| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
|
| |
we need to handle a Scala version like `2.13.0-pre-5e84129`,
as we see in nightly builds before bincompat is locked down
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
e.g `val x2 = Foo[$1] with Bar = boundValue`
is rewritten to `val x2 = (Foo[$1] @uncheckedBounds) with Bar = boundValue`
This is to have refchecks turn a blind eye to the type argument that
doesn't conform the to type parameter bounds.
For regular compilation, without the async transform between patmat
and refchecks, bound conformance is disabled with:
https://github.com/scala/scala/blob/v2.11.7/src/compiler/scala/tools/nsc/typechecker/RefChecks.scala#L1743
Using the `uncheckedBounds` annotation is a newer, more inclusive way
of acheiving the same thing:
https://github.com/scala/scala/blob/v2.11.7/src/compiler/scala/tools/nsc/typechecker/RefChecks.scala#L1677
|
|\
| |
| | |
Fix decision about whether to use a trait or class as the parent
|
| | |
|
| | |
|
|/
|
|
|
| |
Unfortunately I wasn't able to extract a test case, but the patch
has been tested to fix a problem on a real world code base.
|
| |
|
|
|
|
| |
After this change tests pass also on Windows.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- method local lazy vals are now encoded as a single ValDef
rather than a ValDef + DefDef pair. We need to treat ValDef-s
with the LAZY flag in the same way as we used to treat the
DefDef.
- Rename one of the symbols `ANF,anf` in the same scope to avoid
generating anonymous class names that differ only in case. The
compiler warned about this one.
- When patching the LabelDefs to have a `Unit` result type, propagate
this other LabelDefs conclude with a jump to that label. Not sure
why, but without this we now hit an error in the backend about the
nonsensical attempt to emit a coercion from void to int.
- Use crossScalaVersions in the build and update the Scala versions
tested in CI.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Detect cross-state symbol references where the RefTree is nested
in a LabelDef. Failure to do so led to ill-scoped local variable
references which sometimes manifest as VerifyErrors.
- Emit a default case in the Match intended to be a tableswitch.
We have to do this ourselves if we expand after pattern matcher
- Cleanup generated code to avoid redundant blocks
- Avoid unnecessary `matchRes` temporary variable for unit-typed
pattern matches
- Fix the trace level logging in the ANF transform to restore
indented output.
- Emit `{ state = nextState; ... }` rather than
`try { ... } finally { state = nextState }` in state handlers.
This simplifies generated code and has the same meaning, as the
code in the state machine isn't reentrant and can't observe the
"early" transition of the state.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The presentation compiler runs with `-Ymacro-expand:discard`, which
retains the macro expandee in the typechecked trees, rather than
substituting in the expansion. This mode was motivated as a means
to keep IDE functionality working (e.g. completion, navigation,
refactoring) inside macro applications.
However, if one has nested async macro applications, as reported in
the IDE ticket:
https://www.assembla.com/spaces/scala-ide/tickets/1002561
... the expansion of the outer async application was reporting
await calls enclosed by the inner async application.
This change tweaks the traversers used for this analysis to
stop whenever it sees an async.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, the async transformation is performed during the typer
phase, like all other macros.
We have to levy a few artificial restrictions on whern an async
boundary may be: for instance we don't support await within a
pattern guard. A more natural home for the transform would be
after patterns have been translated.
The test case in this commit shows how to use the async transform
from a custom compiler phase after patmat.
The remainder of the commit updates the implementation to handle
the new tree shapes.
For states that correspond to a label definition, we use `-symbol.id`
as the state ID. This made it easier to emit the forward jumps to when
processing the label application before we had seen the label
definition.
I've also made the transformation more efficient in the way it checks
whether a given tree encloses an `await` call: we traverse the input
tree at the start of the macro, and decorate it with tree attachments
containig the answer to this question. Even after the ANF and state
machine transforms introduce new layers of synthetic trees, the
`containsAwait` code need only traverse shallowly through those
trees to find a child that has the cached answer from the original
traversal.
