Methods for implementing virtual method invocation with shared code

[0001] This application claims priority under 35 U.S.C. § 120 as a continuation-in-part of U.S. application Ser. No. 10/803,585, filed Mar. 17, 2004, the disclosure of which is incorporated herein by reference, and furthermore, this application is related to (1) U.S. patent application Ser. No. 10/803,205, filed on Mar. 17, 2004, and entitled “Sharing Runtime Representation of Software Component Methods Across Component Loaders,” (2) U.S. patent application Ser. No. 10/849,468, filed on May 18, 2004, and entitled “Methods for Sharing of Dynamically Compiled Code Across Class Loaders by Making the Compiled Code Loader Reentrant,” and (3) U.S. patent application Ser. No. 10/855,917, filed on May 26, 2004, and entitled “Methods for Sharing Conditionally Across Class Loaders Dynamically Compiled Code,” whereby each of the above related applications is incorporated herein by reference.

[0002] 1. Field of the Invention

[0003] This invention relates generally to the Java™ platform, and more particularly to methods and systems for implementing virtual method invocation.

[0004] 2. Description of the Related Art

[0005] The sharing of runtime representation of classes can help reduce the footprint of applications written in the Java programming language, and can reduce the overhead of both bytecode interpretation and dynamic compilation. However, sharing of compiled code may make ineffective some optimizations commonly employed in object oriented languages. In particular, inline caching is an optimization that may be made ineffective when compiled code is shared. In general, inline caching names a mechanism that reduces the overhead of virtual method dispatch by specializing call sites that repeatedly dispatch to the same method of the same class type. In part, sharing degrades the efficiency of inline caching because sharing turns the inline cache that naturally does not issue a call to more than one different receiver class type into a table driven call site.

[0006] In view of the foregoing, there is a need for methods that can improve the efficiency of inline caching when code is shared across class loaders.

[0007] Broadly speaking, the present invention fills these needs by providing methods and systems for implementing virtual method invocation when a compiled code of an invoked method is shared between class types. It should be appreciated that the present invention can be implemented in numerous ways, including as a process, an apparatus, a system, computer readable media, or a device. Several inventive embodiments of the present invention are described below.

[0008] In accordance with a first aspect of the present invention, a method for implementing virtual method invocation when a compiled code of an invoked method is shared between class types that share a runtime representation is provided. In this method, an entry of an inline cache is set to a reference to an object. Subsequently, instructions for calling code are modified and the inline cache is set to be in one of an unresolved state, a monomorphic state, a shared monomorphic state, or a megamorphic state.

[0009] In accordance with a second aspect of the present invention, a computer readable medium having program instructions for implementing virtual method invocation when a compiled code of an invoked method is shared between class types that share a runtime representation is provided. The computer readable medium includes program instructions for setting an entry of an inline cache to a reference to an object and program instructions for modifying instructions for calling code. Furthermore, program instructions for setting the inline cache to be in one of an unresolved state, a monomorphic state, a shared monomorphic state, or a megamorphic state are also included.

[0010] In accordance with a third aspect of the present invention, a system for implementing virtual method invocation when a compiled code of an invoked method is shared between class types that share a runtime representation is provided. The system includes means for setting an entry of an inline cache to a reference to an object and means for modifying instructions for calling code. Furthermore, the system includes means for setting the inline cache to be in one of an unresolved state, a monomorphic state, a shared monomorphic state, or a megamorphic state.

[0011] Other aspects and advantages of the invention will become apparent from the following detailed description, taken in conjunction with the accompanying drawings, illustrating by way of example the principles of the invention.

[0044] Inventions for reducing main memory usage and runtime costs by sharing a portion of the runtime representation of a software component across a plurality of component loaders are provided. Several exemplary embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0045] The embodiments of the present invention provide methods and systems for sharing a portion of runtime representation of a software component across a plurality of component loaders. In one embodiment, software components are classes and component loaders are class loaders. According to one implementation, the runtime representation of a class is divided into two parts, a class loader independent part, and a class loader dependent part. The class loader independent part can include most of the runtime representation of a class and is sharable across multiple class loaders that define the class and satisfy a set of sharing conditions. The class loader dependent part is replicated for each of the class loaders that define the class.

[0046] According to one embodiment, bytecode interpretation is made loader re-entrant thus allowing the bytecode of methods to be shared across multiple defining class loaders. As used herein, class loader re-entrant bytecode interpretation means that data structures necessary for interpreting bytecodes (including the bytecodes) do not assume states or directly reference data that is specific to a class loader (e.g., the initialization state of a class, a pointer to a resolved symbolic link, etc.). Instead, the bytecode interpreter is capable of efficiently retrieving loader dependent data and states, and of triggering loader dependent events (e.g., class initialization, symbolic link resolution, etc.).

[0047] In one example, re-entrance with respect to class initialization is achieved by adding class initialization barriers to the interpretation of some bytecodes. In another example, re-entrance with respect to symbolic links is achieved by adding link resolution barriers to the interpretation of some bytecodes. Both class initialization and link resolution barriers are achieved by a combination of data structures and sequence of instructions that can exploit the data structures.

[0048] According to another embodiment of the present invention, a dynamic compiler produces loader-reentrant code from bytecodes of methods by generating native code implementing a class initialization barrier for bytecodes that require a class to be initialized, by generating native code implementing a link resolution barrier for bytecodes that require a symbolic link to be resolved, and by generating code to retrieve loader-dependent data from a loader-dependent table when loader-dependent data is needed. In one example, link resolution barrier for a symbolic link to a class is not needed if a class initialization barrier is already required.

[0049] According to still another embodiment of the present invention, bytecodes of a first class type are compiled into a compiled code if the bytecodes have not been compiled, and a table of loader dependencies is generated for the compiled code. The table of loader dependencies is used to compare the loader dependencies of the first class type in the compiled code with the loader dependencies of the second class type. In one embodiment, if the loader dependencies are the same in the two class types, the compiled code can be used for the method of the second class type. In another embodiment, if the loader dependencies are not the same in the two class types, the compiled code is cloned, and portions of the compiled code are modified using the table of dependencies to locate sequence of instructions that need to be modified to enable use of the clone by the second class type.

[0050] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be understood, however, to one skilled in the art, that the present invention may be practiced without some or all of these specific details. In other instances, well known process operations have not been described in detail in order not to unnecessarily obscure the present invention.

I. Environment Description

[0051] As embodiments of the present invention can implement the Java™ platform, an overview of the Java™ programming language (hereinafter “Java”) is provided below. Programs written in Java (hereinafter “Java program”) are composed of one of more classes that are compiled by a Java compiler into class files wherein a respective class file is generated for each corresponding class of a program. Class files have architecturally-neutral binary formats and can contain bytecodes, a symbol table, and ancillary information. The symbol table can comprise symbolic references to fields, methods, and names of the classes used by the class being represented by the class file. Class files are machine-independent and can be executed by a Java™ Virtual Machine (hereinafter “JVM”).

[0052] The JVM is an abstract machine that can provide a level of abstraction between the machine independence of the bytecode in class files and the machine-dependent instruction set of a particular computer hardware. Each JVM is developed to operate in conjunction with one particular operating system on a particular hardware. Class files compiled from Java classes can be executed on any computer system equipped with a JVM. In this manner, Java programs can be ported from one computer system to another without requiring updating of the program code or recompilation of the program classes into class files.

[0053] In operation, a user of a typical Java-based system interacts with an application that embeds a JVM, for instance, a web-browser or a web-server. The application creates an instance of the JVM and supplies the instance of the JVM with an initial class that contains the main method of the application, for instance, a Java web server. The JVM executes the bytecodes of the main method, using a combination of the bytecode interpretation and executed native machine code produced by the compilation of the bytecodes with a dynamic compiler. Classes required for executing the program are loaded, as needed, using class loaders.

II. Sharing Runtime Representation of a Software Component Across a Plurality of Software Component Loaders

[0054] As one embodiment of the present invention is a JVM capable of sharing the runtime representation of classes between a plurality of class loaders, a brief description of classes and class loaders is provided herein. Class loaders are a mechanism for dynamically loading software components in an executing program, wherein the unit of software distribution is a class. Classes are distributed using a standard architecturally-neutral binary representation known as the class file format. Each class is represented by a single class file produced by a Java compiler. Class files can be stored in actual files, in a memory buffer, or be obtained from a network stream. Class files can be loaded by any implementation of the JVM to define a class type.

[0055] Within a program, a class type is uniquely identified by a class name and a namespace. Thus, two classes with the same name can be defined in different namespace and can be treated as different types. Even though the two classes have the same names, the two classes can be defined from different class files and can therefore be completely unrelated.

