multiway_merge.h: Destruct only elements that were have been constructed before.

2007-11-28  Johannes Singler  <singler@ira.uka.de>

        * include/parallel/multiway_merge.h: Destruct only elements that
        were have been constructed before. Code beautifying and formatting.
        * include/parallel/losertree.h: (Copy) construct all loser tree
        item keys, so they can be deconstructed all at once.
        * include/parallel/quicksort.h: Fix memory leak.
        * include/parallel/random_shuffle.h: Use copy constructor instead
        of assignment. Code beautifying and formatting.
        * include/parallel/unique_copy.h: Use assignment instead of copy
        constructor.
        * include/parallel/multiway_mergesort.h: Use copy constructor
        instead of assignment. Code beautifying and formatting.
        * include/parallel/random_shuffle.h: Use copy constructor instead
        of assignment. Code beautifying.

From-SVN: r130490

libstdc++-v3/ChangeLog

+2007-11-28  Johannes Singler  <singler@ira.uka.de>
+
+        * include/parallel/multiway_merge.h: Destruct only elements that 
+        were have been constructed before. Code beautifying and formatting.
+        * include/parallel/losertree.h: (Copy) construct all loser tree 
+        item keys, so they can be deconstructed all at once.
+        * include/parallel/quicksort.h: Fix memory leak.
+        * include/parallel/random_shuffle.h: Use copy constructor instead 
+        of assignment. Code beautifying and formatting.
+        * include/parallel/unique_copy.h: Use assignment instead of copy 
+        constructor.
+        * include/parallel/multiway_mergesort.h: Use copy constructor 
+        instead of assignment. Code beautifying and formatting.
+        * include/parallel/random_shuffle.h: Use copy constructor instead 
+        of assignment. Code beautifying.
+
 2007-11-27  Kaz Kojima  <kkojima@gcc.gnu.org>
 
 	* testsuite/tr1/5_numerical_facilities/special_functions/

libstdc++-v3/include/parallel/losertree.h

@@ -230,6 +230,7 @@ template<typename T, typename Comparator = std::less<T> >
     unsigned int ik, k, offset;
     Loser* losers;
     Comparator comp;
+    bool first_insert;
 
   public:
     inline LoserTree(unsigned int _k, Comparator _comp = std::less<T>())
@@ -240,9 +241,12 @@ template<typename T, typename Comparator = std::less<T> >
       // Next greater power of 2.
       k = 1 << (log2(ik - 1) + 1);
       offset = k;
-      losers = static_cast<Loser*>(::operator new(k * 2 * sizeof(Loser)));
-      for (unsigned int i = ik - 1; i < k; i++)
+      // Avoid default-constructing losers[].key
+      losers = static_cast<Loser*>(::operator new(2 * k * sizeof(Loser)));
+      for (unsigned int i = ik - 1; i < k; ++i)
         losers[i + k].sup = true;
+
+      first_insert = true;
     }
 
     inline ~LoserTree()
@@ -257,9 +261,18 @@ template<typename T, typename Comparator = std::less<T> >
     {
       unsigned int pos = k + source;
 
+      if(first_insert)
+        {
+          // Construct all keys, so we can easily deconstruct them.
+          for (unsigned int i = 0; i < (2 * k); ++i)
+            new(&(losers[i].key)) T(key);
+          first_insert = false;
+        }
+      else
+        new(&(losers[pos].key)) T(key);
+
       losers[pos].sup = sup;
       losers[pos].source = source;
-      new(&(losers[pos].key)) T(key);
     }
 
     unsigned int
@@ -282,7 +295,8 @@ template<typename T, typename Comparator = std::less<T> >
               return left;
             }
           else
-            {	// Right one is less.
+            {
+              // Right one is less.
               losers[root] = losers[left];
               return right;
             }

libstdc++-v3/include/parallel/multiway_mergesort.h

@@ -124,6 +124,8 @@ template<typename RandomAccessIterator, typename _DifferenceTp>
   determine_samples(PMWMSSortingData<RandomAccessIterator>* sd,
                     _DifferenceTp& num_samples)
   {
+    typedef std::iterator_traits<RandomAccessIterator> traits_type;
+    typedef typename traits_type::value_type value_type;
     typedef _DifferenceTp difference_type;
 
