1 // Copyright (c) 2010 Google Inc. All Rights Reserved.
2 //
3 // Redistribution and use in source and binary forms, with or without
4 // modification, are permitted provided that the following conditions are
5 // met:
6 //
7 //     * Redistributions of source code must retain the above copyright
8 // notice, this list of conditions and the following disclaimer.
9 //     * Redistributions in binary form must reproduce the above
10 // copyright notice, this list of conditions and the following disclaimer
11 // in the documentation and/or other materials provided with the
12 // distribution.
13 //     * Neither the name of Google Inc. nor the names of its
14 // contributors may be used to endorse or promote products derived from
15 // this software without specific prior written permission.
16 //
17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
18 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
19 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
20 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
21 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
22 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
23 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 
29 // CFI reader author: Jim Blandy <jimb@mozilla.com> <jimb@red-bean.com>
30 
31 // Implementation of dwarf2reader::LineInfo, dwarf2reader::CompilationUnit,
32 // and dwarf2reader::CallFrameInfo. See dwarf2reader.h for details.
33 
34 #include "common/dwarf/dwarf2reader.h"
35 
36 #include <assert.h>
37 #include <stdint.h>
38 #include <stdio.h>
39 #include <string.h>
40 
41 #include <map>
42 #include <memory>
43 #include <stack>
44 #include <string>
45 #include <utility>
46 
47 #include "common/dwarf/bytereader-inl.h"
48 #include "common/dwarf/bytereader.h"
49 #include "common/dwarf/line_state_machine.h"
50 #include "common/using_std_string.h"
51 
52 namespace dwarf2reader {
53 
CompilationUnit(const SectionMap & sections,uint64 offset,ByteReader * reader,Dwarf2Handler * handler)54 CompilationUnit::CompilationUnit(const SectionMap& sections, uint64 offset,
55                                  ByteReader* reader, Dwarf2Handler* handler)
56     : offset_from_section_start_(offset), reader_(reader),
57       sections_(sections), handler_(handler), abbrevs_(NULL),
58       string_buffer_(NULL), string_buffer_length_(0) {}
59 
60 // Read a DWARF2/3 abbreviation section.
61 // Each abbrev consists of a abbreviation number, a tag, a byte
62 // specifying whether the tag has children, and a list of
63 // attribute/form pairs.
64 // The list of forms is terminated by a 0 for the attribute, and a
65 // zero for the form.  The entire abbreviation section is terminated
66 // by a zero for the code.
67 
ReadAbbrevs()68 void CompilationUnit::ReadAbbrevs() {
69   if (abbrevs_)
70     return;
71 
72   // First get the debug_abbrev section.  ".debug_abbrev" is the name
73   // recommended in the DWARF spec, and used on Linux;
74   // "__debug_abbrev" is the name used in Mac OS X Mach-O files.
75   SectionMap::const_iterator iter = sections_.find(".debug_abbrev");
76   if (iter == sections_.end())
77     iter = sections_.find("__debug_abbrev");
78   assert(iter != sections_.end());
79 
80   abbrevs_ = new std::vector<Abbrev>;
81   abbrevs_->resize(1);
82 
83   // The only way to check whether we are reading over the end of the
84   // buffer would be to first compute the size of the leb128 data by
85   // reading it, then go back and read it again.
86   const char* abbrev_start = iter->second.first +
87                                       header_.abbrev_offset;
88   const char* abbrevptr = abbrev_start;
89 #ifndef NDEBUG
90   const uint64 abbrev_length = iter->second.second - header_.abbrev_offset;
91 #endif
92 
93   while (1) {
94     CompilationUnit::Abbrev abbrev;
95     size_t len;
96     const uint64 number = reader_->ReadUnsignedLEB128(abbrevptr, &len);
97 
98     if (number == 0)
99       break;
100     abbrev.number = number;
101     abbrevptr += len;
102 
103     assert(abbrevptr < abbrev_start + abbrev_length);
104     const uint64 tag = reader_->ReadUnsignedLEB128(abbrevptr, &len);
105     abbrevptr += len;
106     abbrev.tag = static_cast<enum DwarfTag>(tag);
107 
108     assert(abbrevptr < abbrev_start + abbrev_length);
109     abbrev.has_children = reader_->ReadOneByte(abbrevptr);
110     abbrevptr += 1;
111 
112     assert(abbrevptr < abbrev_start + abbrev_length);
113 
114     while (1) {
115       const uint64 nametemp = reader_->ReadUnsignedLEB128(abbrevptr, &len);
116       abbrevptr += len;
117 
118       assert(abbrevptr < abbrev_start + abbrev_length);
119       const uint64 formtemp = reader_->ReadUnsignedLEB128(abbrevptr, &len);
120       abbrevptr += len;
121       if (nametemp == 0 && formtemp == 0)
122         break;
123 
124       const enum DwarfAttribute name =
125         static_cast<enum DwarfAttribute>(nametemp);
126       const enum DwarfForm form = static_cast<enum DwarfForm>(formtemp);
127       abbrev.attributes.push_back(std::make_pair(name, form));
128     }
129     assert(abbrev.number == abbrevs_->size());
130     abbrevs_->push_back(abbrev);
131   }
132 }
133 
134 // Skips a single DIE's attributes.
SkipDIE(const char * start,const Abbrev & abbrev)135 const char* CompilationUnit::SkipDIE(const char* start,
136                                               const Abbrev& abbrev) {
137   for (AttributeList::const_iterator i = abbrev.attributes.begin();
138        i != abbrev.attributes.end();
139        i++)  {
140     start = SkipAttribute(start, i->second);
141   }
142   return start;
143 }
144 
145 // Skips a single attribute form's data.
SkipAttribute(const char * start,enum DwarfForm form)146 const char* CompilationUnit::SkipAttribute(const char* start,
147                                                     enum DwarfForm form) {
148   size_t len;
149 
150   switch (form) {
151     case DW_FORM_indirect:
152       form = static_cast<enum DwarfForm>(reader_->ReadUnsignedLEB128(start,
153                                                                      &len));
154       start += len;
155       return SkipAttribute(start, form);
156 
157     case DW_FORM_flag_present:
158       return start;
159     case DW_FORM_data1:
160     case DW_FORM_flag:
161     case DW_FORM_ref1:
162       return start + 1;
163     case DW_FORM_ref2:
164     case DW_FORM_data2:
165       return start + 2;
166     case DW_FORM_ref4:
167     case DW_FORM_data4:
168       return start + 4;
169     case DW_FORM_ref8:
170     case DW_FORM_data8:
171     case DW_FORM_ref_sig8:
172       return start + 8;
173     case DW_FORM_string:
174       return start + strlen(start) + 1;
175     case DW_FORM_udata:
176     case DW_FORM_ref_udata:
177       reader_->ReadUnsignedLEB128(start, &len);
178       return start + len;
179 
180     case DW_FORM_sdata:
181       reader_->ReadSignedLEB128(start, &len);
182       return start + len;
183     case DW_FORM_addr:
184       return start + reader_->AddressSize();
185     case DW_FORM_ref_addr:
186       // DWARF2 and 3/4 differ on whether ref_addr is address size or
187       // offset size.
188       assert(header_.version >= 2);
189       if (header_.version == 2) {
190         return start + reader_->AddressSize();
191       } else if (header_.version >= 3) {
192         return start + reader_->OffsetSize();
193       }
194       break;
195 
196     case DW_FORM_block1:
197       return start + 1 + reader_->ReadOneByte(start);
198     case DW_FORM_block2:
199       return start + 2 + reader_->ReadTwoBytes(start);
200     case DW_FORM_block4:
201       return start + 4 + reader_->ReadFourBytes(start);
202     case DW_FORM_block:
203     case DW_FORM_exprloc: {
204       uint64 size = reader_->ReadUnsignedLEB128(start, &len);
205       return start + size + len;
206     }
207     case DW_FORM_strp:
208     case DW_FORM_sec_offset:
209       return start + reader_->OffsetSize();
210   }
211   fprintf(stderr,"Unhandled form type");
212   return NULL;
213 }
214 
215 // Read a DWARF2/3 header.
216 // The header is variable length in DWARF3 (and DWARF2 as extended by
217 // most compilers), and consists of an length field, a version number,
218 // the offset in the .debug_abbrev section for our abbrevs, and an
219 // address size.
ReadHeader()220 void CompilationUnit::ReadHeader() {
221   const char* headerptr = buffer_;
222   size_t initial_length_size;
223 
224   assert(headerptr + 4 < buffer_ + buffer_length_);
225   const uint64 initial_length
226     = reader_->ReadInitialLength(headerptr, &initial_length_size);
227   headerptr += initial_length_size;
228   header_.length = initial_length;
229 
230   assert(headerptr + 2 < buffer_ + buffer_length_);
231   header_.version = reader_->ReadTwoBytes(headerptr);
232   headerptr += 2;
233 
234   assert(headerptr + reader_->OffsetSize() < buffer_ + buffer_length_);
235   header_.abbrev_offset = reader_->ReadOffset(headerptr);
236   headerptr += reader_->OffsetSize();
237 
238   assert(headerptr + 1 < buffer_ + buffer_length_);
239   header_.address_size = reader_->ReadOneByte(headerptr);
240   reader_->SetAddressSize(header_.address_size);
241   headerptr += 1;
242 
243   after_header_ = headerptr;
244 
245   // This check ensures that we don't have to do checking during the
246   // reading of DIEs. header_.length does not include the size of the
247   // initial length.
248   assert(buffer_ + initial_length_size + header_.length <=
249         buffer_ + buffer_length_);
250 }
251 
Start()252 uint64 CompilationUnit::Start() {
253   // First get the debug_info section.  ".debug_info" is the name
254   // recommended in the DWARF spec, and used on Linux; "__debug_info"
255   // is the name used in Mac OS X Mach-O files.
256   SectionMap::const_iterator iter = sections_.find(".debug_info");
257   if (iter == sections_.end())
258     iter = sections_.find("__debug_info");
259   assert(iter != sections_.end());
260 
261   // Set up our buffer
262   buffer_ = iter->second.first + offset_from_section_start_;
263   buffer_length_ = iter->second.second - offset_from_section_start_;
264 
265   // Read the header
266   ReadHeader();
267 
268   // Figure out the real length from the end of the initial length to
269   // the end of the compilation unit, since that is the value we
270   // return.
271   uint64 ourlength = header_.length;
272   if (reader_->OffsetSize() == 8)
273     ourlength += 12;
274   else
275     ourlength += 4;
276 
277   // See if the user wants this compilation unit, and if not, just return.
278   if (!handler_->StartCompilationUnit(offset_from_section_start_,
279                                       reader_->AddressSize(),
280                                       reader_->OffsetSize(),
281                                       header_.length,
282                                       header_.version))
283     return ourlength;
284 
285   // Otherwise, continue by reading our abbreviation entries.
286   ReadAbbrevs();
287 
288   // Set the string section if we have one.  ".debug_str" is the name
289   // recommended in the DWARF spec, and used on Linux; "__debug_str"
290   // is the name used in Mac OS X Mach-O files.
291   iter = sections_.find(".debug_str");
292   if (iter == sections_.end())
293     iter = sections_.find("__debug_str");
294   if (iter != sections_.end()) {
295     string_buffer_ = iter->second.first;
296     string_buffer_length_ = iter->second.second;
297   }
298 
299   // Now that we have our abbreviations, start processing DIE's.
300   ProcessDIEs();
301 
302   return ourlength;
303 }
304 
305 // If one really wanted, you could merge SkipAttribute and
306 // ProcessAttribute
307 // This is all boring data manipulation and calling of the handler.
ProcessAttribute(uint64 dieoffset,const char * start,enum DwarfAttribute attr,enum DwarfForm form)308 const char* CompilationUnit::ProcessAttribute(
309     uint64 dieoffset, const char* start, enum DwarfAttribute attr,
310     enum DwarfForm form) {
311   size_t len;
312 
313   switch (form) {
314     // DW_FORM_indirect is never used because it is such a space
315     // waster.
