Loading...
--- dyld/dyld-1165.3/cache_builder/NewSharedCacheBuilder.cpp
+++ dyld/dyld-1330/cache_builder/NewSharedCacheBuilder.cpp
@@ -24,9 +24,11 @@
#include "Defines.h"
#include "NewSharedCacheBuilder.h"
+#include "MachOFile.h"
#include "NewAdjustDylibSegments.h"
#include "CacheDylib.h"
#include "ClosureFileSystem.h"
+#include "JSONReader.h"
#include "JSONWriter.h"
#include "StringUtils.h"
#include "Array.h"
@@ -38,11 +40,26 @@
#include "JustInTimeLoader.h"
#include "OptimizerObjC.h"
#include "OptimizerSwift.h"
+#include "Platform.h"
#include "PrebuiltLoader.h"
#include "DyldProcessConfig.h"
#include "DyldRuntimeState.h"
#include "SwiftVisitor.h"
#include "ParallelUtils.h"
+#include "CString.h"
+#include "Version32.h"
+#include "ExternalGenericMetadataBuilderImport.h"
+#include "SnapshotShared.h"
+#include "AAREncoder.h"
+#include <SharedCacheLinker/SharedCacheLinker.h>
+#include "ThreadLocalVariables.h"
+#include "Algorithm.h"
+
+// mach_o
+#include "ExportsTrie.h"
+#include "Header.h"
+#include "Image.h"
+#include "Misc.h"
// FIXME: Remove this once we don't write to the old objc header struct. See emitObjCOptsHeader()
#include "objc-shared-cache.h"
@@ -51,20 +68,31 @@
#include <list>
#include <mach-o/nlist.h>
#include <sstream>
+#include <sys/fcntl.h>
+#include <sys/stat.h>
#include <unordered_set>
-using dyld3::GradedArchs;
using dyld3::MachOFile;
using dyld4::JustInTimeLoader;
using dyld4::KernelArgs;
using dyld4::Loader;
using dyld4::ProcessConfig;
+using dyld4::RuntimeLocks;
using dyld4::RuntimeState;
using dyld4::SyscallDelegate;
using dyld4::RuntimeLocks;
-using lsl::EphemeralAllocator;
+using lsl::Allocator;
+
+using mach_o::Header;
+using mach_o::Platform;
+using mach_o::Image;
+using mach_o::ExportsTrie;
+using mach_o::Symbol;
+using mach_o::LinkedDylibAttributes;
+using mach_o::Version32;
+using error::Error;
using metadata_visitor::SwiftConformance;
using metadata_visitor::SwiftVisitor;
@@ -90,14 +118,30 @@
// Note, don't change the form of this message without checking in with MRM, as they
// parse it. We really need to add structured errors/warnings some time
std::string reason = "Dylib located at '" + inputFile.path + "' not placed in shared cache because: ";
- callback(reason + inputFile.getError().message());
- }
+ for ( const error::Error& err : inputFile.getErrors() )
+ callback(reason + err.message());
+ }
+ }
+
+ for ( const std::string& warning : this->warnings )
+ callback(warning);
+}
+
+void SharedCacheBuilder::forEachError(void (^callback)(const std::string_view& str)) const
+{
+ for ( const std::string& str : this->errors ) {
+ callback(str);
}
}
void SharedCacheBuilder::forEachCacheDylib(void (^callback)(const std::string_view& path)) const
{
for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
+ // skip Swift prespecialized dylib if it's been built
+ // it's synthesized by the builder, so mrm doesn't need to remove it
+ if ( swiftPrespecializedDylib && &cacheDylib == swiftPrespecializedDylib )
+ continue;
+
// Note this has to return the path, not the install name, as MRM uses this to delete
// the path from disk
callback(cacheDylib.inputFile->path);
@@ -111,39 +155,210 @@
}
}
+struct DylibSizes
+{
+ uint64_t textSegment = 0;
+ uint64_t dataSegment = 0;
+ uint64_t dataConstSegment = 0;
+ uint64_t dataDirtySegment = 0;
+ uint64_t authSegment = 0;
+ uint64_t authConstSegment = 0;
+ uint64_t linkeditSegment = 0;
+
+ // any other segment not in the above
+ uint64_t otherSegment = 0;
+
+ // Just the __TEXT __text section
+ uint64_t textSection = 0;
+
+ // parts of the linkedit
+ uint64_t exportTrie = 0;
+ uint64_t symbolTable = 0;
+ uint64_t symbolStrings = 0;
+};
+
+static DylibSizes getDylibSizes(std::span<const SubCache> subCaches,
+ std::span<const CacheDylib> cacheDylibs)
+{
+ __block DylibSizes dylibSizes;
+ __block std::unordered_set<uint64_t> linkeditSizes;
+
+ // segment sizes
+ for ( const CacheDylib& cacheDylib : cacheDylibs ) {
+ cacheDylib.cacheHdr->forEachSegment(^(const Header::SegmentInfo& info, bool& stop) {
+ if ( info.segmentName == "__TEXT" )
+ dylibSizes.textSegment += info.vmsize;
+ else if ( info.segmentName == "__DATA" )
+ dylibSizes.dataSegment += info.vmsize;
+ else if ( info.segmentName == "__DATA_CONST" )
+ dylibSizes.dataConstSegment += info.vmsize;
+ else if ( info.segmentName == "__DATA_DIRTY" )
+ dylibSizes.dataDirtySegment += info.vmsize;
+ else if ( info.segmentName == "__AUTH" )
+ dylibSizes.authSegment += info.vmsize;
+ else if ( info.segmentName == "__AUTH_CONST" )
+ dylibSizes.authConstSegment += info.vmsize;
+ else if ( info.segmentName == "__LINKEDIT" )
+ linkeditSizes.insert(info.vmsize);
+ else
+ dylibSizes.otherSegment += info.vmsize;
+ });
+ }
+ for ( uint64_t size : linkeditSizes )
+ dylibSizes.linkeditSegment += size;
+
+ // section sizes
+ for ( const CacheDylib& cacheDylib : cacheDylibs ) {
+ cacheDylib.cacheHdr->forEachSection(^(const Header::SectionInfo& info, bool& stop) {
+ if ( (info.segmentName == "__TEXT") && (info.sectionName == "__text") )
+ dylibSizes.textSection += info.size;
+ });
+ }
+
+ // linkedit sizes
+ for ( const CacheDylib& cacheDylib : cacheDylibs ) {
+ for ( const LinkeditDataChunk& chunk : cacheDylib.linkeditChunks ) {
+ if ( chunk.kind == Chunk::Kind::linkeditExportTrie )
+ dylibSizes.exportTrie += chunk.cacheVMSize.rawValue();
+ }
+
+ if ( !cacheDylib.optimizedSymbols.nlist64.empty() ) {
+ dylibSizes.symbolTable += (sizeof(struct nlist_64) * cacheDylib.optimizedSymbols.nlist64.size());
+ } else {
+ dylibSizes.symbolTable += (sizeof(struct nlist) * cacheDylib.optimizedSymbols.nlist64.size());
+ }
+ }
+
+ for ( const SubCache& subCache : subCaches ) {
+ if ( subCache.optimizedSymbolStrings )
+ dylibSizes.symbolStrings += subCache.optimizedSymbolStrings->cacheVMSize.rawValue();
+ }
+
+ return dylibSizes;
+}
+
+std::string SharedCacheBuilder::stats(uint64_t startTimeNanos) const
+{
+ uint64_t endTimeNanos = clock_gettime_nsec_np(CLOCK_UPTIME_RAW);
+ uint64_t totalTime = (endTimeNanos - startTimeNanos) / 1000000000;
+
+ DylibSizes dylibSizes = getDylibSizes(this->subCaches, this->cacheDylibs);
+
+ std::string result;
+ result += "{\n";
+
+ // add all the top level fields
+ result += std::string(" \"version\": 1,\n");
+ result += std::string(" \"minor-version\": 0,\n");
+ result += std::string(" \"update\": \"") + this->options.updateName + "\",\n";
+ result += std::string(" \"device\": \"") + this->options.deviceName + "\",\n";
+ result += std::string(" \"platform\": \"") + this->options.platform.name().c_str() + "\",\n";
+ result += std::string(" \"arch\": \"") + this->options.arch.name() + "\",\n";
+
+ // main stats
+ result += std::string(" \"dylibs-count\": ") + std::to_string(this->cacheDylibs.size()) + ",\n";
+ result += std::string(" \"total-vm-size\": ") + std::to_string(this->totalVMSize.rawValue()) + ",\n";
+ result += std::string(" \"total-text-segment-size\": ") + std::to_string(dylibSizes.textSegment) + ",\n";
+ result += std::string(" \"total-data-segment-size\": ") + std::to_string(dylibSizes.dataSegment) + ",\n";
+ result += std::string(" \"total-data-dirty-segment-size\": ") + std::to_string(dylibSizes.dataDirtySegment) + ",\n";
+ result += std::string(" \"total-data-const-segment-size\": ") + std::to_string(dylibSizes.dataConstSegment) + ",\n";
+ result += std::string(" \"total-auth-segment-size\": ") + std::to_string(dylibSizes.authSegment) + ",\n";
+ result += std::string(" \"total-auth-const-segment-size\": ") + std::to_string(dylibSizes.authConstSegment) + ",\n";
+ result += std::string(" \"total-other-segment-size\": ") + std::to_string(dylibSizes.otherSegment) + ",\n";
+ result += std::string(" \"total-linkedit-segment-size\": ") + std::to_string(dylibSizes.linkeditSegment) + ",\n";
+ result += std::string(" \"total-text-section-size\": ") + std::to_string(dylibSizes.textSection) + ",\n";
+
+ // optimization stats
+ result += std::string(" \"objc-selector-hash-table-size\": ") + std::to_string(this->objcSelectorOptimizer.selectorHashTableTotalByteSize) + ",\n";
+ result += std::string(" \"objc-class-hash-table-size\": ") + std::to_string(this->objcClassOptimizer.classHashTableTotalByteSize) + ",\n";
+ result += std::string(" \"objc-protocol-hash-table-size\": ") + std::to_string(this->objcProtocolOptimizer.protocolHashTableTotalByteSize) + ",\n";
+ result += std::string(" \"objc-selectors-count\": ") + std::to_string(this->objcSelectorOptimizer.selectorsMap.size()) + ",\n";
+ result += std::string(" \"objc-classes-count\": ") + std::to_string(this->objcClassOptimizer.namesMap.size()) + ",\n";
+ result += std::string(" \"objc-protocols-count\": ") + std::to_string(this->objcProtocolOptimizer.namesMap.size()) + ",\n";
+
+ // linkedit breakdown
+ result += std::string(" \"linkedit-export-trie-size\": ") + std::to_string(dylibSizes.exportTrie) + ",\n";
+ result += std::string(" \"linkedit-symbol-table-size\": ") + std::to_string(dylibSizes.symbolTable) + ",\n";
+ result += std::string(" \"linkedit-symbol-strings-size\": ") + std::to_string(dylibSizes.symbolStrings) + ",\n";
+
+ // cache metadata
+ result += std::string(" \"patch-table-size\": ") + std::to_string(this->patchTableOptimizer.patchTableChunk->cacheVMSize.rawValue()) + ",\n";
+ result += std::string(" \"dylib-loaders-size\": ") + std::to_string(this->prebuiltLoaderBuilder.cacheDylibsLoaderSize) + ",\n";
+ result += std::string(" \"executable-loaders-size\": ") + std::to_string(this->prebuiltLoaderBuilder.executablesLoaderSize) + ",\n";
+
+ // Times
+ result += std::string(" \"total-build-time\": ") + std::to_string(totalTime) + "\n";
+
+ result += "}\n";
+ return result;
+}
+
void SharedCacheBuilder::addFile(const void* buffer, size_t bufferSize, std::string_view path,
uint64_t inode, uint64_t modTime, bool forceNotCacheEligible)
{
+ const bool isOSBinary = false;
+ uint64_t sliceOffset = 0;
+ uint64_t sliceLen = 0;
+
+ // On macOS, also allow iOSMac dylibs
+ if ( this->options.platform == Platform::macOS ) {
+ Diagnostics diag;
+ if ( const MachOFile* mf = MachOFile::compatibleSlice(diag, sliceOffset, sliceLen, buffer, bufferSize, path.data(),
+ Platform::macCatalyst, isOSBinary,
+ this->options.gradedArchs) ) {
+ InputFile inputFile;
+ inputFile.mf = mf;
+ inputFile.inode = inode;
+ inputFile.mtime = modTime;
+ inputFile.size = sliceLen;
+ inputFile.path = path;
+ inputFile.forceNotCacheEligible = forceNotCacheEligible;
+ allInputFiles.push_back(std::move(inputFile));
+ return;
+ }
+ }
+
Diagnostics diag;
- const bool isOSBinary = false;
- if ( const MachOFile* mf = MachOFile::compatibleSlice(diag, buffer, bufferSize, path.data(),
+ if ( const MachOFile* mf = MachOFile::compatibleSlice(diag, sliceOffset, sliceLen, buffer, bufferSize, path.data(),
this->options.platform, isOSBinary,
- this->options.archs) ) {
+ this->options.gradedArchs) ) {
InputFile inputFile;
inputFile.mf = mf;
inputFile.inode = inode;
inputFile.mtime = modTime;
+ inputFile.size = sliceLen;
inputFile.path = path;
inputFile.forceNotCacheEligible = forceNotCacheEligible;
allInputFiles.push_back(std::move(inputFile));
return;
}
- // On macOS, also allow iOSMac dylibs
- if ( this->options.platform == dyld3::Platform::macOS ) {
- diag.clearError();
- if ( const MachOFile* mf = MachOFile::compatibleSlice(diag, buffer, bufferSize, path.data(),
- dyld3::Platform::iOSMac, isOSBinary,
- this->options.archs) ) {
- InputFile inputFile;
- inputFile.mf = mf;
- inputFile.inode = inode;
- inputFile.mtime = modTime;
- inputFile.path = path;
- inputFile.forceNotCacheEligible = forceNotCacheEligible;
- allInputFiles.push_back(std::move(inputFile));
+ // If we have an error, then add an input file just in case its useful for later
+ if ( diag.hasError() ) {
+ // Only add dylibs to track for later error emission
+ __block bool isDylib = false;
+ const std::span<uint8_t> bufferSpan = { (uint8_t*)buffer, bufferSize };
+ mach_o::Error parseErr = mach_o::forEachHeader(bufferSpan, path,
+ ^(const mach_o::Header* mh, size_t sliceHeader, bool& stop) {
+ if ( mh->isDylib() ) {
+ isDylib = true;
+ stop = true;
+ }
+ });
+
+ if ( !isDylib )
return;
- }
+
+ InputFile inputFile;
+ inputFile.mf = nullptr;
+ inputFile.inode = 0;
+ inputFile.mtime = 0;
+ inputFile.size = 0;
+ inputFile.path = path;
+ inputFile.forceNotCacheEligible = true;
+ inputFile.addError(Error("%s", (const char*)diag.errorMessageCStr()));
+ allInputFiles.push_back(std::move(inputFile));
+ return;
}
}
@@ -153,6 +368,7 @@
this->inputAliases = aliases;
this->inputIntermediateAliases = intermediateAliases;
}
+
//
// MARK: --- SharedCacheBuilder build methods ---
@@ -167,6 +383,9 @@
if ( this->allInputFiles.empty() )
return Error("Cannot build cache with no inputs");
+ // Reserve a slot for the Swift prespecialized dylib early, so that it can be ordered
+ this->reserveSwiftPrespecializedInputFile();
+
this->categorizeInputs();
this->verifySelfContained();
@@ -174,11 +393,24 @@
return Error("Cannot build cache with no dylibs");
this->sortDylibs();
+ this->sortExecutables();
+
+ // Note this needs to be after sorting, so the order of objc dylibs is consistent with all dylibs list
+ this->findObjCDylibs();
+
+ // ObjC dylibs order is now set, so we can create the Swift prespecialized dylib
+ // Note this needs to happen after order is known because the Swift dylib needs to
+ // known indices of other shared cache dylibs. To create the dylib earlier we would need
+ // to add split seg support for dylib indices.
+ if ( Error error = this->createSwiftPrespecializedDylib() ) {
+ swiftPrespecializedDylibBuildError = error.message();
+ return error;
+ }
// Note this needs to be after sorting, as aliases point to the cache dylibs
this->calculateDylibAliases();
- if ( Error error = this->calculateDylibDependents(); error.hasError() )
+ if ( Error error = this->calculateDylibDependents() )
return error;
this->categorizeDylibSegments();
@@ -195,7 +427,6 @@
Error SharedCacheBuilder::estimateGlobalOptimizations()
{
this->estimateIMPCaches();
- this->findObjCDylibs();
this->findCanonicalObjCSelectors();
this->findCanonicalObjCClassNames();
this->findCanonicalObjCProtocolNames();
@@ -210,8 +441,9 @@
this->estimateSwiftHashTableSizes();
this->calculateCacheDylibsTrie();
- this->estimatePatchTableSize();
+ this->estimateFunctionVariantsSize();
this->estimateCacheLoadersSize();
+ this->estimatePrewarmingSize();
this->setupStubOptimizer();
@@ -224,6 +456,10 @@
// Outputs: subCaches
Error SharedCacheBuilder::createSubCaches()
{
+ Timer::Scope timedScope(this->config, "createSubCaches time");
+ Timer::AggregateTimer aggregateTimerOwner(this->config);
+ auto& aggregateTimer = aggregateTimerOwner;
+
this->computeSubCaches();
// Per-subCache optimizations
@@ -232,8 +468,54 @@
if ( Error error = this->calculateUniqueGOTs(); error.hasError() )
return error;
+ // Finalize the subcaches, by removing any unused regions
+ for ( SubCache& subCache : this->subCaches ) {
+ // Add slide info for each DATA/AUTH segment. Do this after we've added any other DATA*
+ // segments
+ if ( this->config.slideInfo.slideInfoFormat.has_value() )
+ subCache.addSlideInfoChunks();
+
+ subCache.removeEmptyRegions();
+ }
+
+ // Analyze the dylibs to work out binds and patch table info
+ {
+ this->initializePatchTableOptimizer();
+
+ // Because blocks...
+ std::vector<const CacheDylib*> builderCacheDylibsOwner;
+ auto& builderCacheDylibs = builderCacheDylibsOwner;
+ for ( const CacheDylib& cacheDylib : this->cacheDylibs )
+ builderCacheDylibs.push_back(&cacheDylib);
+
+ Error err = parallel::forEach(this->cacheDylibs, ^(size_t index, CacheDylib& cacheDylib) {
+ Diagnostics diag;
+ PatchInfo& dylibPatchInfo = this->patchTableOptimizer.patchInfos[cacheDylib.cacheIndex];
+
+ std::vector<Error> symbolErrors = cacheDylib.calculateBindTargets(diag, this->config, aggregateTimer, builderCacheDylibs,
+ dylibPatchInfo);
+ if ( !symbolErrors.empty() ) {
+ for ( const Error& symbolErr : symbolErrors )
+ this->errors.push_back(symbolErr.message());
+ }
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ cacheDylib.calcuatePatchInfo(diag, this->config, aggregateTimer, dylibPatchInfo);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ return Error();
+ });
+
+ if ( err )
+ return err;
+ }
+
this->sortSubCacheSegments();
this->calculateSlideInfoSize();
+ if ( Error error = this->calculatePatchTableSize(); error.hasError() )
+ return error;
this->calculateCodeSignatureSize();
this->printSubCaches();
if ( Error error = this->computeSubCacheLayout(); error.hasError() )
@@ -278,16 +560,15 @@
Timer::AggregateTimer aggregateTimerOwner(this->config);
auto& aggregateTimer = aggregateTimerOwner;
- // Because blocks...
- std::vector<const CacheDylib*> builderCacheDylibsOwner;
- auto& builderCacheDylibs = builderCacheDylibsOwner;
- for ( const CacheDylib& cacheDylib : this->cacheDylibs )
- builderCacheDylibs.push_back(&cacheDylib);
-
Error err = parallel::forEach(this->cacheDylibs, ^(size_t index, CacheDylib& cacheDylib) {
Diagnostics diag;
cacheDylib.copyRawSegments(this->config, aggregateTimer);
+
+ // patch linked dylibs (load commands) as soon as the raw segments were coppied
+ // so next steps have accurate view of the dylib
+ if ( Error patchErr = this->patchLinkedDylibs(cacheDylib) )
+ return patchErr;
PatchInfo& dylibPatchInfo = this->patchTableOptimizer.patchInfos[cacheDylib.cacheIndex];
cacheDylib.applySplitSegInfo(diag, this->options, this->config,
@@ -299,35 +580,34 @@
if ( diag.hasError() )
return Error("%s", diag.errorMessageCStr());
- cacheDylib.calculateBindTargets(diag, this->config, aggregateTimer, builderCacheDylibs,
- dylibPatchInfo);
+ cacheDylib.bind(diag, this->config, aggregateTimer, dylibPatchInfo, this->functionVariantsOptimizer);
if ( diag.hasError() )
return Error("%s", diag.errorMessageCStr());
- cacheDylib.bind(diag, this->config, aggregateTimer, dylibPatchInfo);
- if ( diag.hasError() )
- return Error("%s", diag.errorMessageCStr());
-
- cacheDylib.updateObjCSelectorReferences(diag, this->config, aggregateTimer, this->objcSelectorOptimizer);
- if ( diag.hasError() )
- return Error("%s", diag.errorMessageCStr());
-
- cacheDylib.convertObjCMethodListsToOffsets(diag, this->config, aggregateTimer, this->objcSelectorOptimizer.selectorStringsChunk);
- if ( diag.hasError() )
- return Error("%s", diag.errorMessageCStr());
-
- // Note, must be after updating selector references and converting relative methods to selector offsets
- cacheDylib.sortObjCMethodLists(diag, this->config, aggregateTimer, this->objcSelectorOptimizer.selectorStringsChunk);
- if ( diag.hasError() )
- return Error("%s", diag.errorMessageCStr());
-
- cacheDylib.optimizeLoadsFromConstants(this->config, aggregateTimer,
- this->objcSelectorOptimizer.selectorStringsChunk);
-
- Error error = cacheDylib.emitObjCIMPCaches(this->config, aggregateTimer, this->objcIMPCachesOptimizer,
- this->objcSelectorOptimizer.selectorStringsChunk);
- if ( error.hasError() )
- return error;
+ // Disable objc optimizations from EK shared cache
+ bool emitObjcOpts = !this->options.platform.isExclaveKit();
+ if ( emitObjcOpts ) {
+ cacheDylib.updateObjCSelectorReferences(diag, this->config, aggregateTimer, this->objcSelectorOptimizer);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ cacheDylib.convertObjCMethodListsToOffsets(diag, this->config, aggregateTimer, this->objcSelectorOptimizer.selectorStringsChunk);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ // Note, must be after updating selector references and converting relative methods to selector offsets
+ cacheDylib.sortObjCMethodLists(diag, this->config, aggregateTimer, this->objcSelectorOptimizer.selectorStringsChunk);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ cacheDylib.optimizeLoadsFromConstants(this->config, aggregateTimer,
+ this->objcSelectorOptimizer.selectorStringsChunk);
+
+ Error error = cacheDylib.emitObjCIMPCaches(this->config, aggregateTimer, this->objcIMPCachesOptimizer,
+ this->objcSelectorOptimizer.selectorStringsChunk);
+ if ( error.hasError() )
+ return error;
+ }
cacheDylib.optimizeStubs(this->options, this->config, aggregateTimer, this->stubOptimizer,
dylibPatchInfo);
@@ -347,47 +627,55 @@
// Outputs: emitted optimizations in the subCache buffers
Error SharedCacheBuilder::postDylibEmitChunks()
{
+ // Disable objc optimizations from EK shared cache
+ bool emitObjcOpts = !this->options.platform.isExclaveKit();
+
+ if ( Error error = this->emitPrewarmingData(); error.hasError() )
+ return error;
this->optimizeTLVs();
if ( Error error = this->emitUniquedGOTs(); error.hasError() )
return error;
- // Note this has to be before we emit the protocol hash table
- if ( Error error = this->emitCanonicalObjCProtocols(); error.hasError() )
+ if ( emitObjcOpts ) {
+ // Note this has to be before we emit the protocol hash table
+ if ( Error error = this->emitCanonicalObjCProtocols(); error.hasError() )
+ return error;
+ }
+
+ this->emitCacheDylibsTrie();
+ if ( Error error = this->emitPatchTable() )
return error;
- this->emitCacheDylibsTrie();
- if ( Error error = this->emitPatchTable(); error.hasError() )
- return error;
+ this->emitFunctionVariants();
// Note, this must be after we emit the patch table
if ( Error error = this->emitCacheDylibsPrebuiltLoaders(); error.hasError() )
return error;
- this->emitObjCHashTables();
-
- bool preAttachedCategories = true;
- if ( preAttachedCategories ) {
+ if ( emitObjcOpts ) {
+ this->emitObjCHashTables();
+
// Note this has to be after anyone walking the objc metadata format
if ( Error error = this->emitPreAttachedObjCCategories(); error.hasError() )
return error;
- }
-
- // Note, this must be after emitCacheDylibsPrebuiltLoaders() as it needs the offset to the SectionLocations*
- // in the PrebuiltLoader*
- this->emitObjCHeaderInfo();
- if ( Error error = this->computeObjCClassLayout(); error.hasError() )
+
+ // Note, this must be after emitCacheDylibsPrebuiltLoaders() as it needs the offset to the SectionLocations*
+ // in the PrebuiltLoader*
+ this->emitObjCHeaderInfo();
+ if ( Error error = this->computeObjCClassLayout(); error.hasError() )
return error;
- // Note this must be after computeObjCClassLayout() as we need it to set the flags for whether
- // we have missing weak superclasses or not
- this->emitObjCOptsHeader();
-
- // Note, this has to be after we've emitted the objc class hash table, and after emitting
- // the objc header info
- if ( Error error = this->emitSwiftHashTables(); error.hasError() )
- return error;
+ // Note this must be after computeObjCClassLayout() as we need it to set the flags for whether
+ // we have missing weak superclasses or not
+ this->emitObjCOptsHeader();
+
+ // Note, this has to be after we've emitted the objc class hash table, and after emitting
+ // the objc header info
+ if ( Error error = this->emitSwiftHashTables(); error.hasError() )
+ return error;
+ }
// Note, this has to be after we've emitted the objc hash tables and the objc header infos
if ( Error error = this->emitExecutablePrebuiltLoaders(); error.hasError() )
@@ -409,12 +697,12 @@
// Outputs: final emitted data in the sub caches
Error SharedCacheBuilder::finalize()
{
-
// Do objc very late, as it adds segments to the mach-o, which aren't in sync with
// the segments on the CacheDylib
this->addObjcSegments();
this->computeCacheHeaders();
this->codeSign();
+ this->buildAtlas();
return Error();
}
@@ -491,61 +779,101 @@
Stats stats(this->config);
Timer::Scope timedScope(this->config, "categorizeInputs time");
- for ( InputFile& inputFile : this->allInputFiles ) {
- if ( inputFile.mf->isDylib() || inputFile.mf->isDyld() ) {
- auto failureHandler = ^(const char* format, ...) __attribute__((format(printf, 1, 2))) {
- char* output_string;
- va_list list;
- va_start(list, format);
- vasprintf(&output_string, format, list);
- va_end(list);
- inputFile.setError(Error("%s", (const char*)output_string));
- free(output_string);
- };
-
- std::string_view installName = inputFile.mf->installName();
- std::string_view dylibPath = inputFile.path;
- if ( (installName != dylibPath) && ((this->options.platform == dyld3::Platform::macOS) || startsWith(dylibPath, "/System/Cryptexes/OS/")) ) {
- // We now typically require that install names and paths match. However symlinks may allow us to bring in a path which
- // doesn't match its install name.
