-annot can result in modified .cmx

[nojb]

With trunk compiler:

$ echo "type t = A" > a.ml
$ echo "let f (_ : A.t) = object end" > b.ml
$ ocamlopt -c a.ml
$ ocamlopt -c b.ml
$ md5sum b.cmx
7bbe2afa5aacd6db86eccde09bbd2593  b.cmx
$ ocamlopt -annot -c b.ml
$ md5sum b.cmx
c791ac713e562561aecff8e655e77f01  b.cmx

[nojb]

The issue seems to have been introduced by c19e8b2 (#2127)

[nojb]

cc @lpw25

[lthls]

It seems that Clambda uses for its identifiers the type Ident, which includes a unique stamp. So if some previous compiler pass uses more (or less) identifiers, the Clambda stamps will differ.
I guess that passing -annot means that the compiler has to print some types, which could create new identifiers. (Flambda is not affected as it uses its own identifier type, with its own stamps.)

Normally producing different cmx files is not a problem, as it can happen for example if you use ocamlopt vs ocamlopt.opt (because of different sharing).
If this causes a problem in your workflow, please elaborate; the reported behaviour is not a bug in my mind.

[nojb]

It seems that Clambda uses for its identifiers the type Ident, which includes a unique stamp. So if some previous compiler pass uses more (or less) identifiers, the Clambda stamps will differ.
I guess that passing -annot means that the compiler has to print some types, which could create new identifiers. (Flambda is not affected as it uses its own identifier type, with its own stamps.)

Do you understand why this behaviour appeared with the indicated commit? Before that commit the compiler did not have this problem.

[lthls]

Unfortunately, no, the type-checker looks like black magic to me.

[nojb]

If this causes a problem in your workflow, please elaborate; the reported behaviour is not a bug in my mind.

We detected the problem in a slightly roundabout way, but should not be a big problem for us. Thanks for investigating.

[gasche]

Im someone wanted to investigate the issue: you don't need to build a .cmx to observe the issue with -annot, just run -dcmm or earlier -dclambda and observe that the identifier stamps are different between the two versions.

(Note: Printtyp uses Ident.rename in its hide function, this may be a source of different behavior from calling -annot, but I don't see how the commit pointed out by @nojb could affect this behavior.)