316     case DW_FORM_indirect:
317       form = static_cast<enum DwarfForm>(reader_->ReadUnsignedLEB128(start,
318                                                                      &len));
319       start += len;
320       return ProcessAttribute(dieoffset, start, attr, form);
321 
322     case DW_FORM_flag_present:
323       handler_->ProcessAttributeUnsigned(dieoffset, attr, form, 1);
324       return start;
325     case DW_FORM_data1:
326     case DW_FORM_flag:
327       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
328                                          reader_->ReadOneByte(start));
329       return start + 1;
330     case DW_FORM_data2:
331       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
332                                          reader_->ReadTwoBytes(start));
333       return start + 2;
334     case DW_FORM_data4:
335       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
336                                          reader_->ReadFourBytes(start));
337       return start + 4;
338     case DW_FORM_data8:
339       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
340                                          reader_->ReadEightBytes(start));
341       return start + 8;
342     case DW_FORM_string: {
343       const char* str = start;
344       handler_->ProcessAttributeString(dieoffset, attr, form,
345                                        str);
346       return start + strlen(str) + 1;
347     }
348     case DW_FORM_udata:
349       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
350                                          reader_->ReadUnsignedLEB128(start,
351                                                                      &len));
352       return start + len;
353 
354     case DW_FORM_sdata:
355       handler_->ProcessAttributeSigned(dieoffset, attr, form,
356                                       reader_->ReadSignedLEB128(start, &len));
357       return start + len;
358     case DW_FORM_addr:
359       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
360                                          reader_->ReadAddress(start));
361       return start + reader_->AddressSize();
362     case DW_FORM_sec_offset:
363       handler_->ProcessAttributeUnsigned(dieoffset, attr, form,
364                                          reader_->ReadOffset(start));
365       return start + reader_->OffsetSize();
366 
367     case DW_FORM_ref1:
368       handler_->ProcessAttributeReference(dieoffset, attr, form,
369                                           reader_->ReadOneByte(start)
370                                           + offset_from_section_start_);
371       return start + 1;
372     case DW_FORM_ref2:
373       handler_->ProcessAttributeReference(dieoffset, attr, form,
374                                           reader_->ReadTwoBytes(start)
375                                           + offset_from_section_start_);
376       return start + 2;
377     case DW_FORM_ref4:
378       handler_->ProcessAttributeReference(dieoffset, attr, form,
379                                           reader_->ReadFourBytes(start)
380                                           + offset_from_section_start_);
381       return start + 4;
382     case DW_FORM_ref8:
383       handler_->ProcessAttributeReference(dieoffset, attr, form,
384                                           reader_->ReadEightBytes(start)
385                                           + offset_from_section_start_);
386       return start + 8;
387     case DW_FORM_ref_udata:
388       handler_->ProcessAttributeReference(dieoffset, attr, form,
389                                           reader_->ReadUnsignedLEB128(start,
390                                                                       &len)
391                                           + offset_from_section_start_);
392       return start + len;
393     case DW_FORM_ref_addr:
394       // DWARF2 and 3/4 differ on whether ref_addr is address size or
395       // offset size.
396       assert(header_.version >= 2);
397       if (header_.version == 2) {
398         handler_->ProcessAttributeReference(dieoffset, attr, form,
399                                             reader_->ReadAddress(start));
400         return start + reader_->AddressSize();
401       } else if (header_.version >= 3) {
402         handler_->ProcessAttributeReference(dieoffset, attr, form,
403                                             reader_->ReadOffset(start));
404         return start + reader_->OffsetSize();
405       }
406       break;
407     case DW_FORM_ref_sig8:
408       handler_->ProcessAttributeSignature(dieoffset, attr, form,
409                                           reader_->ReadEightBytes(start));
410       return start + 8;
411 
412     case DW_FORM_block1: {
413       uint64 datalen = reader_->ReadOneByte(start);
414       handler_->ProcessAttributeBuffer(dieoffset, attr, form, start + 1,
415                                        datalen);
416       return start + 1 + datalen;
417     }
418     case DW_FORM_block2: {
419       uint64 datalen = reader_->ReadTwoBytes(start);
420       handler_->ProcessAttributeBuffer(dieoffset, attr, form, start + 2,
421                                        datalen);
422       return start + 2 + datalen;
423     }
424     case DW_FORM_block4: {
425       uint64 datalen = reader_->ReadFourBytes(start);
426       handler_->ProcessAttributeBuffer(dieoffset, attr, form, start + 4,
427                                        datalen);
428       return start + 4 + datalen;
429     }
430     case DW_FORM_block:
431     case DW_FORM_exprloc: {
432       uint64 datalen = reader_->ReadUnsignedLEB128(start, &len);
433       handler_->ProcessAttributeBuffer(dieoffset, attr, form, start + len,
434                                        datalen);
435       return start + datalen + len;
436     }
437     case DW_FORM_strp: {
438       assert(string_buffer_ != NULL);
439 
440       const uint64 offset = reader_->ReadOffset(start);
441       assert(string_buffer_ + offset < string_buffer_ + string_buffer_length_);
442 
443       const char* str = string_buffer_ + offset;
444       handler_->ProcessAttributeString(dieoffset, attr, form,
445                                        str);
446       return start + reader_->OffsetSize();
447     }
448   }
449   fprintf(stderr, "Unhandled form type\n");
450   return NULL;
451 }
452 
ProcessDIE(uint64 dieoffset,const char * start,const Abbrev & abbrev)453 const char* CompilationUnit::ProcessDIE(uint64 dieoffset,
454                                                  const char* start,
455                                                  const Abbrev& abbrev) {
456   for (AttributeList::const_iterator i = abbrev.attributes.begin();
457        i != abbrev.attributes.end();
458        i++)  {
459     start = ProcessAttribute(dieoffset, start, i->first, i->second);
460   }
461   return start;
462 }
463 
ProcessDIEs()464 void CompilationUnit::ProcessDIEs() {
465   const char* dieptr = after_header_;
466   size_t len;
467 
468   // lengthstart is the place the length field is based on.
469   // It is the point in the header after the initial length field
470   const char* lengthstart = buffer_;
471 
472   // In 64 bit dwarf, the initial length is 12 bytes, because of the
473   // 0xffffffff at the start.
474   if (reader_->OffsetSize() == 8)
475     lengthstart += 12;
476   else
477     lengthstart += 4;
478 
479   std::stack<uint64> die_stack;
480 
481   while (dieptr < (lengthstart + header_.length)) {
482     // We give the user the absolute offset from the beginning of
483     // debug_info, since they need it to deal with ref_addr forms.
484     uint64 absolute_offset = (dieptr - buffer_) + offset_from_section_start_;
485 
486     uint64 abbrev_num = reader_->ReadUnsignedLEB128(dieptr, &len);
487 
488     dieptr += len;
489 
490     // Abbrev == 0 represents the end of a list of children, or padding
491     // at the end of the compilation unit.
492     if (abbrev_num == 0) {
493       if (die_stack.size() == 0)
494         // If it is padding, then we are done with the compilation unit's DIEs.
495         return;
496       const uint64 offset = die_stack.top();
497       die_stack.pop();
498       handler_->EndDIE(offset);
499       continue;
500     }
501 
502     const Abbrev& abbrev = abbrevs_->at(static_cast<size_t>(abbrev_num));
503     const enum DwarfTag tag = abbrev.tag;
504     if (!handler_->StartDIE(absolute_offset, tag)) {
505       dieptr = SkipDIE(dieptr, abbrev);
506     } else {
507       dieptr = ProcessDIE(absolute_offset, dieptr, abbrev);
508     }
509 
510     if (abbrev.has_children) {
511       die_stack.push(absolute_offset);
512     } else {
513       handler_->EndDIE(absolute_offset);
514     }
515   }
516 }
517 
LineInfo(const char * buffer,uint64 buffer_length,ByteReader * reader,LineInfoHandler * handler)518 LineInfo::LineInfo(const char* buffer, uint64 buffer_length,
519                    ByteReader* reader, LineInfoHandler* handler):
520     handler_(handler), reader_(reader), buffer_(buffer),
521     buffer_length_(buffer_length) {
522   header_.std_opcode_lengths = NULL;
523 }
524 
Start()525 uint64 LineInfo::Start() {
526   ReadHeader();
527   ReadLines();
528   return after_header_ - buffer_;
529 }
530 
531 // The header for a debug_line section is mildly complicated, because
532 // the line info is very tightly encoded.
ReadHeader()533 void LineInfo::ReadHeader() {
534   const char* lineptr = buffer_;
535   size_t initial_length_size;
536 
537   const uint64 initial_length
538     = reader_->ReadInitialLength(lineptr, &initial_length_size);
539 
540   lineptr += initial_length_size;
541   header_.total_length = initial_length;
542   assert(buffer_ + initial_length_size + header_.total_length <=
543         buffer_ + buffer_length_);
544 
545   // Address size *must* be set by CU ahead of time.
546   assert(reader_->AddressSize() != 0);
547 
548   header_.version = reader_->ReadTwoBytes(lineptr);
549   lineptr += 2;
550 
551   header_.prologue_length = reader_->ReadOffset(lineptr);
552   lineptr += reader_->OffsetSize();
553 
554   header_.min_insn_length = reader_->ReadOneByte(lineptr);
555   lineptr += 1;
556 
557   header_.default_is_stmt = reader_->ReadOneByte(lineptr);
558   lineptr += 1;
559 
560   header_.line_base = *reinterpret_cast<const int8*>(lineptr);
561   lineptr += 1;
562 
563   header_.line_range = reader_->ReadOneByte(lineptr);
564   lineptr += 1;
565 
566   header_.opcode_base = reader_->ReadOneByte(lineptr);
567   lineptr += 1;
568 
569   header_.std_opcode_lengths = new std::vector<unsigned char>;
570   header_.std_opcode_lengths->resize(header_.opcode_base + 1);
571   (*header_.std_opcode_lengths)[0] = 0;
572   for (int i = 1; i < header_.opcode_base; i++) {
573     (*header_.std_opcode_lengths)[i] = reader_->ReadOneByte(lineptr);
574     lineptr += 1;
575   }
576 
577   // It is legal for the directory entry table to be empty.
578   if (*lineptr) {
579     uint32 dirindex = 1;
580     while (*lineptr) {
581       const char* dirname = lineptr;
582       handler_->DefineDir(dirname, dirindex);
583       lineptr += strlen(dirname) + 1;
584       dirindex++;
585     }
586   }
587   lineptr++;
588 
589   // It is also legal for the file entry table to be empty.
590   if (*lineptr) {
591     uint32 fileindex = 1;
592     size_t len;
593     while (*lineptr) {
594       const char* filename = lineptr;
595       lineptr += strlen(filename) + 1;
596 
597       uint64 dirindex = reader_->ReadUnsignedLEB128(lineptr, &len);
598       lineptr += len;
599 
600       uint64 mod_time = reader_->ReadUnsignedLEB128(lineptr, &len);
601       lineptr += len;
602 
603       uint64 filelength = reader_->ReadUnsignedLEB128(lineptr, &len);
604       lineptr += len;
605       handler_->DefineFile(filename, fileindex, static_cast<uint32>(dirindex),
606                            mod_time, filelength);
607       fileindex++;
608     }
609   }
610   lineptr++;
611 
612   after_header_ = lineptr;
613 }
614 
615 /* static */
ProcessOneOpcode(ByteReader * reader,LineInfoHandler * handler,const struct LineInfoHeader & header,const char * start,struct LineStateMachine * lsm,size_t * len,uintptr pc,bool * lsm_passes_pc)616 bool LineInfo::ProcessOneOpcode(ByteReader* reader,
617                                 LineInfoHandler* handler,
618                                 const struct LineInfoHeader &header,
619                                 const char* start,
620                                 struct LineStateMachine* lsm,
621                                 size_t* len,
622                                 uintptr pc,
623                                 bool *lsm_passes_pc) {
624   size_t oplen = 0;
625   size_t templen;
626   uint8 opcode = reader->ReadOneByte(start);
627   oplen++;
628   start++;
629 
630   // If the opcode is great than the opcode_base, it is a special
631   // opcode. Most line programs consist mainly of special opcodes.
