News for hackers. Stuff that matters.

A comment on Vectorized Emulation: fuzzing at 2 trillion instructions per second. Parent comment

Re:Vectorized Emulation: fuzzing at 2 trillion instructions ... by vnorilo on Wednesday October 17, 2018 @07:51AM (#18238083)
The real problem is when different VMs branch differently. You essentially have to run both sides of a branch in serial with masked lanes. The author mentions this and states that the method works only because the VMs are very deterministic. I don't understand how he deals with indirect jumps.
[ Parent | on HN ]
Re:Vectorized Emulation: fuzzing at 2 trillion instructions ... by DSingularity on Wednesday October 17, 2018 @08:47AM (#18238408)
You mean the possibility that the branch target is different on the different VMs? He can fork the vectorization if the targets diverge. I’m not sure what he does though.
[ Parent | on HN ]
Re:Vectorized Emulation: fuzzing at 2 trillion instructions ... by vnorilo on Wednesday October 17, 2018 @11:30AM (#18240017)
Sure, but how does that not utterly destroy performance as the fork would probably lead to multiple sub-forks down the line.. or lane? :)
[ Parent | on HN ]
Re:Vectorized Emulation: fuzzing at 2 trillion instructions ... by DSingularity on Wednesday October 17, 2018 @02:27PM (#18242059)
When you are fuzzing the ultimate goal is to explore as much as you can of the entire search space. Forks cant hurt performance. They are just another potential execution path to be vectorized.
[ Parent | on HN ]
Re:Vectorized Emulation: fuzzing at 2 trillion instructions ... by OskarS on Wednesday October 17, 2018 @10:48AM (#18239582)
He writes that he's using a special type of IL for this technique, presumably he forbids indirect jumps in it.
[ Parent | on HN ]