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/ADT/OwningPtr.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/Bitcode/ReaderWriter.h"
23 #include "llvm/Support/SystemUtils.h"
24 #include "llvm/Support/MemoryBuffer.h"
25 #include "llvm/Support/MathExtras.h"
26 #include "llvm/Support/Host.h"
27 #include "llvm/Support/Path.h"
28 #include "llvm/Support/Process.h"
29 #include "llvm/Support/system_error.h"
30 #include "llvm/Target/Mangler.h"
31 #include "llvm/Target/SubtargetFeature.h"
32 #include "llvm/MC/MCAsmInfo.h"
33 #include "llvm/MC/MCContext.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetRegistry.h"
36 #include "llvm/Target/TargetSelect.h"
40 bool LTOModule::isBitcodeFile(const void *mem, size_t length) {
41 return llvm::sys::IdentifyFileType((char*)mem, length)
42 == llvm::sys::Bitcode_FileType;
45 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) {
51 MemoryBuffer *buffer = makeBuffer(mem, length);
54 return isTargetMatch(buffer, triplePrefix);
58 bool LTOModule::isBitcodeFileForTarget(const char *path,
59 const char *triplePrefix) {
60 OwningPtr<MemoryBuffer> buffer;
61 if (MemoryBuffer::getFile(path, buffer))
63 return isTargetMatch(buffer.take(), triplePrefix);
66 // Takes ownership of buffer.
67 bool LTOModule::isTargetMatch(MemoryBuffer *buffer, const char *triplePrefix) {
68 std::string Triple = getBitcodeTargetTriple(buffer, getGlobalContext());
70 return (strncmp(Triple.c_str(), triplePrefix,
71 strlen(triplePrefix)) == 0);
75 LTOModule::LTOModule(Module *m, TargetMachine *t)
76 : _module(m), _target(t), _symbolsParsed(false)
80 LTOModule *LTOModule::makeLTOModule(const char *path,
81 std::string &errMsg) {
82 OwningPtr<MemoryBuffer> buffer;
83 if (error_code ec = MemoryBuffer::getFile(path, buffer)) {
84 errMsg = ec.message();
87 return makeLTOModule(buffer.get(), errMsg);
90 LTOModule *LTOModule::makeLTOModule(int fd, const char *path,
92 std::string &errMsg) {
93 OwningPtr<MemoryBuffer> buffer;
94 if (error_code ec = MemoryBuffer::getOpenFile(fd, path, buffer, size)) {
95 errMsg = ec.message();
98 return makeLTOModule(buffer.get(), errMsg);
101 /// makeBuffer - Create a MemoryBuffer from a memory range. MemoryBuffer
102 /// requires the byte past end of the buffer to be a zero. We might get lucky
103 /// and already be that way, otherwise make a copy. Also if next byte is on a
104 /// different page, don't assume it is readable.
105 MemoryBuffer *LTOModule::makeBuffer(const void *mem, size_t length) {
106 const char *startPtr = (char*)mem;
107 const char *endPtr = startPtr+length;
108 if (((uintptr_t)endPtr & (sys::Process::GetPageSize()-1)) == 0 ||
110 return MemoryBuffer::getMemBufferCopy(StringRef(startPtr, length));
112 return MemoryBuffer::getMemBuffer(StringRef(startPtr, length));
116 LTOModule *LTOModule::makeLTOModule(const void *mem, size_t length,
117 std::string &errMsg) {
118 OwningPtr<MemoryBuffer> buffer(makeBuffer(mem, length));
121 return makeLTOModule(buffer.get(), errMsg);
124 LTOModule *LTOModule::makeLTOModule(MemoryBuffer *buffer,
125 std::string &errMsg) {
126 InitializeAllTargets();
128 // parse bitcode buffer
129 OwningPtr<Module> m(ParseBitcodeFile(buffer, getGlobalContext(), &errMsg));
133 std::string Triple = m->getTargetTriple();
135 Triple = sys::getHostTriple();
137 // find machine architecture for this module
138 const Target *march = TargetRegistry::lookupTarget(Triple, errMsg);
142 // construct LTModule, hand over ownership of module and target
143 SubtargetFeatures Features;
144 Features.getDefaultSubtargetFeatures("" /* cpu */, llvm::Triple(Triple));
145 std::string FeatureStr = Features.getString();
146 TargetMachine *target = march->createTargetMachine(Triple, FeatureStr);
147 return new LTOModule(m.take(), target);
151 const char *LTOModule::getTargetTriple() {
152 return _module->getTargetTriple().c_str();
155 void LTOModule::setTargetTriple(const char *triple) {
156 _module->setTargetTriple(triple);
159 void LTOModule::addDefinedFunctionSymbol(Function *f, Mangler &mangler) {
160 // add to list of defined symbols
161 addDefinedSymbol(f, mangler, true);
163 // add external symbols referenced by this function.
