Loading...
launch-cache/update_dyld_shared_cache.cpp dyld-195.5 dyld-353.2.3
--- dyld/dyld-195.5/launch-cache/update_dyld_shared_cache.cpp
+++ dyld/dyld-353.2.3/launch-cache/update_dyld_shared_cache.cpp
@@ -31,7 +31,9 @@
 #include <stdarg.h>
 #include <stdio.h>
 #include <stdlib.h>
+#include <math.h>
 #include <fcntl.h>
+#include <dlfcn.h>
 #include <signal.h>
 #include <errno.h>
 #include <sys/uio.h>
@@ -44,25 +46,29 @@
 #include <servers/bootstrap.h>
 #include <mach-o/loader.h>
 #include <mach-o/fat.h>
+#include <CoreFoundation/CoreFoundation.h>
+#include <Security/Security.h>
+#include <Security/SecCodeSigner.h>
+#include <CommonCrypto/CommonDigest.h>
 
 #include "dyld_cache_format.h"
 
 #include <vector>
 #include <set>
 #include <map>
-#include <ext/hash_map>
+#include <unordered_map>
 
 #include "Architectures.hpp"
 #include "MachOLayout.hpp"
 #include "MachORebaser.hpp"
 #include "MachOBinder.hpp"
 #include "CacheFileAbstraction.hpp"
+#include "dyld_cache_config.h"
 
 #define SELOPT_WRITE
-#include <objc/objc-selopt.h>
-
-#define FIRST_DYLIB_TEXT_OFFSET 0x5000
-#define FIRST_DYLIB_DATA_OFFSET 0x1000
+#include "objc-shared-cache.h"
+
+#define FIRST_DYLIB_TEXT_OFFSET 0x8000
 
 #ifndef LC_FUNCTION_STARTS
     #define LC_FUNCTION_STARTS 0x26
@@ -95,21 +101,34 @@
 }
 
 
-static uint64_t pageAlign(uint64_t addr) { return ( (addr + 4095) & (-4096) ); }
+class CStringHash {
+public:
+	size_t operator()(const char* __s) const {
+		size_t __h = 0;
+		for ( ; *__s; ++__s)
+			__h = 5 * __h + *__s;
+		return __h;
+	};
+};
+class CStringEquals
+{
+public:
+	bool operator()(const char* left, const char* right) const { return (strcmp(left, right) == 0); }
+};
+
+
 
 class ArchGraph
 {
 public:
-	struct CStringEquals {
-		bool operator()(const char* left, const char* right) const { return (strcmp(left, right) == 0); }
-	};
-	typedef __gnu_cxx::hash_map<const char*, const char*, __gnu_cxx::hash<const char*>, CStringEquals> StringToString;
+	typedef std::unordered_map<const char*, const char*, CStringHash, CStringEquals> StringToString;
 
 	static void			addArchPair(ArchPair ap);
 	static void			addRoot(const char* vpath, const std::set<ArchPair>& archs);
 	static void			findSharedDylibs(ArchPair ap);
 	static ArchGraph*	graphForArchPair(ArchPair ap) { return fgPerArchGraph[ap]; }
-	static void			setFileSystemRoot(const char* root, bool usesOverlay) { fgFileSystemRoot = root; fgUsesOverlay = usesOverlay; }
+	static void			setFileSystemRoot(const char* root) { fgFileSystemRoot = root; }
+	static void			setFileSystemOverlay(const std::vector<const char*>& overlays);
 	static const char*	archName(ArchPair ap);
 	
 	ArchPair											getArchPair() { return fArchPair; }
@@ -139,7 +158,7 @@
 		std::set<DependencyNode*>					fRootsDependentOnThis;
 	};
 
-	typedef __gnu_cxx::hash_map<const char*, class DependencyNode*, __gnu_cxx::hash<const char*>, CStringEquals> PathToNode;
+	typedef std::unordered_map<const char*, class DependencyNode*, CStringHash, CStringEquals> PathToNode;
 
 
 								ArchGraph(ArchPair ap) : fArchPair(ap) {}
@@ -151,7 +170,7 @@
 
 	static std::map<ArchPair, ArchGraph*>	fgPerArchGraph;
 	static const char*						fgFileSystemRoot;
-	static bool								fgUsesOverlay;
+	static std::vector<const char*> 		fgFileSystemOverlays;
 	
 	ArchPair									fArchPair;
 	std::set<DependencyNode*>					fRoots;
@@ -161,7 +180,7 @@
 };
 std::map<ArchPair, ArchGraph*>		ArchGraph::fgPerArchGraph;
 const char*							ArchGraph::fgFileSystemRoot = "";
-bool								ArchGraph::fgUsesOverlay = false;
+std::vector<const char*> 			ArchGraph::fgFileSystemOverlays;
 
 void ArchGraph::addArchPair(ArchPair ap)
 {
@@ -169,30 +188,40 @@
 	fgPerArchGraph[ap] = new ArchGraph(ap);
 }
 
+void ArchGraph::setFileSystemOverlay(const std::vector<const char*>& overlays) 
+{ 
+	for (std::vector<const char*>::const_iterator it=overlays.begin(); it != overlays.end(); ++it) 
+		fgFileSystemOverlays.push_back(*it);
+}
+
 void ArchGraph::addRoot(const char* vpath, const std::set<ArchPair>& onlyArchs)
 {
-	char completePath[strlen(fgFileSystemRoot)+strlen(vpath)+2];
+	//fprintf(stderr, "addRoot(%s)\n", vpath);
+	char completePath[MAXPATHLEN];
 	const char* path = NULL;
-	if ( strlen(fgFileSystemRoot) == 0 ) {
-		path = vpath;
-	}
-	else {
+	// check -overlay path first
+	for (std::vector<const char*>::const_iterator it=fgFileSystemOverlays.begin(); it != fgFileSystemOverlays.end(); ++it) {
+		strcpy(completePath, *it);
+		strcat(completePath, vpath);	// assumes vpath starts with '/'
+		struct stat stat_buf;
+		if ( stat(completePath, &stat_buf) == 0 ) {
+			path = completePath;
+			break;
+		}
+	}
+	// if not found in overlay, check for -root
+	if ( (path == NULL) && (fgFileSystemRoot[0] != '\0') ) {
 		strcpy(completePath, fgFileSystemRoot);
 		strcat(completePath, vpath);	// assumes vpath starts with '/'
-		if ( fgUsesOverlay ) {
-			// using -overlay means if /overlay/usr/lib exists use it, otherwise use original path
-			struct stat stat_buf;
-			if ( stat(completePath, &stat_buf) == 0 )
-				path = completePath;
-			else
-				path = vpath;
-		}
-		else {
-			// using -root means alway redirect /usr/lib to /rootpath/usr/lib
+		struct stat stat_buf;
+		if ( stat(completePath, &stat_buf) == 0 )
 			path = completePath;
-		}
-	}
+	}
+	if ( path == NULL ) 
+		path = vpath;
+	
 	try {
+		//fprintf(stderr, "    UniversalMachOLayout::find(%s)\n", path);
 		const UniversalMachOLayout& uni = UniversalMachOLayout::find(path, &onlyArchs);
 		for(std::set<ArchPair>::iterator ait = onlyArchs.begin(); ait != onlyArchs.end(); ++ait) {
 			try {
@@ -230,53 +259,59 @@
 }
 
 // a virtual path does not have the fgFileSystemRoot prefix
-// a virtual path does not have the fgFileSystemRoot prefix
 ArchGraph::DependencyNode* ArchGraph::getNodeForVirtualPath(const char* vpath)
 {
-	if ( fgFileSystemRoot == NULL ) {
-		return this->getNode(vpath);
-	}
-	else {
-		char completePath[strlen(fgFileSystemRoot)+strlen(vpath)+2];
+	//fprintf(stderr, "getNodeForVirtualPath(%s)\n", vpath);
+	char completePath[MAXPATHLEN];
+	for (std::vector<const char*>::const_iterator it=fgFileSystemOverlays.begin(); it != fgFileSystemOverlays.end(); ++it) {
+		const char* overlayPath = *it;
+		// using -overlay means if /overlay/path/dylib exists use it, otherwise use /path/dylib
+		strcpy(completePath, overlayPath);
+		strcat(completePath, vpath);	// assumes vpath starts with '/'
+		struct stat stat_buf;
+		if ( stat(completePath, &stat_buf) == 0 ) {
+			return this->getNode(completePath);
+		}
+		// <rdar://problem/9279770> support when install name is a symlink
+		const char* pathToSymlink = vpath;
+		if ( fgFileSystemRoot[0] != '\0' ) {
+			strcpy(completePath, fgFileSystemRoot);
+			strcat(completePath, vpath);
+			pathToSymlink = completePath;
+		}
+		if ( (lstat(pathToSymlink, &stat_buf) == 0) && S_ISLNK(stat_buf.st_mode) ) {
+			// requested path did not exist in /overlay, but leaf of path is a symlink in /
+			char pathInSymLink[MAXPATHLEN];
+			size_t res = readlink(pathToSymlink, pathInSymLink, sizeof(pathInSymLink));
+			if ( res != -1 ) {
+				pathInSymLink[res] = '\0';
+				if ( pathInSymLink[0] != '/' ) {
+					char symFullPath[MAXPATHLEN];
+					strcpy(symFullPath, vpath);
+					char* lastSlash = strrchr(symFullPath, '/');
+					if ( lastSlash != NULL ) {
+						strcpy(lastSlash+1, pathInSymLink);
+						// (re)try looking for what symlink points to, but in /overlay
+						return this->getNodeForVirtualPath(symFullPath);
+					}
+				} 
+			}
+		}
+	}
+
+	if ( fgFileSystemRoot[0] != '\0' ) {
+		// using -root means always use /rootpath/usr/lib
 		strcpy(completePath, fgFileSystemRoot);
 		strcat(completePath, vpath);	// assumes vpath starts with '/'
-		if ( fgUsesOverlay ) {
-			// using -overlay means if /overlay/path/dylib exists use it, otherwise use /path/dylib
-			struct stat stat_buf;
-			if ( stat(completePath, &stat_buf) == 0 )
-				return this->getNode(completePath);
-			else {
-				// <rdar://problem/9279770> support when install name is a symlink
-				if ( (lstat(vpath, &stat_buf) == 0) && S_ISLNK(stat_buf.st_mode) ) {
-					// requested path did not exist in /overlay, but leaf of path is a symlink in /
-					char pathInSymLink[MAXPATHLEN];
-					size_t res = readlink(vpath, pathInSymLink, sizeof(pathInSymLink));
-					if ( res != -1 ) {
-						pathInSymLink[res] = '\0';
-						if ( pathInSymLink[0] != '/' ) {
-							char symFullPath[MAXPATHLEN];
-							strcpy(symFullPath, vpath);
-							char* lastSlash = strrchr(symFullPath, '/');
-							if ( lastSlash != NULL ) {
-								strcpy(lastSlash+1, pathInSymLink);
-								// (re)try looking for what symlink points to, but in /overlay
-								return this->getNodeForVirtualPath(symFullPath);
-							}
-						} 
-					}
-				}
-				return this->getNode(vpath);
-			}
-		}
-		else {
-			// using -root means always use /rootpath/usr/lib
-			return this->getNode(completePath);
-		}
-	}
+		return this->getNode(completePath);
+	}
+	// not found in -overlay or -root not used
+	return this->getNode(vpath);
 }
 
 ArchGraph::DependencyNode* ArchGraph::getNode(const char* path)
 {
+	//fprintf(stderr, "getNode(%s)\n", path);
 	// look up supplied path to see if node already exists
 	PathToNode::iterator pos = fNodes.find(path);
 	if ( pos != fNodes.end() )
@@ -317,19 +352,40 @@
 		//fprintf(stderr, "adding %s node alias %s for %s\n", archName(fArchPair), node->getLayout()->getID().name, realPath);
 		pos = fNodes.find(node->getLayout()->getID().name);
 		if ( pos != fNodes.end() ) {
-			// <rdar://problem/8305479> warn if two dylib in cache have same install_name
-			char* msg;
-			asprintf(&msg, "update_dyld_shared_cache: warning, found two dylibs with same install path: %s\n\t%s\n\t%s\n", 
-									node->getLayout()->getID().name, pos->second->getPath(), node->getPath());
-			fprintf(stderr, "%s", msg);
-			warnings.push_back(msg);
+			// get uuids of two dylibs to see if this is accidental copy of a dylib or two differnent dylibs with same -install_name
+			uuid_t uuid1;
+			uuid_t uuid2;
+			node->getLayout()->uuid(uuid1);
+			pos->second->getLayout()->uuid(uuid2);
+			if ( memcmp(&uuid1, &uuid2, 16) == 0 ) {
+				// <rdar://problem/8305479> warn if two dylib in cache have same install_name
+				char* msg;
+				asprintf(&msg, "update_dyld_shared_cache: warning, found two copies of the same dylib with same install path: %s\n\t%s\n\t%s\n", 
+										node->getLayout()->getID().name, pos->second->getPath(), node->getPath());
+				fprintf(stderr, "%s", msg);
+				warnings.push_back(msg);
+			}
+			else {
+				// <rdar://problem/12763450> update_dyld_shared_cache should fail if two images have same install name
+				fprintf(stderr, "update_dyld_shared_cache: found two different dylibs with same install path: %s\n\t%s\n\t%s\n", 
+							node->getLayout()->getID().name, pos->second->getPath(), node->getPath());
+				exit(1);
+			}
 		}
 		else
 			fNodes[node->getLayout()->getID().name] = node;
 		// update fAliasesMap with symlinks found
 		const char* aliasPath = realPath;
-		if ( (fgFileSystemRoot != NULL) && (strncmp(realPath, fgFileSystemRoot, strlen(fgFileSystemRoot)) == 0) ) {
+		if ( (fgFileSystemRoot != NULL) && (fgFileSystemRoot[0] != '\0') && (strncmp(realPath, fgFileSystemRoot, strlen(fgFileSystemRoot)) == 0) ) {
 			aliasPath = &realPath[strlen(fgFileSystemRoot)];
+		}
+		// <rdar://problem/11192810> Too many aliases in -overlay mode
+		for (std::vector<const char*>::const_iterator it=fgFileSystemOverlays.begin(); it != fgFileSystemOverlays.end(); ++it) {
+			const char* overlayPath = *it;
+			if ( strncmp(realPath, overlayPath, strlen(overlayPath)) == 0 ) {
+				aliasPath = &realPath[strlen(overlayPath)];
+				break;
+			}
 		}
 		if ( fAliasesMap.find(aliasPath) == fAliasesMap.end() ) {
 			if ( strcmp(aliasPath, node->getLayout()->getID().name) != 0 ) {
@@ -402,7 +458,7 @@
 					fDependsOn.insert(fGraph->getNodeForVirtualPath(dependentPath));
 			}
 			catch (const char* msg) {
-				if ( it->weakImport && ! fLayout->hasSplitSegInfo() ) {
+				if ( it->weakImport || ! fLayout->hasSplitSegInfo() ) {
 					// ok to ignore missing weak imported dylibs from things that are
 					// not going to be in the dyld shared cache
 				}
@@ -476,12 +532,15 @@
 const char*	ArchGraph::archName(ArchPair ap)
 {
 	switch ( ap.arch ) {
-		case CPU_TYPE_POWERPC:
-			return "ppc";
 		case CPU_TYPE_I386:
 			return "i386";
 		case CPU_TYPE_X86_64:
-			return "x86_64";
+			switch ( ap.subtype ) {
+				case CPU_SUBTYPE_X86_64_H:
+					return "x86_64h";
+				default:
+					return "x86_64";
+			}
 		case CPU_TYPE_ARM:
 			switch ( ap.subtype ) {
 				case CPU_SUBTYPE_ARM_V4T:
@@ -494,9 +553,17 @@
 					return "arm-xscale";
 				case CPU_SUBTYPE_ARM_V7:
 					return "armv7";
+				case CPU_SUBTYPE_ARM_V7F:
+					return "armv7f";
+				case CPU_SUBTYPE_ARM_V7K:
+					return "armv7k";
+				case CPU_SUBTYPE_ARM_V7S:
+					return "armv7s";
 				default:
 					return "arm";
 			}
+		case CPU_TYPE_ARM64:
+			return "arm64";
 		default:
 			return "unknown";
 	}
@@ -506,18 +573,18 @@
 {
 	if ( ! layout->isTwoLevelNamespace() ) 
 		asprintf(msg, "can't put %s in shared cache because it was built -flat_namespace", layout->getID().name);
+	else if ( ! layout->inSharableLocation() )
+		asprintf(msg, "can't put %s in shared cache because its -install_name is not in /usr/lib or /System/Library", layout->getID().name);
 	else if ( ! layout->hasSplitSegInfo() ) 
 		asprintf(msg, "can't put %s in shared cache because it was not built for %s or later", layout->getID().name, (iPhoneOS ? "iPhoneOS 3.1" : "MacOSX 10.5"));
 	else if ( ! layout->isRootOwned() )
 		asprintf(msg, "can't put %s in shared cache because it is not owned by root", layout->getID().name);
-	else if ( ! layout->inSharableLocation() )
-		asprintf(msg, "can't put %s in shared cache because it is not in /usr/lib or /System/Library", layout->getID().name);
 	else if ( layout->hasDynamicLookupLinkage() )
 		asprintf(msg, "can't put %s in shared cache because it was built with '-undefined dynamic_lookup'", layout->getID().name);
 	else if ( layout->hasMainExecutableLookupLinkage() )
 		asprintf(msg, "can't put %s in shared cache because it was built with '-bundle_loader'", layout->getID().name);
-	//else if ( ! layout->hasDyldInfo() )
-	//	asprintf(msg, "can't put %s in shared cache because it was built for older OS", layout->getID().name);
+	else if ( layout->hasMultipleReadWriteSegments() )
+		asprintf(msg, "can't put %s in shared cache because it has multiple r/w segments", layout->getID().name);
 	else
 		return true;
 	return false;
@@ -554,6 +621,9 @@
 				if ( nodes.find(realPath) != nodes.end() )
 					continue;
 			}
+			// handle weak imported dylibs not found
+			if ( dit->weakImport )
+				continue;
 			shareableMap[layout] = false;
 			char* msg;
 			asprintf(&msg, "can't put %s in shared cache because it depends on %s which can't be found", layout->getID().name, dit->name);
@@ -578,13 +648,91 @@
 }
 
