1 | //===------------------ mach-o/compact_unwind_encoding.h ------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | // |
8 | // Darwin's alternative to DWARF based unwind encodings. |
9 | // |
10 | //===----------------------------------------------------------------------===// |
11 | |
12 | |
13 | #ifndef __COMPACT_UNWIND_ENCODING__ |
14 | #define __COMPACT_UNWIND_ENCODING__ |
15 | |
16 | #include <stdint.h> |
17 | |
18 | // |
19 | // Compilers can emit standard DWARF FDEs in the __TEXT,__eh_frame section |
20 | // of object files. Or compilers can emit compact unwind information in |
21 | // the __LD,__compact_unwind section. |
22 | // |
23 | // When the linker creates a final linked image, it will create a |
24 | // __TEXT,__unwind_info section. This section is a small and fast way for the |
25 | // runtime to access unwind info for any given function. If the compiler |
26 | // emitted compact unwind info for the function, that compact unwind info will |
27 | // be encoded in the __TEXT,__unwind_info section. If the compiler emitted |
28 | // DWARF unwind info, the __TEXT,__unwind_info section will contain the offset |
29 | // of the FDE in the __TEXT,__eh_frame section in the final linked image. |
30 | // |
31 | // Note: Previously, the linker would transform some DWARF unwind infos into |
32 | // compact unwind info. But that is fragile and no longer done. |
33 | |
34 | |
35 | // |
36 | // The compact unwind endoding is a 32-bit value which encoded in an |
37 | // architecture specific way, which registers to restore from where, and how |
38 | // to unwind out of the function. |
39 | // |
40 | typedef uint32_t compact_unwind_encoding_t; |
41 | |
42 | |
43 | // architecture independent bits |
44 | enum { |
45 | UNWIND_IS_NOT_FUNCTION_START = 0x80000000, |
46 | UNWIND_HAS_LSDA = 0x40000000, |
47 | UNWIND_PERSONALITY_MASK = 0x30000000, |
48 | }; |
49 | |
50 | |
51 | |
52 | |
53 | // |
54 | // x86 |
55 | // |
56 | // 1-bit: start |
57 | // 1-bit: has lsda |
58 | // 2-bit: personality index |
59 | // |
60 | // 4-bits: 0=old, 1=ebp based, 2=stack-imm, 3=stack-ind, 4=DWARF |
61 | // ebp based: |
62 | // 15-bits (5*3-bits per reg) register permutation |
63 | // 8-bits for stack offset |
64 | // frameless: |
65 | // 8-bits stack size |
66 | // 3-bits stack adjust |
67 | // 3-bits register count |
68 | // 10-bits register permutation |
69 | // |
70 | enum { |
71 | UNWIND_X86_MODE_MASK = 0x0F000000, |
72 | UNWIND_X86_MODE_EBP_FRAME = 0x01000000, |
73 | UNWIND_X86_MODE_STACK_IMMD = 0x02000000, |
74 | UNWIND_X86_MODE_STACK_IND = 0x03000000, |
75 | UNWIND_X86_MODE_DWARF = 0x04000000, |
76 | |
77 | UNWIND_X86_EBP_FRAME_REGISTERS = 0x00007FFF, |
78 | UNWIND_X86_EBP_FRAME_OFFSET = 0x00FF0000, |
79 | |
80 | UNWIND_X86_FRAMELESS_STACK_SIZE = 0x00FF0000, |
81 | UNWIND_X86_FRAMELESS_STACK_ADJUST = 0x0000E000, |
82 | UNWIND_X86_FRAMELESS_STACK_REG_COUNT = 0x00001C00, |
83 | UNWIND_X86_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF, |
84 | |
85 | UNWIND_X86_DWARF_SECTION_OFFSET = 0x00FFFFFF, |
86 | }; |
87 | |
88 | enum { |
89 | UNWIND_X86_REG_NONE = 0, |
90 | UNWIND_X86_REG_EBX = 1, |
91 | UNWIND_X86_REG_ECX = 2, |
92 | UNWIND_X86_REG_EDX = 3, |
93 | UNWIND_X86_REG_EDI = 4, |
94 | UNWIND_X86_REG_ESI = 5, |
95 | UNWIND_X86_REG_EBP = 6, |
96 | }; |
