1 //===-- LTOModule.cpp - LLVM Link Time Optimizer --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Link Time Optimization library. This library is
11 // intended to be used by linker to optimize code at link time.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/LTO/LTOModule.h"
16 #include "llvm/ADT/Triple.h"
17 #include "llvm/Bitcode/ReaderWriter.h"
18 #include "llvm/CodeGen/Analysis.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/DiagnosticPrinter.h"
21 #include "llvm/IR/LLVMContext.h"
22 #include "llvm/IR/Mangler.h"
23 #include "llvm/IR/Metadata.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/MC/MCExpr.h"
26 #include "llvm/MC/MCInst.h"
27 #include "llvm/MC/MCInstrInfo.h"
28 #include "llvm/MC/MCParser/MCAsmParser.h"
29 #include "llvm/MC/MCSection.h"
30 #include "llvm/MC/MCSubtargetInfo.h"
31 #include "llvm/MC/MCSymbol.h"
32 #include "llvm/MC/MCTargetAsmParser.h"
33 #include "llvm/MC/SubtargetFeature.h"
34 #include "llvm/Object/IRObjectFile.h"
35 #include "llvm/Object/ObjectFile.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/FileSystem.h"
38 #include "llvm/Support/Host.h"
39 #include "llvm/Support/MemoryBuffer.h"
40 #include "llvm/Support/Path.h"
41 #include "llvm/Support/SourceMgr.h"
42 #include "llvm/Support/TargetRegistry.h"
43 #include "llvm/Support/TargetSelect.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetLoweringObjectFile.h"
46 #include "llvm/Target/TargetRegisterInfo.h"
47 #include "llvm/Target/TargetSubtargetInfo.h"
48 #include "llvm/Transforms/Utils/GlobalStatus.h"
49 #include <system_error>
51 using namespace llvm::object;
53 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
54 llvm::TargetMachine *TM)
55 : IRFile(std::move(Obj)), _target(TM) {}
57 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
58 llvm::TargetMachine *TM,
59 std::unique_ptr<LLVMContext> Context)
60 : OwnedContext(std::move(Context)), IRFile(std::move(Obj)), _target(TM) {}
62 LTOModule::~LTOModule() {}
64 /// isBitcodeFile - Returns 'true' if the file (or memory contents) is LLVM
66 bool LTOModule::isBitcodeFile(const void *Mem, size_t Length) {
67 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
68 MemoryBufferRef(StringRef((const char *)Mem, Length), "<mem>"));
72 bool LTOModule::isBitcodeFile(const char *Path) {
73 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
74 MemoryBuffer::getFile(Path);
78 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
79 BufferOrErr.get()->getMemBufferRef());
83 bool LTOModule::isBitcodeForTarget(MemoryBuffer *Buffer,
84 StringRef TriplePrefix) {
85 ErrorOr<MemoryBufferRef> BCOrErr =
86 IRObjectFile::findBitcodeInMemBuffer(Buffer->getMemBufferRef());
90 std::string Triple = getBitcodeTargetTriple(*BCOrErr, Context);
91 return StringRef(Triple).startswith(TriplePrefix);
94 std::string LTOModule::getProducerString(MemoryBuffer *Buffer) {
95 ErrorOr<MemoryBufferRef> BCOrErr =
96 IRObjectFile::findBitcodeInMemBuffer(Buffer->getMemBufferRef());
100 return getBitcodeProducerString(*BCOrErr, Context);
103 LTOModule *LTOModule::createFromFile(const char *path, TargetOptions options,
104 std::string &errMsg) {
105 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
106 MemoryBuffer::getFile(path);
107 if (std::error_code EC = BufferOrErr.getError()) {
108 errMsg = EC.message();
111 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
112 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
113 &getGlobalContext());
116 LTOModule *LTOModule::createFromOpenFile(int fd, const char *path, size_t size,
117 TargetOptions options,
118 std::string &errMsg) {
119 return createFromOpenFileSlice(fd, path, size, 0, options, errMsg);
122 LTOModule *LTOModule::createFromOpenFileSlice(int fd, const char *path,
123 size_t map_size, off_t offset,
124 TargetOptions options,