- // For example:
- // /usr/lib/libstdc++.6.0.9.dylib is a real file with install name /usr/lib/libstdc++.6.dylib
- // /usr/lib/libstdc++.6.dylib is a symlink to /usr/lib/libstdc++.6.0.9.dylib
- // So long as we add both paths (with one as an alias) then this will work, even if dylibs are removed from disk
- // but the symlink remains.
- // Apply the same symlink crawling for dylibs that will install their contents to Cryptex paths but will have
- // install names with the cryptex paths removed.
- char resolvedSymlinkPath[PATH_MAX];
- if ( fileSystem.getRealPath(installName.data(), resolvedSymlinkPath) ) {
- if ( resolvedSymlinkPath == dylibPath ) {
- // Symlink is the install name and points to the on-disk dylib
- //fprintf(stderr, "Symlink works: %s == %s\n", inputFile.path, installName.c_str());
- dylibPath = installName;
+ struct Chunk
+ {
+ std::vector<CacheDylib> cacheDylibs;
+ std::vector<InputFile*> exeInputFiles;
+ std::vector<InputFile*> nonCacheDylibInputFiles;
+ };
+
+ mapReduce(std::span(allInputFiles), 32, ^(size_t i, Chunk& chunk, std::span<InputFile> files){
+ for ( InputFile& inputFile : files ) {
+ if ( inputFile.mf == nullptr ) continue;
+
+ // only process valid images
+ if ( ((Header*)inputFile.mf)->isDyldManaged() ) {
+ // don't try to validate kernel or firmware mach-o binaries
+ Image image(inputFile.mf, inputFile.size, Image::MappingKind::wholeSliceMapped);
+ if ( mach_o::Error err = image.validate() ) {
+ inputFile.addError(Error("%s", err.message()));
+ continue;
+ }
+ }
+
+ if ( inputFile.mf->isDylib() || inputFile.mf->isDyld() ) {
+ auto failureHandler = ^(const char* format, ...) __attribute__((format(printf, 1, 2))) {
+ char* output_string;
+ va_list list;
+ va_start(list, format);
+ vasprintf(&output_string, format, list);
+ va_end(list);
+ inputFile.addError(Error("%s", (const char*)output_string));
+ free(output_string);
+ };
+
+ std::string_view installName = ((const Header*)inputFile.mf)->installName();
+ std::string_view dylibPath = inputFile.path;
+ if ( (installName != dylibPath) && ((this->options.platform == Platform::macOS) || startsWith(dylibPath, "/System/Cryptexes/OS/")) ) {
+ // We now typically require that install names and paths match. However symlinks may allow us to bring in a path which
+ // doesn't match its install name.
+ // For example:
+ // /usr/lib/libstdc++.6.0.9.dylib is a real file with install name /usr/lib/libstdc++.6.dylib
+ // /usr/lib/libstdc++.6.dylib is a symlink to /usr/lib/libstdc++.6.0.9.dylib
+ // So long as we add both paths (with one as an alias) then this will work, even if dylibs are removed from disk
+ // but the symlink remains.
+ // Apply the same symlink crawling for dylibs that will install their contents to Cryptex paths but will have
+ // install names with the cryptex paths removed.
+ char resolvedSymlinkPath[PATH_MAX];
+ if ( fileSystem.getRealPath(installName.data(), resolvedSymlinkPath) ) {
+ if ( resolvedSymlinkPath == dylibPath ) {
+ // Symlink is the install name and points to the on-disk dylib
+ //fprintf(stderr, "Symlink works: %s == %s\n", inputFile.path, installName.c_str());
+ dylibPath = installName;
+ }
}
}
- }
-
- if ( !inputFile.forceNotCacheEligible && inputFile.mf->canBePlacedInDyldCache(dylibPath.data(), failureHandler) ) {
- CacheDylib cacheDylib(inputFile);
- this->cacheDylibs.push_back(std::move(cacheDylib));
- }
- else {
- this->nonCacheDylibInputFiles.push_back(&inputFile);
- }
- continue;
- }
-
- if ( inputFile.mf->isDynamicExecutable() ) {
- auto failureHandler = ^(const char* reason) {
- inputFile.setError(Error("%s", reason));
- };
- if ( inputFile.mf->canHavePrebuiltExecutableLoader(options.platform, inputFile.path, failureHandler) ) {
- this->exeInputFiles.push_back(&inputFile);
- }
-
- continue;
- }
- }
+
+ if ( !inputFile.forceNotCacheEligible && inputFile.mf->canBePlacedInDyldCache(dylibPath.data(), true /* check objc */, failureHandler) ) {
+ CacheDylib cacheDylib(inputFile);
+ chunk.cacheDylibs.push_back(std::move(cacheDylib));
+ }
+ else {
+ chunk.nonCacheDylibInputFiles.push_back(&inputFile);
+ }
+ continue;
+ }
+
+ if ( inputFile.mf->isDynamicExecutable() ) {
+ auto failureHandler = ^(const char* reason) {
+ inputFile.addError(Error("%s", reason));
+ };
+ if ( inputFile.mf->canHavePrebuiltExecutableLoader(options.platform, inputFile.path, failureHandler) ) {
+ chunk.exeInputFiles.push_back(&inputFile);
+ }
+
+ continue;
+ }
+ }
+ }, ^(std::span<Chunk> chunks) {
+ size_t cacheDylibsSize = 0;
+ size_t nonCacheSize = 0;
+ size_t exeSize = 0;
+ for ( Chunk& chunk : chunks ) {
+ cacheDylibsSize += chunk.cacheDylibs.size();
+ nonCacheSize += chunk.nonCacheDylibInputFiles.size();
+ exeSize += chunk.exeInputFiles.size();
+ }
+
+ cacheDylibs.reserve(cacheDylibsSize);
+ nonCacheDylibInputFiles.reserve(cacheDylibsSize);
+ exeInputFiles.reserve(exeSize);
+
+ for ( Chunk& chunk : chunks ) {
+ std::move(chunk.cacheDylibs.begin(), chunk.cacheDylibs.end(), std::back_inserter(cacheDylibs));
+ std::move(chunk.nonCacheDylibInputFiles.begin(), chunk.nonCacheDylibInputFiles.end(), std::back_inserter(nonCacheDylibInputFiles));
+ std::move(chunk.exeInputFiles.begin(), chunk.exeInputFiles.end(), std::back_inserter(exeInputFiles));
+ }
+ });
if ( this->config.log.printStats ) {
stats.add(" inputs: found %lld cache eligible dylibs\n", (uint64_t)this->cacheDylibs.size());
@@ -558,11 +886,23 @@
{
Timer::Scope timedScope(this->config, "verifySelfContained time");
- __block std::unordered_set<std::string_view> allDylibs;
+ __block std::unordered_set<std::string_view> badDylibs;
+
+ __block std::unordered_map<std::string_view, const InputFile*> allDylibs;
allDylibs.reserve(this->allInputFiles.size());
for ( const InputFile& inputFile : this->allInputFiles ) {
+ if ( inputFile.mf == nullptr ) {
+ // Add the file if its path looks like it might be cache eligible. This is
+ // better than getting no errors at all later
+ if ( inputFile.hasError() && Header::isSharedCacheEligiblePath(inputFile.path.data()) ) {
+ allDylibs.insert({ inputFile.path, &inputFile });
+ badDylibs.insert(inputFile.path);
+ }
+ continue;
+ }
+
if ( inputFile.mf->isDylib() )
- allDylibs.insert(inputFile.mf->installName());
+ allDylibs.insert({ ((const Header*)inputFile.mf)->installName(), &inputFile });
}
__block std::unordered_set<std::string_view> potentialCacheDylibs;
@@ -570,8 +910,10 @@
for ( const CacheDylib& cacheDylib : this->cacheDylibs )
potentialCacheDylibs.insert(cacheDylib.installName);
- __block std::unordered_set<std::string_view> badDylibs;
-
+
+ // Keep track of all errors in case we need them to diagnose libsystem being missing
+ __block std::unordered_map<std::string_view, std::list<std::string>> missingDependencies;
+ __block std::unordered_map<std::string_view, std::list<std::string>> ineligibleDependencies;
// check all dependencies to assure every dylib in cache only depends on other dylibs in cache
__block bool doAgain = true;
@@ -579,6 +921,8 @@
doAgain = false;
// scan dylib list making sure all dependents are in dylib list
for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
+ if ( cacheDylib.inputFile == nullptr ) continue;
+
//Timer::Scope timedScope(this->config, cacheDylib.installName);
// Skip dylibs we marked bad from a previous iteration
if ( cacheDylib.inputFile->hasError() )
@@ -596,37 +940,103 @@
}
}
}
- if ( potentialCacheDylibs.count(loadPath) == 0 ) {
- // Break weak edges, but only if we haven't seen the dylib.
- if ( isWeak && (allDylibs.count(loadPath) == 0) )
- return;
- if ( isWeak && allowedMissingWeakDylibs.count(loadPath) )
- return;
- std::string reason = std::string("Could not find dependency '") + loadPath + "'";
- cacheDylib.inputFile->setError(Error("%s", reason.c_str()));
+
+ // Break weak edges, but only if we haven't seen the dylib.
+ if ( isWeak && (allDylibs.count(loadPath) == 0) )
+ return;
+ if ( isWeak && allowedMissingWeakDylibs.count(loadPath) )
+ return;
+
+ if ( badDylibs.count(loadPath) ) {
+ std::string reason = std::string("Depends on ineligible/bad dylib '") + loadPath + "'";
+ cacheDylib.inputFile->addError(Error("%s", reason.c_str()));
badDylibs.insert(cacheDylib.installName);
+ ineligibleDependencies[cacheDylib.installName].push_back(loadPath);
doAgain = true;
- stop = true;
return;
}
- if ( badDylibs.count(loadPath) ) {
- // Break weak edges, but only if we haven't seen the dylib.
- if ( isWeak && (allDylibs.count(loadPath) == 0) )
- return;
- std::string reason = std::string("Depends on ineligible/bad dylib '") + loadPath + "'";
- cacheDylib.inputFile->setError(Error("%s", reason.c_str()));
+ if ( potentialCacheDylibs.count(loadPath) == 0 ) {
+ std::string reason = std::string("Could not find dependency '") + loadPath + "'";
+ cacheDylib.inputFile->addError(Error("%s", reason.c_str()));
badDylibs.insert(cacheDylib.installName);
+ missingDependencies[cacheDylib.installName].push_back(loadPath);
doAgain = true;
- stop = true;
return;
}
});
}
}
+ // If libsystem is bad, give up, but try make a better error
+ for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
+ if ( cacheDylib.inputFile == nullptr )
+ continue;
+
+ if ( !cacheDylib.inputFile->hasError() )
+ continue;
+
+ bool isLibSystem = cacheDylib.installName.ends_with("/usr/lib/libSystem.dylib") || cacheDylib.installName.ends_with("/usr/lib/libSystem.B.dylib");
+ if ( !isLibSystem )
+ continue;
+
+ // Walk the dependency tree to try find the real error which caused us to fail
+ std::unordered_set<std::string> seenFiles;
+ std::list<std::string> worklist;
+ worklist.push_back(std::string(cacheDylib.installName));
+ while ( !worklist.empty() ) {
+ std::string dylibInstallName = worklist.back();
+ worklist.pop_back();
+
+ // Check if the dylib depends on something missing
+ // If so, we've found an issue
+ if ( auto it = missingDependencies.find(dylibInstallName); it != missingDependencies.end() ) {
+ for ( auto depInstallName : it->second ) {
+ std::string reason = "Dylib located at '";
+ reason += dylibInstallName;
+ reason += "' not placed in shared cache because: ";
+ reason += "Could not find dependency '";
+ reason += depInstallName;
+ reason += "'";
+ this->errors.push_back(reason);
+ }
+ }
+
+ // Add bad dependencies to the worklist, hoping to eventually walk to the root cause
+ if ( auto dylibIt = ineligibleDependencies.find(dylibInstallName); dylibIt != ineligibleDependencies.end() ) {
+ for ( auto depInstallName : dylibIt->second ) {
+ if ( seenFiles.count(depInstallName) )
+ continue;
+
+ seenFiles.insert(depInstallName);
+ worklist.push_back(depInstallName);
+
+ if ( auto depIt = allDylibs.find(depInstallName); depIt != allDylibs.end() ) {
+ const InputFile* inputFile = depIt->second;
+ // Walk the errors on this depenency, skip deps which are just more bad dylibs, and see if we can
+ // find the root cause
+ for ( const error::Error& err : inputFile->getErrors() ) {
+ if ( err.messageContains("Depends on ineligible/bad dylib") )
+ continue;
+
+ std::string reason = "Dylib located at '";
+ reason += depInstallName;
+ reason += "' not placed in shared cache because: ";
+ reason += err.message();
+ this->errors.push_back(reason);
+ }
+ }
+ }
+ }
+ }
+
+ cacheDylibs.clear();
+ }
+
// Add bad dylibs to the "other" dylibs for use in prebuilt loaders
for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
+ if ( cacheDylib.inputFile == nullptr ) continue;
+
if ( cacheDylib.inputFile->hasError() ) {
this->nonCacheDylibInputFiles.push_back(cacheDylib.inputFile);
this->dylibHasMissingDependency = true;
@@ -635,9 +1045,18 @@
this->cacheDylibs.erase(std::remove_if(this->cacheDylibs.begin(), this->cacheDylibs.end(), [&](const CacheDylib& dylib) {
// Dylibs with errors must be removed from the cache
- return dylib.inputFile->hasError();
+ return dylib.inputFile != nullptr && dylib.inputFile->hasError();
}),
this->cacheDylibs.end());
+
+ // verify that there's at least one dylib that has an input file
+ if ( !std::any_of(cacheDylibs.begin(), cacheDylibs.end(), [](const CacheDylib& dylib) {
+ return dylib.inputFile != nullptr;
+ }) ) {
+ // the only remaining dylib is the synthesized Swift prespecialized dylib
+ // so remove it too
+ cacheDylibs.clear();
+ }
}
void SharedCacheBuilder::calculateDylibAliases()
@@ -659,39 +1078,38 @@
{
Timer::Scope timedScope(this->config, "sortDylibs time");
- std::sort(this->cacheDylibs.begin(), this->cacheDylibs.end(), [&](const CacheDylib& a, const CacheDylib& b) {
+ auto sortOrder = [this](const CacheDylib& dylib) -> uint64_t {
// HACK: See addObjCOptimizationsToSubCache() and addObjCTextChunk()
// We put the libobjc __TEXT first in the sub cache so that offsets from it to OBJC_RO are
// positive. But dyld4 and objc HeaderInfo data structures rely on the cache dylibs being
// sorted by mach_header, and moving objc first breaks the order we determine here. So hack
// this too and put libobjc first for now.
- bool isObjCA = (a.installName == "/usr/lib/libobjc.A.dylib");
- bool isObjCB = (b.installName == "/usr/lib/libobjc.A.dylib");
- if ( isObjCA != isObjCB )
- return isObjCA;
-
- const auto& orderA = options.dylibOrdering.find(std::string(a.installName));
- const auto& orderB = options.dylibOrdering.find(std::string(b.installName));
- bool foundA = (orderA != options.dylibOrdering.end());
- bool foundB = (orderB != options.dylibOrdering.end());
-
- // Order all __DATA_DIRTY segments specified in the order file first, in
- // the order specified in the file, followed by any other __DATA_DIRTY
- // segments in lexicographic order.
- if ( foundA && foundB )
- return orderA->second < orderB->second;
- else if ( foundA )
- return true;
- else if ( foundB )
- return false;
+ if ( endsWith(dylib.installName, "/usr/lib/libobjc.A.dylib") )
+ return (1ULL << 32);
+ // rdar://147457741 (dyld shared cache builder should guarnatee that dyld and libdyld end up in the same split cache)
+ if ( endsWith(dylib.installName, "/usr/lib/system/libdyld.dylib") )
+ return (2ULL << 32);
+ if ( endsWith(dylib.installName, "/usr/lib/dyld") )
+ return (3ULL << 32);
+
+ // fixed order first
+ if ( const auto& fixedOrder = options.dylibOrdering.find(std::string(dylib.installName)); fixedOrder != options.dylibOrdering.end() )
+ return (4ULL << 32) + fixedOrder->second;
// Sort mac before iOSMac
- bool isIOSMacA = startsWith(a.installName, "/System/iOSSupport/");
- bool isIOSMacB = startsWith(b.installName, "/System/iOSSupport/");
- if ( isIOSMacA != isIOSMacB )
- return !isIOSMacA;
-
- // Finally sort by install name
+ if ( startsWith(dylib.installName, "/System/iOSSupport/") )
+ return (6ULL << 32);
+ return (5ULL << 32);
+ };
+
+ std::sort(this->cacheDylibs.begin(), this->cacheDylibs.end(), [&](const CacheDylib& a, const CacheDylib& b) {
+ uint64_t aOrder = sortOrder(a);
+ uint64_t bOrder = sortOrder(b);
+
+ if ( aOrder != bOrder )
+ return aOrder < bOrder;
+
+ // Finally sort by install name if there's no explicit ordering
return a.installName < b.installName;
});
@@ -699,6 +1117,46 @@
uint32_t cacheIndex = 0;
for ( CacheDylib& cacheDylib : cacheDylibs )
cacheDylib.cacheIndex = cacheIndex++;
+}
+
+void SharedCacheBuilder::sortExecutables()
+{
+ Timer::Scope timedScope(this->config, "sortExecutables time");
+
+ // make a list of the position on all binaries for anything we don't explicitly sort
+ std::unordered_map<const InputFile*, uint32_t> originalPositions;
+ for ( const InputFile* file : this->exeInputFiles )
+ originalPositions[file] = (uint32_t)originalPositions.size();
+
+ // TODO: Use a file from the perf teams if we have one
+ auto sortOrder = [&originalPositions](const InputFile* file) -> uint64_t {
+ uint32_t originalPos = originalPositions.at(file);
+
+ // Sort some paths first
+ if ( file->path.starts_with("/usr/libexec/") )
+ return (1ULL << 32) + originalPos;
+ if ( file->path.starts_with("/Applications/") )
+ return (2ULL << 32) + originalPos;
+ if ( file->path.starts_with("/private/var/staged_system_apps/") )
+ return (3ULL << 32) + originalPos;
+ if ( file->path.starts_with("/usr/bin/") )
+ return (4ULL << 32) + originalPos;
+
+ // Sort some other paths last
+ if ( file->path.starts_with("/AppleInternal/") )
+ return (62ULL << 32) + originalPos;
+ if ( file->path.starts_with("/usr/local/bin/") )
+ return (63ULL << 32) + originalPos;
+
+ // Any path not in the lists above gets sorted by original position
+ return (32ULL << 32) + originalPos;
+ };
+
+ // Just in case we overflow the buffer in emitExecutablePrebuiltLoaders(), sort the exe files to prioritize
+ // daemons and other common apps
+ std::sort(this->exeInputFiles.begin(), this->exeInputFiles.end(), [&](const InputFile* a, const InputFile* b) {
+ return sortOrder(a) < sortOrder(b);
+ });
}
Error SharedCacheBuilder::calculateDylibDependents()
@@ -752,6 +1210,30 @@
stop = true;
}
});
+
+ // copy the original list of dependents
+ cacheDylib.inputDependents = cacheDylib.dependents;
+
+ // note: below changes to dependents need to be kept in sync with load command patching in `patchLinkedDylibs`
+ // we might want to generalize that if more libraries require patching
+
+ // force swiftCore link the prespecialized dylib
+ if ( swiftPrespecializedDylib && cacheDylib.installName.find("libswiftCore.dylib") != std::string_view::npos ) {
+ CacheDylib::DependentDylib depDylib;
+ depDylib.kind = CacheDylib::DependentDylib::Kind::normal;
+ depDylib.dylib = swiftPrespecializedDylib;
+ cacheDylib.dependents.push_back(std::move(depDylib));
+ }
+
+ // clear all dependents of the prespecialized dylib except libSystem
+ // otherwise loading the library would pull in lots of other dependencies
+ if ( swiftPrespecializedDylib && &cacheDylib == swiftPrespecializedDylib ) {
+ if ( cacheDylib.dependents.empty() || cacheDylib.dependents.front().dylib->installName.find("libSystem") == std::string_view::npos ) {
+ diag.error("expected libSystem as the first linked dylib of %s", cacheDylib.inputHdr->installName());
+ } else {
+ cacheDylib.dependents.erase(cacheDylib.dependents.begin()+1, cacheDylib.dependents.end());
+ }
+ }
if ( diag.hasError() )
return Error("%s", diag.errorMessageCStr());
@@ -884,36 +1366,39 @@
}
static void forEachObjCMethodName(const CacheDylib& cacheDylib,
- void (^callback)(std::string_view str))
-{
- const MachOFile* mf = cacheDylib.inputMF;
- mf->forEachSection(^(const MachOFile::SectionInfo& sectInfo, bool malformedSectionRange, bool& stop) {
- if ( strcmp(sectInfo.segInfo.segName, "__TEXT") != 0 )
+ void (^callback)(const char* str))
+{
+ __block objc_visitor::Visitor objcVisitor = makeInputDylibObjCVisitor(cacheDylib);
+
+ auto visitPointerBasedMethod = ^(const objc_visitor::Method& method) {
+ const char* selectorString = method.getName(objcVisitor);
+ callback(selectorString);
+ };
+
+ auto visitMethodList = ^(const objc_visitor::MethodList& methodList) {
+ if ( methodList.numMethods() == 0 )
return;
- if ( strcmp(sectInfo.sectName, "__objc_methname") != 0 )
+ if ( methodList.usesRelativeOffsets() )
return;
- if ( sectInfo.segInfo.isProtected || ((sectInfo.sectFlags & SECTION_TYPE) != S_CSTRING_LITERALS) ) {
- stop = true;
- return;
- }
- if ( malformedSectionRange ) {
- stop = true;
- return;
- }
-
- // Use the file offset in the section to get the correct content
- const char* content = (const char*)mf + sectInfo.sectFileOffset;
- uint64_t sectionSize = sectInfo.sectSize;
-
- const char* s = (const char*)content;
- const char* end = s + sectionSize;
- while ( s < end ) {
- std::string_view str = s;
- callback(str);
- s += str.size() + 1;
- }
-
- stop = true;
+
+ // Check pointer based method lists
+ uint32_t numMethods = methodList.numMethods();
+ for ( uint32_t i = 0; i != numMethods; ++i ) {
+ const objc_visitor::Method& method = methodList.getMethod(objcVisitor, i);
+ visitPointerBasedMethod(method);
+ }
+ };
+
+ // selector references
+ objcVisitor.forEachSelectorReference(^(VMAddress selRefVMAddr, VMAddress selRefTargetVMAddr,
+ const char* selectorString) {
+ callback(selectorString);
+ });
+
+ // classes/categories/protocols/swift generics
+ objcVisitor.forEachMethodList(^(objc_visitor::MethodList& objcMethodList,
+ std::optional<metadata_visitor::ResolvedValue> extendedMethodTypes) {
+ visitMethodList(objcMethodList);
});
}
@@ -923,7 +1408,7 @@
VMOffset offsetInDylib;
};
-static FoundSymbol findTargetClass(Diagnostics diag,
+static FoundSymbol findTargetClass(Diagnostics& diag,
const std::vector<CacheDylib>& cacheDylibs,
std::string_view symbolName, std::optional<uint32_t> cacheIndex)
{
@@ -961,12 +1446,13 @@
if ( !this->config.layout.is64 )
return;
- if ( this->config.layout.cacheSize.rawValue() > 0x100000000 )
+ // Limited by ImpCacheEntry_v2::impOffset which is 38-bits. For now limit to 16GB
+ // as that is the maximum we know slide info v5 can get to
+ if ( this->config.layout.cacheSize.rawValue() > 16_GB )
return;
// Only arm64* are is supported by the runtime
- std::string_view archName = this->options.archs.name();
- if ( archName != "arm64e" && archName != "arm64")
+ if ( !this->options.arch.sameCpu(mach_o::Architecture::arm64) )
return;
// Skip everything if the JSON file is empty
@@ -1054,6 +1540,9 @@
});
objcVisitor.forEachCategory(^(const objc_visitor::Category& objcCategory, bool& stopCategory) {
+ if ( objcCategory.isForSwiftStubClass() )
+ return;
+
imp_caches::Category impCacheCategory(objcCategory.getName(objcVisitor));
// instance methods
@@ -1121,6 +1610,7 @@
bool isWeakImport = false;
};
+ auto& dylibMethodMap = objcIMPCachesOptimizer.methodMap[cacheDylib.installName];
__block std::vector<BindTarget> bindTargets;
__block Diagnostics diag;
cacheDylib.inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
@@ -1211,7 +1701,6 @@
objc_visitor::Method objcMethod = objcMethodList.getMethod(objcVisitor, i);
imp_caches::BucketMethod methodKey = {
- .installName = cacheDylib.installName,
.className = impCacheClass.name,
.methodName = objcMethod.getName(objcVisitor),
.isInstanceMethod = !impCacheClass.isMetaClass
@@ -1222,7 +1711,7 @@
&cacheDylib,
InputDylibVMAddress(impVMAddr.rawValue())
};
- objcIMPCachesOptimizer.methodMap[methodKey] = inputDylibLocation;
+ dylibMethodMap[methodKey] = inputDylibLocation;
}
}
@@ -1232,6 +1721,9 @@
// Walk each category and set the class pointer
__block uint32_t categoryIndex = 0;
objcVisitor.forEachCategory(^(const objc_visitor::Category& objcCategory, bool& stopCategory) {
+ if ( objcCategory.isForSwiftStubClass() )
+ return;
+
imp_caches::Category& impCacheCategory = dylib.categories[categoryIndex];
const DylibClasses& classMap = dylibClassMaps[cacheDylib.cacheIndex];
@@ -1284,7 +1776,6 @@
objc_visitor::Method objcMethod = objcMethodList.getMethod(objcVisitor, i);
imp_caches::BucketMethod methodKey = {
- .installName = cacheDylib.installName,
.className = impCacheCategory.cls->name,
.methodName = objcMethod.getName(objcVisitor),
.isInstanceMethod = true
@@ -1294,7 +1785,7 @@
&cacheDylib,
InputDylibVMAddress(impVMAddr.rawValue())
};
- objcIMPCachesOptimizer.methodMap[methodKey] = inputDylibLocation;
+ dylibMethodMap[methodKey] = inputDylibLocation;
}
}
@@ -1306,7 +1797,6 @@
objc_visitor::Method objcMethod = objcMethodList.getMethod(objcVisitor, i);
imp_caches::BucketMethod methodKey = {
- .installName = cacheDylib.installName,
.className = impCacheCategory.cls->name,
.methodName = objcMethod.getName(objcVisitor),
.isInstanceMethod = false
@@ -1316,7 +1806,7 @@
&cacheDylib,
InputDylibVMAddress(impVMAddr.rawValue())
};
- objcIMPCachesOptimizer.methodMap[methodKey] = inputDylibLocation;
+ dylibMethodMap[methodKey] = inputDylibLocation;
}
}
}
@@ -1331,6 +1821,10 @@
// We should probably keep the piece here to walk the classes as that can perhaps give us a good estimate of the size of the IMP caches
// themselves, minus the strings which need their own buffer
this->objcIMPCachesOptimizer.builder->buildImpCaches();
+
+ for ( const std::string& warning : this->objcIMPCachesOptimizer.builder->diags.warnings() ) {
+ this->warning("IMP caches: %s", warning.data());
+ }
// Push all the IMP cache selectors in to the main selectors buffer.
