By: Travis (travis.downs.delete@this.gmail.com), April 23, 2017 3:31 pm
Room: Moderated Discussions
rwessel (robertwessel.delete@this.yahoo.com) on April 23, 2017 10:10 am wrote:
> OTOH, a certain density of branches *is* a realistic assumption. Current Intel "5 wide"
> decode of fused instructions is similar (cmp+branch being a common fused pair).
Intel rename is 4-wide in the macro-fused domain (i.e., at most 4 "things" can be renamed when counting a macro-fused test-branch pair as "1"), and it can reach this limit more or less in general (i.e., without some specific instruction mix).
It is 6-wide in the macro-unfused domain (i.e., at most 6 "things" can be renamed if you count a macro-fused test-branch pair as 2), but this isn't general since you can only reach it with an exact mix of 2 instructions + 2 fused test-branch pairs, at most one of which can be taken.
So it's most accurate to say it's 4-wide I think, with 6 coming second place. There isn't really a key way in which it is 5-wide.
> OTOH, a certain density of branches *is* a realistic assumption. Current Intel "5 wide"
> decode of fused instructions is similar (cmp+branch being a common fused pair).
Intel rename is 4-wide in the macro-fused domain (i.e., at most 4 "things" can be renamed when counting a macro-fused test-branch pair as "1"), and it can reach this limit more or less in general (i.e., without some specific instruction mix).
It is 6-wide in the macro-unfused domain (i.e., at most 6 "things" can be renamed if you count a macro-fused test-branch pair as 2), but this isn't general since you can only reach it with an exact mix of 2 instructions + 2 fused test-branch pairs, at most one of which can be taken.
So it's most accurate to say it's 4-wide I think, with 6 coming second place. There isn't really a key way in which it is 5-wide.
