Add an implementation of the delta debugging algorithm.

- This is a pretty slow / memory intensive implementation, and I will likely
   change it to an iterative model, but it works.

llvm-svn: 90447

include/llvm/ADT/DeltaAlgorithm.h

@@ -0,0 +1,89 @@
+//===--- DeltaAlgorithm.h - A Set Minimization Algorithm -------*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_ADT_DELTAALGORITHM_H
+#define LLVM_ADT_DELTAALGORITHM_H
+
+#include <vector>
+#include <set>
+
+namespace llvm {
+
+/// DeltaAlgorithm - Implements the delta debugging algorithm (A. Zeller '99)
+/// for minimizing arbitrary sets using a predicate function.
+///
+/// The result of the algorithm is a subset of the input change set which is
+/// guaranteed to satisfy the predicate, assuming that the input set did. For
+/// well formed predicates, the result set is guaranteed to be such that
+/// removing any single element would falsify the predicate.
+///
+/// For best results the predicate function *should* (but need not) satisfy
+/// certain properties, in particular:
+///  (1) The predicate should return false on an empty set and true on the full
+///  set.
+///  (2) If the predicate returns true for a set of changes, it should return
+///  true for all supersets of that set.
+///
+/// It is not an error to provide a predicate that does not satisfy these
+/// requirements, and the algorithm will generally produce reasonable
+/// results. However, it may run substantially more tests than with a good
+/// predicate.
+class DeltaAlgorithm {
+public:
+  typedef unsigned change_ty;
+  // FIXME: Use a decent data structure.
+  typedef std::set<change_ty> changeset_ty;
+  typedef std::vector<changeset_ty> changesetlist_ty;
+
+private:
+  /// Cache of failed test results. Successful test results are never cached
+  /// since we always reduce following a success.
+  std::set<changeset_ty> FailedTestsCache;
+
+  /// GetTestResult - Get the test result for the \arg Changes from the
+  /// cache, executing the test if necessary.
+  ///
+  /// \param Changes - The change set to test.
+  /// \return - The test result.
+  bool GetTestResult(const changeset_ty &Changes);
+
+  /// Split - Partition a set of changes \arg Sinto one or two subsets.
+  void Split(const changeset_ty &S, changesetlist_ty &Res);
+
+  /// Delta - Minimize a set of \arg Changes which has been partioned into
+  /// smaller sets, by attempting to remove individual subsets.
+  changeset_ty Delta(const changeset_ty &Changes,
+                     const changesetlist_ty &Sets);
+
+  /// Search - Search for a subset (or subsets) in \arg Sets which can be
+  /// removed from \arg Changes while still satisfying the predicate.
+  ///
+  /// \param Res - On success, a subset of Changes which satisfies the
+  /// predicate.
+  /// \return - True on success.
+  bool Search(const changeset_ty &Changes, const changesetlist_ty &Sets,
+              changeset_ty &Res);
+              
+protected:
+  /// UpdatedSearchState - Callback used when the search state changes.
+  virtual void UpdatedSearchState(const changeset_ty &Changes,
+                                  const changesetlist_ty &Sets) {}
+
+  /// ExecuteOneTest - Execute a single test predicate on the change set \arg S.
+  virtual bool ExecuteOneTest(const changeset_ty &S) = 0;
+
+public:
+  /// Run - Minimize the set \arg Changes by executing \see ExecuteOneTest() on
+  /// subsets of changes and returning the smallest set which still satisfies
+  /// the test predicate.
+  changeset_ty Run(const changeset_ty &Changes);
+};
+
+} // end namespace llvm
+
+#endif

lib/Support/CMakeLists.txt

@@ -6,6 +6,7 @@ add_llvm_library(LLVMSupport
   CommandLine.cpp
   ConstantRange.cpp
   Debug.cpp
+  DeltaAlgorithm.cpp
   Dwarf.cpp
   ErrorHandling.cpp
   FileUtilities.cpp

