2016-02-16 18:12 GMT+01:00 Tom Sydney Kerckhove
[...] As suggested by `thomie`, I created a design proposal at https://ghc.haskell.org/trac/ghc/wiki/Proposal/HelpfulImportError and am now looking for feedback.
[ Not sure if the feedback should be submitted here on in the corresponding ticket... ] Just a few quick remarks: * Whatever you do, never walk the file system tree up or down in an uncontrolled way, this will kill basically all benefits and is a show-stopper. File systems like NFS, NTFS, stuff on USB sticks etc. are so *horribly* slow when used that way that the walks will probably dominate your compilation time. And even under Linux it's not uncommon to have a few dozen directory levels and hundreds of thousands of files below our cwd: Just check out a few repositories, have some leftovers from compilations, tons of documentations in small HTML files etc., and this sums up quickly. Git walks up the tree, but only looking for a specific directory and will e.g. not cross mount points under normal circumstances. This is probably the limit of what you can do. * Caching between runs will be tricky: How will you invalidate the cache? People can (and will :-) do all kinds of evil things between runs, so how can you (in-)validate the cache quicker than re-scanning the file system again? * As a general rule to keep in mind during the design: Successful compiler runs should not pay a price. It's OK if things are a little bit slower when an error occurs, but the main use case is successful compilation. This is a bit like exceptions in most programming language implementations: They are more or less for free when you don't use them (yes, they have a cost even then because they complicate/invalidate some compiler optimizations, but let's forget that for now), and are often costly when you actually raise them. Just my 2c, S.