Loading...
--- dyld/dyld-635.2/dyld3/shared-cache/OptimizerObjC.cpp
+++ dyld/dyld-733.6/dyld3/shared-cache/OptimizerObjC.cpp
@@ -36,6 +36,11 @@
#include "FileAbstraction.hpp"
#include "MachOFileAbstraction.hpp"
#include "MachOLoaded.h"
+#include "MachOAnalyzer.h"
+
+#ifndef MH_HAS_OBJC
+ #define MH_HAS_OBJC 0x40000000
+#endif
// Scan a C++ or Swift length-mangled field.
static bool scanMangledField(const char *&string, const char *end,
@@ -124,13 +129,13 @@
uint64_t vmAddrForOnDiskVMAddr(uint64_t vmaddr) {
if ( _chainedFixups ) {
dyld3::MachOLoaded::ChainedFixupPointerOnDisk ptr;
- ptr.raw = vmaddr;
- assert(ptr.authRebase.bind == 0);
- if ( ptr.authRebase.auth ) {
- vmaddr = _cacheUnslideAddr + ptr.authRebase.target;
+ ptr.raw64 = vmaddr;
+ assert(ptr.arm64e.authRebase.bind == 0);
+ if ( ptr.arm64e.authRebase.auth ) {
+ vmaddr = _cacheUnslideAddr + ptr.arm64e.authRebase.target;
}
else {
- vmaddr = ptr.plainRebase.signExtendedTarget();
+ vmaddr = ptr.arm64e.unpackTarget();
}
}
return vmaddr;
@@ -283,6 +288,14 @@
return (pint_t)element->second;
}
+ void visitCoalescedStrings(const CacheBuilder::CacheCoalescedText& coalescedText) {
+ const CacheBuilder::CacheCoalescedText::StringSection& methodNames = coalescedText.getSectionData("__objc_methname");
+ for (const auto& stringAndOffset : methodNames.stringsToOffsets) {
+ uint64_t vmAddr = methodNames.bufferVMAddr + stringAndOffset.second;
+ _selectorStrings[stringAndOffset.first.data()] = vmAddr;
+ }
+ }
+
objc_opt::string_map& strings() {
return _selectorStrings;
}
@@ -341,11 +354,13 @@
private:
typedef typename P::uint_t pint_t;
- objc_opt::string_map _protocolNames;
- objc_opt::protocol_map _protocols;
- size_t _protocolCount;
- size_t _protocolReferenceCount;
- Diagnostics& _diagnostics;
+ objc_opt::string_map _protocolNames;
+ objc_opt::legacy_protocol_map _protocols;
+ objc_opt::protocol_map _protocolsAndHeaders;
+ size_t _protocolCount;
+ size_t _protocolReferenceCount;
+ Diagnostics& _diagnostics;
+ HeaderInfoOptimizer<P, objc_header_info_ro_t<P>>& _hInfos;
friend class ProtocolReferenceWalker<P, ProtocolOptimizer<P>>;
@@ -360,8 +375,8 @@
public:
- ProtocolOptimizer(Diagnostics& diag)
- : _protocolCount(0), _protocolReferenceCount(0), _diagnostics(diag) {
+ ProtocolOptimizer(Diagnostics& diag, HeaderInfoOptimizer<P, objc_header_info_ro_t<P>>& hinfos)
+ : _protocolCount(0), _protocolReferenceCount(0), _diagnostics(diag), _hInfos(hinfos) {
}
void addProtocols(ContentAccessor* cache, const macho_header<P>* header)
@@ -378,13 +393,17 @@
_diagnostics.error("objc protocol is too big");
return;
}
-
uint64_t name_vmaddr = cache->vmAddrForContent((void*)name);
uint64_t proto_vmaddr = cache->vmAddrForContent(proto);
_protocolNames.insert(objc_opt::string_map::value_type(name, name_vmaddr));
- _protocols.insert(objc_opt::protocol_map::value_type(name, proto_vmaddr));
+ _protocols.insert(objc_opt::legacy_protocol_map::value_type(name, proto_vmaddr));
_protocolCount++;
}
+
+ // Note down which header this protocol came from. We'll fill in the proto_vmaddr here later
+ // once we've chosen a single definition for the protocol with this name.