[0056] The JVM implements namespaces as class loaders. Both class types and class loaders have a representation in user code. Specifically, class types are represented as instances of the class java.lang.Class while class loaders are represented as instances of the class java.lang.ClassLoader.

[0057] Class loaders are used by the JVM to obtain the class files necessary to create the class types needed by a program. By way of example, when executing a method, the JVM may need to resolve a symbolic link to a class named “C.” If the JVM cannot find a runtime representation for the class named C in the loader defining the class which method is being executed, the JVM requests that the loader define the class named C. In one example, the loader defining the class which method is being executed is called the initiating loader of class C. The initiating loader returns the instance of the class java.lang.Class representing the defined class type in Java code to the JVM. The JVM obtains the runtime representation of the defined class from the instance of java.lang.Class, and uses same to complete the link resolution.

[0058] The initiating loader of class C can be the defining loader of class C. Alternatively, the initiating loader can forward the request to another class loader. In such scenario, the initiating loader is said to delegate the loading of class C to another loader called the delegatee. The delegatee may in turn forward the request to yet another class loader. Forwarding the request continues until one of the delegatees defines class C. The loader that defines class C is called the defining loader of class C.

[0059] The defining loader L of class C obtains a class file containing the architecture-neutral binary representation of class C from some location. At this point, the defining loader passes the class file to the JVM, which in turn, builds the runtime representation of the class type of class C defined by loader L. The JVM returns an instance of the class java.lang.Class representing the class type in Java code. The loader, in turn, returns the instance of the class java.lang.Class to the loader responsible for delegating the definition of class C to loader L if such a loader exists. However, if such a loader does not exist, the instance of the class java.lang.Class is sent to the JVM if JVM has requested that loader L define class C.

[0060] In one embodiment, the behavior of class loaders can be defined by application programmers through sub-classing the ClassLoader class and overriding the “loadClass” method of the ClassLoader class. The sub-classing mechanism enables application programmers to control the manner class loaders delegate the loading of classes to other loaders as well as the location of the class file that can be used to define a class. For example, a class loader may obtain a class file from a network connection to a remote site, may generate the class file from the onset, or may transparently enhance an existing class file by adding additional features to the existing class file (e.g., automatic persistence, resource management, profiling, etc.), etc. Such user-defined class loaders can also assign appropriate security attributes to classes that are loaded from a certain source. For more information on class loaders, reference can be made to a White Paper entitled “Dynamic Class Loading in the Java™ Virtual Machine,” by Sheng Liang and Gilad Bracha, dated October 1998.

[0061] In one example, programs are executed by at least the boot loader and the system loader, each of which is a system defined class loader. The boot loader can be used at JVM startup to load the so-called “core” classes. In one example, core classes (e.g., java.lang.Object, java.lang.Class, etc.) are closely associated with the JVM implementation. Loading of core classes can be fully controlled by the JVM. For example, the JVM can determine the location of class files to be obtained. The system class loader is the loader created by the JVM to load the class that defines the entry point to a program (i.e., the static main (String [ ]) method of the class). In one embodiment, software programs can create additional class loaders to serve the needs of the programs.

[0062] For simplicity and ease of reference, a class type is described using the notation <C, Ld>^Li, where C denotes the name of the class, Ld denotes the defining loader of the class, and Li denotes the initiating loader. When the initiating class loader is not relevant, the simplified notation <C, Ld> can be used instead of <C, Ld>^Li. When the defining class loader is not relevant, the simplified notation C^Li can be used instead of <C, Ld>^Li. By definition, <C, L1>=<C, L2> implies L1=L2. The notation “˜” is used to denote that two distinct class types can share respective runtime representations (i.e., the two distinct class types satisfy the three sharing conditions). For conciseness, the notation <C, L1>≅<C, L2> is used to denote that <C, L1>˜<C, L2>∥<C, L1>=<C, L2>, wherein “∥” is the Boolean operator “or.” In other words, <C, L1>≅<C, L2> means that <C, L1> and <C, L2> are either the same type, or share the runtime representation.

[0063] Keeping the overviews to Java and class loaders in mind, reference is made to a simplified high level block diagram shown in FIG. 1A depicting an exemplary JVM implementation of the present invention wherein the runtime representation of a class type is shared between different class loaders, in accordance with one embodiment of the present invention. FIG. 1A shows a runtime representation of a class type <A, L1> (i.e., class A defined by a loader L1) 104a, a class type <A, L2> (i.e., class A defined by a loader L2) 104a′, and a class type <A, L3> (i.e., class A defined by a loader L3) 104a″. As can be seen, the runtime representation of each class type 104a-104a″ includes a sharable portion that can be shared with other runtime representation of other class types, and a non-sharable portion. For example, the class type <A, L1> 104a includes a first sharable runtime representation (SRR) of class A 103a that the class type 104a can share with the class type <A, L2> 104a′. In a like manner, the class type <A, L3> 104a″ includes a second SRR of class A 103a′. As illustrated, the second SRR of class A 103a′ is not being shared with any of the class type runtime representations 104a and 104a′. Runtime representations of class A <A, L1> 104a, <A, L2> 104a′, and <A, L3> 104a″ also include non-sharable portions of class type runtime representations 104a-I, 104a′-I, and 104a″-I, respectively.

[0064] In one embodiment of the present invention, the JVM records almost all the SRRs 103a and 103b created by the JVM in a SRR repository 106. In one example, the SRR repository 106 can be implemented as a hash table that may be keyed on class names for fast retrieval. As can be seen, each entry 106a-106n of SRR repository 106 can contain a list of SRRs 103a and 103a′. Specifically, each entry of the SRR repository 106 can include a list of all the SRRs for classes of the same name.

[0065] In one embodiment, the non-sharable portions of class type runtime representations 104a-I, 104a′-I, and 104a″-I are loader dependent. That is, the non-sharable portions of the class type runtime representations 104a-I, 104a′-I, and 104a″-I include states, or refer directly to data structures that are specific to each loader L1, L2, and L3, respectively. The sharable portions of class type runtime representations 103a and 103a′ are loader independent. However, each of the sharable portions of class type runtime representations 103a and 103a′ is bound to a set of sharing conditions (e.g., sharing conditions S1 for the first SRR 103a, and S2 for the second SRR 103b, etc.). In one embodiment, a loader can use an existing SRR of the class type for the class the loader has defined if the class satisfies the sharing conditions bound to the SRR. For example, <A, L3> 104a″ does not satisfy the sharing conditions S1 and therefore, cannot use the SRR of class A 103a. However, <A, L3> 104a″ can satisfy the sharing conditions S2, and as such, can therefore use the SRR of class A 103a′. In one embodiment of the present invention, the loader can use an existing SRR of the class type of the class that has been defined by the class loader if the class satisfies three sharing conditions. Below are more details regarding the three sharing conditions and the manner the sharing conditions can be used.

[0066] According to one embodiment of the present invention, the sharing conditions can simplify sharing of the runtime representation of class types. For instance, the sharing conditions can ensure that two class types satisfying the same sharing conditions can declare the same static and instance variables. That is, the variables, irrespective of being defined directly by the class or inherited, have the same name, protection level, and primitive type (i.e., byte, character, Boolean, integer, float, long, double type or reference type) in both class types. Thus, the variables with the same name can be assigned the same offset within the variables holder regardless of the class type of the holder. An implementation of the JVM can exploit such property to lay out identically the instances of class types that satisfy the same sharing conditions. Such property can be exploited further to share the runtime data structures holding the description of variables among class types. In one example, the runtime data structures holding the description of variables can include offset within the variable holder, protection level, signature, and name of the variables.

[0067] The sharing conditions can further ensure that methods of two class types that satisfy the same sharing conditions have the same name, signature, protection level, and bytecodes, whether the methods are inherited or declared directly by classes. For instance, using such property, an implementation of the JVM can share the runtime representation of methods and lay out, identically, the virtual table of class types satisfying the same sharing conditions. In one embodiment, the runtime representation of methods can include bytecodes as well as other descriptive data (e.g., method name, signature, protection level, etc.). The runtime representation of methods can further arrange for methods to be assigned the same index in the virtual method table of respective class types.

[0068] Yet further, sharing conditions can ensure that two class types satisfying the same sharing conditions implement interfaces that have the same name and the same unimplemented methods. However, in one embodiment, the sharing conditions may not require that the interfaces declare the same methods. An example of such a scenario is shown in Table 1 wherein the sharing conditions allow the runtime representation of the class C to be shared across class loaders L1 and L2, even though the interface loaded by the loader L1 declares a different number of methods than that of the interface A loaded by the loader L2.

