Loading...
--- dyld/dyld-1340/mach_o/CompactUnwind.cpp
+++ dyld/dyld-1245.1/mach_o/CompactUnwind.cpp
@@ -348,9 +348,6 @@
case 1:
permunreg[0] = permutation;
break;
- default:
- strlcat(strBuf, "unsupported registers saved", 128);
- return;
}
// renumber registers back to standard numbers
int registers[6];
@@ -437,10 +434,7 @@
}
const uint32_t firstLevelIndex = low;
const uint32_t firstLevelFunctionOffset = firstLevelTable[firstLevelIndex].functionOffset;
- const uint32_t firstLevelEndFunctionOffset =
- (firstLevelIndex+1) >= _unwindTable->indexCount
- ? firstLevelFunctionOffset + 1
- : firstLevelTable[firstLevelIndex+1].functionOffset;
+ const uint32_t firstLevelEndFunctionOffset = firstLevelTable[firstLevelIndex+1].functionOffset;
const void* secondLevelAddr = (uint8_t*)_unwindTable + firstLevelTable[firstLevelIndex].secondLevelPagesSectionOffset;
if ( targetFunctionOffset > firstLevelEndFunctionOffset )
@@ -460,6 +454,7 @@
if ( entries[mid].functionOffset <= targetFunctionOffset ) {
if ( (mid == last) || (entries[mid+1].functionOffset > targetFunctionOffset) ) {
// next is past target address, so we found it
+ low = mid;
result.funcOffset = entries[mid].functionOffset;
result.encoding = entries[mid].encoding;
result.lsdaOffset = 0;
@@ -488,6 +483,7 @@
uint32_t mid = (low + high)/2;
if ( UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(entries[mid]) <= targetOffset ) {
if ( (mid == last) || (UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(entries[mid+1]) > targetOffset) ) {
+ low = mid;
result.funcOffset = UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(entries[mid]) + firstLevelFunctionOffset;
uint8_t encodingIndex = UNWIND_INFO_COMPRESSED_ENTRY_ENCODING_INDEX(entries[mid]);
if ( encodingIndex < _unwindTable->commonEncodingsArrayCount )
@@ -550,6 +546,372 @@
}
+#if BUILDING_MACHO_WRITER
+
+bool CompactUnwind::encodingMeansUseDwarf(Architecture arch, uint32_t encoding)
+{
+ if ( arch.usesArm64Instructions() )
+ return ((encoding & UNWIND_ARM64_MODE_MASK) == UNWIND_ARM64_MODE_DWARF);
+ else if ( arch.usesx86_64Instructions() )
+ return ((encoding & UNWIND_X86_64_MODE_MASK) == UNWIND_X86_64_MODE_DWARF);
+ assert(0 && "arch not supported for compact unwind");
+}
+
+bool CompactUnwind::encodingCannotBeMerged(Architecture arch, uint32_t encoding)
+{
+ if ( arch.usesx86_64Instructions() )
+ return ((encoding & UNWIND_X86_64_MODE_MASK) == UNWIND_X86_64_MODE_STACK_IND);
+ return false;
+}
+
+// there are two bits in compact unwind that encode which personality function is used
+// this function keeps track of which personality functions are used and when their 2-bit index is
+void CompactUnwind::updatePersonalityForEntry(WriterUnwindInfo& entry, std::vector<UniquePersonality>& personalities)
+{
+ if ( (entry.personalityHandle != nullptr) || (entry.personalityOffset != 0) ) {
+ std::optional<uint32_t> index;
+ for ( const UniquePersonality& personality : personalities ) {
+ if ( personality.handle == entry.personalityHandle ) {
+ index = &personality - personalities.data();
+ break;
+ }
+ else if ( (personality.handle == 0) && (entry.personalityHandle == 0)
+ && (personality.offset != 0) && (personality.offset == entry.personalityOffset) ) {
+ index = &personality - personalities.data();
+ break;
+ }
+ }
+ if ( !index.has_value() ) {
+ index = personalities.size();
+ personalities.push_back({ entry.personalityOffset, entry.personalityHandle });
+ }
+ // update entry with personality index
