1 //===-- LTOModule.cpp - LLVM Link Time Optimizer --------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
12 //
13 //===----------------------------------------------------------------------===//
14 
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>
49 using namespace llvm;
50 using namespace llvm::object;
51 
LTOModule(std::unique_ptr<object::IRObjectFile> Obj,llvm::TargetMachine * TM)52 LTOModule::LTOModule(std::unique_ptr<object::IRObjectFile> Obj,
53                      llvm::TargetMachine *TM)
54     : IRFile(std::move(Obj)), _target(TM) {}
55 
LTOModule(std::unique_ptr<object::IRObjectFile> Obj,llvm::TargetMachine * TM,std::unique_ptr<LLVMContext> Context)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) {}
60 
~LTOModule()61 LTOModule::~LTOModule() {}
62 
63 /// isBitcodeFile - Returns 'true' if the file (or memory contents) is LLVM
64 /// bitcode.
isBitcodeFile(const void * Mem,size_t Length)65 bool LTOModule::isBitcodeFile(const void *Mem, size_t Length) {
66   ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
67       MemoryBufferRef(StringRef((const char *)Mem, Length), "<mem>"));
68   return bool(BCData);
69 }
70 
isBitcodeFile(const char * Path)71 bool LTOModule::isBitcodeFile(const char *Path) {
72   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
73       MemoryBuffer::getFile(Path);
74   if (!BufferOrErr)
75     return false;
76 
77   ErrorOr<MemoryBufferRef> BCData = IRObjectFile::findBitcodeInMemBuffer(
78       BufferOrErr.get()->getMemBufferRef());
79   return bool(BCData);
80 }
81 
isBitcodeForTarget(MemoryBuffer * Buffer,StringRef TriplePrefix)82 bool LTOModule::isBitcodeForTarget(MemoryBuffer *Buffer,
83                                    StringRef TriplePrefix) {
84   ErrorOr<MemoryBufferRef> BCOrErr =
85       IRObjectFile::findBitcodeInMemBuffer(Buffer->getMemBufferRef());
86   if (!BCOrErr)
87     return false;
88   LLVMContext Context;
89   std::string Triple = getBitcodeTargetTriple(*BCOrErr, Context);
90   return StringRef(Triple).startswith(TriplePrefix);
91 }
92 
createFromFile(const char * path,TargetOptions options,std::string & errMsg)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();
99     return nullptr;
100   }
101   std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
102   return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
103                        &getGlobalContext());
104 }
105 
createFromOpenFile(int fd,const char * path,size_t size,TargetOptions options,std::string & errMsg)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);
110 }
111 
createFromOpenFileSlice(int fd,const char * path,size_t map_size,off_t offset,TargetOptions options,std::string & 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();
120     return nullptr;
121   }
122   std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOrErr.get());
123   return makeLTOModule(Buffer->getMemBufferRef(), options, errMsg,
124                        &getGlobalContext());
125 }
126 
createFromBuffer(const void * mem,size_t length,TargetOptions options,std::string & errMsg,StringRef path)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());
132 }
133 
createInLocalContext(const void * mem,size_t length,TargetOptions options,std::string & errMsg,StringRef path)134 LTOModule *LTOModule::createInLocalContext(const void *mem, size_t length,
135                                            TargetOptions options,
136                                            std::string &errMsg,
137                                            StringRef path) {
138   return createInContext(mem, length, options, errMsg, path, nullptr);
139 }
140 
createInContext(const void * mem,size_t length,TargetOptions options,std::string & errMsg,StringRef path,LLVMContext * Context)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);
148 }
149 
parseBitcodeFileImpl(MemoryBufferRef Buffer,LLVMContext & Context,bool ShouldBeLazy,std::string & ErrMsg)150 static Module *parseBitcodeFileImpl(MemoryBufferRef Buffer,
151                                     LLVMContext &Context, bool ShouldBeLazy,
152                                     std::string &ErrMsg) {
153 
154   // Find the buffer.