+ uint64_t hinfo_vmaddr = cache->vmAddrForContent(_hInfos.hinfoForHeader(cache, header));
+ _protocolsAndHeaders.insert(objc_opt::class_map::value_type(name, std::pair<uint64_t, uint64_t>(0, hinfo_vmaddr)));
}
}
@@ -405,9 +424,7 @@
return "libobjc's read-write section is too small (metadata not optimized)";
}
- for (objc_opt::protocol_map::iterator iter = _protocols.begin();
- iter != _protocols.end();
- ++iter)
+ for (auto iter = _protocols.begin(); iter != _protocols.end(); ++iter)
{
objc_protocol_t<P>* oldProto = (objc_protocol_t<P>*)
cache->contentForVMAddr(iter->second);
@@ -450,12 +467,19 @@
proto->setDemangledName(cache, roName, _diagnostics);
}
proto->setFixedUp();
+ proto->setIsCanonical();
// Redirect the protocol table at our new object.
iter->second = cache->vmAddrForContent(proto);
// Add new rebase entries.
proto->addPointers(cache, aslrTracker);
+ }
+
+ // Now that we've chosen the canonical protocols, set the duplicate headers to
+ // point to their protocols.
+ for (auto iter = _protocolsAndHeaders.begin(); iter != _protocolsAndHeaders.end(); ++iter) {
+ iter->second.first = _protocols[iter->first];
}
return NULL;
@@ -467,13 +491,17 @@
refs.walk(cache, header);
}
- objc_opt::string_map& protocolNames() {
+ objc_opt::string_map& protocolNames() {
return _protocolNames;
}
- objc_opt::protocol_map& protocols() {
+ objc_opt::legacy_protocol_map& protocols() {
return _protocols;
- }
+ }
+
+ objc_opt::protocol_map& protocolsAndHeaders() {
+ return _protocolsAndHeaders;
+ }
size_t protocolCount() const { return _protocolCount; }
size_t protocolReferenceCount() const { return _protocolReferenceCount; }
@@ -487,11 +515,87 @@
return 100;
}
+template <typename P>
+void addObjcSegments(Diagnostics& diag, DyldSharedCache* cache, const mach_header* libobjcMH,
+ uint8_t* objcReadOnlyBuffer, uint64_t objcReadOnlyBufferSizeAllocated,
+ uint8_t* objcReadWriteBuffer, uint64_t objcReadWriteBufferSizeAllocated,
+ uint32_t objcRwFileOffset)
+{
+ // validate there is enough free space to add the load commands
+ const dyld3::MachOAnalyzer* libobjcMA = ((dyld3::MachOAnalyzer*)libobjcMH);
+ uint32_t freeSpace = libobjcMA->loadCommandsFreeSpace();
+ const uint32_t segSize = sizeof(macho_segment_command<P>);
+ if ( freeSpace < 2*segSize ) {
+ diag.warning("not enough space in libojbc.dylib to add load commands for objc optimization regions");
+ return;
+ }
+
+ // find location of LINKEDIT LC_SEGMENT load command, we need to insert new segments before it
+ __block uint8_t* linkeditSeg = nullptr;
+ libobjcMA->forEachSegment(^(const dyld3::MachOFile::SegmentInfo& info, bool& stop) {
+ if ( strcmp(info.segName, "__LINKEDIT") == 0 )
+ linkeditSeg = (uint8_t*)libobjcMH + info.loadCommandOffset;
+ });
+ if ( linkeditSeg == nullptr ) {
+ diag.warning("__LINKEDIT not found in libojbc.dylib");
+ return;
+ }
+
+ // move load commands to make room to insert two new ones before LINKEDIT segment load command
+ uint8_t* endOfLoadCommands = (uint8_t*)libobjcMH + sizeof(macho_header<P>) + libobjcMH->sizeofcmds;
