1//===- FunctionImport.cpp - ThinLTO Summary-based Function Import ---------===//
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 Function import based on summaries.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/IPO/FunctionImport.h"
15
16#include "llvm/ADT/StringSet.h"
17#include "llvm/IR/AutoUpgrade.h"
18#include "llvm/IR/DiagnosticPrinter.h"
19#include "llvm/IR/IntrinsicInst.h"
20#include "llvm/IR/Module.h"
21#include "llvm/IRReader/IRReader.h"
22#include "llvm/Linker/Linker.h"
23#include "llvm/Object/FunctionIndexObjectFile.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/SourceMgr.h"
27
28#include <map>
29
30using namespace llvm;
31
32#define DEBUG_TYPE "function-import"
33
34/// Limit on instruction count of imported functions.
35static cl::opt<unsigned> ImportInstrLimit(
36    "import-instr-limit", cl::init(100), cl::Hidden, cl::value_desc("N"),
37    cl::desc("Only import functions with less than N instructions"));
38
39// Load lazily a module from \p FileName in \p Context.
40static std::unique_ptr<Module> loadFile(const std::string &FileName,
41                                        LLVMContext &Context) {
42  SMDiagnostic Err;
43  DEBUG(dbgs() << "Loading '" << FileName << "'\n");
44  std::unique_ptr<Module> Result = getLazyIRFileModule(FileName, Err, Context);
45  if (!Result) {
46    Err.print("function-import", errs());
47    return nullptr;
48  }
49
50  Result->materializeMetadata();
51  UpgradeDebugInfo(*Result);
52
53  return Result;
54}
55
56namespace {
57/// Helper to load on demand a Module from file and cache it for subsequent
58/// queries. It can be used with the FunctionImporter.
59class ModuleLazyLoaderCache {
60  /// Cache of lazily loaded module for import.
61  StringMap<std::unique_ptr<Module>> ModuleMap;
62
63  /// Retrieve a Module from the cache or lazily load it on demand.
64  std::function<std::unique_ptr<Module>(StringRef FileName)> createLazyModule;
65
66public:
67  /// Create the loader, Module will be initialized in \p Context.
68  ModuleLazyLoaderCache(std::function<
69      std::unique_ptr<Module>(StringRef FileName)> createLazyModule)
70      : createLazyModule(createLazyModule) {}
71
72  /// Retrieve a Module from the cache or lazily load it on demand.
73  Module &operator()(StringRef FileName);
74
75  std::unique_ptr<Module> takeModule(StringRef FileName) {
76    auto I = ModuleMap.find(FileName);
77    assert(I != ModuleMap.end());
78    std::unique_ptr<Module> Ret = std::move(I->second);
79    ModuleMap.erase(I);
80    return Ret;
81  }
82};
83
84// Get a Module for \p FileName from the cache, or load it lazily.
85Module &ModuleLazyLoaderCache::operator()(StringRef Identifier) {
86  auto &Module = ModuleMap[Identifier];
87  if (!Module)
88    Module = createLazyModule(Identifier);
89  return *Module;
90}
91} // anonymous namespace
92
93/// Walk through the instructions in \p F looking for external
94/// calls not already in the \p CalledFunctions set. If any are
95/// found they are added to the \p Worklist for importing.
96static void findExternalCalls(const Module &DestModule, Function &F,
97                              const FunctionInfoIndex &Index,
98                              StringSet<> &CalledFunctions,
99                              SmallVector<StringRef, 64> &Worklist) {
100  // We need to suffix internal function calls imported from other modules,
101  // prepare the suffix ahead of time.
102  std::string Suffix;
103  if (F.getParent() != &DestModule)
104    Suffix =
105        (Twine(".llvm.") +
106         Twine(Index.getModuleId(F.getParent()->getModuleIdentifier()))).str();
107
108  for (auto &BB : F) {
109    for (auto &I : BB) {
110      if (isa<CallInst>(I)) {
111        auto CalledFunction = cast<CallInst>(I).getCalledFunction();
112        // Insert any new external calls that have not already been
113        // added to set/worklist.
