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 
17 #include "UpdateCompilerUsed.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/Analysis/Passes.h"
21 #include "llvm/Analysis/TargetLibraryInfo.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/Bitcode/ReaderWriter.h"
24 #include "llvm/CodeGen/ParallelCG.h"
25 #include "llvm/CodeGen/RuntimeLibcalls.h"
26 #include "llvm/Config/config.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/DiagnosticPrinter.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/LegacyPassManager.h"
34 #include "llvm/IR/Mangler.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/InitializePasses.h"
38 #include "llvm/LTO/LTOModule.h"
39 #include "llvm/Linker/Linker.h"
40 #include "llvm/MC/MCAsmInfo.h"
41 #include "llvm/MC/MCContext.h"
42 #include "llvm/MC/SubtargetFeature.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/FileSystem.h"
45 #include "llvm/Support/Host.h"
46 #include "llvm/Support/MemoryBuffer.h"
47 #include "llvm/Support/Signals.h"
48 #include "llvm/Support/TargetRegistry.h"
49 #include "llvm/Support/TargetSelect.h"
50 #include "llvm/Support/ToolOutputFile.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Target/TargetLowering.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "llvm/Target/TargetRegisterInfo.h"
55 #include "llvm/Target/TargetSubtargetInfo.h"
56 #include "llvm/Transforms/IPO.h"
57 #include "llvm/Transforms/IPO/Internalize.h"
58 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
59 #include "llvm/Transforms/ObjCARC.h"
60 #include <system_error>
61 using namespace llvm;
62 
63 const char* LTOCodeGenerator::getVersionString() {
64 #ifdef LLVM_VERSION_INFO
65   return PACKAGE_NAME " version " PACKAGE_VERSION ", " LLVM_VERSION_INFO;
66 #else
67   return PACKAGE_NAME " version " PACKAGE_VERSION;
68 #endif
69 }
70 
71 namespace llvm {
72 cl::opt<bool> LTODiscardValueNames(
73     "lto-discard-value-names",
74     cl::desc("Strip names from Value during LTO (other than GlobalValue)."),
75 #ifdef NDEBUG
76     cl::init(true),
77 #else
78     cl::init(false),
79 #endif
80     cl::Hidden);
81 }
82 
83 LTOCodeGenerator::LTOCodeGenerator(LLVMContext &Context)
84     : Context(Context), MergedModule(new Module("ld-temp.o", Context)),
85       TheLinker(new Linker(*MergedModule)) {
86   Context.setDiscardValueNames(LTODiscardValueNames);
87   initializeLTOPasses();
88 }
89 
90 LTOCodeGenerator::~LTOCodeGenerator() {}
91 
92 // Initialize LTO passes. Please keep this function in sync with
93 // PassManagerBuilder::populateLTOPassManager(), and make sure all LTO
94 // passes are initialized.
95 void LTOCodeGenerator::initializeLTOPasses() {
96   PassRegistry &R = *PassRegistry::getPassRegistry();
97 
98   initializeInternalizePassPass(R);
99   initializeIPSCCPPass(R);
100   initializeGlobalOptPass(R);
101   initializeConstantMergePass(R);
102   initializeDAHPass(R);
103   initializeInstructionCombiningPassPass(R);
104   initializeSimpleInlinerPass(R);
105   initializePruneEHPass(R);
106   initializeGlobalDCEPass(R);
107   initializeArgPromotionPass(R);
108   initializeJumpThreadingPass(R);
109   initializeSROALegacyPassPass(R);
110   initializeSROA_DTPass(R);
111   initializeSROA_SSAUpPass(R);
112   initializePostOrderFunctionAttrsLegacyPassPass(R);
113   initializeReversePostOrderFunctionAttrsPass(R);
114   initializeGlobalsAAWrapperPassPass(R);
115   initializeLICMPass(R);
116   initializeMergedLoadStoreMotionPass(R);
117   initializeGVNLegacyPassPass(R);
118   initializeMemCpyOptPass(R);
119   initializeDCEPass(R);
120   initializeCFGSimplifyPassPass(R);
121 }
122 
123 bool LTOCodeGenerator::addModule(LTOModule *Mod) {
124   assert(&Mod->getModule().getContext() == &Context &&
125          "Expected module in same context");
126 
