dyld3/shared-cache/OptimizerLinkedit.cpp

   1 /* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*-
   2  *
   3  * Copyright (c) 2014 Apple Inc. All rights reserved.
   4  *
   5  * @APPLE_LICENSE_HEADER_START@
   6  *
   7  * This file contains Original Code and/or Modifications of Original Code
   8  * as defined in and that are subject to the Apple Public Source License
   9  * Version 2.0 (the 'License'). You may not use this file except in
  10  * compliance with the License. Please obtain a copy of the License at
  11  * http://www.opensource.apple.com/apsl/ and read it before using this
  12  * file.
  13  *
  14  * The Original Code and all software distributed under the License are
  15  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  16  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  17  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  19  * Please see the License for the specific language governing rights and
  20  * limitations under the License.
  21  *
  22  * @APPLE_LICENSE_HEADER_END@
  23  */
  24
  25
  26 #include <dirent.h>
  27 #include <sys/errno.h>
  28 #include <sys/fcntl.h>
  29 #include <mach-o/loader.h>
  30 #include <mach-o/fat.h>
  31 #include <assert.h>
  32
  33 #include <fstream>
  34 #include <string>
  35 #include <algorithm>
  36 #include <unordered_map>
  37 #include <unordered_set>
  38
  39 #include "MachOFileAbstraction.hpp"
  40 #include "Trie.hpp"
  41 #include "DyldSharedCache.h"
  42 #include "CacheBuilder.h"
  43 #include "MachOLoaded.h"
  44
  45 #define ALIGN_AS_TYPE(value, type) \
  46         ((value + alignof(type) - 1) & (-alignof(type)))
  47
  48 namespace {
  49
  50 template <typename P>
  51 class SortedStringPool
  52 {
  53 public:
  54     // add a string and symbol table entry index to be updated later
  55     void add(uint32_t symbolIndex, const char* symbolName) {
  56         _map[symbolName].push_back(symbolIndex);
  57     }
  58
  59     // copy sorted strings to buffer and update all symbol's string offsets
  60     uint32_t copyPoolAndUpdateOffsets(char* dstStringPool, macho_nlist<P>* symbolTable) {
  61         // walk sorted list of strings
  62         dstStringPool[0] = '\0'; // tradition for start of pool to be empty string
  63         uint32_t poolOffset = 1;
  64         for (auto& entry : _map) {
  65             const std::string& symName = entry.first;
  66             // append string to pool
  67             strcpy(&dstStringPool[poolOffset], symName.c_str());
  68             //  set each string offset of each symbol using it
  69             for (uint32_t symbolIndex : entry.second) {
  70                 symbolTable[symbolIndex].set_n_strx(poolOffset);
  71             }
  72             poolOffset += symName.size() + 1;
  73         }
  74         // return size of pool
  75         return poolOffset;
  76     }
  77
  78     size_t size() {
  79         size_t size = 1;
  80         for (auto& entry : _map) {
  81             size += (entry.first.size() + 1);
  82         }
  83         return size;
  84     }
  85
  86
  87 private:
  88     std::map<std::string, std::vector<uint32_t>> _map;
  89 };
  90
  91
  92 } // anonymous namespace
  93
  94
  95 struct LocalSymbolInfo
  96 {
  97     uint32_t    dylibOffset;
  98     uint32_t    nlistStartIndex;
  99     uint32_t    nlistCount;
 100 };
 101
 102
 103 template <typename P>
 104 class LinkeditOptimizer {
 105 public:
 106                     LinkeditOptimizer(void* cacheBuffer, macho_header<P>* mh, Diagnostics& diag);
 107
 108     uint32_t        linkeditSize() { return _linkeditSize; }
 109     uint64_t        linkeditAddr() { return _linkeditAddr; }
 110     const char*     installName() { return _installName; }
 111     void            copyWeakBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset);
 112     void            copyLazyBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset);
 113     void            copyBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset);
 114     void            copyExportInfo(uint8_t* newLinkEditContent, uint32_t& offset);
 115     void            copyExportedSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex);
 116     void            copyImportedSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex);
 117     void            copyLocalSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex,
 118                                      bool redact, std::vector<LocalSymbolInfo>& localSymbolInfos,
 119                                      std::vector<macho_nlist<P>>& unmappedLocalSymbols, SortedStringPool<P>& localSymbolsStringPool);
 120     void            copyFunctionStarts(uint8_t* newLinkEditContent, uint32_t& offset);
 121     void            copyDataInCode(uint8_t* newLinkEditContent, uint32_t& offset);
 122     void            copyIndirectSymbolTable(uint8_t* newLinkEditContent, uint32_t& offset);
 123     void            updateLoadCommands(uint32_t linkeditStartOffset, uint64_t mergedLinkeditAddr, uint64_t newLinkeditSize,
 124                                        uint32_t sharedSymbolTableStartOffset, uint32_t sharedSymbolTableCount,
 125                                        uint32_t sharedSymbolStringsOffset, uint32_t sharedSymbolStringsSize);
 126
 127     macho_header<P>*                        machHeader() { return _mh; }
 128     const std::vector<const char*>          getDownwardDependents() { return _downDependentPaths; }
 129     const std::vector<const char*>          getAllDependents() { return _allDependentPaths; }
 130     const std::vector<const char*>          getReExportPaths() { return _reExportPaths; }
 131     const std::vector<uint64_t>             initializerAddresses() { return _initializerAddresses; }
 132     const std::vector<macho_section<P>*>    dofSections() { return _dofSections; }
 133     uint32_t                                exportsTrieLinkEditOffset() { return _newExportInfoOffset; }
 134     uint32_t                                exportsTrieLinkEditSize() { return _exportInfoSize; }
 135     uint32_t                                weakBindingLinkEditOffset() { return _newWeakBindingInfoOffset; }
 136     uint32_t                                weakBindingLinkEditSize() { return _newWeakBindingSize; }
 137     uint64_t                                dyldSectionAddress() { return _dyldSectionAddr; }
 138     const std::vector<macho_segment_command<P>*>&  segCmds() { return _segCmds; }
 139
 140
 141     static void optimizeLinkedit(CacheBuilder& builder);
 142     static void mergeLinkedits(CacheBuilder& builder, std::vector<LinkeditOptimizer<P>*>& optimizers);
 143
 144 private:
 145
 146     typedef typename P::uint_t pint_t;
 147     typedef typename P::E E;
 148
 149     macho_header<P>*                        _mh;
 150     void*                                   _cacheBuffer;
 151     Diagnostics&                            _diagnostics;
 152     uint32_t                                _linkeditSize        = 0;
 153     uint64_t                                _linkeditAddr        = 0;
 154     const uint8_t*                          _linkeditBias       = nullptr;
 155     const char*                             _installName        = nullptr;
 156     macho_symtab_command<P>*                _symTabCmd          = nullptr;
 157     macho_dysymtab_command<P>*              _dynSymTabCmd       = nullptr;
 158     macho_dyld_info_command<P>*             _dyldInfo           = nullptr;
 159     macho_linkedit_data_command<P>*         _functionStartsCmd  = nullptr;
 160     macho_linkedit_data_command<P>*         _dataInCodeCmd      = nullptr;
 161     std::vector<macho_segment_command<P>*>  _segCmds;
 162     std::unordered_map<uint32_t,uint32_t>   _oldToNewSymbolIndexes;
 163     std::vector<const char*>                _reExportPaths;
 164     std::vector<const char*>                _downDependentPaths;
