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 "LTOModule.h"
17 #include "llvm/Constants.h"
18 #include "llvm/LLVMContext.h"
19 #include "llvm/Module.h"
20 #include "llvm/ModuleProvider.h"
21 #include "llvm/ADT/OwningPtr.h"
22 #include "llvm/Bitcode/ReaderWriter.h"
23 #include "llvm/Support/SystemUtils.h"
24 #include "llvm/Support/Mangler.h"
25 #include "llvm/Support/MemoryBuffer.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/System/Path.h"
28 #include "llvm/System/Process.h"
29 #include "llvm/Target/SubtargetFeature.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetMachineRegistry.h"
32 #include "llvm/Target/TargetAsmInfo.h"
38 bool LTOModule::isBitcodeFile(const void* mem, size_t length)
40 return ( llvm::sys::IdentifyFileType((char*)mem, length)
41 == llvm::sys::Bitcode_FileType );
44 bool LTOModule::isBitcodeFile(const char* path)
46 return llvm::sys::Path(path).isBitcodeFile();
49 bool LTOModule::isBitcodeFileForTarget(const void* mem, size_t length,
50 const char* triplePrefix)
52 MemoryBuffer* buffer = makeBuffer(mem, length);
55 return isTargetMatch(buffer, triplePrefix);
59 bool LTOModule::isBitcodeFileForTarget(const char* path,
60 const char* triplePrefix)
62 MemoryBuffer *buffer = MemoryBuffer::getFile(path);
65 return isTargetMatch(buffer, triplePrefix);
68 // takes ownership of buffer
69 bool LTOModule::isTargetMatch(MemoryBuffer* buffer, const char* triplePrefix)
71 OwningPtr<ModuleProvider> mp(getBitcodeModuleProvider(buffer,
73 // on success, mp owns buffer and both are deleted at end of this method
78 std::string actualTarget = mp->getModule()->getTargetTriple();
79 return ( strncmp(actualTarget.c_str(), triplePrefix,
80 strlen(triplePrefix)) == 0);
84 LTOModule::LTOModule(Module* m, TargetMachine* t)
85 : _module(m), _target(t), _symbolsParsed(false)
89 LTOModule* LTOModule::makeLTOModule(const char* path,
93 OwningPtr<MemoryBuffer> buffer(MemoryBuffer::getFile(path, &errMsg));
96 return makeLTOModule(buffer.get(), Context, errMsg);
99 /// makeBuffer - create a MemoryBuffer from a memory range.
100 /// MemoryBuffer requires the byte past end of the buffer to be a zero.
101 /// We might get lucky and already be that way, otherwise make a copy.
102 /// Also if next byte is on a different page, don't assume it is readable.
103 MemoryBuffer* LTOModule::makeBuffer(const void* mem, size_t length)
105 const char* startPtr = (char*)mem;
106 const char* endPtr = startPtr+length;
107 if ( (((uintptr_t)endPtr & (sys::Process::GetPageSize()-1)) == 0)
109 return MemoryBuffer::getMemBufferCopy(startPtr, endPtr);
111 return MemoryBuffer::getMemBuffer(startPtr, endPtr);
115 LTOModule* LTOModule::makeLTOModule(const void* mem, size_t length,
116 LLVMContext& Context,
119 OwningPtr<MemoryBuffer> buffer(makeBuffer(mem, length));
122 return makeLTOModule(buffer.get(), Context, errMsg);
125 /// getFeatureString - Return a string listing the features associated with the
128 /// FIXME: This is an inelegant way of specifying the features of a
129 /// subtarget. It would be better if we could encode this information into the
130 /// IR. See <rdar://5972456>.
