1 //===-LTOCodeGenerator.cpp - LLVM Link Time Optimizer ---------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/LTO/LTOCodeGenerator.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/Analysis/Passes.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/Bitcode/ReaderWriter.h"
21 #include "llvm/CodeGen/ParallelCG.h"
22 #include "llvm/CodeGen/RuntimeLibcalls.h"
23 #include "llvm/Config/config.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/DiagnosticInfo.h"
28 #include "llvm/IR/DiagnosticPrinter.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/IR/LegacyPassManager.h"
31 #include "llvm/IR/Mangler.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/LTO/LTOModule.h"
36 #include "llvm/Linker/Linker.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/SubtargetFeature.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/FileSystem.h"
42 #include "llvm/Support/Host.h"
43 #include "llvm/Support/MemoryBuffer.h"
44 #include "llvm/Support/Signals.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Support/TargetSelect.h"
47 #include "llvm/Support/ToolOutputFile.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Target/TargetLowering.h"
50 #include "llvm/Target/TargetOptions.h"
51 #include "llvm/Target/TargetRegisterInfo.h"
52 #include "llvm/Target/TargetSubtargetInfo.h"
53 #include "llvm/Transforms/IPO.h"
54 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
55 #include "llvm/Transforms/ObjCARC.h"
56 #include <system_error>
57 using namespace llvm;
58 
59 const char* LTOCodeGenerator::getVersionString() {
60 #ifdef LLVM_VERSION_INFO
61   return PACKAGE_NAME " version " PACKAGE_VERSION ", " LLVM_VERSION_INFO;
62 #else
63   return PACKAGE_NAME " version " PACKAGE_VERSION;
64 #endif
65 }
66 
67 LTOCodeGenerator::LTOCodeGenerator(LLVMContext &Context)
68     : Context(Context), MergedModule(new Module("ld-temp.o", Context)),
69       IRLinker(new Linker(*MergedModule, [this](const DiagnosticInfo &DI) {
70         MergedModule->getContext().diagnose(DI);
71       })) {
72   initializeLTOPasses();
73 }
74 
75 LTOCodeGenerator::~LTOCodeGenerator() {}
76 
77 // Initialize LTO passes. Please keep this function in sync with
78 // PassManagerBuilder::populateLTOPassManager(), and make sure all LTO
79 // passes are initialized.
80 void LTOCodeGenerator::initializeLTOPasses() {
81   PassRegistry &R = *PassRegistry::getPassRegistry();
82 
83   initializeInternalizePassPass(R);
84   initializeIPSCCPPass(R);
85   initializeGlobalOptPass(R);
86   initializeConstantMergePass(R);
87   initializeDAHPass(R);
88   initializeInstructionCombiningPassPass(R);
89   initializeSimpleInlinerPass(R);
90   initializePruneEHPass(R);
91   initializeGlobalDCEPass(R);
92   initializeArgPromotionPass(R);
93   initializeJumpThreadingPass(R);
94   initializeSROALegacyPassPass(R);
95   initializeSROA_DTPass(R);
96   initializeSROA_SSAUpPass(R);
97   initializeFunctionAttrsPass(R);
98   initializeGlobalsAAWrapperPassPass(R);
99   initializeLICMPass(R);
100   initializeMergedLoadStoreMotionPass(R);
101   initializeGVNPass(R);
102   initializeMemCpyOptPass(R);
103   initializeDCEPass(R);
104   initializeCFGSimplifyPassPass(R);
105 }
106 
107 bool LTOCodeGenerator::addModule(LTOModule *Mod) {
108   assert(&Mod->getModule().getContext() == &Context &&
109          "Expected module in same context");
110 
111   bool ret = IRLinker->linkInModule(Mod->getModule());
112 
113   const std::vector<const char *> &undefs = Mod->getAsmUndefinedRefs();
114   for (int i = 0, e = undefs.size(); i != e; ++i)
115     AsmUndefinedRefs[undefs[i]] = 1;
116 
117   return !ret;
118 }
119 
120 void LTOCodeGenerator::setModule(std::unique_ptr<LTOModule> Mod) {
