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/Metadata.h"
23 #include "llvm/IR/Module.h"
24 #include "llvm/MC/MCExpr.h"
25 #include "llvm/MC/MCInst.h"
26 #include "llvm/MC/MCInstrInfo.h"
27 #include "llvm/MC/MCParser/MCAsmParser.h"
28 #include "llvm/MC/MCSection.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/MC/MCSymbol.h"
31 #include "llvm/MC/MCTargetAsmParser.h"
32 #include "llvm/MC/SubtargetFeature.h"
33 #include "llvm/Object/IRObjectFile.h"
34 #include "llvm/Object/ObjectFile.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/FileSystem.h"
37 #include "llvm/Support/Host.h"
38 #include "llvm/Support/MemoryBuffer.h"
39 #include "llvm/Support/Path.h"
40 #include "llvm/Support/SourceMgr.h"
41 #include "llvm/Support/TargetRegistry.h"
42 #include "llvm/Support/TargetSelect.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetLoweringObjectFile.h"
45 #include "llvm/Target/TargetRegisterInfo.h"
46 #include "llvm/Target/TargetSubtargetInfo.h"
47 #include "llvm/Transforms/Utils/GlobalStatus.h"
48 #include <system_error>
50 using namespace llvm::object;
52 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
53 llvm::TargetMachine *TM)
54 : IRFile(std::move(Obj)), _target(TM) {}
56 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
57 llvm::TargetMachine *TM,
58 std::unique_ptr<LLVMContext> Context)
59 : OwnedContext(std::move(Context)), IRFile(std::move(Obj)), _target(TM) {}
61 LTOModule::~LTOModule() {}
63 /// isBitcodeFile - Returns 'true' if the file (or memory contents) is LLVM
65 bool LTOModule::isBitcodeFile(const void *Mem, size_t Length) {
66 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
67 MemoryBufferRef(StringRef((const char *)Mem, Length), "<mem>"));
71 bool LTOModule::isBitcodeFile(const char *Path) {
72 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
73 MemoryBuffer::getFile(Path);
77 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
78 BufferOrErr.get()->getMemBufferRef());
82 bool LTOModule::isBitcodeForTarget(MemoryBuffer *Buffer,
83 StringRef TriplePrefix) {
84 ErrorOr<MemoryBufferRef> BCOrErr =
85 IRObjectFile::findBitcodeInMemBuffer(Buffer->getMemBufferRef());
89 std::string Triple = getBitcodeTargetTriple(*BCOrErr, Context);
90 return StringRef(Triple).startswith(TriplePrefix);
93 LTOModule *LTOModule::createFromFile(const char *path, TargetOptions options,
94 std::string &errMsg) {
95 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
96 MemoryBuffer::getFile(path);
97 if (std::error_code EC = BufferOrErr.getError()) {
98 errMsg = EC.message();
101 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
102 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
103 &getGlobalContext());
106 LTOModule *LTOModule::createFromOpenFile(int fd, const char *path, size_t size,
107 TargetOptions options,
108 std::string &errMsg) {
109 return createFromOpenFileSlice(fd, path, size, 0, options, errMsg);
112 LTOModule *LTOModule::createFromOpenFileSlice(int fd, const char *path,
113 size_t map_size, off_t offset,
114 TargetOptions options,
115 std::string &errMsg) {
116 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
117 MemoryBuffer::getOpenFileSlice(fd, path, map_size, offset);
118 if (std::error_code EC = BufferOrErr.getError()) {
119 errMsg = EC.message();
122 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
123 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
124 &getGlobalContext());
127 LTOModule *LTOModule::createFromBuffer(const void *mem, size_t length,
128 TargetOptions options,
129 std::string &errMsg, StringRef path) {
130 return createInContext(mem, length, options, errMsg, path,
131 &getGlobalContext());
134 LTOModule *LTOModule::createInLocalContext(const void *mem, size_t length,
135 TargetOptions options,
138 return createInContext(mem, length, options, errMsg, path, nullptr);
141 LTOModule *LTOModule::createInContext(const void *mem, size_t length,
142 TargetOptions options,
143 std::string &errMsg, StringRef path,
144 LLVMContext *Context) {
145 StringRef Data((const char *)mem, length);
146 MemoryBufferRef Buffer(Data, path);
147 return makeLTOModule(Buffer, options, errMsg, Context);
150 static std::unique_ptr<Module> parseBitcodeFileImpl(MemoryBufferRef Buffer,
151 LLVMContext &Context,
153 std::string &ErrMsg) {
156 ErrorOr<MemoryBufferRef> MBOrErr =
157 IRObjectFile::findBitcodeInMemBuffer(Buffer);
158 if (std::error_code EC = MBOrErr.getError()) {
159 ErrMsg = EC.message();
163 std::function<void(const DiagnosticInfo &)> DiagnosticHandler =
164 [&ErrMsg](const DiagnosticInfo &DI) {
165 raw_string_ostream Stream(ErrMsg);
166 DiagnosticPrinterRawOStream DP(Stream);
171 // Parse the full file.