 
+
+class StringPool
+{
+public:
+				StringPool();
+	const char*	getBuffer();
+	uint32_t	size();
+	uint32_t	add(const char* str);
+	uint32_t	addUnique(const char* str);
+	const char* stringAtIndex(uint32_t) const;
+	
+private:
+	typedef std::unordered_map<const char*, uint32_t, CStringHash, CStringEquals> StringToOffset;
+
+	char*			fBuffer;
+	uint32_t		fBufferAllocated;
+	uint32_t		fBufferUsed;
+	StringToOffset	fUniqueStrings;
+};
+
+
+StringPool::StringPool() 
+	: fBufferUsed(0), fBufferAllocated(64*1024*1024)
+{
+	fBuffer = (char*)malloc(fBufferAllocated);
+}
+
+uint32_t StringPool::add(const char* str)
+{
+	uint32_t len = strlen(str);
+	if ( (fBufferUsed + len + 1) > fBufferAllocated ) {
+		// grow buffer
+		throw "string buffer exhausted";
+	}
+	strcpy(&fBuffer[fBufferUsed], str);
+	uint32_t result = fBufferUsed;
+	fUniqueStrings[&fBuffer[fBufferUsed]] = result;
+	fBufferUsed += len+1;
+	return result;
+}
+
+uint32_t StringPool::addUnique(const char* str)
+{
+	StringToOffset::iterator pos = fUniqueStrings.find(str);
+	if ( pos != fUniqueStrings.end() ) 
+		return pos->second;
+	else {
+		//fprintf(stderr, "StringPool::addUnique() new string: %s\n", str);
+		return this->add(str);
+	}
+}
+
+uint32_t StringPool::size()
+{
+	return fBufferUsed;
+}
+
+const char*	StringPool::getBuffer()
+{
+	return fBuffer;
+}
+
+const char* StringPool::stringAtIndex(uint32_t index) const
+{
+	return &fBuffer[index];
+}
+
+
+
+struct LocalSymbolInfo
+{
+	uint32_t	dylibOffset;
+	uint32_t	nlistStartIndex;
+	uint32_t	nlistCount;
+};
+
+
 template <typename A>
 class SharedCache
 {
 public:
-							SharedCache(ArchGraph* graph, const char* rootPath, const char* cacheDir, bool alphaSort, bool verify, bool optimize, bool overlay, uint64_t dyldBaseAddress);
-	bool					update(bool usesOverlay, bool force, bool optimize, bool deleteExistingFirst, int archIndex, 
-										int archCount, bool keepSignatures);
+							SharedCache(ArchGraph* graph, const char* rootPath, const std::vector<const char*>& overlayPaths, const char* cacheDir, bool explicitCacheDir,
+											bool alphaSort, bool verify, bool optimize, uint64_t dyldBaseAddress);
+	bool					update(bool force, bool optimize, bool deleteExistingFirst, int archIndex, 
+										int archCount, bool keepSignatures, bool dontMapLocalSymbols);
 	static const char*		cacheFileSuffix(bool optimized, const char* archName);
 
     // vm address = address AS WRITTEN into the cache
@@ -595,6 +743,8 @@
 	uint64_t				cacheFileOffsetForVMAddress(uint64_t addr) const;
 	uint64_t				VMAddressForCacheFileOffset(uint64_t addr) const;
 
+	static const char*		archName();
+
 private:
 	typedef typename A::P			P;
     typedef typename A::P::E		E;
@@ -602,19 +752,22 @@
 
 	bool					notUpToDate(const char* path, unsigned int aliasCount);
 	bool					notUpToDate(const void* cache, unsigned int aliasCount);
-	uint8_t*				optimizeLINKEDIT(bool keepSignatures);
+	uint8_t*				optimizeLINKEDIT(bool keepSignatures, bool dontMapLocalSymbols);
 	void					optimizeObjC(std::vector<void*>& pointersInData);
 
 	static void				getSharedCacheBasAddresses(cpu_type_t arch, uint64_t* baseReadOnly, uint64_t* baseWritable);
 	static cpu_type_t		arch();
-	static const char*		archName();
-	static uint64_t			sharedRegionReadOnlyStartAddress();
-	static uint64_t			sharedRegionWritableStartAddress();
-	static uint64_t			sharedRegionReadOnlySize();
-	static uint64_t			sharedRegionWritableSize();
+	static uint64_t			sharedRegionStartAddress();
+	static uint64_t			sharedRegionSize();
+	static uint64_t			sharedRegionStartWritableAddress(uint64_t);
+	static uint64_t			sharedRegionStartReadOnlyAddress(uint64_t, uint64_t);
 	static uint64_t			getWritableSegmentNewAddress(uint64_t proposedNewAddress, uint64_t originalAddress, uint64_t executableSlide);
 	static bool				addCacheSlideInfo();
-	
+	static uint64_t			pathHash(const char*);
+	
+	static uint64_t			pageAlign(uint64_t addr);
+	static uint64_t			regionAlign(uint64_t addr);
+	static uint64_t			pageAlign4KB(uint64_t addr);
 	void					assignNewBaseAddresses(bool verify);
 
 	struct LayoutInfo {
@@ -626,6 +779,12 @@
 	struct ByNameSorter {
 		bool operator()(const LayoutInfo& left, const LayoutInfo& right) 
 				{ return (strcmp(left.layout->getID().name, right.layout->getID().name) < 0); }
+	};
+	
+	struct ByAddressSorter {
+		bool operator()(const LayoutInfo& left, const LayoutInfo& right) { 
+			return (left.layout->getSegments()[0].newAddress() < right.layout->getSegments()[0].newAddress()); 
+		}
 	};
 
     struct ByCStringSectionSizeSorter {
@@ -670,6 +829,9 @@
 	std::vector<LayoutInfo>				fDylibs;
 	std::vector<LayoutInfo>				fDylibAliases;
 	std::vector<shared_file_mapping_np>	fMappings;
+	std::vector<macho_nlist<P> >		fUnmappedLocalSymbols;
+	StringPool							fUnmappedLocalsStringPool;
+	std::vector<LocalSymbolInfo>		fLocalSymbolInfos;
 	uint32_t							fHeaderSize;
     uint8_t*							fInMemoryCache;
 	uint64_t							fDyldBaseAddress;
@@ -690,7 +852,10 @@
 	uint32_t							fSizeOfOldStringPoolInCombinedLinkedit;
 	uint32_t							fOffsetOfFunctionStartsInCombinedLinkedit;
 	uint32_t							fSizeOfFunctionStartsInCombinedLinkedit;
+	uint32_t							fOffsetOfDataInCodeInCombinedLinkedit;
+	uint32_t							fSizeOfDataInCodeInCombinedLinkedit;
 	uint32_t							fLinkEditsTotalOptimizedSize;
+	uint32_t							fUnmappedLocalSymbolsSize;
 };
 
 
@@ -783,44 +948,57 @@
 
 
 	
-template <>	 cpu_type_t	SharedCache<ppc>::arch()	{ return CPU_TYPE_POWERPC; }
 template <>	 cpu_type_t	SharedCache<x86>::arch()	{ return CPU_TYPE_I386; }
 template <>	 cpu_type_t	SharedCache<x86_64>::arch()	{ return CPU_TYPE_X86_64; }
 template <>	 cpu_type_t	SharedCache<arm>::arch()	{ return CPU_TYPE_ARM; }
-
-template <>	 uint64_t	SharedCache<ppc>::sharedRegionReadOnlyStartAddress()	{ return 0x90000000; }
-template <>	 uint64_t	SharedCache<x86>::sharedRegionReadOnlyStartAddress()	{ return 0x90000000; }
-template <>	 uint64_t	SharedCache<x86_64>::sharedRegionReadOnlyStartAddress()	{ return 0x7FFF80000000LL; }
-template <>	 uint64_t	SharedCache<arm>::sharedRegionReadOnlyStartAddress()	{ return 0x30000000; }
-
-template <>	 uint64_t	SharedCache<ppc>::sharedRegionWritableStartAddress()	{ return 0xA0000000; }
-template <>	 uint64_t	SharedCache<x86>::sharedRegionWritableStartAddress()	{ return 0xAC000000; }
-template <>	 uint64_t	SharedCache<x86_64>::sharedRegionWritableStartAddress()	{ return 0x7FFF70000000LL; }
-template <>	 uint64_t	SharedCache<arm>::sharedRegionWritableStartAddress()	{ return 0x3E000000; }
-
-template <>	 uint64_t	SharedCache<ppc>::sharedRegionReadOnlySize()			{ return 0x10000000; }
-template <>	 uint64_t	SharedCache<x86>::sharedRegionReadOnlySize()			{ return 0x1C000000; }
-template <>	 uint64_t	SharedCache<x86_64>::sharedRegionReadOnlySize()			{ return 0x40000000; }
-template <>	 uint64_t	SharedCache<arm>::sharedRegionReadOnlySize()			{ return 0x0E000000; }
-
-template <>	 uint64_t	SharedCache<ppc>::sharedRegionWritableSize()			{ return 0x10000000; }
-template <>	 uint64_t	SharedCache<x86>::sharedRegionWritableSize()			{ return 0x04000000; }
-template <>	 uint64_t	SharedCache<x86_64>::sharedRegionWritableSize()			{ return 0x10000000; }
-template <>	 uint64_t	SharedCache<arm>::sharedRegionWritableSize()			{ return 0x02000000; }
-
-
-template <>	 const char*	SharedCache<ppc>::archName()	{ return "ppc"; }
+template <>	 cpu_type_t	SharedCache<arm64>::arch()	{ return CPU_TYPE_ARM64; }
+
+template <>	 uint64_t	SharedCache<x86>::sharedRegionStartAddress()			{ return 0x90000000; }
+template <>	 uint64_t	SharedCache<x86_64>::sharedRegionStartAddress()			{ return 0x7FFF80000000LL; }
+template <>	 uint64_t	SharedCache<arm>::sharedRegionStartAddress()			{ return ARM_SHARED_REGION_START; }
+template <>	 uint64_t	SharedCache<arm64>::sharedRegionStartAddress()			{ return ARM64_SHARED_REGION_START; }
+
+template <>	 uint64_t	SharedCache<x86>::sharedRegionSize()					{ return 0x20000000; }
+template <>	 uint64_t	SharedCache<x86_64>::sharedRegionSize()					{ return 0x40000000; }
+template <>	 uint64_t	SharedCache<arm>::sharedRegionSize()					{ return ARM_SHARED_REGION_SIZE; }
+template <>	 uint64_t	SharedCache<arm64>::sharedRegionSize()					{ return ARM64_SHARED_REGION_SIZE; }
+
+template <>	 uint64_t	SharedCache<x86>::sharedRegionStartWritableAddress(uint64_t exEnd)			{ return exEnd + 0x04000000; }
+template <>	 uint64_t	SharedCache<x86_64>::sharedRegionStartWritableAddress(uint64_t exEnd)		{ return 0x7FFF70000000LL; }
+template <>	 uint64_t	SharedCache<arm>::sharedRegionStartWritableAddress(uint64_t exEnd)			{ return (exEnd + 16383) & (-16384); }
+template <>	 uint64_t	SharedCache<arm64>::sharedRegionStartWritableAddress(uint64_t exEnd)		{ return exEnd; }
+
+template <>	 uint64_t	SharedCache<x86>::sharedRegionStartReadOnlyAddress(uint64_t wrEnd, uint64_t exEnd)	 { return wrEnd + 0x04000000; }
+template <>	 uint64_t	SharedCache<x86_64>::sharedRegionStartReadOnlyAddress(uint64_t wrEnd, uint64_t exEnd){ return exEnd; }
+template <>	 uint64_t	SharedCache<arm>::sharedRegionStartReadOnlyAddress(uint64_t wrEnd, uint64_t exEnd)	 { return (wrEnd + 16383) & (-16384); }
+template <>	 uint64_t	SharedCache<arm64>::sharedRegionStartReadOnlyAddress(uint64_t wrEnd, uint64_t exEnd) { return (wrEnd + 16383) & (-16384); }
+
 template <>	 const char*	SharedCache<x86>::archName()	{ return "i386"; }
 template <>	 const char*	SharedCache<x86_64>::archName()	{ return "x86_64"; }
 template <>	 const char*	SharedCache<arm>::archName()	{ return "arm"; }
-
-template <>	 const char*	SharedCache<ppc>::cacheFileSuffix(bool optimized, const char*)	{ return optimized ? "ppc" : "rosetta"; }
+template <>	 const char*	SharedCache<arm64>::archName()	{ return "arm64"; }
+
 template <>	 const char*	SharedCache<x86>::cacheFileSuffix(bool, const char* archName)	{ return archName; }
 template <>	 const char*	SharedCache<x86_64>::cacheFileSuffix(bool, const char* archName){ return archName; }
 template <>	 const char*	SharedCache<arm>::cacheFileSuffix(bool, const char* archName)	{ return archName; }
+template <>	 const char*	SharedCache<arm64>::cacheFileSuffix(bool, const char* archName)	{ return archName; }
+
+template <>  uint64_t		SharedCache<x86>::pageAlign(uint64_t addr)    { return ( (addr + 4095) & (-4096) ); }
+template <>  uint64_t		SharedCache<x86_64>::pageAlign(uint64_t addr) { return ( (addr + 4095) & (-4096) ); }
+template <>  uint64_t		SharedCache<arm>::pageAlign(uint64_t addr)    { return ( (addr + 4095) & (-4096) ); }
+template <>  uint64_t		SharedCache<arm64>::pageAlign(uint64_t addr)  { return ( (addr + 16383) & (-16384) ); }
+
+template <>  uint64_t		SharedCache<x86>::regionAlign(uint64_t addr)    { return ( (addr + 4095) & (-4096) ); }
+template <>  uint64_t		SharedCache<x86_64>::regionAlign(uint64_t addr) { return ( (addr + 4095) & (-4096) ); }
+template <>  uint64_t		SharedCache<arm>::regionAlign(uint64_t addr)    { return ( (addr + 16383) & (-16384) ); }
+template <>  uint64_t		SharedCache<arm64>::regionAlign(uint64_t addr)  { return ( (addr + 16383) & (-16384) ); }
+
+
+template <typename A>  
+uint64_t SharedCache<A>::pageAlign4KB(uint64_t addr)    { return ( (addr + 4095) & (-4096) ); }
 
 template <typename A>
-SharedCache<A>::SharedCache(ArchGraph* graph, const char* rootPath, const char* cacheDir, bool alphaSort, bool verify, bool optimize, bool overlay, uint64_t dyldBaseAddress) 
+SharedCache<A>::SharedCache(ArchGraph* graph, const char* rootPath, const std::vector<const char*>& overlayPaths, const char* cacheDir, bool explicitCacheDir, bool alphaSort, bool verify, bool optimize, uint64_t dyldBaseAddress) 
   : fArchGraph(graph), fVerify(verify), fExistingIsNotUpToDate(true), 
 	fCacheFileInFinalLocation(rootPath[0] == '\0'), fCacheFilePath(NULL),
 	fExistingCacheForVerification(NULL), fDyldBaseAddress(dyldBaseAddress),
@@ -830,7 +1008,9 @@
 	fOffsetOfOldExternalRelocationsInCombinedLinkedit(0), fSizeOfOldExternalRelocationsInCombinedLinkedit(0),
 	fOffsetOfOldIndirectSymbolsInCombinedLinkedit(0), fSizeOfOldIndirectSymbolsInCombinedLinkedit(0),
 	fOffsetOfOldStringPoolInCombinedLinkedit(0), fSizeOfOldStringPoolInCombinedLinkedit(0),
-	fOffsetOfFunctionStartsInCombinedLinkedit(0), fSizeOfFunctionStartsInCombinedLinkedit(0)
+	fOffsetOfFunctionStartsInCombinedLinkedit(0), fSizeOfFunctionStartsInCombinedLinkedit(0),
+	fOffsetOfDataInCodeInCombinedLinkedit(0), fSizeOfDataInCodeInCombinedLinkedit(0),
+	fUnmappedLocalSymbolsSize(0)
 {
 	if ( fArchGraph->getArchPair().arch != arch() )
 		throwf("SharedCache object is wrong architecture: 0x%08X vs 0x%08X", fArchGraph->getArchPair().arch, arch());
@@ -844,8 +1024,12 @@
 		LayoutInfo temp;
 		temp.layout = lib;
 		temp.info.address = 0;
+		temp.info.inode = lib->getInode();
 		temp.info.modTime = lib->getLastModTime();
-		temp.info.inode = lib->getInode();
+		if ( iPhoneOS ) {
+			temp.info.inode = pathHash(lib->getID().name);
+			temp.info.modTime = 0;
+		}
 		temp.info.pathFileOffset = lib->getNameFileOffset();  // for now this is the offset within the dylib
 		for(ArchGraph::StringToString::iterator ait = aliases.begin(); ait != aliases.end(); ++ait) {
 			if ( strcmp(ait->second, lib->getID().name) == 0 ) {
@@ -857,29 +1041,51 @@
 	}
 