121   assert(&Mod->getModule().getContext() == &Context &&
122          "Expected module in same context");
123 
124   AsmUndefinedRefs.clear();
125 
126   MergedModule = Mod->takeModule();
127   IRLinker = llvm::make_unique<Linker>(*MergedModule,
128                                        IRLinker->getDiagnosticHandler());
129 
130   const std::vector<const char*> &Undefs = Mod->getAsmUndefinedRefs();
131   for (int I = 0, E = Undefs.size(); I != E; ++I)
132     AsmUndefinedRefs[Undefs[I]] = 1;
133 }
134 
135 void LTOCodeGenerator::setTargetOptions(TargetOptions Options) {
136   this->Options = Options;
137 }
138 
139 void LTOCodeGenerator::setDebugInfo(lto_debug_model Debug) {
140   switch (Debug) {
141   case LTO_DEBUG_MODEL_NONE:
142     EmitDwarfDebugInfo = false;
143     return;
144 
145   case LTO_DEBUG_MODEL_DWARF:
146     EmitDwarfDebugInfo = true;
147     return;
148   }
149   llvm_unreachable("Unknown debug format!");
150 }
151 
152 void LTOCodeGenerator::setOptLevel(unsigned Level) {
153   OptLevel = Level;
154   switch (OptLevel) {
155   case 0:
156     CGOptLevel = CodeGenOpt::None;
157     break;
158   case 1:
159     CGOptLevel = CodeGenOpt::Less;
160     break;
161   case 2:
162     CGOptLevel = CodeGenOpt::Default;
163     break;
164   case 3:
165     CGOptLevel = CodeGenOpt::Aggressive;
166     break;
167   }
168 }
169 
170 bool LTOCodeGenerator::writeMergedModules(const char *Path) {
171   if (!determineTarget())
172     return false;
173 
174   // mark which symbols can not be internalized
175   applyScopeRestrictions();
176 
177   // create output file
178   std::error_code EC;
179   tool_output_file Out(Path, EC, sys::fs::F_None);
180   if (EC) {
181     std::string ErrMsg = "could not open bitcode file for writing: ";
182     ErrMsg += Path;
183     emitError(ErrMsg);
184     return false;
185   }
186 
187   // write bitcode to it
188   WriteBitcodeToFile(MergedModule.get(), Out.os(), ShouldEmbedUselists);
189   Out.os().close();
190 
191   if (Out.os().has_error()) {
192     std::string ErrMsg = "could not write bitcode file: ";
193     ErrMsg += Path;
194     emitError(ErrMsg);
195     Out.os().clear_error();
196     return false;
197   }
198 
199   Out.keep();
200   return true;
201 }
202 
203 bool LTOCodeGenerator::compileOptimizedToFile(const char **Name) {
204   // make unique temp output file to put generated code
205   SmallString<128> Filename;
206   int FD;
207 
208   const char *Extension =
209       (FileType == TargetMachine::CGFT_AssemblyFile ? "s" : "o");
210 
211   std::error_code EC =
212       sys::fs::createTemporaryFile("lto-llvm", Extension, FD, Filename);
213   if (EC) {
214     emitError(EC.message());
215     return false;
216   }
217 
218   // generate object file
219   tool_output_file objFile(Filename.c_str(), FD);
220 
221   bool genResult = compileOptimized(&objFile.os());
222   objFile.os().close();
223   if (objFile.os().has_error()) {
224     objFile.os().clear_error();
225     sys::fs::remove(Twine(Filename));
226     return false;
227   }
228 
229   objFile.keep();
230   if (!genResult) {
231     sys::fs::remove(Twine(Filename));
232     return false;
233   }
234 
235   NativeObjectPath = Filename.c_str();
236   *Name = NativeObjectPath.c_str();
237   return true;
238 }
239 
240 std::unique_ptr<MemoryBuffer>
241 LTOCodeGenerator::compileOptimized() {
242   const char *name;
243   if (!compileOptimizedToFile(&name))
244     return nullptr;
245 
246   // read .o file into memory buffer
247   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
248       MemoryBuffer::getFile(name, -1, false);
249   if (std::error_code EC = BufferOrErr.getError()) {
