Re:Tracing JITs and modern CPUs
by TheLoneWolfling on Saturday July 25, 2015 @06:43PM
(#9949333)
Two semi-tangential ideas:
I wish it was possible to do things along the lines of "fetch this data from RAM and at the same time start recomputing it, whichever is faster".
And also, I wish there was a "branch upcoming / branch execute" split. In other words, instead of just "here's a branch", it's more along the lines of "you will have to branch on <x>. <other instructions> now branch on <x>." Effectively a variable number of branch-delay slot(s). Wouldn't help for straight pointer-chasing, but in other situations it could.
I wish it was possible to do things along the lines of "fetch this data from RAM and at the same time start recomputing it, whichever is faster".
And also, I wish there was a "branch upcoming / branch execute" split. In other words, instead of just "here's a branch", it's more along the lines of "you will have to branch on <x>. <other instructions> now branch on <x>." Effectively a variable number of branch-delay slot(s). Wouldn't help for straight pointer-chasing, but in other situations it could.