172 ErrorOr<std::unique_ptr<Module>> M =
173 parseBitcodeFile(*MBOrErr, Context, DiagnosticHandler);
176 return std::move(*M);
180 std::unique_ptr<MemoryBuffer> LightweightBuf =
181 MemoryBuffer::getMemBuffer(*MBOrErr, false);
182 ErrorOr<std::unique_ptr<Module>> M =
183 getLazyBitcodeModule(std::move(LightweightBuf), Context,
184 DiagnosticHandler, true /*ShouldLazyLoadMetadata*/);
187 return std::move(*M);
190 LTOModule *LTOModule::makeLTOModule(MemoryBufferRef Buffer,
191 TargetOptions options, std::string &errMsg,
192 LLVMContext *Context) {
193 std::unique_ptr<LLVMContext> OwnedContext;
195 OwnedContext = llvm::make_unique<LLVMContext>();
196 Context = OwnedContext.get();
199 // If we own a context, we know this is being used only for symbol
200 // extraction, not linking. Be lazy in that case.
201 std::unique_ptr<Module> M = parseBitcodeFileImpl(
203 /* ShouldBeLazy */ static_cast<bool>(OwnedContext), errMsg);
207 std::string TripleStr = M->getTargetTriple();
208 if (TripleStr.empty())
209 TripleStr = sys::getDefaultTargetTriple();
210 llvm::Triple Triple(TripleStr);
212 // find machine architecture for this module
213 const Target *march = TargetRegistry::lookupTarget(TripleStr, errMsg);
217 // construct LTOModule, hand over ownership of module and target
218 SubtargetFeatures Features;
219 Features.getDefaultSubtargetFeatures(Triple);
220 std::string FeatureStr = Features.getString();
221 // Set a default CPU for Darwin triples.
223 if (Triple.isOSDarwin()) {
224 if (Triple.getArch() == llvm::Triple::x86_64)
226 else if (Triple.getArch() == llvm::Triple::x86)
228 else if (Triple.getArch() == llvm::Triple::aarch64)
232 TargetMachine *target = march->createTargetMachine(TripleStr, CPU, FeatureStr,
234 M->setDataLayout(*target->getDataLayout());
236 std::unique_ptr<object::IRObjectFile> IRObj(
237 new object::IRObjectFile(Buffer, std::move(M)));
241 Ret = new LTOModule(std::move(IRObj), target, std::move(OwnedContext));
243 Ret = new LTOModule(std::move(IRObj), target);
245 if (Ret->parseSymbols(errMsg)) {
250 Ret->parseMetadata();
255 /// Create a MemoryBuffer from a memory range with an optional name.
256 std::unique_ptr<MemoryBuffer>
257 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
258 const char *startPtr = (const char*)mem;
259 return MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false);
262 /// objcClassNameFromExpression - Get string that the data pointer points to.
264 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
265 if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
266 Constant *op = ce->getOperand(0);
267 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
268 Constant *cn = gvn->getInitializer();
269 if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
270 if (ca->isCString()) {
271 name = (".objc_class_name_" + ca->getAsCString()).str();
280 /// addObjCClass - Parse i386/ppc ObjC class data structure.