250     emitError(EC.message());
251     sys::fs::remove(NativeObjectPath);
252     return nullptr;
253   }
254 
255   // remove temp files
256   sys::fs::remove(NativeObjectPath);
257 
258   return std::move(*BufferOrErr);
259 }
260 
261 bool LTOCodeGenerator::compile_to_file(const char **Name, bool DisableVerify,
262                                        bool DisableInline,
263                                        bool DisableGVNLoadPRE,
264                                        bool DisableVectorization) {
265   if (!optimize(DisableVerify, DisableInline, DisableGVNLoadPRE,
266                 DisableVectorization))
267     return false;
268 
269   return compileOptimizedToFile(Name);
270 }
271 
272 std::unique_ptr<MemoryBuffer>
273 LTOCodeGenerator::compile(bool DisableVerify, bool DisableInline,
274                           bool DisableGVNLoadPRE, bool DisableVectorization) {
275   if (!optimize(DisableVerify, DisableInline, DisableGVNLoadPRE,
276                 DisableVectorization))
277     return nullptr;
278 
279   return compileOptimized();
280 }
281 
282 bool LTOCodeGenerator::determineTarget() {
283   if (TargetMach)
284     return true;
285 
286   std::string TripleStr = MergedModule->getTargetTriple();
287   if (TripleStr.empty()) {
288     TripleStr = sys::getDefaultTargetTriple();
289     MergedModule->setTargetTriple(TripleStr);
290   }
291   llvm::Triple Triple(TripleStr);
292 
293   // create target machine from info for merged modules
294   std::string ErrMsg;
295   const Target *march = TargetRegistry::lookupTarget(TripleStr, ErrMsg);
296   if (!march) {
297     emitError(ErrMsg);
298     return false;
299   }
300 
301   // Construct LTOModule, hand over ownership of module and target. Use MAttr as
302   // the default set of features.
303   SubtargetFeatures Features(MAttr);
304   Features.getDefaultSubtargetFeatures(Triple);
305   FeatureStr = Features.getString();
306   // Set a default CPU for Darwin triples.
307   if (MCpu.empty() && Triple.isOSDarwin()) {
308     if (Triple.getArch() == llvm::Triple::x86_64)
309       MCpu = "core2";
310     else if (Triple.getArch() == llvm::Triple::x86)
311       MCpu = "yonah";
312     else if (Triple.getArch() == llvm::Triple::aarch64)
313       MCpu = "cyclone";
314   }
315 
316   TargetMach.reset(march->createTargetMachine(TripleStr, MCpu, FeatureStr,
317                                               Options, RelocModel,
318                                               CodeModel::Default, CGOptLevel));
319   return true;
320 }
321 
322 void LTOCodeGenerator::
323 applyRestriction(GlobalValue &GV,
324                  ArrayRef<StringRef> Libcalls,
325                  std::vector<const char*> &MustPreserveList,
326                  SmallPtrSetImpl<GlobalValue*> &AsmUsed,
327                  Mangler &Mangler) {
328   // There are no restrictions to apply to declarations.
329   if (GV.isDeclaration())
330     return;
331 
332   // There is nothing more restrictive than private linkage.
333   if (GV.hasPrivateLinkage())
334     return;
335 
336   SmallString<64> Buffer;
337   TargetMach->getNameWithPrefix(Buffer, &GV, Mangler);
338 
339   if (MustPreserveSymbols.count(Buffer))
340     MustPreserveList.push_back(GV.getName().data());
341   if (AsmUndefinedRefs.count(Buffer))
342     AsmUsed.insert(&GV);
343 
344   // Conservatively append user-supplied runtime library functions to
345   // llvm.compiler.used.  These could be internalized and deleted by
346   // optimizations like -globalopt, causing problems when later optimizations
347   // add new library calls (e.g., llvm.memset => memset and printf => puts).
