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,
92 OwningPtr<MemoryBuffer> buffer(MemoryBuffer::getFile(path, &errMsg));
95 return makeLTOModule(buffer.get(), 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,
117 OwningPtr<MemoryBuffer> buffer(makeBuffer(mem, length));
120 return makeLTOModule(buffer.get(), errMsg);
123 /// getFeatureString - Return a string listing the features associated with the
126 /// FIXME: This is an inelegant way of specifying the features of a
127 /// subtarget. It would be better if we could encode this information into the
128 /// IR. See <rdar://5972456>.
129 std::string getFeatureString(const char *TargetTriple) {
130 SubtargetFeatures Features;
132 if (strncmp(TargetTriple, "powerpc-apple-", 14) == 0) {
133 Features.AddFeature("altivec", true);
134 } else if (strncmp(TargetTriple, "powerpc64-apple-", 16) == 0) {
135 Features.AddFeature("64bit", true);
136 Features.AddFeature("altivec", true);
139 return Features.getString();
142 LTOModule* LTOModule::makeLTOModule(MemoryBuffer* buffer,
145 // parse bitcode buffer
146 OwningPtr<Module> m(ParseBitcodeFile(buffer, getGlobalContext(), &errMsg));
149 // find machine architecture for this module
150 const TargetMachineRegistry::entry* march =
151 TargetMachineRegistry::getClosestStaticTargetForModule(*m, errMsg);
156 // construct LTModule, hand over ownership of module and target
157 std::string FeatureStr = getFeatureString(m->getTargetTriple().c_str());
158 TargetMachine* target = march->CtorFn(*m, FeatureStr);
159 return new LTOModule(m.take(), target);
163 const char* LTOModule::getTargetTriple()
165 return _module->getTargetTriple().c_str();
168 void LTOModule::addDefinedFunctionSymbol(Function* f, Mangler &mangler)
170 // add to list of defined symbols
171 addDefinedSymbol(f, mangler, true);
173 // add external symbols referenced by this function.
174 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
175 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
176 for (unsigned count = 0, total = i->getNumOperands();
177 count != total; ++count) {
178 findExternalRefs(i->getOperand(count), mangler);
184 // get string that data pointer points to
185 bool LTOModule::objcClassNameFromExpression(Constant* c, std::string& name)
187 if (ConstantExpr* ce = dyn_cast<ConstantExpr>(c)) {
188 Constant* op = ce->getOperand(0);
189 if (GlobalVariable* gvn = dyn_cast<GlobalVariable>(op)) {
190 Constant* cn = gvn->getInitializer();
191 if (ConstantArray* ca = dyn_cast<ConstantArray>(cn)) {
192 if ( ca->isCString() ) {
193 name = ".objc_class_name_" + ca->getAsString();
202 // parse i386/ppc ObjC class data structure
203 void LTOModule::addObjCClass(GlobalVariable* clgv)
205 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
206 // second slot in __OBJC,__class is pointer to superclass name
207 std::string superclassName;
208 if ( objcClassNameFromExpression(c->getOperand(1), superclassName) ) {
209 NameAndAttributes info;
210 if ( _undefines.find(superclassName.c_str()) == _undefines.end() ) {
211 const char* symbolName = ::strdup(superclassName.c_str());
212 info.name = ::strdup(symbolName);
213 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
214 // string is owned by _undefines
215 _undefines[info.name] = info;
218 // third slot in __OBJC,__class is pointer to class name
219 std::string className;
220 if ( objcClassNameFromExpression(c->getOperand(2), className) ) {
221 const char* symbolName = ::strdup(className.c_str());
222 NameAndAttributes info;
223 info.name = symbolName;
224 info.attributes = (lto_symbol_attributes)
225 (LTO_SYMBOL_PERMISSIONS_DATA |
226 LTO_SYMBOL_DEFINITION_REGULAR |
227 LTO_SYMBOL_SCOPE_DEFAULT);
228 _symbols.push_back(info);
229 _defines[info.name] = 1;
235 // parse i386/ppc ObjC category data structure
236 void LTOModule::addObjCCategory(GlobalVariable* clgv)
238 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
239 // second slot in __OBJC,__category is pointer to target class name
240 std::string targetclassName;
241 if ( objcClassNameFromExpression(c->getOperand(1), targetclassName) ) {
242 NameAndAttributes info;
243 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ){
244 const char* symbolName = ::strdup(targetclassName.c_str());
245 info.name = ::strdup(symbolName);
246 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
247 // string is owned by _undefines
248 _undefines[info.name] = info;
255 // parse i386/ppc ObjC class list data structure
256 void LTOModule::addObjCClassRef(GlobalVariable* clgv)
258 std::string targetclassName;
259 if ( objcClassNameFromExpression(clgv->getInitializer(), targetclassName) ){
260 NameAndAttributes info;
261 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ) {
262 const char* symbolName = ::strdup(targetclassName.c_str());
263 info.name = ::strdup(symbolName);
264 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
265 // string is owned by _undefines
266 _undefines[info.name] = info;
272 void LTOModule::addDefinedDataSymbol(GlobalValue* v, Mangler& mangler)
274 // add to list of defined symbols
275 addDefinedSymbol(v, mangler, false);
