• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===------------------ mach-o/compact_unwind_encoding.h ------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is dual licensed under the MIT and the University of Illinois Open
6 // Source Licenses. See LICENSE.TXT for details.
7 //
8 //
9 // Darwin's alternative to DWARF based unwind encodings.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 
14 #ifndef __COMPACT_UNWIND_ENCODING__
15 #define __COMPACT_UNWIND_ENCODING__
16 
17 #include <stdint.h>
18 
19 //
20 // Compilers can emit standard DWARF FDEs in the __TEXT,__eh_frame section
21 // of object files. Or compilers can emit compact unwind information in
22 // the __LD,__compact_unwind section.
23 //
24 // When the linker creates a final linked image, it will create a
25 // __TEXT,__unwind_info section.  This section is a small and fast way for the
26 // runtime to access unwind info for any given function.  If the compiler
27 // emitted compact unwind info for the function, that compact unwind info will
28 // be encoded in the __TEXT,__unwind_info section. If the compiler emitted
29 // DWARF unwind info, the __TEXT,__unwind_info section will contain the offset
30 // of the FDE in the __TEXT,__eh_frame section in the final linked image.
31 //
32 // Note: Previously, the linker would transform some DWARF unwind infos into
33 //       compact unwind info.  But that is fragile and no longer done.
34 
35 
36 //
37 // The compact unwind endoding is a 32-bit value which encoded in an
38 // architecture specific way, which registers to restore from where, and how
39 // to unwind out of the function.
40 //
41 typedef uint32_t compact_unwind_encoding_t;
42 
43 
44 // architecture independent bits
45 enum {
46     UNWIND_IS_NOT_FUNCTION_START           = 0x80000000,
47     UNWIND_HAS_LSDA                        = 0x40000000,
48     UNWIND_PERSONALITY_MASK                = 0x30000000,
49 };
50 
51 
52 
53 
54 //
55 // x86
56 //
57 // 1-bit: start
58 // 1-bit: has lsda
59 // 2-bit: personality index
60 //
61 // 4-bits: 0=old, 1=ebp based, 2=stack-imm, 3=stack-ind, 4=DWARF
62 //  ebp based:
63 //        15-bits (5*3-bits per reg) register permutation
64 //        8-bits for stack offset
65 //  frameless:
66 //        8-bits stack size
67 //        3-bits stack adjust
68 //        3-bits register count
69 //        10-bits register permutation
70 //
71 enum {
72     UNWIND_X86_MODE_MASK                         = 0x0F000000,
73     UNWIND_X86_MODE_EBP_FRAME                    = 0x01000000,
74     UNWIND_X86_MODE_STACK_IMMD                   = 0x02000000,
75     UNWIND_X86_MODE_STACK_IND                    = 0x03000000,
76     UNWIND_X86_MODE_DWARF                        = 0x04000000,
77 
78     UNWIND_X86_EBP_FRAME_REGISTERS               = 0x00007FFF,
79     UNWIND_X86_EBP_FRAME_OFFSET                  = 0x00FF0000,
80 
81     UNWIND_X86_FRAMELESS_STACK_SIZE              = 0x00FF0000,
82     UNWIND_X86_FRAMELESS_STACK_ADJUST            = 0x0000E000,
83     UNWIND_X86_FRAMELESS_STACK_REG_COUNT         = 0x00001C00,
84     UNWIND_X86_FRAMELESS_STACK_REG_PERMUTATION   = 0x000003FF,
85 
86     UNWIND_X86_DWARF_SECTION_OFFSET              = 0x00FFFFFF,
87 };
88 
89 enum {
90     UNWIND_X86_REG_NONE     = 0,
91     UNWIND_X86_REG_EBX      = 1,
92     UNWIND_X86_REG_ECX      = 2,
93     UNWIND_X86_REG_EDX      = 3,
94     UNWIND_X86_REG_EDI      = 4,
95     UNWIND_X86_REG_ESI      = 5,
96     UNWIND_X86_REG_EBP      = 6,
97 };
98 
99 //
100 // For x86 there are four modes for the compact unwind encoding:
101 // UNWIND_X86_MODE_EBP_FRAME:
102 //    EBP based frame where EBP is push on stack immediately after return address,
103 //    then ESP is moved to EBP. Thus, to unwind ESP is restored with the current
104 //    EPB value, then EBP is restored by popping off the stack, and the return
105 //    is done by popping the stack once more into the pc.