+ uint32_t remainingSize = (uint32_t)(endOfLoadCommands - linkeditSeg);
+ memmove(linkeditSeg+2*segSize, linkeditSeg, remainingSize);
+
+ // insert new segments
+ macho_segment_command<P>* roSeg = (macho_segment_command<P>*)(linkeditSeg);
+ macho_segment_command<P>* rwSeg = (macho_segment_command<P>*)(linkeditSeg+sizeof(macho_segment_command<P>));
+ roSeg->set_cmd(macho_segment_command<P>::CMD);
+ roSeg->set_cmdsize(segSize);
+ roSeg->set_segname("__OBJC_RO");
+ roSeg->set_vmaddr(cache->unslidLoadAddress() + objcReadOnlyBuffer - (uint8_t*)cache);
+ roSeg->set_vmsize(objcReadOnlyBufferSizeAllocated);
+ roSeg->set_fileoff(objcReadOnlyBuffer - (uint8_t*)cache);
+ roSeg->set_filesize(objcReadOnlyBufferSizeAllocated);
+ roSeg->set_maxprot(VM_PROT_READ);
+ roSeg->set_initprot(VM_PROT_READ);
+ roSeg->set_nsects(0);
+ roSeg->set_flags(0);
+ rwSeg->set_cmd(macho_segment_command<P>::CMD);
+ rwSeg->set_cmdsize(segSize);
+ rwSeg->set_segname("__OBJC_RW");
+ rwSeg->set_vmaddr(cache->unslidLoadAddress() + objcReadWriteBuffer - (uint8_t*)cache);
+ rwSeg->set_vmsize(objcReadWriteBufferSizeAllocated);
+ rwSeg->set_fileoff(objcRwFileOffset);
+ rwSeg->set_filesize(objcReadWriteBufferSizeAllocated);
+ rwSeg->set_maxprot(VM_PROT_WRITE|VM_PROT_READ);
+ rwSeg->set_initprot(VM_PROT_WRITE|VM_PROT_READ);
+ rwSeg->set_nsects(0);
+ rwSeg->set_flags(0);
+
+ // update mach_header to account for new load commands
+ macho_header<P>* mh = (macho_header<P>*)libobjcMH;
+ mh->set_sizeofcmds(mh->sizeofcmds() + 2*segSize);
+ mh->set_ncmds(mh->ncmds()+2);
+
+ // fix up table at start of dyld cache that has pointer into install name for libobjc
+ dyld_cache_image_info* images = (dyld_cache_image_info*)((uint8_t*)cache + cache->header.imagesOffset);
+ uint64_t libobjcUnslidAddress = cache->unslidLoadAddress() + ((uint8_t*)libobjcMH - (uint8_t*)cache);
+ for (uint32_t i=0; i < cache->header.imagesCount; ++i) {
+ if ( images[i].address == libobjcUnslidAddress ) {
+ images[i].pathFileOffset += (2*segSize);
+ break;
+ }
+ }
+}
+
template <typename P>
void doOptimizeObjC(DyldSharedCache* cache, bool forProduction, CacheBuilder::ASLR_Tracker& aslrTracker,
- CacheBuilder::LOH_Tracker& lohTracker,
- const std::map<void*, std::string>& missingWeakImports, Diagnostics& diag)
+ CacheBuilder::LOH_Tracker& lohTracker, const CacheBuilder::CacheCoalescedText& coalescedText,
+ const std::map<void*, std::string>& missingWeakImports, Diagnostics& diag,
+ uint8_t* objcReadOnlyBuffer, uint64_t objcReadOnlyBufferSizeUsed, uint64_t objcReadOnlyBufferSizeAllocated,
+ uint8_t* objcReadWriteBuffer, uint64_t objcReadWriteBufferSizeAllocated,
+ uint32_t objcRwFileOffset)
{
typedef typename P::E E;
typedef typename P::uint_t pint_t;
@@ -509,15 +613,15 @@
//
// Find libobjc's empty sections and build list of images with objc metadata
//
+ __block const mach_header* libobjcMH = nullptr;
__block const macho_section<P> *optROSection = nullptr;
- __block const macho_section<P> *optRWSection = nullptr;
__block const macho_section<P> *optPointerListSection = nullptr;
__block std::vector<const macho_header<P>*> objcDylibs;
cache->forEachImage(^(const mach_header* machHeader, const char* installName) {
