| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The REPL has a long running instance of Global which outputs
classfiles by default to a VirtualDirectory. The inliner did not find
any of these class files when compiling calls to methods defined in
previous runs (ie, previous lines of input.)
This commit:
- Adds a hook to augment the classpath that the optimizer searches,
and uses this in the REPL to add the output directory
- Fixes the implementation of `findClassFile` in VirtualDirectory,
which doesn't seem to have been used in anger before. I've factored out
some common code into a new method on `AbstractFile`.
- Fixes a similar problem getSubDir reported by Li Haoyi
- Adds missing unit test coverage.
This also fixes a bug in REPL autocompletion for types defined
in packages >= 2 level deep (with the `:paste -raw` command).
I've added a test for this case.
|
|
|
|
|
|
|
| |
... by calling javaBinaryNameString, instead.
They all are happy with a throw away String, there is no advantage
to interning this into the name table.
|
|
|
|
|
|
| |
Rather than putting the code of a trait method body into a static method,
leave it in the default method. The static method (needed as the target
of the super calls) now uses `invokespecial` to exactly call that method.
|
|\
| |
| |
| |
| | |
SD-20 Inlcude static methods in the InlineInfo in mixed compilation
Fixes scala/scala-dev#20
|
| |
| |
| |
| |
| |
| |
| |
| | |
In mixed compilation, the InlineInfo for a Java-defined class is created
using the class symbol (vs in separate compilation, where the info is
created by looking at the classfile and its methods). The scala compiler
puts static java methods into the companion symbol, and we forgot to
include them in the list of methods in the InlineInfo.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
There isn't much point to the late* flags in a world where
we're mutating flags left and right in tree and info transformers...
So, lets get rid of the indirection until we can include flags
in a symbol's type history, like we do for its info.
This retires lateDEFERRED (redundant with SYNTHESIZE_IMPL_IN_SUBCLASS).
Since it's introduced so late, it makes little sense to have these
synthetic members go back to DEFERRED. Instead, just set DEFERRED directly.
Also remove unused late* and not* flags.
notPRIVATE subsumes lateFINAL for effective finality (scala/scala-dev#126)
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
One step towards teasing apart the mixin phase, making
each phase that adds members to traits responsible for
mixing in those members into subclasses of said traits.
Another design tenet is to not emit symbols or trees
only to later remove them. Therefore, we model a
val in a trait as its accessor. The underlying field
is an implementation detail. It must be mixed into
subclasses, but has no business in a trait (an interface).
Also trying to reduce tree creation by changing less in subtrees
during tree transforms.
A lot of nice fixes fall out from this rework:
- Correct bridges and more precise generic signatures for
mixed in accessors, since they are now created before erasure.
- Correct enclosing method attribute for classes nested in trait fields.
Trait fields are now created as MethodSymbol (no longer TermSymbol).
This symbol shows up in the `originalOwner` chain of a class declared
within the field initializer. This promoted the field getter to
being the enclosing method of the nested class, which it is not
(the EnclosingMethod attribute is a source-level property).
- Signature inference is now more similar between vals and defs
- No more field for constant-typed vals, or mixed in accessors
for subclasses. A constant val can be fully implemented in a trait.
TODO:
- give same treatment to trait lazy vals (only accessors, no fields)
- remove support for presuper vals in traits
(they don't have the right init semantics in traits anyway)
- lambdalift should emit accessors for captured vals in traits,
not a field
Assorted notes from the full git history before squashing below.
Unit-typed vals: don't suppress field
it affects the memory model -- even a write of unit to a field is relevant...
unit-typed lazy vals should never receive a field
this need was unmasked by test/files/run/t7843-jsr223-service.scala,
which no longer printed the output expected from the `0 to 10 foreach`
Use getter.referenced to track traitsetter
reify's toolbox compiler changes the name of the trait
that owns the accessor between fields and constructors (`$` suffix),
so that the trait setter cannot be found when doing mkAssign in constructors
this could be solved by creating the mkAssign tree immediately during fields
anyway, first experiment: use `referenced` now that fields runs closer
to the constructors phase (I tried this before and something broke)
Infer result type for `val`s, like we do for `def`s
The lack of result type inference caused pos/t6780 to fail
in the new field encoding for traits, as there is no separate accessor,
and method synthesis computes the type signature based on the ValDef tree.
