3 ms·

So I think the inlining decisions in a compiler are some of the most heuristic-y parts of the compiler. There was a paper from Jeff Dean and some other Googler

So I think the inlining decisions in a compiler are some of the most heuristic-y parts of the compiler. There was a paper from Jeff Dean and some other Googlers recently that hinted that machine learned models might be the next leap forward in compiler technology, as a potential means of replacing these heuristics.

rurban 8y ago
Machine learning? If some logic, than precise and reproducible please. Prolog would be nice, not some neural nets. gcc already has some nice lisp. Jeff Dean usually has good ideas, but not this one. ML would be more useful for a JIT, dynamic optimization problems at run-time. But the problem is only the botched architecture, with non-manageable interdependencies. Of course inlining is the mother of all optimizations, most other optimization should be applied after inlining. But for inlining to be effective it has to run after constant folding. And then after inlining you have to run most optimization passes again to get the real benefits. This problem is much simpler: Missing as-integration and a botched constexpr API. They should just take it from llvm.