On Mon, Jan 22, 2018 at 01:11:34PM +0100, Christophe Lyon wrote: > On 19 January 2018 at 22:45, Jakub Jelinek <ja...@redhat.com> wrote: > > That IMHO can't work and the Invalid sum message verifies that. If we want > > the first jump to hit 99times more often than the second one or vice versa, > > I believe first_prob should be .99 * prob or .01 * prob respectively, and > > the second probability then should be (0.01 * prob) / (1 - 0.99 * prob) > > or (0.99 * prob) / (1 - 0.01 * prob) respectively. > > > > With this change the Invalid sum message disappears. > > predict-8.c testcase was apparently trying to match the hardcoded 0.99 > > probability rather than .99 * 65% emitted now. > > > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > > Hi, > > Did you notice any performance impact?
Haven't been looking for that, this is a bugfix, the current code is broken. It changes (fixes) probabilities of some edges, sure, so it could affect something both ways. Jakub