Mach-O: what changed

Old	New		Line
1	1		(macOS only)
2	2
3	3		Mach-O, short for Mach object file format, is a file format for executables, object code, shared libraries, dynamically-loaded code, and core dumps. A derivation of the a.out format, Mach-O offered more extensibility and faster access to information in the symbol table. Source: [Wikipedia](http://en.wikipedia.org/wiki/Mach-O).
4	4
5	5		The Apple file format reference can be found [here](https://developer.apple.com/library/archive/documentation/Performance/Conceptual/CodeFootprint/Articles/MachOOverview.html)
6	6
7	7		The following tools are used in macOS to view Mach-O files:
8	8
9	9		otool - object file displaying tool
10	10
11	11		nm - display name list (symbol table)
12	12
13	13		## Objective-C segment
14	14
15	15		There is no documentation about __OBJC segment and its sections. The following information has been gathered from cctools [sources](http://opensource.apple.com/source/cctools/cctools-667.3/)
16	16
17	17		### Structures strings
18	18
19	19		Some structures in sections contain name pointers. These names are stored in the c-strings section (segment: __TEXT; section: __cstring). The file offset for the string name can be evaluated in the following way:
20	20
21	21		`string_file_offset := cstr_section.offset + (name_addr - cstr_section.addr);`
22	22
23	23		## Sections
24	24
25	25		### __image_info
26	26
27	27		The section contains only the image info information:
28	28
29	29		`imageInfo = packed record`\
30	30		` version : uint32_t; // zero`\
31	31		` flags : uint32_t; // for objc 1.0 - zero`\
32	32		`end;`
33	33
34	34		Flags values:
35	35
36	36		`ImageInfo_F_and_C = $01;`\
37	37		`ImageInfo_GC = $02;`\
38	38		`ImageInfo_GC_only = $04;`
39	39
40	40		### __module_info
41	41
42	42		(objc_module record is declared at objc headers).
43	43
44	44		The number of objc_module structures depends on the number of .m files with objects declarations compiled. _symtab contains the number of classes and categories declared in the module.
45	45
46	46		`objc_module = packed record`\
47	47		` version : culong; // version number = 7 `\
48	48		` size : culong; // sizeof(objc_module)?`\
49	49		` name : PChar; // virtual memory address of the module name`\
50	50		` // Usually mapped to NULL string`\
51	51		` _symtab : Symtab; // virtual memory address of a proper objc_symtab structure (in __symbols section)`\
52	52		`end;`
53	53
54	54		### __symbols
55	55
56	56		the section contains symbol table for a module. (symtable record is declared at objc headers)
57	57
58	58		` objc_symtab = record`\
59	59		` sel_ref_cnt : culong; // zero`\
60	60		` refs : PSEL; // zero`\
61	61		` cls_def_cnt : cushort; // number of declared classes in the module`\
62	62		` cat_def_cnt : cushort; // number of declared categories`\
63	63		` defs: array [0..cls_def_cnt+cat_def_cnt-1] of Pointer; // array of virtual address of declarations`\
64	64		` // 0..cls_def_cnt - 1 : virtual address of class declarations (in "__class" section, can be empty)`\
65	65		` // cls_def_cnt..cat_def_cnt-1 : virtual address of categories declarations (can be empty)`\
66	66		` end;`
67	67
68	68		### __class, __meta_class
69	69
70	70		the section may omit, if none of modules declare any custom classes
71	71
72	72		Both sections use identical structure objc_class. (objc_class record is declared at objc header) Objective-C classes are declared in pair with their meta_classes.
73	73
74	74		`objc_class = record`\
75	75		` isa : PChar; // for class declaration: virtual address of meta-class declaration`\
76	76		` // for meta-class declaration: virtual address of "NSObject" string?`\
77		-	\
	77	+
78	78		` super_class : PChar; // contains the pointer to super_class name`\
79	79		` name : PChar; // pointer to class name `\
80	80		` version : PChar; // = 0 (for obj-c version 1?)`\
81		-	\
	81	+
82	82		` info : culong; // CLS_CLASS for classes`\
83	83		` // CLS_META for meta-classes `\
84		-	\
	84	+
85	85		` instance_size : culong; // size of the instance:`\
86	86		` // class: size of the class instance `\
87	87		` // meta-class: 48 bytes `\
88	88		` `\