155   ErrorOr<MemoryBufferRef> MBOrErr =
156       IRObjectFile::findBitcodeInMemBuffer(Buffer);
157   if (std::error_code EC = MBOrErr.getError()) {
158     ErrMsg = EC.message();
159     return nullptr;
160   }
161 
162   std::function<void(const DiagnosticInfo &)> DiagnosticHandler =
163       [&ErrMsg](const DiagnosticInfo &DI) {
164         raw_string_ostream Stream(ErrMsg);
165         DiagnosticPrinterRawOStream DP(Stream);
166         DI.print(DP);
167       };
168 
169   if (!ShouldBeLazy) {
170     // Parse the full file.
171     ErrorOr<Module *> M =
172         parseBitcodeFile(*MBOrErr, Context, DiagnosticHandler);
173     if (!M)
174       return nullptr;
175     return *M;
176   }
177 
178   // Parse lazily.
179   std::unique_ptr<MemoryBuffer> LightweightBuf =
180       MemoryBuffer::getMemBuffer(*MBOrErr, false);
181   ErrorOr<Module *> M = getLazyBitcodeModule(std::move(LightweightBuf), Context,
182                                              DiagnosticHandler,
183                                              true/*ShouldLazyLoadMetadata*/);
184   if (!M)
185     return nullptr;
186   return *M;
187 }
188 
makeLTOModule(MemoryBufferRef Buffer,TargetOptions options,std::string & errMsg,LLVMContext * Context)189 LTOModule *LTOModule::makeLTOModule(MemoryBufferRef Buffer,
190                                     TargetOptions options, std::string &errMsg,
191                                     LLVMContext *Context) {
192   std::unique_ptr<LLVMContext> OwnedContext;
193   if (!Context) {
194     OwnedContext = llvm::make_unique<LLVMContext>();
195     Context = OwnedContext.get();
196   }
197 
198   // If we own a context, we know this is being used only for symbol
199   // extraction, not linking.  Be lazy in that case.
200   std::unique_ptr<Module> M(parseBitcodeFileImpl(
201       Buffer, *Context,
202       /* ShouldBeLazy */ static_cast<bool>(OwnedContext), errMsg));
203   if (!M)
204     return nullptr;
205 
206   std::string TripleStr = M->getTargetTriple();
207   if (TripleStr.empty())
208     TripleStr = sys::getDefaultTargetTriple();
209   llvm::Triple Triple(TripleStr);
210 
211   // find machine architecture for this module
212   const Target *march = TargetRegistry::lookupTarget(TripleStr, errMsg);
213   if (!march)
214     return nullptr;
215 
216   // construct LTOModule, hand over ownership of module and target
217   SubtargetFeatures Features;
218   Features.getDefaultSubtargetFeatures(Triple);
219   std::string FeatureStr = Features.getString();
220   // Set a default CPU for Darwin triples.
221   std::string CPU;
222   if (Triple.isOSDarwin()) {
223     if (Triple.getArch() == llvm::Triple::x86_64)
224       CPU = "core2";
225     else if (Triple.getArch() == llvm::Triple::x86)
226       CPU = "yonah";
227     else if (Triple.getArch() == llvm::Triple::aarch64)
228       CPU = "cyclone";
229   }
230 
231   TargetMachine *target = march->createTargetMachine(TripleStr, CPU, FeatureStr,
232                                                      options);
233   M->setDataLayout(*target->getDataLayout());
234 
235   std::unique_ptr<object::IRObjectFile> IRObj(
236       new object::IRObjectFile(Buffer, std::move(M)));
237 
238   LTOModule *Ret;
239   if (OwnedContext)
240     Ret = new LTOModule(std::move(IRObj), target, std::move(OwnedContext));
241   else
242     Ret = new LTOModule(std::move(IRObj), target);
243 
244   if (Ret->parseSymbols(errMsg)) {
245     delete Ret;
246     return nullptr;
247   }
248 
249   Ret->parseMetadata();
250 
251   return Ret;
252 }
253 
254 /// Create a MemoryBuffer from a memory range with an optional name.
