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/TargetAsmInfo.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegistry.h"
36 bool LTOModule::isBitcodeFile(const void* mem, size_t length)
38 return ( llvm::sys::IdentifyFileType((char*)mem, length)
39 == llvm::sys::Bitcode_FileType );
42 bool LTOModule::isBitcodeFile(const char* path)
44 return llvm::sys::Path(path).isBitcodeFile();
47 bool LTOModule::isBitcodeFileForTarget(const void* mem, size_t length,
48 const char* triplePrefix)
50 MemoryBuffer* buffer = makeBuffer(mem, length);
53 return isTargetMatch(buffer, triplePrefix);
57 bool LTOModule::isBitcodeFileForTarget(const char* path,
58 const char* triplePrefix)
60 MemoryBuffer *buffer = MemoryBuffer::getFile(path);
63 return isTargetMatch(buffer, triplePrefix);
66 // takes ownership of buffer
67 bool LTOModule::isTargetMatch(MemoryBuffer* buffer, const char* triplePrefix)
69 OwningPtr<ModuleProvider> mp(getBitcodeModuleProvider(buffer,
71 // on success, mp owns buffer and both are deleted at end of this method
76 std::string actualTarget = mp->getModule()->getTargetTriple();
77 return ( strncmp(actualTarget.c_str(), triplePrefix,
78 strlen(triplePrefix)) == 0);
82 LTOModule::LTOModule(Module* m, TargetMachine* t)
83 : _module(m), _target(t), _symbolsParsed(false)
87 LTOModule* LTOModule::makeLTOModule(const char* path,
90 OwningPtr<MemoryBuffer> buffer(MemoryBuffer::getFile(path, &errMsg));
93 return makeLTOModule(buffer.get(), errMsg);
96 /// makeBuffer - create a MemoryBuffer from a memory range.
97 /// MemoryBuffer requires the byte past end of the buffer to be a zero.
98 /// We might get lucky and already be that way, otherwise make a copy.
99 /// Also if next byte is on a different page, don't assume it is readable.
100 MemoryBuffer* LTOModule::makeBuffer(const void* mem, size_t length)
102 const char* startPtr = (char*)mem;
103 const char* endPtr = startPtr+length;
104 if ( (((uintptr_t)endPtr & (sys::Process::GetPageSize()-1)) == 0)
106 return MemoryBuffer::getMemBufferCopy(startPtr, endPtr);
108 return MemoryBuffer::getMemBuffer(startPtr, endPtr);
112 LTOModule* LTOModule::makeLTOModule(const void* mem, size_t length,
115 OwningPtr<MemoryBuffer> buffer(makeBuffer(mem, length));
118 return makeLTOModule(buffer.get(), errMsg);
121 /// getFeatureString - Return a string listing the features associated with the
124 /// FIXME: This is an inelegant way of specifying the features of a
125 /// subtarget. It would be better if we could encode this information into the
126 /// IR. See <rdar://5972456>.
127 std::string getFeatureString(const char *TargetTriple) {
128 SubtargetFeatures Features;
130 if (strncmp(TargetTriple, "powerpc-apple-", 14) == 0) {
131 Features.AddFeature("altivec", true);
132 } else if (strncmp(TargetTriple, "powerpc64-apple-", 16) == 0) {
133 Features.AddFeature("64bit", true);
134 Features.AddFeature("altivec", true);
137 return Features.getString();
140 LTOModule* LTOModule::makeLTOModule(MemoryBuffer* buffer,
143 // parse bitcode buffer
144 OwningPtr<Module> m(ParseBitcodeFile(buffer, getGlobalContext(), &errMsg));
147 // find machine architecture for this module
148 const Target* march = TargetRegistry::lookupTarget(m->getTargetTriple(),
149 /*FallbackToHost=*/true,
150 /*RequireJIT=*/false,
155 // construct LTModule, hand over ownership of module and target
156 std::string FeatureStr = getFeatureString(m->getTargetTriple().c_str());
157 TargetMachine* target = march->createTargetMachine(*m, FeatureStr);
158 return new LTOModule(m.take(), target);
162 const char* LTOModule::getTargetTriple()
164 return _module->getTargetTriple().c_str();
167 void LTOModule::addDefinedFunctionSymbol(Function* f, Mangler &mangler)
169 // add to list of defined symbols
170 addDefinedSymbol(f, mangler, true);
172 // add external symbols referenced by this function.