 165     std::vector<const char*>                _allDependentPaths;
 166     std::vector<uint64_t>                   _initializerAddresses;
 167     std::vector<macho_section<P>*>          _dofSections;
 168     uint32_t                                _newWeakBindingInfoOffset       = 0;
 169     uint32_t                                _newLazyBindingInfoOffset       = 0;
 170     uint32_t                                _newBindingInfoOffset           = 0;
 171     uint32_t                                _newExportInfoOffset            = 0;
 172     uint32_t                                _exportInfoSize                 = 0;
 173     uint32_t                                _newWeakBindingSize             = 0;
 174     uint32_t                                _newExportedSymbolsStartIndex   = 0;
 175     uint32_t                                _newExportedSymbolCount         = 0;
 176     uint32_t                                _newImportedSymbolsStartIndex   = 0;
 177     uint32_t                                _newImportedSymbolCount         = 0;
 178     uint32_t                                _newLocalSymbolsStartIndex      = 0;
 179     uint32_t                                _newLocalSymbolCount            = 0;
 180     uint32_t                                _newFunctionStartsOffset        = 0;
 181     uint32_t                                _newDataInCodeOffset            = 0;
 182     uint32_t                                _newIndirectSymbolTableOffset   = 0;
 183     uint64_t                                _dyldSectionAddr                = 0;
 184 };
 185
 186
 187
 188 template <typename P>
 189 class AcceleratorTables {
 190 public:
 191                 AcceleratorTables(DyldSharedCache* cache, uint64_t linkeditStartAddr, Diagnostics& diag, const std::vector<LinkeditOptimizer<P>*>& optimizers);
 192
 193     uint32_t    totalSize() const;
 194     void        copyTo(uint8_t* buffer);
 195
 196 private:
 197     typedef typename P::E  E;
 198
 199     struct NodeChain;
 200
 201     struct DepNode {
 202         std::vector<DepNode*>   _dependents;
 203         unsigned                _depth;
 204         const char*             _installName;
 205
 206                                 DepNode() : _depth(0), _installName(nullptr) { }
 207         void                    computeDepth();
 208         static void             verifyUnreachable(DepNode* target, NodeChain& chain, Diagnostics& diag, std::unordered_set<DepNode*>& visitedNodes, const std::vector<DepNode*>& from);
 209     };
 210
 211     struct NodeChain {
 212         NodeChain*   prev;
 213         DepNode*     node;
 214     };
 215
 216     std::unordered_map<macho_header<P>*, DepNode>                   _depDAG;
 217     std::vector<dyld_cache_image_info_extra>                        _extraInfo;
 218     std::vector<uint8_t>                                            _trieBytes;
 219     std::vector<uint16_t>                                           _reExportArray;
 220     std::vector<uint16_t>                                           _dependencyArray;
 221     std::vector<uint16_t>                                           _bottomUpArray;
 222     std::vector<dyld_cache_accelerator_initializer>                 _initializers;
 223     std::vector<dyld_cache_accelerator_dof>                         _dofSections;
 224     std::vector<dyld_cache_range_entry>                             _rangeTable;
 225     std::unordered_map<macho_header<P>*, uint32_t>                  _machHeaderToImageIndex;
 226     std::unordered_map<std::string, macho_header<P>*>               _dylibPathToMachHeader;
 227     std::unordered_map<macho_header<P>*, LinkeditOptimizer<P>*>     _machHeaderToOptimizer;
 228     dyld_cache_accelerator_info                                     _acceleratorInfoHeader;
 229 };
 230
 231
 232 template <typename P>
 233 void AcceleratorTables<P>::AcceleratorTables::DepNode::verifyUnreachable(AcceleratorTables<P>::DepNode* target, struct AcceleratorTables<P>::NodeChain& chain, Diagnostics& diag,
 234                                                                         std::unordered_set<DepNode*>& visitedNodes, const std::vector<AcceleratorTables<P>::DepNode*>& from) {
 235     for (DepNode* node : from) {
 236         bool foundCycle = (node == target);
 237         for (NodeChain* c = &chain; c->prev != nullptr; c = c->prev) {
 238             if ( c->node == target ) {
 239                 foundCycle = true;
 240                 break;
 241             }
 242         }
 243         if ( foundCycle ) {
 244             NodeChain* chp = &chain;
 245             std::string msg = std::string("found cycle for ") + target->_installName;
 246             while (chp != nullptr) {
 247                 msg = msg + "\n  " + chp->node->_installName;
 248                 chp = chp->prev;
 249             }
 250             diag.warning("%s", msg.c_str());
 251             return;
 252         }
 253
 254         if ( visitedNodes.count(node) )
 255             continue;
 256         visitedNodes.insert(node);
 257         NodeChain nextChain;
 258         nextChain.prev = &chain;
 259         nextChain.node = node;
 260         verifyUnreachable(target, nextChain, diag, visitedNodes, node->_dependents);
 261     }
 262 }
 263
 264 const uint16_t kBranchIslandDylibIndex = 0x7FFF;
 265
 266 template <typename P>
 267 AcceleratorTables<P>::AcceleratorTables(DyldSharedCache* cache, uint64_t linkeditStartAddr, Diagnostics& diag, const std::vector<LinkeditOptimizer<P>*>& optimizers)
 268 {
 269     // build table mapping tables to map between mach_header, index, and optimizer
 270     for ( LinkeditOptimizer<P>* op : optimizers ) {
 271         _machHeaderToOptimizer[op->machHeader()] = op;
 272     }
 273     const dyld_cache_mapping_info* mappings = (dyld_cache_mapping_info*)((uint8_t*)cache + cache->header.mappingOffset);
 274     uint64_t cacheStartAddress = mappings[0].address;
 275     const dyld_cache_image_info* images = (dyld_cache_image_info*)((uint8_t*)cache + cache->header.imagesOffset);
 276     for (unsigned i=0; i < cache->header.imagesCount; ++i) {
 277         uint64_t segCacheFileOffset = images[i].address - cacheStartAddress;
 278         macho_header<P>* mhMapped = (macho_header<P>*)((uint8_t*)cache+segCacheFileOffset);
 279         const char* path = (char*)cache + images[i].pathFileOffset;
 280         _dylibPathToMachHeader[path] = mhMapped;
 281         // don't add alias entries (path offset in pool near start of cache) to header->index map
 282         if ( images[i].pathFileOffset > segCacheFileOffset )
 283             _machHeaderToImageIndex[mhMapped] = i;
 284     }
 285
 286
 287     // build DAG of image dependencies
 288     for (LinkeditOptimizer<P>* op : optimizers) {
 289         _depDAG[op->machHeader()]._installName = op->installName();
 290     }
 291     for (LinkeditOptimizer<P>* op : optimizers) {
 292         DepNode& node = _depDAG[op->machHeader()];
 293         for (const char* depPath : op->getDownwardDependents()) {
 294             macho_header<P>* depMH = _dylibPathToMachHeader[depPath];
 295             if ( depMH != nullptr ) {
 296                 DepNode* depNode = &_depDAG[depMH];
 297                 node._dependents.push_back(depNode);
 298             }
 299         }
 300     }
 301
 302     // check for cycles in DAG
 303     for (auto& entry : _depDAG) {
 304         DepNode* node = &entry.second;
 305         NodeChain chain;
 306         chain.prev = nullptr;
 307         chain.node = node;
 308         std::unordered_set<DepNode*> visitedNodes;
 309         DepNode::verifyUnreachable(node, chain, diag, visitedNodes, node->_dependents);
 310     }
 311
 312     // compute depth for each DAG node
 313     for (auto& entry : _depDAG) {
 314         entry.second.computeDepth();
 315     }
 316
 317     // build sorted (bottom up) list of images