125 std::string &errMsg) {
126 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
127 MemoryBuffer::getOpenFileSlice(fd, path, map_size, offset);
128 if (std::error_code EC = BufferOrErr.getError()) {
129 errMsg = EC.message();
132 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
133 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
134 &getGlobalContext());
137 LTOModule *LTOModule::createFromBuffer(const void *mem, size_t length,
138 TargetOptions options,
139 std::string &errMsg, StringRef path) {
140 return createInContext(mem, length, options, errMsg, path,
141 &getGlobalContext());
144 LTOModule *LTOModule::createInLocalContext(const void *mem, size_t length,
145 TargetOptions options,
148 return createInContext(mem, length, options, errMsg, path, nullptr);
151 LTOModule *LTOModule::createInContext(const void *mem, size_t length,
152 TargetOptions options,
153 std::string &errMsg, StringRef path,
154 LLVMContext *Context) {
155 StringRef Data((const char *)mem, length);
156 MemoryBufferRef Buffer(Data, path);
157 return makeLTOModule(Buffer, options, errMsg, Context);
160 static std::unique_ptr<Module> parseBitcodeFileImpl(MemoryBufferRef Buffer,
161 LLVMContext &Context,
163 std::string &ErrMsg) {
166 ErrorOr<MemoryBufferRef> MBOrErr =
167 IRObjectFile::findBitcodeInMemBuffer(Buffer);
168 if (std::error_code EC = MBOrErr.getError()) {
169 ErrMsg = EC.message();
173 std::function<void(const DiagnosticInfo &)> DiagnosticHandler =
174 [&ErrMsg](const DiagnosticInfo &DI) {
175 raw_string_ostream Stream(ErrMsg);
176 DiagnosticPrinterRawOStream DP(Stream);
181 // Parse the full file.
182 ErrorOr<std::unique_ptr<Module>> M =
183 parseBitcodeFile(*MBOrErr, Context, DiagnosticHandler);
186 return std::move(*M);
190 std::unique_ptr<MemoryBuffer> LightweightBuf =
191 MemoryBuffer::getMemBuffer(*MBOrErr, false);
192 ErrorOr<std::unique_ptr<Module>> M =
193 getLazyBitcodeModule(std::move(LightweightBuf), Context,
194 DiagnosticHandler, true /*ShouldLazyLoadMetadata*/);
197 return std::move(*M);
200 LTOModule *LTOModule::makeLTOModule(MemoryBufferRef Buffer,
201 TargetOptions options, std::string &errMsg,
202 LLVMContext *Context) {
203 std::unique_ptr<LLVMContext> OwnedContext;
205 OwnedContext = llvm::make_unique<LLVMContext>();
206 Context = OwnedContext.get();
209 // If we own a context, we know this is being used only for symbol
210 // extraction, not linking. Be lazy in that case.
211 std::unique_ptr<Module> M = parseBitcodeFileImpl(
213 /* ShouldBeLazy */ static_cast<bool>(OwnedContext), errMsg);
217 std::string TripleStr = M->getTargetTriple();
218 if (TripleStr.empty())
219 TripleStr = sys::getDefaultTargetTriple();
220 llvm::Triple Triple(TripleStr);
222 // find machine architecture for this module
223 const Target *march = TargetRegistry::lookupTarget(TripleStr, errMsg);
227 // construct LTOModule, hand over ownership of module and target
228 SubtargetFeatures Features;
229 Features.getDefaultSubtargetFeatures(Triple);
230 std::string FeatureStr = Features.getString();
231 // Set a default CPU for Darwin triples.
233 if (Triple.isOSDarwin()) {
234 if (Triple.getArch() == llvm::Triple::x86_64)
236 else if (Triple.getArch() == llvm::Triple::x86)
238 else if (Triple.getArch() == llvm::Triple::aarch64)
242 TargetMachine *target = march->createTargetMachine(TripleStr, CPU, FeatureStr,
244 M->setDataLayout(target->createDataLayout());
246 std::unique_ptr<object::IRObjectFile> IRObj(
247 new object::IRObjectFile(Buffer, std::move(M)));
251 Ret = new LTOModule(std::move(IRObj), target, std::move(OwnedContext));
253 Ret = new LTOModule(std::move(IRObj), target);
255 if (Ret->parseSymbols(errMsg)) {
260 Ret->parseMetadata();
265 /// Create a MemoryBuffer from a memory range with an optional name.