[0069]

TABLE 1
Exemplary Runtime Representation of a Class being Shared
across Loaders having Interfaces with Different Number of Methods
	// In L1
	interface A {
	int foo(int i);
	A foo(A);
	void bar(int i);
	}
	// In L2
	interface A {
	Integer foo(Integer i);
	void bar(int i);
	}
	// Defined by both L1 and L2
	public abstract class C implements A {
	int foo(int i){...}
	Integer foo(Integer i){...}
	A foo(A a){...}
	// bar is an unimplemented method
	}

[0070] In one embodiment of the present invention, the sharing conditions can also be used to ensure that interfaces are substantially identical across class loaders (i.e., that interfaces declare the same methods and that interfaces have identical names). In this manner, the interfaces and respective methods can occupy the same index in the table mapping methods declared by the interface to methods of the class implementing the interface. As a result, the extent of sharing between classes satisfying the sharing conditions can be increased.

[0071] Having discussed the manner and the extent the sharing conditions can simplify sharing of the runtime representation of classes across multiple loaders, reference is made to a flowchart diagram 100 shown in FIG. 1B depicting the method operations performed to determine whether the SRR of a class type can be used to construct the runtime representation of a class type defined by a class loader, in accordance with one embodiment of the present invention. The method begins in operation 108 in which the next SRR for the class having the class name (e.g., “C”) is sought. In one example, upon receiving a request from a loader L to define a class C (i.e., to create class type <C, L>), the JVM locates all SRRs of classes with the name C. In one embodiment of the present invention, the SRRs for classes with the name C can be quickly located by looking up the SRR repository. The method then proceeds to operation 110 in which a determination is made as to whether the next SRR for the class with the name C has been found. If the next SRR for the class with the name C has not been located, the method continues to operation 118 in which a SRR is created from the class file. Next, the method continues to operation 120 in which the runtime representation of the class type is created using the selected SRR of the class type. Creating the runtime representation of the class type includes creating a loader dependent part.

[0072] However, if the next SRR for the class with the name C has been found, the method continues to operation 112 in which a determination is made as to whether the first sharing condition has been satisfied. The first sharing condition is satisfied when the same class file is used by both class loaders. If the first sharing condition has not been satisfied, the method continues to operation 108. Otherwise, the method continues to operation 114 in which a determination is made as to whether the second sharing condition has been satisfied. The second sharing condition is met when the SRR of the super class type of the class type <C, L> is the same as the SRR of the super class type of the first class type. If the second sharing condition has not been satisfied, the method proceeds to operation 108. However, if the second sharing condition has been satisfied, the method continues to operation 116 in which a determination is made as to whether the third sharing condition has been satisfied. The third sharing condition is met when the class types have the same unimplemented methods. If the third sharing condition has not been satisfied, the method continues to operation 108. Alternatively, the method continues to operation 120 in which the runtime representation is created using selected SRR of the class type with the selected SRR referring to the next SRR for the class with the class name “C.”

[0073] By way of example, for each of the SRRs found, the JVM tests whether <C, L> satisfies the three sharing conditions. In one example, the order in which the sharing conditions are evaluated may be of importance. If <C, L> does not satisfy the sharing conditions with any of the SRRs, a new SRR is constructed from the class file used by L to define the class C. Otherwise, the JVM uses the first SRR with which <C, L> satisfies all the sharing conditions. Once the SRR with which <C, L> can satisfy the sharing conditions has been located, the JVM constructs a loader dependent runtime representation for <C, L> using the SRR. In one example, the JVM does not need to fully parse the class file provided by the loader L for the class C. Nor does the JVM need to construct an intermediate representation of <C, L> in order to verify whether <C, L> satisfies the sharing conditions with one of the SRRs of the class C. Additional details regarding the three sharing conditions and the manner the JVM checks whether a given class type <C, L> satisfies the sharing conditions with one of the SRRs of the class C are provided below.

[0074] In accordance with one embodiment of the present invention, the first sharing condition requires that class types whose runtime representation can share the same SRR have been defined from identical class files. The simplified block diagram shown in FIG. 2A illustrates the way the JVM can test whether the first sharing condition has been satisfied, in accordance with one embodiment of the present invention. As can be seen in FIG. 2A, a class file 120′ has been submitted to the JVM by a class loader L 102′ to define a class type <C, L> 104′. As shown, a class file 120c has been used to construct the SRR of class C 103c. The SRR 103c has also been used for the runtime representation 104c of class type <C, L1> and 104c′ of class type <C, L2>, each defined by a loader L1 102c and a loader L2 102d, respectively. To be able to use the SRR of class C 103c for the runtime representation of class type <C, L>, the class files 120c and 120′ should be identical.

[0075] In the embodiment shown in FIG. 2A, the two class files 120c and 120′ are considered identical if the two class files 120c and 120′ are equal byte-per-byte. In the illustrated implementation, bytes BC1-BC8 of the class file 120c are compared with bytes BC1′-BC8′ of the class file 120′, byte-per-byte.

[0076] In another embodiment, the JVM can use a secure Hash Algorithm-1 (SHA-1) to compute digital signatures (i.e., SHA-1 digest) of class files 120c and 120′. For instance, SRRs of classes can include the SHA-1 digest of respective class files. Thus, to search for SRRs to be used as the runtime representation of the class type <C, L>104′, the JVM computes a SHA-1 digest 106′ of the class file 120′ submitted by the loader L 102′ for the class C. In one embodiment, testing the first sharing condition can be implemented by only comparing the digest 106′ with a digest 106c of the SRR 103c. In this manner, the cost associated with byte-per-byte comparison may not be incurred. Comparing the digest 106′ with the digest 106c can almost guarantee that the probability of two different byte streams resulting in the same digest can be very low. In another embodiment, testing the first sharing condition can further include comparing class files 120c and 120′, byte-per-byte, when digests 106 and 106′ have the same value, thus protecting against the probability of having two different byte streams having the same digest value.

[0077] The next sharing condition, the second sharing condition requires that runtime representations of class types sharing the same SRR either have the same super class type or have super class types that share the same SRR. That is, the second sharing condition requires that super (<C1, L1>)≅ super (<C2, L2>), wherein super (<C, L>) denotes the super class type of the class type <C, L>. In one example, among other properties, the second sharing condition can almost guarantee that most super class types super ((C, L>) of all class types <C, L> that use the same SRR have the same SRR. By way of example, a JVM implementation can use such property by including in each SRR of a class Cσ_C the pointer to the SRR shared by all runtime representations of super classes of class C that use the ≅_C. Thus, the SRR pointed to by the SRR σ_C is hereinafter referred to as the super SRR of SRR σ_C.

[0078] In one example, when using software components, the second sharing condition compares the loader independent part of the runtime representation of component types that are super-type of the first and second software component types, respectively. The second condition is satisfied when loader independent parts of the runtime representation of first and second software component types that are the super-type of first and second component types are equal, respectively.

[0079] Continuing to FIG. 2B, an exemplary hierarchy of classes loaded by a software program is illustrated, according to one embodiment of the present invention. Class types <C, L1> 104c, <C, L2> 104c′, and <C, L3> 104c″ are shown to share the same SRR for the class C 103c. As illustrated, <C, L1> 104c and <C, L2> 104c′ have the same super class type <S, Lx> 204-x, whereas the super class type for <C, L3> 104c″ is <S, Ly> 204-y. Both <S. Lx> 204-x and <S, Ly> 204-y use the same SRR for S 203-x (i.e., <S, Lx>˜<S, Ly>), thus satisfying the second sharing condition. However, although the class type <C, L4> 104c′″ has the super class type <S, Ly> 204-y and therefore satisfies the second sharing condition with <C, L1> 104c, <C, L2> 104c′ and <C, L3> 104c″, the class type <C, L4> 104c′″ uses a different SRR for the class C (i.e., 103c′). In one example, while necessary, merely satisfying the second sharing condition may not be sufficient to use a SRR. For instance, although the class type <C, L4> has satisfied the second sharing condition, <C, L4> may violate the first sharing condition or the third sharing condition, or both first and third sharing conditions.

[0080] As can be seen in the embodiment illustrated in FIG. 2B, every SRR 103x-103c″ includes a pointer to the respective super SRR 203-x and 203-y. That is, every SRR 103x-103c″ includes a pointer to the unique SRR used by each super class type of class types that use the first SRR. For example, the SRR for class C 103c includes a pointer to the SRR for S 203-x. In this manner, the second sharing condition can almost guarantee that substantially all class types that use the SRR for the class C 106c can have a super class type which uses the SRR for S 203-x. Similarly, the SRR for C 103c′ includes a pointer to the SRR for S 203-x, and SRR for C 103c″ includes a pointer to the SRR for S 203-y.

