1//===- ExtractFunction.cpp - Extract a function from Program --------------===//
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 several methods that are used to extract functions,
11// loops, or portions of a module from the rest of the module.
12//
13//===----------------------------------------------------------------------===//
14
15#include "BugDriver.h"
16#include "llvm/IR/Constants.h"
17#include "llvm/IR/DataLayout.h"
18#include "llvm/IR/DerivedTypes.h"
19#include "llvm/IR/LLVMContext.h"
20#include "llvm/IR/LegacyPassManager.h"
21#include "llvm/IR/Module.h"
22#include "llvm/IR/Verifier.h"
23#include "llvm/Pass.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/FileUtilities.h"
27#include "llvm/Support/Path.h"
28#include "llvm/Support/Signals.h"
29#include "llvm/Support/ToolOutputFile.h"
30#include "llvm/Transforms/IPO.h"
31#include "llvm/Transforms/Scalar.h"
32#include "llvm/Transforms/Utils/Cloning.h"
33#include "llvm/Transforms/Utils/CodeExtractor.h"
34#include <set>
35using namespace llvm;
36
37#define DEBUG_TYPE "bugpoint"
38
39namespace llvm {
40  bool DisableSimplifyCFG = false;
41  extern cl::opt<std::string> OutputPrefix;
42} // End llvm namespace
43
44namespace {
45  cl::opt<bool>
46  NoDCE ("disable-dce",
47         cl::desc("Do not use the -dce pass to reduce testcases"));
48  cl::opt<bool, true>
49  NoSCFG("disable-simplifycfg", cl::location(DisableSimplifyCFG),
50         cl::desc("Do not use the -simplifycfg pass to reduce testcases"));
51
52  Function* globalInitUsesExternalBA(GlobalVariable* GV) {
53    if (!GV->hasInitializer())
54      return nullptr;
55
56    Constant *I = GV->getInitializer();
57
58    // walk the values used by the initializer
59    // (and recurse into things like ConstantExpr)
60    std::vector<Constant*> Todo;
61    std::set<Constant*> Done;
62    Todo.push_back(I);
63
64    while (!Todo.empty()) {
65      Constant* V = Todo.back();
66      Todo.pop_back();
67      Done.insert(V);
68
69      if (BlockAddress *BA = dyn_cast<BlockAddress>(V)) {
70        Function *F = BA->getFunction();
71        if (F->isDeclaration())
72          return F;
73      }
74
75      for (User::op_iterator i = V->op_begin(), e = V->op_end(); i != e; ++i) {
76        Constant *C = dyn_cast<Constant>(*i);
77        if (C && !isa<GlobalValue>(C) && !Done.count(C))
78          Todo.push_back(C);
79      }
80    }
81    return nullptr;
82  }
83}  // end anonymous namespace
84
85std::unique_ptr<Module>
86BugDriver::deleteInstructionFromProgram(const Instruction *I,
87                                        unsigned Simplification) {
88  // FIXME, use vmap?
89  Module *Clone = CloneModule(Program).release();
90
91  const BasicBlock *PBB = I->getParent();
92  const Function *PF = PBB->getParent();
93
94  Module::iterator RFI = Clone->begin(); // Get iterator to corresponding fn
95  std::advance(RFI, std::distance(PF->getParent()->begin(),
96                                  Module::const_iterator(PF)));
97
98  Function::iterator RBI = RFI->begin();  // Get iterator to corresponding BB
99  std::advance(RBI, std::distance(PF->begin(), Function::const_iterator(PBB)));
100
101  BasicBlock::iterator RI = RBI->begin(); // Get iterator to corresponding inst
102  std::advance(RI, std::distance(PBB->begin(), BasicBlock::const_iterator(I)));
103  Instruction *TheInst = &*RI; // Got the corresponding instruction!
104
105  // If this instruction produces a value, replace any users with null values
106  if (!TheInst->getType()->isVoidTy())
107    TheInst->replaceAllUsesWith(Constant::getNullValue(TheInst->getType()));
108
109  // Remove the instruction from the program.
110  TheInst->getParent()->getInstList().erase(TheInst);
111
112  // Spiff up the output a little bit.
113  std::vector<std::string> Passes;
114
115  /// Can we get rid of the -disable-* options?