127   bool ret = TheLinker->linkInModule(Mod->takeModule());
128 
129   const std::vector<const char *> &undefs = Mod->getAsmUndefinedRefs();
130   for (int i = 0, e = undefs.size(); i != e; ++i)
131     AsmUndefinedRefs[undefs[i]] = 1;
132 
133   return !ret;
134 }
135 
136 void LTOCodeGenerator::setModule(std::unique_ptr<LTOModule> Mod) {
137   assert(&Mod->getModule().getContext() == &Context &&
138          "Expected module in same context");
139 
140   AsmUndefinedRefs.clear();
141 
142   MergedModule = Mod->takeModule();
143   TheLinker = make_unique<Linker>(*MergedModule);
144 
145   const std::vector<const char*> &Undefs = Mod->getAsmUndefinedRefs();
146   for (int I = 0, E = Undefs.size(); I != E; ++I)
147     AsmUndefinedRefs[Undefs[I]] = 1;
148 }
149 
150 void LTOCodeGenerator::setTargetOptions(TargetOptions Options) {
151   this->Options = Options;
152 }
153 
154 void LTOCodeGenerator::setDebugInfo(lto_debug_model Debug) {
155   switch (Debug) {
156   case LTO_DEBUG_MODEL_NONE:
157     EmitDwarfDebugInfo = false;
158     return;
159 
160   case LTO_DEBUG_MODEL_DWARF:
161     EmitDwarfDebugInfo = true;
162     return;
163   }
164   llvm_unreachable("Unknown debug format!");
165 }
166 
167 void LTOCodeGenerator::setOptLevel(unsigned Level) {
168   OptLevel = Level;
169   switch (OptLevel) {
170   case 0:
171     CGOptLevel = CodeGenOpt::None;
172     break;
173   case 1:
174     CGOptLevel = CodeGenOpt::Less;
175     break;
176   case 2:
177     CGOptLevel = CodeGenOpt::Default;
178     break;
179   case 3:
180     CGOptLevel = CodeGenOpt::Aggressive;
181     break;
182   }
183 }
184 
185 bool LTOCodeGenerator::writeMergedModules(const char *Path) {
186   if (!determineTarget())
187     return false;
188 
189   // mark which symbols can not be internalized
190   applyScopeRestrictions();
191 
192   // create output file
193   std::error_code EC;
194   tool_output_file Out(Path, EC, sys::fs::F_None);
195   if (EC) {
196     std::string ErrMsg = "could not open bitcode file for writing: ";
197     ErrMsg += Path;
198     emitError(ErrMsg);
199     return false;
200   }
201 
202   // write bitcode to it
203   WriteBitcodeToFile(MergedModule.get(), Out.os(), ShouldEmbedUselists);
204   Out.os().close();
205 
206   if (Out.os().has_error()) {
207     std::string ErrMsg = "could not write bitcode file: ";
208     ErrMsg += Path;
209     emitError(ErrMsg);
210     Out.os().clear_error();
211     return false;
212   }
213 
214   Out.keep();
215   return true;
216 }
217 
218 bool LTOCodeGenerator::compileOptimizedToFile(const char **Name) {
219   // make unique temp output file to put generated code
220   SmallString<128> Filename;
221   int FD;
222 
223   const char *Extension =
224       (FileType == TargetMachine::CGFT_AssemblyFile ? "s" : "o");
225 
226   std::error_code EC =
227       sys::fs::createTemporaryFile("lto-llvm", Extension, FD, Filename);
228   if (EC) {
229     emitError(EC.message());
230     return false;
231   }
232 
233   // generate object file
234   tool_output_file objFile(Filename.c_str(), FD);
235 
236   bool genResult = compileOptimized(&objFile.os());
237   objFile.os().close();
238   if (objFile.os().has_error()) {
239     objFile.os().clear_error();
240     sys::fs::remove(Twine(Filename));
241     return false;
242   }
243 
244   objFile.keep();
245   if (!genResult) {
246     sys::fs::remove(Twine(Filename));
247     return false;
248   }
249 
250   NativeObjectPath = Filename.c_str();
251   *Name = NativeObjectPath.c_str();
252   return true;
253 }
254 
255 std::unique_ptr<MemoryBuffer>
256 LTOCodeGenerator::compileOptimized() {
257   const char *name;
258   if (!compileOptimizedToFile(&name))
259     return nullptr;
260 
261   // read .o file into memory buffer
262   ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
