| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
|/ |
|
|\
| |
| | |
Add/unpickling
|
| | |
|
| | |
|
| |
| |
| |
| |
| | |
MODULE -> OBJECT, as suggested by Simon.
Header is 4 hex digits, not 8 chars.
|
| | |
|
| |
| |
| |
| |
| | |
1) Avoiud serializing redunant flags that come with modules
2) Drop DEFAULTinit flag - it's not used.
|
| |
| |
| |
| | |
It is always empty anyway. Side benefit: We can get rid of NOTYPE, too.
|
| |
| |
| |
| |
| | |
Remove 3rd argument of BIND types. It can be reconstructed
on unpickling.
|
| | |
|
| |
| |
| |
| |
| | |
1) Drop the length field of BYNAMEparam
2) Eliminate EMPTYTREE
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
1) DefDefs with implicit method types generate implicit parameters
2) Super accessors are inserted after class parameters.
With these changes files in core also print the same after pickling,
with the exception of Types and TypeOps which have some spurious differences:
Types are equal but appear in more simplified form after pickling.
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The unpickled term in a package id is different from the original but the
type is the same.
In a literal prefer constants in the types over thsoe in the terms.
With this change, we get identical typed tree output also for core/pickled.
|
| |
| |
| |
| |
| | |
1) Prefixes in selections are widened if unstable
2) DefDefs get Method flag set
|
| |
| |
| |
| |
| | |
It's not used and is too low-level anyway. Expanded names should
be a semantic concept, the choice of separator is irrelevant.
|
| |
| |
| |
| |
| | |
Would like to do use more semantic names in pickled file, but it's hard without
having semantic names to start with.
|
| |
| |
| |
| |
| |
| |
| |
| | |
We got some spurious differences in the types in TemplateParents
because simplification was done in Typer but not in Unpickler.
With the change we get perfect matches for all files in pickleOK
also if we print their types.
|
| |
| |
| |
| | |
The type of a parameter alias is an ExprType, was the underlying type.
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Shadowed names in types need to be pickled and treated on
unpickling.
We choose to make Shadowed a separate TastyName class, to avoid
the ad-hoc name-mangling in current dotc. When names are redone
Shadowed will also become a special class in the compiler proper.
|
| |
| |
| |
| | |
The new name is more consistent with the usage (e.g. NamedType#shadowed).
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| | |
Need to pickle the type, not the constant stored in the literal,
because constant folding might have adapted the type, but leaves
the constant value unchanged. An example is in pickleOK/unions.scala
where a float 1.0f was adapted to Double, so should be pickled as
a Double 1.0.
|
| |
| |
| |
| |
| | |
Need to enter all top-level symbols before reading any of them, in order
to avoid stale symbols.
|
| |
| |
| |
| |
| |
| |
| | |
We now unpickle all files first, and print afterwards.
This helps in avoiding stale symbols that happen otherwise
when we print a reference to a symbol that has not yet been unpickled.
But it's not yet enough.
|
| |
| |
| |
| |
| | |
Implicits get pickled only for terms. So implicit class
prints differently when unpickled.
|
| | |
|
| |
| |
| |
| |
| |
| | |
Should get an () parameter list in both cases. (In fact, that's pretty annoying,
it woul be better if we did not have this restriction. But in any case it has to
be the same for Namer and Unpickler.)
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Out-of scope references are evil but do arise. Dotc currently produces them
in two cases (t1957.scala and Typer.scala) because it does not correctly
hygenize dependent methods used as closurs. I beleive scalac wil generate them
in more instances.
The changes can produce dependent method types used in closures. Essentially the
problem was that the depndency was not recognized in Typers when the closure
was first created, because it badly interfered with type inference. But when unpickling
the same closure the dependency is recognized. It's too late to fail now, we better
deal with this gracefully. That's why there is now a mode bit "AllowDependentFunctions"
which should be turned on only for unpickling, which reconstitutes dependent functions
as closures without complaining.
|
| |
| |
| |
| | |
The previous setting did not generate same trees for Typer.scala
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
A homogenized view makes sure that there are no spurious
differences between trees before and after pickling.
Currently there are two cases where homogenized is different from
standard view:
1) Backquoted idents are not printed in backticks
2) Instantiated TypeVars so not get an apostroph.
|
| |
| |
| |
| |
| |
| | |
- treated nested Matches correctly
- treat nested packages correctly
- SELECT always needs to select with sig
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Selftypes can be nastily recursive as this example of extmethods shows:
class T[A, This <: That1[A]] extends AnyVal { self: This => ...
The problem, of course, is that This is really `this.This` and to determine
what that is we look up the self type which leads us right back to where we were.
We fix the problem by special casing references to class members in self-types.