const macho_header<P>* mh = (const macho_header<P>*)machHeader;
if ( strstr(installName, "/libobjc.") != nullptr ) {
+ libobjcMH = (mach_header*)mh;
optROSection = mh->getSection("__TEXT", "__objc_opt_ro");
- optRWSection = mh->getSection("__DATA", "__objc_opt_rw");
optPointerListSection = mh->getSection("__DATA", "__objc_opt_ptrs");
}
if ( mh->getSection("__DATA", "__objc_imageinfo") || mh->getSection("__OBJC", "__image_info") ) {
@@ -529,23 +633,21 @@
diag.warning("libobjc's read-only section missing (metadata not optimized)");
return;
}
- if ( optRWSection == nullptr ) {
- diag.warning("libobjc's read/write section missing (metadata not optimized)");
- return;
- }
if ( optPointerListSection == nullptr ) {
diag.warning("libobjc's pointer list section missing (metadata not optimized)");
return;
}
-
- uint8_t* optROData = (uint8_t*)cacheAccessor.contentForVMAddr(optROSection->addr());
+ // point optROData into space allocated in dyld cache
+ uint8_t* optROData = objcReadOnlyBuffer + objcReadOnlyBufferSizeUsed;
+ size_t optRORemaining = objcReadOnlyBufferSizeAllocated - objcReadOnlyBufferSizeUsed;
+ *((uint32_t*)optROData) = objc_opt::VERSION;
if ( optROData == nullptr ) {
diag.warning("libobjc's read-only section has bad content");
return;
}
- size_t optRORemaining = optROSection->size();
- uint8_t* optRWData = (uint8_t*)cacheAccessor.contentForVMAddr(optRWSection->addr());
- size_t optRWRemaining = optRWSection->size();
+
+ uint8_t* optRWData = objcReadWriteBuffer;
+ size_t optRWRemaining = objcReadWriteBufferSizeAllocated;
if (optRORemaining < headerSize) {
diag.warning("libobjc's read-only section is too small (metadata not optimized)");
return;
@@ -581,7 +683,7 @@
//
// First the RO header info
// log("writing out %d RO dylibs at offset %d", (uint32_t)objcDylibs.size(), (uint32_t)(optROSection->size() - optRORemaining));
- uint64_t hinfoROVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
+ uint64_t hinfoROVMAddr = cacheAccessor.vmAddrForContent(optROData);
HeaderInfoOptimizer<P, objc_header_info_ro_t<P>> hinfoROOptimizer;
const char* err = hinfoROOptimizer.init((uint32_t)objcDylibs.size(), optROData, optRORemaining);
if (err) {
@@ -596,7 +698,7 @@
// Then the RW header info
// log("writing out %d RW dylibs at offset %d", (uint32_t)objcDylibs.size(), (uint32_t)(optRWSection->size() - optRWRemaining));
- uint64_t hinfoRWVMAddr = (uint64_t)optRWSection->addr() + (uint64_t)optRWSection->size() - optRWRemaining;
+ uint64_t hinfoRWVMAddr = cacheAccessor.vmAddrForContent(optRWData);
HeaderInfoOptimizer<P, objc_header_info_rw_t<P>> hinfoRWOptimizer;
err = hinfoRWOptimizer.init((uint32_t)objcDylibs.size(), optRWData, optRWRemaining);
if (err) {
@@ -621,6 +723,39 @@
ObjCSelectorUniquer<P> uniq(&cacheAccessor);
std::vector<const macho_header<P>*> sizeSortedDylibs = objcDylibs;
std::sort(sizeSortedDylibs.begin(), sizeSortedDylibs.end(), [](const macho_header<P>* lmh, const macho_header<P>* rmh) -> bool {
+ // Sort a select few heavy hitters first.
