Links for succ/pred/inc/dec were incorrect since they link to a symbol
with `int` as their second type.
Uses local referencing instead so that it links to the correct symbol.
Doesn't change the other links since they worked and doing the same
local referencing for `high`/`low` would've make them link to the group
of procs instead of the specific ones for ordinals
pre-existing issues:
```nim
block:
type
FooObj = object
data: int
Foo = ref ref FooObj
proc delete(self: Foo) =
echo self.data
var s: Foo
new(s, delete)
```
it crashed with arc/orc in 1.6.x and 2.x.x
```nim
block:
type
Foo = ref int
proc delete(self: Foo) =
echo self[]
var s: Foo
new(s, delete)
```
The simple fix is to add a type restriction for the type `T` for arc/orc
versions
```nim
proc new*[T: object](a: var ref T, finalizer: proc (x: T) {.nimcall.})
```
fixes#13417, fixes#19703
When passing an expression to an `openarray` iterator parameter: If the
expression is a statement list (considered "complex"), it's assigned in
a non-deep-copying way to a temporary variable first, then this variable
is used as a parameter. If it's not a statement list, i.e. a call or a
symbol, the parameter is substituted directly with the given expression.
In the case of calls, this results in the call potentially being
executed more than once, or can cause redefined variables in the
codegen.
To fix this, calls are also considered as "complex" assignments to
openarrays, as long as the return type of the call is not `openarray` as
the generated assignment in that case has issues/is unimplemented
(caused a segfault [here in
datamancer](47ba4d81bf/src/datamancer/dataframe.nim (L1580))).
As for why creating a temporary isn't the default only with exceptions
for things like `nkSym`, the "non-deep-copying" way of assignment
apparently still causes arrays to be copied according to a comment in
the code. I'm not sure to what extent this is true: if it still happens
on ARC/ORC, if it happens for every array length, or if we can fix it by
passing arrays by reference. Otherwise, a more general way to assign to
openarrays might be needed, but I'm not sure if the compiler can easily
do this.
This PR is somewhat large, worst case it can be split into one with just
assignments and one with just fields/derefs etc.
Assignments with calls as values have not been touched so they can be
done when calls are implemented. Similarly codegen with complex logic
i.e. `genEnumInfo`, `genTypeInfoV2`, `unaryExpr` is not completely
ported yet so they can be done in standalone PRs.
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>
fixes#24236
Locally tested to generate a 100 KB file for TCC. Empty flexible array
size is standard in C99 but maybe some compilers still don't support it.
At the very least an array size of 1000000 should be rare.
fixes#24338
When unrolling each iteration of a `fields` iterator, the compiler only
opens a new scope for semchecking, but doesn't generate a node that
signals to the codegen that a new scope should be created. This causes
issues for reused template instantiations that reuse variable symbols
between each iteration, which causes the codegen to generate multiple
declarations for them in the same scope (regardless of `inject` or
`gensym`). To fix this, we wrap the unrolled iterations in an `if true:
body` node, which both opens a new scope and doesn't interfere with
`break`.
This fixes several cases of the Nim binding of nfds_t being inconsistent
with the target platform signedness and/or size.
Additionally, it fixes poll's third argument (timeout) being set to Nim
"int" when it should have been "cint".
The former is the same issue that #23045 had attempted to fix, but
failed because it only considered Linux. (Also, it was only applied to
version 2.0, so the two branches now have incompatible versions of the
same bug.)
Notes:
* SVR4's original "unsigned long" definition is cloned by Linux and
Haiku. Nim got this right for Haiku and Linux-amd64, but it was wrong on
non-amd64 Linux.
* Zephyr does not have nfds_t, but simply uses (signed) "int". This was
already correctly reflected by Nim.
* OpenBSD poll.h uses "unsigned int", and other BSD derivatives follow
suit. This being the most commonly copied definition, the fallback case
now returns cuint. (This also seems to be correct for the OS X headers I
could find on the web.)
* This changes Nintendo Switch nfds_t to cuint from culong. It is
purportedly a FreeBSD derivative, so I *think* this is correct, but I
can't tell because I don't have access to the Nintendo Switch headers.
