Clean up ret_mode and region in Lambda.lfunction

[mshinwell]

...and consequentially contains_no_escaping_local_allocs in Flambda 2, which was unused.

Simplif seemed to be the only place relying on this field. @riaqn suggested that ret_mode could be used instead, which seems like it should be correct, but this part of the diff needs checking carefully. The rest is straightforward.

[riaqn]

LGTM - thanks!

Please hold off merging - I'd like to push commit to clean up the comments.

[mshinwell]

I'd like @lpw25 to read the changes to Simplif (not commit-by-commit, read the whole at once) and the changes to the comments. Otherwise this is ok. Thanks.

[riaqn]

@lpw25 and I talked, and we think we want this PR to fix things more thoroughly:

• We will remove ret_mode and region from lfunction.
• We add a field allocate_in_caller_region (or some name similar). It's more precise than ret_mode, because a function with ret_mode=local might not allocate in caller's region (think fun (local_ x) -> x), whereas this new field should give a definite answer. It will be set to true by exclave.
• Propagate similar name changes to flambda2.

I will create JIRA ticket to track this, and hope to do it in the next few weeks.

[mshinwell]

I will have to look at the discussion again but I think there were some problems with inadequate precision of locality information that caused trouble in #1271. Maybe this new idea will solve that. We should revisit that discussion when work recommences on this PR.

[riaqn]

I had another discussion with @lpw25 . Recollection:

• The type checker should track whether a function allocates in caller's region (ACR), and reflect that information in functions' types. This is to reject code that seem like not ACR but actually do and surprise users. In particular, tail-calling an ACR make the current function ACR as well.
• This is currently conservatively approxmated by the return mode of arrow types (and copied to ret_mode in lfunction). In the future once we introduce n-ary arrows, we will decouple ACR from the mode of the return value, and ACR will only be set by exclave.
• With the current conservative approximation, the middle-end can be more precise by inspecting function body.

[riaqn]

I looked at the changes again, and they seem fine to me.

@mshinwell could you look at my last commit "fix probe handler empty parameters", then we can merge.