 	// create path to cache file
-	char cachePathNonOverlay[1024];
-	strcpy(cachePathNonOverlay, cacheDir);
-	if ( cachePathNonOverlay[strlen(cachePathNonOverlay)-1] != '/' )
-		strcat(cachePathNonOverlay, "/");
-	strcat(cachePathNonOverlay, DYLD_SHARED_CACHE_BASE_NAME);
-	strcat(cachePathNonOverlay, cacheFileSuffix(optimize, fArchGraph->archName()));
-	char cachePath[1024];
-	strcpy(cachePath, rootPath);
-	strcat(cachePath, "/");
-	strcat(cachePath, cachePathNonOverlay);
-	if ( !overlay && (rootPath[0] != '\0') )
-		fCacheFilePath = strdup(cachePathNonOverlay);
-	else
+	char cachePathCanonical[MAXPATHLEN];
+	strcpy(cachePathCanonical, cacheDir);
+	if ( cachePathCanonical[strlen(cachePathCanonical)-1] != '/' )
+		strcat(cachePathCanonical, "/");
+	strcat(cachePathCanonical, DYLD_SHARED_CACHE_BASE_NAME);
+	strcat(cachePathCanonical, cacheFileSuffix(optimize, fArchGraph->archName()));
+	char cachePath[MAXPATHLEN];
+	if ( explicitCacheDir ) {
+		fCacheFilePath = strdup(cachePathCanonical);
+	}
+	else if ( overlayPaths.size() == 1 ) {
+		// if no -cache_dir and exactly on -overlay, write cache file into that overlay dir
+		strcpy(cachePath, overlayPaths[0]);
+		strcat(cachePath, "/");
+		strcat(cachePath, cachePathCanonical);
 		fCacheFilePath = strdup(cachePath);
-	if ( overlay ) {
-		// in overlay mode if there already is a cache file in the overlay
-		// check if it is up to date.  If there is no file, check if
-		// the one in the boot volume is up to date.
+	}
+	else if ( rootPath[0] != '\0' ) {
+		strcpy(cachePath, rootPath);
+		strcat(cachePath, "/");
+		strcat(cachePath, cachePathCanonical);
+		fCacheFilePath = strdup(cachePath);
+	}
+	else {
+		fCacheFilePath = strdup(cachePathCanonical);
+	}
+	if ( overlayPaths.size() == 1 ) {
+		// in overlay mode if there already is a cache file in the overlay,
+		// check if it is up to date.  
 		struct stat stat_buf;
-		if ( stat(fCacheFilePath, &stat_buf) == 0 ) 
+		if ( stat(fCacheFilePath, &stat_buf) == 0 ) {
 			fExistingIsNotUpToDate = this->notUpToDate(fCacheFilePath, aliasCount);
-		else
-			fExistingIsNotUpToDate = this->notUpToDate(cachePathNonOverlay, aliasCount);
+		}
+		else if ( rootPath[0] != '\0' ) {
+			// using -root and -overlay, but no cache file in overlay, check one in -root
+			char cachePathRoot[MAXPATHLEN];
+			strcpy(cachePathRoot, rootPath);
+			strcat(cachePathRoot, "/");
+			strcat(cachePathRoot, cachePathCanonical);
+			fExistingIsNotUpToDate = this->notUpToDate(cachePathRoot, aliasCount);
+		}
+		else {
+			// uisng -overlay, but no cache file in overlay, check one in boot volume
+			fExistingIsNotUpToDate = this->notUpToDate(cachePathCanonical, aliasCount);
+		}
 	}
 	else {
 		fExistingIsNotUpToDate = this->notUpToDate(fCacheFilePath, aliasCount);
@@ -908,6 +1114,7 @@
 	fHeaderSize = sizeof(dyld_cache_header) 
 							+ fMappings.size()*sizeof(shared_file_mapping_np) 
 							+ (fDylibs.size()+aliasCount)*sizeof(dyld_cache_image_info);
+	const uint64_t baseHeaderSize = fHeaderSize;
 	//fprintf(stderr, "aliasCount=%d, fHeaderSize=0x%08X\n", aliasCount, fHeaderSize);
 	// build list of aliases and compute where each ones path string will go
 	for(typename std::vector<struct LayoutInfo>::const_iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
@@ -918,6 +1125,10 @@
 			temp.aliases.clear();
 			temp.aliases.push_back(aliasPath);
 			temp.info.pathFileOffset = fHeaderSize; 
+			if ( iPhoneOS ) {
+				temp.info.inode = pathHash(aliasPath); 
+				temp.info.modTime = 0;
+			}
 			fDylibAliases.push_back(temp);
 			fHeaderSize += strlen(aliasPath)+1;
 		}
@@ -931,21 +1142,21 @@
 		// if no existing cache, say so
 		if ( fExistingCacheForVerification == NULL ) {
 			throwf("update_dyld_shared_cache[%u] for arch=%s, could not verify because cache file does not exist in /var/db/dyld/\n",
-			 getpid(), archName());
+			 getpid(), fArchGraph->archName());
 		}
 		const dyldCacheHeader<E>* header = (dyldCacheHeader<E>*)fExistingCacheForVerification;
 		const dyldCacheImageInfo<E>* cacheEntry = (dyldCacheImageInfo<E>*)(fExistingCacheForVerification + header->imagesOffset());
 		for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it, ++cacheEntry) {
 			if ( cacheEntry->address() != it->layout->getSegments()[0].newAddress() ) {
 				throwf("update_dyld_shared_cache[%u] warning: for arch=%s, could not verify cache because start address of %s is 0x%llX in cache, but should be 0x%llX\n",
-							getpid(), archName(), it->layout->getID().name, cacheEntry->address(), it->layout->getSegments()[0].newAddress());
+							getpid(), fArchGraph->archName(), it->layout->getID().name, cacheEntry->address(), it->layout->getSegments()[0].newAddress());
 			}
 		}
 	}
 	
 	
 	if ( fHeaderSize > FIRST_DYLIB_TEXT_OFFSET )
-		throwf("header size miscalculation 0x%08X", fHeaderSize);
+		throwf("header size overflow: allowed=0x%08X, base=0x%08llX, aliases=0x%08llX", FIRST_DYLIB_TEXT_OFFSET, baseHeaderSize, fHeaderSize-baseHeaderSize);
 }
 
 
@@ -955,143 +1166,101 @@
 	return proposedNewAddress;
 }
 
-template <>
-uint64_t SharedCache<ppc>::getWritableSegmentNewAddress(uint64_t proposedNewAddress, uint64_t originalAddress, uint64_t executableSlide)
-{
-	// for ppc writable segments can only move in increments of 64K (so only hi16 instruction needs to be modified)
-	return (((executableSlide & 0x000000000000F000ULL) - ((proposedNewAddress - originalAddress) & 0x000000000000F000ULL)) & 0x000000000000F000ULL) + proposedNewAddress;
-}
-
+template <typename A>
+uint64_t SharedCache<A>::pathHash(const char* path)
+{
+	uint64_t sum = 0;
+	for (const char* s=path; *s != '\0'; ++s)
+		sum += sum*4 + *s;
+	return sum;
+}
+	
 
 template <typename A>
 void SharedCache<A>::assignNewBaseAddresses(bool verify)
 {
-	uint64_t sharedCacheStartAddress = sharedRegionReadOnlyStartAddress();
-#if 0
-	if ( arch() == CPU_TYPE_X86_64 ) {
-		if ( verify ) {
-			if ( fExistingCacheForVerification == NULL ) {
-				throwf("update_dyld_shared_cache[%u] for arch=%s, could not verify because cache file does not exist in /var/db/dyld/\n",
-					   getpid(), archName());
-			}
-			const dyldCacheHeader<E>* header = (dyldCacheHeader<E>*)fExistingCacheForVerification;
-			const dyldCacheFileMapping<E>* mappings = (dyldCacheFileMapping<E>*)(fExistingCacheForVerification + header->mappingOffset());
-			sharedCacheStartAddress = mappings[0].address();
-		}
-		else {
-			// <rdar://problem/5274722> dyld shared cache can be more random
-			uint64_t readOnlySize = 0;
-			for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
-				if ( ! it->layout->hasSplitSegInfo() )
-					continue;
-				std::vector<MachOLayoutAbstraction::Segment>& segs = ((MachOLayoutAbstraction*)(it->layout))->getSegments();
-				for (int i=0; i < segs.size(); ++i) {
-					MachOLayoutAbstraction::Segment& seg = segs[i];
-					if ( ! seg.writable() )
-						readOnlySize += pageAlign(seg.size());
-				}
-			}
-			uint64_t maxSlide = sharedRegionReadOnlySize() - (readOnlySize + FIRST_DYLIB_TEXT_OFFSET);
-			sharedCacheStartAddress = sharedRegionReadOnlyStartAddress() + pageAlign(arc4random() % maxSlide);
-		}
-	}
-#endif
-	uint64_t currentExecuteAddress = sharedCacheStartAddress + FIRST_DYLIB_TEXT_OFFSET;	
-	uint64_t currentWritableAddress = sharedRegionWritableStartAddress() + FIRST_DYLIB_DATA_OFFSET;
-
-	// first layout TEXT and DATA for dylibs
+	// first layout TEXT for dylibs
+	const uint64_t startExecuteAddress = sharedRegionStartAddress();
+	uint64_t currentExecuteAddress = startExecuteAddress + FIRST_DYLIB_TEXT_OFFSET;	
 	for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
 		std::vector<MachOLayoutAbstraction::Segment>& segs = ((MachOLayoutAbstraction*)(it->layout))->getSegments();
-		MachOLayoutAbstraction::Segment* executableSegment = NULL;
+		for (int i=0; i < segs.size(); ++i) {
+			MachOLayoutAbstraction::Segment& seg = segs[i];
+			seg.reset();
+			if ( seg.executable() && !seg.writable() ) {
+				// <rdar://problem/15947734> Some dylib require extra alignment
+				currentExecuteAddress = (currentExecuteAddress + seg.alignment() - 1) & (-seg.alignment());
+				// __TEXT segment
+				if ( it->info.address == 0 )
+					it->info.address = currentExecuteAddress;
+				seg.setNewAddress(currentExecuteAddress);
+				currentExecuteAddress += pageAlign(seg.size());
+			}
+		}
+	}
+	// align __TEXT region
+	currentExecuteAddress = regionAlign(currentExecuteAddress);
+
+	// layout DATA for dylibs
+	const uint64_t startWritableAddress = sharedRegionStartWritableAddress(currentExecuteAddress);
+	uint64_t currentWritableAddress = startWritableAddress;
+	for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
+		std::vector<MachOLayoutAbstraction::Segment>& segs = ((MachOLayoutAbstraction*)(it->layout))->getSegments();
 		for (int i=0; i < segs.size(); ++i) {
 			MachOLayoutAbstraction::Segment& seg = segs[i];
 			seg.reset();
 			if ( seg.writable() ) {
-				if ( seg.executable() && it->layout->hasSplitSegInfo() ) {
-					// skip __IMPORT segments in this pass
-				}
-				else {
-					// __DATA segment
-					// for ppc, writable segments have to move in 64K increments
-					if (  it->layout->hasSplitSegInfo() ) {
-						if ( executableSegment == NULL )
-							throwf("first segment in dylib is not executable for %s", it->layout->getID().name);
-						seg.setNewAddress(getWritableSegmentNewAddress(currentWritableAddress, seg.address(), executableSegment->newAddress() - executableSegment->address()));
-					}
-					else
-						seg.setNewAddress(currentWritableAddress);
-					currentWritableAddress = pageAlign(seg.newAddress() + seg.size());
-				}
-			}
-			else {
-				if ( seg.executable() ) {
-					// __TEXT segment
-					if ( it->info.address == 0 )
-						it->info.address = currentExecuteAddress;
-					executableSegment = &seg;
-					seg.setNewAddress(currentExecuteAddress);
-					currentExecuteAddress += pageAlign(seg.size());
-				}
-				else {
-					// skip read-only segments in this pass
-				}
-			}
-		}
-	}
-
-	// append all read-only (but not LINKEDIT) segments at end of all TEXT segments
-	// append all IMPORT segments at end of all DATA segments rounded to next 2MB 
-	uint64_t currentReadOnlyAddress = currentExecuteAddress;
-	uint64_t startWritableExecutableAddress = (currentWritableAddress + 0x200000 - 1) & (-0x200000);
-	uint64_t currentWritableExecutableAddress = startWritableExecutableAddress;
+				if ( seg.executable() ) 
+					throw "found writable and executable segment";
+				// __DATA segment
+				seg.setNewAddress(currentWritableAddress);
+				// <rdar://problem/13089366> always 4KB align data pages to allow padding to be removed
+				currentWritableAddress = pageAlign4KB(seg.newAddress() + seg.size());
+			}
+		}
+	}
+	// align __DATA region
+	currentWritableAddress = regionAlign(currentWritableAddress);
+
+	// layout all read-only (but not LINKEDIT) segments
+	const uint64_t startReadOnlyAddress = sharedRegionStartReadOnlyAddress(currentWritableAddress, currentExecuteAddress);
+	uint64_t currentReadOnlyAddress = startReadOnlyAddress;
 	for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
 		std::vector<MachOLayoutAbstraction::Segment>& segs = ((MachOLayoutAbstraction*)(it->layout))->getSegments();
 		for(int i=0; i < segs.size(); ++i) {
 			MachOLayoutAbstraction::Segment& seg = segs[i];
-			if ( !seg.writable() && !seg.executable() && (strcmp(seg.name(), "__LINKEDIT") != 0) ) {
-				// allocate non-executable,read-only segments from end of read only shared region
+			if ( seg.readable() && !seg.writable() && !seg.executable() && (strcmp(seg.name(), "__LINKEDIT") != 0) ) {
+				// __UNICODE segment
 				seg.setNewAddress(currentReadOnlyAddress);
 				currentReadOnlyAddress += pageAlign(seg.size());
 			}
-			else if ( seg.writable() && seg.executable() && it->layout->hasSplitSegInfo() ) {
-				// allocate IMPORT segments to end of writable shared region
-				seg.setNewAddress(currentWritableExecutableAddress);
-				currentWritableExecutableAddress += pageAlign(seg.size());
-			}
 		}
 	}	
 
