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);
256 Ret->parseMetadata();
261 /// Create a MemoryBuffer from a memory range with an optional name.
262 std::unique_ptr<MemoryBuffer>
263 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
264 const char *startPtr = (const char*)mem;
265 return MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false);
268 /// objcClassNameFromExpression - Get string that the data pointer points to.
270 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
271 if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
272 Constant *op = ce->getOperand(0);
273 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
274 Constant *cn = gvn->getInitializer();
275 if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
276 if (ca->isCString()) {
277 name = (".objc_class_name_" + ca->getAsCString()).str();
286 /// addObjCClass - Parse i386/ppc ObjC class data structure.
287 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
288 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
291 // second slot in __OBJC,__class is pointer to superclass name
292 std::string superclassName;
293 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
295 _undefines.insert(std::make_pair(superclassName, NameAndAttributes()));
296 if (IterBool.second) {
297 NameAndAttributes &info = IterBool.first->second;
298 info.name = IterBool.first->first().data();
299 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
300 info.isFunction = false;
305 // third slot in __OBJC,__class is pointer to class name
306 std::string className;
307 if (objcClassNameFromExpression(c->getOperand(2), className)) {
308 auto Iter = _defines.insert(className).first;
310 NameAndAttributes info;
311 info.name = Iter->first().data();
312 info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
313 LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
314 info.isFunction = false;
316 _symbols.push_back(info);
320 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
321 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
322 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
325 // second slot in __OBJC,__category is pointer to target class name
326 std::string targetclassName;
327 if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
331 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
333 if (!IterBool.second)
336 NameAndAttributes &info = IterBool.first->second;
337 info.name = IterBool.first->first().data();
338 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
339 info.isFunction = false;
343 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
344 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
345 std::string targetclassName;
346 if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
350 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
352 if (!IterBool.second)
355 NameAndAttributes &info = IterBool.first->second;
356 info.name = IterBool.first->first().data();
357 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
358 info.isFunction = false;
362 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
363 SmallString<64> Buffer;
365 raw_svector_ostream OS(Buffer);
369 const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
370 addDefinedDataSymbol(Buffer.c_str(), V);
373 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
374 // Add to list of defined symbols.
375 addDefinedSymbol(Name, v, false);
377 if (!v->hasSection() /* || !isTargetDarwin */)
380 // Special case i386/ppc ObjC data structures in magic sections:
381 // The issue is that the old ObjC object format did some strange
382 // contortions to avoid real linker symbols. For instance, the
383 // ObjC class data structure is allocated statically in the executable
384 // that defines that class. That data structures contains a pointer to
385 // its superclass. But instead of just initializing that part of the
386 // struct to the address of its superclass, and letting the static and
387 // dynamic linkers do the rest, the runtime works by having that field
388 // instead point to a C-string that is the name of the superclass.
389 // At runtime the objc initialization updates that pointer and sets
390 // it to point to the actual super class. As far as the linker
391 // knows it is just a pointer to a string. But then someone wanted the
392 // linker to issue errors at build time if the superclass was not found.
393 // So they figured out a way in mach-o object format to use an absolute
394 // symbols (.objc_class_name_Foo = 0) and a floating reference
395 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
396 // a class was missing.
397 // The following synthesizes the implicit .objc_* symbols for the linker
398 // from the ObjC data structures generated by the front end.
400 // special case if this data blob is an ObjC class definition
401 std::string Section = v->getSection();
402 if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
403 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
408 // special case if this data blob is an ObjC category definition
409 else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
410 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
415 // special case if this data blob is the list of referenced classes
416 else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
417 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
423 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
424 SmallString<64> Buffer;
426 raw_svector_ostream OS(Buffer);
431 cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
432 addDefinedFunctionSymbol(Buffer.c_str(), F);
435 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
436 // add to list of defined symbols
437 addDefinedSymbol(Name, F, true);
440 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
442 // set alignment part log2() can have rounding errors
443 uint32_t align = def->getAlignment();
444 uint32_t attr = align ? countTrailingZeros(align) : 0;
446 // set permissions part
448 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
450 const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
451 if (gv && gv->isConstant())
452 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
454 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
457 // set definition part
458 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
459 attr |= LTO_SYMBOL_DEFINITION_WEAK;
460 else if (def->hasCommonLinkage())
461 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
463 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
466 if (def->hasLocalLinkage())
467 // Ignore visibility if linkage is local.