[0081] Evaluating whether the class type <C, L> 104′ satisfies the second sharing condition with an existing SRR, can be achieved by comparing the pointer to the SRR of the super class type 207 of class type <C, L> 104b with the pointer to the super SRR of the SRR being considered for sharing. For example, assume that the JVM is testing whether class type <C, L> 104′ satisfies the second sharing condition with the SRR for C 103c″. The JVM finds the super class type of <C, L> 104′ by looking up in the database of the loaded classes having the class name S (i.e., the super class name specified in the class file for <C, L>) as well as the initiating loader L 102′. By way of example, the result can be <S, Lx> (i.e., S^L=<S, Lx>). In such a scenario, the JVM then tests whether the pointer to the SRR used by <S, Lx> 204-x (i.e., the SRR for S 203-x) is equivalent to the pointer to the super SRR of the SRR for class C 103c″, which in this example is SRR for S 203-y. As can be seen, in the illustrated embodiment, the pointers are different, and therefore, the second sharing condition has not been satisfied. Accordingly, the SRR for C 103c″ cannot be used for <C, L> 104′. In the illustrated embodiment, next, the JVM moves on to test the second sharing condition with the remaining SRRs for C, if one exists. In the illustrated embodiment, eventually, the JVM can find that the SRR for C 103c′ satisfies the second condition (the super SRR of SRR for C 103c′ is SRR for S 203-x, which is also the SRR used by the super class type of class type <C, L>.

[0082] Next, satisfying the third sharing condition requires that almost all class types using the same SRR to have the same number of abstract methods. Abstract methods that have been declared by a class, however, do not have any implementation. In one example, abstract methods can be overridden by sub-classes of a class. In the Java programming language, abstract methods can either be declared explicitly or implicitly. For instance, an abstract method can be declared explicitly by the programmer, who specifies the abstract modifier when declaring the abstract method. Alternatively, the abstract method can be declared implicitly, when methods are inherited from an interface that has been declared by a class, or one of the super classes of the class that have not been implemented by the class or any of the super classes of the class. In one example, classes with at least one abstract method are declared abstract using the abstract modifier. Class files of the abstract class include an attribute characterizing the class as an abstract class.

[0083] It must be appreciated that because the first and second sharing conditions have already been satisfied, in one example, a number of conditions associated with the abstract methods may have been satisfied by the time the third sharing condition is being tested. For instance, having satisfied the first sharing condition implies that <C, L> is defined from a class file that is equivalent to the class file used to construct ac. Therefore, the class file from which <C, L> is defined as well as the class file used to construct ac declare the same abstract methods explicitly (including the scenario wherein the class file from which <C, L> is defined as well as the class file used to construct σ_C do not declare any abstract methods, explicitly).

[0084] The second sharing condition assures that the SRR used by super (<C, L>) is the same as the super SRR of σ_C. In this manner, the three sharing conditions have been satisfied and consequently, the abstract methods are the same. The combination of the first and second sharing combinations means that in order for <C, L> and ac to have different abstract methods, <C, L> should declare at least one interface I, and <C, L> cannot implement at least one method that has been declared by the interface I. Alternatively, the combination of the first and second sharing conditions means that if neither <C, L> nor any of super classes of <C, L> declare a non-empty interface, then <C, L> and σ_C have the same abstract methods. As will be explained in more detail below, an implementation of the JVM can be used to accelerate testing of the third sharing condition, in accordance with one embodiment of the present invention.

[0085] Reference is made to a simplified block diagram shown in FIG. 3 depicting a scenario wherein four loaders L1, L2, L3 and LA 102a-102d define the same classes A and C, wherein the class C extends the class A, and the class C implements an interface I. In the illustrated embodiment, the class C is an abstract class, as the class C does not implement all methods defined in the interface I. In one example, each of the loaders L1-LA 102a-d can define a different version of I. In this manner, the JVM needs one SRR for the class A that can be shared across the four loaders L1-LA 102a-d since L1-L4 102a-d satisfy the same three sharing conditions (i.e., <A, L1>˜<A, L2>˜<A, L3>˜<A, L4>).

[0086] Furthermore, the four loaders L1-L4 102a-d satisfy the first two sharing conditions for the class C. Additionally, <C, L1> and <C, L2> can use the same SRR for class C as <C, L1> and <C, L2> satisfy the third sharing condition: Although <C, L1> and <C, L2> define different versions of the interface I, <C, L1> and <C, L2> each has a single unimplemented method, and each has the same name and signature (i.e., method void bar( )). <C, L1> and <C, L2> therefore satisfy the three sharing conditions and can use the same SRR. However, although <C, L3> has a single unimplemented method (void foo(float f)) similar to <C, L1> and <C, L2>, the method does not have the same name and type. In a like manner, <C, L4> has two unimplemented methods (i.e., bar and foo(float f)) and as such, does not satisfy the third sharing condition with any of the SRRs for the class C. Consequently, the JVM needs to create respective SRRs for each of the class types <C, L3> and <C, L4>.

[0087] An exemplary algorithm for determining whether the third sharing condition has been satisfied so as to use the SRR σ_C for the runtime representation of the class type <C, L> can be described as follows: In one embodiment, the JVM searches the runtime representation of each of the interfaces declared by <C, L>. For each interface I^L, the JVM determines which of the methods of I^L does not have an implementation in the SRR σ_C. In one example, such determination is made by searching for a method having the same name and signature in the SRR σ_C. In such a scenario, the third sharing condition is satisfied if the following two criteria have been fulfilled. First, for each unimplemented method of each interface of the class file <C, L> a corresponding method having the same name and signature should exist in the list of unimplemented methods of the SRR σ_C. Second, the number of unimplemented methods of SRR σ_C should be equivalent to the number of unimplemented methods of interfaces of <C, L>.

[0088] In one embodiment of the present invention, testing for the third sharing condition can be expedited by testing, first, whether the SRR σ_C declares any non-empty local interfaces. If the SRR σ_C does not declare any non-empty local interfaces, the third sharing condition is satisfied. At this point, the JVM can use the SRR to build the runtime representation of <C, L>. Otherwise, the JVM can use the above described algorithm.

[0089] In another embodiment of the present invention, determining whether the third sharing condition has been satisfied can be expedited by maintaining an array of pointers to each of the SRRs for the class C and for each of the interfaces I declared by the class C. In such a scenario, the JVM can avoid performing the basic algorithm for testing the third sharing condition if the SRR σ₁ used for the runtime representation of each interface IL declared by <C, L> is equal to the SRR for interface I defined in the array of pointers of the SRR σ_C.

[0090] Keeping in mind the descriptions of the three sharing conditions, reference is made to a high level block diagram shown in FIG. 4A depicting a portion of the runtime representation of classes A and B being shared in an exemplary implementation of the JVM of the present invention, in accordance with one embodiment of the present invention. As will be explained in more detail below, in one embodiment, the runtime representation of a class is configured to mirror an architecturally-neutral binary representation of the class with a format optimized for main-memory use by various sub-systems of the JVM. In one aspect, exemplary subsystems can be the bytecode interpreter, the dynamic compiler, the garbage collector, etc.

[0091] In the embodiment shown in FIG. 4A, the class loader L1 102a and the class loader L2 102b both have defined classes A and B, such that the class B extends class A. In the illustrated embodiment, extending of the class A by the class B results in the creation of class types <A, L1> 104a, <A, L2> 104a′, <B, L1> 104b, <B, L2> 104b′ such that <A, L1>˜<A, L2>, and <B, L1>˜<B, L2> (i.e., L1 and L2 satisfy the same sharing conditions, respectively, for classes A and B). As can be seen, the runtime representation of each class type includes a loader dependent portion implemented by a corresponding instanceKlass object 104a, 104a′, 104b, and 104b′ and a respective loader independent SRR implemented by a corresponding sharedRep object 103a and 103b. Since loaders L1 102a and L2 102b can share the runtime representation of classes A and B, the instanceKlass 104a for the class type <A, L1> and the instanceKlass 104a′ for the class type <A, L2> refer to the same sharedRep 103a implementing a SRR for the class A.

[0092] Similarly, the instanceKlass 104b for the class type <B, L1> and instanceKlass 104b′ for class type <B, L2> refer to the same sharedRep 103b implementing a SRR for the class B. As illustrated, in accordance to one embodiment of the present invention, the sharedRep 103a includes a reference to the sharedRep 103b to implement the super type relationship between SRR, for implementing the second sharing condition. As can be appreciated, a hierarchy exists between the sharedRep objects 103a and 103b that corresponds to the hierarchy defined between the classes A and B. In the illustrated embodiment, the runtime representations of classes A and B include corresponding references to the runtime representation of respective super classes, as illustrated by references from instanceKlass object 104b and instanceKlass object 104b′ to instanceKlass object 104a and instanceKlass object 104a′, respectively.