632   if (opcode >= header.opcode_base) {
633     opcode -= header.opcode_base;
634     const int64 advance_address = (opcode / header.line_range)
635                                   * header.min_insn_length;
636     const int32 advance_line = (opcode % header.line_range)
637                                + header.line_base;
638 
639     // Check if the lsm passes "pc". If so, mark it as passed.
640     if (lsm_passes_pc &&
641         lsm->address <= pc && pc < lsm->address + advance_address) {
642       *lsm_passes_pc = true;
643     }
644 
645     lsm->address += advance_address;
646     lsm->line_num += advance_line;
647     lsm->basic_block = true;
648     *len = oplen;
649     return true;
650   }
651 
652   // Otherwise, we have the regular opcodes
653   switch (opcode) {
654     case DW_LNS_copy: {
655       lsm->basic_block = false;
656       *len = oplen;
657       return true;
658     }
659 
660     case DW_LNS_advance_pc: {
661       uint64 advance_address = reader->ReadUnsignedLEB128(start, &templen);
662       oplen += templen;
663 
664       // Check if the lsm passes "pc". If so, mark it as passed.
665       if (lsm_passes_pc && lsm->address <= pc &&
666           pc < lsm->address + header.min_insn_length * advance_address) {
667         *lsm_passes_pc = true;
668       }
669 
670       lsm->address += header.min_insn_length * advance_address;
671     }
672       break;
673     case DW_LNS_advance_line: {
674       const int64 advance_line = reader->ReadSignedLEB128(start, &templen);
675       oplen += templen;
676       lsm->line_num += static_cast<int32>(advance_line);
677 
678       // With gcc 4.2.1, we can get the line_no here for the first time
679       // since DW_LNS_advance_line is called after DW_LNE_set_address is
680       // called. So we check if the lsm passes "pc" here, not in
681       // DW_LNE_set_address.
682       if (lsm_passes_pc && lsm->address == pc) {
683         *lsm_passes_pc = true;
684       }
685     }
686       break;
687     case DW_LNS_set_file: {
688       const uint64 fileno = reader->ReadUnsignedLEB128(start, &templen);
689       oplen += templen;
690       lsm->file_num = static_cast<uint32>(fileno);
691     }
692       break;
693     case DW_LNS_set_column: {
694       const uint64 colno = reader->ReadUnsignedLEB128(start, &templen);
695       oplen += templen;
696       lsm->column_num = static_cast<uint32>(colno);
697     }
698       break;
699     case DW_LNS_negate_stmt: {
700       lsm->is_stmt = !lsm->is_stmt;
701     }
702       break;
703     case DW_LNS_set_basic_block: {
704       lsm->basic_block = true;
705     }
706       break;
707     case DW_LNS_fixed_advance_pc: {
708       const uint16 advance_address = reader->ReadTwoBytes(start);
709       oplen += 2;
710 
711       // Check if the lsm passes "pc". If so, mark it as passed.
712       if (lsm_passes_pc &&
713           lsm->address <= pc && pc < lsm->address + advance_address) {
714         *lsm_passes_pc = true;
715       }
716 
717       lsm->address += advance_address;
718     }
719       break;
720     case DW_LNS_const_add_pc: {
721       const int64 advance_address = header.min_insn_length
722                                     * ((255 - header.opcode_base)
723                                        / header.line_range);
724 
725       // Check if the lsm passes "pc". If so, mark it as passed.
726       if (lsm_passes_pc &&
727           lsm->address <= pc && pc < lsm->address + advance_address) {
728         *lsm_passes_pc = true;
729       }
730 
731       lsm->address += advance_address;
732     }
733       break;
734     case DW_LNS_extended_op: {
735       const uint64 extended_op_len = reader->ReadUnsignedLEB128(start,
736                                                                 &templen);
737       start += templen;
738       oplen += templen + extended_op_len;
739 
740       const uint64 extended_op = reader->ReadOneByte(start);
741       start++;
742 
743       switch (extended_op) {
744         case DW_LNE_end_sequence: {
745           lsm->end_sequence = true;
746           *len = oplen;
747           return true;
748         }
749           break;
750         case DW_LNE_set_address: {
751           // With gcc 4.2.1, we cannot tell the line_no here since
752           // DW_LNE_set_address is called before DW_LNS_advance_line is
753           // called.  So we do not check if the lsm passes "pc" here.  See
754           // also the comment in DW_LNS_advance_line.
755           uint64 address = reader->ReadAddress(start);
756           lsm->address = address;
757         }
758           break;
759         case DW_LNE_define_file: {
760           const char* filename  = start;
761 
762           templen = strlen(filename) + 1;
763           start += templen;
764 
765           uint64 dirindex = reader->ReadUnsignedLEB128(start, &templen);
766           oplen += templen;
767 
768           const uint64 mod_time = reader->ReadUnsignedLEB128(start,
769                                                              &templen);
770           oplen += templen;
771 
772           const uint64 filelength = reader->ReadUnsignedLEB128(start,
773                                                                &templen);
774           oplen += templen;
775 
776           if (handler) {
777             handler->DefineFile(filename, -1, static_cast<uint32>(dirindex),
778                                 mod_time, filelength);
779           }
780         }
781           break;
782       }
783     }
784       break;
785 
786     default: {
787       // Ignore unknown opcode  silently
788       if (header.std_opcode_lengths) {
789         for (int i = 0; i < (*header.std_opcode_lengths)[opcode]; i++) {
790           reader->ReadUnsignedLEB128(start, &templen);
791           start += templen;
792           oplen += templen;
793         }
794       }
795     }
796       break;
797   }
798   *len = oplen;
799   return false;
800 }
801 
ReadLines()802 void LineInfo::ReadLines() {
803   struct LineStateMachine lsm;
804 
805   // lengthstart is the place the length field is based on.
806   // It is the point in the header after the initial length field
807   const char* lengthstart = buffer_;
808 
809   // In 64 bit dwarf, the initial length is 12 bytes, because of the
810   // 0xffffffff at the start.
811   if (reader_->OffsetSize() == 8)
812     lengthstart += 12;
813   else
814     lengthstart += 4;
815 
816   const char* lineptr = after_header_;
817   lsm.Reset(header_.default_is_stmt);
818 
819   // The LineInfoHandler interface expects each line's length along
820   // with its address, but DWARF only provides addresses (sans
821   // length), and an end-of-sequence address; one infers the length
822   // from the next address. So we report a line only when we get the
823   // next line's address, or the end-of-sequence address.
824   bool have_pending_line = false;
825   uint64 pending_address = 0;
826   uint32 pending_file_num = 0, pending_line_num = 0, pending_column_num = 0;
827 
828   while (lineptr < lengthstart + header_.total_length) {
829     size_t oplength;
830     bool add_row = ProcessOneOpcode(reader_, handler_, header_,
831                                     lineptr, &lsm, &oplength, (uintptr)-1,
832                                     NULL);
833     if (add_row) {
834       if (have_pending_line)
835         handler_->AddLine(pending_address, lsm.address - pending_address,
836                           pending_file_num, pending_line_num,
837                           pending_column_num);
838       if (lsm.end_sequence) {
839         lsm.Reset(header_.default_is_stmt);
840         have_pending_line = false;
841       } else {
842         pending_address = lsm.address;
843         pending_file_num = lsm.file_num;
844         pending_line_num = lsm.line_num;
845         pending_column_num = lsm.column_num;
846         have_pending_line = true;
847       }
848     }
849     lineptr += oplength;
850   }
851 
852   after_header_ = lengthstart + header_.total_length;
853 }
854 
855 // A DWARF rule for recovering the address or value of a register, or
856 // computing the canonical frame address. There is one subclass of this for
857 // each '*Rule' member function in CallFrameInfo::Handler.
858 //
859 // It's annoying that we have to handle Rules using pointers (because
860 // the concrete instances can have an arbitrary size). They're small,
861 // so it would be much nicer if we could just handle them by value
862 // instead of fretting about ownership and destruction.
863 //
864 // It seems like all these could simply be instances of std::tr1::bind,
865 // except that we need instances to be EqualityComparable, too.
866 //
867 // This could logically be nested within State, but then the qualified names
868 // get horrendous.
869 class CallFrameInfo::Rule {
870  public:
~Rule()871   virtual ~Rule() { }
872 
873   // Tell HANDLER that, at ADDRESS in the program, REGISTER can be
874   // recovered using this rule. If REGISTER is kCFARegister, then this rule
875   // describes how to compute the canonical frame address. Return what the
876   // HANDLER member function returned.
877   virtual bool Handle(Handler *handler,
878                       uint64 address, int register) const = 0;
879 
880   // Equality on rules. We use these to decide which rules we need
881   // to report after a DW_CFA_restore_state instruction.
882   virtual bool operator==(const Rule &rhs) const = 0;
883 
operator !=(const Rule & rhs) const884   bool operator!=(const Rule &rhs) const { return ! (*this == rhs); }
885 
886   // Return a pointer to a copy of this rule.
887   virtual Rule *Copy() const = 0;
888 
889   // If this is a base+offset rule, change its base register to REG.
890   // Otherwise, do nothing. (Ugly, but required for DW_CFA_def_cfa_register.)
SetBaseRegister(unsigned reg)891   virtual void SetBaseRegister(unsigned reg) { }
892 
893   // If this is a base+offset rule, change its offset to OFFSET. Otherwise,
894   // do nothing. (Ugly, but required for DW_CFA_def_cfa_offset.)
SetOffset(long long offset)895   virtual void SetOffset(long long offset) { }
896 };
897 
898 // Rule: the value the register had in the caller cannot be recovered.
899 class CallFrameInfo::UndefinedRule: public CallFrameInfo::Rule {
900  public:
UndefinedRule()901   UndefinedRule() { }
~UndefinedRule()902   ~UndefinedRule() { }
Handle(Handler * handler,uint64 address,int reg) const903   bool Handle(Handler *handler, uint64 address, int reg) const {
904     return handler->UndefinedRule(address, reg);
905   }
operator ==(const Rule & rhs) const906   bool operator==(const Rule &rhs) const {
907     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
908     // been carefully considered; cheap RTTI-like workarounds are forbidden.
909     const UndefinedRule *our_rhs = dynamic_cast<const UndefinedRule *>(&rhs);
910     return (our_rhs != NULL);
911   }
Copy() const912   Rule *Copy() const { return new UndefinedRule(*this); }
913 };
914 
915 // Rule: the register's value is the same as that it had in the caller.
916 class CallFrameInfo::SameValueRule: public CallFrameInfo::Rule {
917  public:
SameValueRule()918   SameValueRule() { }
~SameValueRule()919   ~SameValueRule() { }
Handle(Handler * handler,uint64 address,int reg) const920   bool Handle(Handler *handler, uint64 address, int reg) const {
921     return handler->SameValueRule(address, reg);
922   }
operator ==(const Rule & rhs) const923   bool operator==(const Rule &rhs) const {
924     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
925     // been carefully considered; cheap RTTI-like workarounds are forbidden.
926     const SameValueRule *our_rhs = dynamic_cast<const SameValueRule *>(&rhs);
927     return (our_rhs != NULL);
928   }
Copy() const929   Rule *Copy() const { return new SameValueRule(*this); }
930 };
931 
932 // Rule: the register is saved at OFFSET from BASE_REGISTER.  BASE_REGISTER
933 // may be CallFrameInfo::Handler::kCFARegister.
934 class CallFrameInfo::OffsetRule: public CallFrameInfo::Rule {
935  public:
OffsetRule(int base_register,long offset)936   OffsetRule(int base_register, long offset)
937       : base_register_(base_register), offset_(offset) { }
~OffsetRule()938   ~OffsetRule() { }
Handle(Handler * handler,uint64 address,int reg) const939   bool Handle(Handler *handler, uint64 address, int reg) const {
940     return handler->OffsetRule(address, reg, base_register_, offset_);
941   }
operator ==(const Rule & rhs) const942   bool operator==(const Rule &rhs) const {
943     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
944     // been carefully considered; cheap RTTI-like workarounds are forbidden.