-	// append all LINKEDIT segments at end of all read-only segments
+	// layout all LINKEDIT segments at end of all read-only segments
+	currentReadOnlyAddress = regionAlign(currentReadOnlyAddress); // <rdar://problem/16491435>
 	fLinkEditsStartAddress = currentReadOnlyAddress;
 	fFirstLinkEditSegment = NULL;
 	for(typename std::vector<LayoutInfo>::iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
 		std::vector<MachOLayoutAbstraction::Segment>& segs = ((MachOLayoutAbstraction*)(it->layout))->getSegments();
 		for(int i=0; i < segs.size(); ++i) {
 			MachOLayoutAbstraction::Segment& seg = segs[i];
-			if ( !seg.writable() && !seg.executable() && (strcmp(seg.name(), "__LINKEDIT") == 0) ) {
+			if ( seg.readable() && !seg.writable() && !seg.executable() && (strcmp(seg.name(), "__LINKEDIT") == 0) ) {
 				if ( fFirstLinkEditSegment == NULL ) 
 					fFirstLinkEditSegment = &seg;
-				// allocate non-executable,read-only segments from end of read only shared region
 				seg.setNewAddress(currentReadOnlyAddress);
 				currentReadOnlyAddress += pageAlign(seg.size());
 			}
 		}
 	}
-	fLinkEditsTotalUnoptimizedSize = (currentReadOnlyAddress - fLinkEditsStartAddress + 4095) & (-4096);
-
-    // <rdar://problem/9361288> i386 dyld shared cache overflows after adding libclh.dylib
-	if ( (currentReadOnlyAddress -  sharedRegionReadOnlyStartAddress()) > sharedRegionReadOnlySize() )
-		throwf("read-only slice of cache too big: %lluMB (max %lluMB)", 
-				(currentReadOnlyAddress -  sharedRegionReadOnlyStartAddress())/(1024*1024),
-			   sharedRegionReadOnlySize()/(1024*1024));
-		
+	fLinkEditsTotalUnoptimizedSize = pageAlign(currentReadOnlyAddress - fLinkEditsStartAddress);
 
 	// populate large mappings
 	uint64_t cacheFileOffset = 0;
-	if ( currentExecuteAddress > sharedCacheStartAddress + FIRST_DYLIB_TEXT_OFFSET ) {
+	if ( currentExecuteAddress > startExecuteAddress ) {
 		shared_file_mapping_np  executeMapping;
-		executeMapping.sfm_address		= sharedCacheStartAddress;
-		executeMapping.sfm_size			= currentExecuteAddress - sharedCacheStartAddress;
+		executeMapping.sfm_address		= startExecuteAddress;
+		executeMapping.sfm_size			= currentExecuteAddress - startExecuteAddress;
 		executeMapping.sfm_file_offset	= cacheFileOffset;
 		executeMapping.sfm_max_prot		= VM_PROT_READ | VM_PROT_EXECUTE;
 		executeMapping.sfm_init_prot	= VM_PROT_READ | VM_PROT_EXECUTE;
@@ -1099,29 +1268,18 @@
 		cacheFileOffset += executeMapping.sfm_size;
 		
 		shared_file_mapping_np  writableMapping;
-		writableMapping.sfm_address		= sharedRegionWritableStartAddress();
-		writableMapping.sfm_size		= currentWritableAddress - sharedRegionWritableStartAddress();
+		writableMapping.sfm_address		= startWritableAddress;
+		writableMapping.sfm_size		= currentWritableAddress - startWritableAddress;
 		writableMapping.sfm_file_offset	= cacheFileOffset;
 		writableMapping.sfm_max_prot	= VM_PROT_READ | VM_PROT_WRITE;
 		writableMapping.sfm_init_prot	= VM_PROT_READ | VM_PROT_WRITE;
 		fMappings.push_back(writableMapping);
 		cacheFileOffset += writableMapping.sfm_size;
-		
-		if ( currentWritableExecutableAddress > startWritableExecutableAddress ) {
-			shared_file_mapping_np  writableExecutableMapping;
-			writableExecutableMapping.sfm_address	= startWritableExecutableAddress;
-			writableExecutableMapping.sfm_size		= currentWritableExecutableAddress - startWritableExecutableAddress;
-			writableExecutableMapping.sfm_file_offset= cacheFileOffset;
-			writableExecutableMapping.sfm_max_prot	= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
-			writableExecutableMapping.sfm_init_prot	= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 
-			fMappings.push_back(writableExecutableMapping);
-			cacheFileOffset += writableExecutableMapping.sfm_size;
-		}
-		
+				
 		// make read-only (contains LINKEDIT segments) last, so it can be cut back when optimized
 		shared_file_mapping_np  readOnlyMapping;
-		readOnlyMapping.sfm_address		= currentExecuteAddress;
-		readOnlyMapping.sfm_size		= currentReadOnlyAddress - currentExecuteAddress;
+		readOnlyMapping.sfm_address		= startReadOnlyAddress;
+		readOnlyMapping.sfm_size		= currentReadOnlyAddress - startReadOnlyAddress;
 		readOnlyMapping.sfm_file_offset	= cacheFileOffset;
 		readOnlyMapping.sfm_max_prot	= VM_PROT_READ;
 		readOnlyMapping.sfm_init_prot	= VM_PROT_READ;
@@ -1131,7 +1289,7 @@
 	else {
 		// empty cache
 		shared_file_mapping_np  cacheHeaderMapping;
-		cacheHeaderMapping.sfm_address		= sharedRegionWritableStartAddress();
+		cacheHeaderMapping.sfm_address		= startExecuteAddress;
 		cacheHeaderMapping.sfm_size			= FIRST_DYLIB_TEXT_OFFSET;
 		cacheHeaderMapping.sfm_file_offset	= cacheFileOffset;
 		cacheHeaderMapping.sfm_max_prot		= VM_PROT_READ;
@@ -1189,7 +1347,7 @@
 	strcpy(&temp[15-strlen(archPairName)], archPairName);
 	if ( strcmp(header->magic(), temp) != 0 ) {
 		if ( fVerify ) {
-			fprintf(stderr, "update_dyld_shared_cache[%u] cannot verify %s because current cache file has invalid header\n", getpid(), archName());
+			fprintf(stderr, "update_dyld_shared_cache[%u] cannot verify %s because current cache file has invalid header\n", getpid(), archPairName);
 			return false;
 		}
 		else {
@@ -1200,11 +1358,11 @@
 	// not valid if count of images does not match current images needed
 	if ( header->imagesCount() != (fDylibs.size()+aliasCount) ) {
 		if ( fVerify ) {
-			fprintf(stderr, "update_dyld_shared_cache[%u] cannot verify %s because current cache file contains a different set of dylibs\n", getpid(), archName());
+			fprintf(stderr, "update_dyld_shared_cache[%u] cannot verify %s because current cache file contains a different set of dylibs\n", getpid(), archPairName);
 			return false;
 		}
 		else {
-			fprintf(stderr, "update_dyld_shared_cache[%u] updating %s cache because current cache file contains a different set of dylibs\n", getpid(), archName());
+			fprintf(stderr, "update_dyld_shared_cache[%u] updating %s cache because current cache file contains a different set of dylibs\n", getpid(), archPairName);
 			return true;
 		}
 	}
@@ -1223,7 +1381,7 @@
 			if ( fVerify ) {
 				if ( cacheEntry->pathFileOffset() > textSize ) {
 					throwf("update_dyld_shared_cache[%u]: for arch=%s, image entries corrupt, bad path offset in %s\n", 
-								getpid(), archName(), it->layout->getID().name);
+								getpid(), archPairName, it->layout->getID().name);
 				}
 				// in -verify mode, just match by path and warn if file looks different
 				if ( strcmp((char*)cache+cacheEntry->pathFileOffset(), it->layout->getID().name) == 0 ) {
@@ -1231,7 +1389,7 @@
 					sortingMap[it->layout] = cacheEntry-imagesStart;
 					if ( (cacheEntry->inode() != it->info.inode) || (cacheEntry->modTime() != it->info.modTime) ) {
 						fprintf(stderr, "update_dyld_shared_cache[%u] warning: for arch=%s, %s has changed since cache was built\n", 
-								getpid(), archName(), it->layout->getID().name);
+								getpid(), archPairName, it->layout->getID().name);
 					}
 					break;
 				}
@@ -1252,10 +1410,10 @@
 		}
 		if ( !found ) {
 			if ( fVerify ) {
-				throwf("update_dyld_shared_cache[%u] can't verify %s cache because %s is not in existing cache\n", getpid(), archName(), it->layout->getID().name);
+				throwf("update_dyld_shared_cache[%u] can't verify %s cache because %s is not in existing cache\n", getpid(), archPairName, it->layout->getID().name);
 			}
 			else {
-				fprintf(stderr, "update_dyld_shared_cache[%u] updating %s cache because dylib at %s has changed\n", getpid(), archName(), it->layout->getID().name);
+				fprintf(stderr, "update_dyld_shared_cache[%u] updating %s cache because dylib at %s has changed\n", getpid(), archPairName, it->layout->getID().name);
 				return true;
 			}
 		}
@@ -1288,7 +1446,7 @@
 	struct stat stat_buf;
 	::fstat(fd, &stat_buf);
     uint32_t cacheFileSize = stat_buf.st_size;
-    uint32_t cacheAllocatedSize = (cacheFileSize + 4095) & (-4096);
+    uint32_t cacheAllocatedSize = pageAlign(cacheFileSize);
     uint8_t* mappingAddr = NULL;
 	if ( vm_allocate(mach_task_self(), (vm_address_t*)(&mappingAddr), cacheAllocatedSize, VM_FLAGS_ANYWHERE) != KERN_SUCCESS )
         throwf("can't vm_allocate cache of size %u", cacheFileSize);
@@ -1296,7 +1454,7 @@
  	(void)fcntl(fd, F_NOCACHE, 1);
     ssize_t readResult = pread(fd, mappingAddr, cacheFileSize, 0);
     if ( readResult != cacheFileSize )
-        throw "can't read existing cache file";
+        throwf("can't read all of existing cache file (%lu of %u): %s", readResult, cacheFileSize, path);
 	::close(fd);
 
 	// validate it
@@ -1314,76 +1472,6 @@
 	return result;
 }
 
-class CStringEquals
-{
-public:
-	bool operator()(const char* left, const char* right) const { return (strcmp(left, right) == 0); }
-};
-
-class StringPool
-{
-public:
-				StringPool();
-	const char*	getBuffer();
-	uint32_t	size();
-	uint32_t	add(const char* str);
-	uint32_t	addUnique(const char* str);
-	const char* stringAtIndex(uint32_t) const;
-private:
-	typedef __gnu_cxx::hash_map<const char*, uint32_t, __gnu_cxx::hash<const char*>, CStringEquals> StringToOffset;
-
-	char*			fBuffer;
-	uint32_t		fBufferAllocated;
-	uint32_t		fBufferUsed;
-	StringToOffset	fUniqueStrings;
-};
-
-
-StringPool::StringPool() 
-	: fBufferUsed(0), fBufferAllocated(32*1024*1024)
-{
-	fBuffer = (char*)malloc(fBufferAllocated);
-}
-
-uint32_t StringPool::add(const char* str)
-{
-	uint32_t len = strlen(str);
-	if ( (fBufferUsed + len + 1) > fBufferAllocated ) {
-		// grow buffer
-		throw "string buffer exhausted";
-	}
-	strcpy(&fBuffer[fBufferUsed], str);
-	uint32_t result = fBufferUsed;
-	fUniqueStrings[&fBuffer[fBufferUsed]] = result;
-	fBufferUsed += len+1;
-	return result;
-}
-
-uint32_t StringPool::addUnique(const char* str)
-{
-	StringToOffset::iterator pos = fUniqueStrings.find(str);
-	if ( pos != fUniqueStrings.end() ) 
-		return pos->second;
-	else {
-		//fprintf(stderr, "StringPool::addUnique() new string: %s\n", str);
-		return this->add(str);
-	}
-}
-
-uint32_t StringPool::size()
-{
-	return fBufferUsed;
-}
-
-const char*	StringPool::getBuffer()
-{
-	return fBuffer;
-}
-
-const char* StringPool::stringAtIndex(uint32_t index) const
-{
-	return &fBuffer[index];
-}
 
 
 template <typename A>
@@ -1397,12 +1485,16 @@
 		void								copyWeakBindInfo(uint32_t&);
 		void								copyLazyBindInfo(uint32_t&);
 		void								copyExportInfo(uint32_t&);
-		void								copyLocalSymbols(uint32_t symbolTableOffset, uint32_t&);
+		void								copyLocalSymbols(uint32_t symbolTableOffset, uint32_t&, bool dontMapLocalSymbols,
+															uint8_t* cacheStart, StringPool& unmappedLocalsStringPool, 
+															std::vector<macho_nlist<typename A::P> >& unmappedSymbols,
+															std::vector<LocalSymbolInfo>& info);
 		void								copyExportedSymbols(uint32_t symbolTableOffset, uint32_t&);
 		void								copyImportedSymbols(uint32_t symbolTableOffset, uint32_t&);
 		void								copyExternalRelocations(uint32_t& offset);
 		void								copyIndirectSymbolTable(uint32_t& offset);
 		void								copyFunctionStarts(uint32_t& offset);
+		void								copyDataInCode(uint32_t& offset);
 		void								updateLoadCommands(uint64_t newVMAddress, uint64_t size, uint32_t stringPoolOffset, 
 																uint32_t linkEditsFileOffset, bool keepSignatures);
 	
@@ -1422,6 +1514,7 @@
 	macho_dyld_info_command<P>*					fDyldInfo;
 	macho_dysymtab_command<P>*					fDynamicSymbolTable;
 	macho_linkedit_data_command<P>*				fFunctionStarts;
+	macho_linkedit_data_command<P>*				fDataInCode;
 	macho_symtab_command<P>*					fSymbolTableLoadCommand;
 	const macho_nlist<P>*						fSymbolTable;
 	const char*									fStrings;
@@ -1445,6 +1538,9 @@
 	uint32_t									fExternalRelocationsOffsetIntoNewLinkEdit;
 	uint32_t									fIndirectSymbolTableOffsetInfoNewLinkEdit;
 	uint32_t									fFunctionStartsOffsetInNewLinkEdit;
+	uint32_t									fDataInCodeOffsetInNewLinkEdit;
+	uint32_t									fUnmappedLocalSymbolsStartIndexInNewLinkEdit;
+	uint32_t									fUnmappedLocalSymbolsCountInNewLinkEdit;
 };
 
 
@@ -1452,7 +1548,8 @@
 template <typename A>
 LinkEditOptimizer<A>::LinkEditOptimizer(const MachOLayoutAbstraction& layout, const SharedCache<A>& sharedCache, uint8_t* newLinkEdit, StringPool& stringPool)
  : 	fSharedCache(sharedCache), fLayout(layout), fLinkEditBase(NULL), fNewLinkEditStart(newLinkEdit), fDyldInfo(NULL),
-	fDynamicSymbolTable(NULL), fFunctionStarts(NULL), fSymbolTableLoadCommand(NULL), fSymbolTable(NULL), fStrings(NULL), fNewStringPool(stringPool),
+	fDynamicSymbolTable(NULL), fFunctionStarts(NULL), fDataInCode(NULL), 
+	fSymbolTableLoadCommand(NULL), fSymbolTable(NULL), fStrings(NULL), fNewStringPool(stringPool),
 	fBindInfoOffsetIntoNewLinkEdit(0), fBindInfoSizeInNewLinkEdit(0),
 	fWeakBindInfoOffsetIntoNewLinkEdit(0), fWeakBindInfoSizeInNewLinkEdit(0),
 	fLazyBindInfoOffsetIntoNewLinkEdit(0), fLazyBindInfoSizeInNewLinkEdit(0),
@@ -1462,7 +1559,8 @@
 	fExportedSymbolsStartIndexInNewLinkEdit(0), fExportedSymbolsCountInNewLinkEdit(0),
 	fImportSymbolsStartIndexInNewLinkEdit(0), fImportedSymbolsCountInNewLinkEdit(0),
 	fExternalRelocationsOffsetIntoNewLinkEdit(0), fIndirectSymbolTableOffsetInfoNewLinkEdit(0),
-	fFunctionStartsOffsetInNewLinkEdit(0)
+	fFunctionStartsOffsetInNewLinkEdit(0), fDataInCodeOffsetInNewLinkEdit(0),
+	fUnmappedLocalSymbolsStartIndexInNewLinkEdit(0), fUnmappedLocalSymbolsCountInNewLinkEdit(0)
 	
 {
 	fHeader = (const macho_header<P>*)fLayout.getSegments()[0].mappedAddress();
@@ -1497,6 +1595,8 @@
 				break;
 			case LC_FUNCTION_STARTS:
 				fFunctionStarts = (macho_linkedit_data_command<P>*)cmd;
+			case LC_DATA_IN_CODE:
+				fDataInCode = (macho_linkedit_data_command<P>*)cmd;
 				break;
 		}
 		cmd = (const macho_load_command<P>*)(((uint8_t*)cmd)+cmd->cmdsize());
@@ -1569,25 +1669,46 @@
 }
 
