This is the mail archive of the
mailing list for the GCC project.
Re: Combine four insns
On 08/06/10 14:43, Steven Bosscher wrote:
The specific cases I've seen won't combine at the gimple level. For
example, if you look at the H8, you need to have a post-inc memory
operand which I really doubt we're going to ever expose at the gimple level.
If it's patterns we could combine at the GIMPLE level, it wouldn't be
so silly to handle it there. Weren't you once one of the people
talking about a tree-combine pass?
I still think a tree combiner would be a great thing to have as it would
eliminate tree-ssa-forwprop with a more generic engine.
Again, I don't think the majority of the cases where 4->1 or 4->2
combinations occur are going to be anything that we'd expose that early
in the pipeline.
Or if this is a case the compiler could already expand from GIMPLE to
good initial RTL (i.e. TER), that wouldn't be a silly place to put it