I have also moved the platform-specific Tnfds to posix.nim so that we
can reuse the fallback logic on all platforms. (e.g. specifying the size
in posix_linux_amd64 only to then use when defined(linux) in posix_other
seems redundant.)
The only remaining explicit use of `typedef` in the codegen (from my
search) is in `addForwardStructFormat` which from what I understand
won't do anything in NIFC.
refs #6978, refs #6752, refs #21613, refs #24234
The `jsNoInt64`, `whenHasBigInt64`, `whenJsNoBigInt64` templates are
replaced with bool constants to use with `when`. Weird that I didn't do
this in the first place.
The `whenJsNoBigInt64` template was also slightly misleading. The first
branch was compiled for both no bigint64 on JS as well as on C/C++. It
seems only `trandom` depended on this by mistake.
The workaround for #6752 added in #6978 to `times` is also removed with
`--jsbigint64:on`, but #24233 was also encountered with this, so this PR
depends on #24234.
Running `ctags` on files with quoted symbols (e.g. `$`) would list \`
instead of the full ident. Issue was the result getting reassigned at
the end to a \` instead of appending
closes https://github.com/nim-lang/RFCs/issues/554
Adds a symmetric difference operation to the language bitset type. This
maps to a simple `xor` operation on the backend and thus is likely
faster than the current alternatives, namely `(a - b) + (b - a)` or `a +
b - a * b`. The compiler VM implementation of bitsets already
implemented this via `symdiffSets` but it was never used.
The standalone binary operation is added to `setutils`, named
`symmetricDifference` in line with [hash
sets](https://nim-lang.org/docs/sets.html#symmetricDifference%2CHashSet%5BA%5D%2CHashSet%5BA%5D).
An operator version `-+-` and an in-place version like `toggle` as
described in the RFC are also added, implemented as trivial sugar.
refs #8064, refs #24010
Error messages for standalone explicit generic instantiations are
revamped. Failing standalone explicit generic instantiations now only
error after overloading has finished and resolved to the default `[]`
magic (this means `[]` can now be overloaded for procs but this isn't
necessarily intentional, in #24010 it was documented that it isn't
possible). The error messages for failed instantiations are also no
longer a simple `cannot instantiate: foo` message, instead they now give
the same type mismatch error message as overloads with mismatching
explicit generic parameters.
This is now possible due to the changes in #24010 that delegate all
explicit generic proc instantiations to overload resolution. Old code
that worked around this is now removed. `maybeInstantiateGeneric` could
maybe also be removed in favor of just `explicitGenericSym`, the `result
== n` case is due to `explicitGenericInstError` which is only for niche
cases.
Also, to cause "ambiguous identifier" error messages when the explicit
instantiation is a symchoice and the expression context doesn't allow
symchoices, we semcheck the sym/symchoice created by
`explicitGenericSym` with the given expression flags.
#8064 isn't entirely fixed because the error message is still misleading
for the original case which does `values[1]`, as a consequence of
#12664.
The remaining followup from #24259. A body for building the type doesn't
seem necessary here since the types with array fields are generally
atomic/already built from `getTypeDescAux`.
The first commit reverts the revert of #23787.
The second fixes lambdalifting in convolutedly nested
closures/closureiters. This is considered to be the reason of #24094,
though I can't tell for sure, as I was not able to reproduce #24094 for
complicated but irrelevant reasons. Therefore I ask @jmgomez, @metagn or
anyone who could reproduce it to try it again with this PR.
I would suggest this PR to not be squashed if possible, as the history
is already messy enough.
Some theory behind the lambdalifting fix:
- A closureiter that captures anything outside its body will always have
`:up` in its env. This property is now used as a trigger to lift any
proc that captures such a closureiter.
- Instantiating a closureiter involves filling out its `:up`, which was
previously done incorrectly. The fixed algorithm is to use "current" env
if it is the owner of the iter declaration, or traverse through `:up`s
of env param until the common ancestor is found.
---------
Co-authored-by: Andreas Rumpf <rumpf_a@web.de>
It seems exportcpp was implemented in v1.0 but there is no documentation
about it excepts changelog.
`noinline` is used in many procedures in Nim code but there is also no
documentation about it.