 
-
 template <typename A>
-void LinkEditOptimizer<A>::copyLocalSymbols(uint32_t symbolTableOffset, uint32_t& symbolIndex)
+void LinkEditOptimizer<A>::copyLocalSymbols(uint32_t symbolTableOffset, uint32_t& symbolIndex, bool dontMapLocalSymbols, uint8_t* cacheStart, 
+											StringPool&	unmappedLocalsStringPool, std::vector<macho_nlist<P> >& unmappedSymbols,
+											std::vector<LocalSymbolInfo>& dylibInfos)
 {
 	fLocalSymbolsStartIndexInNewLinkEdit = symbolIndex;
+	LocalSymbolInfo localInfo;
+	localInfo.dylibOffset = ((uint8_t*)fHeader) - cacheStart;
+	localInfo.nlistStartIndex = unmappedSymbols.size();
+	localInfo.nlistCount = 0;
 	fSymbolTableStartOffsetInNewLinkEdit = symbolTableOffset + symbolIndex*sizeof(macho_nlist<P>);
 	macho_nlist<P>* const newSymbolTableStart = (macho_nlist<P>*)(fNewLinkEditStart+symbolTableOffset);
 	const macho_nlist<P>* const firstLocal = &fSymbolTable[fDynamicSymbolTable->ilocalsym()];
 	const macho_nlist<P>* const lastLocal  = &fSymbolTable[fDynamicSymbolTable->ilocalsym()+fDynamicSymbolTable->nlocalsym()];
 	uint32_t oldIndex = fDynamicSymbolTable->ilocalsym();
 	for (const macho_nlist<P>* entry = firstLocal; entry < lastLocal; ++entry, ++oldIndex) {
-		if ( (entry->n_type() & N_TYPE) == N_SECT ) {
+		// <rdar://problem/12237639> don't copy stab symbols
+		if ( (entry->n_sect() != NO_SECT) && ((entry->n_type() & N_STAB) == 0) ) {
+			const char* name = &fStrings[entry->n_strx()];
 			macho_nlist<P>* newSymbolEntry = &newSymbolTableStart[symbolIndex];
 			*newSymbolEntry = *entry;
-			newSymbolEntry->set_n_strx(fNewStringPool.addUnique(&fStrings[entry->n_strx()]));
-			++symbolIndex;
+			if ( dontMapLocalSymbols ) {
+				// if local in __text, add <redacted> symbol name to shared cache so backtraces don't have bogus names
+				if ( entry->n_sect() == 1 ) {
+					newSymbolEntry->set_n_strx(fNewStringPool.addUnique("<redacted>"));
+					++symbolIndex;
+				}
+				// copy local symbol to unmmapped locals area
+				unmappedSymbols.push_back(*entry);			
+				unmappedSymbols.back().set_n_strx(unmappedLocalsStringPool.addUnique(name));
+			}
+			else {
+				newSymbolEntry->set_n_strx(fNewStringPool.addUnique(name));
+				++symbolIndex;
+			}
 		}
 	}
 	fLocalSymbolsCountInNewLinkEdit = symbolIndex - fLocalSymbolsStartIndexInNewLinkEdit;
+	localInfo.nlistCount = unmappedSymbols.size() - localInfo.nlistStartIndex;
+	dylibInfos.push_back(localInfo);
 	//fprintf(stderr, "%u locals starting at %u for %s\n", fLocalSymbolsCountInNewLinkEdit, fLocalSymbolsStartIndexInNewLinkEdit, fLayout.getFilePath());
 }
 
@@ -1675,6 +1796,17 @@
 }
 
 template <typename A>
+void LinkEditOptimizer<A>::copyDataInCode(uint32_t& offset)
+{	
+	if ( fDataInCode != NULL ) {
+		fDataInCodeOffsetInNewLinkEdit = offset;
+		memcpy(&fNewLinkEditStart[offset], &fLinkEditBase[fDataInCode->dataoff()], fDataInCode->datasize());
+		offset += fDataInCode->datasize();
+	}
+}
+
+
+template <typename A>
 void LinkEditOptimizer<A>::copyIndirectSymbolTable(uint32_t& offset)
 {	
 	fIndirectSymbolTableOffsetInfoNewLinkEdit = offset;
@@ -1694,7 +1826,7 @@
 }
 
 template <typename A>
-void LinkEditOptimizer<A>::updateLoadCommands(uint64_t newVMAddress, uint64_t size, uint32_t stringPoolOffset, 
+void LinkEditOptimizer<A>::updateLoadCommands(uint64_t newVMAddress, uint64_t leSize, uint32_t stringPoolOffset, 
 												uint32_t linkEditsFileOffset, bool keepSignatures)
 {
 	// set LINKEDIT segment commmand to new merged LINKEDIT
@@ -1706,16 +1838,31 @@
 			macho_segment_command<P>* seg = (macho_segment_command<P>*)cmd;
 			if ( strcmp(seg->segname(), "__LINKEDIT") == 0 ) {
 				seg->set_vmaddr(newVMAddress);
-				seg->set_vmsize(size);
-				seg->set_filesize(size);
+				seg->set_vmsize(leSize);
+				seg->set_filesize(leSize);
 				seg->set_fileoff(linkEditsFileOffset);
 			}
-			// don't alter __TEXT until <rdar://problem/7022345> is fixed
-			else if ( strcmp(seg->segname(), "__TEXT") != 0 ) {
-				// update all other segments fileoff to be offset from start of cache file
+			else {
 				pint_t oldFileOff = seg->fileoff();
-				seg->set_fileoff(fSharedCache.cacheFileOffsetForVMAddress(seg->vmaddr()));
+				// don't alter __TEXT until <rdar://problem/7022345> is fixed
+				if ( strcmp(seg->segname(), "__TEXT") != 0 ) {
+					// update all other segments fileoff to be offset from start of cache file
+					seg->set_fileoff(fSharedCache.cacheFileOffsetForVMAddress(seg->vmaddr()));
+				}
 				pint_t fileOffsetDelta = seg->fileoff() - oldFileOff;
+				const MachOLayoutAbstraction::Segment* layoutSeg = fLayout.getSegment(seg->segname());
+				if ( layoutSeg != NULL ) {
+					//if ( seg->filesize() != layoutSeg->fileSize() ) {
+					//	fprintf(stderr, "LC filesize=0x%08llX, trimmed seg file size=0x%08llX, seg=%s, path=%s\n", 
+					//					seg->filesize(), layoutSeg->fileSize(), seg->segname(), fLayout.getFilePath());
+					//}
+					//if ( seg->vmsize() != layoutSeg->size() ) {
+					//	fprintf(stderr, "LC   vmsize=0x%08llX, trimmed seg      size=0x%08llX, seg=%s, path=%s\n", 
+					//					seg->vmsize(), layoutSeg->size(), seg->segname(), fLayout.getFilePath());
+					//}
+					seg->set_vmsize(layoutSeg->size());
+					seg->set_filesize(layoutSeg->fileSize());
+				}
 				// update all sections in this segment
 				macho_section<P>* const sectionsStart = (macho_section<P>*)((char*)seg + sizeof(macho_segment_command<P>));
 				macho_section<P>* const sectionsEnd = &sectionsStart[seg->nsects()];
@@ -1772,6 +1919,10 @@
 	if ( fFunctionStarts != NULL ) {
 		fFunctionStarts->set_dataoff(linkEditsFileOffset+fFunctionStartsOffsetInNewLinkEdit);
 	}
+	// update data-in-code info
+	if ( fDataInCode != NULL ) {
+		fDataInCode->set_dataoff(linkEditsFileOffset+fDataInCodeOffsetInNewLinkEdit);
+	}
 	
 	// now remove load commands no longer needed
 	const macho_load_command<P>* srcCmd = cmds;
@@ -1781,6 +1932,7 @@
 		uint32_t cmdSize = srcCmd->cmdsize();
 		switch ( srcCmd->cmd() ) {
 			case LC_SEGMENT_SPLIT_INFO:
+			case LC_DYLIB_CODE_SIGN_DRS:
 				// don't copy
 				break;
 			case LC_CODE_SIGNATURE:
@@ -1814,7 +1966,7 @@
 
 
 template <typename A>
-uint8_t* SharedCache<A>::optimizeLINKEDIT(bool keepSignatures)
+uint8_t* SharedCache<A>::optimizeLINKEDIT(bool keepSignatures, bool dontMapLocalSymbols)
 {
 	// allocate space for optimized LINKEDIT area
 	uint8_t* newLinkEdit = new uint8_t[fLinkEditsTotalUnoptimizedSize];
@@ -1860,8 +2012,11 @@
 	fOffsetOfOldSymbolTableInfoInCombinedLinkedit = offset;
 	uint32_t symbolTableOffset = offset;
 	uint32_t symbolTableIndex = 0;
+	if ( dontMapLocalSymbols ) 
+		fUnmappedLocalSymbols.reserve(16384);
 	for(typename std::vector<LinkEditOptimizer<A>*>::iterator it = optimizers.begin(); it != optimizers.end(); ++it) {
-		(*it)->copyLocalSymbols(symbolTableOffset, symbolTableIndex);
+		(*it)->copyLocalSymbols(symbolTableOffset, symbolTableIndex, dontMapLocalSymbols, fInMemoryCache,
+								fUnmappedLocalsStringPool, fUnmappedLocalSymbols, fLocalSymbolInfos);
 		(*it)->copyExportedSymbols(symbolTableOffset, symbolTableIndex);
 		(*it)->copyImportedSymbols(symbolTableOffset, symbolTableIndex);
 	}
@@ -1881,6 +2036,13 @@
 		(*it)->copyFunctionStarts(offset);
 	}
 	fSizeOfFunctionStartsInCombinedLinkedit = offset - fOffsetOfFunctionStartsInCombinedLinkedit;
+
+	// copy data-in-code info
+	fOffsetOfDataInCodeInCombinedLinkedit = offset;
+	for(typename std::vector<LinkEditOptimizer<A>*>::iterator it = optimizers.begin(); it != optimizers.end(); ++it) {
+		(*it)->copyDataInCode(offset);
+	}
+	fSizeOfDataInCodeInCombinedLinkedit = offset - fOffsetOfDataInCodeInCombinedLinkedit;
 
 	// copy indirect symbol tables
 	fOffsetOfOldIndirectSymbolsInCombinedLinkedit = offset;
@@ -1895,15 +2057,15 @@
 	fSizeOfOldStringPoolInCombinedLinkedit = stringPool.size();
 	
 	// total new size round up to page size
-	fLinkEditsTotalOptimizedSize = (fOffsetOfOldStringPoolInCombinedLinkedit + fSizeOfOldStringPoolInCombinedLinkedit + 4095) & (-4096);
+	fLinkEditsTotalOptimizedSize = pageAlign(fOffsetOfOldStringPoolInCombinedLinkedit + fSizeOfOldStringPoolInCombinedLinkedit);
 	
 	// choose new linkedit file offset 
 	uint32_t linkEditsFileOffset = cacheFileOffsetForVMAddress(fLinkEditsStartAddress);
-//	uint32_t linkEditsFileOffset = fLinkEditsStartAddress - sharedRegionReadOnlyStartAddress();	
+//	uint32_t linkEditsFileOffset = fLinkEditsStartAddress - sharedRegionStartAddress();	
 
 	// update load commands so that all dylibs shared different areas of the same LINKEDIT segment
 	for(typename std::vector<LinkEditOptimizer<A>*>::iterator it = optimizers.begin(); it != optimizers.end(); ++it) {
-		(*it)->updateLoadCommands(fLinkEditsStartAddress, fLinkEditsTotalUnoptimizedSize, fOffsetOfOldStringPoolInCombinedLinkedit, linkEditsFileOffset, keepSignatures);
+		(*it)->updateLoadCommands(fLinkEditsStartAddress, fLinkEditsTotalOptimizedSize, fOffsetOfOldStringPoolInCombinedLinkedit, linkEditsFileOffset, keepSignatures);
 	}
 
 	//fprintf(stderr, "fLinkEditsTotalUnoptimizedSize=%llu, fLinkEditsTotalOptimizedSize=%u\n", fLinkEditsTotalUnoptimizedSize, fLinkEditsTotalOptimizedSize);
@@ -1925,12 +2087,11 @@
 				seg.setFileOffset(linkEditsFileOffset);
 			}
 		}
-	}
-	
+	}	
+		
 	// return new end of cache
-	return (uint8_t*)fFirstLinkEditSegment->mappedAddress() + fLinkEditsTotalOptimizedSize;
-}
-
+	return (uint8_t*)fFirstLinkEditSegment->mappedAddress() + regionAlign(fLinkEditsTotalOptimizedSize);
+}
 
 
 template <typename A>
@@ -1966,14 +2127,71 @@
     size_t count() const { return fCount; }
 };
 
+
+template <typename A>
+class ClassListBuilder
+{
+private:
+    typedef typename A::P P;
+
+    objc_opt::string_map fClassNames;
+    objc_opt::class_map fClasses;
+    size_t fCount;
+    HeaderInfoOptimizer<A>& fHinfos;
+
+public:
+
+    ClassListBuilder(HeaderInfoOptimizer<A>& hinfos)
+        : fClassNames()
+        , fClasses()
+        , fCount(0)
+        , fHinfos(hinfos)
+    { }
+
+    void visitClass(SharedCache<A>* cache, 
+                    const macho_header<P>* header,
+                    objc_class_t<A>* cls) 
+    {
+        if (cls->isMetaClass(cache)) return;
+
+        const char *name = cls->getName(cache);
+        uint64_t name_vmaddr = cache->VMAddressForMappedAddress(name);
+        uint64_t cls_vmaddr = cache->VMAddressForMappedAddress(cls);
+        uint64_t hinfo_vmaddr = cache->VMAddressForMappedAddress(fHinfos.hinfoForHeader(cache, header));
+        fClassNames.insert(objc_opt::string_map::value_type(name, name_vmaddr));
+        fClasses.insert(objc_opt::class_map::value_type(name, std::pair<uint64_t, uint64_t>(cls_vmaddr, hinfo_vmaddr)));
+        fCount++;
+    }
+
+    objc_opt::string_map& classNames() { 
+        return fClassNames;
+    }
+
+    objc_opt::class_map& classes() { 
+        return fClasses;
+    }
+
+    size_t count() const { return fCount; }
+};
+
+
+static int percent(size_t num, size_t denom) {
+    if (denom) return (int)(num / (double)denom * 100);
+    else return 100;
+}
+
 template <typename A>
 void SharedCache<A>::optimizeObjC(std::vector<void*>& pointersInData)
 {
     const char *err;
-    size_t headerSize = sizeof(objc_opt::objc_opt_t);
 
     if ( verbose ) {
         fprintf(stderr, "update_dyld_shared_cache: for %s, optimizing objc metadata\n", archName());
+    }
+
+    size_t headerSize = P::round_up(sizeof(objc_opt::objc_opt_t));
+    if (headerSize != sizeof(objc_opt::objc_opt_t)) {
+		warn(archName(), "libobjc's optimization structure size is wrong (metadata not optimized)");
     }
 
     // Find libobjc's empty sections to fill in
@@ -1983,9 +2201,6 @@
         if ( strstr(it->layout->getFilePath(), "libobjc") != NULL ) {
 			const macho_header<P>* mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
 			optROSection = mh->getSection("__TEXT", "__objc_opt_ro");
-			// __objc_selopt is old name for __objc_opt_ro
-			if ( optROSection == NULL )
-				optROSection = mh->getSection("__TEXT", "__objc_selopt");
 			optRWSection = mh->getSection("__DATA", "__objc_opt_rw");
 			break;
 		}
@@ -1996,17 +2211,62 @@
 		return;
 	}
 	
-	objc_opt::objc_opt_t* optROHeader = (objc_opt::objc_opt_t*)mappedAddressForVMAddress(optROSection->addr());
-	if (optROSection->size() < headerSize) {
+	if ( optRWSection == NULL ) {
+		warn(archName(), "libobjc's read/write section missing (metadata not optimized)");
+		return;
+	}
+
+	uint8_t* optROData = (uint8_t*)mappedAddressForVMAddress(optROSection->addr());
+    size_t optRORemaining = optROSection->size();
+
+	uint8_t* optRWData = (uint8_t*)mappedAddressForVMAddress(optRWSection->addr());
+    size_t optRWRemaining = optRWSection->size();
+	
+	if (optRORemaining < headerSize) {
 		warn(archName(), "libobjc's read-only section is too small (metadata not optimized)");
 		return;
 	}
+	objc_opt::objc_opt_t* optROHeader = (objc_opt::objc_opt_t *)optROData;
+    optROData += headerSize;
+    optRORemaining -= headerSize;
 
 	if (E::get32(optROHeader->version) != objc_opt::VERSION) {
 		warn(archName(), "libobjc's read-only section version is unrecognized (metadata not optimized)");
 		return;
 	}
 