97 | |
98 | // |
99 | // For x86 there are four modes for the compact unwind encoding: |
100 | // UNWIND_X86_MODE_EBP_FRAME: |
101 | // EBP based frame where EBP is push on stack immediately after return address, |
102 | // then ESP is moved to EBP. Thus, to unwind ESP is restored with the current |
103 | // EPB value, then EBP is restored by popping off the stack, and the return |
104 | // is done by popping the stack once more into the pc. |
105 | // All non-volatile registers that need to be restored must have been saved |
106 | // in a small range in the stack that starts EBP-4 to EBP-1020. The offset/4 |
107 | // is encoded in the UNWIND_X86_EBP_FRAME_OFFSET bits. The registers saved |
108 | // are encoded in the UNWIND_X86_EBP_FRAME_REGISTERS bits as five 3-bit entries. |
109 | // Each entry contains which register to restore. |
110 | // UNWIND_X86_MODE_STACK_IMMD: |
111 | // A "frameless" (EBP not used as frame pointer) function with a small |
112 | // constant stack size. To return, a constant (encoded in the compact |
113 | // unwind encoding) is added to the ESP. Then the return is done by |
114 | // popping the stack into the pc. |
115 | // All non-volatile registers that need to be restored must have been saved |
116 | // on the stack immediately after the return address. The stack_size/4 is |
117 | // encoded in the UNWIND_X86_FRAMELESS_STACK_SIZE (max stack size is 1024). |
118 | // The number of registers saved is encoded in UNWIND_X86_FRAMELESS_STACK_REG_COUNT. |
119 | // UNWIND_X86_FRAMELESS_STACK_REG_PERMUTATION constains which registers were |
120 | // saved and their order. |
121 | // UNWIND_X86_MODE_STACK_IND: |
122 | // A "frameless" (EBP not used as frame pointer) function large constant |
123 | // stack size. This case is like the previous, except the stack size is too |
124 | // large to encode in the compact unwind encoding. Instead it requires that |
125 | // the function contains "subl $nnnnnnnn,ESP" in its prolog. The compact |
126 | // encoding contains the offset to the nnnnnnnn value in the function in |
127 | // UNWIND_X86_FRAMELESS_STACK_SIZE. |
128 | // UNWIND_X86_MODE_DWARF: |
129 | // No compact unwind encoding is available. Instead the low 24-bits of the |
130 | // compact encoding is the offset of the DWARF FDE in the __eh_frame section. |
131 | // This mode is never used in object files. It is only generated by the |
132 | // linker in final linked images which have only DWARF unwind info for a |
133 | // function. |
134 | // |
135 | // The permutation encoding is a Lehmer code sequence encoded into a |
136 | // single variable-base number so we can encode the ordering of up to |
137 | // six registers in a 10-bit space. |
138 | // |
139 | // The following is the algorithm used to create the permutation encoding used |
140 | // with frameless stacks. It is passed the number of registers to be saved and |
141 | // an array of the register numbers saved. |
142 | // |
143 | //uint32_t permute_encode(uint32_t registerCount, const uint32_t registers[6]) |
144 | //{ |
145 | // uint32_t renumregs[6]; |
146 | // for (int i=6-registerCount; i < 6; ++i) { |
147 | // int countless = 0; |
148 | // for (int j=6-registerCount; j < i; ++j) { |
149 | // if ( registers[j] < registers[i] ) |
150 | // ++countless; |
