diff --git a/CMakeLists.txt b/CMakeLists.txt
index 719a017..faa8520 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -275,6 +275,19 @@ if (LIBDISPATCH)
 	target_link_libraries(objc ${LIBDISPATCH})
 endif ()
 
+# Enable .type and .section assembly directives on platforms other than OS X
+if (NOT APPLE)
+	add_definitions(-DHAS_SECTION_DIRECTIVE=1 -DHAS_TYPE_DIRECTIVE=1)
+endif ()
+
+# Make weak symbols work on OS X
+if (APPLE)
+	set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS
+		"${CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS} -undefined dynamic_lookup")
+	set(CMAKE_C_LINK_FLAGS "${CMAKE_C_LINK_FLAGS} -Wl,-undefined,dynamic_lookup")
+	set(CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS} -Wl,-undefined,dynamic_lookup")
+endif ()
+
 #
 # Installation
 #
diff --git a/block_trampolines.S b/block_trampolines.S
index 879e9a7..2955787 100644
--- a/block_trampolines.S
+++ b/block_trampolines.S
@@ -1,3 +1,5 @@
+#include "common.S"
+
 #
 # This file defines some trampolines for calling blocks.  A block function
 # looks like this:
@@ -15,33 +17,33 @@
 .file	"block_trampolines.S"
 #if __arm__
 .syntax unified
-.globl __objc_block_trampoline_sret
-	.type	__objc_block_trampoline_sret, %function
-.globl __objc_block_trampoline_end_sret
-.globl __objc_block_trampoline
-	.type	__objc_block_trampoline, %function
-.globl __objc_block_trampoline_end
+.globl CDECL(__objc_block_trampoline_sret)
+TYPE_DIRECTIVE(CDECL(__objc_block_trampoline_sret), %function)
+.globl CDECL(__objc_block_trampoline_end_sret)
+.globl CDECL(__objc_block_trampoline)
+TYPE_DIRECTIVE(CDECL(__objc_block_trampoline), %function)
+.globl CDECL(__objc_block_trampoline_end)
 #else
-.globl __objc_block_trampoline_sret
-	.type	__objc_block_trampoline_sret, @function
-.globl __objc_block_trampoline_end_sret
-.globl __objc_block_trampoline
-	.type	__objc_block_trampoline, @function
-.globl __objc_block_trampoline_end
+.globl CDECL(__objc_block_trampoline_sret)
+TYPE_DIRECTIVE(CDECL(__objc_block_trampoline_sret), @function)
+.globl CDECL(__objc_block_trampoline_end_sret)
+.globl CDECL(__objc_block_trampoline)
+TYPE_DIRECTIVE(CDECL(__objc_block_trampoline), @function)
+.globl CDECL(__objc_block_trampoline_end)
 #endif
 #if __x86_64
-__objc_block_trampoline:
+CDECL(__objc_block_trampoline):
 	mov   -15(%rip), %rsi      # Load the block pointer into the second argument
 	xchg  %rdi, %rsi           # Swap the first and second arguments
 	jmp   *-32(%rip)           # Call the block function
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
 	mov   -15(%rip), %rdx      # Load the block pointer into the second argument
 	xchg  %rdx, %rsi           # Swap the first and second arguments
 	jmp   *-32(%rip)           # Call the block function
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline_end_sret):
 #elif __i386
-__objc_block_trampoline:
+CDECL(__objc_block_trampoline):
 	call  next_line            # Store the instruction pointer on the stack
 next_line:
 	pop   %eax                 # Load the old instruction pointer
@@ -50,8 +52,8 @@ next_line:
 	mov   -9(%eax), %ebx       # Load the block pointer to %ebx
 	mov   %ebx, 4(%esp)        # Store the block pointer in the first argument
 	jmp   *-13(%eax)           # Call the block function
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
 	call  next_line2           # Store the instruction pointer on the stack
 next_line2:
 	pop   %eax                 # Load the old instruction pointer
@@ -60,51 +62,53 @@ next_line2:
 	mov   -9(%eax), %ebx       # Load the block pointer to %ebx
 	mov   %ebx, 8(%esp)        # Store the block pointer in the first argument
 	jmp   *-13(%eax)           # Call the block function
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline_end_sret):
 #elif __mips__
 #	ifdef _ABI64
-__objc_block_trampoline:
+CDECL(__objc_block_trampoline):
 	move     $a1, $a0
 	ld       $a0, -16($25)
 	ld       $25, -8($25)
 	jr       $25
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
 	move     $a2, $a1
 	ld       $a1, -16($25)
 	ld       $25, -8($25)
 	jr       $25
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline_end_sret):
 #	else
-__objc_block_trampoline:
+CDECL(__objc_block_trampoline):
 	move   $a1, $a0
 	lw     $a0, -8($25)
 	lw     $25, -4($25)
 	jr     $25
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
 	move   $a2, $a1
 	lw     $a1, -8($25)
 	lw     $25, -4($25)
 	jr     $25
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline_end_sret):
 #	endif
 #elif __arm__
-__objc_block_trampoline:
+CDECL(__objc_block_trampoline):
 	mov r1, r0                 // Move self over _cmd
 	ldr r0, [pc, #-16]         // Load the block pointer over self
 	ldr pc, [pc, #-24]         // Jump to the block function
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
 	mov r2, r1                 // Move self over _cmd
 	ldr r1, [pc, #-16]         // Load the block pointer over self
 	ldr pc, [pc, #-24]         // Jump to the block function
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline_end_sret):
 #else
 #warning imp_implementationWithBlock() not implemented for your architecture
-__objc_block_trampoline:
-__objc_block_trampoline_end:
-__objc_block_trampoline_sret:
-__objc_block_trampoline_end_sret:
+CDECL(__objc_block_trampoline):
+CDECL(__objc_block_trampoline_end):
+CDECL(__objc_block_trampoline_sret):
+CDECL(__objc_block_trampoline_end_sret):
 #endif
+#if HAS_SECTION_DIRECTIVE
 .section .note.GNU-stack,"",%progbits
+#endif
diff --git a/common.S b/common.S
new file mode 100644
index 0000000..521e65d
--- /dev/null
+++ b/common.S
@@ -0,0 +1,12 @@
+#if defined(__WIN32__) || defined(__APPLE__)
+#define CDECL(symbol) _##symbol
+#else
+#define CDECL(symbol) symbol
+#endif
+
+#if HAS_TYPE_DIRECTIVE
+#define TYPE_DIRECTIVE(symbol, symboltype) .type symbol, symboltype
+#else
+#define TYPE_DIRECTIVE(symbol, symboltype)
+#endif
+
diff --git a/objc_msgSend.S b/objc_msgSend.S
index 8f3a004..3cdf15f 100644
--- a/objc_msgSend.S
+++ b/objc_msgSend.S
@@ -1,3 +1,4 @@
+#include "common.S"
 #if __x86_64
 #include "objc_msgSend.x86-64.S"
 #elif __i386
@@ -9,4 +10,6 @@
 #else
 #warning objc_msgSend() not implemented for your architecture
 #endif
+#if HAS_SECTION_DIRECTIVE
 .section .note.GNU-stack,"",%progbits
+#endif
diff --git a/objc_msgSend.arm.S b/objc_msgSend.arm.S
index 9ec7f4b..d0cd25d 100644
--- a/objc_msgSend.arm.S
+++ b/objc_msgSend.arm.S
@@ -77,7 +77,7 @@
 	.save  {\receiver}
 	mov    r1, \sel
 
