refs #24929, partially reverts #23403
Instead of using `Table[ItemId, T]`, the old algorithm is brought back
into `TIdTable[T]` to prevent a performance regression. The inheritance
removal from #23403 still holds, only `ItemId`s are stored.
(cherry picked from commit 82553384d1)
Yet another one of these. Multiple changes piled up in this one. I've
only minimally cleaned it for now (debug code is still here etc). Just
want to start putting this up so I might get feedback. I know this is a
lot and you all are busy with bigger things. As per my last PR, this
might just contain changes that are not ready.
### concept instantiation uniqueness
It has already been said that concepts like `ArrayLike[int]` is not
unique for each matching type of that concept. Likewise the compiler
needs to instantiate a new proc for each unique *bound* type not each
unique invocation of `ArrayLike`
### generic parameter bindings
Couple of things here. The code in sigmatch has to give it's bindings to
the code in concepts, else the information is lost in that step. The
code that prepares the generic variables bound in concepts was also
changed slightly. Net effect is that it works better.
I did choose to use the `LayedIdTable` instead of the `seq`s in
`concepts.nim`. This was mostly to avoid confusing myself. It also
avoids some unnecessary movings around. I wouldn't doubt this is
slightly less performant, but not much in the grand scheme of things and
I would prefer to keep things as easy to understand as possible for as
long as possible because this stuff can get confusing.
### various fixes in the matching logic
Certain forms of modifiers like `var` and generic types like
`tyGenericInst` and `tyGenericInvocation` have logic adjustments based
on my testing and usage
### signature matching method adjustment
This is the weird one, like my last PR. I thought a lot about the
feedback from my last attempt and this is what I came up with. Perhaps
unfortunately I am preoccupied with a slight grey area. consider the
follwing:
```nim
type
C1 = concept
proc p[T](s: Self; x: T)
C2[T] = concept
proc p(s: Self; x: T)
```
It would be temping to say that these are the same, but I don't think
they are. `C2` makes each invocation distinct, and this has important
implications in the type system. eg `C2[int]` is not the same type as
`C2[string]` and this means that signatures are meant to accept a type
that only matches `p` for a single type per unique binding. For `C1` all
are the same and the binding `p` accepts multiple types. There are
multiple variations of this type classes, `tyAnything` and the like.
The make things more complicated, an implementation might match:
```nim
type
A = object
C3 = concept
proc p(s: Self; x: A)
```
if the implementation defines:
```nim
proc p(x: Impl; y: object)
```
while a concept that fits `C2` may be satisfied by something like:
```nim
proc p(x: Impl; y: int)
proc spring[T](x: C2[T])
```
it just depends. None of this is really a problem, it just seems to
provoke some more logic in `concepts.nim` that makes all of this (appear
to?) work. The logic checks for both kinds of matches with a couple of
caveats. The fist is that some unbind-able arrangements may be matched
during overload resolution. I don't think this is avoidable and I
actually think this is a good way to get a failed compilation. So, first
note imo is that failing during binding is preferred to forcing the
programming to write annoying stub procs and putting insane gymnastics
in the compiler. Second thing is: I think this logic is way to accepting
for some parts of overload resolutions. Particularly in `checkGeneric`
when disambiguation is happening. Things get hard to understand for me
here. ~~I made it so the implicit bindings to not count during
disambiguation~~. I still need to test this more, but the thought is
that it would help curb excessive ambiguity errors.
Again, I'm sorry for this being so many changes. It's probably
inconvenient.
---------
Co-authored-by: Andreas Rumpf <rumpf_a@web.de>
(cherry picked from commit dfab30734b)
split from #24198
This is a required refactor for the only good solution I've been able to
think of for #4858 etc. Explanation:
---
`sigmatch` currently [disables
bindings](d6a71a1067/compiler/sigmatch.nim (L1956))
(except for binding to other generic parameters) when matching against
constraints of generic parameters. This is so when the constraint is a
general metatype like `seq`, the type matching will not treat all
following uses of `seq` as the type matched against that generic
parameter.
However to solve #4858 etc we need to bind `or` types with a conversion
match to the type they are supposed to be converted to (i.e. matching
`int literal(123)` against `int8 | int16` should bind `int8`[^1], not
`int`). The generic parameter constraint binding needs some way to keep
track of this so that matching `int literal(123)` against `T: int8 |
int16` also binds `T` to `int8`[^1].
The only good way to do this IMO is to generate a new "binding context"
when matching against constraints, then binding the generic param to
what the constraint was bound to in that context (in #24198 this is
restricted to just `or` types & concrete types with convertible matches,
it doesn't work in general).
---
`semtypinst` already does something similar for bindings of generic
invocations using `LayeredIdTable`, so `LayeredIdTable` is now split
into its own module and used in `sigmatch` for type bindings as well,
rather than a single-layer `TypeMapping`. Other modules which act on
`sigmatch`'s binding map are also updated to use this type instead.
The type is also made into an `object` type rather than a `ref object`
to reduce the pointer indirection when embedding it inside
`TCandidate`/`TReplTypeVars`, but only on arc/orc since there are some
weird aliasing bugs on refc/markAndSweep that cause a segfault when
setting a layer to its previous layer. If we want we can also just
remove the conditional compilation altogether and always use `ref
object` at the cost of some performance.
[^1]: `int8` binding here and not `int16` might seem weird, since they
match equally well. But we need to resolve the ambiguity here, in #24012
I tested disallowing ambiguities like this and it broke many packages
that tries to match int literals to things like `int16 | uint16` or
`int8 | int16`. Instead of making these packages stop working I think
it's better we resolve the ambiguity with a rule like "the earliest `or`
branch with the best match, matches". This is the rule used in #24198.
(cherry picked from commit cad8726907)
