------- Comment #8 from rguenther at suse dot de 2007-11-28 16:13 ------- Subject: Re: [4.3 Regression] VRP/SCEV miscompiles Firefox
On Wed, 28 Nov 2007, rakdver at kam dot mff dot cuni dot cz wrote: > > ------- Comment #6 from rguenth at gcc dot gnu dot org 2007-11-28 15:47 > > ------- > > We could clear the SCEV cache for an SSA_NAME we set a new value range > > (in set_value_range and set_value_range_to_varying), but I > > see that SCEV also caches loop->nb_iterations which we probably would need > > to > > clear unconditionally (for all loops). > > yes, the simplest solution is to call scev_reset somewhere (either in > set_value_range, or in adjust_range_with_scev). However, I fear the > compile time impact would be too large (# of iterations analysis is > fairly time consuming, and not caching the results used to cause > us to spend 10% or more in it). > > More feasible solution seems to be to determine the number of iterations > for all loops before the start of VRP, and to remember the values. Or not use tree_expr_nonnegative_p from inside tree-ssa-loop-niter.c which is where the bad feedback comes from. But yes, computing the number of iterations in advance should fix this as well and is probably best. Richard. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34244