277 // Special case i386/ppc ObjC data structures in magic sections:
278 // The issue is that the old ObjC object format did some strange
279 // contortions to avoid real linker symbols. For instance, the
280 // ObjC class data structure is allocated statically in the executable
281 // that defines that class. That data structures contains a pointer to
282 // its superclass. But instead of just initializing that part of the
283 // struct to the address of its superclass, and letting the static and
284 // dynamic linkers do the rest, the runtime works by having that field
285 // instead point to a C-string that is the name of the superclass.
286 // At runtime the objc initialization updates that pointer and sets
287 // it to point to the actual super class. As far as the linker
288 // knows it is just a pointer to a string. But then someone wanted the
289 // linker to issue errors at build time if the superclass was not found.
290 // So they figured out a way in mach-o object format to use an absolute
291 // symbols (.objc_class_name_Foo = 0) and a floating reference
292 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
293 // a class was missing.
294 // The following synthesizes the implicit .objc_* symbols for the linker
295 // from the ObjC data structures generated by the front end.
296 if ( v->hasSection() /* && isTargetDarwin */ ) {
297 // special case if this data blob is an ObjC class definition
298 if ( v->getSection().compare(0, 15, "__OBJC,__class,") == 0 ) {
299 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
304 // special case if this data blob is an ObjC category definition
305 else if ( v->getSection().compare(0, 18, "__OBJC,__category,") == 0 ) {
306 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
311 // special case if this data blob is the list of referenced classes
312 else if ( v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0 ) {
313 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
319 // add external symbols referenced by this data.
320 for (unsigned count = 0, total = v->getNumOperands();
321 count != total; ++count) {
322 findExternalRefs(v->getOperand(count), mangler);
327 void LTOModule::addDefinedSymbol(GlobalValue* def, Mangler &mangler,
330 // ignore all llvm.* symbols
331 if ( strncmp(def->getNameStart(), "llvm.", 5) == 0 )
334 // string is owned by _defines
335 const char* symbolName = ::strdup(mangler.getMangledName(def).c_str());
337 // set alignment part log2() can have rounding errors
338 uint32_t align = def->getAlignment();
339 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
341 // set permissions part
343 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
345 GlobalVariable* gv = dyn_cast<GlobalVariable>(def);
346 if ( (gv != NULL) && gv->isConstant() )
347 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
349 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
352 // set definition part
353 if ( def->hasWeakLinkage() || def->hasLinkOnceLinkage() ) {
354 attr |= LTO_SYMBOL_DEFINITION_WEAK;
356 else if ( def->hasCommonLinkage()) {
357 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
360 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
364 if ( def->hasHiddenVisibility() )
365 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
366 else if ( def->hasProtectedVisibility() )
367 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
368 else if ( def->hasExternalLinkage() || def->hasWeakLinkage()
369 || def->hasLinkOnceLinkage() || def->hasCommonLinkage() )
370 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
372 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
374 // add to table of symbols
375 NameAndAttributes info;
376 info.name = symbolName;
377 info.attributes = (lto_symbol_attributes)attr;
378 _symbols.push_back(info);
379 _defines[info.name] = 1;
382 void LTOModule::addAsmGlobalSymbol(const char *name) {
383 // only add new define if not already defined
384 if ( _defines.count(name, &name[strlen(name)+1]) == 0 )
387 // string is owned by _defines
388 const char *symbolName = ::strdup(name);
389 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
390 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
391 NameAndAttributes info;
392 info.name = symbolName;
393 info.attributes = (lto_symbol_attributes)attr;
394 _symbols.push_back(info);
395 _defines[info.name] = 1;
398 void LTOModule::addPotentialUndefinedSymbol(GlobalValue* decl, Mangler &mangler)
400 // ignore all llvm.* symbols
401 if ( strncmp(decl->getNameStart(), "llvm.", 5) == 0 )
404 // ignore all aliases
405 if (isa<GlobalAlias>(decl))
408 const char* name = mangler.getMangledName(decl).c_str();
410 // we already have the symbol
411 if (_undefines.find(name) != _undefines.end())
414 NameAndAttributes info;
415 // string is owned by _undefines
416 info.name = ::strdup(name);
417 if (decl->hasExternalWeakLinkage())
418 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
420 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
421 _undefines[name] = info;
426 // Find exeternal symbols referenced by VALUE. This is a recursive function.