+ entry.encoding |= ((index.value()+ 1) << (__builtin_ctz(UNWIND_PERSONALITY_MASK)) );
+ }
+}
+
+void CompactUnwind::compressDuplicates(Architecture arch, std::vector<WriterUnwindInfo>& entries, uint32_t& lsdaCount,
+ CommonEncodingsMap& commonEncodings, std::vector<UniquePersonality>& personalities)
+{
+ lsdaCount = 0;
+ // build a vector removing entries where next function has same encoding
+ WriterUnwindInfo last = { ~0U, ~0U, ~0U, ~0U, nullptr, nullptr, nullptr };
+ // encoding frequency to build common encodings
+ size_t inEntriesSize = entries.size();
+ std::unordered_map<compact_unwind_encoding_t, unsigned int> encodingsUsed;
+ std::erase_if(entries, [&](WriterUnwindInfo& entry) {
+ this->updatePersonalityForEntry(entry, personalities);
+ bool newNeedsDwarf = encodingMeansUseDwarf(arch, entry.encoding);
+ bool cannotBeMerged = encodingCannotBeMerged(arch, entry.encoding);
+ bool duplicate = true;
+ // remove entries which have same encoding and personalityPointer as last one
+ if ( newNeedsDwarf || (entry.encoding != last.encoding) || (entry.personalityHandle != last.personalityHandle)
+ || cannotBeMerged || (entry.lsdaHandle != nullptr) ) {
+ duplicate = false;
+
+ // never put dwarf into common table
+ if ( !newNeedsDwarf )
+ encodingsUsed[entry.encoding] += 1;
+ }
+ if ( entry.encoding & UNWIND_HAS_LSDA ) {
+ ++lsdaCount;
+ assert(entry.lsdaHandle != nullptr);
+ }
+ last = entry;
+ return duplicate;
+ });
+
+ using EncodingsAndUsage = std::pair<compact_unwind_encoding_t, unsigned int>;
+ // put encodings into a vector and sort them descending by frequency and
+ // ascending by the encoding value
+ // there's a limited number of unique encodings but many entries so it's
+ // faster to use an unordered map for encodings and sort it here
+ std::vector<EncodingsAndUsage> encodingsByUsage;
+ encodingsByUsage.resize(encodingsUsed.size());
+ std::copy(encodingsUsed.begin(), encodingsUsed.end(), encodingsByUsage.begin());
+ std::sort(encodingsByUsage.begin(), encodingsByUsage.end(),
+ [](const EncodingsAndUsage& l, const EncodingsAndUsage& r) {
+ if ( l.second != r.second )
+ return l.second > r.second;
+ /* sort by encoding time for same number of usages for deterministic output */
+ return l.first < r.first;
+ });
+ // put the most common encodings into the common table, but at most 127 of them
+ uint32_t maxNumCommonEncodings = std::min((uint32_t)encodingsByUsage.size(), 127u);
+ for ( uint32_t i = 0; i < maxNumCommonEncodings; ++i ) {
+ if ( encodingsByUsage[i].second <= 1 )
+ break;
+ commonEncodings[encodingsByUsage[i].first] = i;
+ }
+ if (_verbose) fprintf(stderr, "compressDuplicates() entries.size()=%lu, uniqueEntries.size()=%lu, lsdaCount=%u\n",
+ inEntriesSize, entries.size(), lsdaCount);
+ if (_verbose) fprintf(stderr, "compressDuplicates() %lu common encodings found\n", commonEncodings.size());
+}
+
+uint8_t CompactUnwind::encodingIndex(uint32_t encoding, const CommonEncodingsMap& commonEncodings, const CommonEncodingsMap& pageSpecificEncodings)
+{
+ const auto& pos = commonEncodings.find(encoding);
+ if ( pos != commonEncodings.end() )
+ return pos->second;
+ else
+ return pageSpecificEncodings.at(encoding);
+}
+
+void CompactUnwind::makeRegularSecondLevelPage(const std::vector<WriterUnwindInfo>& uniqueInfos, uint32_t pageSize,
+ size_t& curInfosIndex, uint8_t*& pageStart, unwind_info_section_header_lsda_index_entry*& lsdaContent)
+{