+ auto getPriority = [](const char* installName) -> int {
+ if (!strcmp(installName, "/usr/lib/libobjc.A.dylib"))
+ return 0;
+ if (!strcmp(installName, "/System/Library/Frameworks/Foundation.framework/Versions/C/Foundation") ||
+ !strcmp(installName, "/System/Library/Frameworks/Foundation.framework/Foundation"))
+ return 1;
+ if (!strcmp(installName, "/System/Library/Frameworks/CoreFoundation.framework/Versions/A/CoreFoundation") ||
+ !strcmp(installName, "/System/Library/Frameworks/CoreFoundation.framework/CoreFoundation"))
+ return 2;
+ // Note we don't sort iOSMac UIKitCore early as we want iOSMac after macOS.
+ if (!strcmp(installName, "/System/Library/PrivateFrameworks/UIKitCore.framework/UIKitCore"))
+ return 3;
+ if (!strcmp(installName, "/System/Library/Frameworks/AppKit.framework/Versions/C/AppKit"))
+ return 4;
+ if (!strcmp(installName, "/System/Library/Frameworks/CFNetwork.framework/Versions/A/CFNetwork") ||
+ !strcmp(installName, "/System/Library/Frameworks/CFNetwork.framework/CFNetwork"))
+ return 5;
+ return INT_MAX;
+ };
+
+ // Sort by priority first
+ int priorityA = getPriority(((const dyld3::MachOFile*)lmh)->installName());
+ int priorityB = getPriority(((const dyld3::MachOFile*)rmh)->installName());
+ if (priorityA != priorityB)
+ return priorityA < priorityB;
+
+ // Sort mac before iOSMac
+ bool isIOSMacA = strncmp(((const dyld3::MachOFile*)lmh)->installName(), "/System/iOSSupport/", 19) == 0;
+ bool isIOSMacB = strncmp(((const dyld3::MachOFile*)rmh)->installName(), "/System/iOSSupport/", 19) == 0;
+ if (isIOSMacA != isIOSMacB)
+ return !isIOSMacA;
+
const macho_section<P>* lSection = lmh->getSection("__TEXT", "__objc_methname");
const macho_section<P>* rSection = rmh->getSection("__TEXT", "__objc_methname");
uint64_t lSelectorSize = (lSection ? lSection->size() : 0);
@@ -628,7 +763,12 @@
return lSelectorSize > rSelectorSize;
});
+ auto alignPointer = [](uint8_t* ptr) -> uint8_t* {
+ return (uint8_t*)(((uintptr_t)ptr + 0x7) & ~0x7);
+ };
+
SelectorOptimizer<P, ObjCSelectorUniquer<P> > selOptimizer(uniq);
+ selOptimizer.visitCoalescedStrings(coalescedText);
for (const macho_header<P>* mh : sizeSortedDylibs) {
LegacySelectorUpdater<P, ObjCSelectorUniquer<P>>::update(&cacheAccessor, mh, uniq);
selOptimizer.optimize(&cacheAccessor, mh);
@@ -637,7 +777,7 @@
diag.verbose(" uniqued %6lu selectors\n", uniq.strings().size());
diag.verbose(" updated %6lu selector references\n", uniq.count());
- uint64_t seloptVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
+ uint64_t seloptVMAddr = cacheAccessor.vmAddrForContent(optROData);
objc_opt::objc_selopt_t *selopt = new(optROData) objc_opt::objc_selopt_t;
err = selopt->write(seloptVMAddr, optRORemaining, uniq.strings());
if (err) {
@@ -645,6 +785,7 @@
return;
}
optROData += selopt->size();
+ optROData = alignPointer(optROData);
optRORemaining -= selopt->size();
uint32_t seloptCapacity = selopt->capacity;
uint32_t seloptOccupied = selopt->occupied;
@@ -658,20 +799,24 @@
//
// Detect classes that have missing weak-import superclasses.
//
- // Production only. Development cache does not do this: a replacement
- // library could omit a class at runtime that was present during
- // cache construction.
+ // Production shared caches don't support roots so we can set this and know
+ // there will definitely not be missing weak superclasses at runtime.
+ // Development shared caches can set this bit as the objc runtime only trusts
+ // this bit if there are no roots at runtime.
//
// This is SAFE: the binaries themselves are unmodified.