+    // Write nothing to optROHeader until everything else is written.
+    // If something fails below, libobjc will not use the section.
+
+    // Find objc-containing dylibs
+    std::vector<LayoutInfo> objcDylibs;
+    for(typename std::vector<LayoutInfo>::const_iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
+        macho_header<P> *mh = (macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        if (mh->getSection("__DATA", "__objc_imageinfo")  ||  mh->getSegment("__OBJC")) {
+            objcDylibs.push_back(*it);
+        }
+    }
+
+    // Build image list
+
+    // This is SAFE: the binaries themselves are unmodified.
+
+    std::vector<LayoutInfo> addressSortedDylibs = objcDylibs;
+    std::sort(addressSortedDylibs.begin(), addressSortedDylibs.end(), ByAddressSorter());
+
+    uint64_t hinfoVMAddr = optRWSection->addr() + optRWSection->size() - optRWRemaining;
+    HeaderInfoOptimizer<A> hinfoOptimizer;
+    err = hinfoOptimizer.init(objcDylibs.size(), optRWData, optRWRemaining);
+    if (err) {
+		warn(archName(), err);
+		return;
+    }
+    for(typename std::vector<LayoutInfo>::const_iterator it = addressSortedDylibs.begin(); it != addressSortedDylibs.end(); ++it) {
+        const macho_header<P> *mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        hinfoOptimizer.update(this, mh, pointersInData);
+    }
+
+
     // Update selector references and build selector list
 
     // This is SAFE: if we run out of room for the selector table, 
@@ -2015,101 +2275,127 @@
     // Heuristic: choose selectors from libraries with more cstring data first.
     // This tries to localize selector cstring memory.
     ObjCSelectorUniquer<A> uniq(this);
-    std::vector<LayoutInfo> sortedDylibs = fDylibs;
-    std::sort(sortedDylibs.begin(), sortedDylibs.end(), ByCStringSectionSizeSorter());
+    std::vector<LayoutInfo> sizeSortedDylibs = objcDylibs;
+    std::sort(sizeSortedDylibs.begin(), sizeSortedDylibs.end(), ByCStringSectionSizeSorter());
 
     SelectorOptimizer<A, ObjCSelectorUniquer<A> > selOptimizer(uniq);
-	for(typename std::vector<LayoutInfo>::const_iterator it = sortedDylibs.begin(); it != sortedDylibs.end(); ++it) {
+	for(typename std::vector<LayoutInfo>::const_iterator it = sizeSortedDylibs.begin(); it != sizeSortedDylibs.end(); ++it) {
         const macho_header<P> *mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
         LegacySelectorUpdater<A, ObjCSelectorUniquer<A> >::update(this, mh, uniq);
         selOptimizer.optimize(this, mh);
 	}
 
-    if ( verbose ) {
-        fprintf(stderr, "update_dyld_shared_cache: for %s, found %zu unique objc selectors\n", archName(), uniq.strings().size());
-    }
-
-    // Write selector table in read-only data.
-    size_t selTableOffset = P::round_up(headerSize);
-    size_t selTableSize;
-    objc_opt::objc_selopt_t *seloptData = (objc_opt::objc_selopt_t *)
-        mappedAddressForVMAddress(optROSection->addr() + selTableOffset);
-    err = objc_opt::write_selopt(seloptData, 
-                                 optROSection->addr() + selTableOffset, 
-                                 optROSection->size() - selTableOffset, 
-                                 uniq.strings(), 
-                                 E::little_endian, &selTableSize);
+    uint64_t seloptVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
+    objc_opt::objc_selopt_t *selopt = new(optROData) objc_opt::objc_selopt_t;
+    err = selopt->write(seloptVMAddr, optRORemaining, uniq.strings());
     if (err) {
         warn(archName(), err);
         return;
     }
+    optROData += selopt->size();
+    optRORemaining -= selopt->size();
+    selopt->byteswap(E::little_endian), selopt = NULL;
+
+
+    // Build class table.
+
+    // This is SAFE: the binaries themselves are unmodified.
+
+    ClassListBuilder<A> classes(hinfoOptimizer);
+    ClassWalker< A, ClassListBuilder<A> > classWalker(classes);
+	for(typename std::vector<LayoutInfo>::const_iterator it = sizeSortedDylibs.begin(); it != sizeSortedDylibs.end(); ++it) {
+        const macho_header<P> *mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        classWalker.walk(this, mh);
+	}
+
+    uint64_t clsoptVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
+    objc_opt::objc_clsopt_t *clsopt = new(optROData) objc_opt::objc_clsopt_t;
+    err = clsopt->write(clsoptVMAddr, optRORemaining, 
+                        classes.classNames(), classes.classes(), verbose);
+    if (err) {
+        warn(archName(), err);
+        return;
+    }
+    optROData += clsopt->size();
+    optRORemaining -= clsopt->size();
+    size_t duplicateCount = clsopt->duplicateCount();
+    clsopt->byteswap(E::little_endian), clsopt = NULL;
+
+
+    // Sort method lists.
+
+    // This is SAFE: modified binaries are still usable as unsorted lists.
+    // This must be done AFTER uniquing selectors.
+
+    MethodListSorter<A> methodSorter;
+    for(typename std::vector<LayoutInfo>::const_iterator it = sizeSortedDylibs.begin(); it != sizeSortedDylibs.end(); ++it) {
+        macho_header<P> *mh = (macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        methodSorter.optimize(this, mh);
+    }
+
+
+    // Repair ivar offsets.
+
+    // This is SAFE: the runtime always validates ivar offsets at runtime.
+
+    IvarOffsetOptimizer<A> ivarOffsetOptimizer;
+	for(typename std::vector<LayoutInfo>::const_iterator it = sizeSortedDylibs.begin(); it != sizeSortedDylibs.end(); ++it) {
+        const macho_header<P> *mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        ivarOffsetOptimizer.optimize(this, mh);
+	}
+    
+
+    // Success. Mark dylibs as optimized.
+	for(typename std::vector<LayoutInfo>::const_iterator it = sizeSortedDylibs.begin(); it != sizeSortedDylibs.end(); ++it) {
+        const macho_header<P> *mh = (const macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
+        const macho_section<P> *imageInfoSection;
+        imageInfoSection = mh->getSection("__DATA", "__objc_imageinfo");
+        if (!imageInfoSection) {
+            imageInfoSection = mh->getSection("__OBJC", "__image_info");
+        }
+        if (imageInfoSection) {
+            objc_image_info<A> *info = (objc_image_info<A> *)
+                mappedAddressForVMAddress(imageInfoSection->addr());
+            info->setOptimizedByDyld();
+        }
+    }
+
+
+    // Success. Update RO header last.
+    E::set32(optROHeader->selopt_offset, seloptVMAddr - optROSection->addr());
+    E::set32(optROHeader->clsopt_offset, clsoptVMAddr - optROSection->addr());
+    E::set32(optROHeader->headeropt_offset, hinfoVMAddr - optROSection->addr());
 
     if ( verbose ) {
-        size_t totalSize = headerSize + selTableSize;
+        size_t roSize = optROSection->size() - optRORemaining;
+        size_t rwSize = optRWSection->size() - optRWRemaining;
         fprintf(stderr, "update_dyld_shared_cache: for %s, %zu/%llu bytes "
                 "(%d%%) used in libobjc read-only optimization section\n", 
-                archName(), totalSize, optROSection->size(), 
-                (int)(totalSize / (double)optROSection->size() * 100));
+                archName(), roSize, optROSection->size(), 
+                percent(roSize, optROSection->size()));
+        fprintf(stderr, "update_dyld_shared_cache: for %s, %zu/%llu bytes "
+                "(%d%%) used in libobjc read/write optimization section\n", 
+                archName(), rwSize, optRWSection->size(), 
+                percent(rwSize, optRWSection->size()));
+        fprintf(stderr, "update_dyld_shared_cache: for %s, "
+                "uniqued %zu selectors\n", 
+                archName(), uniq.strings().size());
         fprintf(stderr, "update_dyld_shared_cache: for %s, "
                 "updated %zu selector references\n", 
                 archName(), uniq.count());
         fprintf(stderr, "update_dyld_shared_cache: for %s, "
+                "updated %zu ivar offsets\n", 
+                archName(), ivarOffsetOptimizer.optimized());
+        fprintf(stderr, "update_dyld_shared_cache: for %s, "
+                "sorted %zu method lists\n", 
+                archName(), methodSorter.optimized());
+        fprintf(stderr, "update_dyld_shared_cache: for %s, "
+                "recorded %zu classes (%zu duplicates)\n", 
+                archName(), classes.classNames().size(), duplicateCount);
+        fprintf(stderr, "update_dyld_shared_cache: for %s, "
                 "wrote objc metadata optimization version %d\n", 
                 archName(), objc_opt::VERSION);
     }
-
-	// if r/w section exists in libojc attempt to optimize categories into classes
-	if ( optRWSection != NULL ) {
-		// Attach categories to classes in the same framework. 
-		// Build aggregated (but unsorted) method lists in read-write data.
-
-		// This is SAFE: if we run out of room while attaching categories in 
-		// a binary then previously-edited binaries are still valid. (This assumes 
-		// each binary is processed all-or-nothing, which CategoryAttacher does.)
-		// This must be done AFTER uniquing selectors.
-		// This must be done BEFORE sorting method lists.
-		
-		size_t categoryOffset = 0;
-		uint8_t *categoryData = (uint8_t*)mappedAddressForVMAddress(optRWSection->addr() + categoryOffset);
-		CategoryAttacher<A> categoryAttacher(categoryData, optRWSection->size() - categoryOffset);
-		for(typename std::vector<LayoutInfo>::const_iterator it = sortedDylibs.begin(); it != sortedDylibs.end(); ++it) {
-			macho_header<P> *mh = (macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
-			err = categoryAttacher.optimize(this, mh, pointersInData);
-			if (err) {
-				warn(archName(), err);
-				return;
-			}
-		}    
-		size_t categorySize = categoryAttacher.bytesUsed();
-
-
-		// Sort method lists.
-		
-		// This is SAFE: modified binaries are still usable as unsorted lists.
-		// This must be done AFTER uniquing selectors.
-		// This must be done AFTER attaching categories.
-
-		MethodListSorter<A> methodSorter;
-		for(typename std::vector<LayoutInfo>::const_iterator it = sortedDylibs.begin(); it != sortedDylibs.end(); ++it) {
-			macho_header<P> *mh = (macho_header<P>*)(*it->layout).getSegments()[0].mappedAddress();
-			methodSorter.optimize(this, mh);
-		}
-
-		if ( verbose ) {
-			size_t totalRWSize = categorySize;
-			fprintf(stderr, "update_dyld_shared_cache: for %s, %zu/%llu bytes "
-                "(%d%%) used in libobjc read-write optimization section\n", 
-                archName(), totalRWSize, optRWSection->size(), 
-                (int)(totalRWSize / (double)optRWSection->size() * 100));
-			fprintf(stderr, "update_dyld_shared_cache: for %s, "
-                "attached %zu categories (%zd bytes used)\n", 
-                archName(), categoryAttacher.count(), 
-                categoryAttacher.bytesUsed());
-		}
-	}
-
-    // Success. Update RO header last
-    E::set32(optROHeader->selopt_offset, headerSize);
 
     return;
 }
@@ -2128,17 +2414,79 @@
 }
 
 
+// <rdar://problem/10730767> update_dyld_shared_cache should use sync_volume_np() instead of sync() 
+static void sync_volume(const char* volumePath)
+{
+#if __MAC_OS_X_VERSION_MIN_REQUIRED >= 1080
+	int error = sync_volume_np(volumePath, SYNC_VOLUME_FULLSYNC|SYNC_VOLUME_FULLSYNC);
+#else
+	int full_sync = 3; // SYNC_VOLUME_FULLSYNC | SYNC_VOLUME_FULLSYNC
+	int error = 0;
+	if ( fsctl(volumePath, 0x80004101 /*FSCTL_SYNC_VOLUME*/, &full_sync, 0) == -1) 
+		error = errno;
+#endif
+	if ( error )
+		::sync();
+}
+
+
+// <rdar://problem/12552226> update shared cache should sign the shared cache
+static bool adhoc_codesign_share_cache(const char* path)
+{
+	CFURLRef target = ::CFURLCreateFromFileSystemRepresentation(NULL, (const UInt8 *)path, strlen(path), FALSE);
+	if ( target == NULL )
+		return false;
+
+	SecStaticCodeRef code;
+	OSStatus status = ::SecStaticCodeCreateWithPath(target, kSecCSDefaultFlags, &code);
+	CFRelease(target);
+	if ( status ) {
+		::fprintf(stderr, "codesign: failed to create url to signed object\n");
+		return false;
+	}
+
+	const void * keys[1] = { (void *)kSecCodeSignerIdentity } ;
+	const void * values[1] = { (void *)kCFNull };
+	CFDictionaryRef params = ::CFDictionaryCreate(NULL, keys, values, 1, &kCFTypeDictionaryKeyCallBacks, &kCFTypeDictionaryValueCallBacks);
+	if ( params == NULL ) {
+		CFRelease(code);
+		return false;
+	}
+	
+	SecCodeSignerRef signer;
+	status = ::SecCodeSignerCreate(params, kSecCSDefaultFlags, &signer);
+	CFRelease(params);
+	if ( status ) {
+		CFRelease(code);
+		::fprintf(stderr, "codesign: failed to create signer object\n");
+		return false;
+	}
+
+	status = ::SecCodeSignerAddSignatureWithErrors(signer, code, kSecCSDefaultFlags, NULL);
+	CFRelease(code);
+	CFRelease(signer);
+	if ( status ) {
+		::fprintf(stderr, "codesign: failed to sign object: %s\n", path);
+		return false;
+	}
+
+	if ( verbose )
+		::fprintf(stderr, "codesigning complete of %s\n", path);
+	
+	return true;
+}
+
+
 
 template <>	 bool	SharedCache<x86_64>::addCacheSlideInfo(){ return true; }
 template <>	 bool	SharedCache<arm>::addCacheSlideInfo()	{ return true; }
 template <>	 bool	SharedCache<x86>::addCacheSlideInfo()	{ return false; }
-template <>	 bool	SharedCache<ppc>::addCacheSlideInfo()	{ return false; }
-
+template <>	 bool	SharedCache<arm64>::addCacheSlideInfo()	{ return true; } 
 
 
 template <typename A>
-bool SharedCache<A>::update(bool usesOverlay, bool force, bool optimize, bool deleteExistingFirst, int archIndex,
-								int archCount, bool keepSignatures)
+bool SharedCache<A>::update(bool force, bool optimize, bool deleteExistingFirst, int archIndex,
+								int archCount, bool keepSignatures, bool dontMapLocalSymbols)
 {
 	bool didUpdate = false;
 	
@@ -2188,6 +2536,8 @@
 			header->set_codeSignatureSize(0);
 			header->set_slideInfoOffset(0);
 			header->set_slideInfoSize(0);
+			header->set_localSymbolsOffset(0);
+			header->set_localSymbolsSize(0);
 			
 			// fill in mappings
 			dyldCacheFileMapping<E>* mapping = (dyldCacheFileMapping<E>*)&inMemoryCache[sizeof(dyldCacheHeader<E>)];
@@ -2239,11 +2589,13 @@
 				struct stat stat_buf;
 				if ( fstat(src, &stat_buf) == -1)
 					throwf("can't stat open file %s, errno=%d", path, errno);
-				if ( (it->layout->getInode() != stat_buf.st_ino) || (it->layout->getLastModTime() != stat_buf.st_mtime) )
-					throwf("file modified during cache creation: %s", path);
+				if ( (it->layout->getInode() != stat_buf.st_ino) )
+					throwf("file inode changed from %llu to %llu during cache creation: %s", it->layout->getInode(), stat_buf.st_ino, path);
+				else if ( it->layout->getLastModTime() != stat_buf.st_mtime )
+					throwf("file mtime changed from 0x%lX to 0x%lX during cache creation: %s", it->layout->getLastModTime(), stat_buf.st_mtime, path);
 
 				if ( verbose )
-					fprintf(stderr, "update_dyld_shared_cache: copying %s to cache\n", it->layout->getID().name);
+					fprintf(stderr, "update_dyld_shared_cache: copying %s to cache\n", it->layout->getFilePath());
 				try {
 					const std::vector<MachOLayoutAbstraction::Segment>& segs = it->layout->getSegments();
 					for (int i=0; i < segs.size(); ++i) {
@@ -2293,10 +2645,6 @@
 			std::vector<void*> pointersInData;
 			pointersInData.reserve(1024);
 			