[0093] In one example, the SRR for a class can include a template for loader dependent data structures. In this manner, the template can be used at class load time to build the loader dependent runtime representation of class types that satisfy the sharing conditions to use the SRR including the template. In one aspect, the template can be used as part of the loader dependent runtime representation of one of the class types that use the SRR so as to minimize space overhead. For example, as illustrated in the embodiment of FIG. 4A, templates 122a and 122b) referenced by sharedReps 103a and 103b, correspondingly, are used as part of the runtime representation for the class type <A, L1> and <B, L1> implemented by instanceKlass objects 104a and 104b, respectively. In the illustrated embodiment, templates 122a and 122b have been used to built the loader dependent data structures 122a′ and 122b′, respectively, for corresponding class types <A, L2> and <B, L2>. If the class type <A, L1> is unloaded (e.g., following a garbage collection operation, etc.), the loader dependent data structures of another class type can be used by the SRR of class A 103a as a template, for instance, the loader dependent data structures 122′a of the class type <A, L2>.

[0094] According to one embodiment of the present invention, the SRR for a class can also include SRRs of methods of the class. By way of example, the SRR of a method, as implemented by a sharedMethod object 123 in one exemplary JVM, includes most of the runtime representation of the method. In one example, the sharedMethod object 123 includes the bytecodes of the method. In another example, the sharedRep object 103b includes the sharedMethod object 123 which encapsulates, among other information, the bytecode of the method foo. The sharedMethod object 123, for instance, can be accessed by the sharedRep object through loader dependent data structures 122b and 122b′. According to one embodiment, instances of a class can contain pointers in respective headers with each pointer being directed to the instanceKlass object representing the corresponding class. For example, instances of the class type <A, L1> can include a pointer to the instanceKlass object 104a, in respective headers.

[0095] Proceeding to FIG. 4B, a detailed block diagram showing certain components of the loader independent SRR 101b of the runtime representation of the class type <B, L1> and the loader dependent part 122b of the same runtime representation of the class type <B, L1> is provided, in accordance with one embodiment of the present invention. In one example, the loader dependent part 122b of the runtime representation of the class type <B, L1> can be used as a template for building the loader dependent part of the runtime representation of other classes capable of using the SRR 101b for runtime representation (e.g., the loader dependent part 122b′ of class <B, L2>).

[0096] As can be seen, the loader dependent part 122b of the runtime representation of class type <B, L1> includes a class pool data structure 132b, a constant pool cache data structure (not shown in this Figure), and an array 134b of pointers to loader dependent representations of methods. Each loader dependent representation of a method can be implemented with a method object (e.g., a method object 136b).

[0097] The loader independent part 101b for the class type <B, L1> includes most of the runtime representation of the class B, such as the sharedRep object 103b. In turn, the sharedRep object includes a reference map 128b, references to the array of fields 130b declared by the class B, the shared constant pool 126b, and the sharedMethod object 123 for each of the methods defined by the class B. In one example, the reference map 128b can be used to assist the garbage collector to find references in instances of the class B. The sharedRep object 103b can also include a reference (not shown in this drawing) to the super sharedRep object of the SRR of the super classes of all classes having the SRR 101b for the shared runtime representation. The sharedRep object 103b can also include references to loader dependent data structures, such as, the class pool 132b and the array of methods 134b.

[0098] References to the loader dependent data structures can help minimizing the space consumed by the SRR. By way of example, the loader dependent data structures enable access to the sharedMethods objects 123 via the arrays of loader dependent method objects 136b, thus evading the use of an array of sharedMethod objects. Similarly, the loader dependent data structures avoid replicating entries for classes defined in the shared constant pool objects by retrieving the loader dependent data structures from the loader dependent class pool 132b. Additionally, the loader dependent data structures 122b can be used as templates for building the loader dependent part of the runtime representation of the class capable of using the loader independent SRR.

[0099] With continued reference to FIG. 4B, as illustrated, each method 136b includes a reference to the shared method object 123, the class pool 132b, and a class-loader re-entrant compiled code 142 (not shown in FIG. 4B), if one has been produced by the dynamic compiler. The sharedMethod object 123, in turn, includes most of the definition of a method. For instance, the sharedMethod object 123 includes the name index, the signature index, and vtable index of the method, the method bytecodes, method maps of on stack references for garbage collection, lines table, exception table, etc. In another embodiment, the sharedMethod object 123 can further include data encoding the protection level of the method as well as information related to the native code of the method that may have been produced by the dynamic compiler.

[0100] In contrast, the loader dependent representation 136b of the same method object includes a pointer to the loader dependent class pool 132b of the instanceKlass owning the method 136b and a pointer to the sharedMethod object 123 for that method. In one example, a pointer to the compiled code which may have been produced by the dynamic compiler for the method may be stored in the method object 136b.

[0101] In one exemplary embodiment, the instanceKlass object 104b includes storage for the static variables of the class B, the virtual method table (vtable), and the interface table (itable). Further included in the loader dependent part 122b are references to some general information about the class <B, L1>. For instance, a reference to the instanceKlass object of class <A, L1> 104a can include references to instances of java.lang.Class, java.lang.ClassLoader, and a reference to the sharedRep object 103b from which all data structures forming the loader independent SRR for <B, L1> 101b can be reached.

[0102] As can be appreciated, in one embodiment, by using the sharing conditions it can be ensured ensure that the vtable index is the same across class loaders. The sharing conditions can further assure that methods are inherited and overridden in the same manner across class loaders. In one example, entries of the vtable can be configured to refer to the class loader dependent representation of methods.

[0103] In one exemplary embodiment, the shared constant pool 126b, the class pool 132b, and the constant pool cache of the runtime representation of a class can be constructed based on information extracted from the constant pool of the class file containing the architecturally-neutral binary representation of the class. In one instance, the constant pool can include all the symbolic links to the variables, methods, interfaces, constants, and classes referenced by the bytecodes of the methods of the class. The constant pool can further include a table, which contains one or more indices to other entries of the table, or an offset within the class file where a symbol is stored. Bytecode instructions capable of using symbolic links (e.g., bytecode instruction used to invoke a method or access a variable, etc.) can include an index to an entry of the constant pool describing the symbolic link required for the bytecode.

[0104] The block diagram shown in FIG. 5 depicts an exemplary system dictionary used by an exemplary implementation of the JVM to keep track of the runtime representation of classes loaded, in accordance with one embodiment of the present invention. As can be seen, the runtime representation of class types <A, L1>, <A, L2>, <B, L1>, and <B, L2> are recorded in a system dictionary 138. In one example, the system dictionary 138 maps keys that include a fully qualified class name and class loader reference to the respective instanceKlass objects 104a, 104b, 104a′, and 104b′. According to one example, multiple entries of the system dictionary 138 can refer to the same instanceKlass object as a result of delegation between class loaders.

[0105] Referring to the block diagram shown in FIG. 6, the relationships between a constant pool 144, bytecodes, and symbols used to encode a symbolic link in an exemplary illustration of the class file 220 can be understood, in accordance with one embodiment of the present invention. A Bytecode instruction 242 of the method bytecode stream 140 includes an index 213 to the entry 248 of the constant pool 144 which describes the symbolic link used by instruction 242. The entry 248 of the constant pool 144 in turn references two additional entries of the constant pool 144, entries 2 and 215. As can be seen, the entry 2 contains an offset to the symbol 252 that encodes the name of the class “A,” and the entry 215 that contains indices to two other entries 241 and 239 of the constant pool 144. As shown, the entry 241 contains an offset to the method name 255 “foo” while entry 241 includes an offset to the signature 256.

[0106] According to one embodiment, to execute a bytecode instruction, the symbolic links required for the execution of the bytecode are configured to be translated into corresponding addresses in the main-memory where the corresponding items (e.g., variable, method, etc.) reside. In one example, the translation from symbolic link to an address in the main-memory is herein referred to as “link resolution.” Resolving of a symbolic link may result in loading the referenced class if the reference class has not been loaded yet. Symbolic link resolution can be expensive as symbolic link resolution can include locating a class loaded by the current class loader and searching the current class loader runtime representation for the object named by the symbol.