I had to special case the ANF transform for expressions that always
lead to a label jump: we avoids trying to push an assignment to a result
variable into `if (cond) jump1() else jump2()`, in trees of the form:
```
% cat sandbox/jump.scala
class Test {
def test = {
(null: Any) match {
case _: String => ""
case _ => ""
}
}
}
% qscalac -Xprint:patmat -Xprint-types sandbox/jump.scala
def test: String = {
case <synthetic> val x1: Any = (null{Null(null)}: Any){Any};
case5(){
if (x1.isInstanceOf{[T0]=> Boolean}[String]{Boolean})
matchEnd4{(x: String)String}(""{String("")}){String}
else
case6{()String}(){String}{String}
}{String};
case6(){
matchEnd4{(x: String)String}(""{String("")}){String}
}{String};
matchEnd4(x: String){
x{String}
}{String}
}{String}
```
|
|
|
|
| |
And remove unused code.
|
|
|
|
|
| |
This avoids leaving .class files in the working directory
after running the test.
|
|
|
|
|
|
| |
By declararing the parameter of `async` as by-name.
Fixes #150 (the bug in the original ticket.)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If we blindly splicing `{..$stats, ..$generatedCode}`, and the last
expression in `$stats` is of type `Nothing`, we'll incur a dead
code warning when typechecking the block.
This commit:
- introduces a helper method to augment user-written stats with
synthetic code
- Emit a try/finally in that code (so we advance the state, even if we
are about to exit the state machine in the async-block global
exception handler
- Hide `Nothing` typed expressions from the dead code analysis
by wrapping them in an `expr: Any`
Fixes #150 (the part reported in the comments, not the original ticket.)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Code like:
val x = if (cond) throw new A else throw new B
Was being transformed to:
val ifRes = ???
if (cond) ifRes = throw new A else ifRes = throw new B
val x = ifRes
by way of the use of `gen.mkZero` which throws `???` if the requested type is `Nothing`
This commit special cases `Nothing` typed expressions in a similar manner to `Unit` type expressions.
The example above is now translated to:
if (cond) throw new A else throw new B
val x = throw new IllegalStateException()
Fixes #120
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
During the ANF transform, we were generating a tree of the shape:
{
val temp: Unit = await(futureOfUnit)
temp
()
}
I tried to simplifiy this to avoid creating the temporary value,
but this proved difficult as it would have required changes to
the subsequent state machine transformation.
Even replacing `temp` with `()` made the state machine transform
harder.
So for now, I've just inserted `temp.asInstanceOf[Unit]` to hide
from the compiler warning.
Fixes #74
|
|
|
|
|
|
| |
This commit disabled live variable analysis for intermediate values of type Nothing.
Fixes #104
|
|
|
|
|
|
|
| |
We were leaking untyped trees out of the macro, which crashed
in refchecks.
This commit proactively typechecks the tree returned by `mkZero`.
|
|\
| |
| |
| |
| |
| |
| |
| |
| | |
merge/2.10.x-to-master-20141219
Conflicts:
src/main/scala/scala/async/internal/AsyncTransform.scala
src/main/scala/scala/async/internal/ExprBuilder.scala
src/test/scala/scala/async/TreeInterrogation.scala
|
| |
| |
| |
| |
| |
| | |
A worthy optimization, suggested by @danarmak.
Closes #73
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Previously, as sequence of state transitions that did not pass through
an asynchrous boundary incurred stack frames. The trivial loop in
the enclosed test case would then overflow the stack.
This commit merges the `resume` and `apply(tr: Try[Any])` methods into
a `apply`. It changes the body of this method to be an infinite loop
with returns at the terminal points in the state machine (or at a
terminal failure.)
To allow merging of these previously separate matches, states that
contain an await are now allocated two state ids: one for the setup
code that calls `onComplete`, and one for the code in the continuation
that records the result and advances the state machine.
Fixes #93
|
| |
| |
| |
| | |
E.g. type param, abstrat type.
|
| | |
|
|\|
| |
| |
| |
| |
| |
| | |
Conflicts:
src/main/scala/scala/async/internal/AnfTransform.scala
src/main/scala/scala/async/internal/AsyncTransform.scala
src/test/scala/scala/async/run/toughtype/ToughType.scala
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
`TreeGen#mkZero` returns `q"null"` for derived value classes.
