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/Host.h"
28 #include "llvm/System/Path.h"
29 #include "llvm/System/Process.h"
30 #include "llvm/Target/SubtargetFeature.h"
31 #include "llvm/Target/TargetAsmInfo.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetRegistry.h"
34 #include "llvm/Target/TargetSelect.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 InitializeAllTargets();
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,
147 InitializeAllTargets();
149 // parse bitcode buffer
150 OwningPtr<Module> m(ParseBitcodeFile(buffer, getGlobalContext(), &errMsg));
154 std::string Triple = m->getTargetTriple();
156 Triple = sys::getHostTriple();
158 // find machine architecture for this module
159 const Target* march = TargetRegistry::lookupTarget(Triple,
160 /*FallbackToHost=*/false,
161 /*RequireJIT=*/false,
166 // construct LTModule, hand over ownership of module and target
167 std::string FeatureStr = getFeatureString(Triple.c_str());
168 TargetMachine* target = march->createTargetMachine(*m, Triple, FeatureStr);
169 return new LTOModule(m.take(), target);
173 const char* LTOModule::getTargetTriple()
175 return _module->getTargetTriple().c_str();
178 void LTOModule::addDefinedFunctionSymbol(Function* f, Mangler &mangler)
180 // add to list of defined symbols
181 addDefinedSymbol(f, mangler, true);
183 // add external symbols referenced by this function.
184 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
185 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
186 for (unsigned count = 0, total = i->getNumOperands();
187 count != total; ++count) {
188 findExternalRefs(i->getOperand(count), mangler);
194 // get string that data pointer points to
195 bool LTOModule::objcClassNameFromExpression(Constant* c, std::string& name)
197 if (ConstantExpr* ce = dyn_cast<ConstantExpr>(c)) {
198 Constant* op = ce->getOperand(0);
199 if (GlobalVariable* gvn = dyn_cast<GlobalVariable>(op)) {
200 Constant* cn = gvn->getInitializer();
201 if (ConstantArray* ca = dyn_cast<ConstantArray>(cn)) {
202 if ( ca->isCString() ) {
203 name = ".objc_class_name_" + ca->getAsString();
212 // parse i386/ppc ObjC class data structure
213 void LTOModule::addObjCClass(GlobalVariable* clgv)
215 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
216 // second slot in __OBJC,__class is pointer to superclass name
217 std::string superclassName;
218 if ( objcClassNameFromExpression(c->getOperand(1), superclassName) ) {
219 NameAndAttributes info;
220 if ( _undefines.find(superclassName.c_str()) == _undefines.end() ) {
221 const char* symbolName = ::strdup(superclassName.c_str());
222 info.name = ::strdup(symbolName);
223 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
224 // string is owned by _undefines
225 _undefines[info.name] = info;
228 // third slot in __OBJC,__class is pointer to class name
229 std::string className;
230 if ( objcClassNameFromExpression(c->getOperand(2), className) ) {
231 const char* symbolName = ::strdup(className.c_str());
232 NameAndAttributes info;
233 info.name = symbolName;
234 info.attributes = (lto_symbol_attributes)
235 (LTO_SYMBOL_PERMISSIONS_DATA |
236 LTO_SYMBOL_DEFINITION_REGULAR |
237 LTO_SYMBOL_SCOPE_DEFAULT);
238 _symbols.push_back(info);
239 _defines[info.name] = 1;
245 // parse i386/ppc ObjC category data structure
246 void LTOModule::addObjCCategory(GlobalVariable* clgv)
248 if (ConstantStruct* c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
249 // second slot in __OBJC,__category is pointer to target class name
250 std::string targetclassName;
251 if ( objcClassNameFromExpression(c->getOperand(1), targetclassName) ) {
252 NameAndAttributes info;
253 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ){
254 const char* symbolName = ::strdup(targetclassName.c_str());
255 info.name = ::strdup(symbolName);
256 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
257 // string is owned by _undefines
258 _undefines[info.name] = info;
265 // parse i386/ppc ObjC class list data structure
266 void LTOModule::addObjCClassRef(GlobalVariable* clgv)
268 std::string targetclassName;
269 if ( objcClassNameFromExpression(clgv->getInitializer(), targetclassName) ){
270 NameAndAttributes info;
271 if ( _undefines.find(targetclassName.c_str()) == _undefines.end() ) {
272 const char* symbolName = ::strdup(targetclassName.c_str());
273 info.name = ::strdup(symbolName);
274 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
275 // string is owned by _undefines
276 _undefines[info.name] = info;
282 void LTOModule::addDefinedDataSymbol(GlobalValue* v, Mangler& mangler)
284 // add to list of defined symbols
285 addDefinedSymbol(v, mangler, false);
287 // Special case i386/ppc ObjC data structures in magic sections:
288 // The issue is that the old ObjC object format did some strange
289 // contortions to avoid real linker symbols. For instance, the
290 // ObjC class data structure is allocated statically in the executable
291 // that defines that class. That data structures contains a pointer to
292 // its superclass. But instead of just initializing that part of the
293 // struct to the address of its superclass, and letting the static and
294 // dynamic linkers do the rest, the runtime works by having that field
295 // instead point to a C-string that is the name of the superclass.
