Re: [PATCH] stdlib: Optimize number of calls to comparison function

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 27/03/24 16:59, Kuan-Wei Chiu wrote:
> On Wed, Mar 27, 2024 at 04:45:35PM -0300, Adhemerval Zanella Netto wrote:
>>
>>
>> On 16/02/24 04:08, Kuan-Wei Chiu wrote:
>>> On Sun, Dec 03, 2023 at 05:48:39AM +0800, Kuan-Wei Chiu wrote:
>>>> The current heapsort implementation in the siftdown function follows
>>>> the standard textbook version, necessitating two comparisons at each
>>>> level. Transitioning to the Bottom-up heapsort version allows us to
>>>> decrease the required comparisons to just one per level. On average,
>>>> this modification significantly reduces the comparison count from
>>>> 2nlog2(n) - 3n + o(n) to nlog2(n) + 0.37*n + o(n).
>>>>
>>>> Refs:
>>>>   BOTTOM-UP-HEAPSORT, a new variant of HEAPSORT beating, on an average,
>>>>   QUICKSORT (if n is not very small)
>>>>   Ingo Wegener
>>>>   Theoretical Computer Science, 118(1); Pages 81-98, 13 September 1993
>>>>   https://doi.org/10.1016/0304-3975(93)90364-Y
>>>>
>>>> Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
>>>> ---
>>>>  stdlib/qsort.c | 36 ++++++++++++++++++------------------
>>>>  1 file changed, 18 insertions(+), 18 deletions(-)
>>>>
>>>> diff --git a/stdlib/qsort.c b/stdlib/qsort.c
>>>> index be01fb5598..f5babef150 100644
>>>> --- a/stdlib/qsort.c
>>>> +++ b/stdlib/qsort.c
>>>> @@ -132,26 +132,26 @@ static inline void
>>>>  siftdown (void *base, size_t size, size_t k, size_t n,
>>>>  	  enum swap_type_t swap_type, __compar_d_fn_t cmp, void *arg)
>>>>  {
>>>> -  /* There can only be a heap condition violation if there are
>>>> -     children.  */
>>>> -  while (2 * k + 1 <= n)
>>>> -    {
>>>> -      /* Left child.  */
>>>> -      size_t j = 2 * k + 1;
>>>> -      /* If the right child is larger, use it.  */
>>>> -      if (j < n && cmp (base + (j * size), base + ((j + 1) * size), arg) < 0)
>>>> -	j++;
>>>> -
>>>> -      /* If k is already >= to its children, we are done.  */
>>>> -      if (j == k || cmp (base + (k * size), base + (j * size), arg) >= 0)
>>>> -	break;
>>>> +  size_t i, j;
>>>> +
>>>> +  /* Find the sift-down path all the way to the leaves. */
>>>> +  for (i = k; j = 2 * i + 1, j + 1 <= n;)
>>>> +    i = cmp (base + j * size, base + (j + 1) * size, arg) >= 0 ? j : (j + 1);
>>>>  
>>>> -      /* Heal the violation.  */
>>>> -      do_swap (base + (size * j), base + (k * size), size, swap_type);
>>>> +  /* Special case for the last leaf with no sibling. */
>>>> +  if (j == n)
>>>> +    i = j;
>>>>  
>>>> -      /* Swapping with j may have introduced a violation at j.  Fix
>>>> -	 it in the next loop iteration.  */
>>>> -      k = j;
>>>> +  /* Backtrack to the correct location. */
>>>> +  while (i != k && cmp (base + k * size, base + i * size, arg) > 0)
>>>> +    i = (i - 1) / 2;
>>>> +
>>>> +  /* Shift the element into its correct place. */
>>>> +  j = i;
>>>> +  while (i != k)
>>>> +    {
>>>> +      i = (i - 1) / 2;
>>>> +      do_swap (base + i * size, base + j * size, size, swap_type);
>>>>      }
>>>>  }
>>>>  
>>>> -- 
>>>> 2.25.1
>>>>
>>>
>>> Since we still retain heapsort as a fallback for mergesort, should we
>>> reconsider applying this patch?
>>>
>>> Thanks,
>>> Kuan-Wei
>>
>> Yes, I think it is work and it looks great to me.
> 
> Should I resend a new patch, or can you directly apply this one?

It applies as-is, so it should be ok. I will run some tests and apply, thanks.