106 //    All non-volatile registers that need to be restored must have been saved
107 //    in a small range in the stack that starts EBP-4 to EBP-1020.  The offset/4
108 //    is encoded in the UNWIND_X86_EBP_FRAME_OFFSET bits.  The registers saved
109 //    are encoded in the UNWIND_X86_EBP_FRAME_REGISTERS bits as five 3-bit entries.
110 //    Each entry contains which register to restore.
111 // UNWIND_X86_MODE_STACK_IMMD:
112 //    A "frameless" (EBP not used as frame pointer) function with a small
113 //    constant stack size.  To return, a constant (encoded in the compact
114 //    unwind encoding) is added to the ESP. Then the return is done by
115 //    popping the stack into the pc.
116 //    All non-volatile registers that need to be restored must have been saved
117 //    on the stack immediately after the return address.  The stack_size/4 is
118 //    encoded in the UNWIND_X86_FRAMELESS_STACK_SIZE (max stack size is 1024).
119 //    The number of registers saved is encoded in UNWIND_X86_FRAMELESS_STACK_REG_COUNT.
120 //    UNWIND_X86_FRAMELESS_STACK_REG_PERMUTATION constains which registers were
121 //    saved and their order.
122 // UNWIND_X86_MODE_STACK_IND:
123 //    A "frameless" (EBP not used as frame pointer) function large constant
124 //    stack size.  This case is like the previous, except the stack size is too
125 //    large to encode in the compact unwind encoding.  Instead it requires that
126 //    the function contains "subl $nnnnnnnn,ESP" in its prolog.  The compact
127 //    encoding contains the offset to the nnnnnnnn value in the function in
128 //    UNWIND_X86_FRAMELESS_STACK_SIZE.
129 // UNWIND_X86_MODE_DWARF:
130 //    No compact unwind encoding is available.  Instead the low 24-bits of the
131 //    compact encoding is the offset of the DWARF FDE in the __eh_frame section.
132 //    This mode is never used in object files.  It is only generated by the
133 //    linker in final linked images which have only DWARF unwind info for a
134 //    function.
135 //
136 // The permutation encoding is a Lehmer code sequence encoded into a
137 // single variable-base number so we can encode the ordering of up to
138 // six registers in a 10-bit space.
139 //
140 // The following is the algorithm used to create the permutation encoding used
141 // with frameless stacks.  It is passed the number of registers to be saved and
142 // an array of the register numbers saved.
143 //
144 //uint32_t permute_encode(uint32_t registerCount, const uint32_t registers[6])
145 //{
146 //    uint32_t renumregs[6];
147 //    for (int i=6-registerCount; i < 6; ++i) {
148 //        int countless = 0;
149 //        for (int j=6-registerCount; j < i; ++j) {
150 //            if ( registers[j] < registers[i] )
151 //                ++countless;
152 //        }
153 //        renumregs[i] = registers[i] - countless -1;
154 //    }
155 //    uint32_t permutationEncoding = 0;
156 //    switch ( registerCount ) {
157 //        case 6:
158 //            permutationEncoding |= (120*renumregs[0] + 24*renumregs[1]
159 //                                    + 6*renumregs[2] + 2*renumregs[3]
160 //                                      + renumregs[4]);
161 //            break;
162 //        case 5:
163 //            permutationEncoding |= (120*renumregs[1] + 24*renumregs[2]
164 //                                    + 6*renumregs[3] + 2*renumregs[4]
165 //                                      + renumregs[5]);
166 //            break;
167 //        case 4:
168 //            permutationEncoding |= (60*renumregs[2] + 12*renumregs[3]
169 //                                   + 3*renumregs[4] + renumregs[5]);
170 //            break;
171 //        case 3:
172 //            permutationEncoding |= (20*renumregs[3] + 4*renumregs[4]
173 //                                     + renumregs[5]);
174 //            break;
175 //        case 2:
176 //            permutationEncoding |= (5*renumregs[4] + renumregs[5]);
177 //            break;
178 //        case 1:
179 //            permutationEncoding |= (renumregs[5]);
180 //            break;
181 //    }
182 //    return permutationEncoding;
183 //}
184 //