151 | // } |
152 | // renumregs[i] = registers[i] - countless -1; |
153 | // } |
154 | // uint32_t permutationEncoding = 0; |
155 | // switch ( registerCount ) { |
156 | // case 6: |
157 | // permutationEncoding |= (120*renumregs[0] + 24*renumregs[1] |
158 | // + 6*renumregs[2] + 2*renumregs[3] |
159 | // + renumregs[4]); |
160 | // break; |
161 | // case 5: |
162 | // permutationEncoding |= (120*renumregs[1] + 24*renumregs[2] |
163 | // + 6*renumregs[3] + 2*renumregs[4] |
164 | // + renumregs[5]); |
165 | // break; |
166 | // case 4: |
167 | // permutationEncoding |= (60*renumregs[2] + 12*renumregs[3] |
168 | // + 3*renumregs[4] + renumregs[5]); |
169 | // break; |
170 | // case 3: |
171 | // permutationEncoding |= (20*renumregs[3] + 4*renumregs[4] |
172 | // + renumregs[5]); |
173 | // break; |
174 | // case 2: |
175 | // permutationEncoding |= (5*renumregs[4] + renumregs[5]); |
176 | // break; |
177 | // case 1: |
178 | // permutationEncoding |= (renumregs[5]); |
179 | // break; |
180 | // } |
181 | // return permutationEncoding; |
182 | //} |
183 | // |
184 | |
185 | |
186 | |
187 | |
188 | // |
189 | // x86_64 |
190 | // |
191 | // 1-bit: start |
192 | // 1-bit: has lsda |
193 | // 2-bit: personality index |
194 | // |
195 | // 4-bits: 0=old, 1=rbp based, 2=stack-imm, 3=stack-ind, 4=DWARF |
196 | // rbp based: |
197 | // 15-bits (5*3-bits per reg) register permutation |
198 | // 8-bits for stack offset |
199 | // frameless: |
200 | // 8-bits stack size |
201 | // 3-bits stack adjust |
202 | // 3-bits register count |
203 | // 10-bits register permutation |
204 | // |
205 | enum { |
206 | UNWIND_X86_64_MODE_MASK = 0x0F000000, |
207 | UNWIND_X86_64_MODE_RBP_FRAME = 0x01000000, |
208 | UNWIND_X86_64_MODE_STACK_IMMD = 0x02000000, |
209 | UNWIND_X86_64_MODE_STACK_IND = 0x03000000, |
210 | UNWIND_X86_64_MODE_DWARF = 0x04000000, |
211 | |
212 | UNWIND_X86_64_RBP_FRAME_REGISTERS = 0x00007FFF, |
213 | UNWIND_X86_64_RBP_FRAME_OFFSET = 0x00FF0000, |
214 | |
215 | UNWIND_X86_64_FRAMELESS_STACK_SIZE = 0x00FF0000, |
216 | UNWIND_X86_64_FRAMELESS_STACK_ADJUST = 0x0000E000, |
217 | UNWIND_X86_64_FRAMELESS_STACK_REG_COUNT = 0x00001C00, |
218 | UNWIND_X86_64_FRAMELESS_STACK_REG_PERMUTATION = 0x000003FF, |
219 | |
220 | UNWIND_X86_64_DWARF_SECTION_OFFSET = 0x00FFFFFF, |
221 | }; |
222 | |
223 | enum { |
224 | UNWIND_X86_64_REG_NONE = 0, |
225 | UNWIND_X86_64_REG_RBX = 1, |
226 | UNWIND_X86_64_REG_R12 = 2, |
227 | UNWIND_X86_64_REG_R13 = 3, |
228 | UNWIND_X86_64_REG_R14 = 4, |
229 | UNWIND_X86_64_REG_R15 = 5, |
230 | UNWIND_X86_64_REG_RBP = 6, |
231 | }; |
232 | // |
233 | // For x86_64 there are four modes for the compact unwind encoding: |
234 | // UNWIND_X86_64_MODE_RBP_FRAME: |
235 | // RBP based frame where RBP is push on stack immediately after return address, |
236 | // then RSP is moved to RBP. Thus, to unwind RSP is restored with the current |
237 | // EPB value, then RBP is restored by popping off the stack, and the return |
238 | // is done by popping the stack once more into the pc. |
239 | // All non-volatile registers that need to be restored must have been saved |
240 | // in a small range in the stack that starts RBP-8 to RBP-2040. The offset/8 |
241 | // is encoded in the UNWIND_X86_64_RBP_FRAME_OFFSET bits. The registers saved |