---------
Co-authored-by: Andreas Rumpf <rumpf_a@web.de>
fixes#24305, refs #23807
Since #23014 `nkHiddenAddr` is produced to fast assign array elements in
iterators. However the array access inside this `nkHiddenAddr` can get
folded at compile time, generating invalid code. In #23807, compile time
folding of regular `addr` expressions was changed to be prevented in
`transf` but `nkHiddenAddr` was not updated alongside it.
The method for preventing folding in `addr` in #23807 was also faulty,
it should only trigger on the immediate child node of the address rather
than all nodes nested inside it. This caused a regression as outlined in
[this
comment](https://github.com/nim-lang/Nim/pull/24322#issuecomment-2419560182).
To fix both issues, `addr` and `nkHiddenAddr` now both shallowly prevent
constant folding for their immediate children.
fixes#24319
`byRefLoc` (`mapType`) requires the Loc `a` to have the right type.
Without `lfEnforceDeref`, it produces the wrong type for `deref (var
array)`, which may come from `mitems`.
fixes#18896fixes#20886
```nim
type
PFile {.importc: "FILE*", header: "<stdio.h>".} = distinct pointer
# import C's FILE* type; Nim will treat it as a new pointer type
```
This is an excerpt from the Nim manual. In the old Nim versions, it
produces a void pointer type instead of the `FILE*` type that should
have been generated. Because these C types tend to be opaque and adapt
to changes on different platforms. It might affect the portability of
Nim on various OS, i.e. `csource_v2` cannot build on the apline platform
because of `Time` relies on Nim types instead of the `time_t` type.
ref https://github.com/nim-lang/Nim/pull/18851
follows up #24259
This is the remaining missing use of `StructInitializer` in
`getDefaultValue` after #24259 and #24302. The only remaining direct C
code in getDefaultValue is [this
line](922f7dfd71/compiler/ccgexprs.nim (L3525))
which creates a global array variable, which isn't implemented yet. Next
steps would be all remaining variable and `typedef` declarations, then
hopefully we can move on to general statements and expressions.
follows up #24259
This was the only use of the `STRING_LITERAL` macro in `nimbase.h`, so
this macro is now removed. We don't have to remove it though, maybe
people use it.
`StructInitializer` is now used for most braced initializers in the C
generation, mostly in `genBracedInit`, `getNullValueAux`,
`getDefaultValue`. The exceptions are:
* the default case branch initializer for objects uses C99 designated
initializers with field names, which are not implemented for
`StructInitializer` yet (`siNamedStruct`)
* the uses in `ccgliterals` are untouched so all of ccgliterals can be
done separately and in 1 go
There is one case where `genBracedInit` does not use cbuilder, which is
the global literal variable for openarrays. The reason for this is
simply that variables with C array type are not implemented, which I
thought would be best to leave out of this PR.
For the simplicity of the implementation, code in `getNullValueAuxT`
that reset the initializer back to its initial state if the `Sup` field
did not have any fields itself, is now disabled. This was so the
compiler does not generate `{}` for the Sup field, i.e. `{{}}`, but
every call to `getNullValueAuxT` still generates `{}` if the struct
doesn't have any fields, so I don't know if it really breaks anything.
The case where the Sup field doesn't have any fields but the struct does
also would have generated `{{}, field}`.
Worst case, we can implement either the "resetting" or just disable the
generation of the `Sup` field if there are no fields total. But a better
fix might be to always generate `{0}` if the struct has no fields, in
line with the `char dummy` field that gets added for all objects with no
fields. This doesn't seem necessary for now but might be for the NIFC
output, in which case we can probably keep the logic contained inside
cbuilder (if no fields generated for `siOrderedStruct`/`siNamedStruct`,
we add a `0` for the `dummy` field). This would stipulate that all uses
of struct initializers are exhaustive for every field in structs.
fixes#24296fixes#24295
Templates use `expectedType` for type inference. It's justified that
when templates don't have an actual return type, i.e., `untyped` etc.