116  if (Simplification > 1 && !NoDCE)
117    Passes.push_back("dce");
118  if (Simplification && !DisableSimplifyCFG)
119    Passes.push_back("simplifycfg");      // Delete dead control flow
120
121  Passes.push_back("verify");
122  std::unique_ptr<Module> New = runPassesOn(Clone, Passes);
123  delete Clone;
124  if (!New) {
125    errs() << "Instruction removal failed.  Sorry. :(  Please report a bug!\n";
126    exit(1);
127  }
128  return New;
129}
130
131std::unique_ptr<Module>
132BugDriver::performFinalCleanups(Module *M, bool MayModifySemantics) {
133  // Make all functions external, so GlobalDCE doesn't delete them...
134  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
135    I->setLinkage(GlobalValue::ExternalLinkage);
136
137  std::vector<std::string> CleanupPasses;
138  CleanupPasses.push_back("globaldce");
139
140  if (MayModifySemantics)
141    CleanupPasses.push_back("deadarghaX0r");
142  else
143    CleanupPasses.push_back("deadargelim");
144
145  std::unique_ptr<Module> New = runPassesOn(M, CleanupPasses);
146  if (!New) {
147    errs() << "Final cleanups failed.  Sorry. :(  Please report a bug!\n";
148    return nullptr;
149  }
150  delete M;
151  return New;
152}
153
154std::unique_ptr<Module> BugDriver::extractLoop(Module *M) {
155  std::vector<std::string> LoopExtractPasses;
156  LoopExtractPasses.push_back("loop-extract-single");
157
158  std::unique_ptr<Module> NewM = runPassesOn(M, LoopExtractPasses);
159  if (!NewM) {
160    outs() << "*** Loop extraction failed: ";
161    EmitProgressBitcode(M, "loopextraction", true);
162    outs() << "*** Sorry. :(  Please report a bug!\n";
163    return nullptr;
164  }
165
166  // Check to see if we created any new functions.  If not, no loops were
167  // extracted and we should return null.  Limit the number of loops we extract
168  // to avoid taking forever.
169  static unsigned NumExtracted = 32;
170  if (M->size() == NewM->size() || --NumExtracted == 0) {
171    return nullptr;
172  } else {
173    assert(M->size() < NewM->size() && "Loop extract removed functions?");
174    Module::iterator MI = NewM->begin();
175    for (unsigned i = 0, e = M->size(); i != e; ++i)
176      ++MI;
177  }
178
179  return NewM;
180}
181
182static void eliminateAliases(GlobalValue *GV) {
183  // First, check whether a GlobalAlias references this definition.
184  // GlobalAlias MAY NOT reference declarations.
185  for (;;) {
186    // 1. Find aliases
187    SmallVector<GlobalAlias*,1> aliases;
188    Module *M = GV->getParent();
189    for (Module::alias_iterator I=M->alias_begin(), E=M->alias_end(); I!=E; ++I)
190      if (I->getAliasee()->stripPointerCasts() == GV)
191        aliases.push_back(&*I);
192    if (aliases.empty())
193      break;
194    // 2. Resolve aliases
195    for (unsigned i=0, e=aliases.size(); i<e; ++i) {
196      aliases[i]->replaceAllUsesWith(aliases[i]->getAliasee());
197      aliases[i]->eraseFromParent();
198    }
199    // 3. Repeat until no more aliases found; there might
200    // be an alias to an alias...
201  }
202}
203
204//
205// DeleteGlobalInitializer - "Remove" the global variable by deleting its initializer,
206// making it external.
207//
208void llvm::DeleteGlobalInitializer(GlobalVariable *GV) {
209  eliminateAliases(GV);
210  GV->setInitializer(nullptr);
211}
212
213// DeleteFunctionBody - "Remove" the function by deleting all of its basic
214// blocks, making it external.
215//
216void llvm::DeleteFunctionBody(Function *F) {
217  eliminateAliases(F);
218  // Function declarations can't have comdats.
219  F->setComdat(nullptr);
220
221  // delete the body of the function...
222  F->deleteBody();
223  assert(F->isDeclaration() && "This didn't make the function external!");
224}
225
226/// GetTorInit - Given a list of entries for static ctors/dtors, return them
227/// as a constant array.
