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, LLVMContext* Context,
92 OwningPtr<MemoryBuffer> buffer(MemoryBuffer::getFile(path, &errMsg));
95 return makeLTOModule(buffer.get(), Context, errMsg);
98 /// makeBuffer - create a MemoryBuffer from a memory range.
99 /// MemoryBuffer requires the byte past end of the buffer to be a zero.
100 /// We might get lucky and already be that way, otherwise make a copy.
101 /// Also if next byte is on a different page, don't assume it is readable.
102 MemoryBuffer* LTOModule::makeBuffer(const void* mem, size_t length)
104 const char* startPtr = (char*)mem;
105 const char* endPtr = startPtr+length;
106 if ( (((uintptr_t)endPtr & (sys::Process::GetPageSize()-1)) == 0)
108 return MemoryBuffer::getMemBufferCopy(startPtr, endPtr);
110 return MemoryBuffer::getMemBuffer(startPtr, endPtr);
114 LTOModule* LTOModule::makeLTOModule(const void* mem, size_t length,
115 LLVMContext* Context,
118 OwningPtr<MemoryBuffer> buffer(makeBuffer(mem, length));
121 return makeLTOModule(buffer.get(), Context, errMsg);
124 /// getFeatureString - Return a string listing the features associated with the
127 /// FIXME: This is an inelegant way of specifying the features of a
128 /// subtarget. It would be better if we could encode this information into the
129 /// IR. See <rdar://5972456>.
130 std::string getFeatureString(const char *TargetTriple) {
131 SubtargetFeatures Features;
133 if (strncmp(TargetTriple, "powerpc-apple-", 14) == 0) {
134 Features.AddFeature("altivec", true);
135 } else if (strncmp(TargetTriple, "powerpc64-apple-", 16) == 0) {
136 Features.AddFeature("64bit", true);
137 Features.AddFeature("altivec", true);
140 return Features.getString();
143 LTOModule* LTOModule::makeLTOModule(MemoryBuffer* buffer, LLVMContext* Context,
146 // parse bitcode buffer
147 OwningPtr<Module> m(ParseBitcodeFile(buffer, Context, &errMsg));
150 // find machine architecture for this module
151 const TargetMachineRegistry::entry* march =
152 TargetMachineRegistry::getClosestStaticTargetForModule(*m, errMsg);
157 // construct LTModule, hand over ownership of module and target
158 std::string FeatureStr = getFeatureString(m->getTargetTriple().c_str());
159 TargetMachine* target = march->CtorFn(*m, FeatureStr);
160 return new LTOModule(m.take(), target);
164 const char* LTOModule::getTargetTriple()
166 return _module->getTargetTriple().c_str();
169 void LTOModule::addDefinedFunctionSymbol(Function* f, Mangler &mangler)
171 // add to list of defined symbols
172 addDefinedSymbol(f, mangler, true);
174 // add external symbols referenced by this function.
175 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
176 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
177 for (unsigned count = 0, total = i->getNumOperands();
178 count != total; ++count) {
179 findExternalRefs(i->getOperand(count), mangler);
185 // get string that data pointer points to
186 bool LTOModule::objcClassNameFromExpression(Constant* c, std::string& name)
188 if (ConstantExpr* ce = dyn_cast<ConstantExpr>(c)) {
189 Constant* op = ce->getOperand(0);
190 if (GlobalVariable* gvn = dyn_cast<GlobalVariable>(op)) {
191 Constant* cn = gvn->getInitializer();
192 if (ConstantArray* ca = dyn_cast<ConstantArray>(cn)) {
193 if ( ca->isCString() ) {
194 name = ".objc_class_name_" + ca->getAsString();
203 // parse i386/ppc ObjC class data structure
204 void LTOModule::addObjCClass(GlobalVariable* clgv)
206 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
207 // second slot in __OBJC,__class is pointer to superclass name
208 std::string superclassName;
209 if ( objcClassNameFromExpression(c->getOperand(1), superclassName) ) {
210 NameAndAttributes info;
211 if ( _undefines.find(superclassName.c_str()) == _undefines.end() ) {
212 const char* symbolName = ::strdup(superclassName.c_str());
213 info.name = ::strdup(symbolName);
214 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
215 // string is owned by _undefines
216 _undefines[info.name] = info;
219 // third slot in __OBJC,__class is pointer to class name
220 std::string className;
221 if ( objcClassNameFromExpression(c->getOperand(2), className) ) {
222 const char* symbolName = ::strdup(className.c_str());
223 NameAndAttributes info;
224 info.name = symbolName;
225 info.attributes = (lto_symbol_attributes)
226 (LTO_SYMBOL_PERMISSIONS_DATA |
227 LTO_SYMBOL_DEFINITION_REGULAR |
228 LTO_SYMBOL_SCOPE_DEFAULT);
229 _symbols.push_back(info);
230 _defines[info.name] = 1;
236 // parse i386/ppc ObjC category data structure
237 void LTOModule::addObjCCategory(GlobalVariable* clgv)
239 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
240 // second slot in __OBJC,__category is pointer to target class name
241 std::string targetclassName;
242 if ( objcClassNameFromExpression(c->getOperand(1), targetclassName) ) {
243 NameAndAttributes info;
244 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ){
245 const char* symbolName = ::strdup(targetclassName.c_str());
246 info.name = ::strdup(symbolName);
247 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
248 // string is owned by _undefines
249 _undefines[info.name] = info;
256 // parse i386/ppc ObjC class list data structure
257 void LTOModule::addObjCClassRef(GlobalVariable* clgv)
259 std::string targetclassName;
260 if ( objcClassNameFromExpression(clgv->getInitializer(), targetclassName) ){
261 NameAndAttributes info;
262 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ) {
