experiment: stable functions

[crusso]

additional tests

• stable-class-bug.mo: demonstrates unsoundness of promoting generic stable functions to generic local functions.
• stable-box-bug.mo: a simplified counter-example.

[github-actions]

Comparing from bb2ae15 to 8a00c4a:
The produced WebAssembly code seems to be completely unchanged.

[crusso]

I added a renaming pass to ir_passes that renames all bound variables apart from @-prefixed ones that the code generator relies on (1c212ed).

As I expected, this appears to break eop, presumably because the free variable annotations on stable closures are not safe for renaming/alpha conversion. Broken, but probably fixable with some work:

I think the stable closure environment need to be represented as an explicit record, with fixed field names but alpha convertible values for each field. The closure body then has to rebind its free variables by pattern matching on the record of fixed field names. Or the moral equivalent of that.

The eop test that fail are

[nix-shell:~/motoko/test/run-drun]$ EXTRA_MOC_ARGS=--enhanced-orthogonal-persistence make
...
Some tests failed:
for-await.mo gc-random-test-force-gc.mo gc-random-test-no-force-gc.mo nested-generic-classes.mo nested-stable-functions.mo stable-box-bug.mo stable-captures-stable.mo stable-class-bug.mo stable-class.mo generic-stable-function.drun incompatible-stable-closure.drun life.drun life-mut.drun migrate-stable-function.drun reachable-stable-functions.drun

I'm pretty sure the tests all succeed when run without eop...checking now...

Yep:

[nix-shell:~/motoko/test/run-drun]$ make
../run.sh -d *.mo *.drun
WARNING: Could not run ic-ref-run, will skip running some tests
....
variant-upgrades:
All tests passed.