// We could try have an IMP cache selectors buffer and a regular selectors buffer, but that complicates
@@ -1361,6 +1855,7 @@
// Add space for the IMP caches themselves
this->objcIMPCachesOptimizer.dylibIMPCaches.resize(dylibs.size());
+ size_t discardedCaches = 0;
for ( uint32_t dylibIndex = 0; dylibIndex != dylibs.size(); ++dylibIndex ) {
imp_caches::Dylib& dylib = dylibs[dylibIndex];
ObjCIMPCachesOptimizer::IMPCacheMap& dylibIMPCaches = objcIMPCachesOptimizer.dylibIMPCaches[dylibIndex];
@@ -1369,6 +1864,24 @@
if ( !impCache.has_value() )
continue;
+ // rdar://147219191 (Selector bucket overflows in IMP caches result in malformed tables)
+ // validate selector offsets point to matching buckets
+ bool discard = false;
+ for ( uint32_t slot = 0; slot < impCache->buckets.size(); ++slot ) {
+ const imp_caches::Bucket& bucket = impCache->buckets[slot];
+ if ( !bucket.isEmptyBucket ) {
+ uint32_t expectedSlot = (bucket.selOffset >> impCache->cache_shift) & impCache->cache_mask;
+ if ( expectedSlot != slot ) {
+ discard = true;
+ break;
+ }
+ }
+ }
+ if ( discard ) {
+ ++discardedCaches;
+ continue;
+ }
+
VMOffset currentOffset((uint64_t)this->objcIMPCachesOptimizer.impCachesTotalByteSize);
assert((this->objcIMPCachesOptimizer.impCachesTotalByteSize % 8) == 0);
this->objcIMPCachesOptimizer.impCachesTotalByteSize += sizeof(ImpCacheHeader_v2);
@@ -1380,6 +1893,10 @@
}
}
+ if ( discardedCaches != 0 ) {
+ this->warning("[ERROR] discarded IMP caches: %lu", discardedCaches);
+ }
+
if ( this->config.log.printStats ) {
stats.add(" objc: found %lld imp cache selectors\n", (uint64_t)this->objcSelectorOptimizer.selectorsMap.size());
stats.add(" objc: using %lld bytes\n", this->objcSelectorOptimizer.selectorStringsTotalByteSize);
@@ -1390,7 +1907,7 @@
void SharedCacheBuilder::findObjCDylibs()
{
// driverKit has no objc
- if ( this->options.platform == dyld3::Platform::driverKit )
+ if ( this->options.platform == Platform::driverKit )
return;
Stats stats(this->config);
@@ -1398,7 +1915,9 @@
assert(this->objcOptimizer.objcDylibs.empty());
for ( CacheDylib& cacheDylib : this->cacheDylibs ) {
- if ( cacheDylib.inputMF->hasObjC() )
+ if ( cacheDylib.inputHdr == nullptr ) continue;
+
+ if ( cacheDylib.inputHdr->hasObjC() )
this->objcOptimizer.objcDylibs.push_back(&cacheDylib);
}
@@ -1439,11 +1958,19 @@
Error err = parallel::forEach(this->cacheDylibs, ^(size_t index, CacheDylib& cacheDylib) {
std::vector<std::string_view>& strings = dylibSelectors[index];
- forEachObjCMethodName(cacheDylib, ^(std::string_view str) {
- strings.push_back(std::move(str));
+ __block std::unordered_set<const void*> seenStrings;
+ forEachObjCMethodName(cacheDylib, ^(const char* str) {
+ auto itAndInserted = seenStrings.insert(str);
+ if ( itAndInserted.second )
+ strings.push_back(str);
});
- // FIXME: Walk selector references, classes, categories, protocols, etc
+ // sort strings as we used to walk the __objc_methname section in order, so by sorting we
+ // should get close to possible to the old order
+ std::sort(strings.begin(), strings.end(), [](const std::string_view& a, const std::string_view& b) {
+ // note pointer comparison as we want the string location in the binary not its contents
+ return a.data() < b.data();
+ });
return Error();
});
@@ -1815,7 +2342,7 @@
size_t objcIndex = 0;
for (size_t cacheIndex = 0; cacheIndex < this->cacheDylibs.size(); cacheIndex++) {
CacheDylib& cacheDylib = this->cacheDylibs[cacheIndex];
- if ( !cacheDylib.inputMF->hasObjC() )
+ if ( !cacheDylib.inputHdr->hasObjC() )
continue;
// Skip dylibs with opcode fixups, as the Category visitor operates on chained fixups to find classes
@@ -1868,6 +2395,10 @@
});
objCVisitor.forEachCategory(^(const objc_visitor::Category &objcCategory, bool &stopCategory) {
+
+ // Skip catlist2 entries. These are only for Swift stub classes
+ if ( objcCategory.isForSwiftStubClass() )
+ return;
__block ObjCCategoryOptimizer::Category objCCategoryInfo(objcCategory.getName(objCVisitor));
objCCategoryInfo.dylibObjcIndex = objcIndex;
@@ -1961,7 +2492,10 @@
uint32_t elementsWithPadding = maxElements * 11 / 10; // if close to power of 2, perfect hash may fail, so don't get within 10% of that
uint32_t powTwoCapacity = 1 << (32 - __builtin_clz(elementsWithPadding - 1));
uint32_t headerSize = 4 * (8 + 256);
- return headerSize + powTwoCapacity / 2 + powTwoCapacity + powTwoCapacity * perElementData;
+ uint32_t totalSize = headerSize + powTwoCapacity / 2 + powTwoCapacity + powTwoCapacity * perElementData;
+
+ // Add in a little extra room for very small tables when we are building unit tests
+ return std::max(totalSize, 4096U);
}
void SharedCacheBuilder::estimateObjCHashTableSizes()
@@ -2098,6 +2632,11 @@
return hashTableSize + (3 * sizeof(uint64_t) * maxElements);
}
+static uint32_t ptrHashTableSize(uint32_t maxElement, uint32_t numPointerKeys)
+{
+ return swiftHashTableSize(maxElement) + numPointerKeys * sizeof(uint64_t);
+}
+
void SharedCacheBuilder::estimateSwiftHashTableSizes()
{
if ( this->objcOptimizer.objcDylibs.empty() )
@@ -2106,7 +2645,7 @@
Stats stats(this->config);
Timer::Scope timedScope(this->config, "estimateSwiftHashTableSizes time");
- this->swiftProtocolConformanceOptimizer.optsHeaderByteSize = sizeof(SwiftOptimizationHeader);
+ this->swiftOptimizer.optsHeaderByteSize = sizeof(SwiftOptimizationHeader);
__block uint32_t numTypeConformances = 0;
__block uint32_t numMetadataConformances = 0;
@@ -2152,7 +2691,29 @@
});
}
- auto& optimizer = this->swiftProtocolConformanceOptimizer;
+ auto& optimizer = this->swiftOptimizer;
+
+ if ( swiftPrespecializedDylib ) {
+ Diagnostics diagVal;
+ Diagnostics& diag = diagVal;
+ SwiftVisitor swiftVisitorVal = makeInputDylibSwiftVisitor(*swiftPrespecializedDylib);
+ SwiftVisitor& swiftVisitor = swiftVisitorVal;
+ swiftVisitor.forEachPointerHashTable(diag, ^(metadata_visitor::ResolvedValue sectionBase, size_t tableIndex, uint8_t *tableStart, size_t numEntries) {
+ assert(optimizer.prespecializedMetadataHashTables.size() == tableIndex);
+
+ PointerHashTableOptimizerInfo& tableInfo = optimizer.prespecializedMetadataHashTables.emplace_back();
+ swiftVisitor.forEachPointerHashTableRelativeEntry(diag, tableStart, VMAddress(0ull), ^(size_t index, std::span<uint64_t> keys, uint64_t value) {
+ assert(!keys.empty() && "pointer keys can't be empty");
+
+ ++tableInfo.numEntries;
+ tableInfo.numPointerKeys += (uint32_t)keys.size();
+ });
+
+ tableInfo.size = ptrHashTableSize(tableInfo.numEntries, tableInfo.numPointerKeys);
+ assert(tableInfo.numEntries == numEntries);
+ });
+ }
+
optimizer.typeConformancesHashTableSize = swiftHashTableSize(numTypeConformances);
optimizer.metadataConformancesHashTableSize = swiftHashTableSize(numMetadataConformances);
optimizer.foreignTypeConformancesHashTableSize = swiftHashTableSize(numForeignConformances);
@@ -2162,6 +2723,12 @@
(uint64_t)optimizer.typeConformancesHashTableSize, numTypeConformances);
stats.add(" swift: metadata hash table estimated size: %lld (from %d entries)\n", (uint64_t)optimizer.metadataConformancesHashTableSize, numMetadataConformances);
stats.add(" swift: foreign metadata hash table estimated size: %lld (from %d entries)\n", (uint64_t)optimizer.foreignTypeConformancesHashTableSize, numForeignConformances);
+
+ stats.add(" swift: prespecialized metadata hash tables %lu\n", optimizer.prespecializedMetadataHashTables.size());
+ for ( int i = 0; i < optimizer.prespecializedMetadataHashTables.size(); ++i ) {
+ const PointerHashTableOptimizerInfo& tableInfo = optimizer.prespecializedMetadataHashTables[i];
+ stats.add(" swift: prespecialized metadata hash table #%d. estimated size: %lld (from %u entries)\n", i, (uint64_t)tableInfo.size, tableInfo.numEntries);
+ }
}
}
@@ -2197,7 +2764,7 @@
// For each alias, also see if we have intermediate aliases
// This is the "Current -> A" symlink in say "/S/L/F/CF.fw/Current/CF"
- if ( this->options.platform == dyld3::Platform::macOS ) {
+ if ( this->options.platform == Platform::macOS ) {
for ( const cache_builder::FileAlias& alias : this->inputIntermediateAliases ) {
const auto& pos = dylibPathToDylibIndex.find(alias.realPath);
if ( pos != dylibPathToDylibIndex.end() ) {
@@ -2217,89 +2784,78 @@
}
}
-void SharedCacheBuilder::estimatePatchTableSize()
-{
+void SharedCacheBuilder::initializePatchTableOptimizer()
+{
+ // Reserve space for the patch infos, one per dylib
+ // TODO: Just move these to the dylib. We don't really need them to be owned by the PatchTableOptimizer
+ this->patchTableOptimizer.patchInfos.resize(this->cacheDylibs.size());
+}
+
+void SharedCacheBuilder::estimateFunctionVariantsSize()
+{
+ this->functionVariantsOptimizer.fvInfoTotalByteSize = sizeof(dyld_cache_function_variant_entry) * 256; // FIXME
+}
+
+void SharedCacheBuilder::estimatePrewarmingSize()
+{
+ // Skip everything if the JSON file is empty
+ if ( this->options.prewarmingOptimizations.empty() )
+ return;
+
+ using json::Node;
+ using json::NodeValueType;
+
Stats stats(this->config);
- Timer::Scope timedScope(this->config, "estimatePatchTableSize time");
-
- // The patch table consists of a series of arrays.
- // For each dylib, we have a list of all clients of that dylib
- // For each dylib we also have a list of used exports
- // For each client we then have a list of symbols used
- // And for each list of symbols, we have a list of locations to patch
- // We need to estimate a patch table based on the above lists
-
- __block uint32_t bindStringsLength = 0;
- __block uint32_t numBindTargets = 0;
- __block uint32_t numBinds = 0;
- uint32_t numClients = 0;
- for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
- __block Diagnostics diag;
- cacheDylib.inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
- mach_o::Fixups fixups(layout);
- fixups.forEachBindTarget(diag, true, 0,
- ^(const mach_o::Fixups::BindTargetInfo& info, bool &stop) {
- ++numBindTargets;
- bindStringsLength += strlen(info.symbolName) + 1;
- }, ^(const mach_o::Fixups::BindTargetInfo &info, bool &stop) {
- ++numBindTargets;
- bindStringsLength += strlen(info.symbolName) + 1;
- });
-
- if ( cacheDylib.inputMF->hasChainedFixups() ) {
- fixups.withChainStarts(diag, ^(const dyld_chained_starts_in_image* starts) {
- fixups.forEachFixupInAllChains(diag, starts, false,
- ^(mach_o::ChainedFixupPointerOnDisk* fixupLoc,
- uint64_t fixupSegmentOffset,
- const dyld_chained_starts_in_segment* segInfo,
- bool& stop) {
- uint32_t bindOrdinal = ~0U;
- int64_t addend = -1;
- if ( fixupLoc->isBind(segInfo->pointer_format, bindOrdinal, addend) )
- ++numBinds;
- });
- });
- } else {
- fixups.forEachBindLocation_Opcodes(diag,
- ^(uint64_t runtimeOffset, uint32_t segmentIndex,
- unsigned int targetIndex, bool &stop) {
- ++numBinds;
- }, ^(uint64_t runtimeOffset, uint32_t segmentIndex, unsigned int overrideBindTargetIndex,
- bool &stop) {
- ++numBinds;
- });
- }
- });
-
- numClients += (uint32_t)cacheDylib.dependents.size();
- }
-
- // Start with the patch header
- uint64_t size = sizeof(dyld_cache_patch_info_v3);
-
- // One of these for each dylib
- size += sizeof(dyld_cache_image_patches_v2) * this->cacheDylibs.size();
-
- // Estimate that 2/3 of exports are used
- size += (sizeof(dyld_cache_image_export_v2) * numBindTargets * 2) / 3;
- size += (bindStringsLength * 2) / 3;
-
- // 1 entry per client
- size += sizeof(dyld_cache_image_clients_v2) * numClients;
-
- // 1 entry per bind target
- size += sizeof(dyld_cache_patchable_export_v2) * numBindTargets;
-
- // 1 entry per location we bind to
- size += sizeof(dyld_cache_patchable_location_v4) * numBinds;
-
- this->patchTableOptimizer.patchTableTotalByteSize = size;
-
- // Reserve space for the patch infos, one per dylib
- this->patchTableOptimizer.patchInfos.resize(this->cacheDylibs.size());
+ Timer::Scope timedScope(this->config, "estimatePrewarmingSize time");
+
+ Diagnostics diag;
+ Node rootNode = json::readJSON(diag, this->options.prewarmingOptimizations.data(),
+ this->options.prewarmingOptimizations.size(),
+ false /* useJSON5 */);
+ if ( diag.hasError() )
+ return;
+
+ // Format is something like:
+ // [
+ // {
+ // "install_name": "..."
+ // "locations": [
+ // {
+ // "name": "symbol name",
+ // "offset": 0
+ // }
+ // ]
+ // },
+ // ...
+ // ]
+ if ( rootNode.type != NodeValueType::Array )
+ return;
+
+ uint64_t numElements = 0;
+ for ( const Node& fileNode : rootNode.array ) {
+ auto locationsIt = fileNode.map.find("locations");
+ if ( locationsIt == fileNode.map.end() ) {
+ // FIXME: Should we error out if the JSON isn't what we expect
+ // For now just skip bad data
+ continue;
+ }
+
+ if ( locationsIt->second.type != NodeValueType::Array ) {
+ // FIXME: Should we error out if the JSON isn't what we expect
+ // For now just skip bad data
+ continue;
+ }
+
+ numElements += locationsIt->second.array.size();
+ }
+
+ this->prewarmingOptimizer.prewarmingByteSize = 0;
+ this->prewarmingOptimizer.prewarmingByteSize += sizeof(dyld_prewarming_header);
+ this->prewarmingOptimizer.prewarmingByteSize += numElements * sizeof(dyld_prewarming_entry);
+ this->prewarmingOptimizer.prewarmingByteSize = (uint32_t)alignTo((uint64_t)this->prewarmingOptimizer.prewarmingByteSize, 8);
if ( this->config.log.printStats ) {
- stats.add(" patch table estimated size: %lld\n", (uint64_t)this->patchTableOptimizer.patchTableTotalByteSize);
+ stats.add(" dyld4 prewarming estimated size: %lld\n", (uint64_t)this->prewarmingOptimizer.prewarmingByteSize);
}
}
@@ -2318,12 +2874,12 @@
size += cacheDylib.inputFile->path.size() + 1;
size = alignTo(size, alignof(dyld4::Loader::LoaderRef));
size += sizeof(dyld4::Loader::LoaderRef) * cacheDylib.dependents.size();
- size += sizeof(Loader::DependentDylibAttributes) * cacheDylib.dependents.size();
+ size += sizeof(Loader::LinkedDylibAttributes) * cacheDylib.dependents.size();
size += sizeof(Loader::FileValidationInfo);
size += sizeof(Loader::Region) * cacheDylib.segments.size();
// iOSMac dylibs likely contain a patch table
- if ( (this->options.platform == dyld3::Platform::macOS)
+ if ( (this->options.platform == Platform::macOS)
&& startsWith(cacheDylib.installName, "/System/iOSSupport") ) {
__block Diagnostics diag;
cacheDylib.inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
@@ -2345,7 +2901,7 @@
// patch tables, etc. For now, 20KB/executable seems about right
// FIXME: Small programs like daemons need just 500 bytes or so, but large /Applications need up to 1MB
// Come up with a better estimate
- this->prebuiltLoaderBuilder.executablesLoaderSize = 20_KB * this->exeInputFiles.size();
+ this->prebuiltLoaderBuilder.executablesLoaderSize = 30_KB * this->exeInputFiles.size();
// Estimate the trie size
// Assume they are all at a high offset
@@ -2358,7 +2914,7 @@
// Assumes that cdHash strings at runtime use lowercase a-f digits
uint32_t codeSignFileOffset = 0;
uint32_t codeSignFileSize = 0;
- if ( inputFile->mf->hasCodeSignature(codeSignFileOffset, codeSignFileSize) ) {
+ if ( ((const Header*)inputFile->mf)->hasCodeSignature(codeSignFileOffset, codeSignFileSize) ) {
auto handler = ^(const uint8_t cdHash[20]) {
std::string cdHashStr = "/cdhash/";
cdHashStr.reserve(24);
@@ -2440,6 +2996,10 @@
// This adds all of them to the given subCache
void SharedCacheBuilder::addObjCOptimizationsToSubCache(SubCache& subCache)
{
+ // Disable objc optimizations for EK
+ if ( this->options.platform.isExclaveKit() )
+ return;
+
// Add objc header info RW
subCache.addObjCHeaderInfoReadWriteChunk(this->config, this->objcOptimizer);
@@ -2450,40 +3010,41 @@
subCache.addObjCCategoriesChunk(this->config, this->objcCategoryOptimizer);
// Add objc opts header
- subCache.addObjCOptsHeaderChunk(this->objcOptimizer);
+ subCache.addObjCOptsHeaderChunk(this->config, this->objcOptimizer);
// Add objc header info RO
- subCache.addObjCHeaderInfoReadOnlyChunk(this->objcOptimizer);
+ subCache.addObjCHeaderInfoReadOnlyChunk(this->config, this->objcOptimizer);
// Add objc image info
- subCache.addObjCImageInfoChunk(this->objcOptimizer);
+ subCache.addObjCImageInfoChunk(this->config, this->objcOptimizer);
// Add selector strings and hash table. These need to be adjacent as the table has offsets in
// to the string section
- subCache.addObjCSelectorStringsChunk(this->objcSelectorOptimizer);
- subCache.addObjCSelectorHashTableChunk(this->objcSelectorOptimizer);
+ subCache.addObjCSelectorStringsChunk(this->config, this->objcSelectorOptimizer);
+ subCache.addObjCSelectorHashTableChunk(this->config, this->objcSelectorOptimizer);
// Add class name strings and hash table
- subCache.addObjCClassNameStringsChunk(this->objcClassOptimizer);
- subCache.addObjCClassHashTableChunk(this->objcClassOptimizer);
+ subCache.addObjCClassNameStringsChunk(this->config, this->objcClassOptimizer);
+ subCache.addObjCClassHashTableChunk(this->config, this->objcClassOptimizer);
// Add protocol name strings and hash table
- subCache.addObjCProtocolNameStringsChunk(this->objcProtocolOptimizer);
- subCache.addObjCProtocolHashTableChunk(this->objcProtocolOptimizer);
+ subCache.addObjCProtocolNameStringsChunk(this->config, this->objcProtocolOptimizer);
+ subCache.addObjCProtocolHashTableChunk(this->config, this->objcProtocolOptimizer);
// Add Swift demangled name strings found in ObjC protocol metadata
- subCache.addObjCProtocolSwiftDemangledNamesChunk(this->objcProtocolOptimizer);
+ subCache.addObjCProtocolSwiftDemangledNamesChunk(this->config, this->objcProtocolOptimizer);
// Add ObjC IMP Caches
- subCache.addObjCIMPCachesChunk(this->objcIMPCachesOptimizer);
+ subCache.addObjCIMPCachesChunk(this->config, this->objcIMPCachesOptimizer);
// Add Swift opts header
- subCache.addSwiftOptsHeaderChunk(this->swiftProtocolConformanceOptimizer);
+ subCache.addSwiftOptsHeaderChunk(this->config, this->swiftOptimizer);
// Add Swift hash tables
- subCache.addSwiftTypeHashTableChunk(this->swiftProtocolConformanceOptimizer);
- subCache.addSwiftMetadataHashTableChunk(this->swiftProtocolConformanceOptimizer);
- subCache.addSwiftForeignHashTableChunk(this->swiftProtocolConformanceOptimizer);
+ subCache.addSwiftTypeHashTableChunk(this->config, this->swiftOptimizer);
+ subCache.addSwiftMetadataHashTableChunk(this->config, this->swiftOptimizer);
+ subCache.addSwiftForeignHashTableChunk(this->config, this->swiftOptimizer);
+ subCache.addSwiftPrespecializedMetadataPointerTableChunks(this->config, this->swiftOptimizer);
}
// The shared cache contains many global optimizations such as dyld4 loaders, trie's, etc.
@@ -2497,6 +3058,9 @@
// Add patch table
subCache.addPatchTableChunk(this->patchTableOptimizer);
+ // Add function-variants table
+ subCache.addFunctionVariantsChunk(this->functionVariantsOptimizer);
+
// Add cache dylib Loader's
subCache.addCacheDylibsLoaderChunk(this->prebuiltLoaderBuilder);
@@ -2505,24 +3069,57 @@
// Add executable trie
subCache.addExecutablesTrieChunk(this->prebuiltLoaderBuilder);
+
+ // Add prewarming data
+ subCache.addPrewarmingDataChunk(this->config, this->prewarmingOptimizer);
}
// Every subCache needs a code signature, and subCache's with DATA* need slide info. This adds
// anything we need, based on whatever else is already in the SubCache.