945     const OffsetRule *our_rhs = dynamic_cast<const OffsetRule *>(&rhs);
946     return (our_rhs &&
947             base_register_ == our_rhs->base_register_ &&
948             offset_ == our_rhs->offset_);
949   }
Copy() const950   Rule *Copy() const { return new OffsetRule(*this); }
951   // We don't actually need SetBaseRegister or SetOffset here, since they
952   // are only ever applied to CFA rules, for DW_CFA_def_cfa_offset, and it
953   // doesn't make sense to use OffsetRule for computing the CFA: it
954   // computes the address at which a register is saved, not a value.
955  private:
956   int base_register_;
957   long offset_;
958 };
959 
960 // Rule: the value the register had in the caller is the value of
961 // BASE_REGISTER plus offset. BASE_REGISTER may be
962 // CallFrameInfo::Handler::kCFARegister.
963 class CallFrameInfo::ValOffsetRule: public CallFrameInfo::Rule {
964  public:
ValOffsetRule(int base_register,long offset)965   ValOffsetRule(int base_register, long offset)
966       : base_register_(base_register), offset_(offset) { }
~ValOffsetRule()967   ~ValOffsetRule() { }
Handle(Handler * handler,uint64 address,int reg) const968   bool Handle(Handler *handler, uint64 address, int reg) const {
969     return handler->ValOffsetRule(address, reg, base_register_, offset_);
970   }
operator ==(const Rule & rhs) const971   bool operator==(const Rule &rhs) const {
972     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
973     // been carefully considered; cheap RTTI-like workarounds are forbidden.
974     const ValOffsetRule *our_rhs = dynamic_cast<const ValOffsetRule *>(&rhs);
975     return (our_rhs &&
976             base_register_ == our_rhs->base_register_ &&
977             offset_ == our_rhs->offset_);
978   }
Copy() const979   Rule *Copy() const { return new ValOffsetRule(*this); }
SetBaseRegister(unsigned reg)980   void SetBaseRegister(unsigned reg) { base_register_ = reg; }
SetOffset(long long offset)981   void SetOffset(long long offset) { offset_ = offset; }
982  private:
983   int base_register_;
984   long offset_;
985 };
986 
987 // Rule: the register has been saved in another register REGISTER_NUMBER_.
988 class CallFrameInfo::RegisterRule: public CallFrameInfo::Rule {
989  public:
RegisterRule(int register_number)990   explicit RegisterRule(int register_number)
991       : register_number_(register_number) { }
~RegisterRule()992   ~RegisterRule() { }
Handle(Handler * handler,uint64 address,int reg) const993   bool Handle(Handler *handler, uint64 address, int reg) const {
994     return handler->RegisterRule(address, reg, register_number_);
995   }
operator ==(const Rule & rhs) const996   bool operator==(const Rule &rhs) const {
997     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
998     // been carefully considered; cheap RTTI-like workarounds are forbidden.
999     const RegisterRule *our_rhs = dynamic_cast<const RegisterRule *>(&rhs);
1000     return (our_rhs && register_number_ == our_rhs->register_number_);
1001   }
Copy() const1002   Rule *Copy() const { return new RegisterRule(*this); }
1003  private:
1004   int register_number_;
1005 };
1006 
1007 // Rule: EXPRESSION evaluates to the address at which the register is saved.
1008 class CallFrameInfo::ExpressionRule: public CallFrameInfo::Rule {
1009  public:
ExpressionRule(const string & expression)1010   explicit ExpressionRule(const string &expression)
1011       : expression_(expression) { }
~ExpressionRule()1012   ~ExpressionRule() { }
Handle(Handler * handler,uint64 address,int reg) const1013   bool Handle(Handler *handler, uint64 address, int reg) const {
1014     return handler->ExpressionRule(address, reg, expression_);
1015   }
operator ==(const Rule & rhs) const1016   bool operator==(const Rule &rhs) const {
1017     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
1018     // been carefully considered; cheap RTTI-like workarounds are forbidden.
1019     const ExpressionRule *our_rhs = dynamic_cast<const ExpressionRule *>(&rhs);
1020     return (our_rhs && expression_ == our_rhs->expression_);
1021   }
Copy() const1022   Rule *Copy() const { return new ExpressionRule(*this); }
1023  private:
1024   string expression_;
1025 };
1026 
1027 // Rule: EXPRESSION evaluates to the address at which the register is saved.
1028 class CallFrameInfo::ValExpressionRule: public CallFrameInfo::Rule {
1029  public:
ValExpressionRule(const string & expression)1030   explicit ValExpressionRule(const string &expression)
1031       : expression_(expression) { }
~ValExpressionRule()1032   ~ValExpressionRule() { }
Handle(Handler * handler,uint64 address,int reg) const1033   bool Handle(Handler *handler, uint64 address, int reg) const {
1034     return handler->ValExpressionRule(address, reg, expression_);
1035   }
operator ==(const Rule & rhs) const1036   bool operator==(const Rule &rhs) const {
1037     // dynamic_cast is allowed by the Google C++ Style Guide, if the use has
1038     // been carefully considered; cheap RTTI-like workarounds are forbidden.
1039     const ValExpressionRule *our_rhs =
1040         dynamic_cast<const ValExpressionRule *>(&rhs);
1041     return (our_rhs && expression_ == our_rhs->expression_);
1042   }
Copy() const1043   Rule *Copy() const { return new ValExpressionRule(*this); }
1044  private:
1045   string expression_;
1046 };
1047 
1048 // A map from register numbers to rules.
1049 class CallFrameInfo::RuleMap {
1050  public:
RuleMap()1051   RuleMap() : cfa_rule_(NULL) { }
RuleMap(const RuleMap & rhs)1052   RuleMap(const RuleMap &rhs) : cfa_rule_(NULL) { *this = rhs; }
~RuleMap()1053   ~RuleMap() { Clear(); }
1054 
1055   RuleMap &operator=(const RuleMap &rhs);
1056 
1057   // Set the rule for computing the CFA to RULE. Take ownership of RULE.
SetCFARule(Rule * rule)1058   void SetCFARule(Rule *rule) { delete cfa_rule_; cfa_rule_ = rule; }
1059 
1060   // Return the current CFA rule. Unlike RegisterRule, this RuleMap retains
1061   // ownership of the rule. We use this for DW_CFA_def_cfa_offset and
1062   // DW_CFA_def_cfa_register, and for detecting references to the CFA before
1063   // a rule for it has been established.
CFARule() const1064   Rule *CFARule() const { return cfa_rule_; }
1065 
1066   // Return the rule for REG, or NULL if there is none. The caller takes
1067   // ownership of the result.
1068   Rule *RegisterRule(int reg) const;
1069 
1070   // Set the rule for computing REG to RULE. Take ownership of RULE.
1071   void SetRegisterRule(int reg, Rule *rule);
1072 
1073   // Make all the appropriate calls to HANDLER as if we were changing from
1074   // this RuleMap to NEW_RULES at ADDRESS. We use this to implement
1075   // DW_CFA_restore_state, where lots of rules can change simultaneously.
1076   // Return true if all handlers returned true; otherwise, return false.
1077   bool HandleTransitionTo(Handler *handler, uint64 address,
1078                           const RuleMap &new_rules) const;
1079 
1080  private:
1081   // A map from register numbers to Rules.
1082   typedef std::map<int, Rule *> RuleByNumber;
1083 
1084   // Remove all register rules and clear cfa_rule_.
1085   void Clear();
1086 
1087   // The rule for computing the canonical frame address. This RuleMap owns
1088   // this rule.
1089   Rule *cfa_rule_;
1090 
1091   // A map from register numbers to postfix expressions to recover
1092   // their values. This RuleMap owns the Rules the map refers to.
1093   RuleByNumber registers_;
1094 };
1095 
operator =(const RuleMap & rhs)1096 CallFrameInfo::RuleMap &CallFrameInfo::RuleMap::operator=(const RuleMap &rhs) {
1097   Clear();
1098   // Since each map owns the rules it refers to, assignment must copy them.
1099   if (rhs.cfa_rule_) cfa_rule_ = rhs.cfa_rule_->Copy();
1100   for (RuleByNumber::const_iterator it = rhs.registers_.begin();
1101        it != rhs.registers_.end(); it++)
1102     registers_[it->first] = it->second->Copy();
1103   return *this;
1104 }
1105 
RegisterRule(int reg) const1106 CallFrameInfo::Rule *CallFrameInfo::RuleMap::RegisterRule(int reg) const {
1107   assert(reg != Handler::kCFARegister);
1108   RuleByNumber::const_iterator it = registers_.find(reg);
1109   if (it != registers_.end())
1110     return it->second->Copy();
1111   else
1112     return NULL;
1113 }
1114 
SetRegisterRule(int reg,Rule * rule)1115 void CallFrameInfo::RuleMap::SetRegisterRule(int reg, Rule *rule) {
1116   assert(reg != Handler::kCFARegister);
1117   assert(rule);
1118   Rule **slot = &registers_[reg];
1119   delete *slot;
1120   *slot = rule;
1121 }
1122 
HandleTransitionTo(Handler * handler,uint64 address,const RuleMap & new_rules) const1123 bool CallFrameInfo::RuleMap::HandleTransitionTo(
1124     Handler *handler,
1125     uint64 address,
1126     const RuleMap &new_rules) const {
1127   // Transition from cfa_rule_ to new_rules.cfa_rule_.
1128   if (cfa_rule_ && new_rules.cfa_rule_) {
1129     if (*cfa_rule_ != *new_rules.cfa_rule_ &&
1130         !new_rules.cfa_rule_->Handle(handler, address,
1131                                      Handler::kCFARegister))
1132       return false;
1133   } else if (cfa_rule_) {
1134     // this RuleMap has a CFA rule but new_rules doesn't.
1135     // CallFrameInfo::Handler has no way to handle this --- and shouldn't;
1136     // it's garbage input. The instruction interpreter should have
1137     // detected this and warned, so take no action here.
1138   } else if (new_rules.cfa_rule_) {
1139     // This shouldn't be possible: NEW_RULES is some prior state, and
1140     // there's no way to remove entries.
1141     assert(0);
1142   } else {
1143     // Both CFA rules are empty.  No action needed.
1144   }
1145 
1146   // Traverse the two maps in order by register number, and report
1147   // whatever differences we find.
1148   RuleByNumber::const_iterator old_it = registers_.begin();
1149   RuleByNumber::const_iterator new_it = new_rules.registers_.begin();
1150   while (old_it != registers_.end() && new_it != new_rules.registers_.end()) {
1151     if (old_it->first < new_it->first) {
1152       // This RuleMap has an entry for old_it->first, but NEW_RULES
1153       // doesn't.
1154       //
1155       // This isn't really the right thing to do, but since CFI generally
1156       // only mentions callee-saves registers, and GCC's convention for
1157       // callee-saves registers is that they are unchanged, it's a good
1158       // approximation.
1159       if (!handler->SameValueRule(address, old_it->first))
1160         return false;
1161       old_it++;
1162     } else if (old_it->first > new_it->first) {
1163       // NEW_RULES has entry for new_it->first, but this RuleMap
1164       // doesn't. This shouldn't be possible: NEW_RULES is some prior
1165       // state, and there's no way to remove entries.
1166       assert(0);
1167     } else {
1168       // Both maps have an entry for this register. Report the new
1169       // rule if it is different.
1170       if (*old_it->second != *new_it->second &&
1171           !new_it->second->Handle(handler, address, new_it->first))
1172         return false;
1173       new_it++, old_it++;
1174     }
1175   }
1176   // Finish off entries from this RuleMap with no counterparts in new_rules.
1177   while (old_it != registers_.end()) {
1178     if (!handler->SameValueRule(address, old_it->first))
1179       return false;
1180     old_it++;
1181   }
1182   // Since we only make transitions from a rule set to some previously
1183   // saved rule set, and we can only add rules to the map, NEW_RULES
1184   // must have fewer rules than *this.