427 void LTOModule::findExternalRefs(Value* value, Mangler &mangler) {
429 if (GlobalValue* gv = dyn_cast<GlobalValue>(value)) {
430 if ( !gv->hasExternalLinkage() )
431 addPotentialUndefinedSymbol(gv, mangler);
432 // If this is a variable definition, do not recursively process
433 // initializer. It might contain a reference to this variable
434 // and cause an infinite loop. The initializer will be
435 // processed in addDefinedDataSymbol().
439 // GlobalValue, even with InternalLinkage type, may have operands with
440 // ExternalLinkage type. Do not ignore these operands.
441 if (Constant* c = dyn_cast<Constant>(value)) {
442 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
443 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
444 findExternalRefs(c->getOperand(i), mangler);
448 void LTOModule::lazyParseSymbols()
450 if ( !_symbolsParsed ) {
451 _symbolsParsed = true;
453 // Use mangler to add GlobalPrefix to names to match linker names.
454 Mangler mangler(*_module, _target->getTargetAsmInfo()->getGlobalPrefix());
455 // add chars used in ObjC method names so method names aren't mangled
456 mangler.markCharAcceptable('[');
457 mangler.markCharAcceptable(']');
458 mangler.markCharAcceptable('(');
459 mangler.markCharAcceptable(')');
460 mangler.markCharAcceptable('-');
461 mangler.markCharAcceptable('+');
462 mangler.markCharAcceptable(' ');
465 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
466 if ( f->isDeclaration() )
467 addPotentialUndefinedSymbol(f, mangler);
469 addDefinedFunctionSymbol(f, mangler);
473 for (Module::global_iterator v = _module->global_begin(),
474 e = _module->global_end(); v != e; ++v) {
475 if ( v->isDeclaration() )
476 addPotentialUndefinedSymbol(v, mangler);
478 addDefinedDataSymbol(v, mangler);
482 const std::string &inlineAsm = _module->getModuleInlineAsm();
483 const std::string glbl = ".globl";
484 std::string asmSymbolName;
485 std::string::size_type pos = inlineAsm.find(glbl, 0);
486 while (pos != std::string::npos) {
490 // skip white space between .globl and symbol name
491 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
492 if (pbegin == std::string::npos)
496 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
497 if (pend == std::string::npos)
500 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
501 addAsmGlobalSymbol(asmSymbolName.c_str());
503 // search next .globl
504 pos = inlineAsm.find(glbl, pend);
507 // make symbols for all undefines
508 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
509 it != _undefines.end(); ++it) {
510 // if this symbol also has a definition, then don't make an undefine
511 // because it is a tentative definition
512 if ( _defines.count(it->getKeyData(), it->getKeyData()+
513 it->getKeyLength()) == 0 ) {
514 NameAndAttributes info = it->getValue();
515 _symbols.push_back(info);
522 uint32_t LTOModule::getSymbolCount()
525 return _symbols.size();
529 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index)
532 if ( index < _symbols.size() )
533 return _symbols[index].attributes;
535 return lto_symbol_attributes(0);
538 const char* LTOModule::getSymbolName(uint32_t index)
541 if ( index < _symbols.size() )
542 return _symbols[index].name;