114        if (!CalledFunction || !CalledFunction->hasName())
115          continue;
116        // Ignore intrinsics early
117        if (CalledFunction->isIntrinsic()) {
118          assert(CalledFunction->getIntrinsicID() != 0);
119          continue;
120        }
121        auto ImportedName = CalledFunction->getName();
122        auto Renamed = (ImportedName + Suffix).str();
123        // Rename internal functions
124        if (CalledFunction->hasInternalLinkage()) {
125          ImportedName = Renamed;
126        }
127        auto It = CalledFunctions.insert(ImportedName);
128        if (!It.second) {
129          // This is a call to a function we already considered, skip.
130          continue;
131        }
132        // Ignore functions already present in the destination module
133        auto *SrcGV = DestModule.getNamedValue(ImportedName);
134        if (SrcGV) {
135          assert(isa<Function>(SrcGV) && "Name collision during import");
136          if (!cast<Function>(SrcGV)->isDeclaration()) {
137            DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Ignoring "
138                         << ImportedName << " already in DestinationModule\n");
139            continue;
140          }
141        }
142
143        Worklist.push_back(It.first->getKey());
144        DEBUG(dbgs() << DestModule.getModuleIdentifier()
145                     << ": Adding callee for : " << ImportedName << " : "
146                     << F.getName() << "\n");
147      }
148    }
149  }
150}
151
152// Helper function: given a worklist and an index, will process all the worklist
153// and decide what to import based on the summary information.
154//
155// Nothing is actually imported, functions are materialized in their source
156// module and analyzed there.
157//
158// \p ModuleToFunctionsToImportMap is filled with the set of Function to import
159// per Module.
160static void GetImportList(Module &DestModule,
161                          SmallVector<StringRef, 64> &Worklist,
162                          StringSet<> &CalledFunctions,
163                          std::map<StringRef, DenseSet<const GlobalValue *>>
164                              &ModuleToFunctionsToImportMap,
165                          const FunctionInfoIndex &Index,
166                          ModuleLazyLoaderCache &ModuleLoaderCache) {
167  while (!Worklist.empty()) {
168    auto CalledFunctionName = Worklist.pop_back_val();
169    DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Process import for "
170                 << CalledFunctionName << "\n");
171
172    // Try to get a summary for this function call.
173    auto InfoList = Index.findFunctionInfoList(CalledFunctionName);
174    if (InfoList == Index.end()) {
175      DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": No summary for "
176                   << CalledFunctionName << " Ignoring.\n");
177      continue;
178    }
179    assert(!InfoList->second.empty() && "No summary, error at import?");
180
181    // Comdat can have multiple entries, FIXME: what do we do with them?
182    auto &Info = InfoList->second[0];
183    assert(Info && "Nullptr in list, error importing summaries?\n");
184
185    auto *Summary = Info->functionSummary();
186    if (!Summary) {
187      // FIXME: in case we are lazyloading summaries, we can do it now.
188      DEBUG(dbgs() << DestModule.getModuleIdentifier()
189                   << ": Missing summary for  " << CalledFunctionName
190                   << ", error at import?\n");
191      llvm_unreachable("Missing summary");
192    }
193
194    if (Summary->instCount() > ImportInstrLimit) {
195      DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Skip import of "
196                   << CalledFunctionName << " with " << Summary->instCount()
197                   << " instructions (limit " << ImportInstrLimit << ")\n");
198      continue;
199    }
200
201    // Get the module path from the summary.
202    auto ModuleIdentifier = Summary->modulePath();
203    DEBUG(dbgs() << DestModule.getModuleIdentifier() << ": Importing "
204                 << CalledFunctionName << " from " << ModuleIdentifier << "\n");
205
206    auto &SrcModule = ModuleLoaderCache(ModuleIdentifier);
207
208    // The function that we will import!
209    GlobalValue *SGV = SrcModule.getNamedValue(CalledFunctionName);
210
211    if (!SGV) {
212      // The destination module is referencing function using their renamed name
213      // when importing a function that was originally local in the source
214      // module. The source module we have might not have been renamed so we try
215      // to remove the suffix added during the renaming to recover the original
216      // name in the source module.