-	bl     slowMsgLookup(PLT)             // This is the only place where the CFI directives have to be accurate...
+	bl     CDECL(slowMsgLookup)(PLT)             // This is the only place where the CFI directives have to be accurate...
 	mov    ip, r0                         // IMP -> ip
 
 	pop    {r5}                           // restore (modified) self to r5
@@ -87,16 +87,16 @@
 	.fnend
 .endm
 
-.globl objc_msgSend_fpret
-	.type	objc_msgSend_fpret, %function
-.globl objc_msgSend
-	.type	objc_msgSend, %function
-objc_msgSend:
-objc_msgSend_fpret:
+.globl CDECL(objc_msgSend_fpret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_fpret), %function)
+.globl CDECL(objc_msgSend)
+TYPE_DIRECTIVE(CDECL(objc_msgSend), %function)
+CDECL(objc_msgSend):
+CDECL(objc_msgSend_fpret):
 	MSGSEND r0, r1
-.globl objc_msgSend_stret
-	.type	objc_msgSend_stret, %function
-objc_msgSend_stret:
+.globl CDECL(objc_msgSend_stret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_stret), %function)
+CDECL(objc_msgSend_stret):
 	MSGSEND r1, r2
 
 LSmallIntClass:
diff --git a/objc_msgSend.mips.S b/objc_msgSend.mips.S
index 1aafc12..c33baaa 100644
--- a/objc_msgSend.mips.S
+++ b/objc_msgSend.mips.S
@@ -127,7 +127,7 @@ lw     $zero, ($zero)
 	# MIPS implementations.
 	# Note: A better linker ought to be able to turn this into a single
 	# jump-immediate, so revisit this decision later...
-	LP     $25, %got_disp(slowMsgLookup)($t8) 
+	LP     $25, %got_disp(CDECL(slowMsgLookup))($t8) 
 
 	daddiu $sp, $sp, -SAVE_SIZE             # We need to preserve all registers that may contain arguments:
 