173 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
174 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
175 for (unsigned count = 0, total = i->getNumOperands();
176 count != total; ++count) {
177 findExternalRefs(i->getOperand(count), mangler);
183 // get string that data pointer points to
184 bool LTOModule::objcClassNameFromExpression(Constant* c, std::string& name)
186 if (ConstantExpr* ce = dyn_cast<ConstantExpr>(c)) {
187 Constant* op = ce->getOperand(0);
188 if (GlobalVariable* gvn = dyn_cast<GlobalVariable>(op)) {
189 Constant* cn = gvn->getInitializer();
190 if (ConstantArray* ca = dyn_cast<ConstantArray>(cn)) {
191 if ( ca->isCString() ) {
192 name = ".objc_class_name_" + ca->getAsString();
201 // parse i386/ppc ObjC class data structure
202 void LTOModule::addObjCClass(GlobalVariable* clgv)
204 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
205 // second slot in __OBJC,__class is pointer to superclass name
206 std::string superclassName;
207 if ( objcClassNameFromExpression(c->getOperand(1), superclassName) ) {
208 NameAndAttributes info;
209 if ( _undefines.find(superclassName.c_str()) == _undefines.end() ) {
210 const char* symbolName = ::strdup(superclassName.c_str());
211 info.name = ::strdup(symbolName);
212 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
213 // string is owned by _undefines
214 _undefines[info.name] = info;
217 // third slot in __OBJC,__class is pointer to class name
218 std::string className;
219 if ( objcClassNameFromExpression(c->getOperand(2), className) ) {
220 const char* symbolName = ::strdup(className.c_str());
221 NameAndAttributes info;
222 info.name = symbolName;
223 info.attributes = (lto_symbol_attributes)
224 (LTO_SYMBOL_PERMISSIONS_DATA |
225 LTO_SYMBOL_DEFINITION_REGULAR |
226 LTO_SYMBOL_SCOPE_DEFAULT);
227 _symbols.push_back(info);
228 _defines[info.name] = 1;
234 // parse i386/ppc ObjC category data structure
235 void LTOModule::addObjCCategory(GlobalVariable* clgv)
237 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
238 // second slot in __OBJC,__category is pointer to target class name
239 std::string targetclassName;
240 if ( objcClassNameFromExpression(c->getOperand(1), targetclassName) ) {
241 NameAndAttributes info;
242 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ){
243 const char* symbolName = ::strdup(targetclassName.c_str());
244 info.name = ::strdup(symbolName);
245 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
246 // string is owned by _undefines
247 _undefines[info.name] = info;
254 // parse i386/ppc ObjC class list data structure
255 void LTOModule::addObjCClassRef(GlobalVariable* clgv)
257 std::string targetclassName;
258 if ( objcClassNameFromExpression(clgv->getInitializer(), targetclassName) ){
259 NameAndAttributes info;
260 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ) {
261 const char* symbolName = ::strdup(targetclassName.c_str());
262 info.name = ::strdup(symbolName);
263 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
264 // string is owned by _undefines
265 _undefines[info.name] = info;
271 void LTOModule::addDefinedDataSymbol(GlobalValue* v, Mangler& mangler)
273 // add to list of defined symbols
274 addDefinedSymbol(v, mangler, false);
276 // Special case i386/ppc ObjC data structures in magic sections:
277 // The issue is that the old ObjC object format did some strange
278 // contortions to avoid real linker symbols. For instance, the
279 // ObjC class data structure is allocated statically in the executable
280 // that defines that class. That data structures contains a pointer to
281 // its superclass. But instead of just initializing that part of the
282 // struct to the address of its superclass, and letting the static and
283 // dynamic linkers do the rest, the runtime works by having that field
284 // instead point to a C-string that is the name of the superclass.