This caused a cyclic error in implicit search, because now the
implicit val's result type was not inferred from the super member,
and inferring it from the RHS would cause implicit search to consider
the member in question, so that a cycle is detected and type checking fails...
Regardless of the new encoding, we should consistently infer result types
for `def`s and `val`s.
Removed test/files/run/t4287inferredMethodTypes.scala and test/files/presentation/t4287c,
since they were relying on inferring argument types from "overridden" constructors
in a test for range positions of default arguments. Constructors don't override,
so that was a mis-feature of -Yinfer-argument-types.
Had to slightly refactor test/files/presentation/doc, as it was relying
on scalac inferring a big intersection type to approximate the anonymous
class that's instantiated for `override lazy val analyzer`.
Now that we infer `Global` as the expected type based on the overridden val,
we make `getComment` private in navigating between good old Skylla and Charybdis.
I'm not sure why we need this restriction for anonymous classes though;
only structural calls are restricted in the way that we're trying to avoid.
The old behavior is maintained nder -Xsource:2.11.
Tests:
- test/files/{pos,neg}/val_infer.scala
- test/files/neg/val_sig_infer_match.scala
- test/files/neg/val_sig_infer_struct.scala
need NMT when inferring sig for accessor
Q: why are we calling valDefSig and not methodSig?
A: traits use defs for vals, but still use valDefSig...
keep accessor and field info in synch
|
|/
|
|
|
| |
The constructor of scala.tools.asm.Handle now takes an additional
boolean parameter to denote whether the owner is an interface.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
And use this as the target of the default methods or
statically resolved super or $init calls.
The call-site change is predicated on `-Yuse-trait-statics`
as a stepping stone for experimentation / bootstrapping.
I have performed this transformation in the backend,
rather than trying to reflect this in the view from
Scala symbols + ASTs.
We also need to add an restriction related to invokespecial to Java
parents: to support a super call to one of these to implement a
super accessor, the interface must be listed as a direct parent
of the class.
The static method names has a trailing $ added to avoid duplicate
name and signature errors in classfiles.
|
| |
|
|
|
|
| |
Keep -Yopt-inline-heuristics and -Yopt-trace unchanged
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This commit fixes various issues with classOf literals and Java
annotations.
- Ensure that a Type within a ConstantType (i.e., a classOf literal)
is erased, so `classOf[List[Int]]` becomes `classOf[List]`.
- Ensure that no non-erased types are passed to `typeToBType` in the
backend. This happens for Java annotations: the annotation type and
`classOf` annotation arguments are not erased, the annotationInfos
of a symbol are not touched in the compiler pipeline.
- If T is an alias to a value class, ensure that `classOf[T]` erases
to the value class by calling `dealiasWiden` in erasure.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Until now, concrete methods in traits were encoded with
"trait implementation classes".
- Such a trait would compile to two class files
- the trait interface, a Java interface, and
- the implementation class, containing "trait implementation methods"
- trait implementation methods are static methods has an explicit self
parameter.
- some methods don't require addition of an interface method, such as
private methods. Calls to these directly call the implementation method
- classes that mixin a trait install "trait forwarders", which implement
the abstract method in the interface by forwarding to the trait
implementation method.
The new encoding:
- no longer emits trait implementation classes or trait implementation
methods.
- instead, concrete methods are simply retained in the interface, as JVM 8
default interface methods (the JVM spec changes in
[JSR-335](http://download.oracle.com/otndocs/jcp/lambda-0_9_3-fr-eval-spec/index.html)
pave the way)
- use `invokespecial` to call private or particular super implementations
of a method (rather `invokestatic`)
- in cases when we `invokespecial` to a method in an indirect ancestor, we add
that ancestor redundantly as a direct parent. We are investigating alternatives
approaches here.
- we still emit trait fowrarders, although we are
[investigating](https://github.com/scala/scala-dev/issues/98) ways to only do
this when the JVM would be unable to resolve the correct method using its rules
for default method resolution.
