Backend support for local allocations

[stedolan]

This PR adds support for local (stack) allocations to the backend, extracted from the branch at https://github.com/ocaml-flambda/ocaml-jst/tree/local-dev. To avoid creating at truly massive PR, this patch contains only the backend changes. (This also means that this PR contains no tests for local allocations. There is an extensive testsuite in the ocaml-jst repo, but the tests require front and middle-end support in order to run).

The original PR series for local allocations is at https://github.com/janestreet/ocaml/pulls?q=is%3Apr.

Concretely, the main changes here are:

• Annotate allocations with a Lambda.alloc_mode. Alloc_heap allocations are normal GC allocations, while the new Alloc_local allocations allocate instead on a separate stack.

• At Mach and beyond, add two new primitives Ibeginregion and Iendregion that delimit regions (roughly, stackframes). All local allocations during a region are freed at the end of the region, by resetting the local stack pointer. Ibeginregion returns the current value of this pointer, while Iendregion resets it.

  (Currently, Iendregion does some additional tracking of stats, for development and testing. This will be removed before release)

• At Cmm level, instead of explicit begin/end primitives we use a block-structured Cregion e. The job of placing Ibeginregion / Iendregion falls to Selectgen.

There is special handling of tail calls: since tail recursive loops must run in constant space, tail calls in tail position of a region end the region before transferring control. This is denoted by an apply_position in Cmm: when it is Apply_tail, the Iendregion will be inserted before the call by Selectgen. Furthermore, there is a new Ctail construct in Cmm to handle the case where an Apply_tail call is inlined, which ends the region before the inlined body is evaluated.

Finally, there is some refactoring of the generation of curry / apply stubs: instead of encoding tupled vs. curried in the sign of the arity, the stub is selected as a pair of a Lambda function kind and an integer. (This refactoring is there mainly to support the middle-end's need for both local and heap versions of the curry stubs)

[xclerc]

(Cannot add a comment on the code outside of the diff.)
In "backend/arm64/proc.ml", shouldn't the op_is_pure
function have the following case:

| Ibeginregion | Iendregion -> false

[xclerc]

I think the current transformations over CFG values are
not affected by these changes (and the CI is green), but
region begins/ends likely impose some restrictions on
what can safely be changed about a CFG value.

@gretay-js This is probably worth adding to your document
about CFG invariants.

(The fact that the CI is green is moot, since the new
features are not currently being tested in this repository.)

[stedolan]

I think the current transformations over CFG values are
not affected by these changes (and the CI is green), but
region begins/ends likely impose some restrictions on
what can safely be changed about a CFG value.

My hope was that marking local allocations as coeffectful and Ibeginregion / Iendregion as effectful would be enough to ensure the right semantics - it is invalid to move a load across a possibly-aliasing store anyway.

[xclerc]

I think the current transformations over CFG values are
not affected by these changes (and the CI is green), but
region begins/ends likely impose some restrictions on
what can safely be changed about a CFG value.

My hope was that marking local allocations as coeffectful and Ibeginregion / Iendregion as effectful would be enough to ensure the right semantics - it is invalid to move a load across a possibly-aliasing store anyway.

I think I am mostly worried about ensuring that no
transformations break the bracketing of begin/end
pairs, but that might be unfounded.

[mshinwell]

The following is needed in Printmach:

!|  | Ialloc { bytes = n; mode = Alloc_heap } ->
 |    fprintf ppf "alloc %i" n;
+|  | Ialloc { bytes = n; mode = Alloc_local } ->
+|    fprintf ppf "alloc_local %i" n;

[xclerc]

For grepping purposes, I think the existing line is:

| Ialloc { bytes = n; } ->

[stedolan]

Thanks, patch applied.