1185   assert(new_it == new_rules.registers_.end());
1186 
1187   return true;
1188 }
1189 
1190 // Remove all register rules and clear cfa_rule_.
Clear()1191 void CallFrameInfo::RuleMap::Clear() {
1192   delete cfa_rule_;
1193   cfa_rule_ = NULL;
1194   for (RuleByNumber::iterator it = registers_.begin();
1195        it != registers_.end(); it++)
1196     delete it->second;
1197   registers_.clear();
1198 }
1199 
1200 // The state of the call frame information interpreter as it processes
1201 // instructions from a CIE and FDE.
1202 class CallFrameInfo::State {
1203  public:
1204   // Create a call frame information interpreter state with the given
1205   // reporter, reader, handler, and initial call frame info address.
State(ByteReader * reader,Handler * handler,Reporter * reporter,uint64 address)1206   State(ByteReader *reader, Handler *handler, Reporter *reporter,
1207         uint64 address)
1208       : reader_(reader), handler_(handler), reporter_(reporter),
1209         address_(address), entry_(NULL), cursor_(NULL) { }
1210 
1211   // Interpret instructions from CIE, save the resulting rule set for
1212   // DW_CFA_restore instructions, and return true. On error, report
1213   // the problem to reporter_ and return false.
1214   bool InterpretCIE(const CIE &cie);
1215 
1216   // Interpret instructions from FDE, and return true. On error,
1217   // report the problem to reporter_ and return false.
1218   bool InterpretFDE(const FDE &fde);
1219 
1220  private:
1221   // The operands of a CFI instruction, for ParseOperands.
1222   struct Operands {
1223     unsigned register_number;  // A register number.
1224     uint64 offset;             // An offset or address.
1225     long signed_offset;        // A signed offset.
1226     string expression;         // A DWARF expression.
1227   };
1228 
1229   // Parse CFI instruction operands from STATE's instruction stream as
1230   // described by FORMAT. On success, populate OPERANDS with the
1231   // results, and return true. On failure, report the problem and
1232   // return false.
1233   //
1234   // Each character of FORMAT should be one of the following:
1235   //
1236   //   'r'  unsigned LEB128 register number (OPERANDS->register_number)
1237   //   'o'  unsigned LEB128 offset          (OPERANDS->offset)
1238   //   's'  signed LEB128 offset            (OPERANDS->signed_offset)
1239   //   'a'  machine-size address            (OPERANDS->offset)
1240   //        (If the CIE has a 'z' augmentation string, 'a' uses the
1241   //        encoding specified by the 'R' argument.)
1242   //   '1'  a one-byte offset               (OPERANDS->offset)
1243   //   '2'  a two-byte offset               (OPERANDS->offset)
1244   //   '4'  a four-byte offset              (OPERANDS->offset)
1245   //   '8'  an eight-byte offset            (OPERANDS->offset)
1246   //   'e'  a DW_FORM_block holding a       (OPERANDS->expression)
1247   //        DWARF expression
1248   bool ParseOperands(const char *format, Operands *operands);
1249 
1250   // Interpret one CFI instruction from STATE's instruction stream, update
1251   // STATE, report any rule changes to handler_, and return true. On
1252   // failure, report the problem and return false.
1253   bool DoInstruction();
1254 
1255   // The following Do* member functions are subroutines of DoInstruction,
1256   // factoring out the actual work of operations that have several
1257   // different encodings.
1258 
1259   // Set the CFA rule to be the value of BASE_REGISTER plus OFFSET, and
1260   // return true. On failure, report and return false. (Used for
1261   // DW_CFA_def_cfa and DW_CFA_def_cfa_sf.)
1262   bool DoDefCFA(unsigned base_register, long offset);
1263 
1264   // Change the offset of the CFA rule to OFFSET, and return true. On
1265   // failure, report and return false. (Subroutine for
1266   // DW_CFA_def_cfa_offset and DW_CFA_def_cfa_offset_sf.)
1267   bool DoDefCFAOffset(long offset);
1268 
1269   // Specify that REG can be recovered using RULE, and return true. On
1270   // failure, report and return false.
1271   bool DoRule(unsigned reg, Rule *rule);
1272 
1273   // Specify that REG can be found at OFFSET from the CFA, and return true.
1274   // On failure, report and return false. (Subroutine for DW_CFA_offset,
1275   // DW_CFA_offset_extended, and DW_CFA_offset_extended_sf.)
1276   bool DoOffset(unsigned reg, long offset);
1277 
1278   // Specify that the caller's value for REG is the CFA plus OFFSET,
1279   // and return true. On failure, report and return false. (Subroutine
1280   // for DW_CFA_val_offset and DW_CFA_val_offset_sf.)
1281   bool DoValOffset(unsigned reg, long offset);
1282 
1283   // Restore REG to the rule established in the CIE, and return true. On
1284   // failure, report and return false. (Subroutine for DW_CFA_restore and
1285   // DW_CFA_restore_extended.)
1286   bool DoRestore(unsigned reg);
1287 
1288   // Return the section offset of the instruction at cursor. For use
1289   // in error messages.
CursorOffset()1290   uint64 CursorOffset() { return entry_->offset + (cursor_ - entry_->start); }
1291 
1292   // Report that entry_ is incomplete, and return false. For brevity.
ReportIncomplete()1293   bool ReportIncomplete() {
1294     reporter_->Incomplete(entry_->offset, entry_->kind);
1295     return false;
1296   }
1297 
1298   // For reading multi-byte values with the appropriate endianness.
1299   ByteReader *reader_;
1300 
1301   // The handler to which we should report the data we find.
1302   Handler *handler_;
1303 
1304   // For reporting problems in the info we're parsing.
1305   Reporter *reporter_;
1306 
1307   // The code address to which the next instruction in the stream applies.
1308   uint64 address_;
1309 
1310   // The entry whose instructions we are currently processing. This is
1311   // first a CIE, and then an FDE.
1312   const Entry *entry_;
1313 
1314   // The next instruction to process.
1315   const char *cursor_;
1316 
1317   // The current set of rules.
1318   RuleMap rules_;
1319 
1320   // The set of rules established by the CIE, used by DW_CFA_restore
1321   // and DW_CFA_restore_extended. We set this after interpreting the
1322   // CIE's instructions.
1323   RuleMap cie_rules_;
1324 
1325   // A stack of saved states, for DW_CFA_remember_state and
1326   // DW_CFA_restore_state.
1327   std::stack<RuleMap> saved_rules_;
1328 };
1329 
InterpretCIE(const CIE & cie)1330 bool CallFrameInfo::State::InterpretCIE(const CIE &cie) {
1331   entry_ = &cie;
1332   cursor_ = entry_->instructions;
1333   while (cursor_ < entry_->end)
1334     if (!DoInstruction())
1335       return false;
1336   // Note the rules established by the CIE, for use by DW_CFA_restore
1337   // and DW_CFA_restore_extended.
1338   cie_rules_ = rules_;
1339   return true;
1340 }
1341 
InterpretFDE(const FDE & fde)1342 bool CallFrameInfo::State::InterpretFDE(const FDE &fde) {
1343   entry_ = &fde;
1344   cursor_ = entry_->instructions;
1345   while (cursor_ < entry_->end)
1346     if (!DoInstruction())
1347       return false;
1348   return true;
1349 }
1350 
ParseOperands(const char * format,Operands * operands)1351 bool CallFrameInfo::State::ParseOperands(const char *format,
1352                                          Operands *operands) {
1353   size_t len;
1354   const char *operand;
1355 
1356   for (operand = format; *operand; operand++) {
1357     size_t bytes_left = entry_->end - cursor_;
1358     switch (*operand) {
1359       case 'r':
1360         operands->register_number = reader_->ReadUnsignedLEB128(cursor_, &len);
1361         if (len > bytes_left) return ReportIncomplete();
1362         cursor_ += len;
1363         break;
1364 
1365       case 'o':
1366         operands->offset = reader_->ReadUnsignedLEB128(cursor_, &len);
1367         if (len > bytes_left) return ReportIncomplete();
1368         cursor_ += len;
1369         break;
1370 
1371       case 's':
1372         operands->signed_offset = reader_->ReadSignedLEB128(cursor_, &len);
1373         if (len > bytes_left) return ReportIncomplete();
1374         cursor_ += len;
1375         break;
1376 
1377       case 'a':
1378         operands->offset =
1379           reader_->ReadEncodedPointer(cursor_, entry_->cie->pointer_encoding,
1380                                       &len);
1381         if (len > bytes_left) return ReportIncomplete();
1382         cursor_ += len;
1383         break;
1384 
1385       case '1':
1386         if (1 > bytes_left) return ReportIncomplete();
1387         operands->offset = static_cast<unsigned char>(*cursor_++);
1388         break;
1389 
1390       case '2':
1391         if (2 > bytes_left) return ReportIncomplete();
1392         operands->offset = reader_->ReadTwoBytes(cursor_);
1393         cursor_ += 2;
1394         break;
1395 
1396       case '4':
1397         if (4 > bytes_left) return ReportIncomplete();
1398         operands->offset = reader_->ReadFourBytes(cursor_);
1399         cursor_ += 4;
1400         break;
1401 
1402       case '8':
1403         if (8 > bytes_left) return ReportIncomplete();
1404         operands->offset = reader_->ReadEightBytes(cursor_);
1405         cursor_ += 8;
1406         break;
1407 
1408       case 'e': {
1409         size_t expression_length = reader_->ReadUnsignedLEB128(cursor_, &len);
1410         if (len > bytes_left || expression_length > bytes_left - len)
1411           return ReportIncomplete();
1412         cursor_ += len;
1413         operands->expression = string(cursor_, expression_length);
1414         cursor_ += expression_length;
1415         break;
1416       }
1417 
1418       default:
1419           assert(0);
1420     }
1421   }
1422 
1423   return true;
1424 }
1425 
DoInstruction()1426 bool CallFrameInfo::State::DoInstruction() {
1427   CIE *cie = entry_->cie;
1428   Operands ops;
1429 
1430   // Our entry's kind should have been set by now.
1431   assert(entry_->kind != kUnknown);
1432 
1433   // We shouldn't have been invoked unless there were more
1434   // instructions to parse.
1435   assert(cursor_ < entry_->end);
1436 
1437   unsigned opcode = *cursor_++;
1438   if ((opcode & 0xc0) != 0) {
1439     switch (opcode & 0xc0) {
1440       // Advance the address.
1441       case DW_CFA_advance_loc: {
1442         size_t code_offset = opcode & 0x3f;
1443         address_ += code_offset * cie->code_alignment_factor;
1444         break;
1445       }
1446 
1447       // Find a register at an offset from the CFA.
1448       case DW_CFA_offset:
1449         if (!ParseOperands("o", &ops) ||
1450             !DoOffset(opcode & 0x3f, ops.offset * cie->data_alignment_factor))
1451           return false;
1452         break;
1453 
1454       // Restore the rule established for a register by the CIE.
1455       case DW_CFA_restore:
1456         if (!DoRestore(opcode & 0x3f)) return false;
1457         break;
1458 
1459       // The 'if' above should have excluded this possibility.
1460       default:
1461         assert(0);
1462     }
1463 
1464     // Return here, so the big switch below won't be indented.
1465     return true;
1466   }
1467 
1468   switch (opcode) {
1469     // Set the address.
1470     case DW_CFA_set_loc:
1471       if (!ParseOperands("a", &ops)) return false;
1472       address_ = ops.offset;
1473       break;
1474 
1475     // Advance the address.
1476     case DW_CFA_advance_loc1:
1477       if (!ParseOperands("1", &ops)) return false;
1478       address_ += ops.offset * cie->code_alignment_factor;
1479       break;
1480 
1481     // Advance the address.
1482     case DW_CFA_advance_loc2:
1483       if (!ParseOperands("2", &ops)) return false;
1484       address_ += ops.offset * cie->code_alignment_factor;
1485       break;
1486 
1487     // Advance the address.
