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 StringMap<NameAndAttributes>::value_type &entry =
199 _undefines.GetOrCreateValue(superclassName.c_str());
200 if (!entry.getValue().name) {
201 const char *symbolName = entry.getKey().data();
202 info.name = symbolName;
203 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
204 entry.setValue(info);
207 // third slot in __OBJC,__class is pointer to class name
208 std::string className;
209 if (objcClassNameFromExpression(c->getOperand(2), className)) {
210 StringSet::value_type &entry =
211 _defines.GetOrCreateValue(className.c_str());
213 NameAndAttributes info;
214 info.name = entry.getKey().data();
215 info.attributes = (lto_symbol_attributes)
216 (LTO_SYMBOL_PERMISSIONS_DATA |
217 LTO_SYMBOL_DEFINITION_REGULAR |
218 LTO_SYMBOL_SCOPE_DEFAULT);
219 _symbols.push_back(info);
225 // Parse i386/ppc ObjC category data structure.
226 void LTOModule::addObjCCategory(GlobalVariable *clgv) {
227 if (ConstantStruct *c = dyn_cast<ConstantStruct>(clgv->getInitializer())) {
228 // second slot in __OBJC,__category is pointer to target class name
229 std::string targetclassName;
230 if (objcClassNameFromExpression(c->getOperand(1), targetclassName)) {
231 NameAndAttributes info;
233 StringMap<NameAndAttributes>::value_type &entry =
234 _undefines.GetOrCreateValue(targetclassName.c_str());
236 if (entry.getValue().name)
239 const char *symbolName = entry.getKey().data();
240 info.name = symbolName;
241 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
242 entry.setValue(info);
248 // Parse i386/ppc ObjC class list data structure.
249 void LTOModule::addObjCClassRef(GlobalVariable *clgv) {
250 std::string targetclassName;
251 if (objcClassNameFromExpression(clgv->getInitializer(), targetclassName)) {
252 NameAndAttributes info;
254 StringMap<NameAndAttributes>::value_type &entry =
255 _undefines.GetOrCreateValue(targetclassName.c_str());
256 if (entry.getValue().name)
259 const char *symbolName = entry.getKey().data();
260 info.name = symbolName;
261 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
262 entry.setValue(info);
267 void LTOModule::addDefinedDataSymbol(GlobalValue *v, Mangler &mangler) {
268 // Add to list of defined symbols.
269 addDefinedSymbol(v, mangler, false);
271 // Special case i386/ppc ObjC data structures in magic sections:
272 // The issue is that the old ObjC object format did some strange
273 // contortions to avoid real linker symbols. For instance, the
274 // ObjC class data structure is allocated statically in the executable
275 // that defines that class. That data structures contains a pointer to
276 // its superclass. But instead of just initializing that part of the
277 // struct to the address of its superclass, and letting the static and
278 // dynamic linkers do the rest, the runtime works by having that field
279 // instead point to a C-string that is the name of the superclass.
280 // At runtime the objc initialization updates that pointer and sets
281 // it to point to the actual super class. As far as the linker
282 // knows it is just a pointer to a string. But then someone wanted the
283 // linker to issue errors at build time if the superclass was not found.
284 // So they figured out a way in mach-o object format to use an absolute
285 // symbols (.objc_class_name_Foo = 0) and a floating reference
286 // (.reference .objc_class_name_Bar) to cause the linker into erroring when
287 // a class was missing.
288 // The following synthesizes the implicit .objc_* symbols for the linker
289 // from the ObjC data structures generated by the front end.
290 if (v->hasSection() /* && isTargetDarwin */) {
291 // special case if this data blob is an ObjC class definition
292 if (v->getSection().compare(0, 15, "__OBJC,__class,") == 0) {
293 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
298 // special case if this data blob is an ObjC category definition
299 else if (v->getSection().compare(0, 18, "__OBJC,__category,") == 0) {
300 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
305 // special case if this data blob is the list of referenced classes
306 else if (v->getSection().compare(0, 18, "__OBJC,__cls_refs,") == 0) {
307 if (GlobalVariable *gv = dyn_cast<GlobalVariable>(v)) {
313 // add external symbols referenced by this data.