[0107] In one example, to avoid the cost of symbolic link resolution every time a bytecode instruction uses the symbolic link, the value of the resolved link can be saved once the link has been resolved for the first time. In such a scenario, an implementation of the JVM can save the resolved link in a runtime representation of the constant pool, at the entry previously indexed by the bytecode instructions. However, saving the resolved link in a runtime representation of the constant pool can cause the constant pool 144 to depend on the class loader resolving the link. The present invention can avoid such problem and allow a part of the constant pool information to be shared across multiple loaders. In one embodiment, sharing a part of the constant pool information across multiple loaders can be achieved by dividing the constant pool 144 into a loader independent part and a loader dependent part. In one example, the loader independent part is a shared constant pool, and the loader dependent part comprises a class pool and a constant pool cache.

[0108] Reference is made to the block diagram shown in FIG. 7A depicting the components of an exemplary runtime representation of the class type <B, L1> involved in a link resolution performed by an interpreter 158, in accordance with one embodiment of the present invention. As depicted, the shared constant pool 126b contains symbolic information configured to be constant across all the class loaders sharing the symbolic information. In one instance, such exemplary information may include indices to entries of the shared constant pool 126b, indices to entries of the class pool 116b, pointers to symbol objects, constant numerical values, pointers to objects implementing literal strings, etc. In one example, the class pool 132b contains symbolic links to classes.

[0109] The symbolic links to classes can be in an unresolved form (e.g., an index to an entry of the shared constant pool 126b containing a pointer to a symbol object), or in a resolved form (e.g., as a pointer to instanceKlass object). The constant pool cache 164b is configured to contain almost all the symbolic links to methods and variables of classes, and methods of interfaces. Unresolved entries of the constant pool cache 164b can contain an index to an entry of the shared constant pool 126b holding the corresponding unresolved symbolic link. In one embodiment of the present invention, resolved entries of the constant pool cache 164b can contain information obtained from the resolved link. In this manner, bytecode interpretation (e.g., the address of a static variable, the vtable index of a virtual method, the address of a method object, or the offset of an instance variable within respective holder, etc.) can be performed faster. However, it should be appreciated by one of ordinary skill in the art that other arrangements can be implemented, so long as loader dependent and loader independent parts are disjoint, and the loader independent part only contains data that is constant across the loaders sharing the data.

[0110] In one example, bytecodes of the class file 220 can refer to a symbolic link via indices to the constant pool 144 of the class file 220. In such a scenario, the main-memory copy of the bytecodes encapsulated in shared method objects can refer to symbolic link via indices of either the constant pool cache 164b or the class pool 132b. In one example, the bytecode instructions including indices to the constant pool 144 of the class file 220 can be rewritten to include indices to corresponding entries of the constant pool cache 164b and class pool 132b. In one embodiment, such bytecode rewriting can take place, for example, at class link time, just after verification of the bytecodes by the verifier of the JVM. It must be noted, however, that bytecode rewriting can be performed at any appropriate time, so long as bytecode rewriting is performed before the bytecodes are executed.

[0111] Still referring to FIG. 7A, the interpreter 158 is shown to keep track of the method being executed as well as the constant pool cache 164b of the class that defines the method in machine registers 160 and 162, respectively. In the embodiment shown in FIG. 7A, the current method is method “foo” of class <B, L1> with the loader dependent part of the runtime representation of foo being the method object 136b, and the shared method object 123 being the loader independent part of the runtime representation. Consequently, machine register 160 holds the pointer to the loader dependent representation 136b of method foo, while machine register 162 holds the constant pool cache 164b. As shown, the interpreter 158 is currently executing bytecode instruction 242 of the bytecodes 203 of method foo encapsulated in sharedMethod object 123 of the loader independent part of the runtime representation of class <B, L1>. The bytecode instruction 242 includes an index to entry 276 of the current constant pool cache, in this case, constant pool cache 164b. The entry of the constant pool cache 164b referenced by bytecode instruction 242 is not resolved and contains an index to an entry 272 of the shared constant pool 126b. The entry 372 in turn contains an index to an entry 274 of the shared constant pool 126b, and an index to an entry of the class pool 132b that contains a pointer to symbol 270 encoding the name of the class D. Entry 274 of the shared constant pool 126b in turn contains two indices to entries of the shared constant pool 126b that contain pointers to symbols 266 and 268 encoding a method name and a method signature, respectively. One must appreciate that all information in the shared constant pool 126b and bytecodes 203 is loader independent. That is, indices stored in the data structures are valid across all class loaders that share the runtime representation of class B.

[0112] FIG. 7B depicts a block diagram illustrating components of the runtime representation of the class type <B, L1> subsequent to resolution of the symbolic link required by the bytecode instruction 242, in accordance with one embodiment of the present invention. As a result of link resolution, the entry of the class pool that included a pointer to the symbol 270 has been overwritten with the pointer to the instanceKlass object 104′d, as the link was resolved. Furthermore, the entry 276 of the constant pool cache 164b has been updated to contain a pointer to the method object 136b. Subsequent execution of bytecode instructions referring to the entry 276 of the constant pool cache 164b of the class type <B, L1> can avoid the cost of link resolution and use the pointer to method 136b′ directly which is available from the constant pool cache entry 276. One must appreciate that in this embodiment, the link resolution has updated loader dependent data structures (e.g., the class pool 132b and constant pool cache 164b).

[0113] According to one embodiment, an implementation of the JVM resolves a symbolic link by locating the class referenced by the symbolic link. If the class referenced by the symbolic link cannot be located, the JVM issues a request to the current class loader to load the missing class. The loader that eventually defines the class then passes the class file for the class to the JVM, which builds a runtime representation of the class. As was described in more detail with respect to FIG. 5, the system dictionary and the shared class repository can be used by the JVM to locate runtime representation of classes and SRR of classes, respectively.

[0114] FIG. 8A depicts a flowchart diagram 400 illustrating method operations performed to load a class using an exemplary implementation of the JVM of the present invention, in accordance with one embodiment of the present invention. The method begins in operation 402 in which an instruction is received from a class loader to construct a runtime representation of a class. Next, in operation 404, the class file of the class is obtained by the JVM from the specified input stream. In operation 406, the SHA-1 digest of the class file is computed.

[0115] Proceeding to operation 408, the linked list of all the shared representations of classes built using class files of equal value are retrieved from the shared class repository, using the SHA-1 digest computed in operation 406. In one embodiment, the SHA-1 digest is a hash key. In one example, the format of the class file may not need verification prior to computing the SHA-1 digest. For instance, if the specified class file does not conform to a valid class file format, the digest of the class file cannot be mapped to an entry in the shared class repository.

[0116] In operation 410, a determination is made as to whether a linked list of sharedRep objects has been found in the shared class repository using the SHA-1 digest of the specified class file. If a linked list of sharedRep objects has not been found, the method continues to “B”. Otherwise, the method proceeds to operation 412 wherein the method scans the list of sharedRep objects to find a sharedRep object together with which the class being defined can satisfy all the sharing conditions. In operation 414, a determination is made as to whether a sharedRep object satisfying the sharing conditions has been found. If such sharedRep object has not been found, the method continues to “B,” otherwise the method continues to “A.”

[0117] Continuing to FIG. 8B, a flowchart diagram illustrating method operations performed to build a new instanceKlass object (i.e., a loader dependent representation of a class type) using the sharedRep object obtained from the shared class repository is provided, in accordance with one embodiment of the present invention. The method begins in operation 416, in which a new instanceKlass object is built using the sharedRep obtained. In one example, the new instanceKlass object is allocated and filled using the information from the sharedRep object obtained. Next, in operation 418, a determination is made as whether a template of loader dependent runtime representation of a class is being used by another instanceKlass object.

[0118] If the template is not being used, the method continues to operation 428 wherein the template is assigned to the instanceKlass object and is initialized for use by the instanceKlass object being created. However, if the template has already been used by another instanceKlass object, the method continues to operation 420 wherein the method replicates the template for the class pool, the constant pool cache, and methods array. The method objects are also created and initialized for use by the new instanceKlass object. Next, in operation 422, a corresponding instance of java.lang.Class is created, followed by operation 424 in which the static variables of the instanceKlass object are created. In operation 426, the instanceKlass object that has been created is entered in the system dictionary.

[0119] FIG. 8C depicts a flowchart diagram illustrating method operations performed when the SHA-1 digest does not map to a sharedRep object, in accordance with one embodiment of the present invention. The method begins in operation 430 in which the format of the class file is verified, followed by operation 432 in which a new sharedRep object is allocated. In the same manner, sharedMethod objects and shared constant pool as well as and templates for class pool, constant pool cache, methods array, and methods objects are allocated filled with information obtained from parsing the class file. In one example, the entries of the class pool are set in an unresolved format. According to one implementation, setting up the remaining components can be postponed until the class is linked (e.g., preparing a constant pool cache, initializing the virtual method table, etc.). Next, in operation 434, the sharedRep object is entered in the shared class repository using the SHA-1 digest computed earlier as a hash key.