 318     std::vector<macho_header<P>*> sortedMachHeaders;
 319     sortedMachHeaders.reserve(optimizers.size());
 320     for (LinkeditOptimizer<P>* op : optimizers) {
 321         if ( strcmp(op->installName(), "dyld_shared_cache_branch_islands") != 0 )
 322             sortedMachHeaders.push_back(op->machHeader());
 323         else
 324             _machHeaderToImageIndex[op->machHeader()] = kBranchIslandDylibIndex;
 325     }
 326     std::sort(sortedMachHeaders.begin(), sortedMachHeaders.end(),
 327               [&](macho_header<P>* lmh, macho_header<P>* rmh) -> bool {
 328                 if ( _depDAG[lmh]._depth != _depDAG[rmh]._depth )
 329                     return (_depDAG[lmh]._depth < _depDAG[rmh]._depth);
 330                 else
 331                     return (lmh < rmh);
 332               });
 333
 334     // build zeroed array of extra infos
 335     dyld_cache_image_info_extra emptyExtra;
 336     emptyExtra.exportsTrieAddr              = 0;
 337     emptyExtra.weakBindingsAddr             = 0;
 338     emptyExtra.exportsTrieSize              = 0;
 339     emptyExtra.weakBindingsSize             = 0;
 340     emptyExtra.dependentsStartArrayIndex    = 0;
 341     emptyExtra.reExportsStartArrayIndex     = 0;
 342     _extraInfo.insert(_extraInfo.begin(), sortedMachHeaders.size(), emptyExtra);
 343
 344     //for ( macho_header<P>* mh : sortedMachHeaders ) {
 345     //    fprintf(stderr, "depth: %3d mh: %p  path: %s\n", _depDAG[mh]._depth, mh, _machHeaderToOptimizer[mh]->installName());
 346     //}
 347
 348     // build dependency table
 349     _dependencyArray.push_back(0xFFFF); // reserve 0 slot to be "no-dependencies"
 350     for (macho_header<P>* mh : sortedMachHeaders) {
 351         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 352         unsigned index = _machHeaderToImageIndex[mh];
 353         auto depPaths = op->getAllDependents();
 354         if ( depPaths.empty() ) {
 355             _extraInfo[index].dependentsStartArrayIndex = 0;
 356         }
 357         else {
 358             _extraInfo[index].dependentsStartArrayIndex = (uint32_t)_dependencyArray.size();
 359             auto downPaths = op->getDownwardDependents();
 360             for (const char* depPath : depPaths) {
 361                 macho_header<P>* depMH = _dylibPathToMachHeader[depPath];
 362                 uint16_t depIndex = _machHeaderToImageIndex[depMH];
 363                 if ( std::find(downPaths.begin(), downPaths.end(), depPath) == downPaths.end())
 364                     depIndex |= 0x8000;
 365                 _dependencyArray.push_back(depIndex);
 366             }
 367             _dependencyArray.push_back(0xFFFF); // mark end of list
 368        }
 369     }
 370
 371     // build re-exports table
 372     _reExportArray.push_back(0xFFFF); // reserve 0 slot to be "no-re-exports"
 373     for (macho_header<P>* mh : sortedMachHeaders) {
 374         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 375         unsigned index = _machHeaderToImageIndex[mh];
 376         auto reExPaths = op->getReExportPaths();
 377         if ( reExPaths.empty() ) {
 378             _extraInfo[index].reExportsStartArrayIndex = 0;
 379         }
 380         else {
 381             _extraInfo[index].reExportsStartArrayIndex = (uint32_t)_reExportArray.size();
 382             for (const char* reExPath : reExPaths) {
 383                 macho_header<P>* reExMH = _dylibPathToMachHeader[reExPath];
 384                 uint32_t reExIndex = _machHeaderToImageIndex[reExMH];
 385                 _reExportArray.push_back(reExIndex);
 386             }
 387             _reExportArray.push_back(0xFFFF); // mark end of list
 388        }
 389     }
 390
 391     // build ordered list of initializers
 392     for (macho_header<P>* mh : sortedMachHeaders) {
 393         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 394         unsigned index = _machHeaderToImageIndex[mh];
 395         _bottomUpArray.push_back(index);
 396         for (uint64_t initializer : op->initializerAddresses()) {
 397             //fprintf(stderr, "0x%08llX %s\n", initializer, op->installName());
 398             dyld_cache_accelerator_initializer entry;
 399             entry.functionOffset = (uint32_t)(initializer-cacheStartAddress);
 400             entry.imageIndex = _machHeaderToImageIndex[mh];
 401             _initializers.push_back(entry);
 402         }
 403     }
 404
 405     // build ordered list of DOF sections
 406     for (macho_header<P>* mh : sortedMachHeaders) {
 407         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 408         assert(op != NULL);
 409         unsigned imageIndex = _machHeaderToImageIndex[mh];
 410         for (auto& sect : op->dofSections()) {
 411             //fprintf(stderr, "0x%08llX %s\n", initializer, op->installName());
 412             dyld_cache_accelerator_dof entry;
 413             entry.sectionAddress = sect->addr();
 414             entry.sectionSize    = (uint32_t)sect->size();
 415             entry.imageIndex     = imageIndex;
 416             _dofSections.push_back(entry);
 417         }
 418     }
 419
 420
 421     // register exports trie and weak binding info in each dylib with image extra info
 422     for (macho_header<P>* mh : sortedMachHeaders) {
 423         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 424         unsigned index = _machHeaderToImageIndex[mh];
 425         _extraInfo[index].exportsTrieAddr  = op->exportsTrieLinkEditOffset() + linkeditStartAddr;
 426         _extraInfo[index].exportsTrieSize  = op->exportsTrieLinkEditSize();
 427         _extraInfo[index].weakBindingsAddr = op->weakBindingLinkEditOffset() + linkeditStartAddr;
 428         _extraInfo[index].weakBindingsSize = op->weakBindingLinkEditSize();
 429     }
 430
 431     // record location of __DATA/__dyld section in libdyld.dylib
 432     macho_header<P>* libdyldMH = _dylibPathToMachHeader["/usr/lib/system/libdyld.dylib"];
 433     LinkeditOptimizer<P>* libdyldOp = _machHeaderToOptimizer[libdyldMH];
 434     uint64_t dyldSectionAddr = libdyldOp->dyldSectionAddress();
 435
 436     // build range table for fast address->image lookups
 437     for (macho_header<P>* mh : sortedMachHeaders) {
 438         LinkeditOptimizer<P>* op = _machHeaderToOptimizer[mh];
 439         unsigned imageIndex = _machHeaderToImageIndex[mh];
 440         for (const macho_segment_command<P>* segCmd : op->segCmds()) {
 441             if ( strcmp(segCmd->segname(), "__LINKEDIT") == 0 )
 442                 continue;
 443             dyld_cache_range_entry entry;
 444             entry.startAddress = segCmd->vmaddr();
 445             entry.size         = (uint32_t)segCmd->vmsize();
 446             entry.imageIndex   = imageIndex;
 447             _rangeTable.push_back(entry);
 448         }
 449     }
 450     std::sort(_rangeTable.begin(), _rangeTable.end(),
 451               [&](const dyld_cache_range_entry& lRange, const dyld_cache_range_entry& rRange) -> bool {
 452                 return (lRange.startAddress < rRange.startAddress);
 453               });
 454
 455     // build trie that maps install names to image index
 456     std::vector<DylibIndexTrie::Entry> dylibEntrys;
 457     for (auto &x : _dylibPathToMachHeader) {
 458         const std::string& path = x.first;
 459         unsigned index = _machHeaderToImageIndex[x.second];
 460         dylibEntrys.push_back(DylibIndexTrie::Entry(path, DylibIndex(index)));
 461     }
 462     DylibIndexTrie dylibsTrie(dylibEntrys);
 463     dylibsTrie.emit(_trieBytes);
 464     while ( (_trieBytes.size() % 4) != 0 )
 465         _trieBytes.push_back(0);
 466
 467     // fill out header
 468     _acceleratorInfoHeader.version              = 1;
 469     _acceleratorInfoHeader.imageExtrasCount     = (uint32_t)_extraInfo.size();