266 std::unique_ptr<MemoryBuffer>
267 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
268 const char *startPtr = (const char*)mem;
269 return MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false);
272 /// objcClassNameFromExpression - Get string that the data pointer points to.
274 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
275 if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
276 Constant *op = ce->getOperand(0);
277 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
278 Constant *cn = gvn->getInitializer();
279 if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
280 if (ca->isCString()) {
281 name = (".objc_class_name_" + ca->getAsCString()).str();
290 /// addObjCClass - Parse i386/ppc ObjC class data structure.
291 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
292 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
295 // second slot in __OBJC,__class is pointer to superclass name
296 std::string superclassName;
297 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
299 _undefines.insert(std::make_pair(superclassName, NameAndAttributes()));
300 if (IterBool.second) {
301 NameAndAttributes &info = IterBool.first->second;
302 info.name = IterBool.first->first().data();
303 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
304 info.isFunction = false;
309 // third slot in __OBJC,__class is pointer to class name
310 std::string className;
311 if (objcClassNameFromExpression(c->getOperand(2), className)) {
312 auto Iter = _defines.insert(className).first;
314 NameAndAttributes info;
315 info.name = Iter->first().data();
316 info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
317 LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
318 info.isFunction = false;
320 _symbols.push_back(info);
324 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
325 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
326 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
329 // second slot in __OBJC,__category is pointer to target class name
330 std::string targetclassName;
331 if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
335 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
337 if (!IterBool.second)
340 NameAndAttributes &info = IterBool.first->second;
341 info.name = IterBool.first->first().data();
342 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
343 info.isFunction = false;
347 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
348 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
349 std::string targetclassName;
350 if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
354 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
356 if (!IterBool.second)
359 NameAndAttributes &info = IterBool.first->second;
360 info.name = IterBool.first->first().data();
361 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
362 info.isFunction = false;
366 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
367 SmallString<64> Buffer;
369 raw_svector_ostream OS(Buffer);
373 const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
374 addDefinedDataSymbol(Buffer.c_str(), V);
377 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
378 // Add to list of defined symbols.
379 addDefinedSymbol(Name, v, false);
381 if (!v->hasSection() /* || !isTargetDarwin */)
384 // Special case i386/ppc ObjC data structures in magic sections:
385 // The issue is that the old ObjC object format did some strange
386 // contortions to avoid real linker symbols. For instance, the
387 // ObjC class data structure is allocated statically in the executable
388 // that defines that class. That data structures contains a pointer to
389 // its superclass. But instead of just initializing that part of the
390 // struct to the address of its superclass, and letting the static and
391 // dynamic linkers do the rest, the runtime works by having that field
392 // instead point to a C-string that is the name of the superclass.
393 // At runtime the objc initialization updates that pointer and sets
394 // it to point to the actual super class. As far as the linker
395 // knows it is just a pointer to a string. But then someone wanted the
396 // linker to issue errors at build time if the superclass was not found.
397 // So they figured out a way in mach-o object format to use an absolute
398 // symbols (.objc_class_name_Foo = 0) and a floating reference
399 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
400 // a class was missing.
401 // The following synthesizes the implicit .objc_* symbols for the linker
402 // from the ObjC data structures generated by the front end.
404 // special case if this data blob is an ObjC class definition
405 std::string Section = v->getSection();
406 if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
407 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
412 // special case if this data blob is an ObjC category definition
413 else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
414 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
419 // special case if this data blob is the list of referenced classes
420 else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
421 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
427 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
428 SmallString<64> Buffer;
430 raw_svector_ostream OS(Buffer);
435 cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
436 addDefinedFunctionSymbol(Buffer.c_str(), F);
439 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
440 // add to list of defined symbols
441 addDefinedSymbol(Name, F, true);
444 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
446 // set alignment part log2() can have rounding errors
447 uint32_t align = def->getAlignment();
448 uint32_t attr = align ? countTrailingZeros(align) : 0;
450 // set permissions part
452 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
454 const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
455 if (gv && gv->isConstant())
456 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
458 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
461 // set definition part
462 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
463 attr |= LTO_SYMBOL_DEFINITION_WEAK;
464 else if (def->hasCommonLinkage())
465 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
467 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
470 if (def->hasLocalLinkage())
471 // Ignore visibility if linkage is local.