242 | // are encoded in the UNWIND_X86_64_RBP_FRAME_REGISTERS bits as five 3-bit entries. |
243 | // Each entry contains which register to restore. |
244 | // UNWIND_X86_64_MODE_STACK_IMMD: |
245 | // A "frameless" (RBP not used as frame pointer) function with a small |
246 | // constant stack size. To return, a constant (encoded in the compact |
247 | // unwind encoding) is added to the RSP. Then the return is done by |
248 | // popping the stack into the pc. |
249 | // All non-volatile registers that need to be restored must have been saved |
250 | // on the stack immediately after the return address. The stack_size/8 is |
251 | // encoded in the UNWIND_X86_64_FRAMELESS_STACK_SIZE (max stack size is 2048). |
252 | // The number of registers saved is encoded in UNWIND_X86_64_FRAMELESS_STACK_REG_COUNT. |
253 | // UNWIND_X86_64_FRAMELESS_STACK_REG_PERMUTATION constains which registers were |
254 | // saved and their order. |
255 | // UNWIND_X86_64_MODE_STACK_IND: |
256 | // A "frameless" (RBP not used as frame pointer) function large constant |
257 | // stack size. This case is like the previous, except the stack size is too |
258 | // large to encode in the compact unwind encoding. Instead it requires that |
259 | // the function contains "subq $nnnnnnnn,RSP" in its prolog. The compact |
260 | // encoding contains the offset to the nnnnnnnn value in the function in |
261 | // UNWIND_X86_64_FRAMELESS_STACK_SIZE. |
262 | // UNWIND_X86_64_MODE_DWARF: |
263 | // No compact unwind encoding is available. Instead the low 24-bits of the |
264 | // compact encoding is the offset of the DWARF FDE in the __eh_frame section. |
265 | // This mode is never used in object files. It is only generated by the |
266 | // linker in final linked images which have only DWARF unwind info for a |
267 | // function. |
268 | // |
269 | |
270 | |
271 | // ARM64 |
272 | // |
273 | // 1-bit: start |
274 | // 1-bit: has lsda |
275 | // 2-bit: personality index |
276 | // |
277 | // 4-bits: 4=frame-based, 3=DWARF, 2=frameless |
278 | // frameless: |
279 | // 12-bits of stack size |
280 | // frame-based: |
281 | // 4-bits D reg pairs saved |
282 | // 5-bits X reg pairs saved |
283 | // DWARF: |
284 | // 24-bits offset of DWARF FDE in __eh_frame section |
285 | // |
286 | enum { |
287 | UNWIND_ARM64_MODE_MASK = 0x0F000000, |
288 | UNWIND_ARM64_MODE_FRAMELESS = 0x02000000, |
289 | UNWIND_ARM64_MODE_DWARF = 0x03000000, |
290 | UNWIND_ARM64_MODE_FRAME = 0x04000000, |
291 | |
292 | UNWIND_ARM64_FRAME_X19_X20_PAIR = 0x00000001, |
293 | UNWIND_ARM64_FRAME_X21_X22_PAIR = 0x00000002, |
294 | UNWIND_ARM64_FRAME_X23_X24_PAIR = 0x00000004, |
295 | UNWIND_ARM64_FRAME_X25_X26_PAIR = 0x00000008, |
296 | UNWIND_ARM64_FRAME_X27_X28_PAIR = 0x00000010, |
297 | UNWIND_ARM64_FRAME_D8_D9_PAIR = 0x00000100, |
298 | UNWIND_ARM64_FRAME_D10_D11_PAIR = 0x00000200, |
299 | UNWIND_ARM64_FRAME_D12_D13_PAIR = 0x00000400, |
300 | UNWIND_ARM64_FRAME_D14_D15_PAIR = 0x00000800, |
301 | |
302 | UNWIND_ARM64_FRAMELESS_STACK_SIZE_MASK = 0x00FFF000, |
303 | UNWIND_ARM64_DWARF_SECTION_OFFSET = 0x00FFFFFF, |
304 | }; |
305 | // For arm64 there are three modes for the compact unwind encoding: |
306 | // UNWIND_ARM64_MODE_FRAME: |
307 | // This is a standard arm64 prolog where FP/LR are immediately pushed on the |
