Tweak the PLT trampoline to avoid generating textrels in LibC. This allows us to share all the LibC mappings, reducing per-process memory consumption by ~200 KB. :^) Patch originally by @nico.