314 for (unsigned count = 0, total = v->getNumOperands();
315 count != total; ++count) {
316 findExternalRefs(v->getOperand(count), mangler);
321 void LTOModule::addDefinedSymbol(GlobalValue *def, Mangler &mangler,
323 // ignore all llvm.* symbols
324 if (def->getName().startswith("llvm."))
327 // ignore available_externally
328 if (def->hasAvailableExternallyLinkage())
331 // string is owned by _defines
332 SmallString<64> Buffer;
333 mangler.getNameWithPrefix(Buffer, def, false);
335 // set alignment part log2() can have rounding errors
336 uint32_t align = def->getAlignment();
337 uint32_t attr = align ? CountTrailingZeros_32(def->getAlignment()) : 0;
339 // set permissions part
341 attr |= LTO_SYMBOL_PERMISSIONS_CODE;
343 GlobalVariable *gv = dyn_cast<GlobalVariable>(def);
344 if (gv && gv->isConstant())
345 attr |= LTO_SYMBOL_PERMISSIONS_RODATA;
347 attr |= LTO_SYMBOL_PERMISSIONS_DATA;
350 // set definition part
351 if (def->hasWeakLinkage() || def->hasLinkOnceLinkage() ||
352 def->hasLinkerPrivateWeakLinkage() ||
353 def->hasLinkerPrivateWeakDefAutoLinkage())
354 attr |= LTO_SYMBOL_DEFINITION_WEAK;
355 else if (def->hasCommonLinkage())
356 attr |= LTO_SYMBOL_DEFINITION_TENTATIVE;
358 attr |= LTO_SYMBOL_DEFINITION_REGULAR;
361 if (def->hasHiddenVisibility())
362 attr |= LTO_SYMBOL_SCOPE_HIDDEN;
363 else if (def->hasProtectedVisibility())
364 attr |= LTO_SYMBOL_SCOPE_PROTECTED;
365 else if (def->hasExternalLinkage() || def->hasWeakLinkage() ||
366 def->hasLinkOnceLinkage() || def->hasCommonLinkage() ||
367 def->hasLinkerPrivateWeakLinkage())
368 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
369 else if (def->hasLinkerPrivateWeakDefAutoLinkage())
370 attr |= LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN;
372 attr |= LTO_SYMBOL_SCOPE_INTERNAL;
374 // add to table of symbols
375 NameAndAttributes info;
376 StringSet::value_type &entry = _defines.GetOrCreateValue(Buffer.c_str());
379 StringRef Name = entry.getKey();
380 info.name = Name.data();
381 assert(info.name[Name.size()] == '\0');
382 info.attributes = (lto_symbol_attributes)attr;
383 _symbols.push_back(info);
386 void LTOModule::addAsmGlobalSymbol(const char *name) {
387 StringSet::value_type &entry = _defines.GetOrCreateValue(name);
389 // only add new define if not already defined
390 if (entry.getValue())
394 const char *symbolName = entry.getKey().data();
395 uint32_t attr = LTO_SYMBOL_DEFINITION_REGULAR;
396 attr |= LTO_SYMBOL_SCOPE_DEFAULT;
397 NameAndAttributes info;
398 info.name = symbolName;
399 info.attributes = (lto_symbol_attributes)attr;
400 _symbols.push_back(info);
403 void LTOModule::addPotentialUndefinedSymbol(GlobalValue *decl,
405 // ignore all llvm.* symbols
406 if (decl->getName().startswith("llvm."))