164 for (Function::iterator b = f->begin(); b != f->end(); ++b) {
165 for (BasicBlock::iterator i = b->begin(); i != b->end(); ++i) {
166 for (unsigned count = 0, total = i->getNumOperands();
167 count != total; ++count) {
168 findExternalRefs(i->getOperand(count), mangler);
174 // Get string that data pointer points to.
175 bool LTOModule::objcClassNameFromExpression(Constant *c, std::string &name) {
176 if (ConstantExpr *ce = dyn_cast<ConstantExpr>(c)) {
177 Constant *op = ce->getOperand(0);
178 if (GlobalVariable *gvn = dyn_cast<GlobalVariable>(op)) {
179 Constant *cn = gvn->getInitializer();
180 if (ConstantArray *ca = dyn_cast<ConstantArray>(cn)) {
181 if (ca->isCString()) {
182 name = ".objc_class_name_" + ca->getAsString();
191 // Parse i386/ppc ObjC class data structure.
192 void LTOModule::addObjCClass(GlobalVariable *clgv) {
193 if (ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
194 // second slot in __OBJC,__class is pointer to superclass name
195 std::string superclassName;
196 if (objcClassNameFromExpression(c->getOperand(1), superclassName)) {
197 NameAndAttributes info;
198 if (_undefines.find(superclassName.c_str()) == _undefines.end()) {
199 const char *symbolName = ::strdup(superclassName.c_str());
200 info.name = symbolName;
201 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
202 // string is owned by _undefines
203 _undefines[info.name] = info;
206 // third slot in __OBJC,__class is pointer to class name
207 std::string className;
208 if (objcClassNameFromExpression(c->getOperand(2), className)) {
209 const char *symbolName = ::strdup(className.c_str());
210 NameAndAttributes info;
211 info.name = symbolName;
212 info.attributes = (lto_symbol_attributes)
213 (LTO_SYMBOL_PERMISSIONS_DATA |
214 LTO_SYMBOL_DEFINITION_REGULAR |
215 LTO_SYMBOL_SCOPE_DEFAULT);
216 _symbols.push_back(info);
217 _defines[info.name] = 1;
223 // Parse i386/ppc ObjC category data structure.
224 void LTOModule::addObjCCategory(GlobalVariable *clgv) {
225 if (ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
226 // second slot in __OBJC,__category is pointer to target class name
227 std::string targetclassName;
228 if (objcClassNameFromExpression(c->getOperand(1), targetclassName)) {
229 NameAndAttributes info;
230 if (_undefines.find(targetclassName.c_str()) == _undefines.end()) {
231 const char *symbolName = ::strdup(targetclassName.c_str());
232 info.name = symbolName;
233 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
234 // string is owned by _undefines
235 _undefines[info.name] = info;
242 // Parse i386/ppc ObjC class list data structure.