89	89		` ivars : Pobjc_ivar_list; // virtual address of objc_ivar_list (stored in "__instance_vars" section)`\
90	90		` // meta-classes don't have ivars list (=0)`\
91		-	\
	91	+
92	92		` methodLists : PPobjc_method_list; // virtual address of objc_method_list`\
93	93		` // class declaration has instance methods list (stored in "__inst_meth" section)`\
94	94		` // meta-class declaration has class methods list (stored in "__cls_meth" section)`\
95	95		` `\
96	96		` cache : Pobjc_cache; // zero`\
97	97		` protocols : Pobjc_protocol_list; // pointer to protocols list. (stored in "__cat_cls_meth" section)`\
98	98		`end;`
99	99
100	100		### __instance_vars
101	101
102	102		the section is optional and may omit if none of classes declares instance variables
103	103
104	104		The section consists of number objc_ivar_list structutres. The number of structures depends upon number of declared classes using instance variables. (objc_ivar_list is declared in objc header)
105	105
106	106		`objc_ivar_list = record`\
107	107		` ivar_count : cint; // number of variables in the list`\
108	108		` ivar_list : array[0..ivar_count-1] of objc_ivar; variable length structure }`\
109	109		`end;`
110	110
111	111		`objc_ivar = record`\
112	112		` ivar_name : PChar; // vm addr of the variable name`\
113	113		` ivar_type : PChar; // vm addr of obj-c variable encoded type `\
114	114		` ivar_offset : cint; // offset from the start of the instance. (the lowest ivar_offset is 40,`\
115	115		` // because sizeof(objc_class) = 40, and objc_class is also part of the instance)`\
116	116		`end;`
117	117
118	118		### __inst_meth, __cls_meth
119	119
120	120		the section is optional and may omit if none of classes declares any methods
121	121
122	122		Both section has identical format. They contain number of objc_method_list records. Number of records is depending on classes declared in modules.
123	123
124	124		`objc_method_list = record`\
125	125		` obsolete : Pobjc_method_list; // not used, always zero`\
126	126		` method_count : cint; // number of objc_method in method_list array `\
127	127		` method_list : array[0..method_count-1] of objc_method; `\
128	128		`end;`
129	129
130	130		`objc_method = record`\
131	131		` method_name : SEL; // virtual address of the selector name `\
132	132		` // (selector name is stored as other names is in __TEXT __cstring section)`\
133	133		` // selector name is short: i.e. "methodName:", and not "+[className methodName:]"`\
134	134		` method_types : PChar; // obj-c encoded function parameters `\
135	135		` method_imp : IMP; // virtual address of method implementation function entry point (in __TEXT __text section)`\
136	136		`end;`
137	137
138	138		### __protocol
139	139
140	140		### __cat_cls_meth (protocols list)
141	141
142	142		### __cat_inst_meth
143	143
144	144		### __cat_cls_meth
145	145
146	146		## Mach-O additional 30Kb size
147	147
148	148		FPC built mach-o executables are somehow larger, compared to the win32 target, for example
149	149
150	150		`begin`\
151	151		` writeln('hello world');`\
152	152		`end.`
153	153
154	154		gives a 30k (stripped) executable for Win, and 60k for macOS.
155	155
156	156		Jonas Maebe: It's because there was a bug in older versions of the Darwin linker that required adding ".reference" assembler directives for routines that have more than one assembler name (most of the compiler helpers in the RTL have that). This fixed the problem, but as a result they are never smart linked out. It's only a fixed overhead of that 30kb (most programs don't contain any extra routines with multiple assembler names)
157	157
158	158		FPC 3.3.1 (trunk r44876) produces a 64 bit binary of 427,920 bytes; add -XX to smart link and it is reduced to 55,656 bytes; now strip it and it's just 47,224 bytes (~46K) which is pretty reasonable.
159	159
160	160		---
161	161		From the Free Pascal wiki, page "Mach-O", as it stood on 2020-05-17 after a change by Trev, out of the wiki's dump of 2024-05-26. Copyright stays with the author who wrote it, according to the old wiki's [[Licensing]] page. If you have concerns about this material, [message the admin](/wiki-contact).
162	162		Category: macOS