409 // ignore all aliases
410 if (isa<GlobalAlias>(decl))
413 SmallString<64> name;
414 mangler.getNameWithPrefix(name, decl, false);
416 StringMap<NameAndAttributes>::value_type &entry =
417 _undefines.GetOrCreateValue(name.c_str());
419 // we already have the symbol
420 if (entry.getValue().name)
423 NameAndAttributes info;
425 info.name = entry.getKey().data();
426 if (decl->hasExternalWeakLinkage())
427 info.attributes = LTO_SYMBOL_DEFINITION_WEAKUNDEF;
429 info.attributes = LTO_SYMBOL_DEFINITION_UNDEFINED;
431 entry.setValue(info);
436 // Find external symbols referenced by VALUE. This is a recursive function.
437 void LTOModule::findExternalRefs(Value *value, Mangler &mangler) {
438 if (GlobalValue *gv = dyn_cast<GlobalValue>(value)) {
439 if (!gv->hasExternalLinkage())
440 addPotentialUndefinedSymbol(gv, mangler);
441 // If this is a variable definition, do not recursively process
442 // initializer. It might contain a reference to this variable
443 // and cause an infinite loop. The initializer will be
444 // processed in addDefinedDataSymbol().
448 // GlobalValue, even with InternalLinkage type, may have operands with
449 // ExternalLinkage type. Do not ignore these operands.
450 if (Constant *c = dyn_cast<Constant>(value)) {
451 // Handle ConstantExpr, ConstantStruct, ConstantArry etc.
452 for (unsigned i = 0, e = c->getNumOperands(); i != e; ++i)
453 findExternalRefs(c->getOperand(i), mangler);
457 void LTOModule::lazyParseSymbols() {
461 _symbolsParsed = true;
463 // Use mangler to add GlobalPrefix to names to match linker names.
464 MCContext Context(*_target->getMCAsmInfo(), NULL);
465 Mangler mangler(Context, *_target->getTargetData());
468 for (Module::iterator f = _module->begin(); f != _module->end(); ++f) {
469 if (f->isDeclaration())
470 addPotentialUndefinedSymbol(f, mangler);
472 addDefinedFunctionSymbol(f, mangler);
476 for (Module::global_iterator v = _module->global_begin(),
477 e = _module->global_end(); v != e; ++v) {
478 if (v->isDeclaration())
479 addPotentialUndefinedSymbol(v, mangler);
481 addDefinedDataSymbol(v, mangler);
485 const std::string &inlineAsm = _module->getModuleInlineAsm();
486 const std::string glbl = ".globl";
487 std::string asmSymbolName;
488 std::string::size_type pos = inlineAsm.find(glbl, 0);
489 while (pos != std::string::npos) {
493 // skip white space between .globl and symbol name
494 std::string::size_type pbegin = inlineAsm.find_first_not_of(' ', pos);
495 if (pbegin == std::string::npos)
499 std::string::size_type pend = inlineAsm.find_first_of('\n', pbegin);
500 if (pend == std::string::npos)
503 asmSymbolName.assign(inlineAsm, pbegin, pend - pbegin);
504 addAsmGlobalSymbol(asmSymbolName.c_str());
506 // search next .globl
507 pos = inlineAsm.find(glbl, pend);
511 for (Module::alias_iterator i = _module->alias_begin(),
512 e = _module->alias_end(); i != e; ++i) {
513 if (i->isDeclaration())
514 addPotentialUndefinedSymbol(i, mangler);
516 addDefinedDataSymbol(i, mangler);
519 // make symbols for all undefines
520 for (StringMap<NameAndAttributes>::iterator it=_undefines.begin();
521 it != _undefines.end(); ++it) {
522 // if this symbol also has a definition, then don't make an undefine
523 // because it is a tentative definition
524 if (_defines.count(it->getKey()) == 0) {
525 NameAndAttributes info = it->getValue();
526 _symbols.push_back(info);
532 uint32_t LTOModule::getSymbolCount() {
534 return _symbols.size();
538 lto_symbol_attributes LTOModule::getSymbolAttributes(uint32_t index) {
540 if (index < _symbols.size())
541 return _symbols[index].attributes;
543 return lto_symbol_attributes(0);
546 const char *LTOModule::getSymbolName(uint32_t index) {
548 if (index < _symbols.size())
549 return _symbols[index].name;