243 void LTOModule::addObjCClassRef(GlobalVariable *clgv) {
244 std::string targetclassName;
245 if (objcClassNameFromExpression(clgv->getInitializer(), targetclassName)) {
246 NameAndAttributes info;
247 if (_undefines.find(targetclassName.c_str()) == _undefines.end()) {
248 const char *symbolName = ::strdup(targetclassName.c_str());
249 info.name = symbolName;
250 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
251 // string is owned by _undefines
252 _undefines[info.name] = info;
258 void LTOModule::addDefinedDataSymbol(GlobalValue *v, Mangler &mangler) {
259 // Add to list of defined symbols.
260 addDefinedSymbol(v, mangler, false);
262 // Special case i386/ppc ObjC data structures in magic sections:
263 // The issue is that the old ObjC object format did some strange
264 // contortions to avoid real linker symbols. For instance, the
265 // ObjC class data structure is allocated statically in the executable
266 // that defines that class. That data structures contains a pointer to
267 // its superclass. But instead of just initializing that part of the
268 // struct to the address of its superclass, and letting the static and
269 // dynamic linkers do the rest, the runtime works by having that field
270 // instead point to a C-string that is the name of the superclass.
271 // At runtime the objc initialization updates that pointer and sets
272 // it to point to the actual super class. As far as the linker
273 // knows it is just a pointer to a string. But then someone wanted the
274 // linker to issue errors at build time if the superclass was not found.
275 // So they figured out a way in mach-o object format to use an absolute
276 // symbols (.objc_class_name_Foo = 0) and a floating reference
277 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
278 // a class was missing.
279 // The following synthesizes the implicit .objc_* symbols for the linker
280 // from the ObjC data structures generated by the front end.
281 if (v->hasSection() /* && isTargetDarwin */) {
282 // special case if this data blob is an ObjC class definition
283 if (v->getSection().compare(0, 15, "__OBJC,__class,") == 0) {
284 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
289 // special case if this data blob is an ObjC category definition
290 else if (v->getSection().compare(0, 18, "__OBJC,__category,") == 0) {
291 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
296 // special case if this data blob is the list of referenced classes
297 else if (v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0) {
298 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
304 // add external symbols referenced by this data.
305 for (unsigned count = 0, total = v->getNumOperands();
306 count != total; ++count) {
307 findExternalRefs(v->getOperand(count), mangler);
312 void LTOModule::addDefinedSymbol(GlobalValue *def, Mangler &mangler,
314 // ignore all llvm.* symbols
315 if (def->getName().startswith("llvm."))
318 // ignore available_externally
319 if (def->hasAvailableExternallyLinkage())
322 // string is owned by _defines
323 const char *symbolName = ::strdup(mangler.getNameWithPrefix(def).c_str());
325 // set alignment part log2() can have rounding errors
326 uint32_t align = def->getAlignment();
327 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
329 // set permissions part
331 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
333 GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
334 if (gv && gv->isConstant())
335 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
337 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
340 // set definition part
341 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage() ||
342 def->hasLinkerPrivateWeakLinkage() ||
343 def->hasLinkerPrivateWeakDefAutoLinkage())
344 attr |= LTO_SYMBOL_DEFINITION_WEAK;
345 else if (def->hasCommonLinkage())
346 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
348 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
351 if (def->hasHiddenVisibility())
352 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
353 else if (def->hasProtectedVisibility())
354 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
355 else if (def->hasExternalLinkage() || def->hasWeakLinkage() ||
356 def->hasLinkOnceLinkage() || def->hasCommonLinkage() ||
357 def->hasLinkerPrivateWeakLinkage())
358 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
359 else if (def->hasLinkerPrivateWeakDefAutoLinkage())
360 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
362 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
364 // add to table of symbols
365 NameAndAttributes info;
366 info.name = symbolName;
367 info.attributes = (lto_symbol_attributes)attr;
368 _symbols.push_back(info);
369 _defines[info.name] = 1;
372 void LTOModule::addAsmGlobalSymbol(const char *name) {
373 // only add new define if not already defined
374 if (_defines.count(name))
377 // string is owned by _defines
378 const char *symbolName = ::strdup(name);
379 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
380 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
381 NameAndAttributes info;
382 info.name = symbolName;
383 info.attributes = (lto_symbol_attributes)attr;
384 _symbols.push_back(info);
385 _defines[info.name] = 1;
388 void LTOModule::addPotentialUndefinedSymbol(GlobalValue *decl,
390 // ignore all llvm.* symbols
391 if (decl->getName().startswith("llvm."))
