> On Fri, Aug 31, 2012 at 7:24 PM, Martin Jambor <mjam...@suse.cz> wrote: > > Hi, > > > > On Thu, Aug 30, 2012 at 05:11:35PM +0200, Martin Jambor wrote: > >> this is a new version of the patch which makes ipa analysis produce > >> predicates for PHI node results, at least at the bottom of the > >> simplest diamond and semi-diamond CFG subgraphs. This time I also > >> analyze the conditions again rather than extracting information from > >> CFG edges, which means I can reason about substantially more PHI > >> nodes. > >> > >> This patch makes us produce loop bounds hint for the pr48636.f90 > >> testcase. > >> > >> Bootstrapped and tested on x86_64-linux. OK for trunk? > >> > >> Thanks, > >> > >> Martin > >> > >> > >> 2012-08-29 Martin Jambor <mjam...@suse.cz> > >> > >> * ipa-inline-analysis.c (phi_result_unknown_predicate): New function. > >> (predicate_for_phi_result): Likewise. > >> (estimate_function_body_sizes): Use the above two functions. > >> > > > > This patch, on top of the one doing loop calculations almost always, > > introduces a number of testsuite failures which somehow I had not > > caught during my testing. The problem is that either > > calculate_dominance_info or loop_optimizer_init introduce new SSA > > names for which there is no index in nonconstant_names which is > > allocated before the dominance and loop computations. I'm currently > > bootstrapping and testing the following fix which simply allocates the > > vector after doing the two computations. If it passes I will commit > > it straight away so that the regression is fixed before I leave for > > the weekend, I hope it's obvious enough for that. > > > > On the other hand, it would really be better if we did not change > > function bodies during IPA summary generation phase... > > Um ... we shouldn't do this. Can you track down where it happens? I > suppose it might come from CFG manipulations loop_optimizer_init > performs when not passing AVOID_CFG_MODIFICATIONS.
I bet it come from loop noromalization :) (i.e. loop closed form or preheader construction both needs new SSA names.) I think it would be best to make pass manager to handle this and make loop normalization to happen once before all SSA IPA analysis Honza
