Speed up monster action planning

[FeepingCreature]

SUMMARY: Performance "Speed up monster action planning."

Noticed that sleeping in the lab was a bit slow, so I attached a profiler to see what was going on.

With GNU libstdc++, shared_ptr<> will always use threadsafe operations when linked against libpthread at all. This is not necessary in CDDA, so replace shared_ptr<> with shared_ptr_fast<> and weak_ptr<> with weak_ptr_fast<>, which is redefined to use libstdc++'s internal single-threaded version.

After that, I noticed that calculating distances in monster::rate_target() was eating up a lot of time to compute trig distances. rate_target doesn't need the actual distance until a way's in, so if we inline rl_dist and allow it to defer the sqrt until later on by adding rl_dist_squared and comparing with the square of best instead, we can make up a bunch more speed.

[ZhilkinSerg]

Do you have a savegame to profile a performance increase (or maybe performance test results as text/screenshots)?

[FeepingCreature]

I should probably do that.

update: impromptu bench timing just sitting around for six hours in my lab cell says this PR probably doesn't do enough to be worthwhile: 87s -> 79s. Gonna try to remove the unwieldy first commit to see how much is the _squared change.

update:

version	time
before	87s
rl_dist_fast	83s
rl_dist_fast + shared_ptr_fast	79s

perf shows lock() eating 15% of cpu while waiting, which fits - non-atomic lock isn't free, but it skips the 40%-of-15% lock cmpxchg.

I think it's still good as a PR. Free 5% twice is nothing to sneeze at.

edit: One important thing to add would probably be making the rl_dist_squared calls fast in the non-trig case by skipping the square/sqrt hack in that case.

edit: Done! rl_dist_squared is now rl_dist_fast and returns a wrapper that automatically calls sqrt() for the int conversion (iff trigdist is set).