 470     _acceleratorInfoHeader.imagesExtrasOffset   = ALIGN_AS_TYPE(sizeof(dyld_cache_accelerator_info), dyld_cache_image_info_extra);
 471     _acceleratorInfoHeader.bottomUpListOffset   = _acceleratorInfoHeader.imagesExtrasOffset + _acceleratorInfoHeader.imageExtrasCount*sizeof(dyld_cache_image_info_extra);
 472     _acceleratorInfoHeader.dylibTrieOffset      = _acceleratorInfoHeader.bottomUpListOffset + _acceleratorInfoHeader.imageExtrasCount*sizeof(uint16_t);
 473     _acceleratorInfoHeader.dylibTrieSize        = (uint32_t)_trieBytes.size();
 474     _acceleratorInfoHeader.initializersOffset   = ALIGN_AS_TYPE(_acceleratorInfoHeader.dylibTrieOffset + _acceleratorInfoHeader.dylibTrieSize, dyld_cache_accelerator_initializer);
 475     _acceleratorInfoHeader.initializersCount    = (uint32_t)_initializers.size();
 476     _acceleratorInfoHeader.dofSectionsOffset    = ALIGN_AS_TYPE(_acceleratorInfoHeader.initializersOffset + _acceleratorInfoHeader.initializersCount*sizeof(dyld_cache_accelerator_initializer), dyld_cache_accelerator_initializer);
 477     _acceleratorInfoHeader.dofSectionsCount     = (uint32_t)_dofSections.size();
 478     _acceleratorInfoHeader.reExportListOffset   = ALIGN_AS_TYPE(_acceleratorInfoHeader.dofSectionsOffset + _acceleratorInfoHeader.dofSectionsCount*sizeof(dyld_cache_accelerator_dof), dyld_cache_accelerator_dof);
 479     _acceleratorInfoHeader.reExportCount        = (uint32_t)_reExportArray.size();
 480     _acceleratorInfoHeader.depListOffset        = ALIGN_AS_TYPE(_acceleratorInfoHeader.reExportListOffset + _acceleratorInfoHeader.reExportCount*sizeof(uint16_t), uint16_t);
 481     _acceleratorInfoHeader.depListCount         = (uint32_t)_dependencyArray.size();
 482     _acceleratorInfoHeader.rangeTableOffset     = ALIGN_AS_TYPE(_acceleratorInfoHeader.depListOffset + _acceleratorInfoHeader.depListCount*sizeof(uint16_t), dyld_cache_range_entry);
 483     _acceleratorInfoHeader.rangeTableCount      = (uint32_t)_rangeTable.size();
 484     _acceleratorInfoHeader.dyldSectionAddr      = dyldSectionAddr;
 485 }
 486
 487
 488 template <typename P>
 489 void AcceleratorTables<P>::DepNode::computeDepth()
 490 {
 491     if ( _depth != 0 )
 492         return;
 493     _depth = 1;
 494     for (DepNode* node : _dependents) {
 495         node->computeDepth();
 496         if ( node->_depth >= _depth )
 497             _depth = node->_depth + 1;
 498     }
 499 }
 500
 501 template <typename P>
 502 uint32_t AcceleratorTables<P>::totalSize() const
 503 {
 504     return (uint32_t)align(_acceleratorInfoHeader.rangeTableOffset + _acceleratorInfoHeader.rangeTableCount*sizeof(dyld_cache_range_entry), 14);
 505 }
 506
 507 template <typename P>
 508 void AcceleratorTables<P>::copyTo(uint8_t* buffer)
 509 {
 510     memcpy(buffer, &_acceleratorInfoHeader, sizeof(dyld_cache_accelerator_info));
 511     memcpy(&buffer[_acceleratorInfoHeader.imagesExtrasOffset], &_extraInfo[0],       _extraInfo.size()*sizeof(dyld_cache_image_info_extra));
 512     memcpy(&buffer[_acceleratorInfoHeader.bottomUpListOffset], &_bottomUpArray[0],   _bottomUpArray.size()*sizeof(uint16_t));
 513     memcpy(&buffer[_acceleratorInfoHeader.initializersOffset], &_initializers[0],    _initializers.size()*sizeof(dyld_cache_accelerator_initializer));
 514     memcpy(&buffer[_acceleratorInfoHeader.reExportListOffset], &_reExportArray[0],   _reExportArray.size()*sizeof(uint16_t));
 515     memcpy(&buffer[_acceleratorInfoHeader.dofSectionsOffset],  &_dofSections[0],     _dofSections.size()*sizeof(dyld_cache_accelerator_dof));
 516     memcpy(&buffer[_acceleratorInfoHeader.depListOffset],      &_dependencyArray[0], _dependencyArray.size()*sizeof(uint16_t));
 517     memcpy(&buffer[_acceleratorInfoHeader.rangeTableOffset],   &_rangeTable[0],      _rangeTable.size()*sizeof(dyld_cache_range_entry));
 518     memcpy(&buffer[_acceleratorInfoHeader.dylibTrieOffset],    &_trieBytes[0],       _trieBytes.size());
 519 }
 520
 521
 522
 523 template <typename P>
 524 LinkeditOptimizer<P>::LinkeditOptimizer(void* cacheBuffer, macho_header<P>* mh, Diagnostics& diag)
 525 : _mh(mh), _cacheBuffer(cacheBuffer), _diagnostics(diag)
 526 {
 527     const unsigned origLoadCommandsSize = mh->sizeofcmds();
 528     unsigned bytesRemaining = origLoadCommandsSize;
 529     unsigned removedCount = 0;
 530     uint64_t    textSegAddr = 0;
 531     int64_t     slide = 0;
 532     const macho_load_command<P>* const cmds = (macho_load_command<P>*)((uint8_t*)mh + sizeof(macho_header<P>));
 533     const uint32_t cmdCount = mh->ncmds();
 534     const macho_load_command<P>* cmd = cmds;
 535     const macho_dylib_command<P>* dylibCmd;
 536     const macho_routines_command<P>* routinesCmd;
 537     macho_segment_command<P>* segCmd;
 538     for (uint32_t i = 0; i < cmdCount; ++i) {
 539         bool remove = false;
 540         switch (cmd->cmd()) {
 541             case LC_ID_DYLIB:
 542                 _installName = ((macho_dylib_command<P>*)cmd)->name();
 543                 break;
 544             case LC_SYMTAB:
 545                 _symTabCmd = (macho_symtab_command<P>*)cmd;
 546                 break;
 547             case LC_DYSYMTAB:
 548                 _dynSymTabCmd = (macho_dysymtab_command<P>*)cmd;
 549                 break;
 550             case LC_DYLD_INFO:
 551             case LC_DYLD_INFO_ONLY:
 552                 _dyldInfo = (macho_dyld_info_command<P>*)cmd;
 553                 _exportInfoSize = _dyldInfo->export_size();
 554                 break;
 555             case LC_FUNCTION_STARTS:
 556                 _functionStartsCmd = (macho_linkedit_data_command<P>*)cmd;
 557                 break;
 558             case LC_DATA_IN_CODE:
 559                 _dataInCodeCmd = (macho_linkedit_data_command<P>*)cmd;
 560                 break;
 561             case LC_ROUTINES:
 562             case LC_ROUTINES_64:
 563                 routinesCmd = (macho_routines_command<P>*)cmd;
 564                 _initializerAddresses.push_back(routinesCmd->init_address());
 565                 break;
 566             case LC_REEXPORT_DYLIB:
 567             case LC_LOAD_DYLIB:
 568             case LC_LOAD_WEAK_DYLIB:
 569             case LC_LOAD_UPWARD_DYLIB:
 570                 dylibCmd = (macho_dylib_command<P>*)cmd;
 571                 _allDependentPaths.push_back(dylibCmd->name());
 572                 if ( cmd->cmd() != LC_LOAD_UPWARD_DYLIB )
 573                     _downDependentPaths.push_back(dylibCmd->name());
 574                 if ( cmd->cmd() == LC_REEXPORT_DYLIB )
 575                     _reExportPaths.push_back(dylibCmd->name());
 576                break;
 577             case macho_segment_command<P>::CMD:
 578                 segCmd = (macho_segment_command<P>*)cmd;
 579                 _segCmds.push_back(segCmd);
 580                 if ( strcmp(segCmd->segname(), "__TEXT") == 0 ) {
 581                     textSegAddr = segCmd->vmaddr();
 582                     slide = (uint64_t)mh - textSegAddr;
 583                 }
 584                 else if ( strcmp(segCmd->segname(), "__LINKEDIT") == 0 ) {
 585                     _linkeditAddr        = segCmd->vmaddr();
 586                     _linkeditBias        = (uint8_t*)mh + (_linkeditAddr - textSegAddr) - segCmd->fileoff();
 587                     _linkeditSize        = (uint32_t)segCmd->vmsize();
 588                 }
 589                 else if ( segCmd->nsects() > 0 ) {
 590                     macho_section<P>* const sectionsStart = (macho_section<P>*)((uint8_t*)segCmd + sizeof(macho_segment_command<P>));
 591                     macho_section<P>* const sectionsEnd = &sectionsStart[segCmd->nsects()];
 592                     for (macho_section<P>* sect=sectionsStart; sect < sectionsEnd; ++sect) {
 593                         const uint8_t type = sect->flags() & SECTION_TYPE;
 594                         if ( type == S_MOD_INIT_FUNC_POINTERS ) {