255 std::unique_ptr<MemoryBuffer>
makeBuffer(const void * mem,size_t length,StringRef name)256 LTOModule::makeBuffer(const void *mem, size_t length, StringRef name) {
257   const char *startPtr = (const char*)mem;
258   return MemoryBuffer::getMemBuffer(StringRef(startPtr, length), name, false);
259 }
260 
261 /// objcClassNameFromExpression - Get string that the data pointer points to.
262 bool
objcClassNameFromExpression(const Constant * c,std::string & name)263 LTOModule::objcClassNameFromExpression(const Constant *c, std::string &name) {
264   if (const ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
265     Constant *op = ce->getOperand(0);
266     if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
267       Constant *cn = gvn->getInitializer();
268       if (ConstantDataArray *ca = dyn_cast<ConstantDataArray>(cn)) {
269         if (ca->isCString()) {
270           name = (".objc_class_name_" + ca->getAsCString()).str();
271           return true;
272         }
273       }
274     }
275   }
276   return false;
277 }
278 
279 /// addObjCClass - Parse i386/ppc ObjC class data structure.
addObjCClass(const GlobalVariable * clgv)280 void LTOModule::addObjCClass(const GlobalVariable *clgv) {
281   const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
282   if (!c) return;
283 
284   // second slot in __OBJC,__class is pointer to superclass name
285   std::string superclassName;
286   if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
287     auto IterBool =
288         _undefines.insert(std::make_pair(superclassName, NameAndAttributes()));
289     if (IterBool.second) {
290       NameAndAttributes &info = IterBool.first->second;
291       info.name = IterBool.first->first().data();
292       info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
293       info.isFunction = false;
294       info.symbol = clgv;
295     }
296   }
297 
298   // third slot in __OBJC,__class is pointer to class name
299   std::string className;
300   if (objcClassNameFromExpression(c->getOperand(2), className)) {
301     auto Iter = _defines.insert(className).first;
302 
303     NameAndAttributes info;
304     info.name = Iter->first().data();
305     info.attributes = LTO_SYMBOL_PERMISSIONS_DATA |
306       LTO_SYMBOL_DEFINITION_REGULAR | LTO_SYMBOL_SCOPE_DEFAULT;
307     info.isFunction = false;
308     info.symbol = clgv;
309     _symbols.push_back(info);
310   }
311 }
312 
313 /// addObjCCategory - Parse i386/ppc ObjC category data structure.
addObjCCategory(const GlobalVariable * clgv)314 void LTOModule::addObjCCategory(const GlobalVariable *clgv) {
315   const ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer());
316   if (!c) return;
317 
318   // second slot in __OBJC,__category is pointer to target class name
319   std::string targetclassName;
320   if (!objcClassNameFromExpression(c->getOperand(1), targetclassName))
321     return;
322 
323   auto IterBool =
324       _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
325 
326   if (!IterBool.second)
327     return;
328 
329   NameAndAttributes &info = IterBool.first->second;
330   info.name = IterBool.first->first().data();
331   info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
332   info.isFunction = false;
333   info.symbol = clgv;
334 }
335 
336 /// addObjCClassRef - Parse i386/ppc ObjC class list data structure.
addObjCClassRef(const GlobalVariable * clgv)337 void LTOModule::addObjCClassRef(const GlobalVariable *clgv) {
338   std::string targetclassName;
339   if (!objcClassNameFromExpression(clgv->getInitializer(), targetclassName))
340     return;
341 
342   auto IterBool =
343       _undefines.insert(std::make_pair(targetclassName, NameAndAttributes()));
344 
345   if (!IterBool.second)
346     return;
347 
348   NameAndAttributes &info = IterBool.first->second;
349   info.name = IterBool.first->first().data();
350   info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
351   info.isFunction = false;
352   info.symbol = clgv;
353 }
354 
addDefinedDataSymbol(const object::BasicSymbolRef & Sym)355 void LTOModule::addDefinedDataSymbol(const object::BasicSymbolRef &Sym) {
356   SmallString<64> Buffer;
357   {
358     raw_svector_ostream OS(Buffer);
359     Sym.printName(OS);
360   }
361 
362   const GlobalValue *V = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
363   addDefinedDataSymbol(Buffer.c_str(), V);
364 }
365 
addDefinedDataSymbol(const char * Name,const GlobalValue * v)366 void LTOModule::addDefinedDataSymbol(const char *Name, const GlobalValue *v) {
367   // Add to list of defined symbols.
