> Hi,. > > this patch converts the variables that hold time benefits and > frequencies in IPA-CP from plain integers to sreals, avoiding the need > to cap them to avoid overflows and also fixing a potential underflows. > > Size costs corresponding to individual constants are left as ints so > that they do not take up too much space. Care must be taken that > adding it up does not overflow, especially in the case of > prop_size_cost, because in cases of extremely long chains of lattice > dependencies it can overflow (e.g. in testsuite/gcc.dg/ipa/pr50744.c). > The overall size is already tracked in long ints. > > Bootstrapped, LTO-bootstrapped and tested on x86_64-linux, OK for trunk? > > Thanks, > > Martin > > > gcc/ChangeLog: > > 2020-11-11 Martin Jambor <mjam...@suse.cz> > > * ipa-cp.c (class ipcp_value_base): Change the type of > local_time_benefit and prop_time_benefit to sreal. Adjust the > constructor initializer. > (ipcp_lattice::print): Dump sreals. > (struct caller_statistics): Change the type of freq_sum to sreal. > (gather_caller_stats): Work with sreal freq_sum. > (incorporate_penalties): Work with sreal evaluation. > (good_cloning_opportunity_p): Adjusted for sreal sreal time_benefit > and freq_sum. Bail out if size_cost is INT_MAX. > (perform_estimation_of_a_value): Work with sreal time_benefit. Avoid > unnecessary capping. > (estimate_local_effects): Pass sreal time benefit to > good_cloning_opportunity_p without capping it. Adjust dumping. > (safe_add): If there can be overflow, return INT_MAX. > (propagate_effects): Work with sreal times. > (get_info_about_necessary_edges): Work with sreal frequencies. > (decide_about_value): Likewise and with sreal time benefits.
OK, thanks! It is bit odd that we work hard enough to overflow sizes, but I guess it is just an extreme case. Definitly capping there makes sense since we do not want such large duplication. Honza
