[flandia]

This artifact is endorsed by Anto

[flandia]

wtf Codex stole my authorship

[LPTK]

Thanks!

As discussed,

• let's add a flag to Assign nodes to determine their effectfulness. We should make sure not to treat local mutation as a nonlocal side effect;
• do not use an effectful annotation (just use actual effects instead);
• make the optimizer not run when the lifter is disabled, so we don't need to care about nested definitions;
• merge this into the eta expansion pass (later: the megapass).

[LPTK]

Please stop force-pushing to this branch, it makes following the latest changes very hard. Just use merge commits like a civilised person.

[flandia]

Please stop force-pushing to this branch, it makes following the latest changes very hard. Just use merge commits like a civilised person.

Oh sorry I wanted to give a clean layer-by-layer review presentation

I'll do that locally next time

[flandia]

So I think this is ready for review.

It adds Pure and UnsafeEta summaries to functions and propagates them through calls from flow analysis (hkmc2/shared/src/test/mlscript/effect-analysis/Basic.mls). It conservatively assumes unsafe unless things can be proven pure. If a function is unsafe to eta-expand, the transformation stops at that point but still eta-expands the calls before it (hkmc2/shared/src/test/mlscript/opt/EtaEffects.mls).

Then a set of mutation analysis rules determine which functions are unsafe (hkmc2/shared/src/test/mlscript/effect-analysis/Mutation.mls). Examples with nested functions depend on the lifter so there are some mysterious and unpredictable (but seem to be sound) behaviors... e.g.,

fun f() =
  let state = 0
  fun observe_mutable() =
    state
  set state = 42
  observe_mutable()
  state
//│ flow-analysis > >>> effect-analysis results >>>
//│ flow-analysis > function f#0 @ <root> -> Pure
//│ flow-analysis > function observe_mutable#0 @ <root> -> Pure
//│ flow-analysis > call observe_mutable@result @ <root> -> Pure
//│ flow-analysis > <<< effect-analysis results <<<

fun g =
  let state = 0
  fun observe_mutable() =
    state
  observe_mutable()
  set state = 42
  state
//│ flow-analysis > >>> effect-analysis results >>>
//│ flow-analysis > function g#0 @ <root> -> UnsafeEta
//│ flow-analysis > function observe_mutable#0 @ <root> -> UnsafeEta
//│ flow-analysis > call observe_mutable@result @ <root> -> UnsafeEta
//│ flow-analysis > <<< effect-analysis results <<<

[flandia]

Then a set of mutation analysis rules determine which functions are unsafe (hkmc2/shared/src/test/mlscript/effect-analysis/Mutation.mls). Examples with nested functions depend on the lifter so there are some mysterious and unpredictable (but seem to be sound) behaviors... e.g.,

OK I managed to figure this out. It's not mysterious or unpredictable but a bit complex

[flandia]

LP has spoken in the meeting; thus we shall change the word Gate

[LPTK]

The PR tries to add hundreds of lines of unwanted extra output. Please be more careful and review your diff properly. We don't want to stage that much noise.

[LPTK]

Regression in eta expansion!

[flandia]

Reason is overly conservative field access classification. Just fixed

[flandia]

I feel like we should have a #config_once or something for simpler test cases. Now adding a #config for a specific case changes all the subsequent blocks which is annoying and fragile!

[flandia]

Just tracking this yet another regression. Codex told me fun f(x) = ... is deforested into

fun f$x_X(X_a)(...consumerVariables) =
  X_a

then X_a becomes a capture, so our analysis classifies it as unsafe. This should be fixed by the capture analysis later

[LPTK]

As discussed, pls ensure that there are no regressions to the existing committed tests – use @pure annotations with FIXMEs if needed. And investigate the problem in #562 (comment) to actually understand it.