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/Support/CommandLine.h"
33 #include "llvm/Support/FileSystem.h"
34 #include "llvm/Support/Host.h"
35 #include "llvm/Support/MemoryBuffer.h"
36 #include "llvm/Support/Path.h"
37 #include "llvm/Support/SourceMgr.h"
38 #include "llvm/Support/TargetRegistry.h"
39 #include "llvm/Support/TargetSelect.h"
40 #include "llvm/Target/TargetLowering.h"
41 #include "llvm/Target/TargetLoweringObjectFile.h"
42 #include "llvm/Target/TargetRegisterInfo.h"
43 #include "llvm/Target/TargetSubtargetInfo.h"
44 #include "llvm/Transforms/Utils/GlobalStatus.h"
45 #include <system_error>
48 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
49 llvm::TargetMachine *TM)
50 : IRFile(std::move(Obj)), _target(TM) {}
52 /// isBitcodeFile - Returns 'true' if the file (or memory contents) is LLVM
54 bool LTOModule::isBitcodeFile(const void *mem, size_t length) {
55 return sys::fs::identify_magic(StringRef((const char *)mem, length)) ==
56 sys::fs::file_magic::bitcode;
59 bool LTOModule::isBitcodeFile(const char *path) {
60 sys::fs::file_magic type;
61 if (sys::fs::identify_magic(path, type))
63 return type == sys::fs::file_magic::bitcode;
66 bool LTOModule::isBitcodeForTarget(MemoryBuffer *buffer,
67 StringRef triplePrefix) {
68 std::string Triple = getBitcodeTargetTriple(buffer, getGlobalContext());
69 return StringRef(Triple).startswith(triplePrefix);
72 LTOModule *LTOModule::createFromFile(const char *path, TargetOptions options,
73 std::string &errMsg) {
74 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
75 MemoryBuffer::getFile(path);
76 if (std::error_code EC = BufferOrErr.getError()) {
77 errMsg = EC.message();
80 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
81 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg);
84 LTOModule *LTOModule::createFromOpenFile(int fd, const char *path, size_t size,
85 TargetOptions options,
86 std::string &errMsg) {
87 return createFromOpenFileSlice(fd, path, size, 0, options, errMsg);
90 LTOModule *LTOModule::createFromOpenFileSlice(int fd, const char *path,
91 size_t map_size, off_t offset,
92 TargetOptions options,
93 std::string &errMsg) {
94 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
95 MemoryBuffer::getOpenFileSlice(fd, path, map_size, offset);
96 if (std::error_code EC = BufferOrErr.getError()) {
97 errMsg = EC.message();
100 std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
101 return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg);
104 LTOModule *LTOModule::createFromBuffer(const void *mem, size_t length,
105 TargetOptions options,
106 std::string &errMsg, StringRef path) {
107 StringRef Data((const char *)mem, length);
108 MemoryBufferRef Buffer(Data, path);
109 return makeLTOModule(Buffer, options, errMsg);
112 LTOModule *LTOModule::makeLTOModule(MemoryBufferRef Buffer,
113 TargetOptions options,
114 std::string &errMsg) {
115 StringRef Data = Buffer.getBuffer();
116 StringRef FileName = Buffer.getBufferIdentifier();
117 std::unique_ptr<MemoryBuffer> MemBuf(
118 makeBuffer(Data.begin(), Data.size(), FileName));
122 ErrorOr<Module *> MOrErr = parseBitcodeFile(MemBuf.get(), getGlobalContext());
123 if (std::error_code EC = MOrErr.getError()) {
124 errMsg = EC.message();
127 std::unique_ptr<Module> M(MOrErr.get());
129 std::string TripleStr = M->getTargetTriple();
130 if (TripleStr.empty())
131 TripleStr = sys::getDefaultTargetTriple();
132 llvm::Triple Triple(TripleStr);
134 // find machine architecture for this module
135 const Target *march = TargetRegistry::lookupTarget(TripleStr, errMsg);
139 // construct LTOModule, hand over ownership of module and target
140 SubtargetFeatures Features;
141 Features.getDefaultSubtargetFeatures(Triple);
142 std::string FeatureStr = Features.getString();
143 // Set a default CPU for Darwin triples.
145 if (Triple.isOSDarwin()) {
146 if (Triple.getArch() == llvm::Triple::x86_64)
148 else if (Triple.getArch() == llvm::Triple::x86)
150 else if (Triple.getArch() == llvm::Triple::aarch64)
154 TargetMachine *target = march->createTargetMachine(TripleStr, CPU, FeatureStr,
156 M->setDataLayout(target->getSubtargetImpl()->getDataLayout());
158 std::unique_ptr<object::IRObjectFile> IRObj(
159 new object::IRObjectFile(Buffer, std::move(M)));
161 LTOModule *Ret = new LTOModule(std::move(IRObj), target);
163 if (Ret->parseSymbols(errMsg)) {
168 Ret->parseMetadata();
173 /// Create a MemoryBuffer from a memory range with an optional name.