1488     case DW_CFA_advance_loc4:
1489       if (!ParseOperands("4", &ops)) return false;
1490       address_ += ops.offset * cie->code_alignment_factor;
1491       break;
1492 
1493     // Advance the address.
1494     case DW_CFA_MIPS_advance_loc8:
1495       if (!ParseOperands("8", &ops)) return false;
1496       address_ += ops.offset * cie->code_alignment_factor;
1497       break;
1498 
1499     // Compute the CFA by adding an offset to a register.
1500     case DW_CFA_def_cfa:
1501       if (!ParseOperands("ro", &ops) ||
1502           !DoDefCFA(ops.register_number, ops.offset))
1503         return false;
1504       break;
1505 
1506     // Compute the CFA by adding an offset to a register.
1507     case DW_CFA_def_cfa_sf:
1508       if (!ParseOperands("rs", &ops) ||
1509           !DoDefCFA(ops.register_number,
1510                     ops.signed_offset * cie->data_alignment_factor))
1511         return false;
1512       break;
1513 
1514     // Change the base register used to compute the CFA.
1515     case DW_CFA_def_cfa_register: {
1516       if (!ParseOperands("r", &ops)) return false;
1517       Rule *cfa_rule = rules_.CFARule();
1518       if (!cfa_rule) {
1519         if (!DoDefCFA(ops.register_number, ops.offset)) {
1520           reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
1521           return false;
1522         }
1523       } else {
1524         cfa_rule->SetBaseRegister(ops.register_number);
1525         if (!cfa_rule->Handle(handler_, address_,
1526                               Handler::kCFARegister))
1527         return false;
1528       }
1529       break;
1530     }
1531 
1532     // Change the offset used to compute the CFA.
1533     case DW_CFA_def_cfa_offset:
1534       if (!ParseOperands("o", &ops) ||
1535           !DoDefCFAOffset(ops.offset))
1536         return false;
1537       break;
1538 
1539     // Change the offset used to compute the CFA.
1540     case DW_CFA_def_cfa_offset_sf:
1541       if (!ParseOperands("s", &ops) ||
1542           !DoDefCFAOffset(ops.signed_offset * cie->data_alignment_factor))
1543         return false;
1544       break;
1545 
1546     // Specify an expression whose value is the CFA.
1547     case DW_CFA_def_cfa_expression: {
1548       if (!ParseOperands("e", &ops))
1549         return false;
1550       Rule *rule = new ValExpressionRule(ops.expression);
1551       rules_.SetCFARule(rule);
1552       if (!rule->Handle(handler_, address_,
1553                         Handler::kCFARegister))
1554         return false;
1555       break;
1556     }
1557 
1558     // The register's value cannot be recovered.
1559     case DW_CFA_undefined: {
1560       if (!ParseOperands("r", &ops) ||
1561           !DoRule(ops.register_number, new UndefinedRule()))
1562         return false;
1563       break;
1564     }
1565 
1566     // The register's value is unchanged from its value in the caller.
1567     case DW_CFA_same_value: {
1568       if (!ParseOperands("r", &ops) ||
1569           !DoRule(ops.register_number, new SameValueRule()))
1570         return false;
1571       break;
1572     }
1573 
1574     // Find a register at an offset from the CFA.
1575     case DW_CFA_offset_extended:
1576       if (!ParseOperands("ro", &ops) ||
1577           !DoOffset(ops.register_number,
1578                     ops.offset * cie->data_alignment_factor))
1579         return false;
1580       break;
1581 
1582     // The register is saved at an offset from the CFA.
1583     case DW_CFA_offset_extended_sf:
1584       if (!ParseOperands("rs", &ops) ||
1585           !DoOffset(ops.register_number,
1586                     ops.signed_offset * cie->data_alignment_factor))
1587         return false;
1588       break;
1589 
1590     // The register is saved at an offset from the CFA.
1591     case DW_CFA_GNU_negative_offset_extended:
1592       if (!ParseOperands("ro", &ops) ||
1593           !DoOffset(ops.register_number,
1594                     -ops.offset * cie->data_alignment_factor))
1595         return false;
1596       break;
1597 
1598     // The register's value is the sum of the CFA plus an offset.
1599     case DW_CFA_val_offset:
1600       if (!ParseOperands("ro", &ops) ||
1601           !DoValOffset(ops.register_number,
1602                        ops.offset * cie->data_alignment_factor))
1603         return false;
1604       break;
1605 
1606     // The register's value is the sum of the CFA plus an offset.
1607     case DW_CFA_val_offset_sf:
1608       if (!ParseOperands("rs", &ops) ||
1609           !DoValOffset(ops.register_number,
1610                        ops.signed_offset * cie->data_alignment_factor))
1611         return false;
1612       break;
1613 
1614     // The register has been saved in another register.
1615     case DW_CFA_register: {
1616       if (!ParseOperands("ro", &ops) ||
1617           !DoRule(ops.register_number, new RegisterRule(ops.offset)))
1618         return false;
1619       break;
1620     }
1621 
1622     // An expression yields the address at which the register is saved.
1623     case DW_CFA_expression: {
1624       if (!ParseOperands("re", &ops) ||
1625           !DoRule(ops.register_number, new ExpressionRule(ops.expression)))
1626         return false;
1627       break;
1628     }
1629 
1630     // An expression yields the caller's value for the register.
1631     case DW_CFA_val_expression: {
1632       if (!ParseOperands("re", &ops) ||
1633           !DoRule(ops.register_number, new ValExpressionRule(ops.expression)))
1634         return false;
1635       break;
1636     }
1637 
1638     // Restore the rule established for a register by the CIE.
1639     case DW_CFA_restore_extended:
1640       if (!ParseOperands("r", &ops) ||
1641           !DoRestore( ops.register_number))
1642         return false;
1643       break;
1644 
1645     // Save the current set of rules on a stack.
1646     case DW_CFA_remember_state:
1647       saved_rules_.push(rules_);
1648       break;
1649 
1650     // Pop the current set of rules off the stack.
1651     case DW_CFA_restore_state: {
1652       if (saved_rules_.empty()) {
1653         reporter_->EmptyStateStack(entry_->offset, entry_->kind,
1654                                    CursorOffset());
1655         return false;
1656       }
1657       const RuleMap &new_rules = saved_rules_.top();
1658       if (rules_.CFARule() && !new_rules.CFARule()) {
1659         reporter_->ClearingCFARule(entry_->offset, entry_->kind,
1660                                    CursorOffset());
1661         return false;
1662       }
1663       rules_.HandleTransitionTo(handler_, address_, new_rules);
1664       rules_ = new_rules;
1665       saved_rules_.pop();
1666       break;
1667     }
1668 
1669     // No operation.  (Padding instruction.)
1670     case DW_CFA_nop:
1671       break;
1672 
1673     // A SPARC register window save: Registers 8 through 15 (%o0-%o7)
1674     // are saved in registers 24 through 31 (%i0-%i7), and registers
1675     // 16 through 31 (%l0-%l7 and %i0-%i7) are saved at CFA offsets
1676     // (0-15 * the register size). The register numbers must be
1677     // hard-coded. A GNU extension, and not a pretty one.
1678     case DW_CFA_GNU_window_save: {
1679       // Save %o0-%o7 in %i0-%i7.
1680       for (int i = 8; i < 16; i++)
1681         if (!DoRule(i, new RegisterRule(i + 16)))
1682           return false;
1683       // Save %l0-%l7 and %i0-%i7 at the CFA.
1684       for (int i = 16; i < 32; i++)
1685         // Assume that the byte reader's address size is the same as
1686         // the architecture's register size. !@#%*^ hilarious.
1687         if (!DoRule(i, new OffsetRule(Handler::kCFARegister,
1688                                       (i - 16) * reader_->AddressSize())))
1689           return false;
1690       break;
1691     }
1692 
1693     // I'm not sure what this is. GDB doesn't use it for unwinding.
1694     case DW_CFA_GNU_args_size:
1695       if (!ParseOperands("o", &ops)) return false;
1696       break;
1697 
1698     // An opcode we don't recognize.
1699     default: {
1700       reporter_->BadInstruction(entry_->offset, entry_->kind, CursorOffset());
1701       return false;
1702     }
1703   }
1704 
1705   return true;
1706 }
1707 
DoDefCFA(unsigned base_register,long offset)1708 bool CallFrameInfo::State::DoDefCFA(unsigned base_register, long offset) {
1709   Rule *rule = new ValOffsetRule(base_register, offset);
1710   rules_.SetCFARule(rule);
1711   return rule->Handle(handler_, address_,
1712                       Handler::kCFARegister);
1713 }
1714 
DoDefCFAOffset(long offset)1715 bool CallFrameInfo::State::DoDefCFAOffset(long offset) {
1716   Rule *cfa_rule = rules_.CFARule();
1717   if (!cfa_rule) {
1718     reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
1719     return false;
1720   }
1721   cfa_rule->SetOffset(offset);
1722   return cfa_rule->Handle(handler_, address_,
1723                           Handler::kCFARegister);
1724 }
1725 
DoRule(unsigned reg,Rule * rule)1726 bool CallFrameInfo::State::DoRule(unsigned reg, Rule *rule) {
1727   rules_.SetRegisterRule(reg, rule);
1728   return rule->Handle(handler_, address_, reg);
1729 }
1730 
DoOffset(unsigned reg,long offset)1731 bool CallFrameInfo::State::DoOffset(unsigned reg, long offset) {
1732   if (!rules_.CFARule()) {
1733     reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
1734     return false;
1735   }
1736   return DoRule(reg,
1737                 new OffsetRule(Handler::kCFARegister, offset));
1738 }
1739 
DoValOffset(unsigned reg,long offset)1740 bool CallFrameInfo::State::DoValOffset(unsigned reg, long offset) {
1741   if (!rules_.CFARule()) {
1742     reporter_->NoCFARule(entry_->offset, entry_->kind, CursorOffset());
1743     return false;
1744   }
1745   return DoRule(reg,
1746                 new ValOffsetRule(Handler::kCFARegister, offset));
1747 }
1748 
DoRestore(unsigned reg)1749 bool CallFrameInfo::State::DoRestore(unsigned reg) {
1750   // DW_CFA_restore and DW_CFA_restore_extended don't make sense in a CIE.
1751   if (entry_->kind == kCIE) {
1752     reporter_->RestoreInCIE(entry_->offset, CursorOffset());
1753     return false;
1754   }
1755   Rule *rule = cie_rules_.RegisterRule(reg);
1756   if (!rule) {
1757     // This isn't really the right thing to do, but since CFI generally
1758     // only mentions callee-saves registers, and GCC's convention for
1759     // callee-saves registers is that they are unchanged, it's a good
1760     // approximation.
1761     rule = new SameValueRule();
1762   }
1763   return DoRule(reg, rule);
1764 }
1765 
ReadEntryPrologue(const char * cursor,Entry * entry)1766 bool CallFrameInfo::ReadEntryPrologue(const char *cursor, Entry *entry) {
1767   const char *buffer_end = buffer_ + buffer_length_;
1768 
1769   // Initialize enough of ENTRY for use in error reporting.
1770   entry->offset = cursor - buffer_;
1771   entry->start = cursor;
1772   entry->kind = kUnknown;
1773   entry->end = NULL;
1774 
1775   // Read the initial length. This sets reader_'s offset size.
1776   size_t length_size;
1777   uint64 length = reader_->ReadInitialLength(cursor, &length_size);
1778   if (length_size > size_t(buffer_end - cursor))
1779     return ReportIncomplete(entry);
1780   cursor += length_size;
1781 
1782   // In a .eh_frame section, a length of zero marks the end of the series
1783   // of entries.
1784   if (length == 0 && eh_frame_) {
1785     entry->kind = kTerminator;
1786     entry->end = cursor;
1787     return true;
1788   }
1789 
1790   // Validate the length.
1791   if (length > size_t(buffer_end - cursor))
1792     return ReportIncomplete(entry);
1793 
1794   // The length is the number of bytes after the initial length field;
1795   // we have that position handy at this point, so compute the end
1796   // now. (If we're parsing 64-bit-offset DWARF on a 32-bit machine,
1797   // and the length didn't fit in a size_t, we would have rejected it
1798   // above.)