368   addDefinedSymbol(Name, v, false);
369 
370   if (!v->hasSection() /* || !isTargetDarwin */)
371     return;
372 
373   // Special case i386/ppc ObjC data structures in magic sections:
374   // The issue is that the old ObjC object format did some strange
375   // contortions to avoid real linker symbols.  For instance, the
376   // ObjC class data structure is allocated statically in the executable
377   // that defines that class.  That data structures contains a pointer to
378   // its superclass.  But instead of just initializing that part of the
379   // struct to the address of its superclass, and letting the static and
380   // dynamic linkers do the rest, the runtime works by having that field
381   // instead point to a C-string that is the name of the superclass.
382   // At runtime the objc initialization updates that pointer and sets
383   // it to point to the actual super class.  As far as the linker
384   // knows it is just a pointer to a string.  But then someone wanted the
385   // linker to issue errors at build time if the superclass was not found.
386   // So they figured out a way in mach-o object format to use an absolute
387   // symbols (.objc_class_name_Foo = 0) and a floating reference
388   // (.reference .objc_class_name_Bar) to cause the linker into erroring when
389   // a class was missing.
390   // The following synthesizes the implicit .objc_* symbols for the linker
391   // from the ObjC data structures generated by the front end.
392 
393   // special case if this data blob is an ObjC class definition
394   std::string Section = v->getSection();
395   if (Section.compare(0, 15, "__OBJC,__class,") == 0) {
396     if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
397       addObjCClass(gv);
398     }
399   }
400 
401   // special case if this data blob is an ObjC category definition
402   else if (Section.compare(0, 18, "__OBJC,__category,") == 0) {
403     if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
404       addObjCCategory(gv);
405     }
406   }
407 
408   // special case if this data blob is the list of referenced classes
409   else if (Section.compare(0, 18, "__OBJC,__cls_refs,") == 0) {
410     if (const GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
411       addObjCClassRef(gv);
412     }
413   }
414 }
415 
addDefinedFunctionSymbol(const object::BasicSymbolRef & Sym)416 void LTOModule::addDefinedFunctionSymbol(const object::BasicSymbolRef &Sym) {
417   SmallString<64> Buffer;
418   {
419     raw_svector_ostream OS(Buffer);
420     Sym.printName(OS);
421   }
422 
423   const Function *F =
424       cast<Function>(IRFile->getSymbolGV(Sym.getRawDataRefImpl()));
425   addDefinedFunctionSymbol(Buffer.c_str(), F);
426 }
427 
addDefinedFunctionSymbol(const char * Name,const Function * F)428 void LTOModule::addDefinedFunctionSymbol(const char *Name, const Function *F) {
429   // add to list of defined symbols
430   addDefinedSymbol(Name, F, true);
431 }
432 
addDefinedSymbol(const char * Name,const GlobalValue * def,bool isFunction)433 void LTOModule::addDefinedSymbol(const char *Name, const GlobalValue *def,
434                                  bool isFunction) {
435   // set alignment part log2() can have rounding errors
436   uint32_t align = def->getAlignment();
437   uint32_t attr = align ? countTrailingZeros(align) : 0;
438 
439   // set permissions part
440   if (isFunction) {
441     attr |= LTO_SYMBOL_PERMISSIONS_CODE;
442   } else {
443     const GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
444     if (gv && gv->isConstant())
445       attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
446     else
447       attr |= LTO_SYMBOL_PERMISSIONS_DATA;
448   }
449 
450   // set definition part
451   if (def->hasWeakLinkage() || def->hasLinkOnceLinkage())
452     attr |= LTO_SYMBOL_DEFINITION_WEAK;
453   else if (def->hasCommonLinkage())
454     attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
455   else
456     attr |= LTO_SYMBOL_DEFINITION_REGULAR;
457 
458   // set scope part
459   if (def->hasLocalLinkage())
460     // Ignore visibility if linkage is local.