lib/Support/DeltaAlgorithm.cpp

@@ -0,0 +1,110 @@
+//===--- DeltaAlgorithm.h - A Set Minimization Algorithm -------*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/DeltaAlgorithm.h"
+#include <algorithm>
+using namespace llvm;
+
+bool DeltaAlgorithm::GetTestResult(const changeset_ty &Changes) {
+  if (FailedTestsCache.count(Changes))
+    return false;
+
+  bool Result = ExecuteOneTest(Changes);
+  if (!Result)
+    FailedTestsCache.insert(Changes);
+
+  return Result;
+}
+
+void DeltaAlgorithm::Split(const changeset_ty &S, changesetlist_ty &Res) {
+  // FIXME: Allow clients to provide heuristics for improved splitting.
+
+  // FIXME: This is really slow.
+  changeset_ty LHS, RHS;
+  unsigned idx = 0;
+  for (changeset_ty::const_iterator it = S.begin(),
+         ie = S.end(); it != ie; ++it, ++idx)
+    ((idx & 1) ? LHS : RHS).insert(*it);
+  if (!LHS.empty())
+    Res.push_back(LHS);
+  if (!RHS.empty())
+    Res.push_back(RHS);
+}
+
+DeltaAlgorithm::changeset_ty
+DeltaAlgorithm::Delta(const changeset_ty &Changes,
+                      const changesetlist_ty &Sets) {
+  // Invariant: union(Res) == Changes
+  UpdatedSearchState(Changes, Sets);
+
+  // If there is nothing left we can remove, we are done.
+  if (Sets.size() <= 1)
+    return Changes;
+
+  // Look for a passing subset.
+  changeset_ty Res;
+  if (Search(Changes, Sets, Res))
+    return Res;
+
+  // Otherwise, partition the sets if possible; if not we are done.
+  changesetlist_ty SplitSets;
+  for (changesetlist_ty::const_iterator it = Sets.begin(),
+         ie = Sets.end(); it != ie; ++it)
+    Split(*it, SplitSets);
+  if (SplitSets.size() == Sets.size())
+    return Changes;
+
+  return Delta(Changes, SplitSets);
+}
+
+bool DeltaAlgorithm::Search(const changeset_ty &Changes,
+                            const changesetlist_ty &Sets,
+                            changeset_ty &Res) {
+  // FIXME: Parallelize.
+  for (changesetlist_ty::const_iterator it = Sets.begin(),
+         ie = Sets.end(); it != ie; ++it) {
+    // If the test passes on this subset alone, recurse.
+    if (GetTestResult(*it)) {
+      changesetlist_ty Sets;
+      Split(*it, Sets);
+      Res = Delta(*it, Sets);
+      return true;
+    }
+
+    // Otherwise, if we have more than two sets, see if test passes on the
+    // complement.
+    if (Sets.size() > 2) {
+      // FIXME: This is really slow.
+      changeset_ty Complement;
+      std::set_difference(
+        Changes.begin(), Changes.end(), it->begin(), it->end(),
+        std::insert_iterator<changeset_ty>(Complement, Complement.begin()));
+      if (GetTestResult(Complement)) {
+        changesetlist_ty ComplementSets;
+        ComplementSets.insert(ComplementSets.end(), Sets.begin(), it);
+        ComplementSets.insert(ComplementSets.end(), it + 1, Sets.end());
+        Res = Delta(Complement, ComplementSets);
+        return true;
+      }
+    }
+  }
+
+  return false;
+}
+
+DeltaAlgorithm::changeset_ty DeltaAlgorithm::Run(const changeset_ty &Changes) {
+  // Check empty set first to quickly find poor test functions.
+  if (GetTestResult(changeset_ty()))
+    return changeset_ty();
+
+  // Otherwise run the real delta algorithm.
+  changesetlist_ty Sets;
+  Split(Changes, Sets);
+
+  return Delta(Changes, Sets);
+}

unittests/ADT/DeltaAlgorithmTest.cpp

@@ -0,0 +1,96 @@
+//===- llvm/unittest/ADT/DeltaAlgorithmTest.cpp ---------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "gtest/gtest.h"
+#include "llvm/ADT/DeltaAlgorithm.h"
+#include <algorithm>
+#include <cstdarg>
+using namespace llvm;
+
+std::ostream &operator<<(std::ostream &OS,
+                         const std::set<unsigned> &S) {
+  OS << "{";
+  for (std::set<unsigned>::const_iterator it = S.begin(),
+         ie = S.end(); it != ie; ++it) {
+    if (it != S.begin())
+      OS << ",";
+    OS << *it;
+  }
+  OS << "}";
+  return OS;
+}
+
+namespace {
+
+class FixedDeltaAlgorithm : public DeltaAlgorithm {
+  changeset_ty FailingSet;
+  unsigned NumTests;
+
+protected:
+  virtual bool ExecuteOneTest(const changeset_ty &Changes) {
+    ++NumTests;
+    return std::includes(Changes.begin(), Changes.end(),
+                         FailingSet.begin(), FailingSet.end());
+  }
+
+public:
+  FixedDeltaAlgorithm(const changeset_ty &_FailingSet)
+    : FailingSet(_FailingSet),
+      NumTests(0) {}
+
+  unsigned getNumTests() const { return NumTests; }
+};
+
+std::set<unsigned> fixed_set(unsigned N, ...) {
+  std::set<unsigned> S;
+  va_list ap;
+  va_start(ap, N);
+  for (unsigned i = 0; i != N; ++i)
+    S.insert(va_arg(ap, unsigned));
+  va_end(ap);
+  return S;
+}
+
+std::set<unsigned> range(unsigned Start, unsigned End) {
+  std::set<unsigned> S;
+  while (Start != End)
+    S.insert(Start++);
+  return S;
+}
+
+std::set<unsigned> range(unsigned N) {
+  return range(0, N);
+}
+
+TEST(DeltaAlgorithmTest, Basic) {
+  // P = {3,5,7} \in S
+  //   [0, 20) should minimize to {3,5,7} in a reasonable number of tests.
+  std::set<unsigned> Fails = fixed_set(3, 3, 5, 7);
+  FixedDeltaAlgorithm FDA(Fails);
+  EXPECT_EQ(fixed_set(3, 3, 5, 7), FDA.Run(range(20)));
+  EXPECT_GE(33U, FDA.getNumTests());
+
+  // P = {3,5,7} \in S
+  //   [10, 20) should minimize to [10,20)
+  EXPECT_EQ(range(10,20), FDA.Run(range(10,20)));
+
+  // P = [0,4) \in S
+  //   [0, 4) should minimize to [0,4) in 11 tests.
+  //
+  // 11 = |{ {},
+  //         {0}, {1}, {2}, {3},
+  //         {1, 2, 3}, {0, 2, 3}, {0, 1, 3}, {0, 1, 2}, 
+  //         {0, 1}, {2, 3} }|
+  FDA = FixedDeltaAlgorithm(range(10));
+  EXPECT_EQ(range(4), FDA.Run(range(4)));
+  EXPECT_EQ(11U, FDA.getNumTests());  
+}
+
+}
+