185 
186 
187 
188 
189 //
190 // x86_64
191 //
192 // 1-bit: start
193 // 1-bit: has lsda
194 // 2-bit: personality index
195 //
196 // 4-bits: 0=old, 1=rbp based, 2=stack-imm, 3=stack-ind, 4=DWARF
197 //  rbp based:
198 //        15-bits (5*3-bits per reg) register permutation
199 //        8-bits for stack offset
200 //  frameless:
201 //        8-bits stack size
202 //        3-bits stack adjust
203 //        3-bits register count
204 //        10-bits register permutation
205 //
206 enum {
207     UNWIND_X86_64_MODE_MASK                         = 0x0F000000,
208     UNWIND_X86_64_MODE_RBP_FRAME                    = 0x01000000,
209     UNWIND_X86_64_MODE_STACK_IMMD                   = 0x02000000,
210     UNWIND_X86_64_MODE_STACK_IND                    = 0x03000000,
211     UNWIND_X86_64_MODE_DWARF                        = 0x04000000,
212 
213     UNWIND_X86_64_RBP_FRAME_REGISTERS               = 0x00007FFF,
214     UNWIND_X86_64_RBP_FRAME_OFFSET                  = 0x00FF0000,
215 
216     UNWIND_X86_64_FRAMELESS_STACK_SIZE              = 0x00FF0000,
217     UNWIND_X86_64_FRAMELESS_STACK_ADJUST            = 0x0000E000,
218     UNWIND_X86_64_FRAMELESS_STACK_REG_COUNT         = 0x00001C00,
219     UNWIND_X86_64_FRAMELESS_STACK_REG_PERMUTATION   = 0x000003FF,
220 
221     UNWIND_X86_64_DWARF_SECTION_OFFSET              = 0x00FFFFFF,
222 };
223 
224 enum {
225     UNWIND_X86_64_REG_NONE       = 0,
226     UNWIND_X86_64_REG_RBX        = 1,
227     UNWIND_X86_64_REG_R12        = 2,
228     UNWIND_X86_64_REG_R13        = 3,
229     UNWIND_X86_64_REG_R14        = 4,
230     UNWIND_X86_64_REG_R15        = 5,
231     UNWIND_X86_64_REG_RBP        = 6,
232 };
233 //
234 // For x86_64 there are four modes for the compact unwind encoding:
235 // UNWIND_X86_64_MODE_RBP_FRAME:
236 //    RBP based frame where RBP is push on stack immediately after return address,
237 //    then RSP is moved to RBP. Thus, to unwind RSP is restored with the current
238 //    EPB value, then RBP is restored by popping off the stack, and the return
239 //    is done by popping the stack once more into the pc.
240 //    All non-volatile registers that need to be restored must have been saved
241 //    in a small range in the stack that starts RBP-8 to RBP-2040.  The offset/8
242 //    is encoded in the UNWIND_X86_64_RBP_FRAME_OFFSET bits.  The registers saved
243 //    are encoded in the UNWIND_X86_64_RBP_FRAME_REGISTERS bits as five 3-bit entries.
244 //    Each entry contains which register to restore.
245 // UNWIND_X86_64_MODE_STACK_IMMD:
246 //    A "frameless" (RBP not used as frame pointer) function with a small
247 //    constant stack size.  To return, a constant (encoded in the compact
248 //    unwind encoding) is added to the RSP. Then the return is done by
249 //    popping the stack into the pc.
250 //    All non-volatile registers that need to be restored must have been saved
251 //    on the stack immediately after the return address.  The stack_size/8 is
252 //    encoded in the UNWIND_X86_64_FRAMELESS_STACK_SIZE (max stack size is 2048).
253 //    The number of registers saved is encoded in UNWIND_X86_64_FRAMELESS_STACK_REG_COUNT.
254 //    UNWIND_X86_64_FRAMELESS_STACK_REG_PERMUTATION constains which registers were
255 //    saved and their order.
256 // UNWIND_X86_64_MODE_STACK_IND:
257 //    A "frameless" (RBP not used as frame pointer) function large constant
258 //    stack size.  This case is like the previous, except the stack size is too
259 //    large to encode in the compact unwind encoding.  Instead it requires that
260 //    the function contains "subq $nnnnnnnn,RSP" in its prolog.  The compact
261 //    encoding contains the offset to the nnnnnnnn value in the function in
262 //    UNWIND_X86_64_FRAMELESS_STACK_SIZE.
263 // UNWIND_X86_64_MODE_DWARF:
264 //    No compact unwind encoding is available.  Instead the low 24-bits of the
265 //    compact encoding is the offset of the DWARF FDE in the __eh_frame section.
266 //    This mode is never used in object files.  It is only generated by the
267 //    linker in final linked images which have only DWARF unwind info for a
268 //    function.
