Skip to main content

Notice: this Wiki will be going read only early in 2024 and edits will no longer be possible. Please see: https://gitlab.eclipse.org/eclipsefdn/helpdesk/-/wikis/Wiki-shutdown-plan for the plan.

Jump to: navigation, search

Difference between revisions of "JDT Core Programmer Guide/ECJ/Lookups"

(ModuleBinding: package lookup through a module)
(additions here and there)
 
Line 1: Line 1:
 
{{note|During name resolving, several structures are used to identify the binding referenced by a given name.}}
 
{{note|During name resolving, several structures are used to identify the binding referenced by a given name.}}
 +
 +
==Global Conventions==
 +
 +
Throughout the following lookup structures a convention exists that methods <code>getPackage0()</code>, <code>getType0</code> and a few more like this will only check, if a sought element already exists in one of the caches (call them "'''passive'''"). Typically, a sibling method exists without the <code>0</code> suffix, which in case of a cache miss will actively search for the sought element ("'''active'''").
 +
 +
To record the fact that an element was not found in a particular location of the tree, special instances <code>TheNotFoundModule</code>, <code>TheNotFoundPackage</code>, <code>TheNotFoundType</code> are inserted into the tree to avoid repeated, unsuccesful resolution attempts.
  
 
==LookupEnvironment==
 
==LookupEnvironment==
Line 7: Line 13:
 
** TypeBinding (<code>knownTypes</code>)
 
** TypeBinding (<code>knownTypes</code>)
  
Since Java 9, however, different instances of LookupEnvironment are used, to implement the perspective of one module each. In this situation
+
Note, that the structure below <code>knownPackages</code> forms a '''tree''', where each edge corresponds to a '''simple name segment'''. Based on this most lookup is done '''incrementally''', i.e., we resolve one simple name at a time, relative to the result so far. The only exception to this rule is field <code>ModuleBinding.declaredPackages</code> (see below).
 +
 
 +
Since Java 9, different instances of LookupEnvironment are used, to implement the perspective of one module each. In this situation
 
* the root lookup environment represents the ''unnamed module''
 
* the root lookup environment represents the ''unnamed module''
 
* module-specific lookup environments are found via the <code>environment</code> field of elements of <code>knownModules</code>
 
* module-specific lookup environments are found via the <code>environment</code> field of elements of <code>knownModules</code>
Line 48: Line 56:
 
==PackageBinding==
 
==PackageBinding==
  
''TBC''
+
As mentioned, package bindings are organized as a tree below <code>LookupEnvironment.knownPackages</code>. Packages bindings can be used to look up child packages (<code>getPackage()</code>) and types (<code>getType</code>). For qualified names of unknown kind method <code>getTypeOrPackage()</code> does all the hard work.
 +
 
 +
Since Java 9 all such lookup should pass the client module (where the reference happens), so that lookup can obey to the rules of visibility and accessibility of JPMS.
 +
 
 +
Java has a bit of a split mind as to whether or not "'''empty packages'''" "exist". We have this tree of packages where leaf packages are "children" of some "parent" package. Many of these "parent" packages are empty. OTOH, some parts of JLS only consider packages that "exist" in terms of being declared by a compilation unit. In that sense those empty "parent" packages do not "exist". Note, that strictly speaking, a package containing a non-Java resource, does not exist either. PackageBinding sports a method <code>hasCompilationUnit</code> to check for the stricter forms of existence. When passing <code>true</code> for the parameter <code>checkCUs</code> the batch compiler will actually parse any found .java file, to check if it indeed declares the expected package (which is of course a bit expensive).
 +
 
 +
===SplitPackageBinding===
 +
In a way this particular strange guy acts as a lookup too: It represents a package with incarnations in different modules and allows:
 +
* navigating to a child package (given a module perspective), which may or may not again be a SplitPackageBinding
 +
* extracting a specific package incarnation corresponding to to the slice of this package within a particular module
 +
 
 +
==ReferenceBinding==
 +
At the level below packages, <code>ReferenceBinding</code> and subclasses represent all classes and interfaces.
 +
 
 +
Lookup of '''fields''' and '''memberTypes''' works just as expected, because still here names are unique.
 +
 
 +
Only when diving into '''methods''', a lot more work is needed to consider type inference and overload resolution, so that's a separate story to be told another time.
  
 +
A special caveat on methods <code>unResolvedFields</code> and <code>unResolvedMethods</code>: initially these have been implemented with the intention to access known members ''without triggering type resolution'' (which can cause detrimental re-entrance in some situations). Unfortunately, the picture is blurred here: in particular the implementation in <code>ReferenceBinding</code> simply delegates to the other <code>getMethods()</code> which ''may'' trigger type resolution!
  
 
==Scopes==
 
==Scopes==

Latest revision as of 17:45, 1 March 2021

Note.png
During name resolving, several structures are used to identify the binding referenced by a given name.


Global Conventions

Throughout the following lookup structures a convention exists that methods getPackage0(), getType0 and a few more like this will only check, if a sought element already exists in one of the caches (call them "passive"). Typically, a sibling method exists without the 0 suffix, which in case of a cache miss will actively search for the sought element ("active").

To record the fact that an element was not found in a particular location of the tree, special instances TheNotFoundModule, TheNotFoundPackage, TheNotFoundType are inserted into the tree to avoid repeated, unsuccesful resolution attempts.