void SharedCacheBuilder::addFinalChunksToSubCache(SubCache& subCache)
{
- subCache.addCacheHeaderChunk(this->cacheDylibs);
-
- // Add slide info for each DATA/AUTH segment. Do this after we've added any other DATA*
- // segments
- if ( this->config.slideInfo.slideInfoFormat.has_value() )
- subCache.addSlideInfoChunks();
+ subCache.addCacheHeaderChunk(this->options, this->config, this->cacheDylibs);
// Add a code signature region
subCache.addCodeSignatureChunk();
-
- // Finalize the SubCache, by removing any unused regions
- subCache.removeEmptyRegions();
+}
+
+static std::optional<Header::SectionInfo> getStubsSection(const DylibSegmentChunk* dylibChunk)
+{
+ if ( dylibChunk == nullptr )
+ return std::nullopt;
+ if ( dylibChunk->kind != Chunk::Kind::dylibText )
+ return std::nullopt;
+
+ const MachOFile* mf = dylibChunk->inputFile->mf;
+ __block std::optional<Header::SectionInfo> stubsSection;
+ ((const Header*)mf)->forEachSection(^(const Header::SectionInfo §Info, bool &stop) {
+ if ( dylibChunk->segmentName != sectInfo.segmentName )
+ return;
+
+ unsigned sectionType = (sectInfo.flags & SECTION_TYPE);
+ if ( sectionType != S_SYMBOL_STUBS )
+ return;
+
+ if ( sectInfo.segmentName != "__TEXT" ) {
+ // stubs aren't in __TEXT. Give up on this one for now
+ return;
+ }
+
+ if ( mf->isArch("arm64e") ) {
+ // For arm64e, we can only optimize __auth_stubs
+ if ( sectInfo.sectionName != "__auth_stubs" )
+ return;
+ } else {
+ // For non-arm64e, we can only optimize __stubs
+ if ( sectInfo.sectionName != "__stubs" )
+ return;
+ }
+
+ stubsSection = sectInfo;
+ });
+
+ return stubsSection;
}
// Add stubs Chunk's for every stubs section in the given text subCache
@@ -2536,60 +3133,29 @@
const DylibSegmentChunk* textChunk = textRegionChunk->isDylibSegmentChunk();
if ( textChunk == nullptr )
continue;
- if ( textChunk->kind != Chunk::Kind::dylibText )
+
+ std::optional<Header::SectionInfo> sectInfo = getStubsSection(textChunk);
+ if ( !sectInfo )
continue;
- const MachOFile* mf = textChunk->inputFile->mf;
- mf->forEachSection(^(const dyld3::MachOFile::SectionInfo §Info,
- bool malformedSectionRange, bool &stop) {
- if ( textChunk->segmentName != sectInfo.segInfo.segName )
- return;
-
- unsigned sectionType = (sectInfo.sectFlags & SECTION_TYPE);
- if ( sectionType != S_SYMBOL_STUBS )
- return;
-
- if ( strcmp(sectInfo.segInfo.segName, "__TEXT") != 0 ) {
- // stubs aren't in __TEXT. Give up on this one for now
- return;
- }
-
- // Make a stubs chunk for this stubs section
- CacheDylib* cacheDylib = fileToDylibMap.at(textChunk->inputFile);
- StubsChunk* devStubsChunk = nullptr;
- StubsChunk* customerStubsChunk = nullptr;
-
- if ( mf->isArch("arm64e") ) {
- // For arm64e, we can only optimize __auth_stubs
- if ( !strcmp(sectInfo.sectName, "__auth_stubs") ) {
- devStubsChunk = &cacheDylib->developmentStubs;
- customerStubsChunk = &cacheDylib->customerStubs;
- }
- } else {
- // For non-arm64e, we can only optimize __stubs
- if ( !strcmp(sectInfo.sectName, "__stubs") ) {
- devStubsChunk = &cacheDylib->developmentStubs;
- customerStubsChunk = &cacheDylib->customerStubs;
- }
- }
-
- if ( devStubsChunk == nullptr )
- return;
-
- assert(devStubsChunk->segmentName.empty());
- devStubsChunk->segmentName = sectInfo.segInfo.segName;
- devStubsChunk->sectionName = sectInfo.sectName;
- devStubsChunk->subCacheFileSize = CacheFileSize(sectInfo.sectSize);
- devStubsChunk->cacheVMSize = CacheVMSize(sectInfo.sectSize);
- devStubsSubCache.addStubsChunk(devStubsChunk);
-
- assert(customerStubsChunk->segmentName.empty());
- customerStubsChunk->segmentName = sectInfo.segInfo.segName;
- customerStubsChunk->sectionName = sectInfo.sectName;
- customerStubsChunk->subCacheFileSize = CacheFileSize(sectInfo.sectSize);
- customerStubsChunk->cacheVMSize = CacheVMSize(sectInfo.sectSize);
- customerStubsSubCache.addStubsChunk(customerStubsChunk);
- });
+ // Make a stubs chunk for this stubs section
+ CacheDylib* cacheDylib = fileToDylibMap.at(textChunk->inputFile);
+ StubsChunk* devStubsChunk = &cacheDylib->developmentStubs;
+ StubsChunk* customerStubsChunk = &cacheDylib->customerStubs;
+
+ assert(devStubsChunk->segmentName.empty());
+ devStubsChunk->segmentName = sectInfo->segmentName;
+ devStubsChunk->sectionName = sectInfo->sectionName;
+ devStubsChunk->subCacheFileSize = CacheFileSize(sectInfo->size);
+ devStubsChunk->cacheVMSize = CacheVMSize(sectInfo->size);
+ devStubsSubCache.addStubsChunk(devStubsChunk);
+
+ assert(customerStubsChunk->segmentName.empty());
+ customerStubsChunk->segmentName = sectInfo->segmentName;
+ customerStubsChunk->sectionName = sectInfo->sectionName;
+ customerStubsChunk->subCacheFileSize = CacheFileSize(sectInfo->size);
+ customerStubsChunk->cacheVMSize = CacheVMSize(sectInfo->size);
+ customerStubsSubCache.addStubsChunk(customerStubsChunk);
}
}
@@ -2696,9 +3262,49 @@
// Nothing to do here
break;
case cache_builder::Region::Kind::dataConst:
+ case cache_builder::Region::Kind::tproConst:
case cache_builder::Region::Kind::data:
case cache_builder::Region::Kind::auth:
- case cache_builder::Region::Kind::authConst: {
+ case cache_builder::Region::Kind::authConst:
+ case cache_builder::Region::Kind::tproAuthConst:{
+ Region& newRegion = newSubCache.regions[(uint32_t)oldRegion.kind];
+ newRegion.chunks = std::move(oldRegion.chunks);
+ break;
+ }
+ case cache_builder::Region::Kind::readOnly:
+ case cache_builder::Region::Kind::linkedit:
+ case cache_builder::Region::Kind::unmapped:
+ case cache_builder::Region::Kind::dynamicConfig:
+ case cache_builder::Region::Kind::codeSignature:
+ case cache_builder::Region::Kind::numKinds:
+ break;
+ }
+ }
+ }
+
+ // Also split the current file so that read-only are in their own files
+ {
+ // Create a new subCache
+ newSubCaches.push_back(SubCache::makeSubCache(options));
+ SubCache& newSubCache = newSubCaches.back();
+
+ // Move all data to the new subCache
+ for ( Region& oldRegion : subCache.regions ) {
+ if ( oldRegion.chunks.empty() )
+ continue;
+
+ // Move all the data regions, leave the rest
+ switch ( oldRegion.kind ) {
+ case cache_builder::Region::Kind::text:
+ case cache_builder::Region::Kind::dataConst:
+ case cache_builder::Region::Kind::tproConst:
+ case cache_builder::Region::Kind::data:
+ case cache_builder::Region::Kind::auth:
+ case cache_builder::Region::Kind::authConst:
+ case cache_builder::Region::Kind::tproAuthConst:
+ // Nothing to do here
+ break;
+ case cache_builder::Region::Kind::readOnly:{
Region& newRegion = newSubCache.regions[(uint32_t)oldRegion.kind];
newRegion.chunks = std::move(oldRegion.chunks);
break;
@@ -2725,56 +3331,104 @@
{
SubCache* currentSubCache = firstSubCache;
- // We'll add LINKEDIT at the end. As the shared region is <= 4GB in size, we can fit
- // all the LINKEDIT in the last subCache and still keep it in range of 32-bit offsets
- bool allLinkeditInLastSubCache = this->config.layout.allLinkeditInLastSubCache;
+ std::string_view libObjcInstallName = "/usr/lib/libobjc.A.dylib";
+ if ( this->options.platform.isExclaveKit() )
+ libObjcInstallName = "/System/ExclaveKit/usr/lib/libobjc.A.dylib";
+
+ // If we have an objc dylib, then gets its optimizations now to work out their size first
+ // The subCache with libobjc gets the header info sections, which for now we assume is the first subcache
+ // Add all the objc tables. This must be done before we add libobjc's __TEXT
+ for ( CacheDylib& cacheDylib : this->cacheDylibs ) {
+ if ( cacheDylib.installName == libObjcInstallName ) {
+ this->addObjCOptimizationsToSubCache(*currentSubCache);
+ break;
+ }
+ }
// Walk all the dylibs, and create a new subCache every time we are about to cross
- // the subCacheTextLimit
- CacheVMSize subCacheTextSize(0ULL);
- for ( CacheDylib& cacheDylib : this->cacheDylibs ) {
- // Peek ahead to find the __TEXT size
- CacheVMSize textSize(0ULL);
- for ( DylibSegmentChunk& segmentInfo : cacheDylib.segments ) {
- if ( segmentInfo.kind == DylibSegmentChunk::Kind::dylibText )
- textSize += segmentInfo.cacheVMSize;
- }
-
- // If we exceed the current limit, then the current subCache is complete and we need
- // to start a new one
- if ( (subCacheTextSize + textSize) > this->config.layout.subCacheTextLimit ) {
- // Create a new subCache
- otherCaches.push_back(SubCache::makeSubCache(this->options));
- currentSubCache = &otherCaches.back();
-
- // Reset the limit for the next subCache
- subCacheTextSize = CacheVMSize(0ULL);
- }
-
- subCacheTextSize += textSize;
-
- // The subCache with libobjc gets the header info sections
- // Add all the objc tables. This must be done before we add libobjc's __TEXT
- std::string_view libObjcInstallName = "/usr/lib/libobjc.A.dylib";
- if ( dyld3::MachOFile::isExclaveKitPlatform(this->options.platform) )
- libObjcInstallName = "/System/ExclaveKit/usr/lib/libobjc.A.dylib";
- if ( cacheDylib.installName == libObjcInstallName )
- this->addObjCOptimizationsToSubCache(*currentSubCache);
-
- // We'll add LINKEDIT at the end. As the shared region is <= 4GB in size, we can fit
- // all the LINKEDIT in the last subCache and still keep it in range of 32-bit offsets
- bool addLinkedit = !allLinkeditInLastSubCache;
- currentSubCache->addDylib(cacheDylib, addLinkedit);
+ // the subCache size limit
+ if ( this->config.layout.discontiguous.has_value() ) {
+ const CacheVMSize subCacheTextLimit = this->config.layout.discontiguous.value().subCacheTextLimit;
+ CacheVMSize subCacheTextSize(0ULL);
+ for ( CacheDylib& cacheDylib : this->cacheDylibs ) {
+ // Peek ahead to find the __TEXT size
+ CacheVMSize textSize(0ULL);
+ for ( DylibSegmentChunk& segmentInfo : cacheDylib.segments ) {
+ if ( segmentInfo.kind == DylibSegmentChunk::Kind::dylibText )
+ textSize += segmentInfo.cacheVMSize;
+ }
+
+ // If we exceed the current limit, then the current subCache is complete and we need
+ // to start a new one
+ if ( (subCacheTextSize + textSize) > subCacheTextLimit ) {
+ // Create a new subCache
+ otherCaches.push_back(SubCache::makeSubCache(this->options));
+ currentSubCache = &otherCaches.back();
+
+ // Reset the limit for the next subCache
+ subCacheTextSize = CacheVMSize(0ULL);
+ }
+
+ subCacheTextSize += textSize;
+
+ currentSubCache->addDylib(this->config, cacheDylib);
+ currentSubCache->addLinkeditFromDylib(cacheDylib);
+ }
+ } else {
+ const CacheVMSize subCacheTextDataLimit = this->config.layout.contiguous.value().subCacheTextDataLimit;
+ const CacheVMSize regionPadding = this->config.layout.contiguous.value().subCachePadding;
+ CacheVMSize subCacheTextDataSize(regionPadding);
+
+ // If we added any objc optimizations, then take their size in to account now
+ for ( const Region& region : currentSubCache->regions ) {
+ if ( region.kind == Region::Kind::linkedit )
+ continue;
+ if ( region.chunks.empty() )
+ continue;
+ for ( const Chunk* chunk : region.chunks ) {
+ // Assume the worst in terms of padding, ie, that every segment is preceded by the max
+ // possible number of padding bytes
+ subCacheTextDataSize += CacheVMSize(chunk->alignment()) + chunk->cacheVMSize;
+ }
+ }
+
+ for ( CacheDylib& cacheDylib : this->cacheDylibs ) {
+ // Peek ahead to find the __TEXT size
+ CacheVMSize textDataSize(0ULL);
+ for ( DylibSegmentChunk& segmentInfo : cacheDylib.segments ) {
+ if ( segmentInfo.kind == DylibSegmentChunk::Kind::dylibLinkedit )
+ continue;
+
+ // Assume the worst in terms of padding, ie, that every segment is preceded by the max
+ // possible number of padding bytes
+ textDataSize += CacheVMSize(segmentInfo.alignment()) + segmentInfo.cacheVMSize;
+
+ // The One Cache stubs will also be included in this limit so account for them too
+ // See addStubsChunks()
+ if ( std::optional<Header::SectionInfo> sectInfo = getStubsSection(&segmentInfo) )
+ textDataSize += CacheVMSize(sectInfo->size);
+ }
+
+ // If we exceed the current limit, then the current subCache is complete and we need
+ // to start a new one
+ if ( (subCacheTextDataSize + textDataSize) > subCacheTextDataLimit ) {
+ // Create a new subCache
+ otherCaches.push_back(SubCache::makeSubCache(this->options));
+ currentSubCache = &otherCaches.back();
+
+ // Reset the limit for the next subCache
+ subCacheTextDataSize = CacheVMSize(regionPadding);
+ }
+
+ subCacheTextDataSize += textDataSize;
+
+ // We'll add LINKEDIT at the end, not here
+ currentSubCache->addDylib(this->config, cacheDylib);
+ currentSubCache->addLinkeditFromDylib(cacheDylib);
+ }
}
// Add all the remaining content in to the final (current) subCache
-
- // Add linkedit chunks from dylibs, if needed
- if ( allLinkeditInLastSubCache ) {
- for ( CacheDylib& cacheDylib : this->cacheDylibs )
- currentSubCache->addLinkeditFromDylib(cacheDylib);
- }
-
// Add all the global optimizations
this->addGlobalOptimizationsToSubCache(*currentSubCache);
@@ -3153,6 +3807,13 @@
sourceStringSize += symbolString.size() + 1;
++sourceStringCount;
+ // rdar://129398821 (dyld cache builder add support for binds relative to dylib segments)
+ // skip synthetic dyld symbols
+ if ( symbolString.find("$dyld$") != std::string_view::npos ) {
+ ++oldSymbolIndex;
+ return;
+ }
+
auto itAndInserted = stringMap.insert({ symbolString, stringBufferSize });
// If we inserted the string, then account for the space
if ( itAndInserted.second )
@@ -3415,17 +4076,54 @@
return Error();
}
+// returns empty string if symbol is in specified boundDylibName,
+// otherwise it return where symbol actually is defined
+static CString getActualSymbolImplDylib(CString symbolName, CString boundDylibName,
+ const std::unordered_map<std::string_view, CacheDylib*>& installNameToDylibMap,
+ bool& symbolImplIsFunctionVariant, uint16_t& functionVariantIndex)
+{
+ // find if bind target is really in another dylib or is a function variant
+ __block CString betterDylib;
+ const auto& posi = installNameToDylibMap.find(boundDylibName);
+ if ( posi != installNameToDylibMap.end() ) {
+ CacheDylib* targetDylib = posi->second;
+ Symbol symbol;
+ if ( targetDylib->inputImage->hasExportsTrie() && targetDylib->inputImage->exportsTrie().hasExportedSymbol(symbolName.c_str(), symbol) ) {
+ betterDylib = (const char*)(targetDylib->installName.data()); // FIXME
+ uint32_t fvi;
+ if (symbol.isFunctionVariant(fvi) ) {
+ symbolImplIsFunctionVariant = true;
+ functionVariantIndex = fvi;
+ }
+ }
+ else {
+ // check if symbol is from re-exported dylib
+ const Header* targetHdr = (Header*)targetDylib->inputMF;
+ targetHdr->forEachLinkedDylib(^(const char* loadPath, LinkedDylibAttributes kind, Version32 compatVersion, Version32 curVersion, bool synthesizedLink, bool &stop) {
+ if ( kind.reExport ) {
+ betterDylib = getActualSymbolImplDylib(symbolName, loadPath, installNameToDylibMap, symbolImplIsFunctionVariant, functionVariantIndex);
+ if ( !betterDylib.empty() ) {
+ stop = true;
+ }
+ }
+ });
+ }
+ }
+ return betterDylib;
+}
+
+
static void parseGOTs(const CacheDylib* dylib, const DylibSegmentChunk* chunk,
std::string_view segmentName, std::string_view sectionName,
- DylibSectionCoalescer::OptimizedSection& dylibOptimizedSection)
-{
- const MachOFile* mf = dylib->inputMF;
+ const std::unordered_map<std::string_view, CacheDylib*>& installNameToDylibMap,
+ OptimizedGOTSection& dylibOptimizedSection)
+{
__block Diagnostics diag;
const bool log = false;
// Skip ineligible dylibs
- if ( !mf->hasChainedFixups() )
+ if ( !dylib->inputMF->hasChainedFixups() )
return;
// Some dylibs have auth gots in segments other than __AUTH_CONST. Skip them for now
@@ -3433,7 +4131,7 @@
return;
__block bool supportsGOTUniquing = false;
- mf->withFileLayout(diag, ^(const mach_o::Layout& layout) {
+ dylib->inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
mach_o::SplitSeg splitSeg(layout);
if ( splitSeg.isV2() )
@@ -3443,13 +4141,11 @@
if ( !supportsGOTUniquing )
return;
- if ( mf->isArch("x86_64") || mf->isArch("x86_64h") ) {
+ if ( dylib->inputHdr->isArch("x86_64") || dylib->inputHdr->isArch("x86_64h") ) {
__block bool oldLinker = false;
- mf->forEachSupportedBuildTool(^(dyld3::Platform platform, uint32_t tool, uint32_t version) {
+ dylib->inputHdr->forEachBuildTool(^(Platform platform, uint32_t tool, uint32_t version) {
uint32_t majorVersion = version >> 16;
- // uint32_t minorVersion = (version >> 8) && 0xFF;
- // uint32_t veryMinorVersion = version && 0xFF;
-
+
if ( tool == TOOL_LD ) {
if ( majorVersion < 803 )
oldLinker = true;
@@ -3461,37 +4157,37 @@
}
// rdar://89319146
- if ( mf->isArch("x86_64") || mf->isArch("x86_64h") ) {
- if ( !strcmp(mf->installName(), "/System/Library/Frameworks/Foundation.framework/Versions/C/Foundation") )
+ if ( dylib->inputHdr->isArch("x86_64") || dylib->inputHdr->isArch("x86_64h") ) {
+ if ( !strcmp(dylib->inputHdr->installName(), "/System/Library/Frameworks/Foundation.framework/Versions/C/Foundation") )
return;
- if ( !strcmp(mf->installName(), "/usr/lib/system/libdispatch.dylib") )
+ if ( !strcmp(dylib->inputHdr->installName(), "/usr/lib/system/libdispatch.dylib") )
return;
}
// rdar://86911139
- if ( mf->builtForPlatform(dyld3::Platform::iOS)
- && !strcmp(mf->installName(), "/System/Library/PrivateFrameworks/CoreUI.framework/CoreUI") )
+ if ( dylib->inputHdr->builtForPlatform(Platform::iOS)
+ && !strcmp(dylib->inputHdr->installName(), "/System/Library/PrivateFrameworks/CoreUI.framework/CoreUI") )
return;
// Dylib segment is eligible. Walk the GOTs
- __block std::optional<dyld3::MachOAnalyzer::SectionInfo> gotSectionInfo;
+ __block std::optional<std::pair<Header::SegmentInfo, Header::SectionInfo>> gotSectionInfo;
__block uint16_t chainedFixupFormat = 0;
- mf->forEachSection(^(const dyld3::MachOAnalyzer::SectionInfo& sectInfo, bool malformedSectionRange, bool& stop) {
- if ( sectInfo.segInfo.segName != segmentName )
+ dylib->inputHdr->forEachSection(^(const Header::SegmentInfo& segInfo, const Header::SectionInfo& sectInfo, bool& stop) {
+ if ( sectInfo.segmentName != segmentName )
return;
- if ( sectInfo.sectName != sectionName)
+ if ( sectInfo.sectionName != sectionName)
return;
- gotSectionInfo = sectInfo;
+ gotSectionInfo = { segInfo, sectInfo };
// As we found the section we want, also get its chained fixup format
- mf->withFileLayout(diag, ^(const mach_o::Layout& layout) {
+ dylib->inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
mach_o::Fixups fixups(layout);
fixups.withChainStarts(diag, ^(const dyld_chained_starts_in_image* starts) {
MachOFile::forEachFixupChainSegment(diag, starts,
^(const dyld_chained_starts_in_segment* segmentInfo,
uint32_t segIndex, bool& stopSegment) {
- if ( segIndex == sectInfo.segInfo.segIndex ) {
+ if ( segIndex == sectInfo.segIndex ) {
chainedFixupFormat = segmentInfo->pointer_format;
stopSegment = true;
}
@@ -3510,7 +4206,7 @@
return;
__block std::vector<mach_o::Fixups::BindTargetInfo> bindTargets;
- mf->withFileLayout(diag, ^(const mach_o::Layout& layout) {
+ dylib->inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
mach_o::Fixups fixups(layout);
fixups.forEachBindTarget(diag, false, 0, ^(const mach_o::Fixups::BindTargetInfo& info, bool& stop) {
@@ -3527,20 +4223,19 @@
return;
__block std::vector<const char*> dependents;
- mf->forEachDependentDylib(^(const char *loadPath, bool isWeak, bool isReExport, bool isUpward, uint32_t compatVersion, uint32_t curVersion, bool &stop) {
+ dylib->inputMF->forEachDependentDylib(^(const char *loadPath, bool isWeak, bool isReExport, bool isUpward, uint32_t compatVersion, uint32_t curVersion, bool &stop) {
dependents.push_back(loadPath);
});
- auto* cacheGotSection = (CoalescedGOTSection*)dylibOptimizedSection.subCacheSection;
- DylibSectionCoalescer::DylibSectionOffsetToCacheSectionOffset& offsetMap = dylibOptimizedSection.offsetMap;
+ bool hasFunctionVariantFixups = dylib->inputHdr->hasFunctionVariantFixups();
// Walk the entries in this section
// File layout so just add the file offset
- const uint8_t* content = (const uint8_t*)mf + gotSectionInfo->sectFileOffset;
+ const uint8_t* content = (const uint8_t*)dylib->inputHdr + gotSectionInfo->second.fileOffset;
const uint8_t* pos = content;
- const uint8_t* end = content + gotSectionInfo->sectSize;
- uint32_t pointerSize = mf->pointerSize();
- assert((gotSectionInfo->sectSize % pointerSize == 0));
+ const uint8_t* end = content + gotSectionInfo->second.size;
+ uint32_t pointerSize = dylib->inputHdr->pointerSize();
+ assert((gotSectionInfo->second.size % pointerSize == 0));
while ( pos != end ) {
const dyld3::MachOLoaded::ChainedFixupPointerOnDisk* fixup = (const dyld3::MachOLoaded::ChainedFixupPointerOnDisk*)pos;
pos += pointerSize;
@@ -3551,56 +4246,83 @@
uint32_t sourceSectionOffset = (uint32_t)((uint64_t)fixup - (uint64_t)content);
// Note down rebases, but otherwise skip them
+ __block bool targetIsFunctionVariant = false;
+ __block uint16_t functionVariantIndex = 0;
if ( !isBind ) {
- dylibOptimizedSection.unoptimizedOffsets.insert(sourceSectionOffset);
- continue;
- }
-
- // We don't support addends right now. But hopefully GOTs don't need them anyway
- if ( addend != 0 )
- continue;
-
- const mach_o::Fixups::BindTargetInfo& bindTarget = bindTargets[bindOrdinal];
-
- // TODO: Weak GOTs. See rdar://86510941
- const char* targetInstallName = nullptr;
- if ( (bindTarget.libOrdinal > 0) && ((unsigned)bindTarget.libOrdinal <= dependents.size()) ) {
- targetInstallName = dependents[bindTarget.libOrdinal - 1];
- } else {
- dylibOptimizedSection.unoptimizedOffsets.insert(sourceSectionOffset);
- continue;
- }
-
- MachOFile::PointerMetaData pmd(fixup, chainedFixupFormat);
-
- typedef CoalescedGOTSection::GOTKey Key;
- Key key = { bindTarget.symbolName, targetInstallName, pmd, bindTarget.weakImport };
-
- int cacheSectionOffset = (int)(cacheGotSection->gotTargetsToOffsets.size() * pointerSize);
- auto itAndInserted = cacheGotSection->gotTargetsToOffsets.insert({ key, cacheSectionOffset });
- if ( itAndInserted.second ) {
- // We inserted the element, so its offset is already valid. Nothing else to do
-
+ if ( hasFunctionVariantFixups ) {
+ // see if this rebase is the local of an internal function variant
+ uint32_t segIndex = gotSectionInfo->first.segmentIndex;
+ dylib->inputImage->functionVariantFixups().forEachFixup(^(mach_o::FunctionVariantFixups::InternalFixup fixupInfo) {
+ if ( fixupInfo.segIndex != segIndex )
+ return;
+ uint64_t segOffset = sourceSectionOffset + gotSectionInfo->second.address - gotSectionInfo->first.vmaddr;
+ if ( fixupInfo.segOffset == segOffset ) {
+ targetIsFunctionVariant = true;
+ functionVariantIndex = fixupInfo.variantIndex;
+ }
+ });
+ }
+ if ( !targetIsFunctionVariant ) {
+ dylibOptimizedSection.addUnoptimizedOffset(sourceSectionOffset);
+ continue;
+ }
+ }
+
+ const CacheDylib* targetDylib = nullptr;
+ CoalescedGOTSection::GOTKey key;
+ if ( targetIsFunctionVariant ) {
+ MachOFile::PointerMetaData pmd(fixup, chainedFixupFormat);
+ char* synthName;
+ asprintf(&synthName, "internal-function-variant#%d", functionVariantIndex);
+ key = { synthName, dylib->installName, pmd, false, true };
+ targetDylib = dylib;
+ }
+ else {
+ // We don't support addends right now. But hopefully GOTs don't need them anyway
+ if ( addend != 0 )
+ continue;
+
+ const mach_o::Fixups::BindTargetInfo& bindTarget = bindTargets[bindOrdinal];
+
+ // TODO: Weak GOTs. See rdar://86510941
+ const char* targetInstallName = nullptr;
+ if ( (bindTarget.libOrdinal > 0) && ((unsigned)bindTarget.libOrdinal <= dependents.size()) ) {
+ targetInstallName = dependents[bindTarget.libOrdinal - 1];
+ } else {
+ dylibOptimizedSection.addUnoptimizedOffset(sourceSectionOffset);
+ continue;
+ }
+
+ // find if bind target is really in another dylib or is a function variant
+ CString betterDylib = getActualSymbolImplDylib(bindTarget.symbolName, targetInstallName, installNameToDylibMap, targetIsFunctionVariant, functionVariantIndex);
+ if ( !betterDylib.empty() && (betterDylib != targetInstallName) ) {
+ targetInstallName = betterDylib.c_str();
+ // fprintf(stderr, "%s found in %s instead of %s\n", bindTarget.symbolName, betterDylib.c_str(), targetInstallName);
+ }
+
+ if ( targetIsFunctionVariant )
+ targetDylib = installNameToDylibMap.at(targetInstallName);
+
+ MachOFile::PointerMetaData pmd(fixup, chainedFixupFormat);
+ key = { bindTarget.symbolName, targetInstallName, pmd, bindTarget.weakImport, targetIsFunctionVariant };
+ }
+
+ if ( dylibOptimizedSection.addOptimizedOffset(sourceSectionOffset, pointerSize, key) ) {
if (log) {
uint64_t gotOffset = ((uint64_t)pos - (uint64_t)content) - pointerSize;
printf("%s[%lld]: %s -> (%s, %s)\n",
- sectionName.data(), gotOffset, mf->installName(),
+ sectionName.data(), gotOffset, dylib->inputHdr->installName(),
key.targetDylibName.data(), key.targetSymbolName.data());
}
- } else {
- // Debugging only. If we didn't include the GOT then we saved that many bytes
- cacheGotSection->savedSpace += pointerSize;
- cacheSectionOffset = itAndInserted.first->second;
- }
-
- // Now keep track of this offset in our source dylib as pointing to this offset
- offsetMap[sourceSectionOffset] = cacheSectionOffset;
- }
-
- // Record which segment/section we just visited
- uint32_t segmentIndex = gotSectionInfo->segInfo.segIndex;
- dylibOptimizedSection.segmentIndex = segmentIndex;
- dylibOptimizedSection.sectionVMOffsetInSegment = VMOffset(gotSectionInfo->sectAddr - gotSectionInfo->segInfo.vmAddr);
+ }
+
+ // store function-variant index in other map
+ if ( targetIsFunctionVariant )
+ dylibOptimizedSection.addFunctionVariantInfo(key, { targetDylib->cacheIndex, functionVariantIndex });
+ }
+
+ // Record which section we just visited
+ dylibOptimizedSection.setSourceSectionInfo(gotSectionInfo->second);
