fixes#5631, fixes#8938, fixes#18855, fixes#19271, fixes#23885,
fixes#24877
`isTupleRecursive`, previously only called to give an error for illegal
recursions for:
* tuple fields
* types declared in type sections
* explicitly instantiated generic types
did not check for recursions in proc types. It now does, meaning proc
types now need a nominal type layer to recurse over themselves. It is
renamed to `isRecursiveStructuralType` to better reflect what it does,
it is different from a recursive type that cannot exist due to a lack of
pointer indirection which is possible for nominal types.
It is now also called to check the param/return types of procs, similar
to how tuple field types are checked. Pointer indirection checks are not
needed since procs are pointers.
I wondered if this would lead to a slowdown in the compiler but since it
only skips structural types it shouldn't take too many iterations, not
to mention only proc types are newly considered and aren't that common.
But maybe something in the implementation could be inefficient, like the
cycle detector using an IntSet.
Note: The name `isRecursiveStructuralType` is not exactly correct
because it still checks for `distinct` types. If it didn't, then the
compiler would accept this:
```nim
type
A = distinct B
B = ref A
```
But this breaks when attempting to write `var x: A`. However this is not
the case for:
```nim
type
A = object
x: B
B = ref A
```
So a better description would be "types that are structural on the
backend".
A future step to deal with #14015 and #23224 might be to check the
arguments of `tyGenericInst` as well but I don't know if this makes
perfect sense.
(cherry picked from commit 7f0e07492f)
fixes#24887 (really just this [1 line
commit](632c7b3397)
would have been enough to fix the issue but it would ignore the general
problem)
When a type definition is encountered where the symbol already has a
type (not a forward type), the type is left alone (not reset to
`tyForward`) and the RHS is handled differently: The RHS is still
semchecked, but the type of the symbol is not updated, and nominal type
nodes are ignored entirely (specifically if they are the same kind as
the symbol's existing type but this restriction is not really needed).
If the existing type of the symbol is an enum and and the RHS has a
nominal enum type node, the enum fields of the existing type are added
to scope rather than creating a new type from the RHS and adding its
symbols instead.
The goal is to prevent any incompatible nominal types from being
generated during resem as in #24887. But it also restricts what macros
can do if they generate type section AST, for example if we have:
```nim
type Foo = int
```
and a macro modifies the type section while keeping the symbol node for
`Foo` like:
```nim
type Foo = float
```
Then the type of `Foo` will still remain `int`, while it previously
became `float`. While we could maybe allow this and make it so only
nominal types cannot be changed, it gets even more complex when
considering generic params and whether or not they get updated. So to
keep it as simple as possible the rule is that the symbol type does not
change, but maybe this behavior was useful for macros.
Only nominal type nodes are ignored for semchecking on the RHS, so that
cases like this do not cause a regression:
```nim
template foo(): untyped =
proc bar() {.inject.} = discard
int
type Foo = foo()
bar() # normally works
```
However this specific code exposed a problem with forward type handling:
---
In specific cases, when the type section is undergoing the final pass,
if the type fits some overly general criteria (it is not an object,
enum, alias or a sink type and its node is not a nominal type node), the
entire RHS is semchecked for a 2nd time as a standalone type (with `nil`
prev) and *maybe* reassigned to the new semchecked type, depending on
its type kind. (for some reason including nominal types when we excluded
them before?) This causes a redefinition error if the RHS defines a
symbol.
This code goes all the way back to the first commit and I could not find
the reason why it was there, but removing it showed a failure in
`thard_tyforward`: If a generic forward type is invoked, it is left as
an unresolved `tyGenericInvocation` on the first run. Semchecking it
again at the end turns it into a `tyGenericInst`. So my understanding is
that it exists to handle these loose forward types, but it is way too
general and there is a similar mechanism `c.skipTypes` which is supposed
to do the same thing but doesn't.
So this is no longer done, and `c.skipTypes` is revamped (and renamed):
It is now a list of types and the nodes that are supposed to evaluate to
them, such that types needing to be updated later due to containing
forward types are added to it along with their nodes. When finishing the
type section, these types are reassigned to the semchecked value of
their nodes so that the forward types in them are fully resolved. The
"reassigning" here works due to updating the data inside the type
pointer directly, and is how forward types work by themselves normally
(`tyForward` types are modified in place as `s.typ`).
For example, as mentioned before, generic invocations of forward types
are first created as `tyGenericInvocation` and need to become
`tyGenericInst` later. So they are now added to this list along with
their node. Object types with forward types as their base types also
need to be updated later to check that the base type is correct/inherit
fields from it: For this the entire object type and its node are added
to the list. Similarly, any case where whether a component type is
`tyGenericInst` or `tyGenericInvocation` matters also needs to cascade
this (`set` does presumably to check the instantiated type).