     thread_index_t iam = omp_get_thread_num();
@@ -137,8 +139,8 @@ template<typename RandomAccessIterator, typename _DifferenceTp>
                   num_samples + 1, es);
 
     for (difference_type i = 0; i < num_samples; i++)
-      sd->samples[iam * num_samples + i] =
-          sd->source[sd->starts[iam] + es[i + 1]];
+      new(&(sd->samples[iam * num_samples + i])) value_type(
+          sd->source[sd->starts[iam] + es[i + 1]]);
 
     delete[] es;
   }
@@ -213,7 +215,8 @@ template<typename RandomAccessIterator, typename Comparator>
               if (num_samples * iam > 0)
                 sd->pieces[iam][s].begin = 
                     std::lower_bound(sd->sorting_places[s],
-                        sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s],
+                        sd->sorting_places[s]
+                            + (sd->starts[s + 1] - sd->starts[s]),
                         sd->samples[num_samples * iam],
                         comp)
                     - sd->sorting_places[s];
@@ -224,8 +227,10 @@ template<typename RandomAccessIterator, typename Comparator>
             if ((num_samples * (iam + 1)) < (num_samples * sd->num_threads))
               sd->pieces[iam][s].end =
                   std::lower_bound(sd->sorting_places[s],
-                                   sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s],
-                                   sd->samples[num_samples * (iam + 1)], comp)
+                          sd->sorting_places[s]
+                              + (sd->starts[s + 1] - sd->starts[s]),
+                          sd->samples[num_samples * (iam + 1)],
+                          comp)
                   - sd->sorting_places[s];
             else
               // Absolute end.
@@ -240,7 +245,8 @@ template<typename RandomAccessIterator, typename Comparator>
             seqs(sd->num_threads);
         for (int s = 0; s < sd->num_threads; s++)
           seqs[s] = std::make_pair(sd->sorting_places[s],
-                                   sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s]);
+                                   sd->sorting_places[s]
+                                       + (sd->starts[s + 1] - sd->starts[s]));
 
         std::vector<SortingPlacesIterator> offsets(sd->num_threads);
 
@@ -256,7 +262,8 @@ template<typename RandomAccessIterator, typename Comparator>
               sd->pieces[iam][seq].end = offsets[seq] - seqs[seq].first;
             else
               // very end of this sequence
-              sd->pieces[iam][seq].end = sd->starts[seq + 1] - sd->starts[seq];
+              sd->pieces[iam][seq].end =
+                  sd->starts[seq + 1] - sd->starts[seq];
           }
 
 #       pragma omp barrier
@@ -284,6 +291,7 @@ template<typename RandomAccessIterator, typename Comparator>
     // Merge to temporary storage, uninitialized creation not possible
     // since there is no multiway_merge calling the placement new
     // instead of the assignment operator.
+    // XXX incorrect (de)construction
     sd->merging_places[iam] = sd->temporaries[iam] =
         static_cast<value_type*>(
         ::operator new(sizeof(value_type) * length_am));
@@ -296,11 +304,13 @@ template<typename RandomAccessIterator, typename Comparator>
 
     for (int s = 0; s < sd->num_threads; s++)
       {
-        seqs[s] = std::make_pair(sd->sorting_places[s] + sd->pieces[iam][s].begin,
+        seqs[s] =
+            std::make_pair(sd->sorting_places[s] + sd->pieces[iam][s].begin,
                            sd->sorting_places[s] + sd->pieces[iam][s].end);
       }
 
-    multiway_merge(seqs.begin(), seqs.end(), sd->merging_places[iam], comp, length_am, sd->stable, false, sequential_tag());
+    multiway_merge(seqs.begin(), seqs.end(), sd->merging_places[iam], comp,
+                   length_am, sd->stable, false, sequential_tag());
 