-			// add pointer in start of __DATA to start of __TEXT to remain compatible with previous dylds
-			pint_t* dataStartPtr = (pint_t*)(&inMemoryCache[fMappings[1].sfm_file_offset]);
-			P::setP(*dataStartPtr, fMappings[0].sfm_address);
-
 			// rebase each dylib in shared cache
 			for(typename std::vector<LayoutInfo>::const_iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
 				try {
@@ -2356,8 +2704,8 @@
 			// merge/optimize all LINKEDIT segments
 			if ( optimize ) {
 				//fprintf(stderr, "update_dyld_shared_cache: original cache file size %uMB\n", cacheFileSize/(1024*1024));
-				cacheFileSize = (this->optimizeLINKEDIT(keepSignatures) - inMemoryCache);
-				//fprintf(stderr, "update_dyld_shared_cache: optimized cache file size %uMB\n", cacheFileSize/(1024*1024));
+				cacheFileSize = (this->optimizeLINKEDIT(keepSignatures, dontMapLocalSymbols) - inMemoryCache);
+				//fprintf(stderr, "update_dyld_shared_cache: optimized cache file size 0x%08X %uMB\n", cacheFileSize, cacheFileSize/(1024*1024));
 				// update header to reduce mapping size
 				dyldCacheHeader<E>* cacheHeader = (dyldCacheHeader<E>*)inMemoryCache;
 				dyldCacheFileMapping<E>* mappings = (dyldCacheFileMapping<E>*)&inMemoryCache[sizeof(dyldCacheHeader<E>)];
@@ -2437,7 +2785,7 @@
 				}
 				slideInfo->set_entries_count(entry_count);
 	
-				int slideInfoPageSize = (slideInfo->entries_offset() + entry_count*entry_size + 4095) & (-4096);
+				int slideInfoPageSize = regionAlign(slideInfo->entries_offset() + entry_count*entry_size);
 				cacheFileSize += slideInfoPageSize;
 			
 				// update mappings to increase RO size
@@ -2458,6 +2806,72 @@
 				memcpy(&inMemoryCache[cacheHeader->slideInfoOffset()], slideInfo, slideInfoPageSize);	
 			}
 			
+			// append local symbol info in an unmapped region
+			if ( dontMapLocalSymbols ) {
+				uint32_t spaceAtEnd = allocatedCacheSize - cacheFileSize;
+				uint32_t localSymbolsOffset = pageAlign(cacheFileSize);
+				dyldCacheLocalSymbolsInfo<E>* infoHeader = (dyldCacheLocalSymbolsInfo<E>*)(&inMemoryCache[localSymbolsOffset]);
+				const uint32_t entriesOffset = sizeof(dyldCacheLocalSymbolsInfo<E>);
+				const uint32_t entriesCount = fLocalSymbolInfos.size();
+				const uint32_t nlistOffset = entriesOffset + entriesCount * sizeof(dyldCacheLocalSymbolEntry<E>);
+				const uint32_t nlistCount = fUnmappedLocalSymbols.size();
+				const uint32_t stringsOffset = nlistOffset + nlistCount * sizeof(macho_nlist<P>);
+				const uint32_t stringsSize = fUnmappedLocalsStringPool.size();
+				if ( stringsOffset+stringsSize > spaceAtEnd ) 
+					throwf("update_dyld_shared_cache[%u] for arch=%s, out of space for local symbols. Have 0x%X, Need 0x%X\n",
+							getpid(), fArchGraph->archName(), spaceAtEnd, stringsOffset+stringsSize);
+				// fill in local symbols info
+				infoHeader->set_nlistOffset(nlistOffset);
+				infoHeader->set_nlistCount(nlistCount);
+				infoHeader->set_stringsOffset(stringsOffset);
+				infoHeader->set_stringsSize(stringsSize);
+				infoHeader->set_entriesOffset(entriesOffset);
+				infoHeader->set_entriesCount(entriesCount);
+				// copy info for each dylib
+				dyldCacheLocalSymbolEntry<E>* entries = (dyldCacheLocalSymbolEntry<E>*)(&inMemoryCache[localSymbolsOffset+entriesOffset]);
+				for (int i=0; i < entriesCount; ++i) {
+					entries[i].set_dylibOffset(fLocalSymbolInfos[i].dylibOffset);
+					entries[i].set_nlistStartIndex(fLocalSymbolInfos[i].nlistStartIndex);
+					entries[i].set_nlistCount(fLocalSymbolInfos[i].nlistCount);
+				}
+				// copy nlists
+				memcpy(&inMemoryCache[localSymbolsOffset+nlistOffset], &fUnmappedLocalSymbols[0], nlistCount*sizeof(macho_nlist<P>));
+				// copy string pool
+				memcpy(&inMemoryCache[localSymbolsOffset+stringsOffset], fUnmappedLocalsStringPool.getBuffer(), stringsSize);
+				
+				// update state
+				fUnmappedLocalSymbolsSize = pageAlign(stringsOffset + stringsSize);
+				cacheFileSize = regionAlign(localSymbolsOffset + fUnmappedLocalSymbolsSize);
+				
+				// update header to show location of slidePointers
+				dyldCacheHeader<E>* cacheHeader = (dyldCacheHeader<E>*)inMemoryCache;
+				cacheHeader->set_localSymbolsOffset(localSymbolsOffset);
+				cacheHeader->set_localSymbolsSize(stringsOffset+stringsSize);
+				cacheHeader->set_codeSignatureOffset(cacheFileSize);
+			}
+			
+			// make sure after all optimizations, that whole cache file fits into shared region address range
+			{
+				dyldCacheHeader<E>* cacheHeader = (dyldCacheHeader<E>*)inMemoryCache;
+				dyldCacheFileMapping<E>* mappings = (dyldCacheFileMapping<E>*)&inMemoryCache[cacheHeader->mappingOffset()];
+				// <rdar://problem/16128830> incorporate code signature size into overflow check
+				uint32_t estCodeSigSize = regionAlign(cacheFileSize/200); // guess 0.5% for code signature
+				for (int i=0; i < cacheHeader->mappingCount(); ++i) {
+					uint64_t endAddr = mappings[i].address() + mappings[i].size() + estCodeSigSize;
+					if ( endAddr > (sharedRegionStartAddress() + sharedRegionSize()) ) {
+						throwf("update_dyld_shared_cache[%u] for arch=%s, shared cache will not fit in shared regionsaddress space.  Overflow amount: %lluKB\n",
+							getpid(), fArchGraph->archName(), (endAddr-(sharedRegionStartAddress() + sharedRegionSize()))/1024);
+					}
+				}
+			}
+			
+			// compute UUID of whole cache
+			uint8_t digest[16];
+			CC_MD5(inMemoryCache, cacheFileSize, digest);
+			// <rdar://problem/6723729> uuids should conform to RFC 4122 UUID version 4 & UUID version 5 formats
+			digest[6] = ( digest[6] & 0x0F ) | ( 3 << 4 );
+			digest[8] = ( digest[8] & 0x3F ) | 0x80;
+			((dyldCacheHeader<E>*)inMemoryCache)->set_uuid(digest);
 			
 			if ( fVerify ) {
 				// if no existing cache, say so
@@ -2556,7 +2970,7 @@
 
 				// write out cache file
 				if ( verbose )
-					fprintf(stderr, "update_dyld_shared_cache: writing cache to disk\n");
+					fprintf(stderr, "update_dyld_shared_cache: writing cache to disk: %s\n", tempCachePath);
 				if ( ::pwrite(fd, inMemoryCache, cacheFileSize, 0) != cacheFileSize )
 					throwf("write() failure creating cache file, errno=%d", errno);
 				if ( progress ) {
@@ -2572,6 +2986,9 @@
 				if ( result != 0 ) 
 					fprintf(stderr, "update_dyld_shared_cache: warning, close() failed with errno=%d for %s\n", errno, tempCachePath);
 				
+				if ( !iPhoneOS )
+					adhoc_codesign_share_cache(tempCachePath);
+
 				// <rdar://problem/7901042> Make life easier for the kernel at shutdown.
 				// If we just move the new cache file over the old, the old file
 				// may need to exist in the open-unlink state.  But because it
@@ -2594,13 +3011,15 @@
 				}
 				
 				// move new cache file to correct location for use after reboot
+				if ( verbose )
+					fprintf(stderr, "update_dyld_shared_cache: atomically moving cache file into place: %s\n", fCacheFilePath);
 				result = ::rename(tempCachePath, fCacheFilePath);
 				if ( result != 0 ) 
 					throwf("can't swap newly create dyld shared cache file: rename(%s,%s) returned errno=%d", tempCachePath, fCacheFilePath, errno);
 				
 				
 				// flush everything to disk to assure rename() gets recorded
-				::sync();
+				sync_volume(fCacheFilePath);
 				didUpdate = true;
 				
 				// restore default signal handlers
@@ -2629,51 +3048,75 @@
 						else if ( it->sfm_init_prot == (VM_PROT_EXECUTE|VM_PROT_WRITE|VM_PROT_READ) )
 							prot = "WX";
 						if ( it->sfm_size > 1024*1024 )
-							fprintf(fmap, "mapping %s %4lluMB 0x%0llX -> 0x%0llX\n", prot, it->sfm_size/(1024*1024),
+							fprintf(fmap, "mapping  %s %4lluMB 0x%0llX -> 0x%0llX\n", prot, it->sfm_size/(1024*1024),
 																it->sfm_address, it->sfm_address+it->sfm_size);
 						else
-							fprintf(fmap, "mapping %s %4lluKB 0x%0llX -> 0x%0llX\n", prot, it->sfm_size/1024,
+							fprintf(fmap, "mapping  %s %4lluKB 0x%0llX -> 0x%0llX\n", prot, it->sfm_size/1024,
 																it->sfm_address, it->sfm_address+it->sfm_size);
 					}
 
-					fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX weak binding info\n",		
+					fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX weak binding info\n",		
 								(fOffsetOfExportInfoInCombinedLinkedit-fOffsetOfWeakBindInfoInCombinedLinkedit)/1024,
 								fLinkEditsStartAddress+fOffsetOfWeakBindInfoInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfExportInfoInCombinedLinkedit);
-					fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX export info\n",		
+					fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX export info\n",		
 								(fOffsetOfBindInfoInCombinedLinkedit-fOffsetOfExportInfoInCombinedLinkedit)/1024,
 								fLinkEditsStartAddress+fOffsetOfExportInfoInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfBindInfoInCombinedLinkedit);
-					fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX binding info\n",		
+					fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX binding info\n",		
 								(fOffsetOfLazyBindInfoInCombinedLinkedit-fOffsetOfBindInfoInCombinedLinkedit)/1024,
 								fLinkEditsStartAddress+fOffsetOfBindInfoInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfLazyBindInfoInCombinedLinkedit);
-					fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX lazy binding info\n",		
+					fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX lazy binding info\n",		
 								(fOffsetOfOldSymbolTableInfoInCombinedLinkedit-fOffsetOfLazyBindInfoInCombinedLinkedit)/1024,
 								fLinkEditsStartAddress+fOffsetOfLazyBindInfoInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfOldSymbolTableInfoInCombinedLinkedit);
-					fprintf(fmap, "linkedit   %4uMB 0x%0llX -> 0x%0llX non-dyld symbol table size\n",		
+					fprintf(fmap, " linkedit   %4uMB 0x%0llX -> 0x%0llX non-dyld symbol table size\n",		
 								(fSizeOfOldSymbolTableInfoInCombinedLinkedit)/(1024*1024),
 								fLinkEditsStartAddress+fOffsetOfOldSymbolTableInfoInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfOldSymbolTableInfoInCombinedLinkedit+fSizeOfOldSymbolTableInfoInCombinedLinkedit);				
 					if ( fSizeOfFunctionStartsInCombinedLinkedit != 0 )
-						fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld functions starts size\n",		
+						fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld functions starts size\n",		
 								fSizeOfFunctionStartsInCombinedLinkedit/1024,
 								fLinkEditsStartAddress+fOffsetOfFunctionStartsInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfFunctionStartsInCombinedLinkedit+fSizeOfFunctionStartsInCombinedLinkedit);				
+					if ( fSizeOfDataInCodeInCombinedLinkedit != 0 )
+						fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld data-in-code info size\n",		
+								fSizeOfDataInCodeInCombinedLinkedit/1024,
+								fLinkEditsStartAddress+fOffsetOfDataInCodeInCombinedLinkedit,
+								fLinkEditsStartAddress+fOffsetOfDataInCodeInCombinedLinkedit+fSizeOfDataInCodeInCombinedLinkedit);				
 					if ( fSizeOfOldExternalRelocationsInCombinedLinkedit != 0 )
-						fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld external relocs size\n",		
+						fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld external relocs size\n",		
 								fSizeOfOldExternalRelocationsInCombinedLinkedit/1024,
 								fLinkEditsStartAddress+fOffsetOfOldExternalRelocationsInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfOldExternalRelocationsInCombinedLinkedit+fSizeOfOldExternalRelocationsInCombinedLinkedit);				
-					fprintf(fmap, "linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld indirect symbol table size\n",		
+					fprintf(fmap, " linkedit   %4uKB 0x%0llX -> 0x%0llX non-dyld indirect symbol table size\n",		
 								fSizeOfOldIndirectSymbolsInCombinedLinkedit/1024,
 								fLinkEditsStartAddress+fOffsetOfOldIndirectSymbolsInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfOldIndirectSymbolsInCombinedLinkedit+fSizeOfOldIndirectSymbolsInCombinedLinkedit);				
-					fprintf(fmap, "linkedit   %4uMB 0x%0llX -> 0x%0llX non-dyld string pool\n",		
+					fprintf(fmap, " linkedit   %4uMB 0x%0llX -> 0x%0llX non-dyld string pool\n",		
 								(fSizeOfOldStringPoolInCombinedLinkedit)/(1024*1024),
 								fLinkEditsStartAddress+fOffsetOfOldStringPoolInCombinedLinkedit,
 								fLinkEditsStartAddress+fOffsetOfOldStringPoolInCombinedLinkedit+fSizeOfOldStringPoolInCombinedLinkedit);				
+					
+					dyldCacheHeader<E>* cacheHeader = (dyldCacheHeader<E>*)inMemoryCache;
+					if ( cacheHeader->slideInfoSize() != 0 ) {
+						fprintf(fmap, " linkedit   %4lluKB kernel slide info\n", (cacheHeader->slideInfoSize())/1024);				
+					}
+
+					fprintf(fmap, "unmapped -- %4uMB local symbol info\n", fUnmappedLocalSymbolsSize/(1024*1024));					
+					
+					uint64_t endMappingAddr = fMappings[2].sfm_address + fMappings[2].sfm_size;
+					fprintf(fmap, "total map   %4lluMB\n", (endMappingAddr - sharedRegionStartAddress())/(1024*1024));	
+					if ( sharedRegionStartWritableAddress(0) == 0x7FFF70000000LL ) {
+						// x86_64 has different slide constraints
+						uint64_t freeSpace = 256*1024*1024 - fMappings[1].sfm_size;
+						fprintf(fmap, "r/w space   %4lluMB -> %d bits of entropy for ASLR\n\n", freeSpace/(1024*1024), (int)log2(freeSpace/4096));
+					}
+					else {
+						uint64_t freeSpace = sharedRegionStartAddress() + sharedRegionSize() - endMappingAddr;
+						fprintf(fmap, "free space  %4lluMB -> %d bits of entropy for ASLR\n\n", freeSpace/(1024*1024), (int)log2(freeSpace/4096));
+					}
 					
 					for(typename std::vector<LayoutInfo>::const_iterator it = fDylibs.begin(); it != fDylibs.end(); ++it) {
 						fprintf(fmap, "%s\n", it->layout->getID().name);
@@ -2771,15 +3214,7 @@
 						*last = '\0';
 						--last;
 					}
-					// <rdar://problem/8305479> images in shared cache are bound against different IOKit than found at runtime
-					// HACK:  Just ignore the known bad IOKit
-					if ( strcmp(symbolStart, "/System/Library/Frameworks/IOKit.framework/IOKit") == 0 ) {
-						fprintf(stderr, "update_dyld_shared_cache: warning, ignoring /System/Library/Frameworks/IOKit.framework/IOKit\n");
-						warnings.push_back("update_dyld_shared_cache: warning, ignoring /System/Library/Frameworks/IOKit.framework/IOKit\n");
-					}
-					else {
-						paths.push_back(symbolStart);
-					}
+					paths.push_back(symbolStart);
 					symbolStart = NULL;
 					state = lineStart;
 				}
@@ -2795,10 +3230,11 @@
 
 
 
-static void setSharedDylibs(const char* rootPath, bool usesOverlay, const std::set<ArchPair>& onlyArchs, std::vector<const char*> rootsPaths)
+static void setSharedDylibs(const char* rootPath, const std::vector<const char*>& overlayPaths, const std::set<ArchPair>& onlyArchs, std::vector<const char*> rootsPaths)
 {
 	// set file system root
-	ArchGraph::setFileSystemRoot(rootPath, usesOverlay);
+	ArchGraph::setFileSystemRoot(rootPath);
+	ArchGraph::setFileSystemOverlay(overlayPaths);
 
 	// initialize all architectures requested
 	for(std::set<ArchPair>::iterator a = onlyArchs.begin(); a != onlyArchs.end(); ++a)
@@ -2814,12 +3250,11 @@
 }
 