461     attr |= LTO_SYMBOL_SCOPE_INTERNAL;
462   else if (def->hasHiddenVisibility())
463     attr |= LTO_SYMBOL_SCOPE_HIDDEN;
464   else if (def->hasProtectedVisibility())
465     attr |= LTO_SYMBOL_SCOPE_PROTECTED;
466   else if (canBeOmittedFromSymbolTable(def))
467     attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
468   else
469     attr |= LTO_SYMBOL_SCOPE_DEFAULT;
470 
471   auto Iter = _defines.insert(Name).first;
472 
473   // fill information structure
474   NameAndAttributes info;
475   StringRef NameRef = Iter->first();
476   info.name = NameRef.data();
477   assert(info.name[NameRef.size()] == '\0');
478   info.attributes = attr;
479   info.isFunction = isFunction;
480   info.symbol = def;
481 
482   // add to table of symbols
483   _symbols.push_back(info);
484 }
485 
486 /// addAsmGlobalSymbol - Add a global symbol from module-level ASM to the
487 /// defined list.
addAsmGlobalSymbol(const char * name,lto_symbol_attributes scope)488 void LTOModule::addAsmGlobalSymbol(const char *name,
489                                    lto_symbol_attributes scope) {
490   auto IterBool = _defines.insert(name);
491 
492   // only add new define if not already defined
493   if (!IterBool.second)
494     return;
495 
496   NameAndAttributes &info = _undefines[IterBool.first->first().data()];
497 
498   if (info.symbol == nullptr) {
499     // FIXME: This is trying to take care of module ASM like this:
500     //
501     //   module asm ".zerofill __FOO, __foo, _bar_baz_qux, 0"
502     //
503     // but is gross and its mother dresses it funny. Have the ASM parser give us
504     // more details for this type of situation so that we're not guessing so
505     // much.
506 
507     // fill information structure
508     info.name = IterBool.first->first().data();
509     info.attributes =
510       LTO_SYMBOL_PERMISSIONS_DATA | LTO_SYMBOL_DEFINITION_REGULAR | scope;
511     info.isFunction = false;
512     info.symbol = nullptr;
513 
514     // add to table of symbols
515     _symbols.push_back(info);
516     return;
517   }
518 
519   if (info.isFunction)
520     addDefinedFunctionSymbol(info.name, cast<Function>(info.symbol));
521   else
522     addDefinedDataSymbol(info.name, info.symbol);
523 
524   _symbols.back().attributes &= ~LTO_SYMBOL_SCOPE_MASK;
525   _symbols.back().attributes |= scope;
526 }
527 
528 /// addAsmGlobalSymbolUndef - Add a global symbol from module-level ASM to the
529 /// undefined list.
addAsmGlobalSymbolUndef(const char * name)530 void LTOModule::addAsmGlobalSymbolUndef(const char *name) {
531   auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
532 
533   _asm_undefines.push_back(IterBool.first->first().data());
534 
535   // we already have the symbol
536   if (!IterBool.second)
537     return;
538 
539   uint32_t attr = LTO_SYMBOL_DEFINITION_UNDEFINED;
540   attr |= LTO_SYMBOL_SCOPE_DEFAULT;
541   NameAndAttributes &info = IterBool.first->second;
542   info.name = IterBool.first->first().data();
543   info.attributes = attr;
544   info.isFunction = false;
545   info.symbol = nullptr;
546 }
547 
548 /// Add a symbol which isn't defined just yet to a list to be resolved later.
