------- Comment #6 from rguenther at suse dot de 2009-04-14 12:24 ------- Subject: Re: data-flow analysis does not discover constant real/imaginary parts
On Tue, 14 Apr 2009, bonzini at gnu dot org wrote: > ------- Comment #5 from bonzini at gnu dot org 2009-04-14 12:16 ------- > Right, sorry -- it is only at -Os that we don't get it. VRP and DOM say that > the jump threading is not performed because the probability is too small, and > PRE does not run at -Os at all. We could add this to CCP if we were able to build partially varying constants, for example with abusing error_mark_node (ugh). So say that the PHI is really constant, but with value __complex__ (error_mark_node, 0) and then be very anal about the error_mark_nodes leaking somewhere ... The same issue exists with VECTOR_CSTs of course. I'll play with that idea. Richard. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39761