}
// This runs after we've assigned Chunk's to SubCache's, but before we've actually
@@ -3619,9 +4341,12 @@
// DylibSegmentChunk's don't have a pointer to their cache dylib. Make a map for them
std::unordered_map<const InputFile*, CacheDylib*> fileToDylibMap;
+ std::unordered_map<std::string_view, CacheDylib*> installNameToDylibMap;
fileToDylibMap.reserve(this->cacheDylibs.size());
- for ( CacheDylib& dylib : this->cacheDylibs )
+ for ( CacheDylib& dylib : this->cacheDylibs ) {
fileToDylibMap[dylib.inputFile] = &dylib;
+ installNameToDylibMap[dylib.installName] = &dylib;
+ }
for ( SubCache& subCache : this->subCaches ) {
// Find the DATA_CONST/AUTH_CONST in each SubCache, if it has any
@@ -3641,20 +4366,44 @@
if ( (dataConstRegion == nullptr) && (authConstRegion == nullptr) )
continue;
- for ( bool auth : { false, true } ) {
- if ( auth && (authConstRegion == nullptr) )
+ for ( UniquedGOTKind sectionKind : { UniquedGOTKind::regular, UniquedGOTKind::authGot, UniquedGOTKind::authPtr } ) {
+
+ Region* region = nullptr;
+ std::string_view segmentName;
+ std::string_view sectionName;
+ const char* kindName = nullptr;
+ CoalescedGOTSection* subCacheUniquedGOTs = nullptr;
+
+ // Skip sections if their segment doesn't exist
+ switch ( sectionKind ) {
+ case UniquedGOTKind::regular:
+ region = dataConstRegion;
+ segmentName = "__DATA_CONST";
+ sectionName = "__got";
+ kindName = "regular";
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.regularGOTs;
+ break;
+ case UniquedGOTKind::authGot:
+ region = authConstRegion;
+ segmentName = "__AUTH_CONST";
+ sectionName = "__auth_got";
+ kindName = "auth-gots";
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.authGOTs;
+ break;
+ case UniquedGOTKind::authPtr:
+ region = authConstRegion;
+ segmentName = "__AUTH_CONST";
+ sectionName = "__auth_ptr";
+ kindName = "auth-ptrs";
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.authPtrs;
+ break;
+ }
+ if ( region == nullptr )
continue;
- if ( !auth && (dataConstRegion == nullptr) )
- continue;
-
- Region& region = auth ? *authConstRegion : *dataConstRegion;
- std::string_view segmentName = auth ? "__AUTH_CONST" : "__DATA_CONST";
- std::string_view sectionName = auth ? "__auth_got" : "__got";
- CoalescedGOTSection& subCacheUniquedGOTs = auth ? subCache.uniquedGOTsOptimizer.authGOTs : subCache.uniquedGOTsOptimizer.regularGOTs;
-
- std::vector<DylibSectionCoalescer::OptimizedSection*> dylibOptimizedSections;
- dylibOptimizedSections.reserve(region.chunks.size());
- for ( const Chunk* chunk : region.chunks ) {
+
+ std::vector<OptimizedSection*> dylibOptimizedSections;
+ dylibOptimizedSections.reserve(region->chunks.size());
+ for ( const Chunk* chunk : region->chunks ) {
const DylibSegmentChunk* segmentChunk = chunk->isDylibSegmentChunk();
if ( !segmentChunk )
continue;
@@ -3662,109 +4411,34 @@
if ( chunk->name() != segmentName )
continue;
- CacheDylib* dylib = fileToDylibMap.at(segmentChunk->inputFile);
- auto& dylibUniquedGOTs = auth ? dylib->optimizedSections.auth_gots : dylib->optimizedSections.gots;
+ CacheDylib* dylib = fileToDylibMap.at(segmentChunk->inputFile);
+ OptimizedGOTSection* dylibUniquedGOTs = nullptr;
+ switch ( sectionKind ) {
+ case UniquedGOTKind::regular:
+ dylibUniquedGOTs = &dylib->optimizedSections.gots;
+ break;
+ case UniquedGOTKind::authGot:
+ dylibUniquedGOTs = &dylib->optimizedSections.auth_gots;
+ break;
+ case UniquedGOTKind::authPtr:
+ dylibUniquedGOTs = &dylib->optimizedSections.auth_ptrs;
+ break;
+ }
// Set the dylib GOTs to point to the subCache they'll be uniqued to
- dylibUniquedGOTs.subCacheSection = &subCacheUniquedGOTs;
- dylibOptimizedSections.push_back(&dylibUniquedGOTs);
-
- parseGOTs(dylib, segmentChunk, segmentName, sectionName, dylibUniquedGOTs);
- }
-
- if ( subCacheUniquedGOTs.gotTargetsToOffsets.empty() )
+ dylibUniquedGOTs->setSubCacheSection(subCacheUniquedGOTs);
+
+ parseGOTs(dylib, segmentChunk, segmentName, sectionName, installNameToDylibMap, *dylibUniquedGOTs);
+ }
+
+ if ( subCacheUniquedGOTs->empty() )
continue;
- // Sort the coalesced GOTs based on the target install name. We find GOTs in the order we parse
- // the fixups in the dylibs, but we want the final cache to keep all GOTs for the same target near
- // each other
- typedef CoalescedGOTSection::GOTKey Key;
- std::vector<Key> sortedKeys;
- sortedKeys.reserve(subCacheUniquedGOTs.gotTargetsToOffsets.size());
- for ( const auto& keyAndValue : subCacheUniquedGOTs.gotTargetsToOffsets )
- sortedKeys.push_back(keyAndValue.first);
-
- std::sort(sortedKeys.begin(), sortedKeys.end(),
- [](const Key& a, const Key& b) {
- // Put libSystem first, then all the /usr/lib/system dylibs
- // That way any GOTs for re-exports from libsystem will be close to similar GOTs
- bool isLibsystemA = a.targetDylibName.find("libSystem.B.dylib") != std::string_view::npos;
- bool isLibsystemB = b.targetDylibName.find("libSystem.B.dylib") != std::string_view::npos;
- if ( isLibsystemA != isLibsystemB )
- return isLibsystemA;
-
- bool isLibsystemReexportA = a.targetDylibName.find("/usr/lib/system") != std::string_view::npos;
- bool isLibsystemReexportB = b.targetDylibName.find("/usr/lib/system") != std::string_view::npos;
- if ( isLibsystemReexportA != isLibsystemReexportB )
- return isLibsystemReexportA;
-
- if ( a.targetDylibName != b.targetDylibName )
- return (a.targetDylibName < b.targetDylibName);
-
- // Install names are the same. Sort by symbol name
- return a.targetSymbolName < b.targetSymbolName;
- });
-
- // Rewrite entries from their original offset to the new offset
- std::unordered_map<uint32_t, uint32_t> oldToNewOffsetMap;
- for ( uint32_t i = 0; i != sortedKeys.size(); ++i ) {
- const Key& key = sortedKeys[i];
- auto it = subCacheUniquedGOTs.gotTargetsToOffsets.find(key);
- assert(it != subCacheUniquedGOTs.gotTargetsToOffsets.end());
-
- uint32_t newCacheSectionOffset = i * pointerSize;
-
- // Record the offset mapping for updating the dylibs
- oldToNewOffsetMap[it->second] = newCacheSectionOffset;
-
- const bool log = false;
- if ( log ) {
- printf("%s[%d]: %s\n", sectionName.data(), newCacheSectionOffset, key.targetSymbolName.data());
- }
-
- it->second = newCacheSectionOffset;
- }
-
- // Also rewrite entries in each dylib
- for ( DylibSectionCoalescer::OptimizedSection* dylibOptimizedSection : dylibOptimizedSections ) {
- for ( auto& keyAndCacheOffset : dylibOptimizedSection->offsetMap ) {
- auto it = oldToNewOffsetMap.find(keyAndCacheOffset.second);
- assert(it != oldToNewOffsetMap.end());
- keyAndCacheOffset.second = it->second;
- }
- }
-
- // Add the new chunks to the subCache
- if ( auth ) {
- subCache.uniquedAuthGOTs = std::make_unique<UniquedGOTsChunk>();
- subCache.uniquedAuthGOTs->cacheVMSize = CacheVMSize((uint64_t)subCacheUniquedGOTs.gotTargetsToOffsets.size() * pointerSize);
- subCache.uniquedAuthGOTs->subCacheFileSize = CacheFileSize((uint64_t)subCacheUniquedGOTs.gotTargetsToOffsets.size() * pointerSize);
-
- region.chunks.push_back(subCache.uniquedAuthGOTs.get());
-
- // FIXME: Do we need this. No-one seems to read it from here, or could get it from the subCache instead
- subCache.uniquedGOTsOptimizer.authGOTsChunk = subCache.uniquedAuthGOTs.get();
- subCache.uniquedGOTsOptimizer.authGOTs.cacheChunk = subCache.uniquedGOTsOptimizer.authGOTsChunk;
- } else {
- subCache.uniquedGOTs = std::make_unique<UniquedGOTsChunk>();
- subCache.uniquedGOTs->cacheVMSize = CacheVMSize((uint64_t)subCacheUniquedGOTs.gotTargetsToOffsets.size() * pointerSize);
- subCache.uniquedGOTs->subCacheFileSize = CacheFileSize((uint64_t)subCacheUniquedGOTs.gotTargetsToOffsets.size() * pointerSize);
-
- region.chunks.push_back(subCache.uniquedGOTs.get());
-
- // FIXME: Do we need this. No-one seems to read it from here, or could get it from the subCache instead
- subCache.uniquedGOTsOptimizer.regularGOTsChunk = subCache.uniquedGOTs.get();
- subCache.uniquedGOTsOptimizer.regularGOTs.cacheChunk = subCache.uniquedGOTsOptimizer.regularGOTsChunk;
- }
+ subCacheUniquedGOTs->finalize(pointerSize, sectionName, config, subCache, *region);
if ( this->config.log.printStats ) {
- uint64_t totalSourceGOTs = 0;
- for ( DylibSectionCoalescer::OptimizedSection* dylibOptimizedSection : dylibOptimizedSections ) {
- totalSourceGOTs += dylibOptimizedSection->offsetMap.size();
- }
- const char* kind = auth ? "auth" : "regular";
stats.add(" got uniquing: uniqued %lld %s GOTs to %lld GOTs\n",
- totalSourceGOTs, kind, (uint64_t)subCacheUniquedGOTs.gotTargetsToOffsets.size());
+ subCacheUniquedGOTs->numSourceGOTs(), kindName, subCacheUniquedGOTs->numCacheGOTs());
}
}
}
@@ -3797,7 +4471,8 @@
const DylibSegmentChunk* segmentA = a->isDylibSegmentChunk();
const DylibSegmentChunk* segmentB = b->isDylibSegmentChunk();
- if ( segmentA->kind == DylibSegmentChunk::Kind::dylibDataDirty ) {
+ // There can be data chunks that aren't dylib segments, e.g. ObjCHeaderInfoReadWriteChunk.
+ if ( segmentA && segmentB && segmentA->kind == DylibSegmentChunk::Kind::dylibDataDirty ) {
const auto& orderA = dirtyDataSegmentOrdering.find(segmentA->inputFile->path);
const auto& orderB = dirtyDataSegmentOrdering.find(segmentB->inputFile->path);
bool foundA = (orderA != dirtyDataSegmentOrdering.end());
@@ -3813,13 +4488,25 @@
return false;
}
+ const DylibSegmentChunk* dylibA = a->isTPROChunk();
+ const DylibSegmentChunk* dylibB = b->isTPROChunk();
+ // Note this shouldn't be possible, but best to be safe and avoid asserting
+ if ( dylibA && dylibB ) {
+ // Sort dyld last so that its allocator gets packed with TPRO from other dylibs
+ bool isDyldA = dylibA->inputFile->path == "/usr/lib/dyld";
+ bool isDyldB = dylibB->inputFile->path == "/usr/lib/dyld";
+ if ( isDyldA != isDyldB )
+ return !isDyldA;
+ }
+
// Note we are using a stable sort, so if the kind's aren't different, return false
// and we'll keep Section's in the order they were added to the vector
return false;
};
- auto linkeditSortOrder = [](const Chunk* a, const Chunk* b) -> bool {
- // Sort read-only segments before LINKEDIT
+ auto dataConstSortOrder = [](const Chunk* a, const Chunk* b) -> bool {
+ // Sort TPRO_CONST before DATA_CONST. This only happens on x86_64
+ // where we put TPRO_CONST and DATA_CONST in the same Region
if ( a->sortOrder() != b->sortOrder() )
return a->sortOrder() < b->sortOrder();
@@ -3828,19 +4515,93 @@
return false;
};
- // Only sort data/auth. Everything else is already in order
+ auto tproConstSortOrder = [](const Chunk* a, const Chunk* b) -> bool {
+ // Sort uniqued GOTs before dylib TPRO_CONST. This puts the uniqued
+ // GOTs before the libdyld TPRO_CONST, but should still be on the same page
+ if ( a->sortOrder() != b->sortOrder() )
+ return a->sortOrder() < b->sortOrder();
+
+ const DylibSegmentChunk* dylibA = a->isTPROChunk();
+ const DylibSegmentChunk* dylibB = b->isTPROChunk();
+ // Note this shouldn't be possible, but best to be safe and avoid asserting
+ if ( !dylibA || !dylibB )
+ return false;
+
+ // Sort dyld last so that its allocator gets packed with TPRO from other dylibs
+ bool isDyldA = dylibA->inputFile->path == "/usr/lib/dyld";
+ bool isDyldB = dylibB->inputFile->path == "/usr/lib/dyld";
+ if ( isDyldA != isDyldB )
+ return !isDyldA;
+
+ // Note we are using a stable sort, so if the kind's aren't different, return false
+ // and we'll keep Section's in the order they were added to the vector
+ return false;
+ };
+
+ auto readOnlySortOrder = [](const Chunk* a, const Chunk* b) -> bool {
+ // Sort read-only segments before LINKEDIT
+ if ( a->sortOrder() != b->sortOrder() )
+ return a->sortOrder() < b->sortOrder();
+
+ // Note we are using a stable sort, so if the kind's aren't different, return false
+ // and we'll keep Section's in the order they were added to the vector
+ return false;
+ };
+
for ( SubCache& subCache : this->subCaches ) {
for ( Region& region : subCache.regions ) {
- if ( region.kind == Region::Kind::text ) {
- std::stable_sort(region.chunks.begin(), region.chunks.end(), textSortOrder);
- }
- else if ( (region.kind == Region::Kind::data) || (region.kind == Region::Kind::auth) ) {
- std::stable_sort(region.chunks.begin(), region.chunks.end(), dataSortOrder);
- }
- else if ( region.kind == Region::Kind::linkedit ) {
- std::stable_sort(region.chunks.begin(), region.chunks.end(), linkeditSortOrder);
- }
- }
+ switch ( region.kind ) {
+ case Region::Kind::text:
+ std::stable_sort(region.chunks.begin(), region.chunks.end(), textSortOrder);
+ break;
+ case Region::Kind::dataConst:
+ case Region::Kind::authConst:
+ std::stable_sort(region.chunks.begin(), region.chunks.end(), dataConstSortOrder);
+ break;
+ case Region::Kind::tproConst:
+ case Region::Kind::tproAuthConst:
+ std::stable_sort(region.chunks.begin(), region.chunks.end(), tproConstSortOrder);
+ break;
+ case Region::Kind::data:
+ case Region::Kind::auth:
+ std::stable_sort(region.chunks.begin(), region.chunks.end(), dataSortOrder);
+ break;
+ case Region::Kind::readOnly:
+ case Region::Kind::linkedit:
+ std::stable_sort(region.chunks.begin(), region.chunks.end(), readOnlySortOrder);
+ break;
+ case Region::Kind::unmapped:
+ case Region::Kind::dynamicConfig:
+ case Region::Kind::codeSignature:
+ case Region::Kind::numKinds:
+ break;
+ }
+ }
+ }
+
+ // After sorting, we have to add alignment chunks before/after x86_64 TPRO
+ if ( this->config.layout.tproIsInData )
+ addAlignmentChunks();
+}
+
+void SharedCacheBuilder::addAlignmentChunks()
+{
+ for ( SubCache& subCache : this->subCaches ) {
+ SubCache::forEachTPRORegionInData(&subCache, {}, ^(Region& region, const Chunk *firstChunk, const Chunk *lastChunk) {
+ // Add alignment before the first chunk
+ {
+ auto firstPos = std::find(region.chunks.begin(), region.chunks.end(), firstChunk);
+ assert(firstPos != region.chunks.end());
+ region.chunks.insert(firstPos, ®ion.alignmentChunks.emplace_back());
+ }
+
+ // Add alignment after the last chunk
+ {
+ auto lastPos = std::find(region.chunks.begin(), region.chunks.end(), lastChunk);
+ assert(lastPos != region.chunks.end());
+ region.chunks.insert(lastPos + 1, ®ion.alignmentChunks.emplace_back());
+ }
+ });
}
}
@@ -3900,8 +4661,10 @@
};
for ( const SubCache& subCache : this->subCaches ) {
+ calculateRegionSlideInfoSize(this->config, Region::Kind::tproConst, subCache.regions, subCache.tproConstSlideInfo);
calculateRegionSlideInfoSize(this->config, Region::Kind::data, subCache.regions, subCache.dataSlideInfo);
calculateRegionSlideInfoSize(this->config, Region::Kind::dataConst, subCache.regions, subCache.dataConstSlideInfo);
+ calculateRegionSlideInfoSize(this->config, Region::Kind::tproAuthConst, subCache.regions, subCache.tproAuthConstSlideInfo);
calculateRegionSlideInfoSize(this->config, Region::Kind::auth, subCache.regions, subCache.authSlideInfo);
calculateRegionSlideInfoSize(this->config, Region::Kind::authConst, subCache.regions, subCache.authConstSlideInfo);
}
@@ -3933,7 +4696,7 @@
void SharedCacheBuilder::printSubCaches() const
{
- const bool printSegments = this->config.log.printDebug;
+ const bool printSegments = this->config.log.printDebugCacheLayout;
if ( !this->config.log.printStats )
return;
@@ -3952,11 +4715,20 @@
case Region::Kind::dataConst:
regionName = "dataConst";
break;
+ case Region::Kind::tproConst:
+ regionName = "tproConst";
+ break;
case Region::Kind::auth:
regionName = "auth";
break;
case Region::Kind::authConst:
regionName = "authConst";
+ break;
+ case Region::Kind::tproAuthConst:
+ regionName = "tproAuthConst";
+ break;
+ case Region::Kind::readOnly:
+ regionName = "readOnly";
break;
case Region::Kind::linkedit:
regionName = "linkedit";
@@ -4158,12 +4930,15 @@
case Region::Kind::codeSignature:
case Region::Kind::numKinds:
break;
+ case Region::Kind::tproConst:
case Region::Kind::data:
case Region::Kind::dataConst:
+ case Region::Kind::tproAuthConst:
case Region::Kind::auth:
case Region::Kind::authConst:
lastReadWriteRegion = ®ion;
break;
+ case Region::Kind::readOnly:
case Region::Kind::dynamicConfig:
case Region::Kind::linkedit:
lastReadOnlyRegion = ®ion;
@@ -4418,7 +5193,6 @@
const Region& lastRegion = subCache.regions.back();
uint64_t bufferSize = (lastRegion.subCacheFileOffset + lastRegion.subCacheFileSize).rawValue();
-#if SUPPORT_CACHE_BUILDER_MEMORY_BUFFERS
// The MRM builder has no file system, so use an in-memory buffer
vm_address_t fullAllocatedBuffer;
if ( kern_return_t kr = vm_allocate(mach_task_self(), &fullAllocatedBuffer, bufferSize, VM_FLAGS_ANYWHERE); kr != 0 ) {
@@ -4428,47 +5202,6 @@
uint8_t *buffer = (uint8_t*)fullAllocatedBuffer;
subCache.buffer = buffer;
subCache.bufferSize = bufferSize;
-#else
- char pathTemplate[] = "/tmp/temp.XXXXXX";
- int fd = mkstemp(pathTemplate);
- if ( fd == -1 ) {
- // Failed to create the file
- return Error("could not create shared cache file because: %s", strerror(errno));
- }
-
- // Resize the file
- if ( int result = ftruncate(fd, bufferSize); result == -1 ) {
- // Failed to resize to the space we need
- return Error("could not truncate shared cache file because: %s", strerror(errno));
- }
-
- void* buffer = mmap(nullptr, (vm_size_t)bufferSize, PROT_READ | PROT_WRITE, MAP_FILE | MAP_SHARED, fd, 0);
- if ( buffer == MAP_FAILED ) {
- // Failed to mmap the file
- return Error("could not mmap shared cache file because: %s", strerror(errno));
- }
-
- // TODO: It would be great to unlink the file, so that it won't be there on disk if the builder crashes
-#if 0
- // Unlink the file. This way, it'll be removed if we crash
- if ( int result = unlinkat(fd, pathTemplate, 0) ) {
- // Failed to unlink the file
- return Error("could not unlink shared cache file because: %s", strerror(errno));
- }
-
- // Close the file as we don't need it now that we have the mmapped buffer
- if ( int result = close(fd) ) {
- // Failed to close the file
- return Error("could not close shared cache file because: %s", strerror(errno));
- }
-#endif
-
- subCache.buffer = (uint8_t*)buffer;
- subCache.bufferSize = bufferSize;
- subCache.fd = fd;
- subCache.tempPath = pathTemplate;
-
-#endif // SUPPORT_CACHE_BUILDER_MEMORY_BUFFERS
if ( log ) {
this->config.log.log("SubCache[%d] allocated (%p..%p)\n",
@@ -4494,6 +5227,7 @@
assert(!cacheDylib.segments.empty());
assert(cacheDylib.segments[0].kind == cache_builder::DylibSegmentChunk::Kind::dylibText);
cacheDylib.cacheMF = (MachOFile*)cacheDylib.segments[0].subCacheBuffer;
+ cacheDylib.cacheHdr = (const Header*)cacheDylib.segments[0].subCacheBuffer;
cacheDylib.cacheLoadAddress = cacheDylib.segments[0].cacheVMAddress;
}
@@ -4509,7 +5243,7 @@
}
// Add a watchpoint for anything we need to debug
-#if DEBUG
+#if 0
{
CacheVMAddress vmAddrToWatch(0x00007FFB40FB4D58ULL);
for ( const SubCache& subCache : this->subCaches ) {
@@ -4531,6 +5265,31 @@
}
#endif
+
+ // Add a watchpoint for anything we need to debug
+#if 0
+ {
+ CacheFileOffset fileOffsetToWatch(0x0012bc0ULL);
+ std::string_view suffixToWatch = ".01";
+ for ( const SubCache& subCache : this->subCaches ) {
+ if ( subCache.fileSuffix != suffixToWatch )
+ continue;
+ for ( const Region& region : subCache.regions ) {
+ for ( const Chunk* chunk : region.chunks ) {
+ if ( fileOffsetToWatch < chunk->subCacheFileOffset )
+ continue;
+ if ( fileOffsetToWatch >= (chunk->subCacheFileOffset + chunk->subCacheFileSize) )
+ continue;
+ uint64_t offsetInChunk = fileOffsetToWatch.rawValue() - chunk->subCacheFileOffset.rawValue();
+ uint8_t* addrToWatch = chunk->subCacheBuffer + offsetInChunk;
+ printf("watchpoint set expression -w w -s 8 -- %p\n", addrToWatch);
+ printf("");
+ }
+ }
+ }
+ }
+#endif
+
return Error();
}
@@ -4581,7 +5340,6 @@
assert(segment.subCacheFileSize.rawValue() == 0);
segment.subCacheFileSize = region->subCacheFileSize;
-
}
}
@@ -4638,7 +5396,10 @@
case Chunk::Kind::linkeditExportTrie:
movedLinkedit.kind = MovedLinkedit::Kind::exportTrie;
break;
- default:
+ case Chunk::Kind::linkeditFunctionVariants:
+ movedLinkedit.kind = MovedLinkedit::Kind::functionVariants;
+ break;
+ default:
assert(0);
break;
}
@@ -4685,6 +5446,40 @@
});
assert(!err.hasError());
+
+ // Now that we have an adjustor, use it to adjust the bindTargets
+ err = parallel::forEach(this->cacheDylibs, ^(size_t index, CacheDylib& cacheDylib) {
+ for ( CacheDylib::BindTarget& bindTarget : cacheDylib.bindTargets ) {
+ // Adjust the bind target. We have a runtime offset for the target input dylib, but we need to know where that runtime Offset will
+ // map to in the target cache dylib
+ switch ( bindTarget.kind ) {
+ case CacheDylib::BindTarget::Kind::absolute:
+ // Skip these. They won't change due to shifting the input dylib in to the cache
+ break;
+ case CacheDylib::BindTarget::Kind::inputImage: {
+ // Convert from an input dylib offset to the cache dylib offset
+ CacheDylib::BindTarget::InputImage inputImage = bindTarget.inputImage;
+ InputDylibVMAddress targetInputVMAddr = inputImage.targetDylib->inputLoadAddress + inputImage.targetRuntimeOffset;
+ CacheVMAddress targetCacheVMAddr = inputImage.targetDylib->adjustor->adjustVMAddr(targetInputVMAddr);
+
+ // Actually change the bindTarget to reflect the new type
+ bindTarget.kind = CacheDylib::BindTarget::Kind::cacheImage;
+ bindTarget.inputImage.~InputImage();
+ bindTarget.cacheImage = (CacheDylib::BindTarget::CacheImage) {
+ VMOffset(targetCacheVMAddr - inputImage.targetDylib->cacheLoadAddress),
+ inputImage.targetDylib, inputImage.isWeakDef,
+ inputImage.isFunctionVariant, inputImage.functionVariantTableIndex
+ };
+ break;
+ }
+ case CacheDylib::BindTarget::Kind::cacheImage:
+ return Error("Shouldn't see cacheImage fixups at this point");
+ }
+ }
+ return Error();
+ });
+
+ assert(!err.hasError());
}
void SharedCacheBuilder::adjustObjCClasses()
@@ -4793,11 +5588,34 @@
}
}
+Error SharedCacheBuilder::calculatePatchTableSize()
+{
+ Stats stats(this->config);
+ Timer::Scope timedScope(this->config, "calculatePatchTableSize time");
+
+ PatchTableBuilder& builder = this->patchTableOptimizer.builder;
+ Error err = builder.prepare(this->cacheDylibs, this->patchTableOptimizer.patchInfos);
+ if ( err.hasError() )
+ return err;
+
+ uint64_t patchTableSize = builder.getPatchTableSize();
+
+ auto* patchTableChunk = this->patchTableOptimizer.patchTableChunk;
+ patchTableChunk->cacheVMSize = CacheVMSize(patchTableSize);
+ patchTableChunk->subCacheFileSize = CacheFileSize(patchTableSize);
+
+ if ( this->config.log.printStats ) {
+ stats.add(" patch table: used %lld bytes\n", patchTableSize);
+ }
+
+ return Error();
+}
+
Error SharedCacheBuilder::emitPatchTable()
{
Stats stats(this->config);
Timer::Scope timedScope(this->config, "emitPatchTable time");
-
+
// Skip this optimization on simulator until we've qualified it there
__block PatchTableBuilder::PatchableClassesSet patchableObjCClasses;
__block PatchTableBuilder::PatchableSingletonsSet patchableCFObj2;
@@ -4809,38 +5627,204 @@
patchableObjCClasses.insert(cacheDylib.adjustor->adjustVMAddr(inputVMAddr));
});
- // Note we have a diagnostic object here, but we don't care if it fails. Then we'll
- // just skip singleton patching on this dylib
- Diagnostics diag;
- cacheDylib.cacheMF->forEachSingletonPatch(diag, ^(MachOFile::SingletonPatchKind kind, uint64_t runtimeOffset) {
+ cacheDylib.cacheHdr->forEachSingletonPatch(^(uint64_t runtimeOffset) {
patchableCFObj2.insert(cacheDylib.cacheLoadAddress + VMOffset(runtimeOffset));
});
}
}
-
- PatchTableBuilder builder;
- Error err = builder.build(this->cacheDylibs, this->patchTableOptimizer.patchInfos,
- patchableObjCClasses,
- patchableCFObj2,
- this->config.layout.cacheBaseAddress);
- if ( err.hasError() )
- return err;
-
- auto* patchTableChunk = this->patchTableOptimizer.patchTableChunk;
+
+ const auto* patchTableChunk = this->patchTableOptimizer.patchTableChunk;
uint8_t* buffer = patchTableChunk->subCacheBuffer;
uint64_t bufferSize = patchTableChunk->subCacheFileSize.rawValue();
uint64_t patchInfoAddr = patchTableChunk->cacheVMAddress.rawValue();
+
+ PatchTableBuilder& builder = this->patchTableOptimizer.builder;
+ if ( Error error = builder.build(this->cacheDylibs, this->patchTableOptimizer.patchInfos, patchableObjCClasses, patchableCFObj2, this->config.layout.cacheBaseAddress) )
+ return error;
if ( Error error = builder.write(buffer, bufferSize, patchInfoAddr); error.hasError() )
return error;
-
+
// We don't need the patchInfos, so clear it to save memory
this->patchTableOptimizer.patchInfos.clear();
- if ( this->config.log.printStats ) {
- uint64_t patchInfoSize = builder.getPatchTableSize();
- stats.add(" patch table: used %lld out of %lld bytes of buffer\n", patchInfoSize, bufferSize);
- }
-
+ return Error();
+}
+
+Error SharedCacheBuilder::emitPrewarmingData()
+{
+ // Skip everything if the JSON file is empty
+ if ( this->options.prewarmingOptimizations.empty() )
+ return Error::none();
+
+ using json::Node;
+ using json::NodeValueType;
+
+ Stats stats(this->config);
+ Timer::Scope timedScope(this->config, "emitPrewarmingData time");
+
+ __block Diagnostics diag;
+ Node rootNode = json::readJSON(diag, this->options.prewarmingOptimizations.data(),
+ this->options.prewarmingOptimizations.size(),
+ false /* useJSON5 */);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ std::unordered_map<std::string_view, const CacheDylib*> dylibMap;
+ for ( const CacheDylib& cacheDylib : cacheDylibs )
+ dylibMap[cacheDylib.installName] = &cacheDylib;
+
+ // Add install names too, just in case dylibs are moving
+ dylibMap.insert(this->dylibAliases.begin(), this->dylibAliases.end());
+
+ // Format is something like:
+ // [
+ // {
+ // "install_name": "..."