```
scala> class V(val a: String) extends AnyVal
defined class V
scala> showRaw(gen.mkZero(typeOf[V]))
res0: String = Literal(Constant(null))
```
We use this API in async to generate the initial value for
ANF-lifted temporary variables.
However, this leads to NPEs, as after posterasure, we call the
unbox method on a null reference:
```
% cat sandbox/Macro.scala; scalac-hash v2.10.4 sandbox/Macro.scala; scala-hash v2.10.4 -e 'val x = Macros.myMacro'
import scala.reflect.macros.Context
import scala.language.experimental.macros
object Macros {
def macroImpl(c: Context): c.Expr[C] = {
import c.universe._
val e1 = c.Expr[C](Literal(Constant(null)).setType(typeOf[C]))
reify(e1.splice.asInstanceOf[C @annotation.unchecked.uncheckedVariance])
}
def myMacro: C = macro macroImpl
}
class C(val a: String) extends AnyVal
java.lang.NullPointerException
at Main$$anon$1.<init>(scalacmd4059893593754060829.scala:1)
at Main$.main(scalacmd4059893593754060829.scala:1)
at Main.main(scalacmd4059893593754060829.scala)
```
This commit installs a custom version of `mkZero` that instead
returns `q"new C[$..targs](${mkZero(wrappedType)})`.
Thanks to @ewiner for pinpointing the problem.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
This progressed along with the fix for #66.
`TreeGen.mkZero` is a bit of a minefield: first with `Nothing` and
now with Value Classes. I wonder if we can provoke the same sort of
bug in the compiler in places where this is used.
Closes #83
|
|\|
| |
| |
| |
| |
| |
| |
| | |
merge/2.10.x-to-master-20140721
Conflicts:
src/main/scala/scala/async/internal/AsyncTransform.scala
src/main/scala/scala/async/internal/Lifter.scala
|
| |\
| | |
| | | |
Fix regression around type skolems and if exprs.
|
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
If we start with:
async({
val res = await[S[_$1 with String]](s);
if (true)
await[Int](0)
res
})
Typechecking the application (before macro expansion) yields
(where the trees are printed in the form `expr{tpe}`):
async[S[_$1#5738 with String#137]]({
val res: S[_$1#5490 with String] forSome { type _$1#5490 } =
await[S[_$1#5487 with String]](
s{S[_$1#5487 with String]}
){S[_$1#5487 with String]};
if (true)
await(0)
else
()
res{S[_$1#5738 with String]}
}{S[_$1#5738 with String]}){S[_$1#5738 with String]}
Note that the type of the second last line contains a skolemized
symbol `_$1#5738` of the existential `_$1#5490`. This is created
by this case in `Typer#adapt`:
case et @ ExistentialType(_, _) if ((mode & (EXPRmode | LHSmode)) == EXPRmode) =>
adapt(tree setType et.skolemizeExistential(context.owner, tree), mode, pt, original)
Our ANF rewrites part of this code to:
<synthetic> val await$1: S[_$1#5487 with String] = await[S[_$1#5487 with String]](awaitable$1);
val res: S[_$1#5490 with String] forSome { type _$1 } = await$1;
And later, the state machine transformation splits the last line into
a blank field and an assignment. Typechecking the `Assign` node
led to the an type error.
This commit manually attributes the types to the `Assign` node so
as to avoid these problem.
It also reigns in an overeager rewriting of `If` nodes in the
ANF transform, which was due to a bug in the label detection
logic introduced in 4fc5463538.
Thanks to @gnovark for yet another devilish test case and
analysis of the problem with label detection.
I worked on a more principled fix on:
https://github.com/retronym/async/compare/ticket/79-2?expand=1
in which I try to use `repackExistential` to convert skolemized
types to existentials for use as the types of synthetic vals
introduced by the ANF transform. This ran into a deeper problem
with existential subtyping in the compiler itself though.
|
| |\ \
| | | |
| | | | |
Fix asymptotic performance issues in live variables analysis.
|
| | |/
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Fix possibly-exponential runtime for DFS graph searches.