 595                             const pint_t* inits = (pint_t*)(sect->addr()+slide);
 596                             const size_t count = sect->size() / sizeof(pint_t);
 597                             for (size_t j=0; j < count; ++j) {
 598                                 uint64_t func = P::getP(inits[j]);
 599                                 _initializerAddresses.push_back(func);
 600                             }
 601                         }
 602                           else if ( type == S_DTRACE_DOF ) {
 603                             _dofSections.push_back(sect);
 604                         }
 605                         else if ( (strcmp(sect->sectname(), "__dyld") == 0) && (strncmp(sect->segname(), "__DATA", 6) == 0) ) {
 606                             _dyldSectionAddr = sect->addr();
 607                         }
 608                     }
 609                 }
 610                 break;
 611             case LC_SEGMENT_SPLIT_INFO:
 612                 remove = true;
 613                 break;
 614         }
 615         uint32_t cmdSize = cmd->cmdsize();
 616         macho_load_command<P>* nextCmd = (macho_load_command<P>*)(((uint8_t*)cmd)+cmdSize);
 617         if ( remove ) {
 618             ::memmove((void*)cmd, (void*)nextCmd, bytesRemaining);
 619             ++removedCount;
 620         }
 621         else {
 622             bytesRemaining -= cmdSize;
 623             cmd = nextCmd;
 624         }
 625     }
 626     // zero out stuff removed
 627     ::bzero((void*)cmd, bytesRemaining);
 628     // update header
 629     mh->set_ncmds(cmdCount - removedCount);
 630     mh->set_sizeofcmds(origLoadCommandsSize - bytesRemaining);
 631 }
 632
 633 /*
 634 static void dumpLoadCommands(const uint8_t* mheader)
 635 {
 636     const mach_header* const mh = (mach_header*)mheader;
 637     const uint32_t cmd_count = mh->ncmds;
 638     bool is64 = (mh->magic == MH_MAGIC_64);
 639     const load_command* cmds = (load_command*)(mheader + (is64 ? sizeof(mach_header_64) : sizeof(mach_header)));
 640     const load_command* cmd = cmds;
 641     const segment_command* segCmd;
 642     const segment_command_64* seg64Cmd;
 643     const symtab_command* symTab;
 644     const linkedit_data_command* leData;
 645     const uint8_t* linkEditBias = NULL;
 646     for (uint32_t i = 0; i < cmd_count; ++i) {
 647         switch (cmd->cmd) {
 648             case LC_SEGMENT:
 649                 segCmd = (const segment_command*)cmd;
 650                 printf("LC_SEGMENT\n");
 651                 printf("  segname  = %s\n",     segCmd->segname);
 652                 printf("  vmaddr   = 0x%08X\n", segCmd->vmaddr);
 653                 printf("  vmsize   = 0x%08X\n", segCmd->vmsize);
 654                 printf("  fileoff  = 0x%08X\n", segCmd->fileoff);
 655                 printf("  filesize = 0x%08X\n", segCmd->filesize);
 656                 if ( strcmp(segCmd->segname, "__TEXT") == 0 ) {
 657                     linkEditBias = mheader - segCmd->fileoff;
 658                 }
 659                  break;
 660             case LC_SEGMENT_64:
 661                 seg64Cmd = (const segment_command_64*)cmd;
 662                 printf("LC_SEGMENT_64\n");
 663                 printf("  segname  = %s\n",        seg64Cmd->segname);
 664                 printf("  vmaddr   = 0x%09llX\n",  seg64Cmd->vmaddr);
 665                 printf("  vmsize   = 0x%09llX\n",  seg64Cmd->vmsize);
 666                 printf("  fileoff  = 0x%09llX\n",  seg64Cmd->fileoff);
 667                 printf("  filesize = 0x%09llX\n",  seg64Cmd->filesize);
 668                 if ( strcmp(seg64Cmd->segname, "__TEXT") == 0 ) {
 669                     linkEditBias = mheader - seg64Cmd->fileoff;
 670                 }
 671                 break;
 672             case LC_SYMTAB:
 673                 symTab = (const symtab_command*)cmd;
 674                 printf("LC_SYMTAB\n");
 675                 printf("  symoff   = 0x%08X\n", symTab->symoff);
 676                 printf("  nsyms    = 0x%08X\n", symTab->nsyms);
 677                 printf("  stroff   = 0x%08X\n", symTab->stroff);
 678                 printf("  strsize  = 0x%08X\n", symTab->strsize);
 679                 {
 680                 const char* strPool = (char*)&linkEditBias[symTab->stroff];
 681                 const nlist_64* sym0 = (nlist_64*)(&linkEditBias[symTab->symoff]);
 682                 printf("    sym[0].n_strx = 0x%08X (%s)\n", sym0->n_un.n_strx, &strPool[sym0->n_un.n_strx]);
 683                 printf("    sym[0].n_type = 0x%02X\n", sym0->n_type);
 684                 printf("    sym[0].n_sect = 0x%02X\n", sym0->n_sect);
 685                 printf("    sym[0].n_desc = 0x%04X\n", sym0->n_desc);
 686                 printf("    sym[0].n_value = 0x%llX\n", sym0->n_value);
 687                 const nlist_64* sym1 = (nlist_64*)(&linkEditBias[symTab->symoff+16]);
 688                 printf("    sym[1].n_strx = 0x%08X (%s)\n", sym1->n_un.n_strx, &strPool[sym1->n_un.n_strx]);
 689                 printf("    sym[1].n_type = 0x%02X\n", sym1->n_type);
 690                 printf("    sym[1].n_sect = 0x%02X\n", sym1->n_sect);
 691                 printf("    sym[1].n_desc = 0x%04X\n", sym1->n_desc);
 692                 printf("    sym[1].n_value = 0x%llX\n", sym1->n_value);
 693                 }
 694                  break;
 695             case LC_FUNCTION_STARTS:
 696                 leData = (const linkedit_data_command*)cmd;
 697                 printf("LC_FUNCTION_STARTS\n");
 698                 printf("  dataoff  = 0x%08X\n", leData->dataoff);
 699                 printf("  datasize = 0x%08X\n", leData->datasize);
 700             default:
 701                 //printf("0x%08X\n", cmd->cmd);
 702                 break;
 703         }
 704         cmd = (const load_command*)(((uint8_t*)cmd)+cmd->cmdsize);
 705     }
 706 }
 707 */
 708
 709 template <typename P>
 710 void LinkeditOptimizer<P>::updateLoadCommands(uint32_t mergedLinkeditStartOffset, uint64_t mergedLinkeditAddr, uint64_t newLinkeditSize,
 711                                               uint32_t sharedSymbolTableStartOffset, uint32_t sharedSymbolTableCount,
 712                                               uint32_t sharedSymbolStringsOffset, uint32_t sharedSymbolStringsSize)
 713 {
 714     // update __LINKEDIT segment in all dylibs to overlap the same shared region
 715     for (macho_segment_command<P>* segCmd : _segCmds) {
 716         if ( strcmp(segCmd->segname(), "__LINKEDIT") == 0 ) {
 717             segCmd->set_vmaddr(mergedLinkeditAddr);
 718             segCmd->set_vmsize(newLinkeditSize);
 719             segCmd->set_fileoff(mergedLinkeditStartOffset);
 720             segCmd->set_filesize(newLinkeditSize);
 721         }
 722         else if ( strcmp(segCmd->segname(), "__TEXT") == 0 ) {
 723             // HACK until lldb fixed in: <rdar://problem/20357466> DynamicLoaderMacOSXDYLD fixes for Monarch dyld shared cache
 724             //segCmd->set_fileoff(0);
 725
 726         }
 727    }
 728
 729     // update symbol table to point to shared symbol table
 730     _symTabCmd->set_symoff(mergedLinkeditStartOffset + sharedSymbolTableStartOffset + _newLocalSymbolsStartIndex*sizeof(macho_nlist<P>));
 731     _symTabCmd->set_nsyms(_newLocalSymbolCount+_newExportedSymbolCount+_newImportedSymbolCount);
 732     _symTabCmd->set_stroff(mergedLinkeditStartOffset + sharedSymbolStringsOffset);
 733     _symTabCmd->set_strsize(sharedSymbolStringsSize);
 734
 735     // update dynamic symbol table to have proper offsets into shared symbol table
 736     _dynSymTabCmd->set_ilocalsym(0);
 737     _dynSymTabCmd->set_nlocalsym(_newLocalSymbolCount);
 738     _dynSymTabCmd->set_iextdefsym(_newExportedSymbolsStartIndex-_newLocalSymbolsStartIndex);
 739     _dynSymTabCmd->set_nextdefsym(_newExportedSymbolCount);
 740     _dynSymTabCmd->set_iundefsym(_newImportedSymbolsStartIndex-_newLocalSymbolsStartIndex);