+ // "locations": [
+ // {
+ // "name": "symbol name",
+ // "offset": 0
+ // }
+ // ]
+ // },
+ // ...
+ // ]
+ if ( rootNode.type != NodeValueType::Array )
+ return Error::none();
+
+ std::vector<dyld_prewarming_entry> prewarmingEntries;
+
+ const uint64_t PrewarmingPageSize = DYLD_CACHE_PREWARMING_DATA_PAGE_SIZE;
+ const uint64_t PrewarmingPageMask = ~(PrewarmingPageSize - 1);
+
+ // FIXME: Do this in parallel if needed, or even in CacheDylib when it goes parallel
+ for ( const Node& fileNode : rootNode.array ) {
+ // Find the install name and locations array
+ auto installNameIt = fileNode.map.find("install_name");
+ auto locationsIt = fileNode.map.find("locations");
+ if ( (installNameIt == fileNode.map.end()) || (locationsIt == fileNode.map.end()) )
+ continue;
+
+ const Node& installNameNode = installNameIt->second;
+ const Node& locationsNode = locationsIt->second;
+
+ // Find the cache dylib for this install name
+ const CacheDylib* cacheDylib = nullptr;
+ if ( auto it = dylibMap.find(installNameNode.value); it != dylibMap.end() )
+ cacheDylib = it->second;
+ else
+ continue;
+
+ // Get the exports trie
+ __block const uint8_t* trieStart = nullptr;
+ __block size_t trieSize = 0;
+ cacheDylib->inputMF->withFileLayout(diag, ^(const mach_o::Layout &layout) {
+ if ( layout.linkedit.exportsTrie.hasValue() ) {
+ trieStart = layout.linkedit.exportsTrie.buffer;
+ trieSize = layout.linkedit.exportsTrie.bufferSize;
+ }
+ });
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ if ( trieStart == nullptr )
+ continue;
+
+ ExportsTrie exportsTrie(trieStart, trieSize);
+
+ // Lazily make a map of local symbols if we need it
+ __block std::unordered_map<std::string_view, uint64_t> localsMap;
+ auto populateLocals = ^() {
+ cacheDylib->inputMF->withFileLayout(diag, ^(const mach_o::Layout& layout) {
+ mach_o::SymbolTable symbolTable(layout);
+
+ symbolTable.forEachLocalSymbol(diag, ^(const char *symbolName, uint64_t n_value, uint8_t n_type,
+ uint8_t n_sect, uint16_t n_desc, bool& stop) {
+ if ( ((n_type & N_TYPE) == N_SECT) && ((n_type & N_STAB) == 0) ) {
+ localsMap[symbolName] = (n_value - cacheDylib->inputLoadAddress.rawValue());
+ }
+ });
+ });
+ };
+
+ // Walk all locations, looking them up in exports/locals
+ for ( const Node& locationNode : locationsNode.array ) {
+ auto nameIt = locationNode.map.find("name");
+ auto offsetIt = locationNode.map.find("offset");
+ if ( (nameIt == locationNode.map.end()) || (offsetIt == locationNode.map.end()) )
+ continue;
+
+ uint64_t offset = parseRequiredInt(diag, offsetIt->second);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ // Check for an export first, then a local
+ uint64_t implOffset = 0;
+ mach_o::Symbol symbol;
+ if ( exportsTrie.hasExportedSymbol(nameIt->second.value.c_str(), symbol) ) {
+ if ( !symbol.isRegular(implOffset) )
+ continue;
+ } else {
+ if ( localsMap.empty() ) {
+ populateLocals();
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+ }
+
+ if ( auto localsIt = localsMap.find(nameIt->second.value); localsIt != localsMap.end() )
+ implOffset = localsIt->second;
+ }
+
+ InputDylibVMAddress inputVMAddr(cacheDylib->inputLoadAddress + VMOffset(implOffset));
+ CacheVMAddress cacheVMAddr = cacheDylib->adjustor->adjustVMAddr(inputVMAddr);
+ cacheVMAddr += VMOffset(offset);
+
+ VMOffset cacheVMOffset = cacheVMAddr - this->config.layout.cacheBaseAddress;
+
+ prewarmingEntries.push_back({ cacheVMOffset.rawValue() & PrewarmingPageMask, 1 });
+ }
+ }
+
+ PrewarmingOptimizer& opt = this->prewarmingOptimizer;
+ dyld_prewarming_header* header = (dyld_prewarming_header*)opt.prewarmingChunk->subCacheBuffer;
+ const void* bufferEnd = opt.prewarmingChunk->subCacheBuffer + opt.prewarmingChunk->subCacheFileSize.rawValue();
+
+ if ( prewarmingEntries.empty() ) {
+ header->version = 1;
+ header->count = 0;
+ return Error::none();
+ }
+
+ // Sort by ascending VM addresses
+ std::sort(prewarmingEntries.begin(), prewarmingEntries.end(),
+ [](const dyld_prewarming_entry& a, const dyld_prewarming_entry& b) {
+ return a.cacheVMOffset < b.cacheVMOffset;
+ });
+
+ // Combine adjacent entries
+ std::vector<dyld_prewarming_entry> combinedEntries;
+ combinedEntries.push_back(prewarmingEntries.front());
+ for ( const dyld_prewarming_entry& entry : std::span(prewarmingEntries).subspan(1) ) {
+ dyld_prewarming_entry& last = combinedEntries.back();
+ if ( entry.cacheVMOffset == last.cacheVMOffset )
+ continue;
+ if ( entry.cacheVMOffset == (last.cacheVMOffset + PrewarmingPageSize) ) {
+ ++last.numPages;
+ } else {
+ combinedEntries.push_back(entry);
+ }
+ }
+
+ // write the results
+ dyld_prewarming_entry* firstOutputEntry = &header->entries[0];
+ dyld_prewarming_entry* lastOutputEntry = &header->entries[combinedEntries.size()];
+ if ( lastOutputEntry > bufferEnd )
+ return Error("overflow in prewarming optimizer (%p vs %p)", lastOutputEntry, bufferEnd);
+
+ header->version = 1;
+ header->count = (uint32_t)combinedEntries.size();
+ memcpy(firstOutputEntry, combinedEntries.data(), sizeof(dyld_prewarming_entry) * header->count);
+
return Error();
}
@@ -4870,7 +5854,7 @@
}
} else {
const char* binPath = "/usr/bin/";
- if ( options.platform == dyld3::Platform::driverKit )
+ if ( options.platform == Platform::driverKit )
binPath = "/System/Library/DriverExtensions/";
for ( const InputFile* exeFile : executableFiles ) {
if ( startsWith(exeFile->path, binPath) )
@@ -4928,8 +5912,8 @@
}
// The cache segments don't have the permissions. Get that from the load commands
- cacheDylib.cacheMF->forEachSegment(^(const MachOFile::SegmentInfo& info, bool& stop) {
- segments[info.segIndex].protections = info.protections;
+ cacheDylib.cacheHdr->forEachSegment(^(const Header::SegmentInfo& info, bool& stop) {
+ segments[info.segmentIndex].protections = info.initProt;
});
this->dylibSegmentLayout.push_back(std::move(segments));
}
@@ -4956,6 +5940,11 @@
linkeditLayout.exportsTrie.entryCount = 0; // Not needed here
linkeditLayout.exportsTrie.hasLinkedit = true;
break;
+ case MovedLinkedit::Kind::functionVariants:
+ linkeditLayout.functionVariants.buffer = kindAndLinkdit.second.cacheLocation;
+ linkeditLayout.functionVariants.bufferSize = (uint32_t)kindAndLinkdit.second.dataSize.rawValue();
+ linkeditLayout.functionVariants.hasLinkedit = true;
+ break;
case MovedLinkedit::Kind::numKinds:
// This should never happen
assert(false);
@@ -4981,21 +5970,21 @@
this->executableSegmentLayout.reserve(executableFiles.size());
for ( const InputFile* executableFile : executableFiles ) {
__block std::vector<mach_o::SegmentLayout> segments;
- executableFile->mf->forEachSegment(^(const MachOFile::SegmentInfo& info, bool& stop) {
+ ((const Header*)executableFile->mf)->forEachSegment(^(const Header::SegmentInfo& info, bool& stop) {
// Note file layout here, not VM layout
uint8_t* segmentBuffer = (uint8_t*)executableFile->mf + info.fileOffset;
mach_o::SegmentLayout segment;
- segment.vmAddr = info.vmAddr;
- segment.vmSize = info.vmSize;
+ segment.vmAddr = info.vmaddr;
+ segment.vmSize = info.vmsize;
segment.fileOffset = info.fileOffset;
segment.fileSize = info.fileSize;
segment.buffer = segmentBuffer;
- segment.protections = info.protections;
+ segment.protections = info.initProt;
segment.kind = mach_o::SegmentLayout::Kind::unknown;
- if ( !strcmp(info.segName, "__TEXT") ) {
+ if ( info.segmentName == "__TEXT" ) {
segment.kind = mach_o::SegmentLayout::Kind::text;
- } else if ( !strcmp(info.segName, "__LINKEDIT") ) {
+ } else if ( info.segmentName == "__LINKEDIT" ) {
segment.kind = mach_o::SegmentLayout::Kind::linkedit;
}
@@ -5087,7 +6076,7 @@
}
}
- Loader::LoadOptions::Finder loaderFinder = ^(Diagnostics& loadDiag, dyld3::Platform, const char* loadPath, const dyld4::Loader::LoadOptions& options) {
+ Loader::LoadOptions::Finder loaderFinder = ^(Diagnostics& loadDiag, Platform, const char* loadPath, const dyld4::Loader::LoadOptions& options) {
auto pos = loadersMap.find(loadPath);
if ( pos != loadersMap.end() ) {
return (const Loader*)pos->second;
@@ -5116,7 +6105,7 @@
Diagnostics loadDiag;
((Loader*)ldr)->loadDependents(loadDiag, state, options);
if ( loadDiag.hasError() ) {
- return Error("%s, loading dependents of %s", loadDiag.errorMessageCStr(), ldr->path());
+ return Error("%s, loading dependents of %s", loadDiag.errorMessageCStr(), ldr->path(state));
}
}
@@ -5148,7 +6137,7 @@
return Error("Could not find a main executable for building cache loaders");
const LayoutBuilder layoutBuilder(cacheDylibs, { });
- EphemeralAllocator processConfigAlloc;
+ STACK_ALLOCATOR(processConfigAlloc, 0);
__block dyld4::Vector<ProcessConfig::DyldCache::CacheDylib> processConfigDylibs(processConfigAlloc);
for ( uint32_t dylibIndex = 0; dylibIndex != this->cacheDylibs.size(); ++dylibIndex ) {
@@ -5171,9 +6160,9 @@
}
// build PrebuiltLoaderSet of all dylibs in cache
+ STACK_ALLOCATOR(alloc, 0);
KernelArgs kernArgs(mainExecutable, { "test.exe" }, {}, {});
SyscallDelegate osDelegate;
- EphemeralAllocator alloc;
ProcessConfig processConfig(&kernArgs, osDelegate, alloc);
RuntimeLocks locks;
RuntimeState state(processConfig, locks, alloc);
@@ -5236,14 +6225,14 @@
__block InputDylibVMAddress inputOptPtrsVMAddress;
__block uint64_t sectionSize = 0;
__block bool found = false;
- cacheDylib->inputMF->forEachSection(^(const dyld3::MachOAnalyzer::SectionInfo& sectInfo, bool malformedSectionRange, bool& stop) {
- if ( (strncmp(sectInfo.segInfo.segName, "__DATA", 6) != 0) && (strncmp(sectInfo.segInfo.segName, "__AUTH", 6) != 0) )
+ cacheDylib->inputHdr->forEachSection(^(const Header::SectionInfo& sectInfo, bool& stop) {
+ if ( !sectInfo.segmentName.starts_with("__DATA") && !sectInfo.segmentName.starts_with("__AUTH") )
return;
- if ( strcmp(sectInfo.sectName, "__objc_opt_ptrs") != 0 )
+ if ( sectInfo.sectionName != "__objc_opt_ptrs" )
return;
- inputOptPtrsVMAddress = InputDylibVMAddress(sectInfo.sectAddr);
- sectionSize = sectInfo.sectSize;
+ inputOptPtrsVMAddress = InputDylibVMAddress(sectInfo.address);
+ sectionSize = sectInfo.size;
found = true;
stop = true;
@@ -5260,7 +6249,7 @@
// T protocolClass;
// };
// typedef struct objc_opt_pointerlist_tt<uintptr_t> objc_opt_pointerlist_t;
- if ( sectionSize < cacheDylib->inputMF->pointerSize() ) {
+ if ( sectionSize < cacheDylib->inputHdr->pointerSize() ) {
return Error("libobjc's pointer list section is too small (metadata not optimized)");
}
@@ -5322,7 +6311,7 @@
const LayoutBuilder layoutBuilder(cacheDylibs, this->exeInputFiles);
const LayoutBuilder* layoutBuilderPtr = &layoutBuilder;
- EphemeralAllocator processConfigAlloc;
+ STACK_ALLOCATOR(processConfigAlloc, 0);
dyld4::Vector<ProcessConfig::DyldCache::CacheDylib> processConfigDylibsOwner(processConfigAlloc);
auto& processConfigDylibs = processConfigDylibsOwner;
@@ -5354,8 +6343,8 @@
}
// Assume last segment file size is the overall file size
- __block uint64_t fileSize = 0;
- inputFile->mf->forEachSegment(^(const MachOFile::SegmentInfo& info, bool& stop) {
+ __block uint32_t fileSize = 0;
+ ((const Header*)inputFile->mf)->forEachSegment(^(const Header::SegmentInfo& info, bool& stop) {
fileSize = std::max(fileSize, info.fileOffset + info.fileSize);
});
otherMapping[inputFile->path] = { inputFile->mf, fileSize };
@@ -5390,16 +6379,16 @@
KernelArgs kernArgs(mainMF, { "test.exe" }, {}, {});
SyscallDelegate osDelegate;
osDelegate._mappedOtherDylibs = otherMapping;
- osDelegate._gradedArchs = &this->options.archs;
+ osDelegate._gradedArchs = &this->options.gradedArchs;
//osDelegate._dyldCache = dyldCache;
- EphemeralAllocator alloc;
+ STACK_ALLOCATOR(alloc, 0);
ProcessConfig processConfig(&kernArgs, osDelegate, alloc);
RuntimeLocks locks;
RuntimeState state(processConfig, locks, alloc);
RuntimeState* statePtr = &state;
Diagnostics launchDiag;
- processConfig.reset(mainMF, exeFile->path.c_str(), nullptr);
+ processConfig.reset(mainMF, exeFile->path.c_str(), exeFile->size, nullptr);
state.resetCachedDylibsArrays(cachedDylibsLoaderSet);
// FIXME: This is terrible and needs to be a real reset method
@@ -5407,7 +6396,9 @@
processConfig.dyldCache.dylibsExpectedOnDisk = !this->options.dylibsRemovedFromDisk;
processConfig.dyldCache.development = isDevelopmentSharedCache(this->options);
- if ( !this->objcOptimizer.objcDylibs.empty() ) {
+ // Disable objc optimizations from EK shared cache
+ bool emitObjcOpts = !this->options.platform.isExclaveKit();
+ if ( !this->objcOptimizer.objcDylibs.empty() && emitObjcOpts ) {
processConfig.dyldCache.objcClassHashTable = (const objc::ClassHashTable*)this->objcClassOptimizer.classHashTableChunk->subCacheBuffer;
processConfig.dyldCache.objcSelectorHashTable = (const objc::SelectorHashTable*)this->objcSelectorOptimizer.selectorHashTableChunk->subCacheBuffer;
processConfig.dyldCache.objcProtocolHashTable = (const objc::ProtocolHashTable*)this->objcProtocolOptimizer.protocolHashTableChunk->subCacheBuffer;
@@ -5418,18 +6409,24 @@
processConfig.dyldCache.unslidLoadAddress = config.layout.cacheBaseAddress.rawValue();
}
- Loader::LoadOptions::Finder loaderFinder = ^(Diagnostics& diag, dyld3::Platform plat, const char* loadPath, const dyld4::Loader::LoadOptions& loadOptions) {
+ // check if main executable is valid mach-o
+ if ( mach_o::Error lerr = state.config.process.mainExecutableHdr->valid(state.config.process.mainExecutableSliceSize) ) {
+ fprintf(stderr, "warning: can't build PrebuiltLoaderSet for '%s': %s\n", exeFile->path.c_str(), lerr.message());
+ return Error();
+ }
+
+ Loader::LoadOptions::Finder loaderFinder = ^(Diagnostics& diag, Platform plat, const char* loadPath, const dyld4::Loader::LoadOptions& loadOptions) {
// when building macOS cache, there may be some incorrect catalyst paths
- if ( (plat == dyld3::Platform::iOSMac) && (strncmp(loadPath, "/System/iOSSupport/", 19) != 0) ) {
+ if ( (plat == Platform::macCatalyst) && (strncmp(loadPath, "/System/iOSSupport/", 19) != 0) ) {
char altPath[PATH_MAX];
strlcpy(altPath, "/System/iOSSupport", PATH_MAX);
strlcat(altPath, loadPath, PATH_MAX);
- if ( const dyld4::PrebuiltLoader* ldr = cachedDylibsLoaderSet->findLoader(altPath) )
+ if ( const dyld4::PrebuiltLoader* ldr = cachedDylibsLoaderSet->findLoader(*statePtr, altPath) )
return (const Loader*)ldr;
}
// check if path is a dylib in the dyld cache, then use its PrebuiltLoader
- if ( const dyld4::PrebuiltLoader* ldr = cachedDylibsLoaderSet->findLoader(loadPath) )
+ if ( const dyld4::PrebuiltLoader* ldr = cachedDylibsLoaderSet->findLoader(*statePtr, loadPath) )
return (const Loader*)ldr;
// call through to getLoader() which will expand @paths
@@ -5458,6 +6455,23 @@
// FIXME: Propagate errors
return Error();
}
+
+ // Set dylibs to be fixedUp before we partition delay init, as it uses this state
+ for ( const Loader* ldr : state.loaded ) {
+ if ( const PrebuiltLoader* prebuiltLdr = ldr->isPrebuiltLoader() )
+ prebuiltLdr->setFixedUp(state);
+ }
+
+ // split off delay loaded dylibs into delayLoaded vector
+ // We have to do this before making the PrebuiltLoaderSet as objc in the closure needs
+ // to know which shared cache dylibs are delay or not.
+ STACK_ALLOC_ARRAY(const Loader*, loadersTemp, state.loaded.size());
+ for (const Loader* ldr : state.loaded)
+ loadersTemp.push_back(ldr);
+ std::span<const Loader*> allLoaders(&loadersTemp[0], (size_t)loadersTemp.count());
+ std::span<const Loader*> topLoaders = allLoaders.subspan(0, 1);
+ state.partitionDelayLoads(allLoaders, topLoaders);
+
state.setMainLoader(mainLoader);
const dyld4::PrebuiltLoaderSet* prebuiltAppSet = dyld4::PrebuiltLoaderSet::makeLaunchSet(launchDiag, state, missingPaths);
if ( launchDiag.hasError() ) {
@@ -5483,40 +6497,39 @@
return Error();
});
+ const PrebuiltLoaderChunk* loaderChunk = this->prebuiltLoaderBuilder.executablesLoaderChunk;
+ const uint64_t loaderBufferSize = loaderChunk->subCacheFileSize.rawValue();
+
assert(!err.hasError());
-
std::map<std::string_view, const dyld4::PrebuiltLoaderSet*> prebuiltsMap;
uint64_t prebuiltsSpace = 0;
+ uint64_t requiredPrebuiltsSpace = 0;
for ( uint64_t i = 0; i != this->exeInputFiles.size(); ++i ) {
const InputFile* exeFile = this->exeInputFiles[i];
const dyld4::PrebuiltLoaderSet* loaderSet = executableLoaders[i];
if ( loaderSet == nullptr )
continue;
+ uint64_t requiredSpace = alignTo(loaderSet->size(), alignof(dyld4::PrebuiltLoaderSet));
+ requiredPrebuiltsSpace += requiredSpace;
+ if ( (prebuiltsSpace + requiredSpace) > loaderBufferSize ) {
+ // skip this loader if it doesn't fit
+ this->warning("cache buffer too small to hold executable PrebuiltLoaderSet '%s' (size=%lldKB)",
+ exeFile->path.c_str(), requiredSpace / 1024);
+ continue;
+ }
+
prebuiltsMap[exeFile->path.c_str()] = loaderSet;
- prebuiltsSpace += alignTo(loaderSet->size(), 8);
- }
-
- const PrebuiltLoaderChunk* loaderChunk = this->prebuiltLoaderBuilder.executablesLoaderChunk;
- uint64_t loaderBufferSize = loaderChunk->subCacheFileSize.rawValue();
+ prebuiltsSpace += requiredSpace;
+ }
+
+ if ( requiredPrebuiltsSpace != prebuiltsSpace ) {
+ this->warning("cache buffer too small to hold executable PrebuiltLoaderSet (prebuiltLoaderSet size=%lluKB, buffer size=%lldKB)",
+ requiredPrebuiltsSpace / 1024, loaderBufferSize / 1024);
+ }
if ( this->config.log.printStats ) {
stats.add(" dyld4 executable Loader's : used %lld out of %lld bytes of buffer\n", prebuiltsSpace, loaderBufferSize);
- }
-
- if ( prebuiltsSpace > loaderBufferSize ) {
- if ( dylibHasMissingDependency ) {
- // At least one dylib was evicted. If it was soemthing common, like UIKit/AppKit, then its going to
- // end up being included in every executable loader and the buffer will overflow
- this->warning("cache buffer too small to hold executable PrebuiltLoaderSet (prebuiltLoaderSet size=%lluKB, buffer size=%lldKB)",
- prebuiltsSpace / 1024, loaderBufferSize / 1024);
-
- // For now, just empty the map. That'll let us emit an empty Trie and PBLS
- prebuiltsMap.clear();
- } else {
- return Error("cache buffer too small to hold executable PrebuiltLoaderSet (prebuiltLoaderSet size=%lluKB, buffer size=%lldKB)",
- prebuiltsSpace / 1024, loaderBufferSize / 1024);
- }
}
// copy all PrebuiltLoaderSets into cache
@@ -5554,7 +6567,7 @@
size_t size = pbls->size();
::memcpy(poolBase + currentPoolOffset, pbls, size);
- currentPoolOffset += alignTo(size, 8);
+ currentPoolOffset += alignTo(size, alignof(dyld4::PrebuiltLoaderSet));
}
const CacheTrieChunk* trieChunk = this->prebuiltLoaderBuilder.executableTrieChunk;
@@ -5869,7 +6882,7 @@
// Get the PrebuiltLoader* for this cache dylib
const PrebuiltLoader* ldr = cachedDylibsLoaderSet->atIndex(cacheDylib.cacheIndex);
- assert(ldr->path() == cacheDylib.installName);
+ //assert(ldr->path(state) == cacheDylib.installName); // can't do assert because state is not passed to this method
CacheVMAddress ldrVMAddr = getVMAddressInSection(*this->prebuiltLoaderBuilder.cacheDylibsLoaderChunk, ldr);
@@ -5998,26 +7011,25 @@
if ( cacheDylib.installName != "/usr/lib/libobjc.A.dylib" )
continue;
- cacheDylib.cacheMF->forEachSection(^(const dyld3::MachOFile::SectionInfo §Info,
- bool malformedSectionRange, bool &stop) {
- if (strcmp(sectInfo.segInfo.segName, "__TEXT") != 0)
+ cacheDylib.cacheHdr->forEachSection(^(const Header::SegmentInfo &segInfo, const Header::SectionInfo §Info, bool &stop) {
+ if ( sectInfo.segmentName != "__TEXT" )
return;
- if (strcmp(sectInfo.sectName, "__objc_opt_ro") != 0)
+ if ( sectInfo.sectionName != "__objc_opt_ro" )
return;
// Find the buffer for the section
stop = true;
- const DylibSegmentChunk& segment = cacheDylib.segments[sectInfo.segInfo.segIndex];
-
- VMAddress sectionVMAddr(sectInfo.sectAddr);
- VMAddress segmentVMAddr(sectInfo.segInfo.vmAddr);
+ const DylibSegmentChunk& segment = cacheDylib.segments[sectInfo.segIndex];
+
+ VMAddress sectionVMAddr(sectInfo.address);
+ VMAddress segmentVMAddr(segInfo.vmaddr);
VMOffset sectionOffsetInSegment = sectionVMAddr - segmentVMAddr;
uint8_t* sectionBuffer = segment.subCacheBuffer + sectionOffsetInSegment.rawValue();
// All fields in the old header are offsets from the header. This is how much to
// shift them by
- uint64_t headerCacheOffset = sectInfo.sectAddr - this->config.layout.cacheBaseAddress.rawValue();
+ uint64_t headerCacheOffset = sectInfo.address - this->config.layout.cacheBaseAddress.rawValue();
// Found the section, now write the content
objc_opt::objc_opt_t* libROHeader = (objc_opt::objc_opt_t *)sectionBuffer;
@@ -6063,22 +7075,12 @@
}
}
-// Struct matching dyld4::LibdyldDyld4Section to be used with a variable pointer size.