This is not complete: Generic invocations with forward types only check
that their base type is a forward type, but not any of their arguments,
which causes #16754 and #24133. The generated invocations also need to
cascade properly: `Foo[Bar[ForwardType]]` for example would see that
`Bar[ForwardType]` is a generic invocation and stay as a generic
invocation itself, but it might not queue itself to be updated later.
Even if it did, only the entire type `Foo[Bar[ForwardType]]` needs to be
queued, updating `Bar[ForwardType]` by itself would be redundant or it
would not change anything at all. But these can be done later.
(cherry picked from commit 525d64fe88)
split from #24204, closes#7674
The `{.size.}` pragma no longer restricts the given size to 1, 2, 4 or 8
if it is used for an imported type. This is not tested very thoroughly
but there's no obvious reason to disallow it.
(cherry picked from commit 1ef9a656d2)
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)
fixes#24725
`lacksMTypeField` doesn't take the base types into consideration. And
for ` {.inheritable, pure.}`, it shouldn't generate a `m_type` field.
(cherry picked from commit e449813c61)
follows up #24425, fixes#18861, fixes#22445
Since #24425 generic object and distinct types now accurately link back
to their generic instantiations. To work around this issue previously,
type matching checked if generic objects/distinct types were
*structurally* equal, which caused false positives with types that
didn't use generic parameters in their structures. This structural check
is now removed, in cases where generic objects/distinct types require a
nontrivial equality check, the generic parameters of the `typeInst`
fields are checked for equality instead.
The check is copied from `tyGenericInst`, but the check in
`tyGenericInst` is not always sufficient as this type can be skipped or
unreachable in the case of `ref object`s.
(cherry picked from commit a2031ec6cf)
fixes#22479, fixes#24374, depends on #24429 and #24430
When instantiating generic types which directly have nominal types
(object, distinct, ref/ptr object but not enums[^1]) as their values,
the nominal type is now copied (in the case of ref objects, its child as
well) so that it receives a fresh ID and `typeInst` field. Previously
this only happened if it contained any generic types in its structure,
as is the case for all other types.
This solves #22479 and #24374 by virtue of the IDs being unique, which
is what destructors check for. Technically types containing generic
param fields work for the same reason. There is also the benefit that
the `typeInst` field is correct. However issues like #22445 aren't
solved because the compiler still uses structural object equality checks
for inheritance etc. which could be removed in a later PR.
Also fixes a pre-existing issue where destructors bound to object types
with generic fields would not error when attempting to define a user
destructor after the fact, but the error message doesn't show where the
implicit destructor was created now since it was only created for
another instance. To do this, a type flag is used that marks the generic
type symbol when a generic instance has a destructor created. Reusing
`tfCheckedForDestructor` for this doesn't work.
Maybe there is a nicer design that isn't an overreliance on the ID
mechanism, but the shortcomings of `tyGenericInst` are too ingrained in
the compiler to use for this. I thought about maybe adding something
like `tyNominalGenericInst`, but it's really much easier if the nominal
type itself directly contains the information of its generic parameters,
or at least its "symbol", which the design is heading towards.
[^1]: See [this
test](21420d8b09/lib/std/enumutils.nim (L102))
in enumutils. The field symbols `b0`/`b1` always have the uninstantiated
type `B` because enum fields don't expect to be generic, so no generic
instance of `B` matches its own symbols. Wouldn't expect anyone to use
generic enums but maybe someone does.
(cherry picked from commit 05c74d6844)
split from #24425
Matching `tyGenericBody` performs a match on the last child of the
generic body, in this case the uninstantied `tyObject` type. If the
object contains no generic fields, this ends up being the same type as
all instantiated ones, but if it does, a new type is created. This fails
the `sameObjectTypes` check that subtype matching for object types uses.
To fix this, also consider that the pattern type could be the generic
uninstantiated object type of the matched type in subtype matching.
(cherry picked from commit 511ab72342)
fixes#24415
Since #23978 (which is in 2.0), all generic types that alias to another
type now insert a `tyAlias` layer in their value. However the
`skipGenericAlias` etc code which `sigmatch` uses is not updated for
this, so `tyAlias` is now skipped in these.
The relevant code in sigmatch is:
67ad1ae159/compiler/sigmatch.nim (L1668-L1673)
This behavior is also suspicious IMO, not skipping a structural
`tyGenericInst` alias can be useful for code like #10220, but this is
currently arbitrarily decided based on "depth" and whether the alias is
to another `tyGenericInst` type or not. Maybe in the future we could
enforce the use of a nominal type.