217      std::pair<StringRef, StringRef> Split =
218          CalledFunctionName.split(".llvm.");
219      SGV = SrcModule.getNamedValue(Split.first);
220      assert(SGV && "Can't find function to import in source module");
221    }
222    if (!SGV) {
223      report_fatal_error(Twine("Can't load function '") + CalledFunctionName +
224                         "' in Module '" + SrcModule.getModuleIdentifier() +
225                         "', error in the summary?\n");
226    }
227
228    Function *F = dyn_cast<Function>(SGV);
229    if (!F && isa<GlobalAlias>(SGV)) {
230      auto *SGA = dyn_cast<GlobalAlias>(SGV);
231      F = dyn_cast<Function>(SGA->getBaseObject());
232      CalledFunctionName = F->getName();
233    }
234    assert(F && "Imported Function is ... not a Function");
235
236    // We cannot import weak_any functions/aliases without possibly affecting
237    // the order they are seen and selected by the linker, changing program
238    // semantics.
239    if (SGV->hasWeakAnyLinkage()) {
240      DEBUG(dbgs() << DestModule.getModuleIdentifier()
241                   << ": Ignoring import request for weak-any "
242                   << (isa<Function>(SGV) ? "function " : "alias ")
243                   << CalledFunctionName << " from "
244                   << SrcModule.getModuleIdentifier() << "\n");
245      continue;
246    }
247
248    // Add the function to the import list
249    auto &Entry = ModuleToFunctionsToImportMap[SrcModule.getModuleIdentifier()];
250    Entry.insert(F);
251
252    // Process the newly imported functions and add callees to the worklist.
253    F->materialize();
254    findExternalCalls(DestModule, *F, Index, CalledFunctions, Worklist);
255  }
256}
257
258// Automatically import functions in Module \p DestModule based on the summaries
259// index.
260//
261// The current implementation imports every called functions that exists in the
262// summaries index.
263bool FunctionImporter::importFunctions(Module &DestModule) {
264  DEBUG(dbgs() << "Starting import for Module "
265               << DestModule.getModuleIdentifier() << "\n");
266  unsigned ImportedCount = 0;
267
268  /// First step is collecting the called external functions.
269  StringSet<> CalledFunctions;
270  SmallVector<StringRef, 64> Worklist;
271  for (auto &F : DestModule) {
272    if (F.isDeclaration() || F.hasFnAttribute(Attribute::OptimizeNone))
273      continue;
274    findExternalCalls(DestModule, F, Index, CalledFunctions, Worklist);
275  }
276  if (Worklist.empty())
277    return false;
278
279  /// Second step: for every call to an external function, try to import it.
280
281  // Linker that will be used for importing function
282  Linker TheLinker(DestModule);
283
284  // Map of Module -> List of Function to import from the Module
285  std::map<StringRef, DenseSet<const GlobalValue *>>
286      ModuleToFunctionsToImportMap;
287
288  // Analyze the summaries and get the list of functions to import by
289  // populating ModuleToFunctionsToImportMap
290  ModuleLazyLoaderCache ModuleLoaderCache(ModuleLoader);
291  GetImportList(DestModule, Worklist, CalledFunctions,
292                ModuleToFunctionsToImportMap, Index, ModuleLoaderCache);
293  assert(Worklist.empty() && "Worklist hasn't been flushed in GetImportList");
294
295  StringMap<std::unique_ptr<DenseMap<unsigned, MDNode *>>>
296      ModuleToTempMDValsMap;
297
298  // Do the actual import of functions now, one Module at a time
299  for (auto &FunctionsToImportPerModule : ModuleToFunctionsToImportMap) {
300    // Get the module for the import
301    auto &FunctionsToImport = FunctionsToImportPerModule.second;
302    std::unique_ptr<Module> SrcModule =
303        ModuleLoaderCache.takeModule(FunctionsToImportPerModule.first);
304    assert(&DestModule.getContext() == &SrcModule->getContext() &&
305           "Context mismatch");
306
307    // Save the mapping of value ids to temporary metadata created when
308    // importing this function. If we have already imported from this module,
309    // add new temporary metadata to the existing mapping.
310    auto &TempMDVals = ModuleToTempMDValsMap[SrcModule->getModuleIdentifier()];
311    if (!TempMDVals)
312      TempMDVals = llvm::make_unique<DenseMap<unsigned, MDNode *>>();
313
314    // Link in the specified functions.
315    if (TheLinker.linkInModule(std::move(SrcModule), Linker::Flags::None,
316                               &Index, &FunctionsToImport, TempMDVals.get()))
317      report_fatal_error("Function Import: link error");
318
319    ImportedCount += FunctionsToImport.size();
320  }
321
322  // Now link in metadata for all modules from which we imported functions.