@@ -193,24 +193,24 @@ lw     $zero, ($zero)
 6:                                          # smallObject:
 #if _ABI64
 	dsll   $t4, $t4, 3                      # Convert tag to pointer offset
-	LP     $t6, %got_disp(SmallObjectClasses)($t8) # Load small object classes array address
+	LP     $t6, %got_disp(CDECL(SmallObjectClasses))($t8) # Load small object classes array address
 	daddu  $t4, $t4, $t6                    # Add the base address to the offset
 	b      1b                               # Return to the normal path
 	LP     $t4, ($t4)                       # Load the class (in delay slot)
 #else
 	b      1b
-	LP     $t4, %got_disp(SmallIntClass)($t8)
+	LP     $t4, %got_disp(CDECL(SmallIntClass))($t8)
 #endif
 	.cfi_endproc
 .endm
-.globl objc_msgSend
-	.type	objc_msgSend, @function
-.globl objc_msgSend_fpret
-	.type	objc_msgSend_fpret, @function
-objc_msgSend_fpret:
-objc_msgSend:
+.globl CDECL(objc_msgSend)
+TYPE_DIRECTIVE(CDECL(objc_msgSend), @function)
+.globl CDECL(objc_msgSend_fpret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_fpret), @function)
+CDECL(objc_msgSend_fpret):
+CDECL(objc_msgSend):
 	MSGSEND $a0, $a1
-.globl objc_msgSend_stret
-	.type	objc_msgSend_stret, @function
-objc_msgSend_stret:
+.globl CDECL(objc_msgSend_stret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_stret), @function)
+CDECL(objc_msgSend_stret):
 	MSGSEND $a1, $a2
diff --git a/objc_msgSend.x86-32.S b/objc_msgSend.x86-32.S
index 8e9f8f6..2b5406f 100644
--- a/objc_msgSend.x86-32.S
+++ b/objc_msgSend.x86-32.S
@@ -87,15 +87,15 @@
 	jmp   1b 
 	.cfi_endproc
 .endm
-.globl objc_msgSend_fpret
-	.type	objc_msgSend_fpret, @function
-objc_msgSend_fpret:
+.globl CDECL(objc_msgSend_fpret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_fpret), @function)
+CDECL(objc_msgSend_fpret):
 	MSGSEND 4, 8, 1
-.globl objc_msgSend
-	.type	objc_msgSend, @function
-objc_msgSend:
+.globl CDECL(objc_msgSend)
+TYPE_DIRECTIVE(CDECL(objc_msgSend), @function)
+CDECL(objc_msgSend):
 	MSGSEND 4, 8, 0
-.globl objc_msgSend_stret
-	.type	objc_msgSend_stret, @function
-objc_msgSend_stret:
+.globl CDECL(objc_msgSend_stret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_stret), @function)
+CDECL(objc_msgSend_stret):
 	MSGSEND 8, 12, 0
diff --git a/objc_msgSend.x86-64.S b/objc_msgSend.x86-64.S
index f519bf3..4be831b 100644
--- a/objc_msgSend.x86-64.S
+++ b/objc_msgSend.x86-64.S
@@ -234,7 +234,7 @@
 .endif
 
 	.cfi_adjust_cfa_offset 0xD8
-	call  slowMsgLookup                  # Call the slow lookup function
+	call  CDECL(slowMsgLookup)           # Call the slow lookup function
 	mov   %rax, %r10                     # Load the returned IMP
 
 	pop   %rdx
@@ -260,20 +260,20 @@
 6:                                        # smallObject:
 	and   \receiver, %r10                 # Find the small int type
 	shll  $3, %r10d
-	lea   SmallObjectClasses(%rip), %r11
+	lea   CDECL(SmallObjectClasses)(%rip), %r11
 	add   %r11, %r10
 	mov   (%r10), %r10
 	jmp   1b 
 	.cfi_endproc
 .endm
-.globl objc_msgSend
-	.type	objc_msgSend, @function
-.globl objc_msgSend_fpret
-	.type	objc_msgSend_fpret, @function
-objc_msgSend_fpret:
-objc_msgSend:
+.globl CDECL(objc_msgSend)
+TYPE_DIRECTIVE(CDECL(objc_msgSend), @function)
+.globl CDECL(objc_msgSend_fpret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_fpret), @function)
+CDECL(objc_msgSend_fpret):
+CDECL(objc_msgSend):
 	MSGSEND %rdi, %rsi
-.globl objc_msgSend_stret
-	.type	objc_msgSend_stret, @function
-objc_msgSend_stret:
+.globl CDECL(objc_msgSend_stret)
+TYPE_DIRECTIVE(CDECL(objc_msgSend_stret), @function)
+CDECL(objc_msgSend_stret):
 	MSGSEND %rsi, %rdx