- bool noMissingWeakSuperclasses = false; // dev cache can't promise otherwise
+ WeakClassDetector<P> weakopt;
+ bool noMissingWeakSuperclasses = weakopt.noMissingWeakSuperclasses(&cacheAccessor,
+ missingWeakImports,
+ sizeSortedDylibs);
+
if (forProduction) {
- WeakClassDetector<P> weakopt;
- noMissingWeakSuperclasses =
- weakopt.noMissingWeakSuperclasses(&cacheAccessor, missingWeakImports, sizeSortedDylibs);
-
- // Shared cache does not currently support unbound weak references.
- // Here we assert that there are none. If support is added later then
+ // Shared cache does not currently support unbound weak references.
+ // Here we assert that there are none. If support is added later then
// this assertion needs to be removed and this path needs to be tested.
+ // FIXME: The internal cache also isn't going to notice that an on-disk
+ // dylib could resolve a weak bind from the shared cache. Should we just
+ // error on all caches, regardless of dev/customer?
if (!noMissingWeakSuperclasses) {
diag.error("Some Objective-C class has a superclass that is "
"weak-import and missing from the cache.");
@@ -691,7 +836,7 @@
diag.verbose(" recorded % 6ld classes\n", classes.classNames().size());
- uint64_t clsoptVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
+ uint64_t clsoptVMAddr = cacheAccessor.vmAddrForContent(optROData);
objc_opt::objc_clsopt_t *clsopt = new(optROData) objc_opt::objc_clsopt_t;
err = clsopt->write(clsoptVMAddr, optRORemaining,
classes.classNames(), classes.classes(), false);
@@ -700,6 +845,7 @@
return;
}
optROData += clsopt->size();
+ optROData = alignPointer(optROData);
optRORemaining -= clsopt->size();
size_t duplicateCount = clsopt->duplicateCount();
uint32_t clsoptCapacity = clsopt->capacity;
@@ -732,7 +878,7 @@
// This is SAFE: no protocol references are updated yet
// This must be done AFTER updating method lists.
- ProtocolOptimizer<P> protocolOptimizer(diag);
+ ProtocolOptimizer<P> protocolOptimizer(diag, hinfoROOptimizer);
for (const macho_header<P>* mh : sizeSortedDylibs) {
protocolOptimizer.addProtocols(&cacheAccessor, mh);
}
@@ -750,24 +896,29 @@
return;
}
- uint64_t protocoloptVMAddr = optROSection->addr() + optROSection->size() - optRORemaining;
- objc_opt::objc_protocolopt_t *protocolopt = new (optROData) objc_opt::objc_protocolopt_t;
- err = protocolopt->write(protocoloptVMAddr, optRORemaining,
- protocolOptimizer.protocolNames(),
- protocolOptimizer.protocols(), true);
+ // Align the buffer again. The new protocols may have added an odd number of name characters
+ optROData = alignPointer(optROData);
+
+ // New protocol table which tracks loaded images.
+ uint64_t protocoloptVMAddr = cacheAccessor.vmAddrForContent(optROData);
+ objc_opt::objc_protocolopt2_t *protocolopt = new (optROData) objc_opt::objc_protocolopt2_t;
+ err = protocolopt->write(protocoloptVMAddr, optRORemaining,
+ protocolOptimizer.protocolNames(),
+ protocolOptimizer.protocolsAndHeaders(), false);
if (err) {
diag.warning("%s", err);
return;
}
optROData += protocolopt->size();
+ optROData = alignPointer(optROData);
optRORemaining -= protocolopt->size();
uint32_t protocoloptCapacity = protocolopt->capacity;
uint32_t protocoloptOccupied = protocolopt->occupied;
protocolopt->byteswap(E::little_endian), protocolopt = NULL;
diag.verbose(" protocol table occupancy %u/%u (%u%%)\n",
- protocoloptOccupied, protocoloptCapacity,
- (unsigned)(protocoloptOccupied/(double)protocoloptCapacity*100));
+ protocoloptOccupied, protocoloptCapacity,
+ (unsigned)(protocoloptOccupied/(double)protocoloptCapacity*100));
// Redirect protocol references to the uniqued protocols.
@@ -815,25 +966,32 @@
}
- // Success. Update RO header last.