131 std::string getFeatureString(const char *TargetTriple) {
132 SubtargetFeatures Features;
134 if (strncmp(TargetTriple, "powerpc-apple-", 14) == 0) {
135 Features.AddFeature("altivec", true);
136 } else if (strncmp(TargetTriple, "powerpc64-apple-", 16) == 0) {
137 Features.AddFeature("64bit", true);
138 Features.AddFeature("altivec", true);
141 return Features.getString();
144 LTOModule* LTOModule::makeLTOModule(MemoryBuffer* buffer,
145 LLVMContext& Context,
148 // parse bitcode buffer
149 OwningPtr<Module> m(ParseBitcodeFile(buffer, Context, &errMsg));
152 // find machine architecture for this module
153 const TargetMachineRegistry::entry* march =
154 TargetMachineRegistry::getClosestStaticTargetForModule(*m, errMsg);
159 // construct LTModule, hand over ownership of module and target
160 std::string FeatureStr = getFeatureString(m->getTargetTriple().c_str());
161 TargetMachine* target = march->CtorFn(*m, FeatureStr);
162 return new LTOModule(m.take(), target);
166 const char* LTOModule::getTargetTriple()
168 return _module->getTargetTriple().c_str();
171 void LTOModule::addDefinedFunctionSymbol(Function* f, Mangler &mangler)
173 // add to list of defined symbols
174 addDefinedSymbol(f, mangler, true);
176 // add external symbols referenced by this function.
177 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
178 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
179 for (unsigned count = 0, total = i->getNumOperands();
180 count != total; ++count) {
181 findExternalRefs(i->getOperand(count), mangler);
187 // get string that data pointer points to
188 bool LTOModule::objcClassNameFromExpression(Constant* c, std::string& name)
190 if (ConstantExpr* ce = dyn_cast<ConstantExpr>(c)) {
191 Constant* op = ce->getOperand(0);
192 if (GlobalVariable* gvn = dyn_cast<GlobalVariable>(op)) {
193 Constant* cn = gvn->getInitializer();
194 if (ConstantArray* ca = dyn_cast<ConstantArray>(cn)) {
195 if ( ca->isCString() ) {
196 name = ".objc_class_name_" + ca->getAsString();
205 // parse i386/ppc ObjC class data structure
206 void LTOModule::addObjCClass(GlobalVariable* clgv)
208 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
209 // second slot in __OBJC,__class is pointer to superclass name
210 std::string superclassName;
211 if ( objcClassNameFromExpression(c->getOperand(1), superclassName) ) {
212 NameAndAttributes info;
213 if ( _undefines.find(superclassName.c_str()) == _undefines.end() ) {
214 const char* symbolName = ::strdup(superclassName.c_str());
215 info.name = ::strdup(symbolName);
216 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
217 // string is owned by _undefines
218 _undefines[info.name] = info;
221 // third slot in __OBJC,__class is pointer to class name
222 std::string className;
223 if ( objcClassNameFromExpression(c->getOperand(2), className) ) {
224 const char* symbolName = ::strdup(className.c_str());
225 NameAndAttributes info;
226 info.name = symbolName;
227 info.attributes = (lto_symbol_attributes)
228 (LTO_SYMBOL_PERMISSIONS_DATA |
229 LTO_SYMBOL_DEFINITION_REGULAR |
230 LTO_SYMBOL_SCOPE_DEFAULT);
231 _symbols.push_back(info);
232 _defines[info.name] = 1;
238 // parse i386/ppc ObjC category data structure
239 void LTOModule::addObjCCategory(GlobalVariable* clgv)
241 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
242 // second slot in __OBJC,__category is pointer to target class name
243 std::string targetclassName;
244 if ( objcClassNameFromExpression(c->getOperand(1), targetclassName) ) {
245 NameAndAttributes info;
246 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ){
247 const char* symbolName = ::strdup(targetclassName.c_str());
248 info.name = ::strdup(symbolName);
249 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
250 // string is owned by _undefines
251 _undefines[info.name] = info;
258 // parse i386/ppc ObjC class list data structure
259 void LTOModule::addObjCClassRef(GlobalVariable* clgv)
261 std::string targetclassName;
262 if ( objcClassNameFromExpression(clgv->getInitializer(), targetclassName) ){
263 NameAndAttributes info;
264 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ) {
265 const char* symbolName = ::strdup(targetclassName.c_str());
266 info.name = ::strdup(symbolName);
267 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
268 // string is owned by _undefines
269 _undefines[info.name] = info;
275 void LTOModule::addDefinedDataSymbol(GlobalValue* v, Mangler& mangler)
277 // add to list of defined symbols
278 addDefinedSymbol(v, mangler, false);
280 // Special case i386/ppc ObjC data structures in magic sections:
281 // The issue is that the old ObjC object format did some strange
282 // contortions to avoid real linker symbols. For instance, the
283 // ObjC class data structure is allocated statically in the executable
284 // that defines that class. That data structures contains a pointer to
285 // its superclass. But instead of just initializing that part of the
286 // struct to the address of its superclass, and letting the static and
287 // dynamic linkers do the rest, the runtime works by having that field
288 // instead point to a C-string that is the name of the superclass.