(cherry picked from commit 45b8434c7d)
closes https://github.com/nim-lang/RFCs/issues/380, fixes#4773, fixes
#14729, fixes#16755, fixes#18150, fixes#22984, refs #11167 (only some
comments fixed), refs #12620 (needs tiny workaround)
The compiler gains a concept of root "nominal" types (i.e. objects,
enums, distincts, direct `Foo = ref object`s, generic versions of all of
these). Exported top-level routines in the same module as the nominal
types that their parameter types derive from (i.e. with
`var`/`sink`/`typedesc`/generic constraints) are considered attached to
the respective type, as the RFC states. This happens for every argument
regardless of placement.
When a call is overloaded and overload matching starts, for all
arguments in the call that already have a type, we add any operation
with the same name in the scope of the root nominal type of each
argument (if it exists) to the overload match. This also happens as
arguments gradually get typed after every overload match. This restricts
the considered overloads to ones attached to the given arguments, as
well as preventing `untyped` arguments from being forcefully typed due
to unrelated overloads. There are some caveats:
* If no overloads with a name are in scope, type bound ops are not
triggered, i.e. if `foo` is not declared, `foo(x)` will not consider a
type bound op for `x`.
* If overloads in scope do not have enough parameters up to the argument
which needs its type bound op considered, then type bound ops are also
not added. For example, if only `foo()` is in scope, `foo(x)` will not
consider a type bound op for `x`.
In the cases of "generic interfaces" like `hash`, `$`, `items` etc. this
is not really a problem since any code using it will have at least one
typed overload imported. For arbitrary versions of these though, as in
the test case for #12620, a workaround is to declare a temporary
"template" overload that never matches:
```nim
# neither have to be exported, just needed for any use of `foo`:
type Placeholder = object
proc foo(_: Placeholder) = discard
```
I don't know what a "proper" version of this could be, maybe something
to do with the new concepts.
Possible directions:
A limitation with the proposal is that parameters like `a: ref Foo` are
not attached to any type, even if `Foo` is nominal. Fixing this for just
`ptr`/`ref` would be a special case, parameters like `seq[Foo]` would
still not be attached to `Foo`. We could also skip any *structural* type
but this could produce more than one nominal type, i.e. `(Foo, Bar)`
(not that this is hard to implement, it just might be unexpected).
Converters do not use type bound ops, they still need to be in scope to
implicitly convert. But maybe they could also participate in the nominal
type consideration: if `Generic[T] = distinct T` has a converter to `T`,
both `Generic` and `T` can be considered as nominal roots.
The other restriction in the proposal, being in the same scope as the
nominal type, could maybe be worked around by explicitly attaching to
the type, i.e.: `proc foo(x: T) {.attach: T.}`, similar to class
extensions in newer OOP languages. The given type `T` needs to be
obtainable from the type of the given argument `x` however, i.e.
something like `proc foo(x: ref T) {.attach: T.}` doesn't work to fix
the `ref` issue since the compiler never obtains `T` from a given `ref
T` argument. Edit: Since the module is queried now, this is likely not
possible.
---------
Co-authored-by: Andreas Rumpf <rumpf_a@web.de>
(cherry picked from commit 2864830941)
fixes#22523
There were 2 problems with the code in `sameType` for
`dcEqIgnoreDistinct`:
1. The code that skipped `{tyDistinct, tyGenericInst}` only ran if the
given types had different kinds. This is fixed by always performing this
skip.
2. The code block below that checks if `tyGenericInst`s have different
values still ran for `dcEqIgnoreDistinct` since it checks if the given
types are generic insts, not the skipped types (and also only the 1st
given type). This is fixed by only invoking this block for `dcEq`;
`dcEqOrDistinctOf` (which is unused) also skips the first given type.
Arguably there is another issue here that `skipGenericAlias` only ever
skips 1 type.
These combined fix the issue (`T` is `GenericInst(V, 1, distinct int)`
and `D[0]` is `GenericInst(D, 0, distinct int)`).
#24037 shouldn't be a dependency but the diff follows it.
follows up #24095
In #24095 a check was added that used `iterOverType` to check if a type
contained unresolved types, with the aim of always treating
`tyGenericBody` as resolved. But the body of the `tyGenericBody` is also
iterated over in `iterOverType`, so if the body of the type actually
used generic parameters (which isn't the case in the test added in
#24095, but is now), the check would still count the type as unresolved.
This is handled by not iterating over the children of `tyGenericBody`,
the only users of `iterOverType` are `containsGenericType` and
`containsUnresolvedType`, the first one always returns true for
`tyGenericBody` and the second one aims to always return false.