296 // At runtime the objc initialization updates that pointer and sets
297 // it to point to the actual super class. As far as the linker
298 // knows it is just a pointer to a string. But then someone wanted the
299 // linker to issue errors at build time if the superclass was not found.
300 // So they figured out a way in mach-o object format to use an absolute
301 // symbols (.objc_class_name_Foo = 0) and a floating reference
302 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
303 // a class was missing.
304 // The following synthesizes the implicit .objc_* symbols for the linker
305 // from the ObjC data structures generated by the front end.
306 if ( v->hasSection() /* && isTargetDarwin */ ) {
307 // special case if this data blob is an ObjC class definition
308 if ( v->getSection().compare(0, 15, "__OBJC,__class,") == 0 ) {
309 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
314 // special case if this data blob is an ObjC category definition
315 else if ( v->getSection().compare(0, 18, "__OBJC,__category,") == 0 ) {
316 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
321 // special case if this data blob is the list of referenced classes
322 else if ( v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0 ) {
323 if (GlobalVariable* gv = dyn_cast<GlobalVariable>(v)) {
329 // add external symbols referenced by this data.
330 for (unsigned count = 0, total = v->getNumOperands();
331 count != total; ++count) {
332 findExternalRefs(v->getOperand(count), mangler);
337 void LTOModule::addDefinedSymbol(GlobalValue* def, Mangler &mangler,
340 // ignore all llvm.* symbols
341 if (def->getName().startswith("llvm."))
344 // string is owned by _defines
345 const char* symbolName = ::strdup(mangler.getMangledName(def).c_str());
347 // set alignment part log2() can have rounding errors
348 uint32_t align = def->getAlignment();
349 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
351 // set permissions part
353 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
355 GlobalVariable* gv = dyn_cast<GlobalVariable>(def);
356 if ( (gv != NULL) && gv->isConstant() )
357 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
359 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
362 // set definition part
363 if ( def->hasWeakLinkage() || def->hasLinkOnceLinkage() ) {
364 attr |= LTO_SYMBOL_DEFINITION_WEAK;
366 else if ( def->hasCommonLinkage()) {
367 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
370 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
374 if ( def->hasHiddenVisibility() )
375 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
376 else if ( def->hasProtectedVisibility() )
377 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
378 else if ( def->hasExternalLinkage() || def->hasWeakLinkage()
379 || def->hasLinkOnceLinkage() || def->hasCommonLinkage() )
380 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
382 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
384 // add to table of symbols
385 NameAndAttributes info;
386 info.name = symbolName;
387 info.attributes = (lto_symbol_attributes)attr;
388 _symbols.push_back(info);
389 _defines[info.name] = 1;
392 void LTOModule::addAsmGlobalSymbol(const char *name) {
393 // only add new define if not already defined
394 if ( _defines.count(name) == 0 )
397 // string is owned by _defines
398 const char *symbolName = ::strdup(name);
399 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
400 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
401 NameAndAttributes info;
402 info.name = symbolName;
403 info.attributes = (lto_symbol_attributes)attr;
404 _symbols.push_back(info);
405 _defines[info.name] = 1;
408 void LTOModule::addPotentialUndefinedSymbol(GlobalValue* decl, Mangler &mangler)
410 // ignore all llvm.* symbols
411 if (decl->getName().startswith("llvm."))