Here's an example:
```
trait T {
println("T")
def m1 = m2
private def m2 = "m2"
}
trait U extends T {
println("T")
override def m1 = super[T].m1
}
class C extends U {
println("C")
def test = m1
}
```
The old and new encodings are displayed and diffed here: https://gist.github.com/retronym/f174d23f859f0e053580
Some notes in the implementation:
- No need to filter members from class decls at all in AddInterfaces
(although we do have to trigger side effecting info transformers)
- We can now emit an EnclosingMethod attribute for classes nested
in private trait methods
- Created a factory method for an AST shape that is used in
a number of places to symbolically bind to a particular
super method without needed to specify the qualifier of
the `Super` tree (which is too limiting, as it only allows
you to refer to direct parents.)
- I also found a similar tree shape created in Delambdafy,
that is better expressed with an existing tree creation
factory method, mkSuperInit.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The previous encodings created static fields in the enclosing class
to host caches. However, this isn't an option once emit code in default
methods of interfaces, as Java interfaces don't allow private static
fields.
We could continue to emit fields, and make them public when added to
traits.
Or, as chosen in this commit, we can emulate a call-site specific
static field by using invokedynamic: when the call site is linked,
our bootstrap methid can perform one-time computation, and we can
capture the result in the CallSite.
To implement this, I've allowed encoding of arbitrary invokedynamic
calls in ApplyDynamic.
The encoding is:
ApplyDynamic(
NoSymbol.newTermSymbol(TermName("methodName")).setInfo(invokedType)
Literal(Constant(bootstrapMethodSymbol)) :: (
Literal(Constant(staticArg0)) :: Literal(Constant(staticArgN)) :: Nil
) :::
(dynArg0 :: dynArgN :: Nil)
)
So far, static args may be `MethodType`, numeric or string literals, or
method symbols, all of which can be converted to constant pool entries.
`MethodTypes` are transformed to the erased JVM type and are converted
to descriptors as String constants.
I've taken advantage of this for symbol literal caching and
for the structural call site cache.
I've also included a test case that shows how a macro could target this
(albeit using private APIs) to cache compiled regexes.
I haven't managed to use this for LambdaMetafactory yet, not sure
if the facility is general enough.
|
|\ |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Until now, the warning was only emitted for enums from Java class files.
This commit fixes it by
- aligning the flags set in JavaParsers with the flags set in
ClassfileParser (which are required by the pattern matcher to
even consider checking exhaustiveness)
- adding the enum members as childs to the class holding the enum
as done in ClassfileParser so that the pattern matcher sees the enum
members when looking for the sealed children of a type
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| | |
With GenBCode being the default and only supported backend for Java 8,
we can get rid of GenASM.
This commit also fixes/migrates/moves to pending/deletes tests which
depended on GenASM before.
|
|/
|
|
|
|
|
|
|
|
|
|
| |
The previous methods to identify method invocations that can be
optimized, such as `isPredefAutoBox`, were String-based. Now we
obtain class and method signatures from symbols through the
BTypes infrastructure.
We also piggy-back on specialization's type transformer to create
all specialized subclasses of Tuple1/Tuple2. We'll do the same in
the future for FunctionN, but the current JFunctionN are written
in Java and specialized artisanally.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The InnerClass attribute needs to contain an entry for every nested
class that is defined or referenced in a class. Details are in a
doc comment in BTypes.scala.
Instead of collecting ClassBTypes of nested classes into a hash map
during code generation, traverse the class before writing it out to
disk. The previous approach was incorrect as soon as the generated
bytecode was modified by the optimzier (DCE, inlining).
Fixes https://github.com/scala/scala-dev/issues/21.
|
|
|
|
|
| |
Until now, there was no good place to hold various utility functions
that are used acrosss the backend / optimizer.
|
|\
| |
| | |
Inliner heuristic for higher-order methods
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Introduce a cache to remember which methods have maxLocals and
maxStack already computed. Before we were computing these values
on every run of eliminateUnreachableCode.