 #   pragma omp barrier
 
@@ -326,7 +336,8 @@ template<typename RandomAccessIterator, typename Comparator>
   inline void
   parallel_sort_mwms(RandomAccessIterator begin, RandomAccessIterator end,
                      Comparator comp, 
-                     typename std::iterator_traits<RandomAccessIterator>::difference_type n,
+                     typename std::iterator_traits<RandomAccessIterator>
+                        ::difference_type n,
                      int num_threads,
                      bool stable)
   {
@@ -368,7 +379,8 @@ template<typename RandomAccessIterator, typename Comparator>
             if (Settings::sort_splitting == Settings::SAMPLING)
               {
                 unsigned int size = 
-                    (Settings::sort_mwms_oversampling * num_threads - 1) * num_threads;
+                    (Settings::sort_mwms_oversampling * num_threads - 1)
+                        * num_threads;
                 sd.samples = static_cast<value_type*>(
                     ::operator new(size * sizeof(value_type)));
               }

libstdc++-v3/include/parallel/partial_sum.h

@@ -73,8 +73,8 @@ template<
       {
         value = bin_op(value, *begin);
         *result = value;
-        result++;
-        begin++;
+        ++result;
+        ++begin;
       }
     return result;
   }
@@ -103,6 +103,9 @@ template<
     typedef typename traits_type::value_type value_type;
     typedef typename traits_type::difference_type difference_type;
 
+    if (begin == end)
+      return result;
+
     thread_index_t num_threads =
         std::min<difference_type>(get_max_threads(), n - 1);
 
@@ -133,7 +136,7 @@ template<
                     ((double)num_threads + Settings::partial_sum_dilatation)),
                     borderstart = n - num_threads * chunk_length;
                 borders[0] = 0;
-                for (int i = 1; i < (num_threads + 1); i++)
+                for (int i = 1; i < (num_threads + 1); ++i)
                   {
                     borders[i] = borderstart;
                     borderstart += chunk_length;
@@ -146,20 +149,21 @@ template<
             OutputIterator target_end;
           } //single
 
-        int iam = omp_get_thread_num();
+        thread_index_t iam = omp_get_thread_num();
         if (iam == 0)
           {
             *result = *begin;
             parallel_partial_sum_basecase(begin + 1, begin + borders[1],
                           result + 1, bin_op, *begin);
-            sums[0] = *(result + borders[1] - 1);
+            new(&(sums[iam])) value_type(*(result + borders[1] - 1));
           }
         else
           {
-            sums[iam] = std::accumulate(begin + borders[iam] + 1,
+            new(&(sums[iam])) value_type(
+                                std::accumulate(begin + borders[iam] + 1,
                                 begin + borders[iam + 1],
                                 *(begin + borders[iam]),
-                            bin_op, __gnu_parallel::sequential_tag());
+                                bin_op, __gnu_parallel::sequential_tag()));
           }
 
 #       pragma omp barrier

libstdc++-v3/include/parallel/quicksort.h

@@ -74,13 +74,13 @@ namespace __gnu_parallel
 
     // Allocate uninitialized, to avoid default constructor.
     value_type* samples = static_cast<value_type*>(
-        operator new(num_samples * sizeof(value_type)));
+      ::operator new(num_samples * sizeof(value_type)));
 
-    for (difference_type s = 0; s < num_samples; s++)
+    for (difference_type s = 0; s < num_samples; ++s)
       {
         const unsigned long long index = static_cast<unsigned long long>(s)
                         * n / num_samples;
-        new(samples + s) value_type(begin[index]);
+        new(&(samples[s])) value_type(begin[index]);
       }
 
     __gnu_sequential::sort(samples, samples + num_samples, comp);
@@ -91,6 +91,8 @@ namespace __gnu_parallel
         pred(comp, pivot);
     difference_type split = parallel_partition(begin, end, pred, num_threads);
 
+    delete[] samples;
+
     return split;
   }
 

libstdc++-v3/include/parallel/random_shuffle.h

@@ -144,23 +144,23 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
     value_type** temporaries = new value_type*[d->num_threads];
 
     // Compute oracles and count appearances.
-    for (bin_index b = 0; b < sd->num_bins + 1; b++)
+    for (bin_index b = 0; b < sd->num_bins + 1; ++b)
       dist[b] = 0;
     int num_bits = sd->num_bits;
 
     random_number rng(d->seed);
 
