strict effects

[Araq]

Todo:

• bootstrapping with --experimental:strictEffects and --warnAsError:effect:on
• test cases
• documentation

Will fix:

• Incorrect possible CatchableError error reported when adding procedure to table #18376
• raises tracking broken for... proc types in tables? #17382
• raises not enforced for proc type parameters #17475
• raises not preseved in proc types, invalid "can raise any" error on compile #13905

[juancarlospaco]

Interesting, ...so this is like strictFuncs but for "Effects" ?. 😲:+1:

[timotheecour]

bugs:

when defined case6: proc fn(a: int, p1, p2: proc()) {.effectsOf: p1.} = # proc fn(a: int, p1, p2: proc()) = # ditto if a == 7: p1() if a<0: raise newException(ValueError, $a) proc main() {.raises: [ValueError].} = # fn(1, proc() = raise newException(OsError, "foo"), proc()=discard) # ok: can raise an unlisted exception: ref OSError is correct fn(1, proc()=discard, proc() = raise newException(IOError, "foo")) # bug: `can raise an unlisted exception: ref IOError` shouldn't be generated because of `effectsOf: p1` main()

also this, slightly different as it involves parent scope:

when defined case8: proc bar(a: proc()) = proc baz(b: proc()) {.effectsOf: a.} = a() # not calling b but calling a param from parent scope proc x()=raise newException(ValueError, "") baz(x) proc main(){.raises: [].} = bar(proc()=raise newException(OsError, "")) main()

produces incorrect diagnostic (in particular, ValueError is never raised)

can raise an unlisted exception: Exception [Effect] can raise an unlisted exception: ref ValueError [Effect] can raise an unlisted exception: ref OSError [Effect] 

[timotheecour]

but this should be legal:

 proc bar(a: proc()) = proc baz(b: proc()) {.effectsOf: [a].} = a()

[Araq]

but this should be legal

Should be done later as it's much harder to implement. I also doubt this will ever be needed.

[timotheecour]

the bugs i mentioned in #18777 (comment) is still there, it looks like the compiler ignores the effectsOf: p1 annotations and assumes every param is called

[Araq]

Your snippet was even added as a test case, it works...

[timotheecour]

ok, i was missing {.experimental: "strictEffects".}; however, the problem remains if i remove the explicit annotaton:

when defined case9: {.push warningAsError[Effect]: on.} {.experimental: "strictEffects".} # proc fn(a: int, p1, p2: proc()) {.effectsOf: p1.} = # proc fn(a: int, p1, p2: proc()) {.effectsOf: p2.} = proc fn(a: int, p1, p2: proc()) = if a == 7: p1() if a<0: raise newException(ValueError, $a) proc main() {.raises: [ValueError].} = fn(1, proc()=discard, proc() = raise newException(IOError, "foo")) main() 

this gives:

t12754.nim(111, 7) Error: fn(1, proc () = discard , proc () = raise
(ref IOError)(msg: "foo", parent: nil)) can raise an unlisted exception: Exception [Effect]

but compiler should infer that this is {.effectsOf: p1.} since only p1 is called, not p2, and then it should not issue any warning since main can't raise anything other than ValueError

[Araq]

There is no inference for the .effectsOf pragma and there is no plan to add any, it would repeat the mistakes of "sink inference".

[Araq]

In a follow up PR we should patch map.