Hacker News Reader

It's definitely still an active project, but I'm on sabbatical this academic year and have had much less time than usual for my regular research. I'll be getting back to it in June. Also, Raimondas Sasnauskas, who implemented the instruction synthesizer, finished his postdoc position and moved on.

In any case, there's been a lot of progress since I did that UW talk, for example here are some early synthesis results:

http://blog.regehr.org/archives/1252

View on HN

That's great to hear! It is a very interesting project

Can you explain why your goal is to implement these optimizations in llvm? Why not have Souper as one of the optimization passes?

It turns out it's harder than you'd think to decide whether or not an optimization is a good idea. Of course there are some relatively simple optimizations that are obviously good, but a lot of those have already been implemented. Also, I'm not sure how many people are going to want to add a solver into their compilation path.

On the other hand, if we simply contribute optimizations then all LLVM users benefit.

We're also using Souper to reveal imprecisions in LLVM dataflow analyses such as known bits -- this turns out to work really well.

Ahh, I see. I assumed that any and all optimizations are good, that makes sense.

Keep up the good work!

Reply on news.ycombinator.com