263 const char* symbolName = ::strdup(targetclassName.c_str());
264 info.name = ::strdup(symbolName);
265 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
266 // string is owned by _undefines
267 _undefines[info.name] = info;
273 void LTOModule::addDefinedDataSymbol(GlobalValue* v, Mangler& mangler)
275 // add to list of defined symbols
276 addDefinedSymbol(v, mangler, false);
278 // Special case i386/ppc ObjC data structures in magic sections:
279 // The issue is that the old ObjC object format did some strange
280 // contortions to avoid real linker symbols. For instance, the
281 // ObjC class data structure is allocated statically in the executable
282 // that defines that class. That data structures contains a pointer to
283 // its superclass. But instead of just initializing that part of the
284 // struct to the address of its superclass, and letting the static and
285 // dynamic linkers do the rest, the runtime works by having that field
286 // instead point to a C-string that is the name of the superclass.
287 // At runtime the objc initialization updates that pointer and sets
288 // it to point to the actual super class. As far as the linker
289 // knows it is just a pointer to a string. But then someone wanted the
290 // linker to issue errors at build time if the superclass was not found.
291 // So they figured out a way in mach-o object format to use an absolute
292 // symbols (.objc_class_name_Foo = 0) and a floating reference
293 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
294 // a class was missing.
295 // The following synthesizes the implicit .objc_* symbols for the linker
296 // from the ObjC data structures generated by the front end.
297 if ( v->hasSection() /* && isTargetDarwin */ ) {
298 // special case if this data blob is an ObjC class definition
299 if ( v->getSection().compare(0, 15, "__OBJC,__class,") == 0 ) {
300 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
305 // special case if this data blob is an ObjC category definition
306 else if ( v->getSection().compare(0, 18, "__OBJC,__category,") == 0 ) {
307 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
312 // special case if this data blob is the list of referenced classes
313 else if ( v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0 ) {
314 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
320 // add external symbols referenced by this data.
321 for (unsigned count = 0, total = v->getNumOperands();
322 count != total; ++count) {
323 findExternalRefs(v->getOperand(count), mangler);
328 void LTOModule::addDefinedSymbol(GlobalValue* def, Mangler &mangler,
331 // ignore all llvm.* symbols
332 if ( strncmp(def->getNameStart(), "llvm.", 5) == 0 )
335 // string is owned by _defines
336 const char* symbolName = ::strdup(mangler.getValueName(def).c_str());
338 // set alignment part log2() can have rounding errors
339 uint32_t align = def->getAlignment();
340 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
342 // set permissions part
344 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
346 GlobalVariable* gv = dyn_cast<GlobalVariable>(def);
347 if ( (gv != NULL) && gv->isConstant() )
348 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
350 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
353 // set definition part
354 if ( def->hasWeakLinkage() || def->hasLinkOnceLinkage() ) {
355 attr |= LTO_SYMBOL_DEFINITION_WEAK;
357 else if ( def->hasCommonLinkage()) {
358 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
361 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
365 if ( def->hasHiddenVisibility() )
366 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
367 else if ( def->hasProtectedVisibility() )
368 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
369 else if ( def->hasExternalLinkage() || def->hasWeakLinkage()
370 || def->hasLinkOnceLinkage() || def->hasCommonLinkage() )
371 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
373 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
375 // add to table of symbols
376 NameAndAttributes info;
377 info.name = symbolName;
378 info.attributes = (lto_symbol_attributes)attr;
379 _symbols.push_back(info);
380 _defines[info.name] = 1;
383 void LTOModule::addAsmGlobalSymbol(const char *name) {
384 // only add new define if not already defined
385 if ( _defines.count(name, &name[strlen(name)+1]) == 0 )
388 // string is owned by _defines
389 const char *symbolName = ::strdup(name);
390 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
391 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
392 NameAndAttributes info;
393 info.name = symbolName;
394 info.attributes = (lto_symbol_attributes)attr;
395 _symbols.push_back(info);
396 _defines[info.name] = 1;
399 void LTOModule::addPotentialUndefinedSymbol(GlobalValue* decl, Mangler &mangler)
401 // ignore all llvm.* symbols
402 if ( strncmp(decl->getNameStart(), "llvm.", 5) == 0 )
405 const char* name = mangler.getValueName(decl).c_str();
407 // we already have the symbol
408 if (_undefines.find(name) != _undefines.end())
411 NameAndAttributes info;
412 // string is owned by _undefines
413 info.name = ::strdup(name);
414 if (decl->hasExternalWeakLinkage())
415 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
417 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
418 _undefines[name] = info;
423 // Find exeternal symbols referenced by VALUE. This is a recursive function.