- E::set32(optROHeader->flags, headerFlags);
- E::set32(optROHeader->selopt_offset, (uint32_t)(seloptVMAddr - optROSection->addr()));
- E::set32(optROHeader->clsopt_offset, (uint32_t)(clsoptVMAddr - optROSection->addr()));
- E::set32(optROHeader->protocolopt_offset, (uint32_t)(protocoloptVMAddr - optROSection->addr()));
- E::set32(optROHeader->headeropt_ro_offset, (uint32_t)(hinfoROVMAddr - optROSection->addr()));
- E::set32(optROHeader->headeropt_rw_offset, (uint32_t)(hinfoRWVMAddr - optROSection->addr()));
+ // Success. Update __objc_opt_ro section in libobjc.dylib to contain offsets to generated optimization structures
+ objc_opt::objc_opt_t* libROHeader = (objc_opt::objc_opt_t *)cacheAccessor.contentForVMAddr(optROSection->addr());
+ E::set32(libROHeader->flags, headerFlags);
+ E::set32(libROHeader->selopt_offset, (uint32_t)(seloptVMAddr - optROSection->addr()));
+ E::set32(libROHeader->clsopt_offset, (uint32_t)(clsoptVMAddr - optROSection->addr()));
+ E::set32(libROHeader->unused_protocolopt_offset, 0);
+ E::set32(libROHeader->headeropt_ro_offset, (uint32_t)(hinfoROVMAddr - optROSection->addr()));
+ E::set32(libROHeader->headeropt_rw_offset, (uint32_t)(hinfoRWVMAddr - optROSection->addr()));
+ E::set32(libROHeader->protocolopt_offset, (uint32_t)(protocoloptVMAddr - optROSection->addr()));
// Log statistics.
- size_t roSize = optROSection->size() - optRORemaining;
- size_t rwSize = optRWSection->size() - optRWRemaining;
- diag.verbose(" %lu/%llu bytes (%d%%) used in libobjc read-only optimization section\n",
- roSize, optROSection->size(), percent(roSize, optROSection->size()));
- diag.verbose(" %lu/%llu bytes (%d%%) used in libobjc read/write optimization section\n",
- rwSize, optRWSection->size(), percent(rwSize, optRWSection->size()));
+ size_t roSize = objcReadOnlyBufferSizeAllocated - optRORemaining;
+ size_t rwSize = objcReadWriteBufferSizeAllocated - optRWRemaining;
+ diag.verbose(" %lu/%llu bytes (%d%%) used in shared cache read-only optimization region\n",
+ roSize, objcReadOnlyBufferSizeAllocated, percent(roSize, objcReadOnlyBufferSizeAllocated));
+ diag.verbose(" %lu/%llu bytes (%d%%) used in shared cache read/write optimization region\n",
+ rwSize, objcReadWriteBufferSizeAllocated, percent(rwSize, objcReadWriteBufferSizeAllocated));
diag.verbose(" wrote objc metadata optimization version %d\n", objc_opt::VERSION);
+
+ // Add segments to libobjc.dylib that cover cache builder allocated r/o and r/w regions
+ addObjcSegments<P>(diag, cache, libobjcMH, objcReadOnlyBuffer, objcReadOnlyBufferSizeAllocated, objcReadWriteBuffer, objcReadWriteBufferSizeAllocated, objcRwFileOffset);
+
// Now that objc has uniqued the selector references, we can apply the LOHs so that ADRP/LDR -> ADRP/ADD
if (forProduction) {
+ const bool logSelectors = false;
uint64_t lohADRPCount = 0;
uint64_t lohLDRCount = 0;
@@ -862,7 +1020,8 @@
if (pass == 0) {
if ( (newPage21 > 2097151) || (newPage21 < -2097151) ) {
- diag.verbose("Out of bounds ADRP selector reference target\n");
+ if (logSelectors)
+ fprintf(stderr, "Out of bounds ADRP selector reference target\n");
instructions.clear();
break;
}
@@ -881,13 +1040,15 @@
if (pass == 0) {
if ( (instruction & 0xC0C00000) != 0xC0400000 ) {
// Not a load, or dest reg isn't xN, or uses sign extension
- diag.verbose("Bad LDR for selector reference optimisation\n");
+ if (logSelectors)
+ fprintf(stderr, "Bad LDR for selector reference optimisation\n");
instructions.clear();
break;
}
if ( (instruction & 0x04000000) != 0 ) {
// Loading a float
- diag.verbose("Bad LDR for selector reference optimisation\n");
+ if (logSelectors)
+ fprintf(stderr, "Bad LDR for selector reference optimisation\n");
instructions.clear();
break;
}
@@ -912,12 +1073,14 @@
if ( (instruction & 0xFFC00000) == 0x91000000 ) {
// ADD imm12
// We don't support ADDs.
