Just a quick remark while skimming through my impressive mail backlog: While splitting this specific phase into 2 parts might solve our problem at hand, I think the overall direction should be even more general, especially in the light of potentially parallel compilation, which would allow more expensive and clever things. I imagine something like e.g. GHC's compilation pipeline, which is very flexible and the optimizer has several passes with some ordering constraints which are repeatedly called, controlled by heuristics, some budgets and even compiler pragmas.
Not something for the short term, but something we should keep in mind to avoid hardwiring even more special cases than we do already now... On Fri, May 18, 2012 at 6:12 PM, <[email protected]> wrote: > another idea: How about splitting Canonicalize phase into two (e.g. > CanonicalizeEarly and CanonicalizeLate)? There are potentially more things > that > depend on later optimization passes and it would be more a general > solution than > patching up the range analysis. [...] > -- v8-dev mailing list [email protected] http://groups.google.com/group/v8-dev