285 // At runtime the objc initialization updates that pointer and sets
286 // it to point to the actual super class. As far as the linker
287 // knows it is just a pointer to a string. But then someone wanted the
288 // linker to issue errors at build time if the superclass was not found.
289 // So they figured out a way in mach-o object format to use an absolute
290 // symbols (.objc_class_name_Foo = 0) and a floating reference
291 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
292 // a class was missing.
293 // The following synthesizes the implicit .objc_* symbols for the linker
294 // from the ObjC data structures generated by the front end.
295 if ( v->hasSection() /* && isTargetDarwin */ ) {
296 // special case if this data blob is an ObjC class definition
297 if ( v->getSection().compare(0, 15, "__OBJC,__class,") == 0 ) {
298 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
303 // special case if this data blob is an ObjC category definition
304 else if ( v->getSection().compare(0, 18, "__OBJC,__category,") == 0 ) {
305 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
310 // special case if this data blob is the list of referenced classes
311 else if ( v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0 ) {
312 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
318 // add external symbols referenced by this data.
319 for (unsigned count = 0, total = v->getNumOperands();
320 count != total; ++count) {
321 findExternalRefs(v->getOperand(count), mangler);
326 void LTOModule::addDefinedSymbol(GlobalValue* def, Mangler &mangler,
329 // ignore all llvm.* symbols
330 if ( strncmp(def->getNameStart(), "llvm.", 5) == 0 )
333 // string is owned by _defines
334 const char* symbolName = ::strdup(mangler.getMangledName(def).c_str());
336 // set alignment part log2() can have rounding errors
337 uint32_t align = def->getAlignment();
338 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
340 // set permissions part
342 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
344 GlobalVariable* gv = dyn_cast<GlobalVariable>(def);
345 if ( (gv != NULL) && gv->isConstant() )
346 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
348 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
351 // set definition part
352 if ( def->hasWeakLinkage() || def->hasLinkOnceLinkage() ) {
353 attr |= LTO_SYMBOL_DEFINITION_WEAK;
355 else if ( def->hasCommonLinkage()) {
356 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
359 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
363 if ( def->hasHiddenVisibility() )
364 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
365 else if ( def->hasProtectedVisibility() )
366 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
367 else if ( def->hasExternalLinkage() || def->hasWeakLinkage()
368 || def->hasLinkOnceLinkage() || def->hasCommonLinkage() )
369 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
371 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
373 // add to table of symbols
374 NameAndAttributes info;
375 info.name = symbolName;
376 info.attributes = (lto_symbol_attributes)attr;
377 _symbols.push_back(info);
378 _defines[info.name] = 1;
381 void LTOModule::addAsmGlobalSymbol(const char *name) {
382 // only add new define if not already defined
383 if ( _defines.count(name) == 0 )
386 // string is owned by _defines
387 const char *symbolName = ::strdup(name);
388 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
389 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
390 NameAndAttributes info;
391 info.name = symbolName;
392 info.attributes = (lto_symbol_attributes)attr;
393 _symbols.push_back(info);
394 _defines[info.name] = 1;
397 void LTOModule::addPotentialUndefinedSymbol(GlobalValue* decl, Mangler &mangler)
399 // ignore all llvm.* symbols
400 if ( strncmp(decl->getNameStart(), "llvm.", 5) == 0 )
403 // ignore all aliases
404 if (isa<GlobalAlias>(decl))
407 const char* name = mangler.getMangledName(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->getKey())) {
512 NameAndAttributes info = it->getValue();
513 _symbols.push_back(info);
520 uint32_t LTOModule::getSymbolCount()
523 return _symbols.size();
527 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index)
530 if ( index < _symbols.size() )
531 return _symbols[index].attributes;
533 return lto_symbol_attributes(0);
536 const char* LTOModule::getSymbolName(uint32_t index)
539 if ( index < _symbols.size() )
540 return _symbols[index].name;