 741     _dynSymTabCmd->set_nundefsym(_newImportedSymbolCount);
 742     _dynSymTabCmd->set_tocoff(0);
 743     _dynSymTabCmd->set_ntoc(0);
 744     _dynSymTabCmd->set_modtaboff(0);
 745     _dynSymTabCmd->set_nmodtab(0);
 746     _dynSymTabCmd->set_indirectsymoff(mergedLinkeditStartOffset + _newIndirectSymbolTableOffset);
 747     _dynSymTabCmd->set_extreloff(0);
 748     _dynSymTabCmd->set_locreloff(0);
 749     _dynSymTabCmd->set_nlocrel(0);
 750
 751     // update dyld info
 752     if ( _dyldInfo != nullptr ) {
 753         _dyldInfo->set_rebase_off(0);
 754         _dyldInfo->set_rebase_size(0);
 755         _dyldInfo->set_bind_off(_dyldInfo->bind_size() ?  mergedLinkeditStartOffset + _newBindingInfoOffset : 0);
 756         _dyldInfo->set_weak_bind_off(_dyldInfo->weak_bind_size() ?  mergedLinkeditStartOffset + _newWeakBindingInfoOffset : 0 );
 757         _dyldInfo->set_lazy_bind_off(_dyldInfo->lazy_bind_size() ?  mergedLinkeditStartOffset + _newLazyBindingInfoOffset : 0 );
 758         _dyldInfo->set_export_off(mergedLinkeditStartOffset + _newExportInfoOffset);
 759     }
 760
 761     // update function-starts
 762     if ( _functionStartsCmd != nullptr )
 763         _functionStartsCmd->set_dataoff(mergedLinkeditStartOffset+_newFunctionStartsOffset);
 764
 765     // update data-in-code
 766     if ( _dataInCodeCmd != nullptr )
 767         _dataInCodeCmd->set_dataoff(mergedLinkeditStartOffset+_newDataInCodeOffset);
 768 }
 769
 770 template <typename P>
 771 void LinkeditOptimizer<P>::copyWeakBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset)
 772 {
 773     if ( _dyldInfo == nullptr )
 774         return;
 775     unsigned size = _dyldInfo->weak_bind_size();
 776     if ( size != 0 ) {
 777         ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_dyldInfo->weak_bind_off()], size);
 778         _newWeakBindingInfoOffset = offset;
 779         _newWeakBindingSize = size;
 780         offset += size;
 781     }
 782 }
 783
 784
 785 template <typename P>
 786 void LinkeditOptimizer<P>::copyLazyBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset)
 787 {
 788     if ( _dyldInfo == nullptr )
 789         return;
 790     unsigned size = _dyldInfo->lazy_bind_size();
 791     if ( size != 0 ) {
 792         ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_dyldInfo->lazy_bind_off()], size);
 793         _newLazyBindingInfoOffset = offset;
 794         offset += size;
 795     }
 796 }
 797
 798 template <typename P>
 799 void LinkeditOptimizer<P>::copyBindingInfo(uint8_t* newLinkEditContent, uint32_t& offset)
 800 {
 801     if ( _dyldInfo == nullptr )
 802         return;
 803     unsigned size = _dyldInfo->bind_size();
 804     if ( size != 0 ) {
 805         ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_dyldInfo->bind_off()], size);
 806         _newBindingInfoOffset = offset;
 807         offset += size;
 808     }
 809 }
 810
 811 template <typename P>
 812 void LinkeditOptimizer<P>::copyExportInfo(uint8_t* newLinkEditContent, uint32_t& offset)
 813 {
 814     if ( _dyldInfo == nullptr )
 815         return;
 816     unsigned size = _dyldInfo->export_size();
 817     if ( size != 0 ) {
 818         ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_dyldInfo->export_off()], size);
 819         _newExportInfoOffset = offset;
 820         offset += size;
 821     }
 822 }
 823
 824
 825 template <typename P>
 826 void LinkeditOptimizer<P>::copyFunctionStarts(uint8_t* newLinkEditContent, uint32_t& offset)
 827 {
 828     if ( _functionStartsCmd == nullptr )
 829         return;
 830     unsigned size = _functionStartsCmd->datasize();
 831     ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_functionStartsCmd->dataoff()], size);
 832     _newFunctionStartsOffset = offset;
 833     offset += size;
 834 }
 835
 836 template <typename P>
 837 void LinkeditOptimizer<P>::copyDataInCode(uint8_t* newLinkEditContent, uint32_t& offset)
 838 {
 839     if ( _dataInCodeCmd == nullptr )
 840         return;
 841     unsigned size = _dataInCodeCmd->datasize();
 842     ::memcpy(&newLinkEditContent[offset], &_linkeditBias[_dataInCodeCmd->dataoff()], size);
 843     _newDataInCodeOffset = offset;
 844     offset += size;
 845 }
 846
 847
 848 template <typename P>
 849 void LinkeditOptimizer<P>::copyLocalSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex,
 850                                             bool redact, std::vector<LocalSymbolInfo>& localSymbolInfos,
 851                                             std::vector<macho_nlist<P>>& unmappedLocalSymbols, SortedStringPool<P>& localSymbolsStringPool)
 852 {
 853     LocalSymbolInfo localInfo;
 854     localInfo.dylibOffset = (uint32_t)(((uint8_t*)_mh) - (uint8_t*)_cacheBuffer);
 855     localInfo.nlistStartIndex = (uint32_t)unmappedLocalSymbols.size();
 856     localInfo.nlistCount = 0;
 857     _newLocalSymbolsStartIndex = symbolIndex;
 858     const char* strings = (char*)&_linkeditBias[_symTabCmd->stroff()];
 859     const macho_nlist<P>* const symbolTable = (macho_nlist<P>*)(&_linkeditBias[_symTabCmd->symoff()]);
 860     const macho_nlist<P>* const firstExport = &symbolTable[_dynSymTabCmd->ilocalsym()];
 861     const macho_nlist<P>* const lastExport  = &symbolTable[_dynSymTabCmd->ilocalsym()+_dynSymTabCmd->nlocalsym()];
 862     for (const macho_nlist<P>* entry = firstExport; entry < lastExport; ++entry) {
 863         if ( (entry->n_type() & N_TYPE) != N_SECT)
 864             continue;
 865          if ( (entry->n_type() & N_STAB) != 0)
 866             continue;
 867         const char* name = &strings[entry->n_strx()];
 868         macho_nlist<P>* newSymbolEntry = (macho_nlist<P>*)&newLinkEditContent[offset];
 869         *newSymbolEntry = *entry;
 870         if ( redact ) {
 871             // if removing local symbols, change __text symbols to "<redacted>" so backtraces don't have bogus names
 872             if ( entry->n_sect() == 1 ) {
 873                 stringPool.add(symbolIndex, "<redacted>");
 874                 ++symbolIndex;
 875                 offset += sizeof(macho_nlist<P>);
 876             }
 877             // copy local symbol to unmmapped locals area
 878             localSymbolsStringPool.add((uint32_t)unmappedLocalSymbols.size(), name);
 879             unmappedLocalSymbols.push_back(*entry);
 880             unmappedLocalSymbols.back().set_n_strx(0);
 881         }
 882         else {
 883             stringPool.add(symbolIndex, name);
 884             ++symbolIndex;
 885             offset += sizeof(macho_nlist<P>);
 886        }
 887     }
 888     _newLocalSymbolCount = symbolIndex - _newLocalSymbolsStartIndex;
 889     localInfo.nlistCount = (uint32_t)unmappedLocalSymbols.size() - localInfo.nlistStartIndex;
 890     localSymbolInfos.push_back(localInfo);
 891 }
 892
 893
 894 template <typename P>
 895 void LinkeditOptimizer<P>::copyExportedSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex)
 896 {
 897     _newExportedSymbolsStartIndex = symbolIndex;
 898     const char* strings = (char*)&_linkeditBias[_symTabCmd->stroff()];
 899     const macho_nlist<P>* const symbolTable = (macho_nlist<P>*)(&_linkeditBias[_symTabCmd->symoff()]);
 900     const macho_nlist<P>* const firstExport = &symbolTable[_dynSymTabCmd->iextdefsym()];
 901     const macho_nlist<P>* const lastExport  = &symbolTable[_dynSymTabCmd->iextdefsym()+_dynSymTabCmd->nextdefsym()];
 902     uint32_t oldSymbolIndex = _dynSymTabCmd->iextdefsym();
 903     for (const macho_nlist<P>* entry = firstExport; entry < lastExport; ++entry, ++oldSymbolIndex) {
 904         if ( (entry->n_type() & N_TYPE) != N_SECT)
 905             continue;
 906         const char* name = &strings[entry->n_strx()];
 907         if ( strncmp(name, ".objc_", 6) == 0 )
 908             continue;
 909         if ( strncmp(name, "$ld$", 4) == 0 )
 910             continue;
