; A minimal Mach-o x64 executable for OS X (also see below Mountain Lion version)
;
; $ nasm -f bin -o tiny_hello tiny_hello.s
; $ chmod +x tiny_hello
; $ ./tiny_hello
; Hello World!
; $

; c.f.
; http://osxbook.com/blog/2009/03/15/crafting-a-tiny-mach-o-executable/ ( the original tiny mach-o executable )
; http://feiri.de/macho/												( improved version, using a very clever trick to put code inside the load command itself)

; Constants (For readability)

%define MH_MAGIC_64					0xfeedfacf

%define CPU_ARCH_ABI64				0x01000000
%define	CPU_TYPE_I386				0x00000007
%define CPU_TYPE_X86_64				CPU_ARCH_ABI64 | CPU_TYPE_I386

%define CPU_SUBTYPE_LIB64			0x80000000
%define CPU_SUBTYPE_I386_ALL		0x00000003

%define MH_EXECUTE					0x2
%define MH_NOUNDEFS					0x1

%define LC_SEGMENT_64				0x19
%define LC_UNIXTHREAD				0x5

%define VM_PROT_READ				0x1
%define VM_PROT_WRITE				0x2
%define VM_PROT_EXECUTE				0x4

%define x86_THREAD_STATE64			0x4
%define	x86_EXCEPTION_STATE64_COUNT	42


%define SYSCALL_CLASS_SHIFT			24
%define SYSCALL_CLASS_MASK			(0xFF << SYSCALL_CLASS_SHIFT)
%define SYSCALL_NUMBER_MASK			(~SYSCALL_CLASS_MASK)

%define SYSCALL_CLASS_UNIX			2

%define SYSCALL_CONSTRUCT_UNIX(syscall_number) \
			((SYSCALL_CLASS_UNIX << SYSCALL_CLASS_SHIFT) | \
			 (SYSCALL_NUMBER_MASK & (syscall_number)))

%define SYS_exit					1
%define SYS_write					4

; NASM directive, not compiled
; Use RIP-Relative addressing for x64
BITS	64
DEFAULT	REL

; Mach-O header
DD		MH_MAGIC_64										; magic
DD		CPU_TYPE_X86_64									; cputype
DD		CPU_SUBTYPE_LIB64 | CPU_SUBTYPE_I386_ALL		; cpusubtype
DD		MH_EXECUTE										; filetype
DD		2												; ncmds
DD		end - load_commands								; sizeofcmds
DD		MH_NOUNDEFS										; flags
DD		0x0												; reserved

; Load commands
load_commands:

; Segment and Sections
DD		LC_SEGMENT_64									; cmd
DD		72												; cmdsize

; store our message in segname
hello_str:
DB		'Hello, World!', 0x0A							; segname
len		EQU		$-hello_str
DB		0x00, 0x00										; last 2-byte of segname

DQ		0x0												; vmaddr
DQ		0x1000											; vmsize
DQ		0												; fileoff
DQ		filesize										; filesize
DD		VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE	; maxprot
DD		VM_PROT_READ | VM_PROT_EXECUTE					; initprot
DD		0x0												; nsects
DD		0x0												; flags

; UNIX Thread Status
DD		LC_UNIXTHREAD									; cmd
DD		184												; cmdsize
DD		x86_THREAD_STATE64								; flavor
DD		x86_EXCEPTION_STATE64_COUNT						; count

; The following should be rax-rdx ( 8 * 4 = 32 bytes)
; Somehow the OS X loader will clear rax-rdx regardless of the initial state
; Thus we can use these 32 bytes to store our code
; The original code is 35 bytes, thus we split the syscalls, store the 2nd part below and uses a jmp
; 
; note: depending on NASM versions, the generated code may vary in size and the padding length has to be adjusted
; the NASM shipped with Xcode 4.0 on Lion generates 64 bit movs ("movq") and a short jump (24 bytes)
; nasm is broken in Mountain Lion and Xcode 4.3 (only i386 binary shipped, however ML is 64-bit only)
; NASM 2.10.04 built from nasm.us by default enables optimization to generate 32 bit movs ("movl") hence more bytes needs to be padded
; however when -O0 is specified, NASM 2.10 will generate movqs but with long jump (27 bytes), hence only 5, instead of 8 bytes needs to be padded
;
start:													; 24 bytes
	; rdi and rsi have already been set in the initial state
	mov		rdx, len
	mov		rax, SYSCALL_CONSTRUCT_UNIX(SYS_write)
	syscall
	jmp		next

DB		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00	; 8 bytes padding
DQ		0x01, hello_str, 0x00, 0x00						; rdi = STDOUT, rsi = address of hello_str,  rbp, rsp

; The exit syscall is 15 bytes, we use the inital sate of r8 and r9 to store the code
next:													; 15 bytes
	xor			rdi, rdi
	mov			rax, SYSCALL_CONSTRUCT_UNIX(SYS_exit)
	syscall
DB		0x00											; 1 byte padding

DQ		0x00, 0x00, 0x00, 0x00, 0x00, 0x00				; r10, r11, r12, r13, r14, r15
DQ		start, 0x00, 0x00, 0x00, 0x00					; rip, rflags, cs, fs, gs

end:
	filesize	EQU	$-$$