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/LLVMContext.h"
21 #include "llvm/IR/Metadata.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/MC/MCExpr.h"
24 #include "llvm/MC/MCInst.h"
25 #include "llvm/MC/MCInstrInfo.h"
26 #include "llvm/MC/MCParser/MCAsmParser.h"
27 #include "llvm/MC/MCSection.h"
28 #include "llvm/MC/MCSubtargetInfo.h"
29 #include "llvm/MC/MCSymbol.h"
30 #include "llvm/MC/MCTargetAsmParser.h"
31 #include "llvm/MC/SubtargetFeature.h"
32 #include "llvm/Object/IRObjectFile.h"
33 #include "llvm/Object/ObjectFile.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/FileSystem.h"
36 #include "llvm/Support/Host.h"
37 #include "llvm/Support/MemoryBuffer.h"
38 #include "llvm/Support/Path.h"
39 #include "llvm/Support/SourceMgr.h"
40 #include "llvm/Support/TargetRegistry.h"
41 #include "llvm/Support/TargetSelect.h"
42 #include "llvm/Target/TargetLowering.h"
43 #include "llvm/Target/TargetLoweringObjectFile.h"
44 #include "llvm/Target/TargetRegisterInfo.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
46 #include "llvm/Transforms/Utils/GlobalStatus.h"
47 #include <system_error>
49 using namespace llvm::object;
51 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
52 llvm::TargetMachine *TM)
53 : IRFile(std::move(Obj)), _target(TM) {}
55 /// isBitcodeFile - Returns 'true' if the file (or memory contents) is LLVM
57 bool LTOModule::isBitcodeFile(const void *Mem, size_t Length) {
58 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
59 MemoryBufferRef(StringRef((const char *)Mem, Length), "<mem>"));
63 bool LTOModule::isBitcodeFile(const char *Path) {
64 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
65 MemoryBuffer::getFile(Path);
69 ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
70 BufferOrErr.get()->getMemBufferRef());
74 bool LTOModule::isBitcodeForTarget(MemoryBuffer *Buffer,
75 StringRef TriplePrefix) {
76 ErrorOr<MemoryBufferRef> BCOrErr =
77 IRObjectFile::findBitcodeInMemBuffer(Buffer->getMemBufferRef());
80 std::string Triple = getBitcodeTargetTriple(*BCOrErr, getGlobalContext());
81 return StringRef(Triple).startswith(TriplePrefix);
84 LTOModule *LTOModule::createFromFile(const char *path, TargetOptions options,
85 std::string &errMsg) {
86 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
87 MemoryBuffer::getFile(path);
88 if (std::error_code EC = BufferOrErr.getError()) {
89 errMsg = EC.message();
92 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
93 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg);
96 LTOModule *LTOModule::createFromOpenFile(int fd, const char *path, size_t size,
97 TargetOptions options,
98 std::string &errMsg) {
99 return createFromOpenFileSlice(fd, path, size, 0, options, errMsg);
102 LTOModule *LTOModule::createFromOpenFileSlice(int fd, const char *path,
103 size_t map_size, off_t offset,
104 TargetOptions options,
105 std::string &errMsg) {
106 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
107 MemoryBuffer::getOpenFileSlice(fd, path, map_size, offset);
108 if (std::error_code EC = BufferOrErr.getError()) {
109 errMsg = EC.message();
112 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
113 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg);
116 LTOModule *LTOModule::createFromBuffer(const void *mem, size_t length,
117 TargetOptions options,
118 std::string &errMsg, StringRef path) {
119 StringRef Data((const char *)mem, length);
120 MemoryBufferRef Buffer(Data, path);
121 return makeLTOModule(Buffer, options, errMsg);
124 LTOModule *LTOModule::makeLTOModule(MemoryBufferRef Buffer,
125 TargetOptions options,
126 std::string &errMsg) {
127 ErrorOr<MemoryBufferRef> MBOrErr =
128 IRObjectFile::findBitcodeInMemBuffer(Buffer);
129 if (std::error_code EC = MBOrErr.getError()) {
130 errMsg = EC.message();
133 ErrorOr<Module *> MOrErr = parseBitcodeFile(*MBOrErr, getGlobalContext());
134 if (std::error_code EC = MOrErr.getError()) {
135 errMsg = EC.message();
138 std::unique_ptr<Module> M(MOrErr.get());
140 std::string TripleStr = M->getTargetTriple();
141 if (TripleStr.empty())
142 TripleStr = sys::getDefaultTargetTriple();
143 llvm::Triple Triple(TripleStr);
145 // find machine architecture for this module
146 const Target *march = TargetRegistry::lookupTarget(TripleStr, errMsg);
150 // construct LTOModule, hand over ownership of module and target
151 SubtargetFeatures Features;
152 Features.getDefaultSubtargetFeatures(Triple);
153 std::string FeatureStr = Features.getString();
154 // Set a default CPU for Darwin triples.