+ const size_t maxEntriesPerPage = (pageSize - sizeof(unwind_info_regular_second_level_page_header))/sizeof(unwind_info_regular_second_level_entry);
+ const size_t entriesToAdd = std::min(maxEntriesPerPage, uniqueInfos.size() - curInfosIndex);
+
+ unwind_info_regular_second_level_page_header* pageHeader = (unwind_info_regular_second_level_page_header*)pageStart;
+ pageHeader->kind = UNWIND_SECOND_LEVEL_REGULAR;
+ pageHeader->entryPageOffset = sizeof(unwind_info_regular_second_level_page_header);
+ pageHeader->entryCount = entriesToAdd;
+
+ unwind_info_regular_second_level_entry* entryArray = (unwind_info_regular_second_level_entry*)((uint8_t*)pageHeader + pageHeader->entryPageOffset);
+ for (uint32_t i=0; i < entriesToAdd; ++i) {
+ const WriterUnwindInfo& info = uniqueInfos[curInfosIndex + i];
+ entryArray[i].functionOffset = info.funcOffset;
+ entryArray[i].encoding = info.encoding;
+ uint64_t entrySectionOffset = (uint8_t*)&entryArray[i].functionOffset - (uint8_t*)&_bytes[0];
+ this->_imageOffsetFixups.push_back({ info.funcHandle, (uint32_t)entrySectionOffset, false });
+ if ( info.encoding & UNWIND_HAS_LSDA ) {
+ lsdaContent->functionOffset = info.funcOffset;
+ lsdaContent->lsdaOffset = info.lsdaOffset;
+ assert(info.lsdaHandle != nullptr);
+
+ uint64_t sectionOffset = (uint8_t*)&lsdaContent->functionOffset - (uint8_t*)&_bytes[0];
+ this->_imageOffsetFixups.push_back({ info.funcHandle, (uint32_t)sectionOffset, false });
+
+ sectionOffset = (uint8_t*)&lsdaContent->lsdaOffset - (uint8_t*)&_bytes[0];
+ this->_imageOffsetFixups.push_back({ info.lsdaHandle, (uint32_t)sectionOffset, false });
+
+ ++lsdaContent;
+ }
+ }
+
+ // update what has been processed
+ curInfosIndex += entriesToAdd;
+ pageStart += (pageHeader->entryPageOffset + pageHeader->entryCount *sizeof(unwind_info_regular_second_level_entry));
+}
+
+void CompactUnwind::makeCompressedSecondLevelPage(const std::vector<WriterUnwindInfo>& uniqueInfos, const CommonEncodingsMap& commonEncodings,
+ uint32_t pageSize, size_t& curInfosIndex, uint8_t*& pageStart, unwind_info_section_header_lsda_index_entry*& lsdaContent)
+{
+ // first pass calculates how many compressed entries we could fit in this sized page
+ // keep adding entries to page until:
+ // 1) encoding table plus entry table plus header exceed page size
+ // 2) the file offset delta from the first to last function > 24 bits
+ // 3) custom encoding index reaches 255
+ // 4) run out of uniqueInfos to encode
+ CommonEncodingsMap pageSpecificEncodings;
+ uint32_t space = pageSize - sizeof(unwind_info_compressed_second_level_page_header);
+ uint32_t entryCount = 0;
+ while ( curInfosIndex + entryCount < uniqueInfos.size() // 4) run out of uniqueInfos to encode
+ && space >= sizeof(uint32_t) ) { // 1) enough room to encode a compressed entry
+ const WriterUnwindInfo& info = uniqueInfos[curInfosIndex + entryCount];
+ if ( commonEncodings.find(info.encoding) == commonEncodings.end() ) {
+ if ( pageSpecificEncodings.find(info.encoding) == pageSpecificEncodings.end() ) {
+ // 1) enough room for the new encoding and the entry, no point adding the encoding
+ // only if there won't be place for the entry
+ if ( space < (sizeof(uint32_t) * 2) )
+ break;
+
+ // need to add page specific encoding
+ uint32_t nextEncodingIndex = (uint32_t)(commonEncodings.size() + pageSpecificEncodings.size());
+ if ( nextEncodingIndex <= 255 ) {
+ pageSpecificEncodings[info.encoding] = nextEncodingIndex;
+ space -= sizeof(uint32_t);
+ } else {