[0120] In one embodiment of the present invention, parsing the class file to create a new sharedRep object may be avoided if a sharedRep object created from a class file having an identical value already exists. Instead of parsing the class file, the new sharedRep object can be obtained by cloning an existing sharedRep object created from a class file having a value identical to the value specified to create the new sharedRep object. In such a scenario, portions of the new sharedRep object that depend on the super sharedRep object (e.g., offset to instance variables, vtable index of methods, etc.) and the unimplemented abstract methods may be changed to reflect the new sharing conditions associated with the new sharedRep object. Specifically, if the sharedRep object selected for cloning differs from the sharedRep object that should be created by the third sharing condition (i.e., objects have the same super sharedRep object), the table of unimplemented method should be changed. If the sharedRep object selected for cloning differs from the sharedRep object to be created by the second sharing condition (i.e., objects have different super sharedRep objects), the table of unimplemented method and data that depends on the super sharedRep object should be changed. In one exemplary cloning implementation, a need to clone the shared constant pool may not exist and the shared constant pool can be shared by the sharedRep object that is being cloned and the sharedRep object that has been cloned.

[0121] FIG. 9A is a flowchart diagram 500 illustrating method operations performed to create a new sharedRep object by cloning a sharedRep object created with a class file having an identical value, in accordance with one embodiment of the present invention. The method begins in operation 502 in which the class pool and constant pool cache templates of the original sharedRep object are replicated and initialized for use by the cloned sharedRep object. Next, in operation 504, the template for the method arrays and the method objects of the original sharedRep object are replicated. Proceeding to operation 506, the sharedMethod objects of the original sharedRep object are replicated and initialized for use by the cloned sharedRep object. In operation 508, a determination is made as to whether the cloned sharedRep object has the same super sharedRep object as the original sharedRep object. If the super sharedRep objects are the same, the method continues to operation 288 in which the table of unimplemented methods is changed. Otherwise, data that depends on the super sharedRep object may be changed in operation 510. The method then proceeds to operation 512 in which the table of unimplemented methods is modified.

[0122] FIG. 9B is a flowchart diagram 500′ depicting method operations performed on an exemplary cloned sharedRep object by an exemplary implementation of the JVM of the present invention, in accordance with another embodiment of the present invention. The method begins in operation 514 in which the reference to the super sharedRep object of the cloned sharedRep object is updated so as to point to the appropriate sharedRep object. In one embodiment of the present invention, the reference to the appropriate super sharedRep object of the cloned sharedRep object is obtained from the instanceKlass object of the super class of the class being defined and on behalf of which the cloned sharedRep object is created. Next, in operation 516, the total number of instance variables of the new super class is determined followed by operation 518 in which the total number of instance variables of the initial sharedRep object is determined.

[0123] Continuing to operation 520, a determination is made as to whether the total number of instance variables of the new super class is equivalent to the total number of instance variables of the super class of initial sharedRep object. If the total number of instances are not equal, the operation continues to operation 522 in which the field table of the initial sharedRep object is scanned followed by the operation 524 in which the offsets of the instance variables are updated. If the total number of instances are equal, the method continues to operation 526 wherein the vtable length, the table of overridden methods, and vtable extension table are updated. Next, in operation 528, the shared methods may be assigned new vtable indices.

[0124] In one embodiment, to share the runtime representation of methods of a class, particularly, the bytecodes of methods, bytecode interpretation is made loader re-entrant. Loader reentrant bytecode interpretation requires that the main memory representation of bytecodes not include loader dependent data (e.g., operands of bytecode instruction is designed to be meaningful for all loaders sharing the bytecodes, etc.), or reflect loader-dependent states (e.g., a bytecode is configured to reflect a particular class initialization state, etc.). Instead, any particular loader-dependent data or loader dependent state is retrieved and evaluated at bytecode interpretation time using loader-independent data. Furthermore, techniques utilized to eliminate class initialization tests (e.g., “quickening,” etc.) may not be used to support loader reentrance.

[0125] In one example, efficient access to loader dependent values of a resolved symbolic link can be performed by using the loader dependent constant pool cache and the class pool associated with the class loader dependent representation of the current class (i.e., the class defining the method being executed). For instance, upon entering an invoked method, a reference to the loader dependent constant pool cache is stored at a dedicated location of the stack frame being prepared to execute the invoked method. In one embodiment, the dedicated location can be a register assigned to cache the constant pool cache. From the constant pool cache, the interpreter can efficiently retrieve the associated loader dependent class pool. To trigger link resolutions and class initializations, in one embodiment, “barriers” which are defined as short sequences of instructions can be used.

[0126] According to one implementation, when utilizing an appropriate encoding of the content of the constant pool cache, implementing a barrier can incur minor overhead. For instance, when implemented on SPARC processors, barriers can add one instruction to the interpretation of bytecode instructions. In one aspect of the present invention, to use barriers, resolved information is initialized to a distinguishable value. It must be appreciated that the resolved information is normally obtained from a constant pool cache upon interpreting a bytecode instruction. However, in accordance with one exemplary embodiment, the resolved information can be initialized to a value that denotes an unresolved link, for when link resolution barriers (LRB) are used. In another embodiment, the resolved information can be initialized to a value that denotes an uninitialized class, for when class initialization barriers (CIB) are used.

[0127] In one embodiment of the present invention, a LRB can be added to parts of an exemplary implementation of the bytecode interpreter performing the interpretation of bytecode instructions that refer to symbolic links. In an exemplary implementation of the JVM of the present invention, such bytecode instructions include the quickened versions of getfield, putfield, invokevirtual, invokespecial, and invokeinterface. The LRB is configured to efficiently test whether the symbol being referred to by a bytecode instruction has been resolved by the current class loader.

[0128] FIG. 10A includes a flow chart diagram 600 of method operations performed to determine whether the symbolic link referenced by a bytecode instruction manipulating an instance variable is resolved by the current class loader, in accordance with one embodiment of the present invention. By way of example, instruction manipulating instance variables can be getfield or putfield, or the quickened version of getfield or putfield. The method begins in operation 602 in which an index to the current constant pool cache is computed from the operand of the bytecode instruction. Next, in operation 604, the value of an offset within the instance variable holder is loaded from the current constant pool cache at the index obtained. In one example, the index can be obtained from the bytecode instruction operand.

[0129] Proceeding to operation 606, a determination is made as to whether the offset value loaded from the current constant pool cache is equal to an unresolved link marker. In one example, the unresolved link marker is a distinguishable value used to denote the unresolved link. If the values are not equal, the method continues to operation 610 wherein instance variable manipulation is performed. However, if the offset loaded from the current constant pool is equivalent to an unresolved link marker, the method continues to operation 608 wherein a call to the runtime is issued so as to resolve the symbolic link for the current class loader. In one embodiment, the runtime can replace the unresolved link maker at the entry of the constant pool cache with the resolved offset value. In this manner, subsequent uses of the constant pool cache entry that may result in a call to the runtime can be avoided. Furthermore, the resolved offset value can be returned. Thereafter, the method proceeds to operation 610 in which instance variable operation is performed using the offset value obtained. In one example, by using a value of −1 as the value of the unresolved marker, the LRB for bytecode instructions can be implemented to manipulate instance variables on the SPARC processor as a single branch on negative register value. As a result, one instruction is added to the interpretation of the bytecode, as shown in the exemplary code in Table 2.

[0130]

TABLE 2
Exemplary code including link resolution barrier
ld [Rcache + (header size + 2*wordSize)],	// Retrieve offset to field
Roffset
brgz,pt Roffset, resolved	// LRB
ld [ Robject + Roffset], Rvalue	// load field

[0131] Reference is made to FIG. 10B including a flow chart diagram 600′ of method operations performed to obtain the resolved information encoded by the symbolic link in an operand of an invokevirtual method, in accordance with one embodiment of the present invention. An invokevirtual bytecode instruction implements a virtual method dispatch mechanism, wherein the method to be invoked is obtained via a table driven lookup mechanism. In one example, the resolved information expected by the bytecode is an index to a virtual table. The pointer to the method object encapsulating the bytecode of the method to be invoked is stored in the virtual table.

