RE: [Caml-list] Justifying a breaking 4.03 change, strong dependency on modules with "external" declarations (was: )

[David Allsopp <dra-news@metastack.com>]

Hongbo Zhang wrote:

> My 2 cents: I did not like this change, for modules which has side effect, it should always
> be explicitly listed in the link line, instead of relying on such fragile implicit behavior.

I think Gabriel's explanation has missed an important subtlety of the original problem. One of the reasons you might "omit" the .cmx file when linking is because you thought you'd given it, because the .cmx file was in a .cmxa file which you had linked (or more likely, one which ocamlfind was adding thanks to -package and -linkpkg instructions...).

Even worse, even if you did specify Bar.cmx, the linker did nothing to ensure that you'd placed it *before* the first call to an external value - meaning that the primitive could be called before the required initialisation code.

The only workarounds were to use val instead of external (giving both a slight performance penalty and forcing a .mli file to be written - although many, me included, think .mli files should be mandatory anyway!) or to use -linkall, which is not necessarily feasible.

I remember wasting a lot of time the first time I got stung by that... even though it is in the manual. In fact, that means the first paragraph of the manual needs updating!


David


PS I noticed that the old behaviour still happens with normal primitives (e.g. %identity), but I guess that's intentional?