Re: Thoughts and Opinions or something like that

[njn@cs.mu.oz.au (Nicholas James NETHERCOTE)]

jmv16@cornell.edu (Jeffrey M. Vinocur) writes:

>Tarjei T. Jensen <tarjei.jensen@kvaerner.com> wrote:
>>
>>As far as instructions are concerned I always fancied deviding branch
>>instructions into two groups: Those who branch most of the time and those
>>who seldom branch. The idea was that one could skimp on branch prediction on
>>the CPU. For a while or for loop one would use the former type of
>>instructions. Since I have heard of nobody who have done this I assume that
>>the idea is not particularly good.

>Well, the goal isn't to lessen branch prediction -- the silicon
>is there either way -- but rather to reduce mispredicted
>branches.  The most basic approach, I guess, is assume all
>backwards branches will be taken -- this gets loops.  One
>not-very-complicated approach is to cache whether a branch was
>taken last time through and predict the same.  This approach,
>however, does exactly what you suggest; it gets loop iterations
>2,3,4,...,n-1 all correct.

I remember seeing a suggestion somewhere for a pragma in a C-- implementation
that allowed you to estimate the probability that a condition succeeded.  So it
sounds like it must be a worthwhile idea, assuming your estimates are accurate.

--
Nick Nethercote
njn[AT]cs.mu.oz.au