Re:Dataflow Analyses and Compiler Optimizations That Use The...
by moonchild on Sunday April 21, 2024 @07:39PM
(#40110168)
> repository of analyzed code
yes, this is obviously what you want. there shouldn't be a notion of 'running' the compiler; it should simply be a persistent process. importantly, there are two issues, which are orthogonal but both critically important: iterativity—that is, produce an acceptable result immediately, but then continue to refine it—and incrementality—that is, when making a small change to the code, take advantage of information learned about the previous version.
there is this recent paper on incrementality https://arxiv.org/pdf/2104.01270.pdf, though it has some caveats (e.g. it does not handle iterativity at all, so there is only so far it can go; can do much better)—but it is interesting non-technically because the authors work for amazon, so perhaps some of this is headed for industry
yes, this is obviously what you want. there shouldn't be a notion of 'running' the compiler; it should simply be a persistent process. importantly, there are two issues, which are orthogonal but both critically important: iterativity—that is, produce an acceptable result immediately, but then continue to refine it—and incrementality—that is, when making a small change to the code, take advantage of information learned about the previous version.
there is this recent paper on incrementality https://arxiv.org/pdf/2104.01270.pdf, though it has some caveats (e.g. it does not handle iterativity at all, so there is only so far it can go; can do much better)—but it is interesting non-technically because the authors work for amazon, so perhaps some of this is headed for industry