424 void LTOModule::findExternalRefs(Value* value, Mangler &mangler) {
426 if (GlobalValue* gv = dyn_cast<GlobalValue>(value)) {
427 if ( !gv->hasExternalLinkage() )
428 addPotentialUndefinedSymbol(gv, mangler);
429 // If this is a variable definition, do not recursively process
430 // initializer. It might contain a reference to this variable
431 // and cause an infinite loop. The initializer will be
432 // processed in addDefinedDataSymbol().
436 // GlobalValue, even with InternalLinkage type, may have operands with
437 // ExternalLinkage type. Do not ignore these operands.
438 if (Constant* c = dyn_cast<Constant>(value)) {
439 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
440 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
441 findExternalRefs(c->getOperand(i), mangler);
445 void LTOModule::lazyParseSymbols()
447 if ( !_symbolsParsed ) {
448 _symbolsParsed = true;
450 // Use mangler to add GlobalPrefix to names to match linker names.
451 Mangler mangler(*_module, _target->getTargetAsmInfo()->getGlobalPrefix());
452 // add chars used in ObjC method names so method names aren't mangled
453 mangler.markCharAcceptable('[');
454 mangler.markCharAcceptable(']');
455 mangler.markCharAcceptable('(');
456 mangler.markCharAcceptable(')');
457 mangler.markCharAcceptable('-');
458 mangler.markCharAcceptable('+');
459 mangler.markCharAcceptable(' ');
462 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
463 if ( f->isDeclaration() )
464 addPotentialUndefinedSymbol(f, mangler);
466 addDefinedFunctionSymbol(f, mangler);
470 for (Module::global_iterator v = _module->global_begin(),
471 e = _module->global_end(); v != e; ++v) {
472 if ( v->isDeclaration() )
473 addPotentialUndefinedSymbol(v, mangler);
475 addDefinedDataSymbol(v, mangler);
479 const std::string &inlineAsm = _module->getModuleInlineAsm();
480 const std::string glbl = ".globl";
481 std::string asmSymbolName;
482 std::string::size_type pos = inlineAsm.find(glbl, 0);
483 while (pos != std::string::npos) {
487 // skip white space between .globl and symbol name
488 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
489 if (pbegin == std::string::npos)
493 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
494 if (pend == std::string::npos)
497 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
498 addAsmGlobalSymbol(asmSymbolName.c_str());
500 // search next .globl
501 pos = inlineAsm.find(glbl, pend);
504 // make symbols for all undefines
505 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
506 it != _undefines.end(); ++it) {
507 // if this symbol also has a definition, then don't make an undefine
508 // because it is a tentative definition
509 if ( _defines.count(it->getKeyData(), it->getKeyData()+
510 it->getKeyLength()) == 0 ) {
511 NameAndAttributes info = it->getValue();
512 _symbols.push_back(info);
519 uint32_t LTOModule::getSymbolCount()
522 return _symbols.size();
526 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index)
529 if ( index < _symbols.size() )
530 return _symbols[index].attributes;
532 return lto_symbol_attributes(0);
535 const char* LTOModule::getSymbolName(uint32_t index)
538 if ( index < _symbols.size() )
539 return _symbols[index].name;