jbuilder uses timestamps only for change detection

[mmottl]

It happens sometimes that a developer edits a file, saves it, and then quickly undos the change followed by another save before building. jbuilder will perceive the file as changed regardless of contents, because it apparently just uses timestamps to detect changes. This problem can also happen if, for example, an intermediate file is equivalent to the previous version, maybe because a change didn't affect code generation.

I think it would be better to hash all input files and (intermediate) targets to reduce build times in these cases. Other build tools support this, too, e.g. ocamlbuild, and probably omake. Hashing is usually very much cheaper than any compilation step. An unchanged timestamp should, of course, be interpreted as "no change" to avoid hashing all files all the time.

[jaredly]

Would also be nice to diff at the AST level as a second early bailout

[ghost]

There a re a couple of things to sort to make this work:

1. change the way aliases are handled; currently creating an alias essentially does touch <some-file>. They would now have to produce a file whose contents reflects the contents of the dependencies of the alias
2. decide how to checksum files. We can't rely on external system tools as this is not portable, and calling Digest.file synchronously for every file is likely to be noticeable on large trees. We could create a auxiliary tool for this

I'm also thinking of adding a bridge with jenga so that we can use jenga with jbuilder projects. This will solve this issue as jenga already handles all this properly.

[ghost]

Jbuilder is now hashing files