468 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
469 else if (def->hasHiddenVisibility())
470 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
471 else if (def->hasProtectedVisibility())
472 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
473 else if (canBeOmittedFromSymbolTable(def))
474 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
476 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
478 if (def->hasComdat())
479 attr |= LTO_SYMBOL_COMDAT;
481 if (isa<GlobalAlias>(def))
482 attr |= LTO_SYMBOL_ALIAS;
484 auto Iter = _defines.insert(Name).first;
486 // fill information structure
487 NameAndAttributes info;
488 StringRef NameRef = Iter->first();
489 info.name = NameRef.data();
490 assert(info.name[NameRef.size()] == '\0');
491 info.attributes = attr;
492 info.isFunction = isFunction;
495 // add to table of symbols
496 _symbols.push_back(info);
499 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
501 void LTOModule::addAsmGlobalSymbol(const char *name,
502 lto_symbol_attributes scope) {
503 auto IterBool = _defines.insert(name);
505 // only add new define if not already defined
506 if (!IterBool.second)
509 NameAndAttributes &info = _undefines[IterBool.first->first().data()];
511 if (info.symbol == nullptr) {
512 // FIXME: This is trying to take care of module ASM like this:
514 // module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
516 // but is gross and its mother dresses it funny. Have the ASM parser give us
517 // more details for this type of situation so that we're not guessing so
520 // fill information structure
521 info.name = IterBool.first->first().data();
523 LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
524 info.isFunction = false;
525 info.symbol = nullptr;
527 // add to table of symbols
528 _symbols.push_back(info);
533 addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
535 addDefinedDataSymbol(info.name, info.symbol);
537 _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
538 _symbols.back().attributes |= scope;
541 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
543 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
544 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
546 _asm_undefines.push_back(IterBool.first->first().data());
548 // we already have the symbol
549 if (!IterBool.second)
552 uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
553 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
554 NameAndAttributes &info = IterBool.first->second;
555 info.name = IterBool.first->first().data();
556 info.attributes = attr;
557 info.isFunction = false;
558 info.symbol = nullptr;
561 /// Add a symbol which isn't defined just yet to a list to be resolved later.
562 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
564 SmallString<64> name;
566 raw_svector_ostream OS(name);
570 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
572 // we already have the symbol
573 if (!IterBool.second)
576 NameAndAttributes &info = IterBool.first->second;
578 info.name = IterBool.first->first().data();
580 const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
582 if (decl->hasExternalWeakLinkage())
583 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
585 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
587 info.isFunction = isFunc;
591 void LTOModule::parseSymbols() {
592 for (auto &Sym : IRFile->symbols()) {
593 const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
594 uint32_t Flags = Sym.getFlags();
595 if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
598 bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
601 SmallString<64> Buffer;
603 raw_svector_ostream OS(Buffer);
606 const char *Name = Buffer.c_str();
609 addAsmGlobalSymbolUndef(Name);
610 else if (Flags & object::BasicSymbolRef::SF_Global)
611 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
613 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
617 auto *F = dyn_cast<Function>(GV);
619 addPotentialUndefinedSymbol(Sym, F != nullptr);
624 addDefinedFunctionSymbol(Sym);
628 if (isa<GlobalVariable>(GV)) {
629 addDefinedDataSymbol(Sym);
633 assert(isa<GlobalAlias>(GV));
634 addDefinedDataSymbol(Sym);
637 // make symbols for all undefines
638 for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
639 e = _undefines.end(); u != e; ++u) {
640 // If this symbol also has a definition, then don't make an undefine because
641 // it is a tentative definition.
642 if (_defines.count(u->getKey())) continue;
643 NameAndAttributes info = u->getValue();
644 _symbols.push_back(info);
648 /// parseMetadata - Parse metadata from the module
649 void LTOModule::parseMetadata() {
650 raw_string_ostream OS(LinkerOpts);
653 if (Metadata *Val = getModule().getModuleFlag("Linker Options")) {
654 MDNode *LinkerOptions = cast<MDNode>(Val);
655 for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
656 MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
657 for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
658 MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
659 OS << " " << MDOption->getString();
666 for (const NameAndAttributes &Sym : _symbols) {
669 _target->getObjFileLowering()->emitLinkerFlagsForGlobal(OS, Sym.symbol,
673 // Add other interesting metadata here.