Unfortunately this means `iterOverType` isn't as generic of an API
anymore but maybe it shouldn't be used anymore for these procs.
fixes regression remaining after #24092
In #24092 `prepareNode` was updated so it wouldn't try to instantiate
generic type symbols (like `Generic` when `type Generic[T] = object`,
and `prepareNode` is what `tyFromExpr` uses in most of the compiler. An
exception is in sigmatch, which is now changed to use `prepareNode` to
make generic type symbols work in the same way as usual. However this
requires another change to work:
Dot fields and matches to `typedesc` on generic types generate
`tyFromExpr` in generic contexts since #24005, including generic type
symbols. But this means when we try to instantiate the `tyFromExpr` in
sigmatch, which increases `c.inGenericContext` for potentially remaining
unresolved expressions, dotcalls stay as `tyFromExpr` and so never
match. To fix this, we change the "generic type" check in dot fields and
`typedesc` matching to an "unresolved type" check which excludes generic
body types; and for generic body types, we only generate `tyFromExpr` if
the dot field is a generic parameter of the generic type (so that it
gets resolved only at instantiation).
Notes for the future:
* Sigmatch shouldn't have to `inc c.inGenericContext`, if a `tyFromExpr`
can't instantiate it's fine if we just fail the match (i.e. redirect the
instantiation errors from `semtypinst` to a match failure). Then again
maybe this is the best way to check for inability to instantiate.
* The `elif c.inGenericContext > 0 and t.containsUnresolvedType` check
in dotfields could maybe be simplified to just checking for `tyFromExpr`
and `tyGenericParam`, but I don't know if this is an exhaustive list.
refs #24032, split from #24036
Conversion from variables of range types or base types of range types to
the other are now considered mutable for `var` params, similar to how
distinct types are mutable when converted to their base type or vice
versa. There are 2 main differences:
1. Conversions from base types to range types need to emit
`nkChckRange`, which is not generated for things like tuple/object
fields.
2. Range types can still correspond to different types in the backend
when nested in other types, such as `set[range[3..5]]` vs
`set[range[0..5]]`.
Since the convertibility check for `var` params and a check whether to
emit a no-op for `nkConv` (and now also `nkChckRange`) so that the
output is still addressable both use `sameType`, we accomplish this by
adding a new flag to `sameType` that ignores range types, but only when
they're not nested in other types. The implementation for this might be
flawed, I didn't include children of some metatypes as "nested in other
types", but stuff like `tyGenericInst` params are respected.
updated version of #22193
After #22029 and the followups #23983 and #24005 which fixed issues with
it, `tyFromExpr` no longer match any proc params in generic type bodies
but delay all non-matching calls until the type is instantiated.
Previously the mechanism `fauxMatch` was used to pretend that any
failing match against `tyFromExpr` actually matched, but prevented the
instantiation of the type until later.
Since this mechanism is not needed anymore for `tyFromExpr`, it is now
only used for `tyError` to prevent cascading errors and changed to a
bool field for simplicity. A change in `semtypes` was also needed to
prevent calling `fitNode` on default param values resolving to type
`tyFromExpr` in generic procs for params with non-generic types, as this
would try to coerce the expression into a concrete type when it can't be
instantiated yet.
The aliases `tyProxy` and `tyUnknown` for `tyError` and `tyFromExpr` are
also removed for uniformity.
This fixes a nimsuggest crash when opening:
beacon_chain/consensus_object_pools/blockchain_dag.nim
from the nimbus-eth2 project and many other .nim files (44 files, to be
precise) in the same project.
Replaces: https://github.com/nim-lang/Nim/pull/23402
Close#22826
I am not sure why this code skips generic insts, so letting CI tell me.
Update: It has told me nothing. Maybe someone knows during review.
Issue itself seems to be that the generic instance is skipped thus it
ends up being just `float` which makes it use the wrong generic instance
of the proc because it matches the one in cache
---------
Co-authored-by: SirOlaf <>
Close#21742
Checking if there's any side-effects and if just changing typeRel is
adequate for this issue before trying to look into related ones.
`skipBoth` is also not that great, it can lead to code that works
sometimes but fails when the proc is instantiated with branching
aliases. This is mostly an issue with error clarity though.
---------
Co-authored-by: SirOlaf <unknown>
Co-authored-by: SirOlaf <>
* sacrifice "tgenericshardcases" for working statics
* legacy switch for CI, maybe experimental later
* convert to experimental
* apparently untyped needs the experimental switch
* try special case call semcheck
* try fix
* fix compilation
* final cleanup, not experimental, make `when` work
* remove last needed use of untyped
* fix unused warning in test
* remove untyped feature