472 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
473 else if (def->hasHiddenVisibility())
474 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
475 else if (def->hasProtectedVisibility())
476 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
477 else if (canBeOmittedFromSymbolTable(def))
478 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
480 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
482 if (def->hasComdat())
483 attr |= LTO_SYMBOL_COMDAT;
485 if (isa<GlobalAlias>(def))
486 attr |= LTO_SYMBOL_ALIAS;
488 auto Iter = _defines.insert(Name).first;
490 // fill information structure
491 NameAndAttributes info;
492 StringRef NameRef = Iter->first();
493 info.name = NameRef.data();
494 assert(info.name[NameRef.size()] == '\0');
495 info.attributes = attr;
496 info.isFunction = isFunction;
499 // add to table of symbols
500 _symbols.push_back(info);
503 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
505 void LTOModule::addAsmGlobalSymbol(const char *name,
506 lto_symbol_attributes scope) {
507 auto IterBool = _defines.insert(name);
509 // only add new define if not already defined
510 if (!IterBool.second)
513 NameAndAttributes &info = _undefines[IterBool.first->first().data()];
515 if (info.symbol == nullptr) {
516 // FIXME: This is trying to take care of module ASM like this:
518 // module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
520 // but is gross and its mother dresses it funny. Have the ASM parser give us
521 // more details for this type of situation so that we're not guessing so
524 // fill information structure
525 info.name = IterBool.first->first().data();
527 LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
528 info.isFunction = false;
529 info.symbol = nullptr;
531 // add to table of symbols
532 _symbols.push_back(info);
537 addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
539 addDefinedDataSymbol(info.name, info.symbol);
541 _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
542 _symbols.back().attributes |= scope;
545 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
547 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
548 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
550 _asm_undefines.push_back(IterBool.first->first().data());
552 // we already have the symbol
553 if (!IterBool.second)
556 uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
557 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
558 NameAndAttributes &info = IterBool.first->second;
559 info.name = IterBool.first->first().data();
560 info.attributes = attr;
561 info.isFunction = false;
562 info.symbol = nullptr;
565 /// Add a symbol which isn't defined just yet to a list to be resolved later.
566 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
568 SmallString<64> name;
570 raw_svector_ostream OS(name);
574 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
576 // we already have the symbol
577 if (!IterBool.second)
580 NameAndAttributes &info = IterBool.first->second;
582 info.name = IterBool.first->first().data();
584 const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
586 if (decl->hasExternalWeakLinkage())
587 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
589 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
591 info.isFunction = isFunc;
595 /// parseSymbols - Parse the symbols from the module and model-level ASM and add
596 /// them to either the defined or undefined lists.
597 bool LTOModule::parseSymbols(std::string &errMsg) {
598 for (auto &Sym : IRFile->symbols()) {
599 const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
600 uint32_t Flags = Sym.getFlags();
601 if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
604 bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
607 SmallString<64> Buffer;
609 raw_svector_ostream OS(Buffer);
612 const char *Name = Buffer.c_str();
615 addAsmGlobalSymbolUndef(Name);
616 else if (Flags & object::BasicSymbolRef::SF_Global)
617 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
619 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
623 auto *F = dyn_cast<Function>(GV);
625 addPotentialUndefinedSymbol(Sym, F != nullptr);
630 addDefinedFunctionSymbol(Sym);
634 if (isa<GlobalVariable>(GV)) {
635 addDefinedDataSymbol(Sym);
639 assert(isa<GlobalAlias>(GV));
640 addDefinedDataSymbol(Sym);
643 // make symbols for all undefines
644 for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
645 e = _undefines.end(); u != e; ++u) {
646 // If this symbol also has a definition, then don't make an undefine because
647 // it is a tentative definition.
648 if (_defines.count(u->getKey())) continue;
649 NameAndAttributes info = u->getValue();
650 _symbols.push_back(info);
656 /// parseMetadata - Parse metadata from the module
657 void LTOModule::parseMetadata() {
658 raw_string_ostream OS(LinkerOpts);
661 if (Metadata *Val = getModule().getModuleFlag("Linker Options")) {
662 MDNode *LinkerOptions = cast<MDNode>(Val);
663 for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
664 MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
665 for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
666 MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
667 OS << " " << MDOption->getString();
674 for (const NameAndAttributes &Sym : _symbols) {
677 _target->getObjFileLowering()->emitLinkerFlagsForGlobal(OS, Sym.symbol,
681 // Add other interesting metadata here.