generics: error with multiple types and constraints

[brentp]

Update: @krux02 has found this happens when constraints are specified. When the constraints are removed the code will compile.

the code below fails with: Error: cannot instantiate Data got: <type int, type float64> but expected: <T: SomeNumber, U: SomeNumber>

type
  Element*[U:SomeNumber] = ref object
    value*: U

  Data*[T:SomeNumber, U:SomeNumber] = ref object
    xs*: seq[T]
    e*: Element[U]

var e = Element[float64](value:2'f64)
var xs = new_seq[int](10)
var d = Data[int, float64](xs:xs, e:e)

[narimiran]

Even this doesn't work:

type
  Element*[U: float64] = ref object
    value*: U

  Data*[T: int, U: float64] = ref object
    xs*: seq[T]
    e*: Element[U]

let e = Element[float64](value:2'f64)
let xs = new_seq[int](10)
var d = Data[int, float64](xs: xs, e: e)

[brentp]

or this:

type
  Data*[T:SomeNumber, U:SomeReal] = ref object
    x*: T
    value*: U

var d = Data[int, float64](x:10.int, value:2'f64)

[krux02]

well as close as I can get between working and non working with non-essential parts removed.

type
  # this works
  DataA*[T, U] = object
    xs*: seq[T]
    e*: U

  # this doesn't
  DataB*[T:SomeNumber; U:SomeNumber] = object
    xs*: seq[T]
    e*: U

var a = DataA[int, float64](xs: @[1,2,3], e: 12.3)
var b = DataB[int, float64](xs: @[1,2,3], e: 12.3)

[LemonBoy]

There are two factors contributing to this:

• In matchesAux, before paramTypesMatch we should really reset m.typedescMatched = false
• The typeRel handling of tyOr (and tyAnd/tyNot I think) is fishy: when T:SomeNumber is run first trough typerel the bindingRet ties it to the type it matched (tyInt in the snippet above). This means that when it's time to process U:SomeNumber the tyOr is fetched from the cache as tyInt (WTF?) and then is compared to tyFloat64, yielding isNone and making the instantiation fail.

CC @zah

[zah]

m.typedescMatched is a very hacky work-around trying to compensate for one very old historic flaw of the compiler internal design. The generic params "signature" of a type is still consisting of tyGenericParams slots as formal requirement, while it should really be using tyTypeDesc. The meaning of tyGenericParam is "match any value", while tyTypeDesc has the proper meaning of "match any type". Unfortunately, fixing this has proven to be quite hard in the few times I've attempted to solve the problem, so the compiler is still relying on these subtly wrong m.typedescMatched hacks.

Regarding your point 2, you seem to be missing the fact that each named type class is a "bind once" type by default. This is briefly explained in the type classes section of the manual:
https://nim-lang.org/docs/manual.html#generics-type-classes

In this sense, Brent's example is the best when it comes to solving this issue, because it doesn't invoke this extra complexity:

type
  Data*[T:SomeNumber, U:SomeReal] = ref object
    x*: T
    value*: U

var d = Data[int, float64](x:10.int, value:2'f64)

With that said, I'll mention for completeness that there are indeed problems with the or handling in the compiler. Properly handling or types in all situations requires back-tracking in the type binding process. We'll eventually introduce some support for this in the compiler to deal with some more easily triggered cases involving concepts.

[LemonBoy]

There's definitely more than it meets the eye.

very hacky work-around

Yeah, I noticed some comments about this in the compiler.

Regarding your point 2, you seem to be missing the fact that each named type class is a "bind once" type by default.

Oh, I've completely missed that bit.

In this sense, Brent's example is the best when trying to solve this issue, because it doesn't invoke this extra complexity:

Definitely

Resetting m.typedescMatched fixes the issue at hand, do you think it may have other unintended consequences?

[zah]

Sounds like a plausible solution, but as usual the test suite will tell you if there are problems.

[Araq]

The generic params "signature" of a type is still consisting of tyGenericParams slots as formal requirement, while it should really be using tyTypeDesc.

That's not true, tyTypeDesc is just as flawed for this as typically you want to allow e.g. int but not typeDesc[int] in your generic.