394 // ignore all aliases
395 if (isa<GlobalAlias>(decl))
398 std::string name = mangler.getNameWithPrefix(decl);
400 // we already have the symbol
401 if (_undefines.find(name) != _undefines.end())
404 NameAndAttributes info;
405 // string is owned by _undefines
406 info.name = ::strdup(name.c_str());
407 if (decl->hasExternalWeakLinkage())
408 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
410 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
411 _undefines[name] = info;
416 // Find external symbols referenced by VALUE. This is a recursive function.
417 void LTOModule::findExternalRefs(Value *value, Mangler &mangler) {
418 if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) {
419 if (!gv->hasExternalLinkage())
420 addPotentialUndefinedSymbol(gv, mangler);
421 // If this is a variable definition, do not recursively process
422 // initializer. It might contain a reference to this variable
423 // and cause an infinite loop. The initializer will be
424 // processed in addDefinedDataSymbol().
428 // GlobalValue, even with InternalLinkage type, may have operands with
429 // ExternalLinkage type. Do not ignore these operands.
430 if (Constant *c = dyn_cast<Constant>(value)) {
431 // Handle ConstantExpr, ConstantStruct, ConstantArry etc.
432 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
433 findExternalRefs(c->getOperand(i), mangler);
437 void LTOModule::lazyParseSymbols() {
441 _symbolsParsed = true;
443 // Use mangler to add GlobalPrefix to names to match linker names.
444 MCContext Context(*_target->getMCAsmInfo(), NULL);
445 Mangler mangler(Context, *_target->getTargetData());
448 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
449 if (f->isDeclaration())
450 addPotentialUndefinedSymbol(f, mangler);
452 addDefinedFunctionSymbol(f, mangler);
456 for (Module::global_iterator v = _module->global_begin(),
457 e = _module->global_end(); v != e; ++v) {
458 if (v->isDeclaration())
459 addPotentialUndefinedSymbol(v, mangler);
461 addDefinedDataSymbol(v, mangler);
465 const std::string &inlineAsm = _module->getModuleInlineAsm();
466 const std::string glbl = ".globl";
467 std::string asmSymbolName;
468 std::string::size_type pos = inlineAsm.find(glbl, 0);
469 while (pos != std::string::npos) {
473 // skip white space between .globl and symbol name
474 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
475 if (pbegin == std::string::npos)
479 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
480 if (pend == std::string::npos)
483 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
484 addAsmGlobalSymbol(asmSymbolName.c_str());
486 // search next .globl
487 pos = inlineAsm.find(glbl, pend);
491 for (Module::alias_iterator i = _module->alias_begin(),
492 e = _module->alias_end(); i != e; ++i) {
493 if (i->isDeclaration())
494 addPotentialUndefinedSymbol(i, mangler);
496 addDefinedDataSymbol(i, mangler);
499 // make symbols for all undefines
500 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
501 it != _undefines.end(); ++it) {
502 // if this symbol also has a definition, then don't make an undefine
503 // because it is a tentative definition
504 if (_defines.count(it->getKey()) == 0) {
505 NameAndAttributes info = it->getValue();
506 _symbols.push_back(info);
512 uint32_t LTOModule::getSymbolCount() {
514 return _symbols.size();
518 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index) {
520 if (index < _symbols.size())
521 return _symbols[index].attributes;
523 return lto_symbol_attributes(0);
526 const char *LTOModule::getSymbolName(uint32_t index) {
528 if (index < _symbols.size())
529 return _symbols[index].name;