std::pair<ssize_t,ssize_t> PartitionTernaryLL(SortArray& A, ssize_t lo, ssize_t hi)
{
// pick pivot and swap to back
ssize_t p = QuickSortSelectPivot(A, lo, hi);
value_type pivot = A[p];
A.swap(p, hi-1);
A.mark(hi-1);
volatile ssize_t i = lo, k = hi-1;
A.watch(&i, 3);
for (ssize_t j = lo; j < k; ++j)
{
int cmp = A[j].cmp(pivot); // ternary comparison
if (cmp == 0) {
A.swap(--k, j);
--j; // reclassify A[j]
A.mark(k,4);
}
else if (cmp < 0) {
A.swap(i++, j);
}
}
// unwatch i, because the pivot is swapped there
// in the first step of the following swap loop.
A.unwatch_all();
ssize_t j = i + (hi-k);
for (ssize_t s = 0; s < hi-k; ++s) {
A.swap(i+s, hi-1-s);
A.mark_swap(i+s, hi-1-s);
}
A.unmark_all();
return std::make_pair(i,j);
}
void QuickSortTernaryLR(SortArray& A, ssize_t lo, ssize_t hi)
{
if (hi <= lo) return;
int cmp;
// pick pivot and swap to back
ssize_t piv = QuickSortSelectPivot(A, lo, hi+1);
A.swap(piv, hi);
A.mark(hi);
const value_type& pivot = A[hi];
// schema: |p === |i <<< | ??? |j >>> |q === |piv
volatile ssize_t i = lo, j = hi-1;
volatile ssize_t p = lo, q = hi-1;
A.watch(&i, 3);
A.watch(&j, 3);
for (;;)
{
// partition on left
while (i <= j && (cmp = A[i].cmp(pivot)) <= 0)
{
if (cmp == 0) {
A.mark(p,4);
A.swap(i, p++);
}
++i;
}
// partition on right
while (i <= j && (cmp = A[j].cmp(pivot)) >= 0)
{
if (cmp == 0) {
A.mark(q,4);
A.swap(j, q--);
}
--j;
}
if (i > j) break;
// swap item between < > regions
A.swap(i++, j--);
}
// swap pivot to right place
A.swap(i,hi);
A.mark_swap(i,hi);
ssize_t num_less = i - p;
ssize_t num_greater = q - j;
// swap equal ranges into center, but avoid swapping equal elements
j = i-1; i = i+1;
ssize_t pe = lo + std::min(p-lo, num_less);
for (ssize_t k = lo; k < pe; k++, j--) {
A.swap(k,j);
A.mark_swap(k,j);
}
ssize_t qe = hi-1 - std::min(hi-1-q, num_greater-1); // one already greater at end
for (ssize_t k = hi-1; k > qe; k--, i++) {
A.swap(i,k);
A.mark_swap(i,k);
}
A.unwatch_all();
A.unmark_all();
QuickSortTernaryLR(A, lo, lo + num_less - 1);
QuickSortTernaryLR(A, hi - num_greater + 1, hi);
}