Added Changes entry.

It would be nice to have a review by someone familiar with assembly (I'm not competent here) and the general backend implementation/style; @bschommer, can you vouch for the patch correctness? Otherwise maybe @alainfrisch, @xavierleroy?

I am wondering if the same change is necessary in the other alloc* functions (eg the comment at https://github.com/ocaml/ocaml/blob/trunk/asmrun/amd64.S#L387).

To expand on my previous comment, it seems that all the functions alloc* are invoked via call so their stack is 8-aligned on entry (due to call pushing the return address). Therefore it should be made 16-aligned before invoking caml_call_gc.

(On closer reading, the code seems OK as it stands - the macros ENTER_FUNCTION and LEAVE_FUNCTION take care to 16-align the stack on the functions alloc1, alloc2 and alloc3).

I can confirm what @christoph-cullmann said. We now have the patch applied for about a week and have not encountered any problems on any of our internal-tests so far.

Yes, at all other places I think the stack is properly aligned before we issue the call instruction (and if I don't misread the ABI docs, it must be 16 byte aligned before the call instruction, not counting the effect of the call itself)

I went ahead and merged, thanks.

Thanks, I pinged Xavier if he has any feedback.
Hope that makes --with-frame-pointers work more reliable.

I'm always late to the party, sorry about that. But I'm OK with the change. It is a very good idea to always conform to the ABI, indeed!