323  for (StringMapEntry<std::unique_ptr<DenseMap<unsigned, MDNode *>>> &SME :
324       ModuleToTempMDValsMap) {
325    // Load the specified source module.
326    auto &SrcModule = ModuleLoaderCache(SME.getKey());
327
328    // Link in all necessary metadata from this module.
329    if (TheLinker.linkInMetadata(SrcModule, SME.getValue().get()))
330      return false;
331  }
332
333  DEBUG(dbgs() << "Imported " << ImportedCount << " functions for Module "
334               << DestModule.getModuleIdentifier() << "\n");
335  return ImportedCount;
336}
337
338/// Summary file to use for function importing when using -function-import from
339/// the command line.
340static cl::opt<std::string>
341    SummaryFile("summary-file",
342                cl::desc("The summary file to use for function importing."));
343
344static void diagnosticHandler(const DiagnosticInfo &DI) {
345  raw_ostream &OS = errs();
346  DiagnosticPrinterRawOStream DP(OS);
347  DI.print(DP);
348  OS << '\n';
349}
350
351/// Parse the function index out of an IR file and return the function
352/// index object if found, or nullptr if not.
353static std::unique_ptr<FunctionInfoIndex>
354getFunctionIndexForFile(StringRef Path, std::string &Error,
355                        DiagnosticHandlerFunction DiagnosticHandler) {
356  std::unique_ptr<MemoryBuffer> Buffer;
357  ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
358      MemoryBuffer::getFile(Path);
359  if (std::error_code EC = BufferOrErr.getError()) {
360    Error = EC.message();
361    return nullptr;
362  }
363  Buffer = std::move(BufferOrErr.get());
364  ErrorOr<std::unique_ptr<object::FunctionIndexObjectFile>> ObjOrErr =
365      object::FunctionIndexObjectFile::create(Buffer->getMemBufferRef(),
366                                              DiagnosticHandler);
367  if (std::error_code EC = ObjOrErr.getError()) {
368    Error = EC.message();
369    return nullptr;
370  }
371  return (*ObjOrErr)->takeIndex();
372}
373
374/// Pass that performs cross-module function import provided a summary file.
375class FunctionImportPass : public ModulePass {
376  /// Optional function summary index to use for importing, otherwise
377  /// the summary-file option must be specified.
378  const FunctionInfoIndex *Index;
379
380public:
381  /// Pass identification, replacement for typeid
382  static char ID;
383
384  /// Specify pass name for debug output
385  const char *getPassName() const override {
386    return "Function Importing";
387  }
388
389  explicit FunctionImportPass(const FunctionInfoIndex *Index = nullptr)
390      : ModulePass(ID), Index(Index) {}
391
392  bool runOnModule(Module &M) override {
393    if (SummaryFile.empty() && !Index)
394      report_fatal_error("error: -function-import requires -summary-file or "
395                         "file from frontend\n");
396    std::unique_ptr<FunctionInfoIndex> IndexPtr;
397    if (!SummaryFile.empty()) {
398      if (Index)
399        report_fatal_error("error: -summary-file and index from frontend\n");
400      std::string Error;
401      IndexPtr = getFunctionIndexForFile(SummaryFile, Error, diagnosticHandler);
402      if (!IndexPtr) {
403        errs() << "Error loading file '" << SummaryFile << "': " << Error
404               << "\n";
405        return false;
406      }
407      Index = IndexPtr.get();
408    }
409
410    // Perform the import now.
411    auto ModuleLoader = [&M](StringRef Identifier) {
412      return loadFile(Identifier, M.getContext());
413    };
414    FunctionImporter Importer(*Index, ModuleLoader);
415    return Importer.importFunctions(M);
416
417    return false;
418  }
419};
420
421char FunctionImportPass::ID = 0;
422INITIALIZE_PASS_BEGIN(FunctionImportPass, "function-import",
423                      "Summary Based Function Import", false, false)
424INITIALIZE_PASS_END(FunctionImportPass, "function-import",
425                    "Summary Based Function Import", false, false)
426
427namespace llvm {
428Pass *createFunctionImportPass(const FunctionInfoIndex *Index = nullptr) {
429  return new FunctionImportPass(Index);
430}
431}
432