Also update the implementation of eliminateUnreachableCode to keep
correct max values.
|
| |
| |
| |
| |
| | |
Make InlinerHeuristics a backend component like the others, instead
of nested within the Inliner component.
|
| |
| |
| |
| |
| | |
If a class (trait) is a SAM type, store the name and descriptor of the
SAM in the ClassBType's InlineInfo.
|
| | |
|
|/
|
|
|
|
|
|
|
|
| |
A DefDef owned by a trait may now have have a method body.
Such a method must be emitted without ACC_ABSTRACT, and
with the code attribute.
Tested by intercepting the compile pipeline and adding the
DEFAULTMETHOD flag and a method body before running
the backend.
|
|
|
|
|
| |
Similar to the new JAVA_ANNOTATION flag, be more explicit about flags
for java entities.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The Scala classfile and java source parsers make Java annotation
classes (which are actually interfaces at the classfile level) look
like Scala annotation classes:
- the INTERFACE / ABSTRACT flags are not added
- scala.annotation.Annotation is added as superclass
- scala.annotation.ClassfileAnnotation is added as interface
This makes type-checking @Annot uniform, whether it is defined in Java
or Scala.
This is a hack that leads to various bugs (SI-9393, SI-9400). Instead
the type-checking should be special-cased for Java annotations.
This commit fixes SI-9393 and a part of SI-9400, but it's still easy
to generate invalid classfiles. Restores the assertions that were
disabled in #4621. I'd like to leave these assertions in: they
are valuable and helped uncovering the issue being fixed here.
A new flag JAVA_ANNOTATION is introduced for Java annotation
ClassSymbols, similar to the existing ENUM flag. When building
ClassBTypes for Java annotations, the flags, superclass and interfaces
are recovered to represent the situation in the classfile.
Cleans up and documents the flags space in the area of "late" and
"anti" flags.
The test for SI-9393 is extended to test both the classfile and the
java source parser.
|
| |
|
|\
| |
| | |
SI-9392 Avoid crash in GenBCode for incoherent trees
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
A macro in shapeless was generating a tree of the form:
```
{
class C#2
new C#2
}.setType(C#1)
```
This happened due to an error in the macro; it used untypecheck
to try to fix the owner-chain consistency problem, but kept a
reference to the previous version of the block-local class symbol
`C` and used this in the resulting tree.
This commit detects the particular situation we encountered, and
avoids the crash by not creating the `NestedInfo` for the
`BType` corresponding to `C#1`. The code comment discusses why I
think this is safe, and suggests a refactoring that would mean
we only ever try to construct `NestedInfo` when we are going to
need them.
|
|/ |
|
|\
| |
| | |
Closure elimination for new backend
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
When an indylambda closure is allocated and invoked within the same
method, rewrite the invocation to the implementation method.
This works for any indylambda / SAM type, not only Scala functions.
However, the Scala compiler (under -Xexperimental) currently desugars
function literals for non-FunctionN types to an anonymous class during
typer.
No testing yet, waiting for FunctionN to become SAMs first.
The feature requires scala-java8-compat to be on the classpath and a
number of compiler flags:
-Ydelambdafy:method -Ybackend:GenBCode -Yopt:closure-elimination -target:jvm-1.8
➜ scala git:(opt/closureInlining) ant -Dscala-java8-compat.package=1 -Dlocker.skip=1
➜ scala git:(opt/closureInlining) cd sandbox
➜ sandbox git:(opt/closureInlining) cat Fun.java
public interface Fun<T> {
T apply(T x);
}
➜ sandbox git:(opt/closureInlining) javac Fun.java
➜ sandbox git:(opt/closureInlining) cat Test.scala
class C {
val z = "too"
def f = {
val kap = "me! me!"