263       MemoryBuffer::getFile(name, -1, false);
264   if (std::error_code EC = BufferOrErr.getError()) {
265     emitError(EC.message());
266     sys::fs::remove(NativeObjectPath);
267     return nullptr;
268   }
269 
270   // remove temp files
271   sys::fs::remove(NativeObjectPath);
272 
273   return std::move(*BufferOrErr);
274 }
275 
276 bool LTOCodeGenerator::compile_to_file(const char **Name, bool DisableVerify,
277                                        bool DisableInline,
278                                        bool DisableGVNLoadPRE,
279                                        bool DisableVectorization) {
280   if (!optimize(DisableVerify, DisableInline, DisableGVNLoadPRE,
281                 DisableVectorization))
282     return false;
283 
284   return compileOptimizedToFile(Name);
285 }
286 
287 std::unique_ptr<MemoryBuffer>
288 LTOCodeGenerator::compile(bool DisableVerify, bool DisableInline,
289                           bool DisableGVNLoadPRE, bool DisableVectorization) {
290   if (!optimize(DisableVerify, DisableInline, DisableGVNLoadPRE,
291                 DisableVectorization))
292     return nullptr;
293 
294   return compileOptimized();
295 }
296 
297 bool LTOCodeGenerator::determineTarget() {
298   if (TargetMach)
299     return true;
300 
301   std::string TripleStr = MergedModule->getTargetTriple();
302   if (TripleStr.empty()) {
303     TripleStr = sys::getDefaultTargetTriple();
304     MergedModule->setTargetTriple(TripleStr);
305   }
306   llvm::Triple Triple(TripleStr);
307 
308   // create target machine from info for merged modules
309   std::string ErrMsg;
310   const Target *march = TargetRegistry::lookupTarget(TripleStr, ErrMsg);
311   if (!march) {
312     emitError(ErrMsg);
313     return false;
314   }
315 
316   // Construct LTOModule, hand over ownership of module and target. Use MAttr as
317   // the default set of features.
318   SubtargetFeatures Features(MAttr);
319   Features.getDefaultSubtargetFeatures(Triple);
320   FeatureStr = Features.getString();
321   // Set a default CPU for Darwin triples.
322   if (MCpu.empty() && Triple.isOSDarwin()) {
323     if (Triple.getArch() == llvm::Triple::x86_64)
324       MCpu = "core2";
325     else if (Triple.getArch() == llvm::Triple::x86)
326       MCpu = "yonah";
327     else if (Triple.getArch() == llvm::Triple::aarch64)
328       MCpu = "cyclone";
329   }
330 
331   TargetMach.reset(march->createTargetMachine(TripleStr, MCpu, FeatureStr,
332                                               Options, RelocModel,
333                                               CodeModel::Default, CGOptLevel));
334   return true;
335 }
336 
337 void LTOCodeGenerator::applyScopeRestrictions() {
338   if (ScopeRestrictionsDone || !ShouldInternalize)
339     return;
340 
341   if (ShouldRestoreGlobalsLinkage) {
342     // Record the linkage type of non-local symbols so they can be restored
343     // prior
344     // to module splitting.
345     auto RecordLinkage = [&](const GlobalValue &GV) {
346       if (!GV.hasAvailableExternallyLinkage() && !GV.hasLocalLinkage() &&
347           GV.hasName())
348         ExternalSymbols.insert(std::make_pair(GV.getName(), GV.getLinkage()));
349     };
350     for (auto &GV : *MergedModule)
351       RecordLinkage(GV);
352     for (auto &GV : MergedModule->globals())
353       RecordLinkage(GV);
354     for (auto &GV : MergedModule->aliases())
355       RecordLinkage(GV);
356   }
357 
358   // Update the llvm.compiler_used globals to force preserving libcalls and
359   // symbols referenced from asm
360   UpdateCompilerUsed(*MergedModule, *TargetMach, AsmUndefinedRefs);
361 
362   // Declare a callback for the internalize pass that will ask for every
363   // candidate GlobalValue if it can be internalized or not.
