Speed up convergence significantly and also reduce the size of testcases by making large portions of a function's CFG dead at a time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@5915 91177308-0d34-0410-b5e6-96231b3b80d8

tools/bugpoint/BugDriver.h

@@ -21,6 +21,7 @@ class DebugCrashes;
 class ReduceMiscompilingPasses;
 class ReduceMiscompilingFunctions;
 class ReduceCrashingFunctions;
+class ReduceCrashingBlocks;
 
 class BugDriver {
   const std::string ToolName;  // Name of bugpoint
@@ -33,6 +34,7 @@ class BugDriver {
   friend class ReduceMiscompilingPasses;
   friend class ReduceMiscompilingFunctions;
   friend class ReduceCrashingFunctions;
+  friend class ReduceCrashingBlocks;
 public:
   BugDriver(const char *toolname)
     : ToolName(toolname), Program(0), Interpreter(0) {}

tools/bugpoint/CrashDebugger.cpp

@@ -8,9 +8,17 @@
 #include "SystemUtils.h"
 #include "ListReducer.h"
 #include "llvm/Module.h"
+#include "llvm/PassManager.h"
+#include "llvm/Pass.h"
+#include "llvm/Constant.h"
+#include "llvm/iTerminators.h"
+#include "llvm/Type.h"
+#include "llvm/SymbolTable.h"
+#include "llvm/Support/CFG.h"
+#include "llvm/Analysis/Verifier.h"
+#include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/Bytecode/Writer.h"
-#include "llvm/Pass.h"
 #include <fstream>
 #include <set>
 
@@ -117,6 +125,108 @@ bool ReduceCrashingFunctions::TestFuncs(std::vector<Function*> &Funcs) {
 }
 
 
+/// ReduceCrashingBlocks reducer - This works by setting the terminators of all
+/// terminators except the specified basic blocks to a 'ret' instruction, then
+/// running the simplify-cfg pass.  This has the effect of chopping up the CFG
+/// really fast which can reduce large functions quickly.
+///
+class ReduceCrashingBlocks : public ListReducer<BasicBlock*> {
+  BugDriver &BD;
+public:
+  ReduceCrashingBlocks(BugDriver &bd) : BD(bd) {}
+    
+  virtual TestResult doTest(std::vector<BasicBlock*> &Prefix,
+                            std::vector<BasicBlock*> &Kept) {
+    if (TestBlocks(Kept))
+      return KeepSuffix;
+    if (!Prefix.empty() && TestBlocks(Prefix))
+      return KeepPrefix;
+    return NoFailure;
+  }
+    
+  bool TestBlocks(std::vector<BasicBlock*> &Prefix);
+};
+
+bool ReduceCrashingBlocks::TestBlocks(std::vector<BasicBlock*> &BBs) {
+  // Clone the program to try hacking it appart...
+  Module *M = CloneModule(BD.Program);
+  
+  // Convert list to set for fast lookup...
+  std::set<BasicBlock*> Blocks;
+  for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
+    // Convert the basic block from the original module to the new module...
+    Function *F = BBs[i]->getParent();
+    Function *CMF = M->getFunction(F->getName(), F->getFunctionType());
+    assert(CMF && "Function not in module?!");
+
+    // Get the mapped basic block...
+    Function::iterator CBI = CMF->begin();
+    std::advance(CBI, std::distance(F->begin(), Function::iterator(BBs[i])));
+    Blocks.insert(CBI);
+  }
+
+  std::cout << "Checking for crash with only these blocks:";
+  for (unsigned i = 0, e = Blocks.size(); i != e; ++i)
+    std::cout << " " << BBs[i]->getName();
+  std::cout << ": ";
+
+  // Loop over and delete any hack up any blocks that are not listed...
+  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I)
+    for (Function::iterator BB = I->begin(), E = I->end(); BB != E; ++BB)
+      if (!Blocks.count(BB) && !isa<ReturnInst>(BB->getTerminator())) {
+        // Loop over all of the successors of this block, deleting any PHI nodes
+        // that might include it.
+        for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
+          (*SI)->removePredecessor(BB);
+
+        // Delete the old terminator instruction...
+        BB->getInstList().pop_back();
+        
+        // Add a new return instruction of the appropriate type...
+        const Type *RetTy = BB->getParent()->getReturnType();
+        ReturnInst *RI = new ReturnInst(RetTy == Type::VoidTy ? 0 :
+                                        Constant::getNullValue(RetTy));
+        BB->getInstList().push_back(RI);
+      }
+
+  // The CFG Simplifier pass may delete one of the basic blocks we are
+  // interested in.  If it does we need to take the block out of the list.  Make
+  // a "persistent mapping" by turning basic blocks into <function, name> pairs.
+  // This won't work well if blocks are unnamed, but that is just the risk we
+  // have to take.
+  std::vector<std::pair<Function*, std::string> > BlockInfo;
+
+  for (std::set<BasicBlock*>::iterator I = Blocks.begin(), E = Blocks.end();
+       I != E; ++I)
+    BlockInfo.push_back(std::make_pair((*I)->getParent(), (*I)->getName()));
+
+  // Now run the CFG simplify pass on the function...
+  PassManager Passes;
+  Passes.add(createCFGSimplificationPass());
+  Passes.add(createVerifierPass());
+  Passes.run(*M);
+
+  // Try running on the hacked up program...
+  std::swap(BD.Program, M);
+  if (BD.runPasses(BD.PassesToRun)) {
+    delete M;         // It crashed, keep the trimmed version...
+
+    // Make sure to use basic block pointers that point into the now-current
+    // module, and that they don't include any deleted blocks.
+    BBs.clear();
+    for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) {
+      SymbolTable &ST = BlockInfo[i].first->getSymbolTable();
+      SymbolTable::iterator I = ST.find(Type::LabelTy);
+      if (I != ST.end() && I->second.count(BlockInfo[i].second))
+        BBs.push_back(cast<BasicBlock>(I->second[BlockInfo[i].second]));
+    }
+    return true;
+  }
+  delete BD.Program;  // It didn't crash, revert...
+  BD.Program = M;
+  return false;
+}
+
 /// debugCrash - This method is called when some pass crashes on input.  It
 /// attempts to prune down the testcase to something reasonable, and figure
 /// out exactly which pass is crashing.
@@ -154,8 +264,16 @@ bool BugDriver::debugCrash() {
     }
   }
 
-  // FIXME: This should attempt to delete entire basic blocks at a time to speed
-  // up convergence...
+  // Attempt to delete entire basic blocks at a time to speed up
+  // convergence... this actually works by setting the terminator of the blocks
+  // to a return instruction then running simplifycfg, which can potentially
+  // shrinks the code dramatically quickly
+  //
+  std::vector<BasicBlock*> Blocks;
+  for (Module::iterator I = Program->begin(), E = Program->end(); I != E; ++I)
+    for (Function::iterator FI = I->begin(), E = I->end(); FI != E; ++FI)
+      Blocks.push_back(FI);
+  ReduceCrashingBlocks(*this).reduceList(Blocks);
 
   // FIXME: This should use the list reducer to converge faster by deleting
   // larger chunks of instructions at a time!