LookupEnvironment

The central place where the following bindings are globally stored:

  • ModuleBinding (knownModules)
  • PackageBinding (knownPackages, recursively)
    • TypeBinding (knownTypes)

Note, that the structure below knownPackages forms a tree, where each edge corresponds to a simple name segment. Based on this most lookup is done incrementally, i.e., we resolve one simple name at a time, relative to the result so far. The only exception to this rule is field ModuleBinding.declaredPackages (see below).

Since Java 9, different instances of LookupEnvironment are used, to implement the perspective of one module each. In this situation

  • the root lookup environment represents the unnamed module
  • module-specific lookup environments are found via the environment field of elements of knownModules
  • each lookup environment has a back link root to the root environment
  • fields of LookupEnvironment that are not module specific are documented as either ROOT_ONLY or SHARED (see javadoc of LookupEnvironment#root).

Other "singletons"

The root lookup environment also links a few unique instances:

  • globalOptions : all compiler options in effect
  • nameEnvironment : sometimes called the oracle, since it's an external/opaque entity that can answer sought types and packages.
  • typeRequestor : callback for newly discovered types
  • typeSystem : manages variants of known type bindings (parameterizations, arrays, annotations)
  • verifier :
  • problemReporter : a fall-back problem reporter, which is used for reporting errors that cannot be associated to a particular AST node.
  • classFilePool :

TBC

ModuleBinding

Since Java 9 also module bindings act as lookups.

The current design has been developed in bug 547181:

Declared Packages

Each module binding stores a flat map of declaredPackages.

  • This map is indexed by qualified names
  • Values in this map are PlainPackageBinding

The map represents exactly those packages that are declared in this module. It does not take into account possible name clashes with packages declared in another module (SplitPackageBinding, see below).

NB: Much of the complexity of package lookup is necessary to admit and distinguish same-named packages in different modules. Interestingly, the JVM will not be able to handle such programs unless a custom Layer implementation is provided.

During compilation, the map declaredPackages is eagerly filled with all packages mentioned in exports and opens directives. This is done as to avoid re-entrant package lookup as we observed before bug 547181 due to crazy traversal of the module graph etc. For automatic modules, which have no such declarations, the structure is filled using a lazy, one-time scan using IModuleAwareNameEnvironment.listPackages().

For explicit modules additional (non-exported) packages may be added lazily, but those have no impact on resolving of references outside this module.

Visible and Accessible Packages

Additionally, two methods of name getVisiblePackage exist, which will search the current module and all other modules read by it. If more than one visible module declares the requested package, the result will be represented as a SplitPackageBinding. JLS distinguishes visibility and accessibility. The former is defined just by the module graph, whereas the latter, stronger form also takes exports declarations into account. To query accessibility, the method canAccess(PackageBinding) can be used.

PackageBinding

As mentioned, package bindings are organized as a tree below LookupEnvironment.knownPackages. Packages bindings can be used to look up child packages (getPackage()) and types (getType). For qualified names of unknown kind method getTypeOrPackage() does all the hard work.

Since Java 9 all such lookup should pass the client module (where the reference happens), so that lookup can obey to the rules of visibility and accessibility of JPMS.

Java has a bit of a split mind as to whether or not "empty packages" "exist". We have this tree of packages where leaf packages are "children" of some "parent" package. Many of these "parent" packages are empty. OTOH, some parts of JLS only consider packages that "exist" in terms of being declared by a compilation unit. In that sense those empty "parent" packages do not "exist". Note, that strictly speaking, a package containing a non-Java resource, does not exist either. PackageBinding sports a method hasCompilationUnit to check for the stricter forms of existence. When passing true for the parameter checkCUs the batch compiler will actually parse any found .java file, to check if it indeed declares the expected package (which is of course a bit expensive).

SplitPackageBinding

In a way this particular strange guy acts as a lookup too: It represents a package with incarnations in different modules and allows:

  • navigating to a child package (given a module perspective), which may or may not again be a SplitPackageBinding
  • extracting a specific package incarnation corresponding to to the slice of this package within a particular module

ReferenceBinding

At the level below packages, ReferenceBinding and subclasses represent all classes and interfaces.

Lookup of fields and memberTypes works just as expected, because still here names are unique.

Only when diving into methods, a lot more work is needed to consider type inference and overload resolution, so that's a separate story to be told another time.

A special caveat on methods unResolvedFields and unResolvedMethods: initially these have been implemented with the intention to access known members without triggering type resolution (which can cause detrimental re-entrance in some situations). Unfortunately, the picture is blurred here: in particular the implementation in ReferenceBinding simply delegates to the other getMethods() which may trigger type resolution!

Scopes

In addition to the global lookup (which serves source and binary elements), the following AST nodes have their dedicated Scope which performs the initial, location-aware part of name lookup:

  • CompilationUnitDeclaration -> CompilationUnitScope
  • ModuleDeclaration -> ModuleScope
  • TypeDeclaration -> ClassScope
  • AbstractMethodDeclaration -> MethodScope
  • Block -> BlockScope (also used by block-like nodes like ForeachStatement)

A TypeDeclaration may hold additional scopes initializerScope and staticInitializerScope for resolving field initializers (implementation uses MethodScope also here).

Each scope has a parent link to support inside-out searches during resolution. TBC

Back to the top