freestanding/bits.h
rename to arch/x86/include/libucontext/bits.h
diff --git a/arch/x86/makecontext.c b/arch/x86/makecontext.c
index fc292f9..8d7fbf5 100644
--- a/arch/x86/makecontext.c
+++ b/arch/x86/makecontext.c
@@ -11,12 +11,15 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 
 extern void libucontext_trampoline(void);
@@ -53,5 +56,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	*argp++ = (uintptr_t) ucp->uc_link;
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/x86/trampoline.c b/arch/x86/trampoline.c
index 83eb453..699a050 100644
--- a/arch/x86/trampoline.c
+++ b/arch/x86/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/arch/x86_64/defs.h b/arch/x86_64/defs.h
index 4a53f96..c5fd126 100644
--- a/arch/x86_64/defs.h
+++ b/arch/x86_64/defs.h
@@ -32,7 +32,7 @@
 #define REG_SZ		(8)
 
 #define FETCH_LINKPTR(dest) \
-	asm("movq (%%rbx), %0" : "=rm" ((dest)));
+	asm("movq (%%rbx), %0" : "=r" ((dest)));
 
 #include "common-defs.h"
 
diff --git a/arch/x86_64/freestanding/bits.h b/arch/x86_64/include/libucontext/bits.h
similarity index 100%
rename from arch/x86_64/freestanding/bits.h
rename to arch/x86_64/include/libucontext/bits.h
diff --git a/arch/x86_64/makecontext.c b/arch/x86_64/makecontext.c
index 0550ff6..66b809e 100644
--- a/arch/x86_64/makecontext.c
+++ b/arch/x86_64/makecontext.c
@@ -10,13 +10,16 @@
  * from the use of this software.
  */
 
+#ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#endif
 #include <stddef.h>
 #include <stdarg.h>
 #include <string.h>
 #include <stdint.h>
 
 #include "defs.h"
+#include <libucontext/libucontext.h>
 
 extern void libucontext_trampoline(void);
 
@@ -73,5 +76,6 @@ libucontext_makecontext(libucontext_ucontext_t *ucp, void (*func)(void), int arg
 	va_end(va);
 }
 
-
+#ifdef EXPORT_UNPREFIXED
 extern __typeof(libucontext_makecontext) makecontext __attribute__((weak, __alias__("libucontext_makecontext")));
+#endif
diff --git a/arch/x86_64/trampoline.c b/arch/x86_64/trampoline.c
index 83eb453..699a050 100644
--- a/arch/x86_64/trampoline.c
+++ b/arch/x86_64/trampoline.c
@@ -1,2 +1,3 @@
 #include "defs.h"
+#include <libucontext/libucontext.h>
 #include "common-trampoline.c"
diff --git a/include/libucontext/libucontext.h b/include/libucontext/libucontext.h
index 8207126..037da0e 100644
--- a/include/libucontext/libucontext.h
+++ b/include/libucontext/libucontext.h
@@ -1,6 +1,7 @@
 #ifndef LIBUCONTEXT_LIBUCONTEXT_H
 #define LIBUCONTEXT_LIBUCONTEXT_H
 