1799   entry->end = cursor + length;
1800 
1801   // Parse the next field: either the offset of a CIE or a CIE id.
1802   size_t offset_size = reader_->OffsetSize();
1803   if (offset_size > size_t(entry->end - cursor)) return ReportIncomplete(entry);
1804   entry->id = reader_->ReadOffset(cursor);
1805 
1806   // Don't advance cursor past id field yet; in .eh_frame data we need
1807   // the id's position to compute the section offset of an FDE's CIE.
1808 
1809   // Now we can decide what kind of entry this is.
1810   if (eh_frame_) {
1811     // In .eh_frame data, an ID of zero marks the entry as a CIE, and
1812     // anything else is an offset from the id field of the FDE to the start
1813     // of the CIE.
1814     if (entry->id == 0) {
1815       entry->kind = kCIE;
1816     } else {
1817       entry->kind = kFDE;
1818       // Turn the offset from the id into an offset from the buffer's start.
1819       entry->id = (cursor - buffer_) - entry->id;
1820     }
1821   } else {
1822     // In DWARF CFI data, an ID of ~0 (of the appropriate width, given the
1823     // offset size for the entry) marks the entry as a CIE, and anything
1824     // else is the offset of the CIE from the beginning of the section.
1825     if (offset_size == 4)
1826       entry->kind = (entry->id == 0xffffffff) ? kCIE : kFDE;
1827     else {
1828       assert(offset_size == 8);
1829       entry->kind = (entry->id == 0xffffffffffffffffULL) ? kCIE : kFDE;
1830     }
1831   }
1832 
1833   // Now advance cursor past the id.
1834    cursor += offset_size;
1835 
1836   // The fields specific to this kind of entry start here.
1837   entry->fields = cursor;
1838 
1839   entry->cie = NULL;
1840 
1841   return true;
1842 }
1843 
ReadCIEFields(CIE * cie)1844 bool CallFrameInfo::ReadCIEFields(CIE *cie) {
1845   const char *cursor = cie->fields;
1846   size_t len;
1847 
1848   assert(cie->kind == kCIE);
1849 
1850   // Prepare for early exit.
1851   cie->version = 0;
1852   cie->augmentation.clear();
1853   cie->code_alignment_factor = 0;
1854   cie->data_alignment_factor = 0;
1855   cie->return_address_register = 0;
1856   cie->has_z_augmentation = false;
1857   cie->pointer_encoding = DW_EH_PE_absptr;
1858   cie->instructions = 0;
1859 
1860   // Parse the version number.
1861   if (cie->end - cursor < 1)
1862     return ReportIncomplete(cie);
1863   cie->version = reader_->ReadOneByte(cursor);
1864   cursor++;
1865 
1866   // If we don't recognize the version, we can't parse any more fields of the
1867   // CIE. For DWARF CFI, we handle versions 1 through 3 (there was never a
1868   // version 2 of CFI data). For .eh_frame, we handle versions 1 and 3 as well;
1869   // the difference between those versions seems to be the same as for
1870   // .debug_frame.
1871   if (cie->version < 1 || cie->version > 3) {
1872     reporter_->UnrecognizedVersion(cie->offset, cie->version);
1873     return false;
1874   }
1875 
1876   const char *augmentation_start = cursor;
1877   const void *augmentation_end =
1878       memchr(augmentation_start, '\0', cie->end - augmentation_start);
1879   if (! augmentation_end) return ReportIncomplete(cie);
1880   cursor = static_cast<const char *>(augmentation_end);
1881   cie->augmentation = string(augmentation_start,
1882                                   cursor - augmentation_start);
1883   // Skip the terminating '\0'.
1884   cursor++;
1885 
1886   // Is this CFI augmented?
1887   if (!cie->augmentation.empty()) {
1888     // Is it an augmentation we recognize?
1889     if (cie->augmentation[0] == DW_Z_augmentation_start) {
1890       // Linux C++ ABI 'z' augmentation, used for exception handling data.
1891       cie->has_z_augmentation = true;
1892     } else {
1893       // Not an augmentation we recognize. Augmentations can have arbitrary
1894       // effects on the form of rest of the content, so we have to give up.
1895       reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation);
1896       return false;
1897     }
1898   }
1899 
1900   // Parse the code alignment factor.
1901   cie->code_alignment_factor = reader_->ReadUnsignedLEB128(cursor, &len);
1902   if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
1903   cursor += len;
1904 
1905   // Parse the data alignment factor.
1906   cie->data_alignment_factor = reader_->ReadSignedLEB128(cursor, &len);
1907   if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
1908   cursor += len;
1909 
1910   // Parse the return address register. This is a ubyte in version 1, and
1911   // a ULEB128 in version 3.
1912   if (cie->version == 1) {
1913     if (cursor >= cie->end) return ReportIncomplete(cie);
1914     cie->return_address_register = uint8(*cursor++);
1915   } else {
1916     cie->return_address_register = reader_->ReadUnsignedLEB128(cursor, &len);
1917     if (size_t(cie->end - cursor) < len) return ReportIncomplete(cie);
1918     cursor += len;
1919   }
1920 
1921   // If we have a 'z' augmentation string, find the augmentation data and
1922   // use the augmentation string to parse it.
1923   if (cie->has_z_augmentation) {
1924     uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &len);
1925     if (size_t(cie->end - cursor) < len + data_size)
1926       return ReportIncomplete(cie);
1927     cursor += len;
1928     const char *data = cursor;
1929     cursor += data_size;
1930     const char *data_end = cursor;
1931 
1932     cie->has_z_lsda = false;
1933     cie->has_z_personality = false;
1934     cie->has_z_signal_frame = false;
1935 
1936     // Walk the augmentation string, and extract values from the
1937     // augmentation data as the string directs.
1938     for (size_t i = 1; i < cie->augmentation.size(); i++) {
1939       switch (cie->augmentation[i]) {
1940         case DW_Z_has_LSDA:
1941           // The CIE's augmentation data holds the language-specific data
1942           // area pointer's encoding, and the FDE's augmentation data holds
1943           // the pointer itself.
1944           cie->has_z_lsda = true;
1945           // Fetch the LSDA encoding from the augmentation data.
1946           if (data >= data_end) return ReportIncomplete(cie);
1947           cie->lsda_encoding = DwarfPointerEncoding(*data++);
1948           if (!reader_->ValidEncoding(cie->lsda_encoding)) {
1949             reporter_->InvalidPointerEncoding(cie->offset, cie->lsda_encoding);
1950             return false;
1951           }
1952           // Don't check if the encoding is usable here --- we haven't
1953           // read the FDE's fields yet, so we're not prepared for
1954           // DW_EH_PE_funcrel, although that's a fine encoding for the
1955           // LSDA to use, since it appears in the FDE.
1956           break;
1957 
1958         case DW_Z_has_personality_routine:
1959           // The CIE's augmentation data holds the personality routine
1960           // pointer's encoding, followed by the pointer itself.
1961           cie->has_z_personality = true;
1962           // Fetch the personality routine pointer's encoding from the
1963           // augmentation data.
1964           if (data >= data_end) return ReportIncomplete(cie);
1965           cie->personality_encoding = DwarfPointerEncoding(*data++);
1966           if (!reader_->ValidEncoding(cie->personality_encoding)) {
1967             reporter_->InvalidPointerEncoding(cie->offset,
1968                                               cie->personality_encoding);
1969             return false;
1970           }
1971           if (!reader_->UsableEncoding(cie->personality_encoding)) {
1972             reporter_->UnusablePointerEncoding(cie->offset,
1973                                                cie->personality_encoding);
1974             return false;
1975           }
1976           // Fetch the personality routine's pointer itself from the data.
1977           cie->personality_address =
1978             reader_->ReadEncodedPointer(data, cie->personality_encoding,
1979                                         &len);
1980           if (len > size_t(data_end - data))
1981             return ReportIncomplete(cie);
1982           data += len;
1983           break;
1984 
1985         case DW_Z_has_FDE_address_encoding:
1986           // The CIE's augmentation data holds the pointer encoding to use
1987           // for addresses in the FDE.
1988           if (data >= data_end) return ReportIncomplete(cie);
1989           cie->pointer_encoding = DwarfPointerEncoding(*data++);
1990           if (!reader_->ValidEncoding(cie->pointer_encoding)) {
1991             reporter_->InvalidPointerEncoding(cie->offset,
1992                                               cie->pointer_encoding);
1993             return false;
1994           }
1995           if (!reader_->UsableEncoding(cie->pointer_encoding)) {
1996             reporter_->UnusablePointerEncoding(cie->offset,
1997                                                cie->pointer_encoding);
1998             return false;
1999           }
2000           break;
2001 
2002         case DW_Z_is_signal_trampoline:
2003           // Frames using this CIE are signal delivery frames.
2004           cie->has_z_signal_frame = true;
2005           break;
2006 
2007         default:
2008           // An augmentation we don't recognize.
2009           reporter_->UnrecognizedAugmentation(cie->offset, cie->augmentation);
2010           return false;
2011       }
2012     }
2013   }
2014 
2015   // The CIE's instructions start here.
2016   cie->instructions = cursor;
2017 
2018   return true;
2019 }
2020 
ReadFDEFields(FDE * fde)2021 bool CallFrameInfo::ReadFDEFields(FDE *fde) {
2022   const char *cursor = fde->fields;
2023   size_t size;
2024 
2025   fde->address = reader_->ReadEncodedPointer(cursor, fde->cie->pointer_encoding,
2026                                              &size);
2027   if (size > size_t(fde->end - cursor))
2028     return ReportIncomplete(fde);
2029   cursor += size;
2030   reader_->SetFunctionBase(fde->address);
2031 
2032   // For the length, we strip off the upper nybble of the encoding used for
2033   // the starting address.
2034   DwarfPointerEncoding length_encoding =
2035     DwarfPointerEncoding(fde->cie->pointer_encoding & 0x0f);
2036   fde->size = reader_->ReadEncodedPointer(cursor, length_encoding, &size);
2037   if (size > size_t(fde->end - cursor))
2038     return ReportIncomplete(fde);
2039   cursor += size;
2040 
2041   // If the CIE has a 'z' augmentation string, then augmentation data
2042   // appears here.
2043   if (fde->cie->has_z_augmentation) {
2044     uint64_t data_size = reader_->ReadUnsignedLEB128(cursor, &size);
2045     if (size_t(fde->end - cursor) < size + data_size)
2046       return ReportIncomplete(fde);
2047     cursor += size;
2048 
2049     // In the abstract, we should walk the augmentation string, and extract
2050     // items from the FDE's augmentation data as we encounter augmentation
2051     // string characters that specify their presence: the ordering of items
2052     // in the augmentation string determines the arrangement of values in
2053     // the augmentation data.
2054     //
2055     // In practice, there's only ever one value in FDE augmentation data
2056     // that we support --- the LSDA pointer --- and we have to bail if we
2057     // see any unrecognized augmentation string characters. So if there is
2058     // anything here at all, we know what it is, and where it starts.
2059     if (fde->cie->has_z_lsda) {
2060       // Check whether the LSDA's pointer encoding is usable now: only once
2061       // we've parsed the FDE's starting address do we call reader_->
2062       // SetFunctionBase, so that the DW_EH_PE_funcrel encoding becomes
2063       // usable.
2064       if (!reader_->UsableEncoding(fde->cie->lsda_encoding)) {
2065         reporter_->UnusablePointerEncoding(fde->cie->offset,
2066                                            fde->cie->lsda_encoding);
2067         return false;
2068       }
2069 
2070       fde->lsda_address =
2071         reader_->ReadEncodedPointer(cursor, fde->cie->lsda_encoding, &size);
2072       if (size > data_size)
2073         return ReportIncomplete(fde);
2074       // Ideally, we would also complain here if there were unconsumed
2075       // augmentation data.
