Re: [RFC] overflow safe scaling of 64bit values

2017-10-11 Thread Jan Hubicka
> > Any ideas for better version? If not I will go ahead with this variant and > > increase profile probability base. > > Why not use GCC wide int? Because I would like to keep operations fast on in common case where overflows do not happen. There was at least one case (the propagation of probabi

Re: [RFC] overflow safe scaling of 64bit values

2017-10-11 Thread Richard Biener
On Tue, 10 Oct 2017, Jan Hubicka wrote: > Hi, > in order to drop frequencies from basic blocks and counts from edges I need > to make probabilities more precise (so we do not get all those roundoff errors > from 1-base fixpoint arithmetics). Increasing base is easy now, but it > means that in

Re: [RFC] overflow safe scaling of 64bit values

2017-10-10 Thread David Edelsohn
> Any ideas for better version? If not I will go ahead with this variant and > increase profile probability base. Why not use GCC wide int? - David

[RFC] overflow safe scaling of 64bit values

2017-10-10 Thread Jan Hubicka
Hi, in order to drop frequencies from basic blocks and counts from edges I need to make probabilities more precise (so we do not get all those roundoff errors from 1-base fixpoint arithmetics). Increasing base is easy now, but it means that in temporaries one can get overflows easily. I need