269 //
270 
271 
272 // ARM64
273 //
274 // 1-bit: start
275 // 1-bit: has lsda
276 // 2-bit: personality index
277 //
278 // 4-bits: 4=frame-based, 3=DWARF, 2=frameless
279 //  frameless:
280 //        12-bits of stack size
281 //  frame-based:
282 //        4-bits D reg pairs saved
283 //        5-bits X reg pairs saved
284 //  DWARF:
285 //        24-bits offset of DWARF FDE in __eh_frame section
286 //
287 enum {
288     UNWIND_ARM64_MODE_MASK                     = 0x0F000000,
289     UNWIND_ARM64_MODE_FRAMELESS                = 0x02000000,
290     UNWIND_ARM64_MODE_DWARF                    = 0x03000000,
291     UNWIND_ARM64_MODE_FRAME                    = 0x04000000,
292 
293     UNWIND_ARM64_FRAME_X19_X20_PAIR            = 0x00000001,
294     UNWIND_ARM64_FRAME_X21_X22_PAIR            = 0x00000002,
295     UNWIND_ARM64_FRAME_X23_X24_PAIR            = 0x00000004,
296     UNWIND_ARM64_FRAME_X25_X26_PAIR            = 0x00000008,
297     UNWIND_ARM64_FRAME_X27_X28_PAIR            = 0x00000010,
298     UNWIND_ARM64_FRAME_D8_D9_PAIR              = 0x00000100,
299     UNWIND_ARM64_FRAME_D10_D11_PAIR            = 0x00000200,
300     UNWIND_ARM64_FRAME_D12_D13_PAIR            = 0x00000400,
301     UNWIND_ARM64_FRAME_D14_D15_PAIR            = 0x00000800,
302 
303     UNWIND_ARM64_FRAMELESS_STACK_SIZE_MASK     = 0x00FFF000,
304     UNWIND_ARM64_DWARF_SECTION_OFFSET          = 0x00FFFFFF,
305 };
306 // For arm64 there are three modes for the compact unwind encoding:
307 // UNWIND_ARM64_MODE_FRAME:
308 //    This is a standard arm64 prolog where FP/LR are immediately pushed on the
309 //    stack, then SP is copied to FP. If there are any non-volatile registers
310 //    saved, then are copied into the stack frame in pairs in a contiguous
311 //    range right below the saved FP/LR pair.  Any subset of the five X pairs
312 //    and four D pairs can be saved, but the memory layout must be in register
313 //    number order.
314 // UNWIND_ARM64_MODE_FRAMELESS:
315 //    A "frameless" leaf function, where FP/LR are not saved. The return address
316 //    remains in LR throughout the function. If any non-volatile registers
317 //    are saved, they must be pushed onto the stack before any stack space is
318 //    allocated for local variables.  The stack sized (including any saved
319 //    non-volatile registers) divided by 16 is encoded in the bits
320 //    UNWIND_ARM64_FRAMELESS_STACK_SIZE_MASK.
321 // UNWIND_ARM64_MODE_DWARF:
322 //    No compact unwind encoding is available.  Instead the low 24-bits of the
323 //    compact encoding is the offset of the DWARF FDE in the __eh_frame section.
324 //    This mode is never used in object files.  It is only generated by the
325 //    linker in final linked images which have only DWARF unwind info for a
326 //    function.
327 //
328 
329 
330 
331 
332 
333 ////////////////////////////////////////////////////////////////////////////////
334 //
335 //  Relocatable Object Files: __LD,__compact_unwind
336 //
337 ////////////////////////////////////////////////////////////////////////////////
338 
339 //
340 // A compiler can generated compact unwind information for a function by adding
341 // a "row" to the __LD,__compact_unwind section.  This section has the
342 // S_ATTR_DEBUG bit set, so the section will be ignored by older linkers.
343 // It is removed by the new linker, so never ends up in final executables.
344 // This section is a table, initially with one row per function (that needs
345 // unwind info).  The table columns and some conceptual entries are:
346 //
347 //     range-start               pointer to start of function/range
348 //     range-length
349 //     compact-unwind-encoding   32-bit encoding
350 //     personality-function      or zero if no personality function
351 //     lsda                      or zero if no LSDA data
352 //
353 // The length and encoding fields are 32-bits.  The other are all pointer sized.
