------- Comment #3 from rguenther at suse dot de 2007-12-10 12:32 ------- Subject: Re: Tree optimization pipeline needs re-tuning
On Mon, 10 Dec 2007, steven at gcc dot gnu dot org wrote: > Insert usual rant about running more and more and more and more and more and > more and more and more and more and more and more passes, with total disregard > for compilation time of the common cases where optimization doesn't do > anything > but increasing compile time. > > Before adding more passes, add some mechanism/heuristics to only run passes > when benefit is expected. Not the whole world is tramp3d. Please take glibc > and the linux kernel (or other codes with almost no abstraction) as compile > time indicators for a change, too. :) Apart from the early loop unrolling which has benefits for Fortran code, I was only suggesting to re-order things, not to add additional passes. But yes, adding passes is usually far easier to make regression-free from a runtime performance point. And removing anything from GCC is at least 10 times harder than adding something - which you know, I bet ;) Richard. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34416