289 // At runtime the objc initialization updates that pointer and sets
290 // it to point to the actual super class. As far as the linker
291 // knows it is just a pointer to a string. But then someone wanted the
292 // linker to issue errors at build time if the superclass was not found.
293 // So they figured out a way in mach-o object format to use an absolute
294 // symbols (.objc_class_name_Foo = 0) and a floating reference
295 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
296 // a class was missing.
297 // The following synthesizes the implicit .objc_* symbols for the linker
298 // from the ObjC data structures generated by the front end.
299 if ( v->hasSection() /* && isTargetDarwin */ ) {
300 // special case if this data blob is an ObjC class definition
301 if ( v->getSection().compare(0, 15, "__OBJC,__class,") == 0 ) {
302 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
307 // special case if this data blob is an ObjC category definition
308 else if ( v->getSection().compare(0, 18, "__OBJC,__category,") == 0 ) {
309 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
314 // special case if this data blob is the list of referenced classes
315 else if ( v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0 ) {
316 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
322 // add external symbols referenced by this data.
323 for (unsigned count = 0, total = v->getNumOperands();
324 count != total; ++count) {
325 findExternalRefs(v->getOperand(count), mangler);
330 void LTOModule::addDefinedSymbol(GlobalValue* def, Mangler &mangler,
333 // ignore all llvm.* symbols
334 if ( strncmp(def->getNameStart(), "llvm.", 5) == 0 )
337 // string is owned by _defines
338 const char* symbolName = ::strdup(mangler.getValueName(def).c_str());
340 // set alignment part log2() can have rounding errors
341 uint32_t align = def->getAlignment();
342 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
344 // set permissions part
346 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
348 GlobalVariable* gv = dyn_cast<GlobalVariable>(def);
349 if ( (gv != NULL) && gv->isConstant() )
350 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
352 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
355 // set definition part
356 if ( def->hasWeakLinkage() || def->hasLinkOnceLinkage() ) {
357 attr |= LTO_SYMBOL_DEFINITION_WEAK;
359 else if ( def->hasCommonLinkage()) {
360 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
363 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
367 if ( def->hasHiddenVisibility() )
368 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
369 else if ( def->hasProtectedVisibility() )
370 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
371 else if ( def->hasExternalLinkage() || def->hasWeakLinkage()
372 || def->hasLinkOnceLinkage() || def->hasCommonLinkage() )
373 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
375 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
377 // add to table of symbols
378 NameAndAttributes info;
379 info.name = symbolName;
380 info.attributes = (lto_symbol_attributes)attr;
381 _symbols.push_back(info);
382 _defines[info.name] = 1;
385 void LTOModule::addAsmGlobalSymbol(const char *name) {
386 // only add new define if not already defined
387 if ( _defines.count(name, &name[strlen(name)+1]) == 0 )
390 // string is owned by _defines
391 const char *symbolName = ::strdup(name);
392 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
393 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
394 NameAndAttributes info;
395 info.name = symbolName;
396 info.attributes = (lto_symbol_attributes)attr;
397 _symbols.push_back(info);
398 _defines[info.name] = 1;
401 void LTOModule::addPotentialUndefinedSymbol(GlobalValue* decl, Mangler &mangler)
403 // ignore all llvm.* symbols
404 if ( strncmp(decl->getNameStart(), "llvm.", 5) == 0 )
407 const char* name = mangler.getValueName(decl).c_str();
409 // we already have the symbol
410 if (_undefines.find(name) != _undefines.end())
413 NameAndAttributes info;
414 // string is owned by _undefines
415 info.name = ::strdup(name);
416 if (decl->hasExternalWeakLinkage())
417 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
419 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
420 _undefines[name] = info;
425 // Find exeternal symbols referenced by VALUE. This is a recursive function.
