Re: Simple benchmarks with gcc-3_0-branch

[Daniel Berlin <dberlin at redhat dot com>]

Daniel Berlin <dberlin@redhat.com> writes:
> 
> Right now, we only consider registers over their entire lifetime as
> candidates for allocation, which is basically the real cause of all
> the problems you are seeing.
> Combined with the fact that it's pretty conservative, too, and you end
> up with way too much stuff spilled.

Oh, and the first part of this isn't *quite* true, of course. I'm
simplifying.


Howevr, my point the problems you are specifically specifying should
be mainly taken care of in a register allocator, since LCM already
minimizes register pressure when compared to most other options.
Or so, the papers claim.
In fact, I can't find a single paper where register pressure is talked
about as something that needs to be solved by their particular pass,
except in scheduling and  register allocation papers (and one on lazy
code motion, and one on lazy strength reduction).
IE everything i see points to the fact that it's something to be
solved in register allocators, not by limiting optimization
opportunities. 

--Dan