348   // Leave it to the linker to remove any dead code (e.g. with -dead_strip).
349   if (isa<Function>(GV) &&
350       std::binary_search(Libcalls.begin(), Libcalls.end(), GV.getName()))
351     AsmUsed.insert(&GV);
352 }
353 
354 static void findUsedValues(GlobalVariable *LLVMUsed,
355                            SmallPtrSetImpl<GlobalValue*> &UsedValues) {
356   if (!LLVMUsed) return;
357 
358   ConstantArray *Inits = cast<ConstantArray>(LLVMUsed->getInitializer());
359   for (unsigned i = 0, e = Inits->getNumOperands(); i != e; ++i)
360     if (GlobalValue *GV =
361         dyn_cast<GlobalValue>(Inits->getOperand(i)->stripPointerCasts()))
362       UsedValues.insert(GV);
363 }
364 
365 // Collect names of runtime library functions. User-defined functions with the
366 // same names are added to llvm.compiler.used to prevent them from being
367 // deleted by optimizations.
368 static void accumulateAndSortLibcalls(std::vector<StringRef> &Libcalls,
369                                       const TargetLibraryInfo& TLI,
370                                       const Module &Mod,
371                                       const TargetMachine &TM) {
372   // TargetLibraryInfo has info on C runtime library calls on the current
373   // target.
374   for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs);
375        I != E; ++I) {
376     LibFunc::Func F = static_cast<LibFunc::Func>(I);
377     if (TLI.has(F))
378       Libcalls.push_back(TLI.getName(F));
379   }
380 
381   SmallPtrSet<const TargetLowering *, 1> TLSet;
382 
383   for (const Function &F : Mod) {
384     const TargetLowering *Lowering =
385         TM.getSubtargetImpl(F)->getTargetLowering();
386 
387     if (Lowering && TLSet.insert(Lowering).second)
388       // TargetLowering has info on library calls that CodeGen expects to be
389       // available, both from the C runtime and compiler-rt.
390       for (unsigned I = 0, E = static_cast<unsigned>(RTLIB::UNKNOWN_LIBCALL);
391            I != E; ++I)
392         if (const char *Name =
393                 Lowering->getLibcallName(static_cast<RTLIB::Libcall>(I)))
394           Libcalls.push_back(Name);
395   }
396 
397   array_pod_sort(Libcalls.begin(), Libcalls.end());
398   Libcalls.erase(std::unique(Libcalls.begin(), Libcalls.end()),
399                  Libcalls.end());
400 }
401 
402 void LTOCodeGenerator::applyScopeRestrictions() {
403   if (ScopeRestrictionsDone || !ShouldInternalize)
404     return;
405 
406   // Start off with a verification pass.
