I didn’t mean to explicitly ask about __builtin_expect. I only meant to reference it as something with a similar function. Articles on the net say programmers are more likely to mispredict frequencies unless they use the proper tools in specific circumstances.
My intent was to suggest using a similar annotation or instruction to inform a cold predictor. Being in a base state, a programmer can probably have a reasonable idea about the likely code path. And then once the predictor is warmed up, the annotation should be ignored.
Of course support for __builtin_expect is nice. But I think that something like __builtin_cold_expect would be generally more practical. Once things are primed, I wouldn’t want to second guess predictors that do 90+%.
As far as entropy cost, it is understable that an additional instruction might be too costly. What about an extra, optional or unused bit on a branch instruction?
Another related question. When speculating on a branch, might there any added benefit to knowing the probability within a granularity greater than 50% accuracy? It seems like if the decision were not simple and a weight could be applied, e.g. one branch had a more expensive subsequent instruction than the alternative, this too might be useful. But I’m probably speculating too much myself. I’ll stop this line of questioning as I give it a probability of .05 of being potentially useful.
- This reply was modified 9 years, 6 months ago by casten.