Stefan Sperling:

> Alternatively, as quick fix you could try whether qsort is faster than
> mergesort on average. But I don't have much hope that this helps much.

It's actually slower in my specific test case.

Baseline (mergesort):
   12m35.32s real    12m03.32s user     0m32.00s system

qsort:
   19m29.20s real    19m02.10s user     0m26.89s system

>  rewrite parse_tree_entry() to avoid use of memchr and strnlen

That has negligible effect:
   12m22.08s real    11m49.87s user     0m32.16s system

-- 
Christian "naddy" Weisgerber                          naddy@mips.inka.de