addPotentialUndefinedSymbol(const object::BasicSymbolRef & Sym,bool isFunc)549 void LTOModule::addPotentialUndefinedSymbol(const object::BasicSymbolRef &Sym,
550                                             bool isFunc) {
551   SmallString<64> name;
552   {
553     raw_svector_ostream OS(name);
554     Sym.printName(OS);
555   }
556 
557   auto IterBool = _undefines.insert(std::make_pair(name, NameAndAttributes()));
558 
559   // we already have the symbol
560   if (!IterBool.second)
561     return;
562 
563   NameAndAttributes &info = IterBool.first->second;
564 
565   info.name = IterBool.first->first().data();
566 
567   const GlobalValue *decl = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
568 
569   if (decl->hasExternalWeakLinkage())
570     info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
571   else
572     info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
573 
574   info.isFunction = isFunc;
575   info.symbol = decl;
576 }
577 
578 /// parseSymbols - Parse the symbols from the module and model-level ASM and add
579 /// them to either the defined or undefined lists.
parseSymbols(std::string & errMsg)580 bool LTOModule::parseSymbols(std::string &errMsg) {
581   for (auto &Sym : IRFile->symbols()) {
582     const GlobalValue *GV = IRFile->getSymbolGV(Sym.getRawDataRefImpl());
583     uint32_t Flags = Sym.getFlags();
584     if (Flags & object::BasicSymbolRef::SF_FormatSpecific)
585       continue;
586 
587     bool IsUndefined = Flags & object::BasicSymbolRef::SF_Undefined;
588 
589     if (!GV) {
590       SmallString<64> Buffer;
591       {
592         raw_svector_ostream OS(Buffer);
593         Sym.printName(OS);
594       }
595       const char *Name = Buffer.c_str();
596 
597       if (IsUndefined)
598         addAsmGlobalSymbolUndef(Name);
599       else if (Flags & object::BasicSymbolRef::SF_Global)
600         addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_DEFAULT);
601       else
602         addAsmGlobalSymbol(Name, LTO_SYMBOL_SCOPE_INTERNAL);
603       continue;
604     }
605 
606     auto *F = dyn_cast<Function>(GV);
607     if (IsUndefined) {
608       addPotentialUndefinedSymbol(Sym, F != nullptr);
609       continue;
610     }
611 
612     if (F) {
613       addDefinedFunctionSymbol(Sym);
614       continue;
615     }
616 
617     if (isa<GlobalVariable>(GV)) {
618       addDefinedDataSymbol(Sym);
619       continue;
620     }
621 
622     assert(isa<GlobalAlias>(GV));
623     addDefinedDataSymbol(Sym);
624   }
625 
626   // make symbols for all undefines
627   for (StringMap<NameAndAttributes>::iterator u =_undefines.begin(),
628          e = _undefines.end(); u != e; ++u) {
629     // If this symbol also has a definition, then don't make an undefine because
630     // it is a tentative definition.
631     if (_defines.count(u->getKey())) continue;
632     NameAndAttributes info = u->getValue();
633     _symbols.push_back(info);
634   }
635 
636   return false;
637 }
638 
639 /// parseMetadata - Parse metadata from the module
parseMetadata()640 void LTOModule::parseMetadata() {
641   // Linker Options
642   if (Metadata *Val = getModule().getModuleFlag("Linker Options")) {
643     MDNode *LinkerOptions = cast<MDNode>(Val);
644     for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
645       MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
646       for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
647         MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
648         // FIXME: Make StringSet::insert match Self-Associative Container
649         // requirements, returning <iter,bool> rather than bool, and use that
650         // here.
651         StringRef Op =
652             _linkeropt_strings.insert(MDOption->getString()).first->first();
653         StringRef DepLibName =
654             _target->getObjFileLowering()->getDepLibFromLinkerOpt(Op);
655         if (!DepLibName.empty())
656           _deplibs.push_back(DepLibName.data());
657         else if (!Op.empty())
658           _linkeropts.push_back(Op.data());
659       }
660     }
661   }
662 
663   // Add other interesting metadata here.
664 }
665