281 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
282 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
285 // second slot in __OBJC,__class is pointer to superclass name
286 std::string superclassName;
287 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
289 _undefines.insert(std::make_pair(superclassName, NameAndAttributes()));
290 if (IterBool.second) {
291 NameAndAttributes &info = IterBool.first->second;
292 info.name = IterBool.first->first().data();
293 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
294 info.isFunction = false;
299 // third slot in __OBJC,__class is pointer to class name
300 std::string className;
301 if (objcClassNameFromExpression(c->getOperand(2), className)) {
302 auto Iter = _defines.insert(className).first;
304 NameAndAttributes info;
305 info.name = Iter->first().data();
306 info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
307 LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
308 info.isFunction = false;
310 _symbols.push_back(info);
314 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
315 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
316 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
319 // second slot in __OBJC,__category is pointer to target class name
320 std::string targetclassName;
321 if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
325 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
327 if (!IterBool.second)
330 NameAndAttributes &info = IterBool.first->second;
331 info.name = IterBool.first->first().data();
332 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
333 info.isFunction = false;
337 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
338 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
339 std::string targetclassName;
340 if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
344 _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
346 if (!IterBool.second)
349 NameAndAttributes &info = IterBool.first->second;
350 info.name = IterBool.first->first().data();
351 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
352 info.isFunction = false;
356 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
357 SmallString<64> Buffer;
359 raw_svector_ostream OS(Buffer);
363 const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
364 addDefinedDataSymbol(Buffer.c_str(), V);
367 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
368 // Add to list of defined symbols.
369 addDefinedSymbol(Name, v, false);
371 if (!v->hasSection() /* || !isTargetDarwin */)
374 // Special case i386/ppc ObjC data structures in magic sections:
375 // The issue is that the old ObjC object format did some strange
376 // contortions to avoid real linker symbols. For instance, the
377 // ObjC class data structure is allocated statically in the executable
378 // that defines that class. That data structures contains a pointer to
379 // its superclass. But instead of just initializing that part of the
380 // struct to the address of its superclass, and letting the static and
381 // dynamic linkers do the rest, the runtime works by having that field
382 // instead point to a C-string that is the name of the superclass.
383 // At runtime the objc initialization updates that pointer and sets
384 // it to point to the actual super class. As far as the linker
385 // knows it is just a pointer to a string. But then someone wanted the
386 // linker to issue errors at build time if the superclass was not found.
387 // So they figured out a way in mach-o object format to use an absolute
388 // symbols (.objc_class_name_Foo = 0) and a floating reference
389 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
390 // a class was missing.
391 // The following synthesizes the implicit .objc_* symbols for the linker
392 // from the ObjC data structures generated by the front end.
394 // special case if this data blob is an ObjC class definition
395 std::string Section = v->getSection();
396 if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
397 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
402 // special case if this data blob is an ObjC category definition
403 else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
404 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
409 // special case if this data blob is the list of referenced classes
410 else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
411 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
417 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
418 SmallString<64> Buffer;
420 raw_svector_ostream OS(Buffer);
425 cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
426 addDefinedFunctionSymbol(Buffer.c_str(), F);
429 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
430 // add to list of defined symbols
431 addDefinedSymbol(Name, F, true);
434 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
436 // set alignment part log2() can have rounding errors
437 uint32_t align = def->getAlignment();
438 uint32_t attr = align ? countTrailingZeros(align) : 0;
440 // set permissions part
442 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
444 const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
445 if (gv && gv->isConstant())
446 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
448 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
451 // set definition part
452 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
453 attr |= LTO_SYMBOL_DEFINITION_WEAK;
454 else if (def->hasCommonLinkage())
455 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
457 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
460 if (def->hasLocalLinkage())
461 // Ignore visibility if linkage is local.