val f: Tuple2[String, String] => String = (o => z + kap + o.toString)
f(("a", "b"))
}
def g = {
val f: Int => String = x => x.toString
f(10)
}
def h = {
val f: Fun[Int] = x => x + 100 // Java SAM, requires -Xexperimental, will create an anonymous class in typer
f(10)
}
def i = {
val l = 10l
val f: (Long, String) => String = (x, s) => s + l + z + x
f(20l, "n")
}
def j = {
val f: Int => Int = x => x + 101 // specialized
f(33)
}
}
➜ sandbox git:(opt/closureInlining) ../build/quick/bin/scalac -target:jvm-1.8 -Yopt:closure-elimination -Ydelambdafy:method -Ybackend:GenBCode -Xexperimental -cp ../build/quick/scala-java8-compat:. Test.scala
➜ sandbox git:(opt/closureInlining) asm -a C.class
➜ sandbox git:(opt/closureInlining) cat C.asm
[...]
public g()Ljava/lang/String;
L0
INVOKEDYNAMIC apply()Lscala/compat/java8/JFunction1; [
// handle kind 0x6 : INVOKESTATIC
java/lang/invoke/LambdaMetafactory.altMetafactory(Ljava/lang/invoke/MethodHandles$Lookup;Ljava/lang/String;Ljava/lang/invoke/MethodType;[Ljava/lang/Object;)Ljava/lang/invoke/CallSite;
// arguments:
(Ljava/lang/Object;)Ljava/lang/Object;,
// handle kind 0x6 : INVOKESTATIC
C.C$$$anonfun$2$adapted(Ljava/lang/Object;)Ljava/lang/String;,
(Ljava/lang/Object;)Ljava/lang/String;,
3,
1,
Lscala/Serializable;.class,
0
]
CHECKCAST scala/Function1
L1
ASTORE 1
L2
ALOAD 1
BIPUSH 10
INVOKESTATIC scala/runtime/BoxesRunTime.boxToInteger (I)Ljava/lang/Integer;
ASTORE 2
POP
ALOAD 2
INVOKESTATIC C.C$$$anonfun$2$adapted (Ljava/lang/Object;)Ljava/lang/String;
CHECKCAST java/lang/String
L3
ARETURN
[...]
|
|\ \
| |/
|/| |
SI-9359 Fix InnerClass entry flags for nested Java enums
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The access flags in InnerClass entries for nested Java enums were
basically completely off.
A first step is to use the recently introduced backend method
`javaClassfileFlags`, which is now moved to BCodeAsmCommon.
See its doc for an explanation.
Then the flags of the enum class symbol were off. An enum is
- final if none of its values has a class body
- abstract if it has an abstract method
(https://docs.oracle.com/javase/specs/jls/se7/html/jls-8.html#jls-8.9)
When using the ClassfileParser:
- ENUM was never added. I guess that's just an oversight.
- ABSTRACT (together with SEALED) was always added. This is to
enable exhaustiveness checking, see 3f7b8b5. This is a hack and we
have to go through the class members in the backend to find out if
the enum actually has the `ACC_ABSTRACT` flag or not.
When using the JavaParser:
- FINAL was never added.
- ABSTRACT was never added.
This commit fixes all of the above and tests cases (Java enum read
from the classfile and from source).
|
|/ |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There were two issues in the new inliner that would cause a
VerifyError and an IllegalAccessError.
First, an access to a public member of package protected class C can
only be inlined if the destination class can access C. This is tested
by t7582b.
Second, an access to a protected member requires the receiver object
to be a subtype of the class where the instruction is located. So
when inlining such an access, we need to know the type of the receiver
object - which we don't have. Therefore we don't inline in this case
for now. This can be fixed once we have a type propagation analyis.
https://github.com/scala-opt/scala/issues/13.
This case is tested by t2106.
Force kmpSliceSearch test to delambdafy:inline
See discussion on https://github.com/scala/scala/pull/4505. The issue
will go away when moving to indy-lambda.
|
|
|
|
|
|
|
|
|
|
|
|
| |
Many backend components don't have access to the compiler instance
holding the settings.
Before this change, individual settings required in these parts of the
backend were made available as members of trait BTypes, implemented
in the subclass BTypesFromSymbols (which has access to global).
This change exposes a single value of type ScalaSettings in the super
trait BTypes, which gives access to all compiler settings.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Running an ASM analyzer returns null frames for unreachable
instructions in the analyzed method. The inliner (and other components
of the optimizer) require unreachable code to be eliminated to avoid
null frames.
