Effects / double translation: keep more functions in direct style

[vouillon]

We don't require escaping functions nor functions called from a CPS call site to have a CPS variant. This makes it possible to reduce the amount of code duplication.
This implies that CPS call sites have to be able do deal with direct style functions. This makes CPS calls a little bit slower, but we don't care too much about performance in this case.

[hhugo]

@vouillon, do you want to merge this before a release ?

[vouillon]

@vouillon, do you want to merge this before a release ?

No, this can be merged later.

[OlivierNicole]

I’m looking at this, but I think I need to reflect a bit more to understand this statement:

We don't require escaping functions nor functions called from a CPS call site to have a CPS variant.

[OlivierNicole]

This looks good to me! I only have a question on an unrelated point.

[OlivierNicole]

Here’s how this PR affects the (bzip2-ed) code size increase factor induced by --effects=double-translation compared to --effects=cps:

Revision	master (3d645b9)	this PR
ocamlc (doesn't use effect handlers)	161 %	155 %
generator.ml	111.3 %	111.0 %

On generator.ml which is a small program using effect handlers, the effect is very small.

[OlivierNicole]

If the calling convention can be chosen dynamically, doesn’t it open the way to being able to link together files compiled with --effects=cps and files compiled with --effects=double-translation?

[vouillon]

If the calling convention can be chosen dynamically, doesn’t it open the way to being able to link together files compiled with --effects=cps and files compiled with --effects=double-translation?

With --effects=double-translation , we expect that there will always be a direct style version of all functions, which we can call directly. This is not the case with --effects=cps.