Long story short, even if the type contains no reference at all to its
generic parameters we still have to walk its AST and evaluate any
nkRecWhen nodes that semRecordNodeAux skipped due to the type being a
generic one.
We also must be careful to modify the type `n` node in place since it
may be referenced by the caller as seen in the tillegaltyperecursion
test.
Moreover we also can't have the nkSym drift away from their original
values in order not to break the JS nkObjConstr codegen.
This fixes `jsondoc0` so that it has comments in the output as described
by the documentation. It also fixes `jsondoc`/`jsondoc2` as it stored
it's output in a file with an html extension. Along with this it also
adds a new field `moduleDescription` to `jsondoc`/`jsondoc2` which
contains the module description.
* Early evaluation of mIs
The `evalIs` implementation was just a broken copy of `isOpImpl` so
let's just avoid it alltogether: `mIs` nodes are either resolved during
the semantic phase or bust.
* Remove dead code and tidy it up
* Add `toOpenArray[T](ptr UncheckedArray[T])` for clarity. `ptr array[0,T]`
for some unchecked type already works but A) `UncheckedArray` seems to be
the intended future way for this kind of access, and B) essentially all use
cases will have a `ptr` for that kind of array source and this call signature
lets callers drop the trailing `[]` corresponding to that `ptr` deref.
This PR relates to issue https://github.com/nim-lang/Nim/issues/9001 .
* Add a test for toOpenArray() for UncheckedArray[T]s.
- Go's write barriers are now plugged-in in all the relevant points
- "gcGo" is correctly classified by usesWriteBarrier()
- some gogc structures and functions now use golib wrappers to keep GCC
version-specific conditions out of the compiler/stdlib code
- we no longer allow mixing the C malloc with Go's
- fix a problem with string copying