When the return type of templates is specified, we should not infer the
type
```nim
template g(): string = ""
let c: cstring = g()
```
In this example, it is not reasonable to annotate the templates
expression with the `cstring` type before the `fitNode` check with its
specified return type.
follows up #24279
`discard finishTest` was wrong if the test still had a `retries` option:
it would just ignore the result of the test. This is an unlikely mistake
but we safeguard against it by splitting `finishTest` into two, one that
completely ignores the retries option and `finishTestRetryable` which
has to be checked for a retry. This also makes the code look slightly
better.
Testament now retries a test by a specified amount if it fails in any
way other than an invalid spec. This is to deal with the flaky GC tests
on Windows CI that fail in many different ways, from the linker randomly
erroring, segfaults, etc.
Unfortunately I couldn't do this cleanly in testament's current code.
The proc `addResult`, which is the "final" proc called in a test run's
lifetime, is now wrapped in a proc `finishTest` that returns a bool
`true` if the test failed and has to be retried. This result is
propagated up from `cmpMsgs` and `compilerOutputTests` until it reaches
`testSpecHelper`, which handles these results by recursing if the test
has to be retried. Since calling `testSpecHelper` means "run this test
with one given configuration", this means every single matrix
option/target etc. receive an equal amount of retries each.
The result of `finishTest` is ignored in cases where it's known that it
won't be retried due to passing, being skipped, having an invalid spec
etc. It's also ignored in `testNimblePackages` because it's not
necessary for those specific tests yet and similar retry behavior is
already implemented for part of it.
This was a last resort for the flaky GC tests but they've been a problem
for years at this point, they give us more work to do and turn off
contributors. Ideally GC tests failing should mark as "needs review" in
the CI rather than "failed" but I don't know if Github supports
something like this.
fixes#24288, refs #23625
Since #23625 "cannot evaluate" errors during VM code generation are
"soft" errors with `nim check`, i.e. the code generation isn't halted
(except for the current proc which `return`s which can cause wrong
codegen) and the expression is still attempted to be evaluated. Now,
these errors signal to the VM that the current generated VM code cannot
be evaluated, and so instead of evaluating, an error node is returned.
This keeps the benefit of the "soft" errors without potentially crashing
the compiler on improperly generated VM code. Although maybe the
compiler might not be able to handle the generated error node in some
cases.
This fixes the chame example in #24288 but this is not tested in CI.
Presumably it or the compiler was doing something like `compiles()` on
code that can't run in the VM.
I would accept nicer ways of tracking non-evaluability than
`c.cannotEval = true` but I tried to keep it as harmless as possible.
fixes#24021
The field checking for case object branches at some point generates a
negated set `contains` check for the object discriminator. For enum
types, this tries to generate a complement set and convert to a
`contains` check in that instead. It obtains this type from the type of
the element node in the `contains` check.
`buildProperFieldCheck` creates the element node by changing a field
access expression like `foo.z` into `foo.kind`. In order to do this, it
copies the node `foo.z` and sets the field name in the node to the
symbol `kind`. But when copying the node, the type of the original
`foo.z` is retained. This means that the complement is performed on the
type of the accessed field rather than the type of the discriminator,
which causes problems when the accessed field is also an enum.
To fix this, we properly set the type of the copied node to the type of
the kind field. An alternative is just to make a new node instead.
A lot of text for a single line change, I know, but this part of the
codebase could use more explanation.
fixes#24087, refs https://forum.nim-lang.org/t/341, refs #14222, refs
#14221
The Nim compiler calls `cl` for linking as well as compilation. This
means that options to the linker have to be passed after a `/link`
argument. But the Nim compiler doesn't include this option normally,
because users may still want to pass non-linker options to `cl` at link
time.
To deal with this, a workaround is used: every single library link
option adds `/link` before it. The linker simply ignores extraneous
`/link` arguments and gives a warning instead, since it's an
unrecognized option to the linker. This is really hacky but otherwise we
need to separate linker arguments into arguments passed either to the
compiler or to the linker at link time, and this behavior wouldn't be
meaningful outside of MSVC.
I can't really test this manually but I did test that the linker ignores
`/link`. I also can't really do more than this, I don't really use MSVC
so I wouldn't know how to navigate it, or how people use it. Ideally
someone who knows more about/uses MSVC can give their input or take
over.
