Re: [pure-lang-users] Patch to compile Pure with LLVM upstream

[Jiri S. <jir...@bl...>]

Albert Graef wrote:
> I found that adding special case rules to avltree.pure only yields
> marginal improvements (<10%), so you might be able to remove them from
> your implementation, too, without taking a major performance hit. That
> alone will make your modules load much faster, and there's some other
> refactoring that we can do (like checking the balance factor in a guard
> instead of matching it on the lhs) that will further reduce the number
> of overlapping rules.
> 
OK. Testing set is now in 'examples/set_test.pure'.

Removed type annotations don't hurt.

I changed pattern matches of constants to guards where possible. The 
result does not seem to compile any faster (at least on my PC) and the 
execution is substantially slower :-( .

Do you see any other possible rewrites of the code?

I looked at other similar implementation of AVL trees (I already 
mentioned it somewhere in the Q discussion). It doesn't use those 
'table' and 'table2' to calculate the new balances, there is some 
div/mod/&/| magic instead. The rest of the code is almost the same. This 
variant of the algorithm was only by about 30 % faster than Bird/Wadler 
in Q. I don't think it would compile much faster than the current one.

Jiri