462 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
463 else if (def->hasHiddenVisibility())
464 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
465 else if (def->hasProtectedVisibility())
466 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
467 else if (canBeOmittedFromSymbolTable(def))
468 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
470 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
472 if (def->hasComdat())
473 attr |= LTO_SYMBOL_COMDAT;
475 auto Iter = _defines.insert(Name).first;
477 // fill information structure
478 NameAndAttributes info;
479 StringRef NameRef = Iter->first();
480 info.name = NameRef.data();
481 assert(info.name[NameRef.size()] == '\0');
482 info.attributes = attr;
483 info.isFunction = isFunction;
486 // add to table of symbols
487 _symbols.push_back(info);
490 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
492 void LTOModule::addAsmGlobalSymbol(const char *name,
493 lto_symbol_attributes scope) {
494 auto IterBool = _defines.insert(name);
496 // only add new define if not already defined
497 if (!IterBool.second)
500 NameAndAttributes &info = _undefines[IterBool.first->first().data()];
502 if (info.symbol == nullptr) {
503 // FIXME: This is trying to take care of module ASM like this:
505 // module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
507 // but is gross and its mother dresses it funny. Have the ASM parser give us
508 // more details for this type of situation so that we're not guessing so
511 // fill information structure
512 info.name = IterBool.first->first().data();
514 LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
515 info.isFunction = false;
516 info.symbol = nullptr;
518 // add to table of symbols
519 _symbols.push_back(info);
524 addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
526 addDefinedDataSymbol(info.name, info.symbol);
528 _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
529 _symbols.back().attributes |= scope;
532 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
534 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
535 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
537 _asm_undefines.push_back(IterBool.first->first().data());
539 // we already have the symbol
540 if (!IterBool.second)
543 uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
544 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
545 NameAndAttributes &info = IterBool.first->second;
546 info.name = IterBool.first->first().data();
547 info.attributes = attr;
548 info.isFunction = false;
549 info.symbol = nullptr;
552 /// Add a symbol which isn't defined just yet to a list to be resolved later.
553 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
555 SmallString<64> name;
557 raw_svector_ostream OS(name);
561 auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
563 // we already have the symbol
564 if (!IterBool.second)
567 NameAndAttributes &info = IterBool.first->second;
569 info.name = IterBool.first->first().data();
571 const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
573 if (decl->hasExternalWeakLinkage())
574 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
576 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
578 info.isFunction = isFunc;
582 /// parseSymbols - Parse the symbols from the module and model-level ASM and add
583 /// them to either the defined or undefined lists.
584 bool LTOModule::parseSymbols(std::string &errMsg) {
585 for (auto &Sym : IRFile->symbols()) {
586 const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
587 uint32_t Flags = Sym.getFlags();
588 if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
591 bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
594 SmallString<64> Buffer;
596 raw_svector_ostream OS(Buffer);
599 const char *Name = Buffer.c_str();
602 addAsmGlobalSymbolUndef(Name);
603 else if (Flags & object::BasicSymbolRef::SF_Global)
604 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
606 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
610 auto *F = dyn_cast<Function>(GV);
612 addPotentialUndefinedSymbol(Sym, F != nullptr);
617 addDefinedFunctionSymbol(Sym);
621 if (isa<GlobalVariable>(GV)) {
622 addDefinedDataSymbol(Sym);
626 assert(isa<GlobalAlias>(GV));
627 addDefinedDataSymbol(Sym);
630 // make symbols for all undefines
631 for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
632 e = _undefines.end(); u != e; ++u) {
633 // If this symbol also has a definition, then don't make an undefine because
634 // it is a tentative definition.
635 if (_defines.count(u->getKey())) continue;
636 NameAndAttributes info = u->getValue();
637 _symbols.push_back(info);
643 /// parseMetadata - Parse metadata from the module
644 void LTOModule::parseMetadata() {
646 if (Metadata *Val = getModule().getModuleFlag("Linker Options")) {
647 MDNode *LinkerOptions = cast<MDNode>(Val);
648 for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
649 MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
650 for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
651 MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
652 // FIXME: Make StringSet::insert match Self-Associative Container
653 // requirements, returning <iter,bool> rather than bool, and use that
656 _linkeropt_strings.insert(MDOption->getString()).first->first();
657 StringRef DepLibName =
658 _target->getObjFileLowering()->getDepLibFromLinkerOpt(Op);
659 if (!DepLibName.empty())
660 _deplibs.push_back(DepLibName.data());
661 else if (!Op.empty())
662 _linkeropts.push_back(Op.data());
667 // Add other interesting metadata here.