426 void LTOModule::findExternalRefs(Value* value, Mangler &mangler) {
428 if (GlobalValue* gv = dyn_cast<GlobalValue>(value)) {
429 if ( !gv->hasExternalLinkage() )
430 addPotentialUndefinedSymbol(gv, mangler);
431 // If this is a variable definition, do not recursively process
432 // initializer. It might contain a reference to this variable
433 // and cause an infinite loop. The initializer will be
434 // processed in addDefinedDataSymbol().
438 // GlobalValue, even with InternalLinkage type, may have operands with
439 // ExternalLinkage type. Do not ignore these operands.
440 if (Constant* c = dyn_cast<Constant>(value)) {
441 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
442 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
443 findExternalRefs(c->getOperand(i), mangler);
447 void LTOModule::lazyParseSymbols()
449 if ( !_symbolsParsed ) {
450 _symbolsParsed = true;
452 // Use mangler to add GlobalPrefix to names to match linker names.
453 Mangler mangler(*_module, _target->getTargetAsmInfo()->getGlobalPrefix());
454 // add chars used in ObjC method names so method names aren't mangled
455 mangler.markCharAcceptable('[');
456 mangler.markCharAcceptable(']');
457 mangler.markCharAcceptable('(');
458 mangler.markCharAcceptable(')');
459 mangler.markCharAcceptable('-');
460 mangler.markCharAcceptable('+');
461 mangler.markCharAcceptable(' ');
464 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
465 if ( f->isDeclaration() )
466 addPotentialUndefinedSymbol(f, mangler);
468 addDefinedFunctionSymbol(f, mangler);
472 for (Module::global_iterator v = _module->global_begin(),
473 e = _module->global_end(); v != e; ++v) {
474 if ( v->isDeclaration() )
475 addPotentialUndefinedSymbol(v, mangler);
477 addDefinedDataSymbol(v, mangler);
481 const std::string &inlineAsm = _module->getModuleInlineAsm();
482 const std::string glbl = ".globl";
483 std::string asmSymbolName;
484 std::string::size_type pos = inlineAsm.find(glbl, 0);
485 while (pos != std::string::npos) {
489 // skip white space between .globl and symbol name
490 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
491 if (pbegin == std::string::npos)
495 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
496 if (pend == std::string::npos)
499 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
500 addAsmGlobalSymbol(asmSymbolName.c_str());
502 // search next .globl
503 pos = inlineAsm.find(glbl, pend);
506 // make symbols for all undefines
507 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
508 it != _undefines.end(); ++it) {
509 // if this symbol also has a definition, then don't make an undefine
510 // because it is a tentative definition
511 if ( _defines.count(it->getKeyData(), it->getKeyData()+
512 it->getKeyLength()) == 0 ) {
513 NameAndAttributes info = it->getValue();
514 _symbols.push_back(info);
521 uint32_t LTOModule::getSymbolCount()
524 return _symbols.size();
528 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index)
531 if ( index < _symbols.size() )
532 return _symbols[index].attributes;
534 return lto_symbol_attributes(0);
537 const char* LTOModule::getSymbolName(uint32_t index)
540 if ( index < _symbols.size() )
541 return _symbols[index].name;