414 // ignore all aliases
415 if (isa<GlobalAlias>(decl))
418 std::string name = mangler.getMangledName(decl);
420 // we already have the symbol
421 if (_undefines.find(name) != _undefines.end())
424 NameAndAttributes info;
425 // string is owned by _undefines
426 info.name = ::strdup(name.c_str());
427 if (decl->hasExternalWeakLinkage())
428 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
430 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
431 _undefines[name] = info;
436 // Find external symbols referenced by VALUE. This is a recursive function.
437 void LTOModule::findExternalRefs(Value* value, Mangler &mangler) {
439 if (GlobalValue* gv = dyn_cast<GlobalValue>(value)) {
440 if ( !gv->hasExternalLinkage() )
441 addPotentialUndefinedSymbol(gv, mangler);
442 // If this is a variable definition, do not recursively process
443 // initializer. It might contain a reference to this variable
444 // and cause an infinite loop. The initializer will be
445 // processed in addDefinedDataSymbol().
449 // GlobalValue, even with InternalLinkage type, may have operands with
450 // ExternalLinkage type. Do not ignore these operands.
451 if (Constant* c = dyn_cast<Constant>(value)) {
452 // Handle ConstantExpr, ConstantStruct, ConstantArry etc..
453 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
454 findExternalRefs(c->getOperand(i), mangler);
458 void LTOModule::lazyParseSymbols()
460 if ( !_symbolsParsed ) {
461 _symbolsParsed = true;
463 // Use mangler to add GlobalPrefix to names to match linker names.
464 Mangler mangler(*_module, _target->getTargetAsmInfo()->getGlobalPrefix());
465 // add chars used in ObjC method names so method names aren't mangled
466 mangler.markCharAcceptable('[');
467 mangler.markCharAcceptable(']');
468 mangler.markCharAcceptable('(');
469 mangler.markCharAcceptable(')');
470 mangler.markCharAcceptable('-');
471 mangler.markCharAcceptable('+');
472 mangler.markCharAcceptable(' ');
475 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
476 if ( f->isDeclaration() )
477 addPotentialUndefinedSymbol(f, mangler);
479 addDefinedFunctionSymbol(f, mangler);
483 for (Module::global_iterator v = _module->global_begin(),
484 e = _module->global_end(); v != e; ++v) {
485 if ( v->isDeclaration() )
486 addPotentialUndefinedSymbol(v, mangler);
488 addDefinedDataSymbol(v, mangler);
492 const std::string &inlineAsm = _module->getModuleInlineAsm();
493 const std::string glbl = ".globl";
494 std::string asmSymbolName;
495 std::string::size_type pos = inlineAsm.find(glbl, 0);
496 while (pos != std::string::npos) {
500 // skip white space between .globl and symbol name
501 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
502 if (pbegin == std::string::npos)
506 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
507 if (pend == std::string::npos)
510 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
511 addAsmGlobalSymbol(asmSymbolName.c_str());
513 // search next .globl
514 pos = inlineAsm.find(glbl, pend);
517 // make symbols for all undefines
518 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
519 it != _undefines.end(); ++it) {
520 // if this symbol also has a definition, then don't make an undefine
521 // because it is a tentative definition
522 if ( _defines.count(it->getKey()) == 0 ) {
523 NameAndAttributes info = it->getValue();
524 _symbols.push_back(info);
531 uint32_t LTOModule::getSymbolCount()
534 return _symbols.size();
538 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index)
541 if ( index < _symbols.size() )
542 return _symbols[index].attributes;
544 return lto_symbol_attributes(0);
547 const char* LTOModule::getSymbolName(uint32_t index)
550 if ( index < _symbols.size() )
551 return _symbols[index].name;