-// This is so we can use it in the shared cache builder, which is always
-// 64-bit but can emit 32-bit structs
-template <typename P>
-struct FixedSizeLibdyldDyld4Section {
- P apis;
- P allImageInfos;
- P defaultVars[5];
- P dyldLookupFuncAddr;
- P tlv_get_addrAddr;
-};
-
-static_assert(sizeof(FixedSizeLibdyldDyld4Section<intptr_t>) == sizeof(dyld4::LibdyldDyld4Section));
-
void SharedCacheBuilder::optimizeTLVs()
{
+ // driverkit does not support thread-locals, all other platforms do
+ if ( options.platform == Platform::driverKit )
+ return;
+
Stats stats(this->config);
Timer::Scope timedScope(this->config, "optimizeTLVs time");
@@ -6105,60 +7107,20 @@
return;
}
- // Find the tlv_get_addrAddr from inside the __dyld4 section
- __block CacheVMAddress getAddrVMAddr;
- __block bool foundTLVGetAddr = false;
- libdyldDylib->cacheMF->forEachSection(^(const MachOFile::SectionInfo §Info,
- bool malformedSectionRange, bool &stop) {
- if ( strcmp(sectInfo.sectName, "__dyld4") != 0 )
- return;
-
- if ( (strncmp(sectInfo.segInfo.segName, "__DATA", 6) != 0)
- && (strncmp(sectInfo.segInfo.segName, "__AUTH", 6) != 0) )
- return;
-
- // Found the section we need. Now to check if its valid
- stop = true;
-
- const DylibSegmentChunk& segment = libdyldDylib->segments[sectInfo.segInfo.segIndex];
-
- VMAddress sectionVMAddr(sectInfo.sectAddr);
- VMAddress segmentVMAddr(sectInfo.segInfo.vmAddr);
- VMOffset sectionOffsetInSegment = sectionVMAddr - segmentVMAddr;
- uint8_t* sectionBuffer = segment.subCacheBuffer + sectionOffsetInSegment.rawValue();
-
-
- if ( this->config.layout.is64 ) {
- typedef FixedSizeLibdyldDyld4Section<uint64_t> dyld4_section_t;
- if ( sectInfo.sectSize < sizeof(dyld4_section_t) ) {
- // Old libdyld without the field we need
- return;
- }
-
- const dyld4_section_t* dyldSection = (dyld4_section_t*)sectionBuffer;
- CacheVMAddress cacheBaseAddress = this->config.layout.cacheBaseAddress;
- getAddrVMAddr = Fixup::Cache64::getCacheVMAddressFromLocation(cacheBaseAddress,
- &dyldSection->tlv_get_addrAddr);
- } else {
- typedef FixedSizeLibdyldDyld4Section<uint32_t> dyld4_section_t;
- if ( sectInfo.sectSize < sizeof(dyld4_section_t) ) {
- // Old libdyld without the field we need
- return;
- }
-
- const dyld4_section_t* dyldSection = (dyld4_section_t*)sectionBuffer;
- CacheVMAddress cacheBaseAddress = this->config.layout.cacheBaseAddress;
- getAddrVMAddr = Fixup::Cache32::getCacheVMAddressFromLocation(cacheBaseAddress,
- &dyldSection->tlv_get_addrAddr);
- }
-
- foundTLVGetAddr = true;
- });
-
- if ( !foundTLVGetAddr ) {
- this->warning("Could not find tlv_get_addr (TLVs not optimized)");
+ // Find _tlv_get_addr function in libdyld.dylib
+ const Image& libdyldImg = *libdyldDylib->inputImage;
+
+ Symbol tlvBootstrapSymbol;
+ if ( !libdyldImg.hasExportsTrie() || !libdyldImg.exportsTrie().hasExportedSymbol("__tlv_bootstrap", tlvBootstrapSymbol) ) {
+ this->warning("Could not find '_tlv_get_addr' in libdyld");
return;
}
+ // NOTE: magic here:
+ // To cleanly error of if TLVs are not set up, the on-disk thunks bind to _tlv_bootstrap, which aborts in invoked.
+ // The thunks when setup are changed to point to _tlv_get_addr, which is not an exported symbol. Therefore
+ // the cache builder has no way to find it. The fix/hack is that _tlv_bootstrap and _tlv_get_addr are written
+ // in assembly and always 8 bytes apart.
+ CacheVMAddress getAddrVMAddr(libdyldDylib->cacheLoadAddress.rawValue() + tlvBootstrapSymbol.implOffset() + 8);
// We read the value for this symbol to know the first key we can allocate for TLVs
// We then have to stop optimizing if and when we reach "end", that's the
@@ -6210,7 +7172,7 @@
return;
}
- if ( !cacheDylib.cacheMF->hasThreadLocalVariables() )
+ if ( !cacheDylib.cacheHdr->hasThreadLocalVariables() )
continue;
// Get the next available key (one key per dylib)
@@ -6227,69 +7189,117 @@
return;
}
- cacheDylib.cacheMF->forEachSection(^(const MachOFile::SectionInfo& sectInfo,
- bool malformedSectionRange, bool& stop) {
- if ( (sectInfo.sectFlags & SECTION_TYPE) != S_THREAD_LOCAL_VARIABLES )
+ // find initial content for all thread locals in this dylib
+ __block bool initialContentNonZero = false;
+ __block uint64_t initialContentAddr = 0;
+ __block size_t initialContentSize = 0;
+ cacheDylib.cacheHdr->forEachSection(^(const Header::SectionInfo& sectInfo, bool& stop) {
+ switch ( sectInfo.flags & SECTION_TYPE ) {
+ case S_THREAD_LOCAL_REGULAR:
+ initialContentNonZero = true;
+ [[clang::fallthrough]];
+ case S_THREAD_LOCAL_ZEROFILL:
+ if ( initialContentSize == 0 ) {
+ // first of N contiguous TLV template sections, record as if this was only section
+ initialContentAddr = sectInfo.address;
+ initialContentSize = sectInfo.size;
+ }
+ else {
+ // non-first of N contiguous TLV template sections, accumlate values
+ initialContentSize = sectInfo.address + sectInfo.size - initialContentAddr;
+ }
+ break;
+ }
+ });
+
+ cacheDylib.cacheHdr->forEachSection(^(const Header::SegmentInfo &segInfo, const Header::SectionInfo& sectInfo, bool& stop) {
+ if ( (sectInfo.flags & SECTION_TYPE) != S_THREAD_LOCAL_VARIABLES )
return;
- DylibSegmentChunk& segment = cacheDylib.segments[sectInfo.segInfo.segIndex];
-
- VMAddress sectionVMAddr(sectInfo.sectAddr);
- VMAddress segmentVMAddr(sectInfo.segInfo.vmAddr);
+ DylibSegmentChunk& segment = cacheDylib.segments[sectInfo.segIndex];
+
+ VMAddress sectionVMAddr(sectInfo.address);
+ VMAddress segmentVMAddr(segInfo.vmaddr);
VMOffset sectionOffsetInSegment = sectionVMAddr - segmentVMAddr;
uint8_t* sectionBuffer = segment.subCacheBuffer + sectionOffsetInSegment.rawValue();
if ( this->config.layout.is64 ) {
- struct tlv_thunk_t
+ struct tlv_disk_thunk64_t
{
- uint64_t thunk;
+ uint64_t func;
uint64_t key;
uint64_t offset;
};
- uint32_t count = (uint32_t)(sectInfo.sectSize / sizeof(tlv_thunk_t));
- tlv_thunk_t* thunkBuffer = (tlv_thunk_t*)sectionBuffer;
- for ( uint32_t i = 0; i != count; ++i ) {
- tlv_thunk_t& tlvThunk = thunkBuffer[i];
-
- // Set the key to the next available key
- tlvThunk.key = key;
+ uint32_t count = (uint32_t)(sectInfo.size / sizeof(tlv_disk_thunk64_t));
+ tlv_disk_thunk64_t* thunkBuffer = (tlv_disk_thunk64_t*)sectionBuffer;
+ for ( uint32_t i = 0; i < count; ++i ) {
+ size_t offset = thunkBuffer[i].offset; // only value needed from disk format
+ dyld::ThreadLocalVariables::TLV_Thunkv2& newThunk = *((dyld::ThreadLocalVariables::TLV_Thunkv2*)&thunkBuffer[i]);
+
+ // Reset the fields using TLV_Thunkv2 format
+ newThunk.key = key;
+ newThunk.offset = (uint32_t)offset;
+ newThunk.initialContentSize = (uint32_t)initialContentSize;
+ newThunk.initialContentDelta = 0;
+ if ( initialContentNonZero ) {
+ uint64_t newThunkDeltaFieldAddr = (sectInfo.address + i*sizeof(tlv_disk_thunk64_t) + offsetof(dyld::ThreadLocalVariables::TLV_Thunkv2,initialContentDelta));
+ newThunk.initialContentDelta = (int32_t)(initialContentAddr - newThunkDeltaFieldAddr);
+ }
// Set the thunk to tlv_get_addr()
- uint8_t high8 = 0;
- uint16_t authDiversity = 0;
- bool authHasAddrDiv = false;
- uint8_t authKey = ptrauth_key_asia;
- bool isAuth = this->config.layout.hasAuthRegion;
- Fixup::Cache64::setLocation(this->config.layout.cacheBaseAddress, &tlvThunk.thunk,
+ uint8_t high8 = 0;
+ uint16_t authDiversity = 0;
+ bool authHasAddrDiv = false;
+ uint8_t authKey = ptrauth_key_asia;
+ bool isAuth = this->config.layout.hasAuthRegion;
+ Fixup::Cache64::setLocation(this->config.layout.cacheBaseAddress, &newThunk.func,
getAddrVMAddr, high8,
authDiversity, authHasAddrDiv, authKey, isAuth);
// Add to ASLR tracker
- segment.tracker.add(&tlvThunk.thunk);
+ segment.tracker.add(&newThunk.func);
}
- } else {
- struct tlv_thunk_t
+ }
+ else {
+ struct tlv_disk_thunk32_t
{
- uint32_t thunk;
+ uint32_t func;
uint32_t key;
uint32_t offset;
};
-
- uint32_t count = (uint32_t)(sectInfo.sectSize / sizeof(tlv_thunk_t));
- tlv_thunk_t* thunkBuffer = (tlv_thunk_t*)sectionBuffer;
- for ( uint32_t i = 0; i != count; ++i ) {
- tlv_thunk_t& tlvThunk = thunkBuffer[i];
-
- // Set the key to the next available key
- tlvThunk.key = key;
+ struct tlv_cache_thunk32_t
+ {
+ uint32_t func;
+ uint16_t key;
+ uint16_t offset;
+ int32_t machHeaderDelta; // if < 0, content is found by walking load commands. If > 0, then it is size and content is all zeros
+ };
+ static_assert(sizeof(tlv_disk_thunk32_t) == sizeof(tlv_cache_thunk32_t));
+ uint32_t loadAddr = (uint32_t)(cacheDylib.cacheHdr->preferredLoadAddress());
+ uint32_t count = (uint32_t)(sectInfo.size / sizeof(tlv_disk_thunk32_t));
+ tlv_disk_thunk32_t* thunkBuffer = (tlv_disk_thunk32_t*)sectionBuffer;
+ for ( uint32_t i = 0; i < count; ++i ) {
+ uint32_t offset = thunkBuffer[i].offset; // only value needed from disk format
+ tlv_cache_thunk32_t& newThunk = *((tlv_cache_thunk32_t*)&thunkBuffer[i]);
+
+ if ( offset > 0xFFFF )
+ this->warning("thread-local too large (%u max 65535) in %s", offset, cacheDylib.inputFile->path.c_str());
+
+ // Reset the fields using TLV_Thunkv2 format
+ newThunk.key = (uint16_t)key;
+ newThunk.offset = (uint16_t)offset; // FIXME: error if offset > 0xFFFF
+ newThunk.machHeaderDelta = (uint32_t)initialContentSize;
+ if ( initialContentNonZero ) {
+ uint64_t newThunkDeltaFieldAddr = (sectInfo.address + i*sizeof(tlv_disk_thunk32_t) + offsetof(tlv_cache_thunk32_t,machHeaderDelta));
+ newThunk.machHeaderDelta = (int32_t)(loadAddr - newThunkDeltaFieldAddr);
+ }
// Set the thunk to tlv_get_addr()
- Fixup::Cache32::setLocation(this->config.layout.cacheBaseAddress, &tlvThunk.thunk,
- getAddrVMAddr);
+ Fixup::Cache32::setLocation(this->config.layout.cacheBaseAddress, &newThunk.func, getAddrVMAddr);
// Add to ASLR tracker
- segment.tracker.add(&tlvThunk.thunk);
+ segment.tracker.add(&newThunk.func);
}
}
});
@@ -6329,22 +7339,36 @@
if ( (dataConstRegion == nullptr) && (authConstRegion == nullptr) )
continue;
- for ( bool auth : { false, true } ) {
- if ( auth && (authConstRegion == nullptr) )
+ for ( UniquedGOTKind sectionKind : { UniquedGOTKind::regular, UniquedGOTKind::authGot, UniquedGOTKind::authPtr } ) {
+
+ Region* region = nullptr;
+ CoalescedGOTSection* subCacheUniquedGOTs = nullptr;
+
+ // Skip sections if their segment doesn't exist
+ switch ( sectionKind ) {
+ case UniquedGOTKind::regular:
+ region = dataConstRegion;
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.regularGOTs;
+ break;
+ case UniquedGOTKind::authGot:
+ region = authConstRegion;
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.authGOTs;
+ break;
+ case UniquedGOTKind::authPtr:
+ region = authConstRegion;
+ subCacheUniquedGOTs = &subCache.uniquedGOTsOptimizer.authPtrs;
+ break;
+ }
+
+ if ( region == nullptr )
continue;
- if ( !auth && (dataConstRegion == nullptr) )
+
+ if ( subCacheUniquedGOTs->empty() )
continue;
-
- Region& region = auth ? *authConstRegion : *dataConstRegion;
- CoalescedGOTSection& subCacheUniquedGOTs = auth ? subCache.uniquedGOTsOptimizer.authGOTs : subCache.uniquedGOTsOptimizer.regularGOTs;
- if ( subCacheUniquedGOTs.cacheChunk == nullptr )
- continue;
-
- UniquedGOTsChunk* subCacheGOTChunk = subCacheUniquedGOTs.cacheChunk->isUniquedGOTsChunk();
std::set<const void*> seenFixups;
std::vector<PatchInfo::GOTInfo> gots;
- for ( const Chunk* chunk : region.chunks ) {
+ for ( const Chunk* chunk : region->chunks ) {
const DylibSegmentChunk* segmentChunk = chunk->isDylibSegmentChunk();
if ( !segmentChunk )
continue;
@@ -6355,46 +7379,55 @@
// Walk all the binds in this dylib, looking for GOT uses of the bind
assert(cacheDylib->bindTargets.size() == dylibPatchInfo.bindGOTUses.size());
assert(cacheDylib->bindTargets.size() == dylibPatchInfo.bindAuthGOTUses.size());
+ assert(cacheDylib->bindTargets.size() == dylibPatchInfo.bindAuthPtrUses.size());
for ( uint32_t bindIndex = 0; bindIndex != cacheDylib->bindTargets.size(); ++bindIndex ) {
const CacheDylib::BindTarget& bindTarget = cacheDylib->bindTargets[bindIndex];
- std::vector<PatchInfo::GOTInfo>* bindUses = nullptr;
- if ( auth ) {
- bindUses = &dylibPatchInfo.bindAuthGOTUses[bindIndex];
- } else {
- bindUses = &dylibPatchInfo.bindGOTUses[bindIndex];
+ std::span<PatchInfo::GOTInfo> bindUses;
+ switch ( sectionKind ) {
+ case UniquedGOTKind::regular:
+ bindUses = dylibPatchInfo.bindGOTUses[bindIndex];
+ break;
+ case UniquedGOTKind::authGot:
+ bindUses = dylibPatchInfo.bindAuthGOTUses[bindIndex];
+ break;
+ case UniquedGOTKind::authPtr:
+ bindUses = dylibPatchInfo.bindAuthPtrUses[bindIndex];
+ break;
}
+
+ if ( bindUses.empty() )
+ continue;
// For absolute binds, just set the pointers and move on
if ( bindTarget.kind == CacheDylib::BindTarget::Kind::absolute ) {
- for ( const PatchInfo::GOTInfo& got : *bindUses ) {
- CacheVMAddress gotVMAddr = got.patchInfo.cacheVMAddr;
- assert(gotVMAddr >= subCacheGOTChunk->cacheVMAddress);
- assert(gotVMAddr < (subCacheGOTChunk->cacheVMAddress + subCacheGOTChunk->cacheVMSize));
- VMOffset cacheSectionVMOffset = gotVMAddr - subCacheGOTChunk->cacheVMAddress;
-
- const void* fixupLoc = subCacheGOTChunk->subCacheBuffer + cacheSectionVMOffset.rawValue();
+ for ( const PatchInfo::GOTInfo& got : bindUses ) {
+ CacheVMAddress gotVMAddr = got.useLocation.clientGOT->cacheVMAddress + got.useLocation.clientGOTOffset;
+ const void* fixupLoc = subCacheUniquedGOTs->gotLocation(gotVMAddr);
+ uint64_t targetValue = std::get<uint64_t>(got.targetValue);
if ( this->config.layout.is64 ) {
- *(uint64_t*)fixupLoc = got.targetValue.rawValue();
+ *(uint64_t*)fixupLoc = targetValue;
} else {
- *(uint32_t*)fixupLoc = (uint32_t)got.targetValue.rawValue();
+ *(uint32_t*)fixupLoc = (uint32_t)targetValue;
}
}
continue;
}
- gots.insert(gots.end(), bindUses->begin(), bindUses->end());
+ assert(bindTarget.kind == CacheDylib::BindTarget::Kind::cacheImage);
+ gots.insert(gots.end(), bindUses.begin(), bindUses.end());
}
}
// Found all the GOTs/authGOTS for this subCache. Now we need to emit them
for ( const PatchInfo::GOTInfo& got : gots ) {
- CacheVMAddress gotVMAddr = got.patchInfo.cacheVMAddr;
- assert(gotVMAddr >= subCacheGOTChunk->cacheVMAddress);
- assert(gotVMAddr < (subCacheGOTChunk->cacheVMAddress + subCacheGOTChunk->cacheVMSize));
- VMOffset cacheSectionVMOffset = gotVMAddr - subCacheGOTChunk->cacheVMAddress;
-
- void* rawFixupLoc = subCacheGOTChunk->subCacheBuffer + cacheSectionVMOffset.rawValue();
+ CacheVMAddress gotVMAddr = got.useLocation.clientGOT->cacheVMAddress + got.useLocation.clientGOTOffset;
+
+ // Function variants don't get emitted right now
+ if ( !subCacheUniquedGOTs->shouldEmitGOT(gotVMAddr) )
+ continue;
+
+ void* rawFixupLoc = subCacheUniquedGOTs->gotLocation(gotVMAddr);
// Ignore dupes
if ( seenFixups.count(rawFixupLoc) )
@@ -6402,11 +7435,13 @@
seenFixups.insert(rawFixupLoc);
- CacheVMAddress targetVMAddr = config.layout.cacheBaseAddress + got.targetValue;
+ const DylibOffset& dylibOffset = std::get<DylibOffset>(got.targetValue);
+ InputDylibVMAddress inputVMAddr = dylibOffset.cacheDylib->inputLoadAddress + dylibOffset.vmOffset;
+ CacheVMAddress targetVMAddr = dylibOffset.cacheDylib->adjustor->adjustVMAddr(inputVMAddr);
if ( this->config.layout.is64 ) {
uint64_t high8 = 0;
uint64_t finalVMAddr = targetVMAddr.rawValue();
- if ( !got.patchInfo.authenticated ) {
+ if ( !got.useLocation.authenticated ) {
high8 = (finalVMAddr >> 56);
if ( high8 != 0 ) {
// Remove high8 from the vmAddr
@@ -6417,16 +7452,16 @@
rawFixupLoc,
CacheVMAddress(finalVMAddr),
high8,
- got.patchInfo.discriminator,
- got.patchInfo.usesAddressDiversity, got.patchInfo.key,
- got.patchInfo.authenticated);
+ got.useLocation.discriminator,
+ got.useLocation.usesAddressDiversity, got.useLocation.key,
+ got.useLocation.authenticated);
} else {
Fixup::Cache32::setLocation(this->config.layout.cacheBaseAddress,
rawFixupLoc,
CacheVMAddress(targetVMAddr));
}
- subCacheGOTChunk->tracker.add(rawFixupLoc);
+ subCacheUniquedGOTs->trackFixup(rawFixupLoc);
}
}
}
@@ -6483,7 +7518,7 @@
objc_visitor::Visitor& objcVisitor = objcVisitors[dylibObjCIndex];
if ( log ) {
- printf(" at 0x%llx in %s\n", protocolVMAddr, objcVisitor.mf()->installName());
+ printf(" at 0x%llx in %s\n", protocolVMAddr, objcVisitor.hdr()->installName());
}
__block bool foundProtocol = false;
@@ -6641,7 +7676,7 @@
error = Error("Superclass of class '%s' is weak-import"
"and missing. Referenced in %s",
objcClass.getName(objcVisitor),
- objcVisitor.mf()->installName());
+ objcVisitor.hdr()->installName());
stopClass = true;
this->objcOptimizer.foundMissingWeakSuperclass = true;
}
@@ -6674,7 +7709,7 @@
classMap[classVMAddr] = (uint32_t)classInfos.size();
if ( log ) {
- printf("%s: [0x%08llx] %s%s\n", objcVisitor.mf()->installName(), classVMAddr.rawValue(),
+ printf("%s: [0x%08llx] %s%s\n", objcVisitor.hdr()->installName(), classVMAddr.rawValue(),
objcClass.getName(objcVisitor), objcClass.isMetaClass ? " (meta)" : "");
}
@@ -6729,7 +7764,7 @@
worklist.insert(worklist.end(), classInfo->subClasses.begin(), classInfo->subClasses.end());
bool elidedSomething = false;
const objc_visitor::Class& objcClass = classInfo->classPos;
-
+ const bool isSwiftClass = objcClass.isSwift(*classInfo->objcVisitor);
auto& map = objcClass.isMetaClass ? metaclassMap : classMap;
std::optional<VMAddress> superclassVMAddr = objcClass.getSuperclassVMAddr(*classInfo->objcVisitor);
@@ -6770,7 +7805,7 @@
continue;
// skip ivars that swiftc has optimized away
- if ( ivar.elided(*classInfo->objcVisitor) ) {
+ if ( isSwiftClass && ivar.elided(*classInfo->objcVisitor) ) {
if ( log ) {
if ( !elidedSomething )
printf("adjusting ivars for %s\n", objcClass.getName(*classInfo->objcVisitor));
@@ -7160,7 +8195,7 @@
size_t objcIndex = 0;
for (size_t cacheIndex = 0; cacheIndex < this->cacheDylibs.size(); cacheIndex++) {
CacheDylib& cacheDylib = this->cacheDylibs[cacheIndex];
- if ( !cacheDylib.inputMF->hasObjC() )
+ if ( !cacheDylib.inputHdr->hasObjC() )
continue;
this->objcCategoryOptimizer.preAttachedDylibs.insert(objcIndex);
@@ -7217,13 +8252,65 @@
this->objcOptimizer.headerInfoReadOnlyChunk->subCacheBuffer,
this->objcOptimizer.headerInfoReadWriteChunk->subCacheBuffer,
this->objcOptimizer.headerInfoReadOnlyChunk->cacheVMAddress,
- this->swiftProtocolConformanceOptimizer);
-
+ swiftPrespecializedDylib,
+ this->swiftOptimizer);
if ( diag.hasError() )
return Error("Couldn't build Swift protocol opts because: %s", diag.errorMessageCStr());
return Error();
}
+
+static void getFunctionVariantTable(const CacheDylib& cacheDylib, uint64_t& vmAddr, uint64_t& vmSize)
+{
+ vmAddr = 0;
+ vmSize = 0;
+ for ( const LinkeditDataChunk& chunk : cacheDylib.linkeditChunks ) {
+ if ( chunk.isFunctionVariantsTable() ) {
+ vmAddr = chunk.cacheVMAddress.rawValue();
+ vmSize = chunk.cacheVMSize.rawValue();
+ }
+ }
+}
+
+void SharedCacheBuilder::emitFunctionVariants()
+{
+ // find uniqued GOT slots in each subcache that point to function variants
+ for ( SubCache& subCache : this->subCaches ) {
+ subCache.uniquedGOTsOptimizer.forEachFunctionVariant(^(const CoalescedGOTSection::FunctionVariantInfo& fv, uint64_t gotVMAddr,
+ dyld3::MachOFile::PointerMetaData pmd) {
+ uint32_t targetDylibIndex = fv.dylibIndex;
+ uint64_t fvTableVmAddr;
+ uint64_t fvTableVmSize;
+ getFunctionVariantTable(this->cacheDylibs[targetDylibIndex], fvTableVmAddr, fvTableVmSize);
+ dyld_cache_function_variant_entry entry;
+ entry.fixupLocVmAddr = gotVMAddr;
+ entry.functionVariantTableVmAddr = fvTableVmAddr;
+ entry.functionVariantTableSizeDiv4 = fvTableVmSize/4;
+ entry.dylibHeaderVmAddr = this->cacheDylibs[targetDylibIndex].cacheLoadAddress.rawValue();
+ entry.variantIndex = fv.variantIndex;
+ entry.pacAuth = pmd.authenticated;
+ entry.pacAddress = pmd.usesAddrDiversity;
+ entry.pacKey = pmd.key;
+ entry.pacDiversity = pmd.diversity;
+ entry.targetDylibIndex = targetDylibIndex;
+ assert(entry.variantIndex == fv.variantIndex);
+ this->functionVariantsOptimizer.infos.push_back(entry);
+ });
+ }
+
+ // make sure it fits in estimated size
+ size_t actualSize = offsetof(dyld_cache_function_variant_info, entries[this->functionVariantsOptimizer.infos.size()]);
+ assert(actualSize < this->functionVariantsOptimizer.fvInfoTotalByteSize);
+ this->functionVariantsOptimizer.fvInfoTotalByteSize = actualSize;
+
+ // write table to LINKEDIT
+ dyld_cache_function_variant_info* table = (dyld_cache_function_variant_info*)(this->functionVariantsOptimizer.chunk->subCacheBuffer);
+ table->version = 1;
+ table->count = (uint32_t)this->functionVariantsOptimizer.infos.size();
+ if ( table->count > 0 )
+ memcpy(table->entries, &this->functionVariantsOptimizer.infos[0], sizeof(dyld_cache_function_variant_entry)*this->functionVariantsOptimizer.infos.size());
+}
+
void SharedCacheBuilder::emitCacheDylibsTrie()
{
@@ -7275,11 +8362,14 @@
switch ( region.kind ) {
case Region::Kind::text:
case Region::Kind::dynamicConfig:
+ case Region::Kind::readOnly:
case Region::Kind::linkedit:
maxSlide = std::min(maxSlide, subCacheLimit - region.subCacheVMSize);
break;
+ case Region::Kind::tproConst:
case Region::Kind::data:
case Region::Kind::dataConst:
+ case Region::Kind::tproAuthConst:
case Region::Kind::auth:
case Region::Kind::authConst:
if ( firstDataRegion == nullptr )
@@ -7304,6 +8394,11 @@
// We must be a largeContiguous cache. Others were dealt with above in the x86_64 and/or sim cases
assert(this->config.layout.contiguous.has_value());
+ // Some caches have a fixed max slide
+ if ( this->config.layout.cacheFixedSlide.has_value() ) {
+ return this->config.layout.cacheFixedSlide.value();
+ }
+
// Start off making sure we can't slide past the end of the cache
CacheVMAddress maxVMAddress(0ULL);
for ( const Region& region : this->subCaches.back().regions ) {
@@ -7359,30 +8454,223 @@
}
}
+Error SharedCacheBuilder::patchLinkedDylibs(CacheDylib& cacheDylib)
+{
+ if ( swiftPrespecializedDylib == nullptr )
+ return Error::none();
+
+ Diagnostics diag;
+ if ( &cacheDylib == swiftPrespecializedDylib ) {
+ // remove all but libSystem
+ cacheDylib.removeLinkedDylibs(diag);
+ } else if ( cacheDylib.installName.find("libswiftCore.dylib") != std::string_view::npos ) {
+ // add Swift prespecialized dylib dependency to libswiftCore
+ cacheDylib.addLinkedDylib(diag, *swiftPrespecializedDylib);
+ }
+
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ return Error::none();
+}
+
+void SharedCacheBuilder::addSubCacheFileInfo(uint64_t cacheVMAddress, PropertyList::Array &files, const SubCache &subCache) {
+ using Array = PropertyList::Array;
+ using Integer = PropertyList::Integer;
+ using String = PropertyList::String;
+ using Dictionary = PropertyList::Dictionary;
+
+ uuid_t subcacheUUID;
+ uuid_parse(subCache.uuidString, subcacheUUID);
+ auto& subCacheFile = files.addObject<Dictionary>();
+
+ subCacheFile.addObjectForKey<String>("name", options.mainCacheFileName + subCache.fileSuffix);
+ subCacheFile.addObjectForKey<PropertyList::UUID>(kDyldAtlasSharedCacheUUIDKey,subcacheUUID);
+ subCacheFile.addObjectForKey<Integer>("voff",subCache.subCacheVMAddress.rawValue()-cacheVMAddress);
+ subCacheFile.addObjectForKey<Integer>("fsze", subCache.bufferSize);
+ subCacheFile.addObjectForKey<Integer>("padr", subCache.subCacheVMAddress.rawValue());
+ auto& mappingsArray = subCacheFile.addObjectForKey<Array>(kDyldAtlasSharedCacheMappingArrayKey);
+
+ dyld_cache_header* subCacheHeader = (dyld_cache_header*)subCache.buffer;
+ auto* mappings = (dyld_cache_mapping_info*)((uint8_t*)subCacheHeader + subCacheHeader->mappingOffset);
+
+ uint64_t lastAddress = 0;
+ for ( auto i = 0; i < subCacheHeader->mappingCount; ++i) {
+ auto& mapping = mappingsArray.addObject<Dictionary>();
+ mapping.addObjectForKey<Integer>(kDyldAtlasSharedCacheMappingsSizeKey, mappings[i].size);
+ mapping.addObjectForKey<Integer>(kDyldAtlasSharedCacheMappingsPreferredLoadAddressKey, mappings[i].address);
+ mapping.addObjectForKey<Integer>(kDyldAtlasSharedCacheMappingsFileOffsetKey, mappings[i].fileOffset);
+ mapping.addObjectForKey<Integer>(kDyldAtlasSharedCacheMappingsMaxProtKey, mappings[i].maxProt);
+ if (mappings[i].address + mappings[i].size > lastAddress) {
+ lastAddress = mappings[i].address + mappings[i].size;
+ }
+ }
+ subCacheFile.addObjectForKey<Integer>("size", lastAddress-subCacheHeader->sharedRegionStart);
+}
+
+void SharedCacheBuilder::addCacheAtlasInfo(PropertyList::Dictionary *cacheAtlas, const SubCache &mainCache) {
+ using Array = PropertyList::Array;
+ using Integer = PropertyList::Integer;
+ using String = PropertyList::String;
+
+ uint64_t cacheVMAddress = mainCache.subCacheVMAddress.rawValue();
+ uuid_t cacheUUID;
+ uuid_parse(mainCache.uuidString, cacheUUID);
+ cacheAtlas->addObjectForKey<PropertyList::UUID>(kDyldAtlasSharedCacheUUIDKey, cacheUUID);
+ cacheAtlas->addObjectForKey<Integer>(kDyldAtlasSharedCachePreferredLoadAddressKey, mainCache.subCacheVMAddress.rawValue());
+ cacheAtlas->addObjectForKey<Integer>(kDyldAtlasSharedCacheVMSizeKey, this->totalVMSize.rawValue());
+ dyld_cache_header* cacheHeader = (dyld_cache_header*)mainCache.buffer;
+
+ if (!uuid_is_null(cacheHeader->symbolFileUUID)) {
+ cacheAtlas->addObjectForKey<String>(kDyldAtlasSharedCacheSymbolFileName, options.mainCacheFileName + ".symbols");
+ cacheAtlas->addObjectForKey<PropertyList::UUID>(kDyldAtlasSharedCacheSymbolFileName, cacheHeader->symbolFileUUID);
+ }
+
+ auto& files = cacheAtlas->addObjectForKey<Array>("dscs");
+ addSubCacheFileInfo(cacheVMAddress, files, mainCache);
+ for (auto& subCache : mainCache.subCaches) {
+ addSubCacheFileInfo(cacheVMAddress, files, *subCache);
+ }
+
+ // FIXME: Remove once deprecate legacy APIs that need this
+ if (strncmp(cacheHeader->magic, "dyld_v1arm64_32", 15) != 0) {
+ cacheAtlas->addObjectForKey<Integer>("psze", 4);
+ } else {
+ cacheAtlas->addObjectForKey<Integer>("psze", 8);
+ }
+}
+
+// This generates the atlas binary plist included with the shared cache
+void SharedCacheBuilder::buildAtlas() {
+ STACK_ALLOCATOR(allocator, 0);
+ using Array = PropertyList::Array;
+ using Dictionary = PropertyList::Dictionary;
+ using Integer = PropertyList::Integer;
+ using String = PropertyList::String;
+ auto propertyListEncoder = PropertyList(allocator);
+ auto& rootDictionary = propertyListEncoder.rootDictionary();
+ // The same plist contains both the customer and shared cache data, since they share layouts
+ // We include dictionaries at the root so they can be lookup by leaf name or UUID
+ auto& byUuidDictionary = rootDictionary.addObjectForKey<Dictionary>("uuids");
+ auto& byNameDictionary = rootDictionary.addObjectForKey<Dictionary>("names");
+ Dictionary* customerCacheAtlas = nullptr;
+ Dictionary* devCacheAtlas = nullptr;
+ Array* images = nullptr;
+ uuid_string_t customerCacheUUID = {0};
+ uuid_string_t devCacheUUID = {0};
+ std::string customerCacheName;
+ std::string devCacheName;
+
+ for ( const SubCache& subCache : this->subCaches ) {
+ if ( subCache.isMainCustomerCache() ) {
+ customerCacheName = options.mainCacheFileName + subCache.fileSuffix;
+ customerCacheAtlas = &byUuidDictionary.addObjectForKey<Dictionary>(subCache.uuidString);
+ byNameDictionary.insertObjectForKey(customerCacheName, *customerCacheAtlas);
+ addCacheAtlasInfo(customerCacheAtlas, subCache);
+ strcpy(customerCacheUUID, subCache.uuidString);
+ } else if ( subCache.isMainDevelopmentCache() ) {
+ devCacheName = options.mainCacheFileName + subCache.fileSuffix;
+ devCacheAtlas = &byUuidDictionary.addObjectForKey<Dictionary>(subCache.uuidString);
+ byNameDictionary.insertObjectForKey(devCacheName, *devCacheAtlas);
+ addCacheAtlasInfo(devCacheAtlas, subCache);
+ strcpy(devCacheUUID, subCache.uuidString);
+ }
+ }
+ assert((customerCacheAtlas != nullptr) || (devCacheAtlas != nullptr));
+
+ // The bplist00 format supports uniquing objects, but uniquing collections is slow. Since we know a priori the images for both cache variants
+ // will be the same we create it int he first cache atlas, and if there is a second we insert a reference to the already existing one.