407   legacy::PassManager passes;
408   passes.add(createVerifierPass());
409 
410   // mark which symbols can not be internalized
411   Mangler Mangler;
412   std::vector<const char*> MustPreserveList;
413   SmallPtrSet<GlobalValue*, 8> AsmUsed;
414   std::vector<StringRef> Libcalls;
415   TargetLibraryInfoImpl TLII(Triple(TargetMach->getTargetTriple()));
416   TargetLibraryInfo TLI(TLII);
417 
418   accumulateAndSortLibcalls(Libcalls, TLI, *MergedModule, *TargetMach);
419 
420   for (Function &f : *MergedModule)
421     applyRestriction(f, Libcalls, MustPreserveList, AsmUsed, Mangler);
422   for (GlobalVariable &v : MergedModule->globals())
423     applyRestriction(v, Libcalls, MustPreserveList, AsmUsed, Mangler);
424   for (GlobalAlias &a : MergedModule->aliases())
425     applyRestriction(a, Libcalls, MustPreserveList, AsmUsed, Mangler);
426 
427   GlobalVariable *LLVMCompilerUsed =
428     MergedModule->getGlobalVariable("llvm.compiler.used");
429   findUsedValues(LLVMCompilerUsed, AsmUsed);
430   if (LLVMCompilerUsed)
431     LLVMCompilerUsed->eraseFromParent();
432 
433   if (!AsmUsed.empty()) {
434     llvm::Type *i8PTy = llvm::Type::getInt8PtrTy(Context);
435     std::vector<Constant*> asmUsed2;
436     for (auto *GV : AsmUsed) {
437       Constant *c = ConstantExpr::getBitCast(GV, i8PTy);
438       asmUsed2.push_back(c);
439     }
440 
441     llvm::ArrayType *ATy = llvm::ArrayType::get(i8PTy, asmUsed2.size());
442     LLVMCompilerUsed =
443       new llvm::GlobalVariable(*MergedModule, ATy, false,
444                                llvm::GlobalValue::AppendingLinkage,
445                                llvm::ConstantArray::get(ATy, asmUsed2),
446                                "llvm.compiler.used");
447 
448     LLVMCompilerUsed->setSection("llvm.metadata");
449   }
450 
451   passes.add(createInternalizePass(MustPreserveList));
452 
453   // apply scope restrictions
454   passes.run(*MergedModule);
455 
456   ScopeRestrictionsDone = true;
457 }
458 
459 /// Optimize merged modules using various IPO passes
460 bool LTOCodeGenerator::optimize(bool DisableVerify, bool DisableInline,
461                                 bool DisableGVNLoadPRE,
462                                 bool DisableVectorization) {
463   if (!this->determineTarget())
464     return false;
465 
466   // Mark which symbols can not be internalized
467   this->applyScopeRestrictions();
468 
469   // Instantiate the pass manager to organize the passes.
470   legacy::PassManager passes;
471 
472   // Add an appropriate DataLayout instance for this module...
473   MergedModule->setDataLayout(TargetMach->createDataLayout());
474 
475   passes.add(
476       createTargetTransformInfoWrapperPass(TargetMach->getTargetIRAnalysis()));
477 
478   Triple TargetTriple(TargetMach->getTargetTriple());
479   PassManagerBuilder PMB;
480   PMB.DisableGVNLoadPRE = DisableGVNLoadPRE;
481   PMB.LoopVectorize = !DisableVectorization;
482   PMB.SLPVectorize = !DisableVectorization;
483   if (!DisableInline)
484     PMB.Inliner = createFunctionInliningPass();
485   PMB.LibraryInfo = new TargetLibraryInfoImpl(TargetTriple);
486   PMB.OptLevel = OptLevel;
487   PMB.VerifyInput = !DisableVerify;
488   PMB.VerifyOutput = !DisableVerify;
489 
490   PMB.populateLTOPassManager(passes);
491 
492   // Run our queue of passes all at once now, efficiently.
493   passes.run(*MergedModule);
494 
495   return true;
496 }
497 
498 bool LTOCodeGenerator::compileOptimized(ArrayRef<raw_pwrite_stream *> Out) {
499   if (!this->determineTarget())
500     return false;
501 
502   legacy::PassManager preCodeGenPasses;
503 
504   // If the bitcode files contain ARC code and were compiled with optimization,
505   // the ObjCARCContractPass must be run, so do it unconditionally here.
506   preCodeGenPasses.add(createObjCARCContractPass());
507   preCodeGenPasses.run(*MergedModule);
508 
509   // Do code generation. We need to preserve the module in case the client calls
510   // writeMergedModules() after compilation, but we only need to allow this at
511   // parallelism level 1. This is achieved by having splitCodeGen return the
512   // original module at parallelism level 1 which we then assign back to
513   // MergedModule.