These are always written symbolically, and compelmentary logic in unpickling
makes sure we do not need to compute the denotation of such members, but
instead take the denotation directly from the symbol.
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
ElimLocals becomes a slightly less trivial transform: NormalizeFlags.
It also computes PureInterface flag, thus relieving Namer and Unpickler
from doing the same in two different ways. Besides, the computation in
Namer/TreeInfo was flawed because it did not take into account that
nested non-static classes are not allowed in an interface (only static
classes are, but these would not be members of the interface in the Scala
sense).
|
| |
| |
| |
| |
| |
| | |
- treat applied higher-kinded types specially when pickling (
drop #Apply projection)
- pickle the original self info sintead of the processed self type
|
| |
| |
| |
| |
| |
| |
| | |
Use fewer length fields in type encodings.
- create new category of tags that take exactly one argument tree.
- avoid using length where a tag fits in one of theopther categories.
|
| |
| |
| |
| |
| |
| |
| |
| | |
Also: Make Pickler a plain phase; it is neither a macro transformer nor
a miniphase.
Add tests to pickleOK that are known to be stable under pickling/unpicking
by comparing tree representations via show.
|
| | |
|
| | |
|
| |
| |
| |
| | |
To avoid spurious differences and allow all flags in modifiers to be printed.
|
| |
| |
| |
| |
| |
| | |
Only roots should have their infos overwritten; other symbols in the
root scope should be replaced by same-named symbols in the Tasty file.
If we do not do this, we end up with inconsistent caches in root symbols.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The Unpickler visits nodes in a certain order and (so far) expects to see
definitions before uses. This commit makes sure that the Unpickler generates
definitions in an order that matches the pickler's behavior. Every definition
should be "pre-registered" before it can be used. This still allows for mutual
recursion between symbols because preregistering enters all symbols of a scope
in bulk before generating any references to these symbols.
It would be nice if this was the end of the story, but unfortunately that's not the case.
It turned out that dotc produced references that were not legal according to the
implemented model, and that are also not legal in the formal type system. Each of these
violations referred to a symbol outside its scope. There were two sources:
1) Pattern-bound variables were retained in the type of a case block and, consequently,
the type of the enclosing match. This has been fixed by a previous commit a0f8ec21c9ce5381bea780e7be89286653fc676e.
2) Dependent anonymous functions led to (illegal) dependent closures with references
to their parameters leaking out in the environment. This has been mostly fixed by the previous commits,
in particular 1c70842036b083652c3eeab83aad0b2490674bfe. But there are still two problems remaining, see:
89c8bd8a1eb9fb3f0f09f25bedb68de1ef2e2ae8.
We might fix the two problems. But it's inconceivable to me that scalac could also
produce only "hygienic" trees that do not have escaping references. There are too many situations
where we know this is not true (existential skolems, for a start).
So we choose to flag escaping variables in logs instead of treating them as errors, and to deal with the
situation in the Unpickler.
|
| |
| |
| |
| |
| |
| |
| | |
We have two unhygienic closures left - one in t1957.scala the other in Typer.scala.
This commit leaves some printlns that can be uncommented to get more info on these.
It would be better to fix them but I am running out of time to do so. Maybe someone
else can pick up with the info this commit allows to recover.
|
| |
| |
| |
| |
| | |
We now make sure that a closure's result type does avoid
references to parameter types.
|
| |
| |
| |
| |
| |
| |
| | |
If a result type is not fully defined (i.e. has type variables),
dependency status depends on the instantiation of the variables.
In this case we now still ignore the variable but do not cache
the status for future uses.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The previous logic for avoiding dependent method types in closures had a hole.
The problem arose when the expected return type of a closure was a type variable.
Then, the that type variable would be taken as the declared result type of the
closure without (at first) checking the body. The type variable would not yet
be bounded and therefore would not represent a dependent method type. Afterwards
when typechecking the closure the type variable woul dbe bounded and instantiated.
But at that time, all checking and possibly avoiding of depenencies has already
happened.
We solve the problem by typing the body of a closure during Namer whenever the expected
type is not fully defined.
Doing so uncovered a problem that anonymus function were seen as the target of returns
(tehy shoul be skipped instead). Thsi problem is fixed by a patch to SymDenotations#isSourceMethod.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
According to the "new" interpretation of T#U as \exists x: T. x.U,
x.U is not a subtype of T#U. Hence variance in prefixes of named
types should be 0.
There is one case where this leads to problems. (Typing failure in
Desugar.scala, to be exact). This is when computing variances
in local type inference. We solve this problem locally
by overriding accumulator application to prefixes in the specific
variances accumulator. Because the "variances" accumulator is now
irregular wrt to the rest of variance handling, we move it
from Types to Inferencing, which is the only place where it is needed.
|