>What I'm now puzzled about is why just exercising the right kind of
>sort didn't exhibit the slowdown.
I wonder if the size of your sorted dataset ( i.e. all 1000000 rows) is
the reason - too big to fit into sort_mem, so that temporary files are
needed. The resulting file management *may* have obscured the difference
in sort speed - although that does not explain how your resuts were
consistently *faster* for the Solaris qsort.
regards
Mark