 
-static void scanForSharedDylibs(const char* rootPath, bool usesOverlay, const char* dirOfPathFiles, const std::set<ArchPair>& onlyArchs)
+static void scanForSharedDylibs(const char* rootPath, const std::vector<const char*>& overlayPaths, const char* dirOfPathFiles, const std::set<ArchPair>& onlyArchs)
 {
 	char rootDirOfPathFiles[strlen(rootPath)+strlen(dirOfPathFiles)+2];
-	// in -overlay mode, still look for roots in /var/db/dyld
 	// in -root mode, look for roots in /rootpath/var/db/dyld
-	if ( !usesOverlay && (strlen(rootPath) != 0) ) {
+	if ( rootPath[0] != '\0' ) {
 		strcpy(rootDirOfPathFiles, rootPath);
 		strcat(rootDirOfPathFiles, dirOfPathFiles);
 		dirOfPathFiles = rootDirOfPathFiles;
@@ -2860,14 +3295,14 @@
 	
 	if ( rootsPaths.size() == 0 )
 		fprintf(stderr, "update_dyld_shared_cache: warning, no entries found in shared_region_roots\n");
-	setSharedDylibs(rootPath, usesOverlay, onlyArchs, rootsPaths);
-}
-
-static void setSharedDylibs(const char* rootPath, bool usesOverlay, const char* pathsFile, const std::set<ArchPair>& onlyArchs)
+	setSharedDylibs(rootPath, overlayPaths, onlyArchs, rootsPaths);
+}
+
+static void setSharedDylibs(const char* rootPath, const std::vector<const char*>& overlayPaths, const char* pathsFile, const std::set<ArchPair>& onlyArchs)
 {
 	std::vector<const char*> rootsPaths;
 	parsePathsFile(pathsFile, rootsPaths);
-	setSharedDylibs(rootPath, usesOverlay, onlyArchs, rootsPaths);
+	setSharedDylibs(rootPath, overlayPaths, onlyArchs, rootsPaths);
 }
 
 
@@ -2914,8 +3349,8 @@
 
 
 
-static bool updateSharedeCacheFile(const char* rootPath, bool usesOverlay, const char* cacheDir, const std::set<ArchPair>& onlyArchs, 
-									bool force, bool alphaSort, bool optimize, bool deleteExistingFirst, bool verify, bool keepSignatures)
+static bool updateSharedeCacheFile(const char* rootPath, const std::vector<const char*>& overlayPaths, const char* cacheDir, bool explicitCacheDir, const std::set<ArchPair>& onlyArchs, 
+									bool force, bool alphaSort, bool optimize, bool deleteExistingFirst, bool verify, bool keepSignatures, bool dontMapLocalSymbols)
 {
 	bool didUpdate = false;
 	// get dyld load address info
@@ -2940,34 +3375,28 @@
 		else
 			fprintf(stderr, "update_dyld_shared_cache: warning, dyld not available for specified architectures\n");
 		switch ( a->arch ) {
-			case CPU_TYPE_POWERPC:
-				{
-		#if __i386__ || __x86_64__
-					// <rdar://problem/5217377> Rosetta does not work with optimized dyld shared cache
-					SharedCache<ppc> cache(ArchGraph::graphForArchPair(*a), rootPath, cacheDir, alphaSort, verify, false, usesOverlay, dyldBaseAddress);
-					didUpdate |= cache.update(usesOverlay, force, false, deleteExistingFirst, index, archCount, keepSignatures);
-		#else
-					SharedCache<ppc> cache(ArchGraph::graphForArchPair(*a), rootPath, cacheDir, alphaSort, verify, optimize, usesOverlay, dyldBaseAddress);
-					didUpdate |= cache.update(usesOverlay, force, optimize, deleteExistingFirst, index, archCount, keepSignatures);
-		#endif
-				}
-				break;
 			case CPU_TYPE_I386:
 				{
-					SharedCache<x86> cache(ArchGraph::graphForArchPair(*a), rootPath, cacheDir, alphaSort, verify, optimize, usesOverlay, dyldBaseAddress);
-					didUpdate |= cache.update(usesOverlay, force, optimize, deleteExistingFirst, index, archCount, keepSignatures);
+					SharedCache<x86> cache(ArchGraph::graphForArchPair(*a), rootPath, overlayPaths, cacheDir, explicitCacheDir, alphaSort, verify, optimize, dyldBaseAddress);
+					didUpdate |= cache.update(force, optimize, deleteExistingFirst, index, archCount, keepSignatures, dontMapLocalSymbols);
 				}
 				break;
 			case CPU_TYPE_X86_64:
 				{
-					SharedCache<x86_64> cache(ArchGraph::graphForArchPair(*a), rootPath, cacheDir, alphaSort, verify, optimize, usesOverlay, dyldBaseAddress);
-					didUpdate |= cache.update(usesOverlay, force, optimize, deleteExistingFirst, index, archCount, keepSignatures);
+					SharedCache<x86_64> cache(ArchGraph::graphForArchPair(*a), rootPath, overlayPaths, cacheDir, explicitCacheDir, alphaSort, verify, optimize, dyldBaseAddress);
+					didUpdate |= cache.update(force, optimize, deleteExistingFirst, index, archCount, keepSignatures, dontMapLocalSymbols);
 				}
 				break;
 			case CPU_TYPE_ARM:
 				{
-					SharedCache<arm> cache(ArchGraph::graphForArchPair(*a), rootPath, cacheDir, alphaSort, verify, optimize, usesOverlay, dyldBaseAddress);
-					didUpdate |= cache.update(usesOverlay, force, optimize, deleteExistingFirst, index, archCount, keepSignatures);
+					SharedCache<arm> cache(ArchGraph::graphForArchPair(*a), rootPath, overlayPaths, cacheDir, explicitCacheDir, alphaSort, verify, optimize, dyldBaseAddress);
+					didUpdate |= cache.update(force, optimize, deleteExistingFirst, index, archCount, keepSignatures, dontMapLocalSymbols);
+				}
+				break;
+			case CPU_TYPE_ARM64:
+				{
+					SharedCache<arm64> cache(ArchGraph::graphForArchPair(*a), rootPath, overlayPaths, cacheDir, explicitCacheDir, alphaSort, verify, optimize, dyldBaseAddress);
+					didUpdate |= cache.update(force, optimize, deleteExistingFirst, index, archCount, keepSignatures, dontMapLocalSymbols);
 				}
 				break;
 		}
@@ -2990,14 +3419,16 @@
 {
 	std::set<ArchPair> onlyArchs;
 	const char* rootPath = "";
+	std::vector<const char*> overlayPaths;
 	const char* dylibListFile = NULL;
 	bool force = false;
 	bool alphaSort = false;
 	bool optimize = true;
-	bool hasRoot = false;
-	bool hasOverlay = false;
 	bool verify = false;
 	bool keepSignatures = false;
+	bool explicitCacheDir = false;
+	bool dontMapLocalSymbols = false;
+	bool relaunchForHaswell = false;
 	const char* cacheDir = NULL;
 	
 	try {
@@ -3026,8 +3457,12 @@
 				else if ( strcmp(arg, "-no_opt") == 0 ) {
 					optimize = false;
 				}
+				else if ( strcmp(arg, "-dont_map_local_symbols") == 0 ) {
+					dontMapLocalSymbols = true;
+				}
 				else if ( strcmp(arg, "-iPhone") == 0 ) {
 					iPhoneOS = true;
+					alphaSort = true;
 				}
 				else if ( strcmp(arg, "-dylib_list") == 0 ) {
 					dylibListFile = argv[++i];
@@ -3035,34 +3470,30 @@
 						throw "-dylib_list missing path argument";
 				}
 				else if ( (strcmp(arg, "-root") == 0) || (strcmp(arg, "--root") == 0) ) {
-					if ( hasOverlay )
-						throw "cannot use both -root and -overlay";
 					rootPath = argv[++i];
 					if ( rootPath == NULL )
 						throw "-root missing path argument";
-					hasRoot = true;
 				}
 				else if ( strcmp(arg, "-overlay") == 0 ) {
-					if ( hasRoot )
-						throw "cannot use both -root and -overlay";
-					rootPath = argv[++i];
-					if ( rootPath == NULL )
-						throw "-root missing path argument";
-					hasOverlay = true;
+					const char* path = argv[++i];
+					if ( path == NULL )
+						throw "-overlay missing path argument";
+					overlayPaths.push_back(path);
 				}
 				else if ( strcmp(arg, "-cache_dir") == 0 ) {
 					cacheDir = argv[++i];
 					if ( cacheDir == NULL )
 						throw "-cache_dir missing path argument";
+					explicitCacheDir = true;
 				}
 				else if ( strcmp(arg, "-arch") == 0 ) {
 					const char* arch = argv[++i];
-					if ( strcmp(arch, "ppc") == 0 ) 
-						onlyArchs.insert(ArchPair(CPU_TYPE_POWERPC, CPU_SUBTYPE_POWERPC_ALL));
-					else if ( strcmp(arch, "i386") == 0 )
+					if ( strcmp(arch, "i386") == 0 )
 						onlyArchs.insert(ArchPair(CPU_TYPE_I386, CPU_SUBTYPE_I386_ALL));
 					else if ( strcmp(arch, "x86_64") == 0 )
 						onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_64_ALL));
+					else if ( strcmp(arch, "x86_64h") == 0 )
+						onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_64_H));
 					else if ( strcmp(arch, "armv4t") == 0 )
 						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V4T));
 					else if ( strcmp(arch, "armv5") == 0 )
@@ -3071,15 +3502,21 @@
 						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V6));
 					else if ( strcmp(arch, "armv7") == 0 )
 						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V7));
+					else if ( strcmp(arch, "armv7f") == 0 )
+						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V7F));
+					else if ( strcmp(arch, "armv7k") == 0 )
+						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V7K));
+					else if ( strcmp(arch, "armv7s") == 0 )
+						onlyArchs.insert(ArchPair(CPU_TYPE_ARM, CPU_SUBTYPE_ARM_V7S));
+					else if ( strcmp(arch, "arm64") == 0 )
+						onlyArchs.insert(ArchPair(CPU_TYPE_ARM64, CPU_SUBTYPE_ARM64_ALL));
 					else 
 						throwf("unknown architecture %s", arch);
 				}
 				else if ( strcmp(arg, "-universal_boot") == 0 ) {
-			#if __ppc__
-					throwf("universal_boot option can only be used on Intel machines");
-			#endif
 					onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_64_ALL));
 					onlyArchs.insert(ArchPair(CPU_TYPE_I386, CPU_SUBTYPE_I386_ALL));
+					relaunchForHaswell = true;
 				}
 				else {
 					usage();
@@ -3092,7 +3529,7 @@
 			}
 		}
 				
-		// strip tailing slashes on -root or -overlay
+		// strip tailing slashes on -root 
 		// make it a real path so as to not make all dylibs look like symlink aliases
 		if ( rootPath[0] != '\0' ) {
 			char realRootPath[MAXPATHLEN];
@@ -3101,15 +3538,17 @@
 			rootPath = strdup(realRootPath);
 		}
 		
-		// set location to write cache dir
-		if ( cacheDir == NULL ) {
-			if ( (rootPath[0] == '\0') || hasOverlay ) {
-				cacheDir =  (iPhoneOS ? IPHONE_DYLD_SHARED_CACHE_DIR : MACOSX_DYLD_SHARED_CACHE_DIR);
-			}
-			else {
-				asprintf((char**)&cacheDir, "%s/%s", rootPath, (iPhoneOS ? IPHONE_DYLD_SHARED_CACHE_DIR : MACOSX_DYLD_SHARED_CACHE_DIR));
-			}
-		}
+		// strip tailing slashes on -overlay
+		for (std::vector<const char*>::iterator it=overlayPaths.begin(); it != overlayPaths.end(); ++it) {
+			char realOverlayPath[MAXPATHLEN];
+			if ( realpath(*it, realOverlayPath) == NULL )
+				throwf("realpath() failed on %s\n", *it);
+			*it = strdup(realOverlayPath);
+		}
+
+		// set default location to write cache dir
+		if ( cacheDir == NULL ) 
+			cacheDir = (iPhoneOS ? IPHONE_DYLD_SHARED_CACHE_DIR : MACOSX_DYLD_SHARED_CACHE_DIR);
 
 		// if no restrictions specified, use architectures that work on this machine
 		if ( onlyArchs.size() == 0 ) {
@@ -3122,22 +3561,21 @@
 				size_t len = sizeof(int);
 			#if __i386__ || __x86_64__
 				onlyArchs.insert(ArchPair(CPU_TYPE_I386, CPU_SUBTYPE_I386_ALL));
-				// check rosetta is installed
-				char rosettaPath[1024];
-				strlcpy(rosettaPath, rootPath, 1024);
-				strlcat(rosettaPath, "/usr/libexec/oah/translate", 1024);
-				struct stat stat_buf;
-				if ( stat(rosettaPath, &stat_buf) == 0 ) {
-					onlyArchs.insert(ArchPair(CPU_TYPE_POWERPC, CPU_SUBTYPE_POWERPC_ALL));
-				}
-				else if ( hasOverlay ) {
-					// in overlay mode, rosetta may be installed on base system, but is not in update root
-					if ( stat("/usr/libexec/oah/translate", &stat_buf) == 0 ) 
-						onlyArchs.insert(ArchPair(CPU_TYPE_POWERPC, CPU_SUBTYPE_POWERPC_ALL));
-				}
 				// check system is capable of running 64-bit programs
-				if ( (sysctlbyname("hw.optional.x86_64", &available, &len, NULL, 0) == 0) && available )
-					onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_64_ALL));
+				if ( (sysctlbyname("hw.optional.x86_64", &available, &len, NULL, 0) == 0) && available ) {
+					// check system is capable of running x86_64h code
+					struct host_basic_info info;
+					mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT;
+					mach_port_t hostPort = mach_host_self();
+					kern_return_t result = host_info(hostPort, HOST_BASIC_INFO, (host_info_t)&info, &count);
+					mach_port_deallocate(mach_task_self(), hostPort);
+					if ( result != KERN_SUCCESS )
+						throw "host_info() failed";
+					if ( info.cpu_subtype == CPU_SUBTYPE_X86_64_H )
+						onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_64_H));
+					else
+						onlyArchs.insert(ArchPair(CPU_TYPE_X86_64, CPU_SUBTYPE_X86_ALL));
+				}
 			#else
 				#error unsupported architecture
 			#endif
@@ -3149,11 +3587,44 @@
 		
 		// build list of shared dylibs
 		if ( dylibListFile != NULL )
-			setSharedDylibs(rootPath, hasOverlay, dylibListFile, onlyArchs);
+			setSharedDylibs(rootPath, overlayPaths, dylibListFile, onlyArchs);
 		else
-			scanForSharedDylibs(rootPath, hasOverlay, "/var/db/dyld/shared_region_roots/", onlyArchs);
-		updateSharedeCacheFile(rootPath, hasOverlay, cacheDir, onlyArchs, force, alphaSort, optimize, 
-								false, verify, keepSignatures);
+			scanForSharedDylibs(rootPath, overlayPaths, "/var/db/dyld/shared_region_roots/", onlyArchs);
+		bool didUpdate = updateSharedeCacheFile(rootPath, overlayPaths, cacheDir, explicitCacheDir, onlyArchs, force, alphaSort, optimize,
+								false, verify, keepSignatures, dontMapLocalSymbols);
+								
+		if ( didUpdate && !iPhoneOS ) {
+			void* handle = dlopen("/usr/lib/libspindump.dylib", RTLD_LAZY);
+			if ( handle != NULL ) {
+				typedef bool (*dscsym_proc_t)(const char *root);
+				dscsym_proc_t proc = (dscsym_proc_t)dlsym(handle, "dscsym_save_nuggets_for_current_caches");
+				const char* nuggetRootPath = "/";
+				if ( !overlayPaths.empty() ) 
+					nuggetRootPath = overlayPaths[0];
+				else if ( rootPath[0] != '\0' )
+					nuggetRootPath = rootPath;
+				(*proc)(nuggetRootPath);
+			}
+			dlclose(handle);
+		}
+
+		if ( relaunchForHaswell ) {
+			char cmd[2048];
+			strlcpy(cmd, argv[0], 2048);
+			strlcat(cmd, " -arch x86_64h", 2048);
+			if ( force )
+				strlcat(cmd, " -force", 2048);
+			if ( verify )
+				strlcat(cmd, " -verify", 2048);
+			if ( alphaSort )
+				strlcat(cmd, " -sort_by_name", 2048);
+			if ( (rootPath != NULL) && (rootPath[0] != '\0') ) {
+				strlcat(cmd, " -root ", 2048);
+				strlcat(cmd, rootPath, 2048);
+			}
+			return system(cmd);
+		}
+
 	}
 	catch (const char* msg) {
 		fprintf(stderr, "update_dyld_shared_cache failed: %s\n", msg);