+#include <stddef.h>
 #include <libucontext/bits.h>
 
 int  libucontext_getcontext(libucontext_ucontext_t *);
diff --git a/meson.build b/meson.build
new file mode 100644
index 0000000..e72dfa3
--- /dev/null
+++ b/meson.build
@@ -0,0 +1,175 @@
+project(
+  'libucontext',
+  'c',
+  meson_version : '>=0.55.0',
+  version : run_command('head', files('VERSION')).stdout()
+)
+
+cpu = host_machine.cpu_family()
+if cpu == 'sh4'
+  cpu = 'sh'
+endif
+
+project_description = 'Portable implementation of ucontext'
+
+project_headers = [
+  'include/libucontext/libucontext.h'
+]
+
+project_source_files = [
+  'arch' / cpu / 'getcontext.S',
+  'arch' / cpu / 'setcontext.S',
+  'arch' / cpu / 'swapcontext.S',
+]
+if cpu in ['mips', 'mips64']
+  project_source_files += [
+    'arch' / cpu / 'makecontext.S'
+  ]
+else
+  project_source_files += [
+    'arch' / cpu / 'makecontext.c'
+  ]
+endif
+if cpu in ['ppc', 'ppc64']
+  project_source_files += [
+    'arch' / cpu / 'retfromsyscall.c'
+  ]
+endif
+if cpu not in ['mips', 'mips64', 'ppc', 'ppc64', 's390x']
+  project_source_files += [
+    'arch' / cpu / 'trampoline.c'
+  ]
+endif
+
+project_includes = [
+  'include',
+  'arch/common'
+]
+
+build_args = [
+  '-std=gnu99',
+  '-D_BSD_SOURCE',
+  '-fPIC',
+  '-DPIC'
+]
+
+
+# ===================================================================
+
+# ======
+# Options
+# ======
+
+freestanding = get_option('freestanding')
+export_unprefixed = get_option('export_unprefixed')
+build_posix = true
+
+if freestanding
+  build_args += '-DFREESTANDING'
+  build_posix = false
+  export_unprefixed = false
+  project_headers += ['arch' / cpu / 'include/libucontext/bits.h']
+  project_includes += ['arch' / cpu / 'include']
+else
+  project_headers += ['arch/common/include/libucontext/bits.h']
+  project_includes += ['arch/common/include']
+endif
+
+if export_unprefixed
+  build_args += '-DEXPORT_UNPREFIXED'
+endif
+
+# ======
+# Target
+# ======
+
+headers = include_directories(project_includes)
+is_subproject = meson.is_subproject()
+# Build only static library if subproject
+
+libucontext_target = both_libraries(
+  'ucontext',
+  project_source_files,
+  version: meson.project_version(),
+  install : not is_subproject,
+  c_args : build_args,
+  include_directories : headers,
+)
+if is_subproject
+  libucontext_target = libucontext_target.get_static_lib()
+endif
+libucontext_dep = declare_dependency(
+  include_directories: headers,
+  link_with : libucontext_target
+)
+
+if build_posix
+  libucontext_posix_target = both_libraries(
+    'ucontext_posix',
+    project_source_files + ['libucontext_posix.c'],
+    version: meson.project_version(),
+    install : not is_subproject,
+    c_args : build_args,
+    include_directories : headers,
+  )
+  if is_subproject
+    libucontext_posix_target = libucontext_posix_target.get_static_lib()
+  endif
+  libucontext_posix_dep = declare_dependency(
+    include_directories: headers,
+    link_with : libucontext_posix_target
+  )
+endif
+
+# =======
+# Project
+# =======
+
+if not is_subproject
+  # Make this library usable from the system's
+  # package manager.
+  install_headers(project_headers, subdir : meson.project_name())
+
+  pkg_mod = import('pkgconfig')
+  pkg_mod.generate(
+    name : meson.project_name(),
+    filebase : meson.project_name(),
+    description : project_description,
+    subdirs : meson.project_name(),
+    libraries : libucontext_target,
+  )
+endif
+
+# ====
+# Docs
+# ====
+
+# TODO: meson.build for docs
+if not is_subproject
+  #subdir('docs')
+endif
+
+# ==========
+# Unit Tests
+# ==========
+
+if not is_subproject
+  test('test_libucontext',
+    executable(
+      'test_libucontext',
+      files('test_libucontext.c'),
+      dependencies : libucontext_dep,
+      install : false
+    )
+  )
+  if build_posix
+    test('test_libucontext_posix',
+      executable(
+        'test_libucontext_posix',
+        files('test_libucontext_posix.c'),
+        dependencies : libucontext_posix_dep,
+        install : false
+      )
+    )
+  endif
+endif
diff --git a/meson_options.txt b/meson_options.txt
new file mode 100644
index 0000000..d4201d1
--- /dev/null
+++ b/meson_options.txt
@@ -0,0 +1,4 @@
+option('freestanding', type : 'boolean', value : false,
+       description: 'Do not use system headers')
+option('export_unprefixed', type : 'boolean', value : true,
+       description: 'Export POSIX 2004 ucontext names as alises')
\ No newline at end of file