2076     }
2077 
2078     cursor += data_size;
2079   }
2080 
2081   // The FDE's instructions start after those.
2082   fde->instructions = cursor;
2083 
2084   return true;
2085 }
2086 
Start()2087 bool CallFrameInfo::Start() {
2088   const char *buffer_end = buffer_ + buffer_length_;
2089   const char *cursor;
2090   bool all_ok = true;
2091   const char *entry_end;
2092   bool ok;
2093 
2094   // Traverse all the entries in buffer_, skipping CIEs and offering
2095   // FDEs to the handler.
2096   for (cursor = buffer_; cursor < buffer_end;
2097        cursor = entry_end, all_ok = all_ok && ok) {
2098     FDE fde;
2099 
2100     // Make it easy to skip this entry with 'continue': assume that
2101     // things are not okay until we've checked all the data, and
2102     // prepare the address of the next entry.
2103     ok = false;
2104 
2105     // Read the entry's prologue.
2106     if (!ReadEntryPrologue(cursor, &fde)) {
2107       if (!fde.end) {
2108         // If we couldn't even figure out this entry's extent, then we
2109         // must stop processing entries altogether.
2110         all_ok = false;
2111         break;
2112       }
2113       entry_end = fde.end;
2114       continue;
2115     }
2116 
2117     // The next iteration picks up after this entry.
2118     entry_end = fde.end;
2119 
2120     // Did we see an .eh_frame terminating mark?
2121     if (fde.kind == kTerminator) {
2122       // If there appears to be more data left in the section after the
2123       // terminating mark, warn the user. But this is just a warning;
2124       // we leave all_ok true.
2125       if (fde.end < buffer_end) reporter_->EarlyEHTerminator(fde.offset);
2126       break;
2127     }
2128 
2129     // In this loop, we skip CIEs. We only parse them fully when we
2130     // parse an FDE that refers to them. This limits our memory
2131     // consumption (beyond the buffer itself) to that needed to
2132     // process the largest single entry.
2133     if (fde.kind != kFDE) {
2134       ok = true;
2135       continue;
2136     }
2137 
2138     // Validate the CIE pointer.
2139     if (fde.id > buffer_length_) {
2140       reporter_->CIEPointerOutOfRange(fde.offset, fde.id);
2141       continue;
2142     }
2143 
2144     CIE cie;
2145 
2146     // Parse this FDE's CIE header.
2147     if (!ReadEntryPrologue(buffer_ + fde.id, &cie))
2148       continue;
2149     // This had better be an actual CIE.
2150     if (cie.kind != kCIE) {
2151       reporter_->BadCIEId(fde.offset, fde.id);
2152       continue;
2153     }
2154     if (!ReadCIEFields(&cie))
2155       continue;
2156 
2157     // We now have the values that govern both the CIE and the FDE.
2158     cie.cie = &cie;
2159     fde.cie = &cie;
2160 
2161     // Parse the FDE's header.
2162     if (!ReadFDEFields(&fde))
2163       continue;
2164 
2165     // Call Entry to ask the consumer if they're interested.
2166     if (!handler_->Entry(fde.offset, fde.address, fde.size,
2167                          cie.version, cie.augmentation,
2168                          cie.return_address_register)) {
2169       // The handler isn't interested in this entry. That's not an error.
2170       ok = true;
2171       continue;
2172     }
2173 
2174     if (cie.has_z_augmentation) {
2175       // Report the personality routine address, if we have one.
2176       if (cie.has_z_personality) {
2177         if (!handler_
2178             ->PersonalityRoutine(cie.personality_address,
2179                                  IsIndirectEncoding(cie.personality_encoding)))
2180           continue;
2181       }
2182 
2183       // Report the language-specific data area address, if we have one.
2184       if (cie.has_z_lsda) {
2185         if (!handler_
2186             ->LanguageSpecificDataArea(fde.lsda_address,
2187                                        IsIndirectEncoding(cie.lsda_encoding)))
2188           continue;
2189       }
2190 
2191       // If this is a signal-handling frame, report that.
2192       if (cie.has_z_signal_frame) {
2193         if (!handler_->SignalHandler())
2194           continue;
2195       }
2196     }
2197 
2198     // Interpret the CIE's instructions, and then the FDE's instructions.
2199     State state(reader_, handler_, reporter_, fde.address);
2200     ok = state.InterpretCIE(cie) && state.InterpretFDE(fde);
2201 
2202     // Tell the ByteReader that the function start address from the
2203     // FDE header is no longer valid.
2204     reader_->ClearFunctionBase();
2205 
2206     // Report the end of the entry.
2207     handler_->End();
2208   }
2209 
2210   return all_ok;
2211 }
2212 
KindName(EntryKind kind)2213 const char *CallFrameInfo::KindName(EntryKind kind) {
2214   if (kind == CallFrameInfo::kUnknown)
2215     return "entry";
2216   else if (kind == CallFrameInfo::kCIE)
2217     return "common information entry";
2218   else if (kind == CallFrameInfo::kFDE)
2219     return "frame description entry";
2220   else {
2221     assert (kind == CallFrameInfo::kTerminator);
2222     return ".eh_frame sequence terminator";
2223   }
2224 }
2225 
ReportIncomplete(Entry * entry)2226 bool CallFrameInfo::ReportIncomplete(Entry *entry) {
2227   reporter_->Incomplete(entry->offset, entry->kind);
2228   return false;
2229 }
2230 
Incomplete(uint64 offset,CallFrameInfo::EntryKind kind)2231 void CallFrameInfo::Reporter::Incomplete(uint64 offset,
2232                                          CallFrameInfo::EntryKind kind) {
2233   fprintf(stderr,
2234           "%s: CFI %s at offset 0x%llx in '%s': entry ends early\n",
2235           filename_.c_str(), CallFrameInfo::KindName(kind), offset,
2236           section_.c_str());
2237 }
2238 
EarlyEHTerminator(uint64 offset)2239 void CallFrameInfo::Reporter::EarlyEHTerminator(uint64 offset) {
2240   fprintf(stderr,
2241           "%s: CFI at offset 0x%llx in '%s': saw end-of-data marker"
2242           " before end of section contents\n",
2243           filename_.c_str(), offset, section_.c_str());
2244 }
2245 
CIEPointerOutOfRange(uint64 offset,uint64 cie_offset)2246 void CallFrameInfo::Reporter::CIEPointerOutOfRange(uint64 offset,
2247                                                    uint64 cie_offset) {
2248   fprintf(stderr,
2249           "%s: CFI frame description entry at offset 0x%llx in '%s':"
2250           " CIE pointer is out of range: 0x%llx\n",
2251           filename_.c_str(), offset, section_.c_str(), cie_offset);
2252 }
2253 
BadCIEId(uint64 offset,uint64 cie_offset)2254 void CallFrameInfo::Reporter::BadCIEId(uint64 offset, uint64 cie_offset) {
2255   fprintf(stderr,
2256           "%s: CFI frame description entry at offset 0x%llx in '%s':"
2257           " CIE pointer does not point to a CIE: 0x%llx\n",
2258           filename_.c_str(), offset, section_.c_str(), cie_offset);
2259 }
2260 
UnrecognizedVersion(uint64 offset,int version)2261 void CallFrameInfo::Reporter::UnrecognizedVersion(uint64 offset, int version) {
2262   fprintf(stderr,
2263           "%s: CFI frame description entry at offset 0x%llx in '%s':"
2264           " CIE specifies unrecognized version: %d\n",
2265           filename_.c_str(), offset, section_.c_str(), version);
2266 }
2267 
UnrecognizedAugmentation(uint64 offset,const string & aug)2268 void CallFrameInfo::Reporter::UnrecognizedAugmentation(uint64 offset,
2269                                                        const string &aug) {
2270   fprintf(stderr,
2271           "%s: CFI frame description entry at offset 0x%llx in '%s':"
2272           " CIE specifies unrecognized augmentation: '%s'\n",
2273           filename_.c_str(), offset, section_.c_str(), aug.c_str());
2274 }
2275 
InvalidPointerEncoding(uint64 offset,uint8 encoding)2276 void CallFrameInfo::Reporter::InvalidPointerEncoding(uint64 offset,
2277                                                      uint8 encoding) {
2278   fprintf(stderr,
2279           "%s: CFI common information entry at offset 0x%llx in '%s':"
2280           " 'z' augmentation specifies invalid pointer encoding: 0x%02x\n",
2281           filename_.c_str(), offset, section_.c_str(), encoding);
2282 }
2283 
UnusablePointerEncoding(uint64 offset,uint8 encoding)2284 void CallFrameInfo::Reporter::UnusablePointerEncoding(uint64 offset,
2285                                                       uint8 encoding) {
2286   fprintf(stderr,
2287           "%s: CFI common information entry at offset 0x%llx in '%s':"
2288           " 'z' augmentation specifies a pointer encoding for which"
2289           " we have no base address: 0x%02x\n",
2290           filename_.c_str(), offset, section_.c_str(), encoding);
2291 }
2292 
RestoreInCIE(uint64 offset,uint64 insn_offset)2293 void CallFrameInfo::Reporter::RestoreInCIE(uint64 offset, uint64 insn_offset) {
2294   fprintf(stderr,
2295           "%s: CFI common information entry at offset 0x%llx in '%s':"
2296           " the DW_CFA_restore instruction at offset 0x%llx"
2297           " cannot be used in a common information entry\n",
2298           filename_.c_str(), offset, section_.c_str(), insn_offset);
2299 }
2300 
BadInstruction(uint64 offset,CallFrameInfo::EntryKind kind,uint64 insn_offset)2301 void CallFrameInfo::Reporter::BadInstruction(uint64 offset,
2302                                              CallFrameInfo::EntryKind kind,
2303                                              uint64 insn_offset) {
2304   fprintf(stderr,
2305           "%s: CFI %s at offset 0x%llx in section '%s':"
2306           " the instruction at offset 0x%llx is unrecognized\n",
2307           filename_.c_str(), CallFrameInfo::KindName(kind),
2308           offset, section_.c_str(), insn_offset);
2309 }
2310 
NoCFARule(uint64 offset,CallFrameInfo::EntryKind kind,uint64 insn_offset)2311 void CallFrameInfo::Reporter::NoCFARule(uint64 offset,
2312                                         CallFrameInfo::EntryKind kind,
2313                                         uint64 insn_offset) {
2314   fprintf(stderr,
2315           "%s: CFI %s at offset 0x%llx in section '%s':"
2316           " the instruction at offset 0x%llx assumes that a CFA rule has"
2317           " been set, but none has been set\n",
2318           filename_.c_str(), CallFrameInfo::KindName(kind), offset,
2319           section_.c_str(), insn_offset);
2320 }
2321 
EmptyStateStack(uint64 offset,CallFrameInfo::EntryKind kind,uint64 insn_offset)2322 void CallFrameInfo::Reporter::EmptyStateStack(uint64 offset,
2323                                               CallFrameInfo::EntryKind kind,
2324                                               uint64 insn_offset) {
2325   fprintf(stderr,
2326           "%s: CFI %s at offset 0x%llx in section '%s':"
2327           " the DW_CFA_restore_state instruction at offset 0x%llx"
2328           " should pop a saved state from the stack, but the stack is empty\n",
2329           filename_.c_str(), CallFrameInfo::KindName(kind), offset,
2330           section_.c_str(), insn_offset);
2331 }
2332 
ClearingCFARule(uint64 offset,CallFrameInfo::EntryKind kind,uint64 insn_offset)2333 void CallFrameInfo::Reporter::ClearingCFARule(uint64 offset,
2334                                               CallFrameInfo::EntryKind kind,
2335                                               uint64 insn_offset) {
2336   fprintf(stderr,
2337           "%s: CFI %s at offset 0x%llx in section '%s':"
2338           " the DW_CFA_restore_state instruction at offset 0x%llx"
2339           " would clear the CFA rule in effect\n",
2340           filename_.c_str(), CallFrameInfo::KindName(kind), offset,
2341           section_.c_str(), insn_offset);
2342 }
2343 
2344 }  // namespace dwarf2reader
2345