228static Constant *GetTorInit(std::vector<std::pair<Function*, int> > &TorList) {
229  assert(!TorList.empty() && "Don't create empty tor list!");
230  std::vector<Constant*> ArrayElts;
231  Type *Int32Ty = Type::getInt32Ty(TorList[0].first->getContext());
232
233  StructType *STy =
234      StructType::get(Int32Ty, TorList[0].first->getType(), nullptr);
235  for (unsigned i = 0, e = TorList.size(); i != e; ++i) {
236    Constant *Elts[] = {
237      ConstantInt::get(Int32Ty, TorList[i].second),
238      TorList[i].first
239    };
240    ArrayElts.push_back(ConstantStruct::get(STy, Elts));
241  }
242  return ConstantArray::get(ArrayType::get(ArrayElts[0]->getType(),
243                                           ArrayElts.size()),
244                            ArrayElts);
245}
246
247/// SplitStaticCtorDtor - A module was recently split into two parts, M1/M2, and
248/// M1 has all of the global variables.  If M2 contains any functions that are
249/// static ctors/dtors, we need to add an llvm.global_[cd]tors global to M2, and
250/// prune appropriate entries out of M1s list.
251static void SplitStaticCtorDtor(const char *GlobalName, Module *M1, Module *M2,
252                                ValueToValueMapTy &VMap) {
253  GlobalVariable *GV = M1->getNamedGlobal(GlobalName);
254  if (!GV || GV->isDeclaration() || GV->hasLocalLinkage() ||
255      !GV->use_empty()) return;
256
257  std::vector<std::pair<Function*, int> > M1Tors, M2Tors;
258  ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
259  if (!InitList) return;
260
261  for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i) {
262    if (ConstantStruct *CS = dyn_cast<ConstantStruct>(InitList->getOperand(i))){
263      if (CS->getNumOperands() != 2) return;  // Not array of 2-element structs.
264
265      if (CS->getOperand(1)->isNullValue())
266        break;  // Found a null terminator, stop here.
267
268      ConstantInt *CI = dyn_cast<ConstantInt>(CS->getOperand(0));
269      int Priority = CI ? CI->getSExtValue() : 0;
270
271      Constant *FP = CS->getOperand(1);
272      if (ConstantExpr *CE = dyn_cast<ConstantExpr>(FP))
273        if (CE->isCast())
274          FP = CE->getOperand(0);
275      if (Function *F = dyn_cast<Function>(FP)) {
276        if (!F->isDeclaration())
277          M1Tors.push_back(std::make_pair(F, Priority));
278        else {
279          // Map to M2's version of the function.
280          F = cast<Function>(VMap[F]);
281          M2Tors.push_back(std::make_pair(F, Priority));
282        }
283      }
284    }
285  }
286
287  GV->eraseFromParent();
288  if (!M1Tors.empty()) {
289    Constant *M1Init = GetTorInit(M1Tors);
290    new GlobalVariable(*M1, M1Init->getType(), false,
291                       GlobalValue::AppendingLinkage,
292                       M1Init, GlobalName);
293  }
294
295  GV = M2->getNamedGlobal(GlobalName);
296  assert(GV && "Not a clone of M1?");
297  assert(GV->use_empty() && "llvm.ctors shouldn't have uses!");
298
299  GV->eraseFromParent();
300  if (!M2Tors.empty()) {
301    Constant *M2Init = GetTorInit(M2Tors);
302    new GlobalVariable(*M2, M2Init->getType(), false,
303                       GlobalValue::AppendingLinkage,
304                       M2Init, GlobalName);
305  }
306}
307
308std::unique_ptr<Module>
309llvm::SplitFunctionsOutOfModule(Module *M, const std::vector<Function *> &F,
310                                ValueToValueMapTy &VMap) {
311  // Make sure functions & globals are all external so that linkage
312  // between the two modules will work.