+ if (customerCacheAtlas) {
+ images = &customerCacheAtlas->addObjectForKey<Array>(kDyldAtlasSharedCacheImageArrayKey);
+ if (devCacheAtlas) {
+ devCacheAtlas->insertObjectForKey(kDyldAtlasSharedCacheImageArrayKey, *images);
+ }
+ } else {
+ images = &devCacheAtlas->addObjectForKey<Array>(kDyldAtlasSharedCacheImageArrayKey);
+ }
+
+ for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
+ auto& image = images->addObject<Dictionary>();
+ auto& segments = image.addObjectForKey<Array>(kDyldAtlasImageSegmentArrayKey);
+ image.addObjectForKey<String>(kDyldAtlasImageInstallnameKey, cacheDylib.installName);
+ image.addObjectForKey<Integer>(kDyldAtlasImagePreferredLoadAddressKey, cacheDylib.cacheLoadAddress.rawValue());
+ uuid_t uuid;
+ assert(cacheDylib.cacheHdr->getUuid(uuid));
+ image.addObjectForKey<PropertyList::UUID>(kDyldAtlasImageUUIDKey, uuid);
+
+ for ( const DylibSegmentChunk& segmentChunk : cacheDylib.segments ) {
+ auto& segment = segments.addObject<Dictionary>();
+ segment.addObjectForKey<String>(kDyldAtlasSegmentNameKey, segmentChunk.name());
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentPreferredLoadAddressKey, segmentChunk.cacheVMAddress.rawValue());
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentSizeKey, segmentChunk.cacheVMSize.rawValue());
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentFileOffsetKey, segmentChunk.subCacheFileOffset.rawValue());
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentFileSizeKey, segmentChunk.subCacheFileSize.rawValue());
+ if ( strcmp("__TEXT", segmentChunk.name()) == 0 ) {
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentPermissionsKey, VM_PROT_READ | VM_PROT_EXECUTE);
+ } else if ( strcmp("__LINKEDIT", segmentChunk.name()) == 0 ) {
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentPermissionsKey, VM_PROT_READ);
+ } else {
+ segment.addObjectForKey<Integer>(kDyldAtlasSegmentPermissionsKey, VM_PROT_READ | VM_PROT_WRITE);
+ }
+ }
+ }
+
+ ByteStream fileStream(allocator);
+ propertyListEncoder.encode(fileStream);
+ AAREncoder aarEncoder(allocator);
+
+ if (customerCacheUUID[0]) {
+ std::string plistPath = std::string("caches/uuids/") + customerCacheUUID + ".plist";
+ std::string symlinkTarget = std::string("../uuids/") + customerCacheUUID + ".plist";
+ std::string symlinkSource = std::string("caches/names/") + customerCacheName + ".plist";
+
+ aarEncoder.addFile(plistPath, fileStream.span());
+ aarEncoder.addSymLink(symlinkSource, symlinkTarget);
+ if (devCacheUUID[0]) {
+ std::string devPlistPath = std::string("caches/uuids/") + devCacheUUID + ".plist";
+ std::string devSymlinkSource = std::string("caches/names/") + devCacheName + ".plist";
+ aarEncoder.addSymLink(devPlistPath, symlinkTarget);
+ aarEncoder.addSymLink(devSymlinkSource, symlinkTarget);
+ }
+ } else {
+ assert(devCacheUUID[0] != 0);
+ std::string plistPath = std::string("caches/uuids/") + devCacheUUID + ".plist";
+ std::string symlinkTarget = std::string("../uuids/") + devCacheUUID + ".plist";
+ std::string symlinkSource = std::string("caches/names/") + devCacheName + ".plist";
+
+ aarEncoder.addFile(plistPath, fileStream.span());
+ aarEncoder.addSymLink(symlinkSource, symlinkTarget);
+ }
+
+ ByteStream outputStream(allocator);
+ aarEncoder.encode(outputStream);
+ std::copy(outputStream.begin(), outputStream.end(), std::back_insert_iterator(atlasData));
+}
+
void SharedCacheBuilder::computeCacheHeaders()
{
Timer::Scope timedScope(this->config, "computeCacheHeaders time");
- for ( SubCache& subCache : this->subCaches )
- subCache.writeCacheHeader(this->options, this->config, this->cacheDylibs);
-
- // Content for the first (main) subCache only
+ // Content for all subcaches
__block uint32_t osVersion = 0;
__block uint32_t altPlatform = 0;
__block uint32_t altOsVersion = 0;
+
+ // Content for the first (main) subCache only
CacheVMAddress dyldInCacheUnslidAddr = CacheVMAddress(0ULL);
CacheVMAddress dyldInCacheEntryUnslidAddr = CacheVMAddress(0ULL);
+
{
// look for libdyld.dylib and record OS verson info into cache header
for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
if ( endsWith(cacheDylib.installName, "/libdyld.dylib") ) {
- cacheDylib.inputMF->forEachSupportedPlatform(^(dyld3::Platform platform, uint32_t minOS, uint32_t sdk) {
- if ( platform == options.platform ) {
- osVersion = minOS;
+ cacheDylib.inputHdr->platformAndVersions().unzip(^(mach_o::PlatformAndVersions pvs) {
+ if ( pvs.platform == options.platform ) {
+ osVersion = pvs.minOS.value();
}
else {
- altPlatform = (uint32_t)platform;
- altOsVersion = minOS;
+ altPlatform = pvs.platform.value();
+ altOsVersion = pvs.minOS.value();
}
});
}
@@ -7391,13 +8679,17 @@
dyldInCacheUnslidAddr = cacheDylib.cacheLoadAddress;
uint64_t dyldEntryOffset;
bool usesCRT;
- if ( cacheDylib.cacheMF->getEntry(dyldEntryOffset, usesCRT) ) {
+ if ( cacheDylib.cacheHdr->getEntry(dyldEntryOffset, usesCRT) ) {
// the "pc" value in the LC_UNIXTHREAD was adjusted when dyld was placed in the cache
dyldInCacheEntryUnslidAddr = dyldInCacheUnslidAddr + VMOffset(dyldEntryOffset);
}
}
}
}
+
+ for ( SubCache& subCache : this->subCaches )
+ subCache.writeCacheHeader(this->options, this->config, this->cacheDylibs,
+ osVersion, altPlatform, altOsVersion);
// Fill in info for the main caches. This must be after addCacheHeaderImageInfo().
for ( SubCache& subCache : this->subCaches ) {
@@ -7405,11 +8697,11 @@
subCache.addMainCacheHeaderInfo(this->options, this->config,
this->cacheDylibs,
this->totalVMSize, getMaxSlide(),
- osVersion, altPlatform, altOsVersion,
dyldInCacheUnslidAddr, dyldInCacheEntryUnslidAddr,
this->dylibTrieOptimizer,
- this->objcOptimizer, this->swiftProtocolConformanceOptimizer,
- this->patchTableOptimizer, this->prebuiltLoaderBuilder);
+ this->objcOptimizer, this->swiftOptimizer,
+ this->patchTableOptimizer, this->functionVariantsOptimizer,
+ this->prebuiltLoaderBuilder, this->prewarmingOptimizer);
continue;
}
@@ -7492,7 +8784,12 @@
return this->evictedDylibs;
}
-void SharedCacheBuilder::getResults(std::vector<CacheBuffer>& results) const
+std::string_view SharedCacheBuilder::getSwiftPrespecializedDylibBuildError() const
+{
+ return swiftPrespecializedDylibBuildError;
+}
+
+void SharedCacheBuilder::getResults(std::vector<CacheBuffer>& results, std::vector<std::byte>& atlas) const
{
for ( const SubCache& subCache : this->subCaches ) {
CacheBuffer buffer;
@@ -7500,6 +8797,8 @@
buffer.bufferSize = subCache.bufferSize;
buffer.cdHash = cdHashToString(subCache.cdHash);
+ if ( this->config.codeSign.mode == CodeSign::Mode::agile )
+ buffer.agilecdHash = cdHashToString(subCache.agilecdHash);
buffer.uuid = subCache.uuidString;
buffer.cacheFileSuffix = subCache.fileSuffix;
@@ -7507,13 +8806,8 @@
buffer.usedByCustomerConfig = subCache.shouldKeepCache(false, true);
buffer.usedByDevelopmentConfig = subCache.shouldKeepCache(true, false);
- // The builder executable also passes back the fd. This should typically be used instead of the data buffer
-#if !SUPPORT_CACHE_BUILDER_MEMORY_BUFFERS
- buffer.fd = subCache.fd;
- buffer.tempPath = subCache.tempPath;
-#endif
-
results.push_back(std::move(buffer));
+ atlas = std::move(this->atlasData);
}
}
@@ -7529,12 +8823,15 @@
case Region::Kind::text:
prot = "EX";
break;
+ case Region::Kind::tproConst:
case Region::Kind::data:
case Region::Kind::dataConst:
+ case Region::Kind::tproAuthConst:
case Region::Kind::auth:
case Region::Kind::authConst:
prot = "RW";
break;
+ case Region::Kind::readOnly:
case Region::Kind::linkedit:
prot = "RO";
break;
@@ -7584,31 +8881,32 @@
assert(mainSubCache.isMainCache());
- dyld3::json::Node cacheNode;
+ json::Node cacheNode;
cacheNode.map["version"].value = "1";
cacheNode.map["disposition"].value = disposition;
- cacheNode.map["base-address"].value = dyld3::json::hex(baseAddress);
+ cacheNode.map["arch"].value = this->options.arch.name();
+ cacheNode.map["platform"].value = mach_o::Platform(this->options.platform).name();
+ cacheNode.map["base-address"].value = json::hex(baseAddress);
cacheNode.map["uuid"].value = mainSubCache.uuidString;
- dyld3::json::Node imagesNode;
+ json::Node imagesNode;
for ( const CacheDylib& cacheDylib : this->cacheDylibs ) {
- dyld3::json::Node imageNode;
+ json::Node imageNode;
imageNode.map["path"].value = cacheDylib.installName;
- const dyld3::MachOFile* mf = cacheDylib.cacheMF;
uuid_t uuid;
- if ( mf->getUuid(uuid) ) {
+ if ( cacheDylib.cacheHdr->getUuid(uuid) ) {
uuid_string_t uuidStr;
uuid_unparse(uuid, uuidStr);
imageNode.map["uuid"].value = uuidStr;
}
- __block dyld3::json::Node segmentsNode;
- mf->forEachSegment(^(const dyld3::MachOAnalyzer::SegmentInfo &info, bool &stop) {
- dyld3::json::Node segmentNode;
- segmentNode.map["name"].value = info.segName;
- segmentNode.map["start-vmaddr"].value = dyld3::json::hex(info.vmAddr);
- segmentNode.map["end-vmaddr"].value = dyld3::json::hex(info.vmAddr + info.vmSize);
+ __block json::Node segmentsNode;
+ cacheDylib.cacheHdr->forEachSegment(^(const Header::SegmentInfo &info, bool &stop) {
+ json::Node segmentNode;
+ segmentNode.map["name"].value = info.segmentName;
+ segmentNode.map["start-vmaddr"].value = json::hex(info.vmaddr);
+ segmentNode.map["end-vmaddr"].value = json::hex(info.vmaddr + info.vmsize);
// Add sections in verbose mode
segmentsNode.array.push_back(segmentNode);
@@ -7689,9 +8987,9 @@
{
va_list list;
va_start(list, format);
- void* buffer = _simple_salloc();
+ _SIMPLE_STRING buffer = _simple_salloc();
_simple_vsprintf(buffer, format, list);
- this->warnings.push_back((const char*)buffer);
+ this->warnings.push_back(_simple_string(buffer));
_simple_sfree(buffer);
va_end(list);
@@ -7715,3 +9013,219 @@
fprintf(stderr, "Didn't find a dylib with install name: %s\n", installName);
}
+
+// This name is used only to create a placeholder input file and determine the library order.
+const std::string_view swiftPrespecializedDylibInstallName = "/usr/lib/libswiftPrespecialized.dylib";
+
+bool SharedCacheBuilder::shouldBuildSwiftPrespecializedDylib()
+{
+ if ( options.platform == Platform::driverKit )
+ return false;
+
+ if ( options.platform.isExclaveKit() )
+ return false;
+
+ // build the dylib, only if the order file is defined
+ if ( options.swiftGenericMetadataFile.empty() )
+ return false;
+
+ // check if the metadata builder is available
+#if !BUILDING_CACHE_BUILDER_UNIT_TESTS && !BUILDING_SIM_CACHE_BUILDER
+ if ( swift_externalMetadataBuilder_create == nullptr )
+ return false;
+#endif // !BUILDING_CACHE_BUILDER_UNIT_TESTS
+
+ return true;
+}
+
+Error SharedCacheBuilder::buildSwiftPrespecializedDylibJSON()
+{
+#if !BUILDING_CACHE_BUILDER_UNIT_TESTS && !BUILDING_SIM_CACHE_BUILDER
+ Timer::Scope timedScope(this->config, "buildSwiftPrespecializedDylibJSON time");
+
+ SwiftExternalMetadataBuilder* builder = swift_externalMetadataBuilder_create(options.platform.value(), options.arch.name());
+ if ( !builder )
+ return Error("swift_externalMetadataBuilder_create failed");
+
+ for ( const CacheDylib& dylib : this->cacheDylibs ) {
+ if ( dylib.inputHdr == nullptr ) continue;
+
+ // TODO: rdar://132262275 (dyld shared cache builder should tell Swift Metadata builder also about dyld)
+ if ( dylib.inputHdr->isDylinker() ) continue;
+
+ if ( const char* err = swift_externalMetadataBuilder_addDylib(builder, dylib.inputHdr->installName(),
+ (const struct mach_header*)dylib.inputHdr, dylib.inputFile->size) )
+ return Error("swift_externalMetadataBuilder_addDylib failed: %s", err);
+ }
+
+ if ( const char* err = swift_externalMetadataBuilder_readNamesJSON(builder, options.swiftGenericMetadataFile.c_str()) )
+ return Error("swift_externalMetadataBuilder_readNamesJSON failed: %s", err);
+
+ if ( const char* err = swift_externalMetadataBuilder_buildMetadata(builder) )
+ return Error("swift_externalMetadataBuilder_buildMetadata failed: %s", err);
+
+ if ( const char* json = swift_externalMetadataBuilder_getMetadataJSON(builder) )
+ swiftPrespecializedDylibJSON = json;
+ else
+ return Error("swift_externalMetadataBuilder_getMetadataJSON returned an empty JSON");
+
+ const std::string_view placeholderVersion = R"("platformVersion": "1.0")";
+ // Patch platformVersion if it's 1.0 until rdar://122585868 is fixed
+ if ( auto pos = swiftPrespecializedDylibJSON.find(placeholderVersion);
+ pos != swiftPrespecializedDylibJSON.npos ) {
+
+ __block mach_o::Version32 newMinOS;
+ // determine new deployment target based on dyld's version
+ for ( const InputFile& inputFile : allInputFiles ) {
+ if ( !inputFile.mf )
+ continue;
+
+ if ( !endsWith(inputFile.path, "dyld") )
+ continue;
+
+ ((Header*)inputFile.mf)->platformAndVersions().unzip(^(mach_o::PlatformAndVersions pvs) {
+ if ( pvs.platform == options.platform )
+ newMinOS = pvs.minOS;
+ });
+ break;
+ }
+
+ if ( newMinOS > mach_o::Version32(1, 0) ) {
+ char verStr[32];
+ newMinOS.toString(verStr);
+ std::string newVersion = "\"platformVersion\": \"";
+ newVersion += verStr;
+ newVersion += "\"";
+
+ swiftPrespecializedDylibJSON.replace(pos, placeholderVersion.size(), newVersion);
+ }
+ }
+
+ swift_externalMetadataBuilder_destroy(builder);
+
+ if ( options.debug ) {
+ std::string path;
+ if ( const char* dir = getenv("TMPDIR") )
+ path = dir;
+ if ( path.empty() )
+ path = "/tmp";
+ path += "/swift-prespecialized.json-XXXXXX";
+
+ int outFileFd = mkstemp(path.data());
+ if ( outFileFd != -1 ) {
+ write(outFileFd, swiftPrespecializedDylibJSON.data(), swiftPrespecializedDylibJSON.size());
+ }
+ }
+#endif // !BUILDING_CACHE_BUILDER_UNIT_TESTS
+
+ return Error::none();
+}
+
+bool SharedCacheBuilder::reserveSwiftPrespecializedInputFile()
+{
+ if ( !shouldBuildSwiftPrespecializedDylib() )
+ return false;
+
+ InputFile inputFile;
+ inputFile.mf = nullptr;
+ inputFile.inode = 0;
+ inputFile.mtime = 0;
+ inputFile.path = swiftPrespecializedDylibInstallName;
+ allInputFiles.push_back(std::move(inputFile));
+ cacheDylibs.push_back(CacheDylib(swiftPrespecializedDylibInstallName));
+ return true;
+}
+
+Error SharedCacheBuilder::createSwiftPrespecializedDylib()
+{
+ if ( !shouldBuildSwiftPrespecializedDylib() )
+ return Error::none();
+
+ if ( Error err = buildSwiftPrespecializedDylibJSON() )
+ return err;
+
+ Timer::Scope timedScope(this->config, "createSwiftPrespecializedDylib time");
+
+ InputFile* inputFile = nullptr;
+ if ( allInputFiles.empty() || allInputFiles.back().path != swiftPrespecializedDylibInstallName )
+ return Error("missing input file placeholder for Swift prespecialized dylib");
+ inputFile = &allInputFiles.back();
+
+ std::vector<const char*> dylibsList;
+ // the dylib list needs to be in order of objc dylibs
+ for ( const CacheDylib* dylib : this->objcOptimizer.objcDylibs )
+ dylibsList.push_back(CString::dup(dylib->installName).c_str());
+
+ // TODO: support in-memory file buffer
+ std::string path;
+ if ( const char* dir = getenv("TMPDIR") )
+ path = dir;
+ if ( path.empty() )
+ path = "/tmp";
+ path += "/libswiftPrespecialized.dylib-XXXXXX";
+
+ int outFileFd = mkstemp(path.data());
+ if ( outFileFd == -1 )
+ return Error("couldn't create a temporary file for Swift prespecialized dylib: %s", (const char*)strerror(errno));
+
+ close(outFileFd);
+ if ( const char* err = ldMakeDylibFromJSON(swiftPrespecializedDylibJSON, dylibsList, path.c_str()) )
+ return Error("%s", err);
+
+ // cleanup dylibs list
+ for ( const char* str : dylibsList )
+ free((void*)str);
+
+ // re-open output file
+ outFileFd = open(path.c_str(), O_RDONLY);
+ if ( outFileFd < 0 )
+ return Error("could not open swift dylib file because: %s", (const char*)strerror(errno));
+
+ struct stat stat_buf;
+ if ( fstat(outFileFd, &stat_buf) == -1 )
+ return Error("could not stat swift dylib file because: %s", (const char*)strerror(errno));
+
+ vm_size_t bufferSize = stat_buf.st_size;
+ void* buffer = mmap(nullptr, bufferSize, PROT_READ, MAP_FILE | MAP_SHARED, outFileFd, 0);
+ if ( buffer == MAP_FAILED ) {
+ // Failed to mmap the file
+ return Error("could not mmap swift dylib file because: %s", (const char*)strerror(errno));
+ }
+
+ Diagnostics diag;
+ uint64_t sliceOffset = 0;
+ inputFile->mf = MachOFile::compatibleSlice(diag, sliceOffset, inputFile->size, buffer, bufferSize, path.data(),
+ this->options.platform, /* isOSBinary */ false,
+ this->options.gradedArchs);
+ if ( diag.hasError() )
+ return Error("%s", diag.errorMessageCStr());
+
+ // recreate cache dylib at the reserved slot
+ auto cacheDylibIt = std::find_if(cacheDylibs.begin(), cacheDylibs.end(), [](CacheDylib& dylib) {
+ return dylib.inputHdr == nullptr && dylib.installName == swiftPrespecializedDylibInstallName;
+ });
+ if ( cacheDylibIt == cacheDylibs.end() )
+ return Error("missing cache dylib slot for Swift prespecialized dylib");
+
+ // save previously computed cache index
+ uint32_t index = cacheDylibIt->cacheIndex;
+ // recreate cache dylib with the updated input file
+ *cacheDylibIt = CacheDylib(*inputFile);
+ cacheDylibIt->cacheIndex = index;
+ // rdar://122906481 (Shared cache builder - explicitly model dylibs without a need for a patch table)
+ cacheDylibIt->needsPatchTable = false;
+ this->swiftPrespecializedDylib = &*cacheDylibIt;
+
+ // sanity check Swift dylib compatibility
+ __block Error err = Error::none();
+ inputFile->mf->withFileLayout(diag, ^(const mach_o::Layout& layout) {
+ mach_o::SplitSeg splitSeg(layout);
+
+ if ( !splitSeg.isV2() )
+ err = Error("Swift prespecialized dylib must use split seg V2");
+ });
+ if ( !inputFile->mf->hasChainedFixups() )
+ err = Error("Swift prespecialized dylib must use chained fixups");
+
+ return std::move(err);
+}