354 //
355 // In x86_64 assembly, these entry would look like:
356 //
357 //     .section __LD,__compact_unwind,regular,debug
358 //
359 //     #compact unwind for _foo
360 //     .quad    _foo
361 //     .set     L1,LfooEnd-_foo
362 //     .long    L1
363 //     .long    0x01010001
364 //     .quad    0
365 //     .quad    0
366 //
367 //     #compact unwind for _bar
368 //     .quad    _bar
369 //     .set     L2,LbarEnd-_bar
370 //     .long    L2
371 //     .long    0x01020011
372 //     .quad    __gxx_personality
373 //     .quad    except_tab1
374 //
375 //
376 // Notes: There is no need for any labels in the the __compact_unwind section.
377 //        The use of the .set directive is to force the evaluation of the
378 //        range-length at assembly time, instead of generating relocations.
379 //
380 // To support future compiler optimizations where which non-volatile registers
381 // are saved changes within a function (e.g. delay saving non-volatiles until
382 // necessary), there can by multiple lines in the __compact_unwind table for one
383 // function, each with a different (non-overlapping) range and each with
384 // different compact unwind encodings that correspond to the non-volatiles
385 // saved at that range of the function.
386 //
387 // If a particular function is so wacky that there is no compact unwind way
388 // to encode it, then the compiler can emit traditional DWARF unwind info.
389 // The runtime will use which ever is available.
390 //
391 // Runtime support for compact unwind encodings are only available on 10.6
392 // and later.  So, the compiler should not generate it when targeting pre-10.6.
393 
394 
395 
396 
397 ////////////////////////////////////////////////////////////////////////////////
398 //
399 //  Final Linked Images: __TEXT,__unwind_info
400 //
401 ////////////////////////////////////////////////////////////////////////////////
402 
403 //
404 // The __TEXT,__unwind_info section is laid out for an efficient two level lookup.
405 // The header of the section contains a coarse index that maps function address
406 // to the page (4096 byte block) containing the unwind info for that function.
407 //
408 
409 #define UNWIND_SECTION_VERSION 1
410 struct unwind_info_section_header
411 {
412     uint32_t    version;            // UNWIND_SECTION_VERSION
413     uint32_t    commonEncodingsArraySectionOffset;
414     uint32_t    commonEncodingsArrayCount;
415     uint32_t    personalityArraySectionOffset;
416     uint32_t    personalityArrayCount;
417     uint32_t    indexSectionOffset;
418     uint32_t    indexCount;
419     // compact_unwind_encoding_t[]
420     // uint32_t personalities[]
421     // unwind_info_section_header_index_entry[]
422     // unwind_info_section_header_lsda_index_entry[]
423 };
424 
425 struct unwind_info_section_header_index_entry
426 {
427     uint32_t        functionOffset;
428     uint32_t        secondLevelPagesSectionOffset;  // section offset to start of regular or compress page
429     uint32_t        lsdaIndexArraySectionOffset;    // section offset to start of lsda_index array for this range
430 };
431 
432 struct unwind_info_section_header_lsda_index_entry
433 {
434     uint32_t        functionOffset;
435     uint32_t        lsdaOffset;
436 };
437 
438 //
439 // There are two kinds of second level index pages: regular and compressed.
440 // A compressed page can hold up to 1021 entries, but it cannot be used
441 // if too many different encoding types are used.  The regular page holds
442 // 511 entries.
443 //
444 
445 struct unwind_info_regular_second_level_entry
446 {
447     uint32_t                    functionOffset;
448     compact_unwind_encoding_t    encoding;
449 };
450 
451 #define UNWIND_SECOND_LEVEL_REGULAR 2
452 struct unwind_info_regular_second_level_page_header
453 {
454     uint32_t    kind;    // UNWIND_SECOND_LEVEL_REGULAR
455     uint16_t    entryPageOffset;
456     uint16_t    entryCount;
457     // entry array
458 };
459 
460 #define UNWIND_SECOND_LEVEL_COMPRESSED 3
461 struct unwind_info_compressed_second_level_page_header
462 {
463     uint32_t    kind;    // UNWIND_SECOND_LEVEL_COMPRESSED
464     uint16_t    entryPageOffset;
465     uint16_t    entryCount;
466     uint16_t    encodingsPageOffset;
467     uint16_t    encodingsCount;
468     // 32-bit entry array
469     // encodings array
470 };
471 
472 #define UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(entry)            (entry & 0x00FFFFFF)
473 #define UNWIND_INFO_COMPRESSED_ENTRY_ENCODING_INDEX(entry)        ((entry >> 24) & 0xFF)
474 
475 
476 
477 #endif
478 
479