156 if (Triple.isOSDarwin()) {
157 if (Triple.getArch() == llvm::Triple::x86_64)
159 else if (Triple.getArch() == llvm::Triple::x86)
161 else if (Triple.getArch() == llvm::Triple::aarch64)
165 TargetMachine *target = march->createTargetMachine(TripleStr, CPU, FeatureStr,
167 M->setDataLayout(target->getSubtargetImpl()->getDataLayout());
169 std::unique_ptr<object::IRObjectFile> IRObj(
170 new object::IRObjectFile(Buffer, std::move(M)));
172 LTOModule *Ret = new LTOModule(std::move(IRObj), target);
174 if (Ret->parseSymbols(errMsg)) {
179 Ret->parseMetadata();
184 /// Create a MemoryBuffer from a memory range with an optional name.
185 std::unique_ptr<MemoryBuffer>
186 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
187 const char *startPtr = (const char*)mem;
188 return MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false);
191 /// objcClassNameFromExpression - Get string that the data pointer points to.
193 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
194 if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
195 Constant *op = ce->getOperand(0);
196 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
197 Constant *cn = gvn->getInitializer();
198 if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
199 if (ca->isCString()) {
200 name = ".objc_class_name_" + ca->getAsCString().str();
209 /// addObjCClass - Parse i386/ppc ObjC class data structure.
210 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
211 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
214 // second slot in __OBJC,__class is pointer to superclass name
215 std::string superclassName;
216 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
217 NameAndAttributes info;
218 StringMap<NameAndAttributes>::value_type &entry =
219 _undefines.GetOrCreateValue(superclassName);
220 if (!entry.getValue().name) {
221 const char *symbolName = entry.getKey().data();
222 info.name = symbolName;
223 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
224 info.isFunction = false;
226 entry.setValue(info);
230 // third slot in __OBJC,__class is pointer to class name
231 std::string className;
232 if (objcClassNameFromExpression(c->getOperand(2), className)) {
233 StringSet::value_type &entry = _defines.GetOrCreateValue(className);
236 NameAndAttributes info;
237 info.name = entry.getKey().data();
238 info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
239 LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
240 info.isFunction = false;
242 _symbols.push_back(info);
246 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
247 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
248 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
251 // second slot in __OBJC,__category is pointer to target class name
252 std::string targetclassName;
253 if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
256 NameAndAttributes info;
257 StringMap<NameAndAttributes>::value_type &entry =
258 _undefines.GetOrCreateValue(targetclassName);
260 if (entry.getValue().name)
263 const char *symbolName = entry.getKey().data();
264 info.name = symbolName;
265 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
266 info.isFunction = false;
268 entry.setValue(info);
271 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
272 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
273 std::string targetclassName;
274 if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
277 NameAndAttributes info;
278 StringMap<NameAndAttributes>::value_type &entry =
279 _undefines.GetOrCreateValue(targetclassName);
280 if (entry.getValue().name)
283 const char *symbolName = entry.getKey().data();
284 info.name = symbolName;
285 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
286 info.isFunction = false;
288 entry.setValue(info);
291 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
292 SmallString<64> Buffer;
294 raw_svector_ostream OS(Buffer);
298 const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
299 addDefinedDataSymbol(Buffer.c_str(), V);
302 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
303 // Add to list of defined symbols.
304 addDefinedSymbol(Name, v, false);
306 if (!v->hasSection() /* || !isTargetDarwin */)
309 // Special case i386/ppc ObjC data structures in magic sections:
310 // The issue is that the old ObjC object format did some strange
311 // contortions to avoid real linker symbols. For instance, the
312 // ObjC class data structure is allocated statically in the executable
313 // that defines that class. That data structures contains a pointer to
314 // its superclass. But instead of just initializing that part of the
315 // struct to the address of its superclass, and letting the static and
316 // dynamic linkers do the rest, the runtime works by having that field
317 // instead point to a C-string that is the name of the superclass.
318 // At runtime the objc initialization updates that pointer and sets
319 // it to point to the actual super class. As far as the linker
320 // knows it is just a pointer to a string. But then someone wanted the
321 // linker to issue errors at build time if the superclass was not found.
322 // So they figured out a way in mach-o object format to use an absolute
323 // symbols (.objc_class_name_Foo = 0) and a floating reference
324 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
325 // a class was missing.
326 // The following synthesizes the implicit .objc_* symbols for the linker
327 // from the ObjC data structures generated by the front end.
329 // special case if this data blob is an ObjC class definition
330 std::string Section = v->getSection();
331 if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
332 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
337 // special case if this data blob is an ObjC category definition
338 else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
339 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
344 // special case if this data blob is the list of referenced classes
345 else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
346 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
352 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
353 SmallString<64> Buffer;
355 raw_svector_ostream OS(Buffer);
360 cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
361 addDefinedFunctionSymbol(Buffer.c_str(), F);
364 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
365 // add to list of defined symbols
366 addDefinedSymbol(Name, F, true);
369 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
371 // set alignment part log2() can have rounding errors
372 uint32_t align = def->getAlignment();
373 uint32_t attr = align ? countTrailingZeros(align) : 0;
375 // set permissions part
377 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
379 const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
380 if (gv && gv->isConstant())
381 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
383 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
386 // set definition part
387 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
388 attr |= LTO_SYMBOL_DEFINITION_WEAK;
389 else if (def->hasCommonLinkage())
390 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
392 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
395 if (def->hasLocalLinkage())
396 // Ignore visibility if linkage is local.