308 | // stack, then SP is copied to FP. If there are any non-volatile registers |
309 | // saved, then are copied into the stack frame in pairs in a contiguous |
310 | // range right below the saved FP/LR pair. Any subset of the five X pairs |
311 | // and four D pairs can be saved, but the memory layout must be in register |
312 | // number order. |
313 | // UNWIND_ARM64_MODE_FRAMELESS: |
314 | // A "frameless" leaf function, where FP/LR are not saved. The return address |
315 | // remains in LR throughout the function. If any non-volatile registers |
316 | // are saved, they must be pushed onto the stack before any stack space is |
317 | // allocated for local variables. The stack sized (including any saved |
318 | // non-volatile registers) divided by 16 is encoded in the bits |
319 | // UNWIND_ARM64_FRAMELESS_STACK_SIZE_MASK. |
320 | // UNWIND_ARM64_MODE_DWARF: |
321 | // No compact unwind encoding is available. Instead the low 24-bits of the |
322 | // compact encoding is the offset of the DWARF FDE in the __eh_frame section. |
323 | // This mode is never used in object files. It is only generated by the |
324 | // linker in final linked images which have only DWARF unwind info for a |
325 | // function. |
326 | // |
327 | |
328 | |
329 | |
330 | |
331 | |
332 | //////////////////////////////////////////////////////////////////////////////// |
333 | // |
334 | // Relocatable Object Files: __LD,__compact_unwind |
335 | // |
336 | //////////////////////////////////////////////////////////////////////////////// |
337 | |
338 | // |
339 | // A compiler can generated compact unwind information for a function by adding |
340 | // a "row" to the __LD,__compact_unwind section. This section has the |
341 | // S_ATTR_DEBUG bit set, so the section will be ignored by older linkers. |
342 | // It is removed by the new linker, so never ends up in final executables. |
343 | // This section is a table, initially with one row per function (that needs |
344 | // unwind info). The table columns and some conceptual entries are: |
345 | // |
346 | // range-start pointer to start of function/range |
347 | // range-length |
348 | // compact-unwind-encoding 32-bit encoding |
349 | // personality-function or zero if no personality function |
350 | // lsda or zero if no LSDA data |
351 | // |
352 | // The length and encoding fields are 32-bits. The other are all pointer sized. |
353 | // |
354 | // In x86_64 assembly, these entry would look like: |
355 | // |
356 | // .section __LD,__compact_unwind,regular,debug |
357 | // |
358 | // #compact unwind for _foo |
359 | // .quad _foo |
360 | // .set L1,LfooEnd-_foo |
361 | // .long L1 |
362 | // .long 0x01010001 |
363 | // .quad 0 |
364 | // .quad 0 |
365 | // |
366 | // #compact unwind for _bar |
367 | // .quad _bar |
368 | // .set L2,LbarEnd-_bar |
369 | // .long L2 |
370 | // .long 0x01020011 |
371 | // .quad __gxx_personality |
372 | // .quad except_tab1 |
373 | // |
374 | // |
375 | // Notes: There is no need for any labels in the the __compact_unwind section. |
376 | // The use of the .set directive is to force the evaluation of the |
377 | // range-length at assembly time, instead of generating relocations. |
378 | // |
379 | // To support future compiler optimizations where which non-volatile registers |