 911         macho_nlist<P>* newSymbolEntry = (macho_nlist<P>*)&newLinkEditContent[offset];
 912         *newSymbolEntry = *entry;
 913         newSymbolEntry->set_n_strx(0);
 914         stringPool.add(symbolIndex, name);
 915         _oldToNewSymbolIndexes[oldSymbolIndex] = symbolIndex - _newLocalSymbolsStartIndex;
 916         ++symbolIndex;
 917         offset += sizeof(macho_nlist<P>);
 918     }
 919     _newExportedSymbolCount = symbolIndex - _newExportedSymbolsStartIndex;
 920 }
 921
 922 template <typename P>
 923 void LinkeditOptimizer<P>::copyImportedSymbols(uint8_t* newLinkEditContent, SortedStringPool<P>& stringPool, uint32_t& offset, uint32_t& symbolIndex)
 924 {
 925     _newImportedSymbolsStartIndex = symbolIndex;
 926     const char* strings = (char*)&_linkeditBias[_symTabCmd->stroff()];
 927     const macho_nlist<P>* const symbolTable = (macho_nlist<P>*)(&_linkeditBias[_symTabCmd->symoff()]);
 928     const macho_nlist<P>* const firstImport = &symbolTable[_dynSymTabCmd->iundefsym()];
 929     const macho_nlist<P>* const lastImport  = &symbolTable[_dynSymTabCmd->iundefsym()+_dynSymTabCmd->nundefsym()];
 930     uint32_t oldSymbolIndex = _dynSymTabCmd->iundefsym();
 931     for (const macho_nlist<P>* entry = firstImport; entry < lastImport; ++entry, ++oldSymbolIndex) {
 932         if ( (entry->n_type() & N_TYPE) != N_UNDF)
 933             continue;
 934         const char* name = &strings[entry->n_strx()];
 935         macho_nlist<P>* newSymbolEntry = (macho_nlist<P>*)&newLinkEditContent[offset];
 936         *newSymbolEntry = *entry;
 937         newSymbolEntry->set_n_strx(0);
 938         stringPool.add(symbolIndex, name);
 939         _oldToNewSymbolIndexes[oldSymbolIndex] = symbolIndex - _newLocalSymbolsStartIndex;
 940         ++symbolIndex;
 941         offset += sizeof(macho_nlist<P>);
 942     }
 943     _newImportedSymbolCount = symbolIndex - _newImportedSymbolsStartIndex;
 944 }
 945
 946 template <typename P>
 947 void LinkeditOptimizer<P>::copyIndirectSymbolTable(uint8_t* newLinkEditContent, uint32_t& offset)
 948 {
 949     _newIndirectSymbolTableOffset = offset;
 950      const uint32_t* const indirectTable = (uint32_t*)&_linkeditBias[_dynSymTabCmd->indirectsymoff()];
 951     uint32_t* newIndirectTable = (uint32_t*)&newLinkEditContent[offset];
 952     for (uint32_t i=0; i < _dynSymTabCmd->nindirectsyms(); ++i) {
 953         uint32_t symbolIndex = E::get32(indirectTable[i]);
 954         if ( (symbolIndex == INDIRECT_SYMBOL_ABS) || (symbolIndex == INDIRECT_SYMBOL_LOCAL) )
 955             E::set32(newIndirectTable[i], symbolIndex);
 956         else
 957             E::set32(newIndirectTable[i], _oldToNewSymbolIndexes[symbolIndex]);
 958         offset += sizeof(uint32_t);
 959     }
 960 }
 961
 962 template <typename P>
 963 void LinkeditOptimizer<P>::mergeLinkedits(CacheBuilder& builder, std::vector<LinkeditOptimizer<P>*>& optimizers)
 964 {
 965     // allocate space for new linkedit data
 966     uint64_t totalUnoptLinkeditsSize = builder._readOnlyRegion.sizeInUse - builder._nonLinkEditReadOnlySize;
 967     uint8_t* newLinkEdit = (uint8_t*)calloc(totalUnoptLinkeditsSize, 1);
 968     SortedStringPool<P> stringPool;
 969     uint32_t offset = 0;
 970
 971     builder._diagnostics.verbose("Merged LINKEDIT:\n");
 972
 973     // copy weak binding info
 974     uint32_t startWeakBindInfosOffset = offset;
 975     for (LinkeditOptimizer<P>* op : optimizers) {
 976         // Skip chained fixups as the in-place linked list isn't valid any more
 977         const dyld3::MachOFile* mf = (dyld3::MachOFile*)op->machHeader();
 978         if (!mf->hasChainedFixups())
 979             op->copyWeakBindingInfo(newLinkEdit, offset);
 980     }
 981     builder._diagnostics.verbose("  weak bindings size:      %5uKB\n", (uint32_t)(offset-startWeakBindInfosOffset)/1024);
 982
 983     // copy export info
 984     uint32_t startExportInfosOffset = offset;
 985     for (LinkeditOptimizer<P>* op : optimizers) {
 986         op->copyExportInfo(newLinkEdit, offset);
 987     }
 988     builder._diagnostics.verbose("  exports info size:       %5uKB\n", (uint32_t)(offset-startExportInfosOffset)/1024);
 989
 990     // in theory, an optimized cache can drop the binding info
 991     if ( true ) {
 992         // copy binding info
 993         uint32_t startBindingsInfosOffset = offset;
 994         for (LinkeditOptimizer<P>* op : optimizers) {
 995             // Skip chained fixups as the in-place linked list isn't valid any more
 996             const dyld3::MachOFile* mf = (dyld3::MachOFile*)op->machHeader();
 997             if (!mf->hasChainedFixups())
 998                 op->copyBindingInfo(newLinkEdit, offset);
 999         }
1000         builder._diagnostics.verbose("  bindings size:           %5uKB\n", (uint32_t)(offset-startBindingsInfosOffset)/1024);
1001
1002        // copy lazy binding info
1003         uint32_t startLazyBindingsInfosOffset = offset;
1004         for (LinkeditOptimizer<P>* op : optimizers) {
1005             // Skip chained fixups as the in-place linked list isn't valid any more
1006             const dyld3::MachOFile* mf = (dyld3::MachOFile*)op->machHeader();
1007             if (!mf->hasChainedFixups())
1008                 op->copyLazyBindingInfo(newLinkEdit, offset);
1009         }
1010         builder._diagnostics.verbose("  lazy bindings size:      %5uKB\n", (offset-startLazyBindingsInfosOffset)/1024);
1011     }
1012
1013     // copy symbol table entries
1014     std::vector<macho_nlist<P>> unmappedLocalSymbols;
1015     if ( builder._options.excludeLocalSymbols )
1016         unmappedLocalSymbols.reserve(0x01000000);
1017     std::vector<LocalSymbolInfo> localSymbolInfos;
1018         localSymbolInfos.reserve(optimizers.size());
1019     SortedStringPool<P> localSymbolsStringPool;
1020     uint32_t symbolIndex = 0;
1021     const uint32_t sharedSymbolTableStartOffset = offset;
1022     uint32_t sharedSymbolTableExportsCount = 0;
1023     uint32_t sharedSymbolTableImportsCount = 0;
1024     for (LinkeditOptimizer<P>* op : optimizers) {
1025          op->copyLocalSymbols(newLinkEdit, stringPool, offset, symbolIndex, builder._options.excludeLocalSymbols,
1026                              localSymbolInfos, unmappedLocalSymbols, localSymbolsStringPool);
1027         uint32_t x = symbolIndex;
1028         op->copyExportedSymbols(newLinkEdit, stringPool, offset, symbolIndex);
1029         sharedSymbolTableExportsCount += (symbolIndex-x);
1030         uint32_t y = symbolIndex;
1031         op->copyImportedSymbols(newLinkEdit, stringPool, offset, symbolIndex);
1032         sharedSymbolTableImportsCount += (symbolIndex-y);
1033     }
1034     uint32_t sharedSymbolTableCount = symbolIndex;
1035     const uint32_t sharedSymbolTableEndOffset = offset;
1036
1037     // copy function starts
1038     uint32_t startFunctionStartsOffset = offset;
1039     for (LinkeditOptimizer<P>* op : optimizers) {
1040         op->copyFunctionStarts(newLinkEdit, offset);
1041     }
1042     builder._diagnostics.verbose("  function starts size:    %5uKB\n", (offset-startFunctionStartsOffset)/1024);
1043
1044     // copy data-in-code info
1045     uint32_t startDataInCodeOffset = offset;
1046     for (LinkeditOptimizer<P>* op : optimizers) {
1047         op->copyDataInCode(newLinkEdit, offset);
1048     }
1049     builder._diagnostics.verbose("  data in code size:       %5uKB\n", (offset-startDataInCodeOffset)/1024);
1050
1051     // copy indirect symbol tables
1052     for (LinkeditOptimizer<P>* op : optimizers) {
1053         op->copyIndirectSymbolTable(newLinkEdit, offset);
1054     }
1055     // if indirect table has odd number of entries, end will not be 8-byte aligned
1056     if ( (offset % sizeof(typename P::uint_t)) != 0 )
1057         offset += 4;
1058
1059     // copy string pool
1060     uint32_t sharedSymbolStringsOffset = offset;
