Use TLS for big temporary structures

[Oppen]

Given that we are already not reentrant (malloc can't be used in signal handlers and we make extensive use of it), nor can I think of reasonable scenarios where we would want that, and that we avoid recursion as much as we can (even if not explicitly a goal, it just happens to be the case), we may avoid many allocations by just using thread local storage globals (or statics if we want cleaner code) for them. I think it's worth benchmarking to see whether it makes sense. I'm thinking mostly of the places where we would store a bitmap for mass operations.

[lemire]

Can you elaborate? We definitively can use thread local globals, but how do you propose coding this up?

[Oppen]

I'd need to check the code again to look for places where it would be useful, but I was thinking mostly of lazy operations or those working on multiple bitmaps at a time that may use a container as an accumulator and then convert it to whatever is appropriate depending on its cardinality. The temporary bitmap container could be recycled by sharing a TLS global. I just thought of asking first rather than come up with code that wouldn't have been accepted in any variant (as opposed to regular back and forth to improve the patch).

[Oppen]

BTW, sorry for the long delay to answer, I was actually on vacation and wrote the issue just so I wouldn't forget about it.

[lemire]

There is no doubt that we are sometimes limited by memory allocation. I think it is worth pursuing.