+ break; // 3) custom encoding index reaches 255
+ }
+ }
+ }
+ // compute function offset
+ assert(info.funcOffset >= uniqueInfos[curInfosIndex].funcOffset);
+ uint32_t funcOffsetWithInPage = info.funcOffset - uniqueInfos[curInfosIndex].funcOffset;
+ if ( funcOffsetWithInPage > 0x00FF0000 ) {
+ // don't use 0x00FFFFFF because addresses may vary after atoms are laid out again
+ break; // 2) the file offset delta from the first to last function > 24 bits
+ }
+ ++entryCount;
+ space -= sizeof(uint32_t);
+ }
+
+ // fallback to regular encoding when eligible compressed entries don't use all the available page space,
+ // this isn't the last page and the number of the eligible entries is smaller
+ // than the number of regular entries that can be encoded in this page
+ if ( space >= minPageSize && (curInfosIndex + entryCount) < uniqueInfos.size() ) {
+ const size_t maxEntriesPerPage = (pageSize - sizeof(unwind_info_regular_second_level_page_header))/sizeof(unwind_info_regular_second_level_entry);
+ if ( entryCount < maxEntriesPerPage ) {
+ makeRegularSecondLevelPage(uniqueInfos, pageSize, curInfosIndex, pageStart, lsdaContent);
+ return;
+ }
+ }
+
+ // second pass fills in page
+ unwind_info_compressed_second_level_page_header* pageHeader = (unwind_info_compressed_second_level_page_header*)pageStart;
+ pageHeader->kind = UNWIND_SECOND_LEVEL_COMPRESSED;
+ pageHeader->entryPageOffset = sizeof(unwind_info_compressed_second_level_page_header);
+ pageHeader->entryCount = entryCount;
+ pageHeader->encodingsPageOffset = pageHeader->entryPageOffset + entryCount*sizeof(uint32_t);
+ pageHeader->encodingsCount = pageSpecificEncodings.size();
+ uint32_t* const entriesArray = (uint32_t*)((uint8_t*)pageHeader + pageHeader->entryPageOffset);
+ uint32_t firstFuncOffset = uniqueInfos[curInfosIndex].funcOffset;
+ const void* firstFuncHandle = uniqueInfos[curInfosIndex].funcHandle;
+ for (uint32_t i=0; i < entryCount; ++i) {
+ const WriterUnwindInfo& info = uniqueInfos[curInfosIndex + i];
+ uint32_t offset = info.funcOffset - firstFuncOffset;
+ uint8_t eIndex = encodingIndex(info.encoding, commonEncodings, pageSpecificEncodings);
+ entriesArray[i] = (offset & 0x00FFFFFF) | (eIndex << 24);
+ uint64_t sectionOffset = (uint8_t*)&entriesArray[i] - (uint8_t*)&_bytes[0];
+ this->_diff24Fixups.push_back({ info.funcHandle, firstFuncHandle, (uint32_t)sectionOffset });
+
+ if ( info.encoding & UNWIND_HAS_LSDA ) {
+ lsdaContent->functionOffset = info.funcOffset;
+ lsdaContent->lsdaOffset = info.lsdaOffset;
+ assert(info.lsdaHandle != nullptr);
+
+ sectionOffset = (uint8_t*)&lsdaContent->functionOffset - (uint8_t*)&_bytes[0];
+ this->_imageOffsetFixups.push_back({ info.funcHandle, (uint32_t)sectionOffset, false });
+
+ sectionOffset = (uint8_t*)&lsdaContent->lsdaOffset - (uint8_t*)&_bytes[0];
+ this->_imageOffsetFixups.push_back({ info.lsdaHandle, (uint32_t)sectionOffset, false });
+
+ ++lsdaContent;
+ }
+ }
+ uint32_t* const encodingsArray = (uint32_t*)((uint8_t*)pageHeader + pageHeader->encodingsPageOffset);
+ uint32_t const commonEncodingsSize = (uint32_t)commonEncodings.size();
+ for (const auto& enc : pageSpecificEncodings) {
+ encodingsArray[enc.second - commonEncodingsSize] = enc.first;
+ }
+
+ // update what has been processed
+ curInfosIndex += entryCount;
+ pageStart += (pageHeader->encodingsPageOffset + pageHeader->encodingsCount *sizeof(uint32_t));
+}
+
+
+//
+// FIXME CompactUnwind needs two modes: fast and optimized.