Improve DFA fixpoint algorithm to correctly compute worklist
of only changed nodes for each iteration.
Added test that takes > 2 minutes to compile without these
improvements.
|
| |/
| |
| |
| |
| |
| |
| | |
`gen.mkZero(NothingTpe)` gives the tree `Predef.???`. Instead, we should leave
the `await` field uninitialized with `ValDef(..., rhs = EmptyTree)`.
Fixes #66
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
- remove unneeded `setType(NoType)`, which was leftover from my
first attempts to find this bug.
- fix typo in error message
- optimize imports
(cherry picked from commit 5c6ea29966fa80faae13892da50fc68ed1bf9ae7)
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
We were incorrectly typechecking the `ClassDef` of the state machine
in the macro in a way that discarded the resulting trees, and only
kept around the symbol.
The led to the the macro engine retypechecking
that node, which somehow led to duplicated lazy val initiaializer
`DefDef`-s in the template, which manifest as a `VerifyError`.
This commit:
- rescues the typechecked `ClassDef` node from the eager
typechecking by the macro
- loosens the restriction on lazy vals in async blocks. They are
still prohibited if they contain an await on the RHS
- Adds a test that shows evalution is indeed lazy.
(cherry picked from commit cc4587b1985519f7049d0feb0783d8e22c10f792)
Conflicts:
src/main/scala/scala/async/internal/AsyncAnalysis.scala
src/main/scala/scala/async/internal/AsyncTransform.scala
|
| |
| |
| |
| |
| |
| | |
Works on the 2.10.x branch, so just backprting the test.
Cherry picked from 6f6546ebfc26564843621e79d840209a5103d3c8.
|
| | |
|
| |
| |
| |
| |
| |
| |
| | |
- remove unneeded `setType(NoType)`, which was leftover from my
first attempts to find this bug.
- fix typo in error message
- optimize imports
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
We were incorrectly typechecking the `ClassDef` of the state machine
in the macro in a way that discarded the resulting trees, and only
kept around the symbol.
The led to the the macro engine retypechecking
that node, which somehow led to duplicated lazy val initiaializer
`DefDef`-s in the template, which manifest as a `VerifyError`.
This commit:
- rescues the typechecked `ClassDef` node from the eager
typechecking by the macro
- loosens the restriction on lazy vals in async blocks. They are
still prohibited if they contain an await on the RHS
- Adds a test that shows evalution is indeed lazy.
Fixes #52
|
| |
| |
| |
| |
| |
| | |
Predicate the `asClass` cast with an `isClass` check.
Fixes #63
|
| | |
|
|/ |
|
|
|
|
| |
2013 must have been unlucky.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The new tree shapes handled for do/while look like:
// type checked
async({
val b = false;
doWhile$1(){
await(());
if (b)
doWhile$1()
else
()
};
()
})
We had to change ExprBuilder to create states for the if/else
that concludes the doWhile body, and also loosen the assertion
that the label jump must be the last thing we see.
We also have to look for more than just `containsAwait` when
deciding whether an `If` needs to be transformed into states;
it might also contain a jump to the enclosing label that is on
the other side of an `await`, and hence needs to be a state
transition instead.
|
|
|
|
|
|
|
|
|
| |
These stem from the handling of the internal/external view
or method type parameters by `thisMethodType` in `Namers`.
I've now preseversed the orginal ValDefs favoured the latter
when constructing the new DefDef, and made construction of
all liftables consistent in this regard.
|
|
|
|
|
| |
Once they escape, we leave the references in the state
machines fields untouched.
|
|
|
|
|
|
|
|
|
|
| |
- Remove the CPS fallback version of async. That was not intended
to be part of 1.0.
- Lookup the await method beside the macro, rather than requiring
all calls to go to AsyncBase.await.
- Create a minimal version of Async that just contains await/async
and delegates to the macro implementation in internal._
- Add scaladoc.
|
|
|
|
|
|
| |
We were relying on an internal API that no longer exists.
We also need to tweak the way our tests infer scalaBinaryVersion.
|