     // First main loop.
-    for (difference_type i = 0; i < length; i++)
+    for (difference_type i = 0; i < length; ++i)
       {
         bin_index oracle = random_number_pow2(num_bits, rng);
         oracles[i] = oracle;
 
         // To allow prefix (partial) sum.
-        dist[oracle + 1]++;
+        ++(dist[oracle + 1]);
       }
 
-    for (bin_index b = 0; b < sd->num_bins + 1; b++)
+    for (bin_index b = 0; b < sd->num_bins + 1; ++b)
       sd->dist[b][iam + 1] = dist[b];
 
 #   pragma omp barrier
@@ -169,7 +169,7 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
     {
       // Sum up bins, sd->dist[s + 1][d->num_threads] now contains the
       // total number of items in bin s
-      for (bin_index s = 0; s < sd->num_bins; s++)
+      for (bin_index s = 0; s < sd->num_bins; ++s)
         __gnu_sequential::partial_sum(sd->dist[s + 1],
                                       sd->dist[s + 1] + d->num_threads + 1,
                                       sd->dist[s + 1]);
@@ -178,14 +178,14 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
 #   pragma omp barrier
 
     sequence_index_t offset = 0, global_offset = 0;
-    for (bin_index s = 0; s < d->bins_begin; s++)
+    for (bin_index s = 0; s < d->bins_begin; ++s)
       global_offset += sd->dist[s + 1][d->num_threads];
 
 #   pragma omp barrier
 
-    for (bin_index s = d->bins_begin; s < d->bins_end; s++)
+    for (bin_index s = d->bins_begin; s < d->bins_end; ++s)
       {
-	for (int t = 0; t < d->num_threads + 1; t++)
+	for (int t = 0; t < d->num_threads + 1; ++t)
 	  sd->dist[s + 1][t] += offset;
 	offset = sd->dist[s + 1][d->num_threads];
       }
@@ -196,24 +196,25 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
 #   pragma omp barrier
 
     // Draw local copies to avoid false sharing.
-    for (bin_index b = 0; b < sd->num_bins + 1; b++)
+    for (bin_index b = 0; b < sd->num_bins + 1; ++b)
       dist[b] = sd->dist[b][iam];
-    for (bin_index b = 0; b < sd->num_bins; b++)
+    for (bin_index b = 0; b < sd->num_bins; ++b)
       bin_proc[b] = sd->bin_proc[b];
-    for (thread_index_t t = 0; t < d->num_threads; t++)
+    for (thread_index_t t = 0; t < d->num_threads; ++t)
       temporaries[t] = sd->temporaries[t];
 
     RandomAccessIterator source = sd->source;
     difference_type start = sd->starts[iam];
 
     // Distribute according to oracles, second main loop.
-    for (difference_type i = 0; i < length; i++)
+    for (difference_type i = 0; i < length; ++i)
       {
         bin_index target_bin = oracles[i];
         thread_index_t target_p = bin_proc[target_bin];
 
         // Last column [d->num_threads] stays unchanged.
-        temporaries[target_p][dist[target_bin + 1]++] = *(source + i + start);
+        new(&(temporaries[target_p][dist[target_bin + 1]++])) value_type(
+              *(source + i + start));
       }
 
     delete[] oracles;
@@ -224,7 +225,7 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
 #   pragma omp barrier
 