174 std::unique_ptr<MemoryBuffer>
175 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
176 const char *startPtr = (const char*)mem;
177 return std::unique_ptr<MemoryBuffer>(
178 MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false));
181 /// objcClassNameFromExpression - Get string that the data pointer points to.
183 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
184 if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
185 Constant *op = ce->getOperand(0);
186 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
187 Constant *cn = gvn->getInitializer();
188 if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
189 if (ca->isCString()) {
190 name = ".objc_class_name_" + ca->getAsCString().str();
199 /// addObjCClass - Parse i386/ppc ObjC class data structure.
200 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
201 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
204 // second slot in __OBJC,__class is pointer to superclass name
205 std::string superclassName;
206 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
207 NameAndAttributes info;
208 StringMap<NameAndAttributes>::value_type &entry =
209 _undefines.GetOrCreateValue(superclassName);
210 if (!entry.getValue().name) {
211 const char *symbolName = entry.getKey().data();
212 info.name = symbolName;
213 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
214 info.isFunction = false;
216 entry.setValue(info);
220 // third slot in __OBJC,__class is pointer to class name
221 std::string className;
222 if (objcClassNameFromExpression(c->getOperand(2), className)) {
223 StringSet::value_type &entry = _defines.GetOrCreateValue(className);
226 NameAndAttributes info;
227 info.name = entry.getKey().data();
228 info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
229 LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
230 info.isFunction = false;
232 _symbols.push_back(info);
236 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
237 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
238 const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
241 // second slot in __OBJC,__category is pointer to target class name
242 std::string targetclassName;
243 if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
246 NameAndAttributes info;
247 StringMap<NameAndAttributes>::value_type &entry =
248 _undefines.GetOrCreateValue(targetclassName);
250 if (entry.getValue().name)
253 const char *symbolName = entry.getKey().data();
254 info.name = symbolName;
255 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
256 info.isFunction = false;
258 entry.setValue(info);
261 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
262 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
263 std::string targetclassName;
264 if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
267 NameAndAttributes info;
268 StringMap<NameAndAttributes>::value_type &entry =
269 _undefines.GetOrCreateValue(targetclassName);
270 if (entry.getValue().name)
273 const char *symbolName = entry.getKey().data();
274 info.name = symbolName;
275 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
276 info.isFunction = false;
278 entry.setValue(info);
281 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
282 SmallString<64> Buffer;
284 raw_svector_ostream OS(Buffer);
288 const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
289 addDefinedDataSymbol(Buffer.c_str(), V);
292 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
293 // Add to list of defined symbols.
294 addDefinedSymbol(Name, v, false);
296 if (!v->hasSection() /* || !isTargetDarwin */)
299 // Special case i386/ppc ObjC data structures in magic sections:
300 // The issue is that the old ObjC object format did some strange
301 // contortions to avoid real linker symbols. For instance, the
302 // ObjC class data structure is allocated statically in the executable
303 // that defines that class. That data structures contains a pointer to
304 // its superclass. But instead of just initializing that part of the
305 // struct to the address of its superclass, and letting the static and
306 // dynamic linkers do the rest, the runtime works by having that field
307 // instead point to a C-string that is the name of the superclass.
308 // At runtime the objc initialization updates that pointer and sets
309 // it to point to the actual super class. As far as the linker
310 // knows it is just a pointer to a string. But then someone wanted the
311 // linker to issue errors at build time if the superclass was not found.
312 // So they figured out a way in mach-o object format to use an absolute
313 // symbols (.objc_class_name_Foo = 0) and a floating reference
314 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
315 // a class was missing.
316 // The following synthesizes the implicit .objc_* symbols for the linker
317 // from the ObjC data structures generated by the front end.
319 // special case if this data blob is an ObjC class definition
320 std::string Section = v->getSection();
321 if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
322 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
327 // special case if this data blob is an ObjC category definition
328 else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
329 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
334 // special case if this data blob is the list of referenced classes
335 else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
336 if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
342 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
343 SmallString<64> Buffer;
345 raw_svector_ostream OS(Buffer);
350 cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
351 addDefinedFunctionSymbol(Buffer.c_str(), F);
354 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
355 // add to list of defined symbols
356 addDefinedSymbol(Name, F, true);
359 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
361 // set alignment part log2() can have rounding errors
362 uint32_t align = def->getAlignment();
363 uint32_t attr = align ? countTrailingZeros(align) : 0;
365 // set permissions part
367 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
369 const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
370 if (gv && gv->isConstant())
371 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
373 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
376 // set definition part
377 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
378 attr |= LTO_SYMBOL_DEFINITION_WEAK;
379 else if (def->hasCommonLinkage())
380 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
382 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
385 if (def->hasLocalLinkage())
386 // Ignore visibility if linkage is local.