1

*(macOS only)*

2

3

*Mach-O, short for Mach object file format, is a file format for executables, object code, shared libraries, dynamically-loaded code, and core dumps. A derivation of the a.out format, Mach-O offered more extensibility and faster access to information in the symbol table.* Source: [Wikipedia](http://en.wikipedia.org/wiki/Mach-O).

4

5

The Apple file format reference can be found [here](https://developer.apple.com/library/archive/documentation/Performance/Conceptual/CodeFootprint/Articles/MachOOverview.html)

6

7

The following tools are used in macOS to view Mach-O files:

8

9

**otool** - object file displaying tool

10

11

**nm** - display name list (symbol table)

12

13

## Objective-C segment

14

15

There is no documentation about __OBJC segment and its sections. The following information has been gathered from cctools [sources](http://opensource.apple.com/source/cctools/cctools-667.3/)

16

17

### Structures strings

18

19

Some structures in sections contain name pointers. These names are stored in the c-strings section (segment: __TEXT; section: __cstring). The file offset for the string name can be evaluated in the following way:

20

21

`string_file_offset := cstr_section.offset + (name_addr - cstr_section.addr);`

22

23

## Sections

24

25

### __image_info

26

27

The section contains only the image info information:

28

29

`imageInfo = packed record`\

30

` version : uint32_t; // zero`\

31

` flags : uint32_t; // for objc 1.0 - zero`\

32

`end;`

33

34

Flags values:

35

36

`ImageInfo_F_and_C = $01;`\

37

`ImageInfo_GC = $02;`\

38

`ImageInfo_GC_only = $04;`

39

40

### __module_info

41

42

(objc_module record is declared at objc headers).

43

44

The number of objc_module structures depends on the number of .m files with objects declarations compiled. _symtab contains the number of classes and categories declared in the module.

45

46

`objc_module = packed record`\

47

` version : culong; // version number = 7 `\

48

` size : culong; // sizeof(objc_module)?`\

49

` name : PChar; // virtual memory address of the module name`\

50

` // Usually mapped to NULL string`\

51

` _symtab : Symtab; // virtual memory address of a proper objc_symtab structure (in __symbols section)`\

52

`end;`

53

54

### __symbols

55

56

the section contains symbol table for a module. (symtable record is declared at objc headers)

57

58

` objc_symtab = record`\

59

` sel_ref_cnt : culong; // zero`\

60

` refs : PSEL; // zero`\

61

` cls_def_cnt : cushort; // number of declared classes in the module`\

62

` cat_def_cnt : cushort; // number of declared categories`\

63

` defs: array [0..cls_def_cnt+cat_def_cnt-1] of Pointer; // array of virtual address of declarations`\

64

` // 0..cls_def_cnt - 1 : virtual address of class declarations (in "__class" section, can be empty)`\

65

` // cls_def_cnt..cat_def_cnt-1 : virtual address of categories declarations (can be empty)`\

66

` end;`

67

68

### __class, __meta_class

69

70

*the section may omit, if none of modules declare any custom classes*

71

72

Both sections use identical structure objc_class. (objc_class record is declared at objc header) Objective-C classes are declared in pair with their meta_classes.