380 | // are saved changes within a function (e.g. delay saving non-volatiles until |
381 | // necessary), there can by multiple lines in the __compact_unwind table for one |
382 | // function, each with a different (non-overlapping) range and each with |
383 | // different compact unwind encodings that correspond to the non-volatiles |
384 | // saved at that range of the function. |
385 | // |
386 | // If a particular function is so wacky that there is no compact unwind way |
387 | // to encode it, then the compiler can emit traditional DWARF unwind info. |
388 | // The runtime will use which ever is available. |
389 | // |
390 | // Runtime support for compact unwind encodings are only available on 10.6 |
391 | // and later. So, the compiler should not generate it when targeting pre-10.6. |
392 | |
393 | |
394 | |
395 | |
396 | //////////////////////////////////////////////////////////////////////////////// |
397 | // |
398 | // Final Linked Images: __TEXT,__unwind_info |
399 | // |
400 | //////////////////////////////////////////////////////////////////////////////// |
401 | |
402 | // |
403 | // The __TEXT,__unwind_info section is laid out for an efficient two level lookup. |
404 | // The header of the section contains a coarse index that maps function address |
405 | // to the page (4096 byte block) containing the unwind info for that function. |
406 | // |
407 | |
408 | #define UNWIND_SECTION_VERSION 1 |
409 | struct |
410 | { |
411 | uint32_t ; // UNWIND_SECTION_VERSION |
412 | uint32_t ; |
413 | uint32_t ; |
414 | uint32_t ; |
415 | uint32_t ; |
416 | uint32_t ; |
417 | uint32_t ; |
418 | // compact_unwind_encoding_t[] |
419 | // uint32_t personalities[] |
420 | // unwind_info_section_header_index_entry[] |
421 | // unwind_info_section_header_lsda_index_entry[] |
422 | }; |
423 | |
424 | struct |
425 | { |
426 | uint32_t ; |
427 | uint32_t ; // section offset to start of regular or compress page |
428 | uint32_t ; // section offset to start of lsda_index array for this range |
429 | }; |
430 | |
431 | struct |
432 | { |
433 | uint32_t ; |
434 | uint32_t ; |
435 | }; |
436 | |
437 | // |
438 | // There are two kinds of second level index pages: regular and compressed. |
439 | // A compressed page can hold up to 1021 entries, but it cannot be used |
440 | // if too many different encoding types are used. The regular page holds |
441 | // 511 entries. |
442 | // |
443 | |
444 | struct unwind_info_regular_second_level_entry |
445 | { |
446 | uint32_t functionOffset; |
447 | compact_unwind_encoding_t encoding; |
448 | }; |
449 | |
450 | #define UNWIND_SECOND_LEVEL_REGULAR 2 |
451 | struct |
452 | { |
453 | uint32_t ; // UNWIND_SECOND_LEVEL_REGULAR |
454 | uint16_t ; |
455 | uint16_t ; |
456 | // entry array |
457 | }; |
458 | |
459 | #define UNWIND_SECOND_LEVEL_COMPRESSED 3 |
460 | struct |
461 | { |
462 | uint32_t ; // UNWIND_SECOND_LEVEL_COMPRESSED |
463 | uint16_t ; |
464 | uint16_t ; |
465 | uint16_t ; |
466 | uint16_t ; |
467 | // 32-bit entry array |
468 | // encodings array |
469 | }; |
470 | |
471 | #define UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(entry) (entry & 0x00FFFFFF) |
472 | #define UNWIND_INFO_COMPRESSED_ENTRY_ENCODING_INDEX(entry) ((entry >> 24) & 0xFF) |
473 | |
474 | |
475 | |
476 | #endif |
477 | |
478 | |