1061     uint32_t sharedSymbolStringsSize = stringPool.copyPoolAndUpdateOffsets((char*)&newLinkEdit[sharedSymbolStringsOffset], (macho_nlist<P>*)&newLinkEdit[sharedSymbolTableStartOffset]);
1062     offset += sharedSymbolStringsSize;
1063     uint32_t newLinkeditUnalignedSize = offset;
1064     uint64_t newLinkeditAlignedSize = align(offset, 14);
1065     builder._diagnostics.verbose("  symbol table size:       %5uKB (%d exports, %d imports)\n", (sharedSymbolTableEndOffset-sharedSymbolTableStartOffset)/1024, sharedSymbolTableExportsCount, sharedSymbolTableImportsCount);
1066     builder._diagnostics.verbose("  symbol string pool size: %5uKB\n", sharedSymbolStringsSize/1024);
1067     builder._sharedStringsPoolVmOffset = (uint32_t)((builder._readOnlyRegion.unslidLoadAddress - builder._readExecuteRegion.unslidLoadAddress) + builder._nonLinkEditReadOnlySize + sharedSymbolStringsOffset);
1068
1069     // overwrite mapped LINKEDIT area in cache with new merged LINKEDIT content
1070     builder._diagnostics.verbose("LINKEDITS optimized from %uMB to %uMB\n", (uint32_t)totalUnoptLinkeditsSize/(1024*1024), (uint32_t)newLinkeditUnalignedSize/(1024*1024));
1071     ::memcpy(builder._readOnlyRegion.buffer+builder._nonLinkEditReadOnlySize, newLinkEdit, newLinkeditAlignedSize);
1072     ::free(newLinkEdit);
1073     builder._readOnlyRegion.sizeInUse = builder._nonLinkEditReadOnlySize + newLinkeditAlignedSize;
1074
1075     // If making cache for customers, add extra accelerator tables for dyld
1076     DyldSharedCache* cacheHeader = (DyldSharedCache*)builder._readExecuteRegion.buffer;
1077     if ( builder._options.optimizeStubs ) {
1078         uint64_t addrWhereAccTablesWillBe     = builder._readOnlyRegion.unslidLoadAddress+builder._readOnlyRegion.sizeInUse;
1079         uint64_t addrWhereMergedLinkWillStart = builder._readOnlyRegion.unslidLoadAddress+builder._nonLinkEditReadOnlySize;
1080         AcceleratorTables<P> tables(cacheHeader, addrWhereMergedLinkWillStart, builder._diagnostics, optimizers);
1081         uint32_t tablesSize = tables.totalSize();
1082         if ( tablesSize < (builder._readOnlyRegion.bufferSize - builder._readOnlyRegion.sizeInUse) ) {
1083             tables.copyTo(builder._readOnlyRegion.buffer+builder._readOnlyRegion.sizeInUse);
1084             cacheHeader->header.accelerateInfoAddr = addrWhereAccTablesWillBe;
1085             cacheHeader->header.accelerateInfoSize = tablesSize;
1086             builder._readOnlyRegion.sizeInUse += align(tablesSize, 14);
1087             builder._diagnostics.verbose("Accelerator tables %uMB\n", (uint32_t)tablesSize/(1024*1024));
1088        }
1089         else {
1090             builder._diagnostics.warning("not enough room to add dyld accelerator tables");
1091         }
1092     }
1093
1094     // overwrite end of un-opt linkedits to create a new unmapped region for local symbols
1095     if ( builder._options.excludeLocalSymbols ) {
1096         const uint32_t entriesOffset = sizeof(dyld_cache_local_symbols_info);
1097         const uint32_t entriesCount  = (uint32_t)localSymbolInfos.size();
1098         const uint32_t nlistOffset   = (uint32_t)align(entriesOffset + entriesCount * sizeof(dyld_cache_local_symbols_info), 4); // 16-byte align start
1099         const uint32_t nlistCount    = (uint32_t)unmappedLocalSymbols.size();
1100         const uint32_t stringsSize   = (uint32_t)localSymbolsStringPool.size();
1101         const uint32_t stringsOffset = nlistOffset + nlistCount * sizeof(macho_nlist<P>);
1102         // allocate buffer for local symbols
1103         const size_t localsBufferSize = align(stringsOffset + stringsSize, 14);
1104         vm_address_t localsBuffer;
1105         if ( ::vm_allocate(mach_task_self(), &localsBuffer, localsBufferSize, VM_FLAGS_ANYWHERE) == 0 ) {
1106             dyld_cache_local_symbols_info* infoHeader = (dyld_cache_local_symbols_info*)localsBuffer;
1107             // fill in header info
1108             infoHeader->nlistOffset       = nlistOffset;
1109             infoHeader->nlistCount        = nlistCount;
1110             infoHeader->stringsOffset     = stringsOffset;
1111             infoHeader->stringsSize       = stringsSize;
1112             infoHeader->entriesOffset     = entriesOffset;
1113             infoHeader->entriesCount      = entriesCount;
1114             // copy info for each dylib
1115             dyld_cache_local_symbols_entry* entries = (dyld_cache_local_symbols_entry*)(((uint8_t*)infoHeader)+entriesOffset);
1116             for (uint32_t i=0; i < entriesCount; ++i) {
1117                 entries[i].dylibOffset        = localSymbolInfos[i].dylibOffset;
1118                 entries[i].nlistStartIndex    = localSymbolInfos[i].nlistStartIndex;
1119                 entries[i].nlistCount         = localSymbolInfos[i].nlistCount;
1120             }
1121             // copy nlists
1122             macho_nlist<P>* newLocalsSymbolTable = (macho_nlist<P>*)(localsBuffer+nlistOffset);
1123             ::memcpy(newLocalsSymbolTable, &unmappedLocalSymbols[0], nlistCount*sizeof(macho_nlist<P>));
1124             // copy string pool
1125             localSymbolsStringPool.copyPoolAndUpdateOffsets(((char*)infoHeader)+stringsOffset, newLocalsSymbolTable);
1126             // update cache header
1127             cacheHeader->header.localSymbolsSize    = localsBufferSize;
1128             // return buffer of local symbols, caller to free() it
1129             builder._localSymbolsRegion.buffer      = (uint8_t*)localsBuffer;
1130             builder._localSymbolsRegion.bufferSize  = localsBufferSize;
1131             builder._localSymbolsRegion.sizeInUse   = localsBufferSize;
1132         }
1133         else {
1134             builder._diagnostics.warning("could not allocate local symbols");
1135         }
1136     }
1137
1138     // update all load commands to new merged layout
1139     uint64_t linkeditsUnslidStartAddr = builder._readOnlyRegion.unslidLoadAddress + builder._nonLinkEditReadOnlySize;
1140     uint32_t linkeditsCacheFileOffset = (uint32_t)(builder._readOnlyRegion.cacheFileOffset + builder._nonLinkEditReadOnlySize);
1141     for (LinkeditOptimizer<P>* op : optimizers) {
1142         op->updateLoadCommands(linkeditsCacheFileOffset, linkeditsUnslidStartAddr, newLinkeditUnalignedSize,
1143                                sharedSymbolTableStartOffset, sharedSymbolTableCount,
1144                                sharedSymbolStringsOffset, sharedSymbolStringsSize);
1145     }
1146 }
1147
1148
1149 template <typename P>
1150 void LinkeditOptimizer<P>::optimizeLinkedit(CacheBuilder& builder)
1151 {
1152     DyldSharedCache* cache = (DyldSharedCache*)builder._readExecuteRegion.buffer;
1153     // construct a LinkeditOptimizer for each image
1154     __block std::vector<LinkeditOptimizer<P>*> optimizers;
1155     cache->forEachImage(^(const mach_header* mh, const char*) {
1156         optimizers.push_back(new LinkeditOptimizer<P>(cache, (macho_header<P>*)mh, builder._diagnostics));
1157     });
1158 #if 0
1159     // add optimizer for each branch pool
1160     for (uint64_t poolOffset : branchPoolOffsets) {
1161         macho_header<P>* mh =  (macho_header<P>*)((char*)cache + poolOffset);
1162         optimizers.push_back(new LinkeditOptimizer<P>(cache, mh, diag));
1163     }
1164 #endif
1165     // merge linkedit info
1166     mergeLinkedits(builder, optimizers);
1167
1168     // delete optimizers
1169     for (LinkeditOptimizer<P>* op : optimizers)
1170         delete op;
1171 }
1172
1173 void CacheBuilder::optimizeLinkedit(const std::vector<uint64_t>& branchPoolOffsets)
1174 {
1175     if ( _archLayout->is64 ) {
1176         return LinkeditOptimizer<Pointer64<LittleEndian>>::optimizeLinkedit(*this);
1177     }
1178     else {
1179          return LinkeditOptimizer<Pointer32<LittleEndian>>::optimizeLinkedit(*this);
1180     }
1181 }
1182
1183
1184