73

74

`objc_class = record`\

75

` isa : PChar; // for class declaration: virtual address of meta-class declaration`\

76

` // for meta-class declaration: virtual address of "NSObject" string?`\

78

` super_class : PChar; // contains the pointer to super_class name`\

79

` name : PChar; // pointer to class name `\

80

` version : PChar; // = 0 (for obj-c version 1?)`\

82

` info : culong; // CLS_CLASS for classes`\

83

` // CLS_META for meta-classes `\

85

` instance_size : culong; // size of the instance:`\

86

` // class: size of the class instance `\

87

` // meta-class: 48 bytes `\

88

` `\

89

` ivars : Pobjc_ivar_list; // virtual address of objc_ivar_list (stored in "__instance_vars" section)`\

90

` // meta-classes don't have ivars list (=0)`\

92

` methodLists : PPobjc_method_list; // virtual address of objc_method_list`\

93

` // class declaration has instance methods list (stored in "__inst_meth" section)`\

94

` // meta-class declaration has class methods list (stored in "__cls_meth" section)`\

95

` `\

96

` cache : Pobjc_cache; // zero`\

97

` protocols : Pobjc_protocol_list; // pointer to protocols list. (stored in "__cat_cls_meth" section)`\

98

`end;`

99

100

### __instance_vars

101

102

*the section is optional and may omit if none of classes declares instance variables*

103

104

The section consists of number objc_ivar_list structutres. The number of structures depends upon number of declared classes using instance variables. (objc_ivar_list is declared in objc header)

105

106

`objc_ivar_list = record`\

107

` ivar_count : cint; // number of variables in the list`\

108

` ivar_list : array[0..ivar_count-1] of objc_ivar; variable length structure }`\

109

`end;`

110

111

`objc_ivar = record`\

112

` ivar_name : PChar; // vm addr of the variable name`\

113

` ivar_type : PChar; // vm addr of obj-c variable encoded type `\

114

` ivar_offset : cint; // offset from the start of the instance. (the lowest ivar_offset is 40,`\

115

` // because sizeof(objc_class) = 40, and objc_class is also part of the instance)`\

116

`end;`

117

118

### __inst_meth, __cls_meth

119

120

*the section is optional and may omit if none of classes declares any methods*

121

122

Both section has identical format. They contain number of objc_method_list records. Number of records is depending on classes declared in modules.

123

124

`objc_method_list = record`\

125

` obsolete : Pobjc_method_list; // not used, always zero`\

126

` method_count : cint; // number of objc_method in method_list array `\

127

` method_list : array[0..method_count-1] of objc_method; `\

128

`end;`

129

130

`objc_method = record`\

131

` method_name : SEL; // virtual address of the selector name `\

132

` // (selector name is stored as other names is in __TEXT __cstring section)`\

133

` // selector name is short: i.e. "methodName:", and not "+[className methodName:]"`\

134

` method_types : PChar; // obj-c encoded function parameters `\

135

` method_imp : IMP; // virtual address of method implementation function entry point (in __TEXT __text section)`\

136

`end;`

137

138

### __protocol

139

140

### __cat_cls_meth (protocols list)

141

142

### __cat_inst_meth

143

144

### __cat_cls_meth

145

146

## Mach-O additional 30Kb size

147

148

FPC built mach-o executables are somehow larger, compared to the win32 target, for example

149

150

`begin`\

151

` writeln('hello world');`\

152

`end.`

153

154

gives a 30k (stripped) executable for Win, and 60k for macOS.

155

156

Jonas Maebe: *It's because there was a bug in older versions of the Darwin linker that required adding ".reference" assembler directives for routines that have more than one assembler name (most of the compiler helpers in the RTL have that). This fixed the problem, but as a result they are never smart linked out. It's only a fixed overhead of that 30kb (most programs don't contain any extra routines with multiple assembler names)*

157

158

FPC 3.3.1 (trunk r44876) produces a 64 bit binary of 427,920 bytes; add -XX to smart link and it is reduced to 55,656 bytes; now strip it and it's just 47,224 bytes (~46K) which is pretty reasonable.

159

160

---

161

*From the Free Pascal wiki, page "Mach-O", as it stood on 2020-05-17 after a change by Trev, out of the wiki's dump of 2024-05-26. Copyright stays with the author who wrote it, according to the old wiki's [[Licensing]] page. If you have concerns about this material, [message the admin](/wiki-contact).*

162

Category: macOS