caml_output_value_to_malloc: really use malloc

[gasche]

External users (unison, hhvm, pyre-check) assume that the memory they get from caml_output_value_to_malloc can be freed with free. (See #12491 (comment) for some user codebase examples .)

The function was changed to use caml_stat_alloc in 4.06 ( 02a8b99 ); this breaks user code, as one should then use caml_stat_free.

There is no use of caml_output_value_to_malloc in the runtime itself, so we do not need to provide a caml_output_value_to_stat_alloc variant.

[abbysmal]

Change looks good to me: I assume it was changed by mistake in 02a8b99 when grepping for malloc occurrences?

[abbysmal]

(Changes entry needs update, will merge afterward.)

[gasche]

Thanks! I rebased the PR.

[xavierleroy]

OK for this change: ownership of the allocated block is transferred to 3rd-party C code, so the automatic deallocation that can occur with caml_stat_alloc + cleanup-at-exit mode is not right.

Symmetrically, there's caml_input_value_from_malloc, which currently uses caml_stat_free to free the buffer passed to it, implying that is must be allocated with caml_stat_alloc. The way it's used in OCamlMPI currently is inconsistent (allocation is done sometimes with caml_stat_alloc, sometimes with malloc).
\

[gasche]

Ah, good catch, I didn't know about that function and will change it as well.

[xavierleroy]

Ah, good catch, I didn't know about that function and will change it as well.

As you'll see it's a bit more difficult...

[gasche]

I pushed a new commit to use malloc/free to handle the intern_input of input.c, which in particular prevents a bug where we would call caml_stat_free on the user-provided input from caml_input_value_from_malloc.

[xavierleroy]

LGTM! That was easier than I feared...

[gasche]

Thanks both, this is now merged.