397 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
398 else if (def->hasHiddenVisibility())
399 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
400 else if (def->hasProtectedVisibility())
401 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
402 else if (canBeOmittedFromSymbolTable(def))
403 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
405 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
407 StringSet::value_type &entry = _defines.GetOrCreateValue(Name);
410 // fill information structure
411 NameAndAttributes info;
412 StringRef NameRef = entry.getKey();
413 info.name = NameRef.data();
414 assert(info.name[NameRef.size()] == '\0');
415 info.attributes = attr;
416 info.isFunction = isFunction;
419 // add to table of symbols
420 _symbols.push_back(info);
423 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
425 void LTOModule::addAsmGlobalSymbol(const char *name,
426 lto_symbol_attributes scope) {
427 StringSet::value_type &entry = _defines.GetOrCreateValue(name);
429 // only add new define if not already defined
430 if (entry.getValue())
435 NameAndAttributes &info = _undefines[entry.getKey().data()];
437 if (info.symbol == nullptr) {
438 // FIXME: This is trying to take care of module ASM like this:
440 // module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
442 // but is gross and its mother dresses it funny. Have the ASM parser give us
443 // more details for this type of situation so that we're not guessing so
446 // fill information structure
447 info.name = entry.getKey().data();
449 LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
450 info.isFunction = false;
451 info.symbol = nullptr;
453 // add to table of symbols
454 _symbols.push_back(info);
459 addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
461 addDefinedDataSymbol(info.name, info.symbol);
463 _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
464 _symbols.back().attributes |= scope;
467 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
469 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
470 StringMap<NameAndAttributes>::value_type &entry =
471 _undefines.GetOrCreateValue(name);
473 _asm_undefines.push_back(entry.getKey().data());
475 // we already have the symbol
476 if (entry.getValue().name)
479 uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
480 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
481 NameAndAttributes info;
482 info.name = entry.getKey().data();
483 info.attributes = attr;
484 info.isFunction = false;
485 info.symbol = nullptr;
487 entry.setValue(info);
490 /// Add a symbol which isn't defined just yet to a list to be resolved later.
491 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
493 SmallString<64> name;
495 raw_svector_ostream OS(name);
499 StringMap<NameAndAttributes>::value_type &entry =
500 _undefines.GetOrCreateValue(name);
502 // we already have the symbol
503 if (entry.getValue().name)
506 NameAndAttributes info;
508 info.name = entry.getKey().data();
510 const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
512 if (decl->hasExternalWeakLinkage())
513 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
515 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
517 info.isFunction = isFunc;
520 entry.setValue(info);
523 /// parseSymbols - Parse the symbols from the module and model-level ASM and add
524 /// them to either the defined or undefined lists.
525 bool LTOModule::parseSymbols(std::string &errMsg) {
526 for (auto &Sym : IRFile->symbols()) {
527 const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
528 uint32_t Flags = Sym.getFlags();
529 if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
532 bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
535 SmallString<64> Buffer;
537 raw_svector_ostream OS(Buffer);
540 const char *Name = Buffer.c_str();
543 addAsmGlobalSymbolUndef(Name);
544 else if (Flags & object::BasicSymbolRef::SF_Global)
545 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
547 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
551 auto *F = dyn_cast<Function>(GV);
553 addPotentialUndefinedSymbol(Sym, F != nullptr);
558 addDefinedFunctionSymbol(Sym);
562 if (isa<GlobalVariable>(GV)) {
563 addDefinedDataSymbol(Sym);
567 assert(isa<GlobalAlias>(GV));
568 addDefinedDataSymbol(Sym);
571 // make symbols for all undefines
572 for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
573 e = _undefines.end(); u != e; ++u) {
574 // If this symbol also has a definition, then don't make an undefine because
575 // it is a tentative definition.
576 if (_defines.count(u->getKey())) continue;
577 NameAndAttributes info = u->getValue();
578 _symbols.push_back(info);
584 /// parseMetadata - Parse metadata from the module
585 void LTOModule::parseMetadata() {
587 if (Value *Val = getModule().getModuleFlag("Linker Options")) {
588 MDNode *LinkerOptions = cast<MDNode>(Val);
589 for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
590 MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
591 for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
592 MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
593 StringRef Op = _linkeropt_strings.
594 GetOrCreateValue(MDOption->getString()).getKey();
595 StringRef DepLibName = _target->getSubtargetImpl()
596 ->getTargetLowering()
597 ->getObjFileLowering()
598 .getDepLibFromLinkerOpt(Op);
599 if (!DepLibName.empty())
600 _deplibs.push_back(DepLibName.data());
601 else if (!Op.empty())
602 _linkeropts.push_back(Op.data());
607 // Add other interesting metadata here.