364   Mangler Mangler;
365   SmallString<64> MangledName;
366   auto MustPreserveGV = [&](const GlobalValue &GV) -> bool {
367     // Need to mangle the GV as the "MustPreserveSymbols" StringSet is filled
368     // with the linker supplied name, which on Darwin includes a leading
369     // underscore.
370     MangledName.clear();
371     MangledName.reserve(GV.getName().size() + 1);
372     Mangler::getNameWithPrefix(MangledName, GV.getName(),
373                                MergedModule->getDataLayout());
374     return MustPreserveSymbols.count(MangledName);
375   };
376 
377   internalizeModule(*MergedModule, MustPreserveGV);
378 
379   ScopeRestrictionsDone = true;
380 }
381 
382 /// Restore original linkage for symbols that may have been internalized
383 void LTOCodeGenerator::restoreLinkageForExternals() {
384   if (!ShouldInternalize || !ShouldRestoreGlobalsLinkage)
385     return;
386 
387   assert(ScopeRestrictionsDone &&
388          "Cannot externalize without internalization!");
389 
390   if (ExternalSymbols.empty())
391     return;
392 
393   auto externalize = [this](GlobalValue &GV) {
394     if (!GV.hasLocalLinkage() || !GV.hasName())
395       return;
396 
397     auto I = ExternalSymbols.find(GV.getName());
398     if (I == ExternalSymbols.end())
399       return;
400 
401     GV.setLinkage(I->second);
402   };
403 
404   std::for_each(MergedModule->begin(), MergedModule->end(), externalize);
405   std::for_each(MergedModule->global_begin(), MergedModule->global_end(),
406                 externalize);
407   std::for_each(MergedModule->alias_begin(), MergedModule->alias_end(),
408                 externalize);
409 }
410 
411 /// Optimize merged modules using various IPO passes
412 bool LTOCodeGenerator::optimize(bool DisableVerify, bool DisableInline,
413                                 bool DisableGVNLoadPRE,
414                                 bool DisableVectorization) {
415   if (!this->determineTarget())
416     return false;
417 
418   // We always run the verifier once on the merged module, the `DisableVerify`
419   // parameter only applies to subsequent verify.
420   if (verifyModule(*MergedModule, &dbgs()))
421     report_fatal_error("Broken module found, compilation aborted!");
422 
423   // Mark which symbols can not be internalized
424   this->applyScopeRestrictions();
425 
426   // Instantiate the pass manager to organize the passes.
427   legacy::PassManager passes;
428 
429   // Add an appropriate DataLayout instance for this module...
430   MergedModule->setDataLayout(TargetMach->createDataLayout());
431 
432   passes.add(
433       createTargetTransformInfoWrapperPass(TargetMach->getTargetIRAnalysis()));
434 
435   Triple TargetTriple(TargetMach->getTargetTriple());
436   PassManagerBuilder PMB;
437   PMB.DisableGVNLoadPRE = DisableGVNLoadPRE;
438   PMB.LoopVectorize = !DisableVectorization;
439   PMB.SLPVectorize = !DisableVectorization;
440   if (!DisableInline)
441     PMB.Inliner = createFunctionInliningPass();
442   PMB.LibraryInfo = new TargetLibraryInfoImpl(TargetTriple);
443   PMB.OptLevel = OptLevel;
444   PMB.VerifyInput = !DisableVerify;
445   PMB.VerifyOutput = !DisableVerify;
446 
447   PMB.populateLTOPassManager(passes);
448 
449   // Run our queue of passes all at once now, efficiently.
450   passes.run(*MergedModule);
451 
452   return true;
453 }
454 
455 bool LTOCodeGenerator::compileOptimized(ArrayRef<raw_pwrite_stream *> Out) {
456   if (!this->determineTarget())
457     return false;
458 
459   legacy::PassManager preCodeGenPasses;
460 
461   // If the bitcode files contain ARC code and were compiled with optimization,
462   // the ObjCARCContractPass must be run, so do it unconditionally here.
463   preCodeGenPasses.add(createObjCARCContractPass());
464   preCodeGenPasses.run(*MergedModule);
465 
466   // Re-externalize globals that may have been internalized to increase scope
467   // for splitting
468   restoreLinkageForExternals();
469 
470   // Do code generation. We need to preserve the module in case the client calls
471   // writeMergedModules() after compilation, but we only need to allow this at
472   // parallelism level 1. This is achieved by having splitCodeGen return the
473   // original module at parallelism level 1 which we then assign back to
474   // MergedModule.