387 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
388 else if (def->hasHiddenVisibility())
389 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
390 else if (def->hasProtectedVisibility())
391 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
392 else if (canBeOmittedFromSymbolTable(def))
393 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
395 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
397 StringSet::value_type &entry = _defines.GetOrCreateValue(Name);
400 // fill information structure
401 NameAndAttributes info;
402 StringRef NameRef = entry.getKey();
403 info.name = NameRef.data();
404 assert(info.name[NameRef.size()] == '\0');
405 info.attributes = attr;
406 info.isFunction = isFunction;
409 // add to table of symbols
410 _symbols.push_back(info);
413 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
415 void LTOModule::addAsmGlobalSymbol(const char *name,
416 lto_symbol_attributes scope) {
417 StringSet::value_type &entry = _defines.GetOrCreateValue(name);
419 // only add new define if not already defined
420 if (entry.getValue())
425 NameAndAttributes &info = _undefines[entry.getKey().data()];
427 if (info.symbol == nullptr) {
428 // FIXME: This is trying to take care of module ASM like this:
430 // module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
432 // but is gross and its mother dresses it funny. Have the ASM parser give us
433 // more details for this type of situation so that we're not guessing so
436 // fill information structure
437 info.name = entry.getKey().data();
439 LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
440 info.isFunction = false;
441 info.symbol = nullptr;
443 // add to table of symbols
444 _symbols.push_back(info);
449 addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
451 addDefinedDataSymbol(info.name, info.symbol);
453 _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
454 _symbols.back().attributes |= scope;
457 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
459 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
460 StringMap<NameAndAttributes>::value_type &entry =
461 _undefines.GetOrCreateValue(name);
463 _asm_undefines.push_back(entry.getKey().data());
465 // we already have the symbol
466 if (entry.getValue().name)
469 uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
470 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
471 NameAndAttributes info;
472 info.name = entry.getKey().data();
473 info.attributes = attr;
474 info.isFunction = false;
475 info.symbol = nullptr;
477 entry.setValue(info);
480 /// Add a symbol which isn't defined just yet to a list to be resolved later.
481 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
483 SmallString<64> name;
485 raw_svector_ostream OS(name);
489 StringMap<NameAndAttributes>::value_type &entry =
490 _undefines.GetOrCreateValue(name);
492 // we already have the symbol
493 if (entry.getValue().name)
496 NameAndAttributes info;
498 info.name = entry.getKey().data();
500 const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
502 if (decl->hasExternalWeakLinkage())
503 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
505 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
507 info.isFunction = isFunc;
510 entry.setValue(info);
513 /// parseSymbols - Parse the symbols from the module and model-level ASM and add
514 /// them to either the defined or undefined lists.
515 bool LTOModule::parseSymbols(std::string &errMsg) {
516 for (auto &Sym : IRFile->symbols()) {
517 const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
518 uint32_t Flags = Sym.getFlags();
519 if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
522 bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
525 SmallString<64> Buffer;
527 raw_svector_ostream OS(Buffer);
530 const char *Name = Buffer.c_str();
533 addAsmGlobalSymbolUndef(Name);
534 else if (Flags & object::BasicSymbolRef::SF_Global)
535 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
537 addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
541 auto *F = dyn_cast<Function>(GV);
543 addPotentialUndefinedSymbol(Sym, F != nullptr);
548 addDefinedFunctionSymbol(Sym);
552 if (isa<GlobalVariable>(GV)) {
553 addDefinedDataSymbol(Sym);
557 assert(isa<GlobalAlias>(GV));
558 addDefinedDataSymbol(Sym);
561 // make symbols for all undefines
562 for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
563 e = _undefines.end(); u != e; ++u) {
564 // If this symbol also has a definition, then don't make an undefine because
565 // it is a tentative definition.
566 if (_defines.count(u->getKey())) continue;
567 NameAndAttributes info = u->getValue();
568 _symbols.push_back(info);
574 /// parseMetadata - Parse metadata from the module
575 void LTOModule::parseMetadata() {
577 if (Value *Val = getModule().getModuleFlag("Linker Options")) {
578 MDNode *LinkerOptions = cast<MDNode>(Val);
579 for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
580 MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
581 for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
582 MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
583 StringRef Op = _linkeropt_strings.
584 GetOrCreateValue(MDOption->getString()).getKey();
585 StringRef DepLibName = _target->getSubtargetImpl()
586 ->getTargetLowering()
587 ->getObjFileLowering()
588 .getDepLibFromLinkerOpt(Op);
589 if (!DepLibName.empty())
590 _deplibs.push_back(DepLibName.data());
591 else if (!Op.empty())
592 _linkeropts.push_back(Op.data());
597 // Add other interesting metadata here.