+// Fast uses regular pages every and is easy to size and layout
+// Optimize tries to make the table as small as possible, but that means the size estimation will be expensive
+//
+size_t CompactUnwind::estimateCompactUnwindTableSize(std::span<const WriterUnwindInfo> unwindInfos)
+{
+ std::unordered_set<uint32_t> uniqueEncodings;
+ unsigned lsdaCount = 0;
+ for (const WriterUnwindInfo& entry : unwindInfos) {
+ uniqueEncodings.insert(entry.encoding);
+ if ( entry.encoding & UNWIND_HAS_LSDA )
+ ++lsdaCount;
+ }
+ //fprintf(stderr, "ext: unwindInfos.size=%lu uniqueEncodings.size=%lu\n", unwindInfos.size(), uniqueEncodings.size());
+ // calculate worst case size where all pages are regular
+ return 64 + 20 + unwindInfos.size()*8 + lsdaCount*8 + unwindInfos.size()/32 + uniqueEncodings.size()*4;
+}
+
+// Note: unwindInfos must come in sorted by functionOffset
+CompactUnwind::CompactUnwind(Architecture arch, std::vector<WriterUnwindInfo> unwindInfos)
+{
+ // build new compressed list by removing entries where next function has same encoding
+ // put the most common encodings into the common table, but at most 127 of them
+ // build up vector of personality functions used, with an index for each
+ uint32_t lsdaCount;
+ CommonEncodingsMap commonEncodings;
+ std::vector<UniquePersonality> personalities;
+ compressDuplicates(arch, unwindInfos, lsdaCount, commonEncodings, personalities);
+ // FIXME: need a way to error out if there are more than 3 personality functions used
+
+ // calculate worst case size for all unwind info pages when allocating buffer
+ const size_t entriesPerRegularPage = (maxPageSize-sizeof(unwind_info_regular_second_level_page_header))/sizeof(unwind_info_regular_second_level_entry);
+ const size_t pageCountUpperBound = ((unwindInfos.size() - 1)/entriesPerRegularPage) + 3;
+ _bytes.resize(estimateCompactUnwindTableSize(unwindInfos));
+
+ // fill in section header
+ unwind_info_section_header* header = (unwind_info_section_header*)&_bytes[0];
+ header->version = UNWIND_SECTION_VERSION;
+ header->commonEncodingsArraySectionOffset = sizeof(unwind_info_section_header);
+ header->commonEncodingsArrayCount = (uint32_t)commonEncodings.size();
+ header->personalityArraySectionOffset = header->commonEncodingsArraySectionOffset + (uint32_t)(commonEncodings.size()*sizeof(compact_unwind_encoding_t));
+ header->personalityArrayCount = (uint32_t)personalities.size();
+ header->indexSectionOffset = header->personalityArraySectionOffset + (uint32_t)(personalities.size()*sizeof(uint32_t));
+ header->indexCount = 0; // fill in after second level pages built
+
+ // fill in commmon encodings
+ uint32_t* commonEncodingsArray = (uint32_t*)&_bytes[header->commonEncodingsArraySectionOffset];
+ for (const auto& enc : commonEncodings ) {
+ assert(enc.second < header->commonEncodingsArrayCount);
+ commonEncodingsArray[enc.second] = enc.first;
+ }
+
+ // fill in personalities
+ uint32_t* personalityArray = (uint32_t*)&_bytes[header->personalityArraySectionOffset];
+ for (const auto& p : personalities) {
