Hello Sergey,
> The option sorting in argp-help was based on the stability
> of sort algorithm, therefore the test failed on systems with
> quicksort qsort implementations. It went unnoticed on glibc,
> because its qsort uses adaptive algorithm, defaulting to
> insertion sort if the number of partitions is smaller than a
> predefined value, which happens to be true for the argp tests.
Indeed, with your latest changes, the argp tests pass on:
- glibc 2.3.6, IRIX 6.5 (where it failed before)
- MacOS X 10.3.9, OSF/1 4.0d (not tested last time).
> Use argument addresses to ensure they never compare as equal.
However, I think this is not portable. While earlier you were assuming
a stable sorting algorithm, you are now assuming an in-place sorting
algorithm. If a libc were to implement qsort by use of mergesort (and
temporary storage), such a sorting algorithm would pass you two pointers
to hol_entrys in different array. Then HOL_ENTRY_PTRCMP's result depends
on the temporary storage allocation strategy.
To make this portable, I see two options:
- Sort an array of 'hol_entry*' instead of an array of 'hol_entry';
then you can be sure the hol_entrys stay in place. Or
- Add a field 'array_index' to the hol_entry struct. Fill these fields
right before calling qsort.
Bruno