514   MergedModule =
515       splitCodeGen(std::move(MergedModule), Out, MCpu, FeatureStr, Options,
516                    RelocModel, CodeModel::Default, CGOptLevel, FileType);
517 
518   return true;
519 }
520 
521 /// setCodeGenDebugOptions - Set codegen debugging options to aid in debugging
522 /// LTO problems.
523 void LTOCodeGenerator::setCodeGenDebugOptions(const char *Options) {
524   for (std::pair<StringRef, StringRef> o = getToken(Options); !o.first.empty();
525        o = getToken(o.second))
526     CodegenOptions.push_back(o.first);
527 }
528 
529 void LTOCodeGenerator::parseCodeGenDebugOptions() {
530   // if options were requested, set them
531   if (!CodegenOptions.empty()) {
532     // ParseCommandLineOptions() expects argv[0] to be program name.
533     std::vector<const char *> CodegenArgv(1, "libLLVMLTO");
534     for (std::string &Arg : CodegenOptions)
535       CodegenArgv.push_back(Arg.c_str());
536     cl::ParseCommandLineOptions(CodegenArgv.size(), CodegenArgv.data());
537   }
538 }
539 
540 void LTOCodeGenerator::DiagnosticHandler(const DiagnosticInfo &DI,
541                                          void *Context) {
542   ((LTOCodeGenerator *)Context)->DiagnosticHandler2(DI);
543 }
544 
545 void LTOCodeGenerator::DiagnosticHandler2(const DiagnosticInfo &DI) {
546   // Map the LLVM internal diagnostic severity to the LTO diagnostic severity.
547   lto_codegen_diagnostic_severity_t Severity;
548   switch (DI.getSeverity()) {
549   case DS_Error:
550     Severity = LTO_DS_ERROR;
551     break;
552   case DS_Warning:
553     Severity = LTO_DS_WARNING;
554     break;
555   case DS_Remark:
556     Severity = LTO_DS_REMARK;
557     break;
558   case DS_Note:
559     Severity = LTO_DS_NOTE;
560     break;
561   }
562   // Create the string that will be reported to the external diagnostic handler.
563   std::string MsgStorage;
564   raw_string_ostream Stream(MsgStorage);
565   DiagnosticPrinterRawOStream DP(Stream);
566   DI.print(DP);
567   Stream.flush();
568 
569   // If this method has been called it means someone has set up an external
570   // diagnostic handler. Assert on that.
571   assert(DiagHandler && "Invalid diagnostic handler");
572   (*DiagHandler)(Severity, MsgStorage.c_str(), DiagContext);
573 }
574 
575 void
576 LTOCodeGenerator::setDiagnosticHandler(lto_diagnostic_handler_t DiagHandler,
577                                        void *Ctxt) {
578   this->DiagHandler = DiagHandler;
579   this->DiagContext = Ctxt;
580   if (!DiagHandler)
581     return Context.setDiagnosticHandler(nullptr, nullptr);
582   // Register the LTOCodeGenerator stub in the LLVMContext to forward the
583   // diagnostic to the external DiagHandler.
584   Context.setDiagnosticHandler(LTOCodeGenerator::DiagnosticHandler, this,
585                                /* RespectFilters */ true);
586 }
587 
588 namespace {
589 class LTODiagnosticInfo : public DiagnosticInfo {
590   const Twine &Msg;
591 public:
592   LTODiagnosticInfo(const Twine &DiagMsg, DiagnosticSeverity Severity=DS_Error)
593       : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {}
594   void print(DiagnosticPrinter &DP) const override { DP << Msg; }
595 };
596 }
597 
598 void LTOCodeGenerator::emitError(const std::string &ErrMsg) {
599   if (DiagHandler)
600     (*DiagHandler)(LTO_DS_ERROR, ErrMsg.c_str(), DiagContext);
601   else
602     Context.diagnose(LTODiagnosticInfo(ErrMsg));
603 }
604