+ size_t index = &p - personalities.data();
+ personalityArray[index] = p.offset;
+
+ uint64_t sectionOffset = (uint8_t*)&personalityArray[index] - (uint8_t*)header;
+ this->_imageOffsetFixups.push_back({ p.handle, (uint32_t)sectionOffset, false });
+ }
+
+ // build second level pages and fill in first level as each is built
+ unwind_info_section_header_index_entry* firstLevelTable = (unwind_info_section_header_index_entry*)&_bytes[header->indexSectionOffset];
+ unwind_info_section_header_lsda_index_entry* lsdaContent = (unwind_info_section_header_lsda_index_entry*)&_bytes[header->indexSectionOffset+pageCountUpperBound*sizeof(unwind_info_section_header_index_entry)];
+ uint8_t* secondLevelContent = (uint8_t*)&lsdaContent[lsdaCount];
+ uint8_t* const firstSecondContent = secondLevelContent;
+ size_t curInfosIndex = 0;
+ // reserve approximate buffers for fixup vectors
+ this->_imageOffsetFixups.reserve(unwindInfos.size() / 2);
+ this->_diff24Fixups.reserve(unwindInfos.size() / 2);
+
+ while (curInfosIndex < unwindInfos.size()) {
+ uint64_t sectionOffset = (uint8_t*)&firstLevelTable[header->indexCount].functionOffset - (uint8_t*)header;
+ this->_imageOffsetFixups.push_back({ unwindInfos[curInfosIndex].funcHandle,
+ (uint32_t)sectionOffset,
+ false });
+
+ firstLevelTable[header->indexCount].functionOffset = unwindInfos[curInfosIndex].funcOffset;
+ firstLevelTable[header->indexCount].secondLevelPagesSectionOffset = (uint32_t)(secondLevelContent - &_bytes[0]);
+ firstLevelTable[header->indexCount].lsdaIndexArraySectionOffset = (uint32_t)((uint8_t*)lsdaContent - &_bytes[0]);
+ makeCompressedSecondLevelPage(unwindInfos, commonEncodings, maxPageSize, curInfosIndex, secondLevelContent, lsdaContent);
+
+ header->indexCount++;
+ // 8-byte align next page
+ secondLevelContent = (uint8_t*)(((uintptr_t)secondLevelContent+7) & (-8));
+ }
+ // add extra top level index to denote the end
+ {
+ firstLevelTable[header->indexCount].functionOffset = unwindInfos.back().funcOffset;
+ firstLevelTable[header->indexCount].secondLevelPagesSectionOffset = 0;
+ firstLevelTable[header->indexCount].lsdaIndexArraySectionOffset = (uint32_t)(firstSecondContent - &_bytes[0]);
+
+ uint64_t sectionOffset = (uint8_t*)&firstLevelTable[header->indexCount].functionOffset - (uint8_t*)header;
+ this->_imageOffsetFixups.push_back({
+ unwindInfos.back().funcHandle,
+ (uint32_t)sectionOffset,
+ true
+ });
+
+ header->indexCount++;
+ }
+
+ assert(header->indexCount <= pageCountUpperBound && "not enough space reserved for compact unwind first level table");
+
+ // update pointers to the constructed table can be used
+ //fprintf(stderr, "est-size=%lu, act-size=%lu, ext2=%lu\n", _bytes.size(), secondLevelContent-&_bytes[0], estimateCompactUnwindTableSize(unwindInfos));
+ assert(secondLevelContent <= &_bytes[_bytes.size()]);
+ _bytes.resize(secondLevelContent-&_bytes[0]);
+ _unwindTable = header;
+ _unwindTableSize = _bytes.size();
+}
+
+
+
+#endif // BUILDING_MACHO_WRITER
+
+
} // namespace mach_o