1//===- LoopExtractor.cpp - Extract each loop into a new function ----------===//
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// A pass wrapper around the ExtractLoop() scalar transformation to extract each
11// top-level loop into its own new function. If the loop is the ONLY loop in a
12// given function, it is not touched. This is a pass most useful for debugging
13// via bugpoint.
14//
15//===----------------------------------------------------------------------===//
16
17#define DEBUG_TYPE "loop-extract"
18#include "llvm/Transforms/IPO.h"
19#include "llvm/Instructions.h"
20#include "llvm/Module.h"
21#include "llvm/Pass.h"
22#include "llvm/Analysis/Dominators.h"
23#include "llvm/Analysis/LoopPass.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Transforms/Scalar.h"
26#include "llvm/Transforms/Utils/BasicBlockUtils.h"
27#include "llvm/Transforms/Utils/CodeExtractor.h"
28#include "llvm/ADT/Statistic.h"
29#include <fstream>
30#include <set>
31using namespace llvm;
32
33STATISTIC(NumExtracted, "Number of loops extracted");
34
35namespace {
36  struct LoopExtractor : public LoopPass {
37    static char ID; // Pass identification, replacement for typeid
38    unsigned NumLoops;
39
40    explicit LoopExtractor(unsigned numLoops = ~0)
41      : LoopPass(ID), NumLoops(numLoops) {
42        initializeLoopExtractorPass(*PassRegistry::getPassRegistry());
43      }
44
45    virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
46
47    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
48      AU.addRequiredID(BreakCriticalEdgesID);
49      AU.addRequiredID(LoopSimplifyID);
50      AU.addRequired<DominatorTree>();
51    }
52  };
53}
54
55char LoopExtractor::ID = 0;
56INITIALIZE_PASS_BEGIN(LoopExtractor, "loop-extract",
57                      "Extract loops into new functions", false, false)
58INITIALIZE_PASS_DEPENDENCY(BreakCriticalEdges)
59INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
60INITIALIZE_PASS_DEPENDENCY(DominatorTree)
61INITIALIZE_PASS_END(LoopExtractor, "loop-extract",
62                    "Extract loops into new functions", false, false)
63
64namespace {
65  /// SingleLoopExtractor - For bugpoint.
66  struct SingleLoopExtractor : public LoopExtractor {
67    static char ID; // Pass identification, replacement for typeid
68    SingleLoopExtractor() : LoopExtractor(1) {}
69  };
70} // End anonymous namespace
71
72char SingleLoopExtractor::ID = 0;
73INITIALIZE_PASS(SingleLoopExtractor, "loop-extract-single",
74                "Extract at most one loop into a new function", false, false)
75
76// createLoopExtractorPass - This pass extracts all natural loops from the
77// program into a function if it can.
78//
79Pass *llvm::createLoopExtractorPass() { return new LoopExtractor(); }
80
81bool LoopExtractor::runOnLoop(Loop *L, LPPassManager &LPM) {
82  // Only visit top-level loops.
83  if (L->getParentLoop())
84    return false;
85
86  // If LoopSimplify form is not available, stay out of trouble.
87  if (!L->isLoopSimplifyForm())
88    return false;
89
90  DominatorTree &DT = getAnalysis<DominatorTree>();
91  bool Changed = false;
92
93  // If there is more than one top-level loop in this function, extract all of
94  // the loops. Otherwise there is exactly one top-level loop; in this case if
95  // this function is more than a minimal wrapper around the loop, extract
96  // the loop.
97  bool ShouldExtractLoop = false;
98
99  // Extract the loop if the entry block doesn't branch to the loop header.
100  TerminatorInst *EntryTI =
101    L->getHeader()->getParent()->getEntryBlock().getTerminator();
102  if (!isa<BranchInst>(EntryTI) ||
103      !cast<BranchInst>(EntryTI)->isUnconditional() ||
104      EntryTI->getSuccessor(0) != L->getHeader()) {
105    ShouldExtractLoop = true;
106  } else {
107    // Check to see if any exits from the loop are more than just return
108    // blocks.
109    SmallVector<BasicBlock*, 8> ExitBlocks;
110    L->getExitBlocks(ExitBlocks);
111    for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
112      if (!isa<ReturnInst>(ExitBlocks[i]->getTerminator())) {
113        ShouldExtractLoop = true;
114        break;
115      }
116  }
117
118  if (ShouldExtractLoop) {
119    // We must omit landing pads. Landing pads must accompany the invoke
120    // instruction. But this would result in a loop in the extracted
121    // function. An infinite cycle occurs when it tries to extract that loop as
122    // well.
123    SmallVector<BasicBlock*, 8> ExitBlocks;
124    L->getExitBlocks(ExitBlocks);
125    for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
126      if (ExitBlocks[i]->isLandingPad()) {
127        ShouldExtractLoop = false;
128        break;
129      }
130  }
131
132  if (ShouldExtractLoop) {
133    if (NumLoops == 0) return Changed;
134    --NumLoops;
135    CodeExtractor Extractor(DT, *L);
136    if (Extractor.extractCodeRegion() != 0) {
137      Changed = true;
138      // After extraction, the loop is replaced by a function call, so
139      // we shouldn't try to run any more loop passes on it.
140      LPM.deleteLoopFromQueue(L);
141    }
142    ++NumExtracted;
143  }
144
145  return Changed;
146}
147
148// createSingleLoopExtractorPass - This pass extracts one natural loop from the
149// program into a function if it can.  This is used by bugpoint.
150//
151Pass *llvm::createSingleLoopExtractorPass() {
152  return new SingleLoopExtractor();
153}
154
155
156// BlockFile - A file which contains a list of blocks that should not be
157// extracted.
