On Tue, Dec 16, 2014 at 2:08 PM, Jan Hubicka <hubi...@ucw.cz> wrote: > Hi, > conversion to sreal makes it possible to compute badness in more streamlined > manner. Together with the sreal::normalize change this patch finally makes > fibheap badness calcualtion to be out of radar in profiling and I hope it > makes it more maintainable by eliminating the issues from roundoff errors and > overflows that was quite painful. > > Incrementally it is also possible to turn time computation into sreals. > > Bootstrapped/regtested x86_64-linux, wll wait with the commit for the > preivous changes to show up in the benchmark testers. > > Thanks for Martin and Trevor for working on sreals&fibheap changes. > > Honza > > * ipa-inline.c (cgraph_freq_base_rec, percent_rec): New functions. > (compute_uninlined_call_time): Return sreal. > (compute_inlined_call_time): Return sreal. > (big_speedup_p): Update to be computed with sreals. > (relative_time_benefit): Turn to sreal computation; return > value as numerator/denominator to save division. > (edge_badness): Rewrite to sreals; remove overflow checks > and cleanup. > (ipa_inline): Initialize cgraph_freq_base_rec and percent_rec. > (inline_small_functions): Update dumping; speedup fibheap maintenance. > (update_edge_key): Update dumping.
This caused: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64694 -- H.J.
