none (none.delete@this.none.com) on August 14, 2012 6:56 am wrote:
[snip]
> In theory yes, predication looks great.

Maybe not great, but certainly useful.

>In practice, as I said in another post, it might constrain too
> many other things in your micro-architecture, which actually might reduce
> achievable performance. I really think spending the budget transistor and
> design specification in doing a smarter branch predictor is better.

I do not mean to imply that the ARM architects were ignorant or stupid. However, I think providing conditional stores is more attractive than forcing software to implement them by conditionally selecting a store address. An implementation of predicated stores would only have to be equal to the software workaround (and such could probably be implemented much like the software version by cracking the store into an address generation µop and a data µop where the address generation could produce a null address--which would not match with any loads and would commit the store data to no-place--No Big Deal [maybe]).

Wish branches (that inform hardware of a hammock that is likely to benefit from predication) might be better (moving such into the microarchitecture--which might treat such as an ordinary branch, predicate if "sufficiently short", or make a dynamic choice on handling based on prediction confidence information--while still allowing software to communicate information that could avoid expensive mispredictions), but predicated stores (in my ignorance) do not seem especially limiting.