158static cl::opt<std::string>
159BlockFile("extract-blocks-file", cl::value_desc("filename"),
160          cl::desc("A file containing list of basic blocks to not extract"),
161          cl::Hidden);
162
163namespace {
164  /// BlockExtractorPass - This pass is used by bugpoint to extract all blocks
165  /// from the module into their own functions except for those specified by the
166  /// BlocksToNotExtract list.
167  class BlockExtractorPass : public ModulePass {
168    void LoadFile(const char *Filename);
169    void SplitLandingPadPreds(Function *F);
170
171    std::vector<BasicBlock*> BlocksToNotExtract;
172    std::vector<std::pair<std::string, std::string> > BlocksToNotExtractByName;
173  public:
174    static char ID; // Pass identification, replacement for typeid
175    BlockExtractorPass() : ModulePass(ID) {
176      if (!BlockFile.empty())
177        LoadFile(BlockFile.c_str());
178    }
179
180    bool runOnModule(Module &M);
181  };
182}
183
184char BlockExtractorPass::ID = 0;
185INITIALIZE_PASS(BlockExtractorPass, "extract-blocks",
186                "Extract Basic Blocks From Module (for bugpoint use)",
187                false, false)
188
189// createBlockExtractorPass - This pass extracts all blocks (except those
190// specified in the argument list) from the functions in the module.
191//
192ModulePass *llvm::createBlockExtractorPass() {
193  return new BlockExtractorPass();
194}
195
196void BlockExtractorPass::LoadFile(const char *Filename) {
197  // Load the BlockFile...
198  std::ifstream In(Filename);
199  if (!In.good()) {
200    errs() << "WARNING: BlockExtractor couldn't load file '" << Filename
201           << "'!\n";
202    return;
203  }
204  while (In) {
205    std::string FunctionName, BlockName;
206    In >> FunctionName;
207    In >> BlockName;
208    if (!BlockName.empty())
209      BlocksToNotExtractByName.push_back(
210          std::make_pair(FunctionName, BlockName));
211  }
212}
213
214/// SplitLandingPadPreds - The landing pad needs to be extracted with the invoke
215/// instruction. The critical edge breaker will refuse to break critical edges
216/// to a landing pad. So do them here. After this method runs, all landing pads
217/// should have only one predecessor.
218void BlockExtractorPass::SplitLandingPadPreds(Function *F) {
219  for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I) {
220    InvokeInst *II = dyn_cast<InvokeInst>(I);
221    if (!II) continue;
222    BasicBlock *Parent = II->getParent();
223    BasicBlock *LPad = II->getUnwindDest();
224
225    // Look through the landing pad's predecessors. If one of them ends in an
226    // 'invoke', then we want to split the landing pad.
227    bool Split = false;
228    for (pred_iterator
229           PI = pred_begin(LPad), PE = pred_end(LPad); PI != PE; ++PI) {
230      BasicBlock *BB = *PI;
231      if (BB->isLandingPad() && BB != Parent &&
232          isa<InvokeInst>(Parent->getTerminator())) {
233        Split = true;
234        break;
235      }
236    }
237
238    if (!Split) continue;
239
240    SmallVector<BasicBlock*, 2> NewBBs;
241    SplitLandingPadPredecessors(LPad, Parent, ".1", ".2", 0, NewBBs);
242  }
243}
244
245bool BlockExtractorPass::runOnModule(Module &M) {
246  std::set<BasicBlock*> TranslatedBlocksToNotExtract;
247  for (unsigned i = 0, e = BlocksToNotExtract.size(); i != e; ++i) {
248    BasicBlock *BB = BlocksToNotExtract[i];
249    Function *F = BB->getParent();
250
251    // Map the corresponding function in this module.
252    Function *MF = M.getFunction(F->getName());
253    assert(MF->getFunctionType() == F->getFunctionType() && "Wrong function?");
254
255    // Figure out which index the basic block is in its function.
256    Function::iterator BBI = MF->begin();
257    std::advance(BBI, std::distance(F->begin(), Function::iterator(BB)));
258    TranslatedBlocksToNotExtract.insert(BBI);
259  }
260
261  while (!BlocksToNotExtractByName.empty()) {
262    // There's no way to find BBs by name without looking at every BB inside
263    // every Function. Fortunately, this is always empty except when used by
264    // bugpoint in which case correctness is more important than performance.
265
266    std::string &FuncName  = BlocksToNotExtractByName.back().first;
267    std::string &BlockName = BlocksToNotExtractByName.back().second;
268
269    for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
270      Function &F = *FI;
271      if (F.getName() != FuncName) continue;
272
273      for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
274        BasicBlock &BB = *BI;
275        if (BB.getName() != BlockName) continue;
276
277        TranslatedBlocksToNotExtract.insert(BI);
278      }
279    }
280
281    BlocksToNotExtractByName.pop_back();
282  }
283
284  // Now that we know which blocks to not extract, figure out which ones we WANT
285  // to extract.
286  std::vector<BasicBlock*> BlocksToExtract;
287  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
288    SplitLandingPadPreds(&*F);
289    for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
290      if (!TranslatedBlocksToNotExtract.count(BB))
291        BlocksToExtract.push_back(BB);
292  }
293
294  for (unsigned i = 0, e = BlocksToExtract.size(); i != e; ++i) {
295    SmallVector<BasicBlock*, 2> BlocksToExtractVec;
296    BlocksToExtractVec.push_back(BlocksToExtract[i]);
297    if (const InvokeInst *II =
298        dyn_cast<InvokeInst>(BlocksToExtract[i]->getTerminator()))
299      BlocksToExtractVec.push_back(II->getUnwindDest());
300    CodeExtractor(BlocksToExtractVec).extractCodeRegion();
301  }
302
303  return !BlocksToExtract.empty();
304}
305