Before this change, unreachable code was eliminated before building
the call graph, but not again before inlining: the inliner assumed
that methods in the call graph have no unreachable code.
This invariant can break when inlining a method. Example:
def f = throw e
def g = f; println()
When building the call graph, both f and g contain no unreachable
code. After inlining f, the println() call becomes unreachable. This
breaks the inliner's assumption if it tries to inline a call to g.
This change intruduces a cache to remember methods that have no
unreachable code. This allows invoking DCE every time no dead code is
required, and bail out fast. This also simplifies following the
control flow in the optimizer (call DCE whenever no dead code is
required).
|
|
|
|
|
| |
Introduces a stress-test mode "everything" in which the inliner tries
to inline every calliste that can be statically resolved.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Issue precise warnings when the inliner fails to inline or analyze a
callsite. Inline failures may have various causes, for example because
some class cannot be found on the classpath when building the call
graph. So we need to store problems that happen early in the optimizer
(when building the necessary data structures, call graph, ClassBTypes)
to be able to report them later in case the inliner accesses the
related data.
We use Either to store these warning messages. The commit introduces
an implicit class `RightBiasedEither` to make Either easier to use for
error propagation. This would be subsumed by a biased either in the
standard library (or could use a Validation).
The `info` of each ClassBType is now an Either. There are two cases
where the info is not available:
- The type info should be parsed from a classfile, but the class
cannot be found on the classpath
- SI-9111, the type of a Java source originating class symbol cannot
be completed
This means that the operations on ClassBType that query the info now
return an Either, too.
Each Callsite in the call graph now stores the source position of the
call instruction. Since the call graph is built after code generation,
we build a map from invocation nodes to positions during code gen and
query it when building the call graph.
The new inliner can report a large number of precise warnings when a
callsite cannot be inlined, or if the inlining metadata cannot be
computed precisely, for example due to a missing classfile.
The new -Yopt-warnings multi-choice option allows configuring inliner
warnings.
By default (no option provided), a one-line summary is issued in case
there were callsites annotated @inline that could not be inlined.
|
|
|
|
|
| |
I observed cases (eg Scaladoc tests) where we end up with 17k+
ClassNodes, which makes 500 MB.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In order to inline a final trait method, callsites of such methods are
first re-written from interface calls to static calls of the trait's
implementation class. Then inlining proceeds as ususal.
One problem that came up during development was that mixin methods are
added to class symbols only for classes being compiled, but not for
others. In order to inline a mixin method, we need the InlineInfo,
which so far was built using the class (and method) symbols. So we had
a problem with separate compilation.
Looking up the symbol from a given classfile name was already known to
be brittle (it's also one of the weak points of the current inliner),
so we changed the strategy. Now the InlineInfo for every class is
encoded in a new classfile attribute.
This classfile attribute is relatively small, because all strings it
references (class internal names, method names, method descriptors)
would exist anyway in the constant pool, so it just adds a few
references.
When building the InlineInfo for a class symbol, we only look at the
symbol properties for symbols being compiled in the current run. For
unpickled symbols, we build the InlineInfo by reading the classfile
attribute.
This change also adds delambdafy:method classes to currentRun.symSource.
Otherwise, currentRun.compiles(lambdaClass) is false.
|
|
|
|
|
|
|
|
|
| |
The inliner forces some method symbols to complete that would not be
completed otherwise. This triggers SI-9111, in which the completer of
a valid Java method definition reports an error in mixed compilation.
The workaround disables error reporting while completing lazy method
and class symbols in the backend.
|
|
|
|
|
|
|
| |
The `ByteCodeRepository.classNode(InternalName)` method now returns an
option. Concretely, in mixed compilation, the compiler does not create
a ClassNode for Java classes, and they may not exist on the classpath
either.
|
|
|
|
|
|
|
|
|
|
| |
Inlining decisions will be taken by analyzing the ASM bytecode. This
commit adds tools to build a call graph representation that can be
used for these decisions.
The call graph is currently built by considering method descriptors of
callsite instructions. It will become more precise by using data flow
analyses.
|