     // Shuffle bins internally.
-    for (bin_index b = d->bins_begin; b < d->bins_end; b++)
+    for (bin_index b = d->bins_begin; b < d->bins_end; ++b)
       {
         value_type* begin =
                     sd->temporaries[iam] +
@@ -338,9 +339,9 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
             sd.temporaries = new value_type*[num_threads];
             sd.dist = new difference_type*[num_bins + 1];
             sd.bin_proc = new thread_index_t[num_bins];
-            for (bin_index b = 0; b < num_bins + 1; b++)
+            for (bin_index b = 0; b < num_bins + 1; ++b)
               sd.dist[b] = new difference_type[num_threads + 1];
-            for (bin_index b = 0; b < (num_bins + 1); b++)
+            for (bin_index b = 0; b < (num_bins + 1); ++b)
               {
                 sd.dist[0][0] = 0;
                 sd.dist[b][0] = 0;
@@ -354,7 +355,7 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
                             split = n % num_threads, start = 0;
             difference_type bin_chunk_length = num_bins / num_threads,
                             bin_split = num_bins % num_threads;
-            for (thread_index_t i = 0; i < num_threads; i++)
+            for (thread_index_t i = 0; i < num_threads; ++i)
               {
                 starts[i] = start;
                 start += (i < split) ? (chunk_length + 1) : chunk_length;
@@ -364,7 +365,7 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
                 bin_cursor += (i < bin_split) ?
                     (bin_chunk_length + 1) : bin_chunk_length;
                 pus[i].bins_end = bin_cursor;
-                for (; j < bin_cursor; j++)
+                for (; j < bin_cursor; ++j)
                   sd.bin_proc[j] = i;
                 pus[i].num_threads = num_threads;
                 pus[i].seed = rng(std::numeric_limits<uint32>::max());
@@ -378,7 +379,7 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
 
     delete[] starts;
     delete[] sd.bin_proc;
-    for (int s = 0; s < (num_bins + 1); s++)
+    for (int s = 0; s < (num_bins + 1); ++s)
       delete[] sd.dist[s];
     delete[] sd.dist;
     delete[] sd.temporaries;
@@ -455,31 +456,31 @@ template<typename RandomAccessIterator, typename RandomNumberGenerator>
         difference_type* dist0 = new difference_type[num_bins + 1],
                        * dist1 = new difference_type[num_bins + 1];
 
-        for (int b = 0; b < num_bins + 1; b++)
+        for (int b = 0; b < num_bins + 1; ++b)
           dist0[b] = 0;
 
         random_number bitrng(rng(0xFFFFFFFF));
 
-        for (difference_type i = 0; i < n; i++)
+        for (difference_type i = 0; i < n; ++i)
           {
             bin_index oracle = random_number_pow2(num_bits, bitrng);
             oracles[i] = oracle;
 
             // To allow prefix (partial) sum.
-            dist0[oracle + 1]++;
+            ++(dist0[oracle + 1]);
           }
 
         // Sum up bins.
         __gnu_sequential::partial_sum(dist0, dist0 + num_bins + 1, dist0);
 
-        for (int b = 0; b < num_bins + 1; b++)
+        for (int b = 0; b < num_bins + 1; ++b)
           dist1[b] = dist0[b];
 
         // Distribute according to oracles.
-        for (difference_type i = 0; i < n; i++)
-          target[(dist0[oracles[i]])++] = *(begin + i);
+        for (difference_type i = 0; i < n; ++i)
+          new(&(target[(dist0[oracles[i]])++])) value_type(*(begin + i));
 
-        for (int b = 0; b < num_bins; b++)
+        for (int b = 0; b < num_bins; ++b)
           {
             sequential_random_shuffle(target + dist1[b],
                                       target + dist1[b + 1],

libstdc++-v3/include/parallel/unique_copy.h

@@ -100,16 +100,14 @@ template<
           end = borders[iam + 1];
 
           i++;
-          new (static_cast<void *>(&*out)) value_type(*first);
-          out++;
+          *out++ = *first;
 
           for (InputIterator iter = first + begin; iter < first + end; ++iter)
             {
               if (!binary_pred(*iter, *(iter-1)))
                 {
                   i++;
-                  new (static_cast<void *>(&*out)) value_type(*iter);
-                  out++;
+                  *out++ = *iter;
                 }
             }
         }
@@ -153,8 +151,7 @@ template<
               if (iter == first || !binary_pred(*iter, *(iter-1)))
                 {
                   i++;
-                  new (static_cast<void *>(&*iter_out)) value_type(*iter);
-                  iter_out++;
+                  *iter_out++ = *iter;
                 }
             }
 
@@ -170,8 +167,7 @@ template<
             {
               if (!binary_pred(*iter, *(iter-1)))
                 {
-                  new (static_cast<void *> (&*iter_out)) value_type(*iter);
-                  iter_out++;
+                  *iter_out++ = *iter;
                 }
             }
         }