- diag.verbose("Bad ADD for selector reference optimisation\n");
+ if (logSelectors)
+ fprintf(stderr, "Bad ADD for selector reference optimisation\n");
instructions.clear();
break;
}
- diag.verbose("Unknown instruction for selref optimisation\n");
+ if (logSelectors)
+ fprintf(stderr, "Unknown instruction for selref optimisation\n");
instructions.clear();
break;
}
@@ -941,10 +1104,36 @@
void CacheBuilder::optimizeObjC()
{
+ uint32_t objcRwFileOffset = (uint32_t)((_objcReadWriteBuffer - _readWriteRegion.buffer) + _readWriteRegion.cacheFileOffset);
if ( _archLayout->is64 )
- doOptimizeObjC<Pointer64<LittleEndian>>((DyldSharedCache*)_readExecuteRegion.buffer, _options.optimizeStubs, _aslrTracker, _lohTracker, _missingWeakImports, _diagnostics);
+ doOptimizeObjC<Pointer64<LittleEndian>>((DyldSharedCache*)_readExecuteRegion.buffer, _options.optimizeStubs, _aslrTracker, _lohTracker,
+ _coalescedText, _missingWeakImports,
+ _diagnostics, _objcReadOnlyBuffer, _objcReadOnlyBufferSizeUsed, _objcReadOnlyBufferSizeAllocated,
+ _objcReadWriteBuffer, _objcReadWriteBufferSizeAllocated, objcRwFileOffset);
else
- doOptimizeObjC<Pointer32<LittleEndian>>((DyldSharedCache*)_readExecuteRegion.buffer, _options.optimizeStubs, _aslrTracker, _lohTracker, _missingWeakImports, _diagnostics);
+ doOptimizeObjC<Pointer32<LittleEndian>>((DyldSharedCache*)_readExecuteRegion.buffer, _options.optimizeStubs, _aslrTracker, _lohTracker,
+ _coalescedText, _missingWeakImports,
+ _diagnostics, _objcReadOnlyBuffer, _objcReadOnlyBufferSizeUsed, _objcReadOnlyBufferSizeAllocated,
+ _objcReadWriteBuffer, _objcReadWriteBufferSizeAllocated, objcRwFileOffset);
}
-
+static uint32_t hashTableSize(uint32_t maxElements, uint32_t perElementData)
+{
+ uint32_t elementsWithPadding = maxElements*11/10; // if close to power of 2, perfect hash may fail, so don't get within 10% of that
+ uint32_t powTwoCapacity = 1 << (32 - __builtin_clz(elementsWithPadding - 1));
+ uint32_t headerSize = 4*(8+256);
+ return headerSize + powTwoCapacity/2 + powTwoCapacity + powTwoCapacity*perElementData;
+}
+
+// The goal here is to allocate space in the dyld shared cache (while it is being laid out) that will contain
+// the objc structures that previously were in the __objc_opt_ro section.
+uint32_t CacheBuilder::computeReadOnlyObjC(uint32_t selRefCount, uint32_t classDefCount, uint32_t protocolDefCount)
+{
+ return 0xA000 + hashTableSize(selRefCount, 5) + hashTableSize(classDefCount, 12) + hashTableSize(protocolDefCount, 8);
+}
+
+// Space to replace the __objc_opt_rw section.
+uint32_t CacheBuilder::computeReadWriteObjC(uint32_t imageCount, uint32_t protocolDefCount)
+{
+ return 8*imageCount + protocolDefCount*12*(_archLayout->is64 ? 8 : 4);
+}