313  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
314    I->setLinkage(GlobalValue::ExternalLinkage);
315  for (Module::global_iterator I = M->global_begin(), E = M->global_end();
316       I != E; ++I) {
317    if (I->hasName() && I->getName()[0] == '\01')
318      I->setName(I->getName().substr(1));
319    I->setLinkage(GlobalValue::ExternalLinkage);
320  }
321
322  ValueToValueMapTy NewVMap;
323  std::unique_ptr<Module> New = CloneModule(M, NewVMap);
324
325  // Remove the Test functions from the Safe module
326  std::set<Function *> TestFunctions;
327  for (unsigned i = 0, e = F.size(); i != e; ++i) {
328    Function *TNOF = cast<Function>(VMap[F[i]]);
329    DEBUG(errs() << "Removing function ");
330    DEBUG(TNOF->printAsOperand(errs(), false));
331    DEBUG(errs() << "\n");
332    TestFunctions.insert(cast<Function>(NewVMap[TNOF]));
333    DeleteFunctionBody(TNOF);       // Function is now external in this module!
334  }
335
336
337  // Remove the Safe functions from the Test module
338  for (Function &I : *New)
339    if (!TestFunctions.count(&I))
340      DeleteFunctionBody(&I);
341
342  // Try to split the global initializers evenly
343  for (GlobalVariable &I : M->globals()) {
344    GlobalVariable *GV = cast<GlobalVariable>(NewVMap[&I]);
345    if (Function *TestFn = globalInitUsesExternalBA(&I)) {
346      if (Function *SafeFn = globalInitUsesExternalBA(GV)) {
347        errs() << "*** Error: when reducing functions, encountered "
348                  "the global '";
349        GV->printAsOperand(errs(), false);
350        errs() << "' with an initializer that references blockaddresses "
351                  "from safe function '" << SafeFn->getName()
352               << "' and from test function '" << TestFn->getName() << "'.\n";
353        exit(1);
354      }
355      DeleteGlobalInitializer(&I); // Delete the initializer to make it external
356    } else {
357      // If we keep it in the safe module, then delete it in the test module
358      DeleteGlobalInitializer(GV);
359    }
360  }
361
362  // Make sure that there is a global ctor/dtor array in both halves of the
363  // module if they both have static ctor/dtor functions.
364  SplitStaticCtorDtor("llvm.global_ctors", M, New.get(), NewVMap);
365  SplitStaticCtorDtor("llvm.global_dtors", M, New.get(), NewVMap);
366
367  return New;
368}
369
370//===----------------------------------------------------------------------===//
371// Basic Block Extraction Code
372//===----------------------------------------------------------------------===//
373
374std::unique_ptr<Module>
375BugDriver::extractMappedBlocksFromModule(const std::vector<BasicBlock *> &BBs,
376                                         Module *M) {
377  SmallString<128> Filename;
378  int FD;
379  std::error_code EC = sys::fs::createUniqueFile(
380      OutputPrefix + "-extractblocks%%%%%%%", FD, Filename);
381  if (EC) {
382    outs() << "*** Basic Block extraction failed!\n";
383    errs() << "Error creating temporary file: " << EC.message() << "\n";
384    EmitProgressBitcode(M, "basicblockextractfail", true);
385    return nullptr;
386  }
387  sys::RemoveFileOnSignal(Filename);
388
389  tool_output_file BlocksToNotExtractFile(Filename.c_str(), FD);
390  for (std::vector<BasicBlock*>::const_iterator I = BBs.begin(), E = BBs.end();
391       I != E; ++I) {
392    BasicBlock *BB = *I;
393    // If the BB doesn't have a name, give it one so we have something to key
394    // off of.
395    if (!BB->hasName()) BB->setName("tmpbb");
396    BlocksToNotExtractFile.os() << BB->getParent()->getName() << " "
397                                << BB->getName() << "\n";
398  }
399  BlocksToNotExtractFile.os().close();
400  if (BlocksToNotExtractFile.os().has_error()) {
401    errs() << "Error writing list of blocks to not extract\n";
402    EmitProgressBitcode(M, "basicblockextractfail", true);
403    BlocksToNotExtractFile.os().clear_error();
404    return nullptr;
405  }
406  BlocksToNotExtractFile.keep();
407
408  std::string uniqueFN = "--extract-blocks-file=";
409  uniqueFN += Filename.str();
410  const char *ExtraArg = uniqueFN.c_str();
411
412  std::vector<std::string> PI;
413  PI.push_back("extract-blocks");
414  std::unique_ptr<Module> Ret = runPassesOn(M, PI, 1, &ExtraArg);
415
416  sys::fs::remove(Filename.c_str());
417
418  if (!Ret) {
419    outs() << "*** Basic Block extraction failed, please report a bug!\n";
420    EmitProgressBitcode(M, "basicblockextractfail", true);
421  }
422  return Ret;
423}
424