475   MergedModule = splitCodeGen(
476       std::move(MergedModule), Out, {}, MCpu, FeatureStr, Options, RelocModel,
477       CodeModel::Default, CGOptLevel, FileType, ShouldRestoreGlobalsLinkage);
478 
479   // If statistics were requested, print them out after codegen.
480   if (llvm::AreStatisticsEnabled())
481     llvm::PrintStatistics();
482 
483   return true;
484 }
485 
486 /// setCodeGenDebugOptions - Set codegen debugging options to aid in debugging
487 /// LTO problems.
488 void LTOCodeGenerator::setCodeGenDebugOptions(const char *Options) {
489   for (std::pair<StringRef, StringRef> o = getToken(Options); !o.first.empty();
490        o = getToken(o.second))
491     CodegenOptions.push_back(o.first);
492 }
493 
494 void LTOCodeGenerator::parseCodeGenDebugOptions() {
495   // if options were requested, set them
496   if (!CodegenOptions.empty()) {
497     // ParseCommandLineOptions() expects argv[0] to be program name.
498     std::vector<const char *> CodegenArgv(1, "libLLVMLTO");
499     for (std::string &Arg : CodegenOptions)
500       CodegenArgv.push_back(Arg.c_str());
501     cl::ParseCommandLineOptions(CodegenArgv.size(), CodegenArgv.data());
502   }
503 }
504 
505 void LTOCodeGenerator::DiagnosticHandler(const DiagnosticInfo &DI,
506                                          void *Context) {
507   ((LTOCodeGenerator *)Context)->DiagnosticHandler2(DI);
508 }
509 
510 void LTOCodeGenerator::DiagnosticHandler2(const DiagnosticInfo &DI) {
511   // Map the LLVM internal diagnostic severity to the LTO diagnostic severity.
512   lto_codegen_diagnostic_severity_t Severity;
513   switch (DI.getSeverity()) {
514   case DS_Error:
515     Severity = LTO_DS_ERROR;
516     break;
517   case DS_Warning:
518     Severity = LTO_DS_WARNING;
519     break;
520   case DS_Remark:
521     Severity = LTO_DS_REMARK;
522     break;
523   case DS_Note:
524     Severity = LTO_DS_NOTE;
525     break;
526   }
527   // Create the string that will be reported to the external diagnostic handler.
528   std::string MsgStorage;
529   raw_string_ostream Stream(MsgStorage);
530   DiagnosticPrinterRawOStream DP(Stream);
531   DI.print(DP);
532   Stream.flush();
533 
534   // If this method has been called it means someone has set up an external
535   // diagnostic handler. Assert on that.
536   assert(DiagHandler && "Invalid diagnostic handler");
537   (*DiagHandler)(Severity, MsgStorage.c_str(), DiagContext);
538 }
539 
540 void
541 LTOCodeGenerator::setDiagnosticHandler(lto_diagnostic_handler_t DiagHandler,
542                                        void *Ctxt) {
543   this->DiagHandler = DiagHandler;
544   this->DiagContext = Ctxt;
545   if (!DiagHandler)
546     return Context.setDiagnosticHandler(nullptr, nullptr);
547   // Register the LTOCodeGenerator stub in the LLVMContext to forward the
548   // diagnostic to the external DiagHandler.
549   Context.setDiagnosticHandler(LTOCodeGenerator::DiagnosticHandler, this,
550                                /* RespectFilters */ true);
551 }
552 
553 namespace {
554 class LTODiagnosticInfo : public DiagnosticInfo {
555   const Twine &Msg;
556 public:
557   LTODiagnosticInfo(const Twine &DiagMsg, DiagnosticSeverity Severity=DS_Error)
558       : DiagnosticInfo(DK_Linker, Severity), Msg(DiagMsg) {}
559   void print(DiagnosticPrinter &DP) const override { DP << Msg; }
560 };
561 }
562 
563 void LTOCodeGenerator::emitError(const std::string &ErrMsg) {
564   if (DiagHandler)
565     (*DiagHandler)(LTO_DS_ERROR, ErrMsg.c_str(), DiagContext);
566   else
567     Context.diagnose(LTODiagnosticInfo(ErrMsg));
568 }
569