[0132] The method begins with operation 612 in which an index to the current constant pool cache is computed from the operand of the bytecode instruction. Next, in operation 614, the value of a vtable index is loaded from the current constant pool cache at the index obtained. In operation 616, a determination is made as to whether the vtable index loaded from the current constant pool cache is equal to a distinguishable value used to denote an unresolved link. If the values are not equivalent, the method continues to operation 620 wherein the virtual method dispatch is performed. In one example, the virtual method dispatch is performed by using the vtable index to index the virtual table. In one instance, the virtual table is embedded in the instanceKlass object and is referenced from the object targeted by the method invocation. However, if the vtable index loaded from the current constant pool is an unresolved link marker, the method continues to operation 618 in which a call to the runtime is issued so as to resolve the symbolic link for the current class loader. For instance, the symbolic link can be resolved into a loader dependent method object, from which a sharedMethod object can be reached to obtain the vtable index for the method. The runtime can then replace the unresolved link maker at the entry of the constant pool cache with the vtable index obtained in operation 620. In this manner, subsequent uses of the constant pool cache entry can avoid calling the runtime.

[0133] According to one embodiment, link resolution barriers may be redundant in the presence of class initialization barriers (CIB). Specifically, when a class initialization barrier is required for interpreting a bytecode instruction, a link resolution barrier may not be needed. In one example, class initialization barrier may be needed in the implementation of bytecode instructions that may result in a first use of a class by a software program. FIG. 11A depicts a flow chart diagram 700 of exemplary method operations performed to determine whether a link resolution barrier or a class initialization barrier should be implemented, in accordance with one embodiment of the present invention. The method begins in operation 702 in which a decision is made as to whether a quickened version of getstatic, putstatic, invokestatic, or new is being generated. If the quickened bytecode for getstatic, putstatic, invokestatic, or new is being generated, the method continues to operation 704 in which class initialization barrier is implemented. Otherwise, the method continues to operation 706 in which a determination is made as to whether the quickened bytecode being interpreted refers to a class loader dependent symbolic link. If a class loader dependent symbolic link is being implemented, the method continues to operation 708 in which link resolution barrier is implemented.

[0134] FIG. 11B depicts a flowchart diagram 700′ of the method operations performed while interpreting a bytecode instruction manipulating a static variable, in accordance with one embodiment of the present invention. By way of example, bytecode instructions manipulating static variable may be getstatic or putstatic, or the quickened versions of getstatic or putstatic. The method begins in operation 710, in which an index to the current constant pool cache is computed from the operand of the bytecode instruction. Next, in operation 712, the address to a static variable is loaded from the current constant pool cache. In one example, the index is obtained from the bytecode instruction operand.

[0135] In operation 714, a determination is made as to whether the static variable address loaded from the current constant pool cache is equal to a distinguishable value used to denote an uninitialized class. If the values are not equal, the method continues to operation 718 wherein the static variable manipulation is performed. In one instance, static variable manipulation may be loading of the value of the static variable onto the expression stack when performing a getstatic, or storing a value from the expression stack onto the static variable when performing a putstatic. However, if the address loaded from the current constant pool is equivalent to an uninitialized class marker, the method continues to operation 716 wherein a call is issued to the runtime so as to initialize the class for the current class loader. In one example, calling the runtime may also resolve the symbolic link to the class before initializing the class. The runtime then replaces the uninitialized class marker at the entry of the constant pool cache with the address of the static variable, thus avoiding subsequent calling of the runtime when constant pool cache entry is used. Next, the method proceeds to operation 718. In one example, the NULL pointer value can be used as the uninitialized class marker. In this manner, the CIB for bytecode instructions that manipulate static variables can be implemented on the SPARC processor as a branch on non-zero register value, as shown in the exemplary code depicted in Table 3.

[0136]

TABLE 3
Exemplary Code using CIB
ld [Rcache + (header size + 2*wordSize)], Rvar  // load static variable
address
brnz,a,pt Rvar, initialized //CIB
ld [Rvar],Rvalue // Load static variable

[0137] Bytecode instructions for invoking a static method, such as invokestatic and the quickened version of invokestatic, refer to a constant pool cache entry in respective operands. The constant cache pool entry holds a reference to the method representing the static method. In one example, the CIB for such bytecode instructions is implemented by loading the pointer to the method object from the constant pool cache using the index computed from the operand of the invokestatic bytecode instruction. Then, the pointer to the method object from the constant pool cache is compared with an uninitialized class marker. If the value of the pointer equal to that of uninitialized marker, execution can be routed to runtime, enabling initialization of the class defining the static method. Runtime may also resolve the symbolic link to the class before initializing the class. In such a scenario, the constant pool cache entry can be filled with the pointer to the method object representing the static method, before runtime returns the method object pointer to the interpreter. In one example, a NULL pointer value can be used for the uninitialized class marker. In such a situation, the CIB for bytecode instructions invoke a static method that can be implemented on SPARC with a single branch on non-zero register value.

[0138] The quickened version of “new” bytecode instructions refers to the class pool entry holding a reference to the instanceKlass object that represents the class for which an instance is to be allocated. In one example, the class pool can be organized so as to hold two sets of information per entry. For instance, the first set of information includes the pointer to the instanceKlass object representing a class type while the second set of information includes the size of the instance of the class. A CIB for the new bytecode instruction is implemented by initializing the size information to the uninitialized class marker whose value is distinguishable from any of the possible instance sizes.

[0139] FIG. 11CD shows a flow chart diagram 700′ of method operations performed when interpreting the quickened version of the new bytecode, in accordance with one embodiment of the present invention. The method begins in operation 730, in which an index to the current class pool is computed from the operand of the bytecode instructions. Next, in operation 732, the instance size is loaded from the current class pool, at the index obtained from the bytecode instruction operand. In a different embodiment, the address of the current class pool can be obtained from the header of the current constant pool cache which address has been stored at a dedicated location upon entering the current method. Proceeding to operation 734, a determination is made as to whether the instance size value loaded from the current class pool is equal to the uninitialized class marker. If the values are not equal, the method continues to operation 738 where the allocation of heap space for the new instance is performed. In one example, the heap spaced allowed is an amount equivalent to the instance size obtained from the class pool.

[0140] Continuing to operation 740, the pointer to the instanceKlass object describing the instantiated class is obtained from the class pool entry defined adjacent to the class pool entry from which the instance size was obtained. Then, in operation 742, the pointer to the instanceKlass object is stored in the header of the allocated object. However, if in operation 734 a determination was made that the instance size loaded from the current constant pool is equal to an uninitialized class marker, the method continues to operation 736 wherein a call to the runtime is issued so as to initialize the class for the current class loader. The runtime may also resolve the symbolic link to the class before initializing the class. The runtime then replaces the uninitialized class maker at the entry of the class pool with the size of instances of the class, thus avoiding calling the runtime during subsequent uses of the class pool entry. Next, the method proceeds to operation 738. In one embodiment, the value of −1 can be used as the value of the uninitialized marker. In this manner, the CIB for the new bytecode instruction can be implemented on the SPARC processor as a single branch on negative register value.

III. Sharing of Dynamically Compiled Code—Loader Re-Entrant Code

[0141] Making compiled code loader re-entrant means that the same compiled code can be executed on behalf of different class loaders, regardless of how these class loaders resolve symbolic links to other classes. In other words, making the compiled code loader re-entrant allows different defining class loaders of the same method to share the same compiled code. This strategy has several advantages: (1) compilation costs are paid only once per shared representation of a method, hence the cost of compilation is amortized across the class loaders; (2) since the compiled code is loader re-entrant, the compiled code can be used immediately by any class loader defining the method, thereby eliminating bytecode interpretation; and (3) memory is saved by sharing the compiled code across the class loaders.

[0142] In one embodiment, the compiled code can be made loader re-entrant by avoiding the inlining of information derived from symbolic links whose resolved value depends on a particular class loader (e.g., pointers to static variables or addresses of methods to invoke) and by guarding against uninitialized classes and unresolved symbolic links because the same classes and symbolic links may be initialized/resolved in one loader but not in another loader.

[0143] Instead of inlining loader-dependent information, the information can be retrieved at runtime. The code produced by the dynamic compiler accesses information derived from the symbolic links by inlining loader-independent indexes to a table, hereinafter referred to as the constant pool cache, that contains loader-dependent information derived from resolved symbolic links. The address of the constant pool cache is determined at method invocation from the invoked method. Each loader-independent index identifies the same symbolic link in class types that share a same runtime representation of a class.

[0144] Specifically, let us assume two class types <C, L1> and <C, L2> such that <C, L1>˜<C, L2> and class C refers to a method m of a class B (denoted as B.m). Thus, in loader L1 (respectively, L2), B.m refers to method m of class type B^L1 (respectively, B^L2). The B.m is identified by the same index i_B.min both loaders L1 and L2. Instead of inlining the loader-dependent addresses of B^L1.m or B^L2.m in compiled code, the compiler generates instructions to retrieve these addresses from a loader-dependent table of resolved symbolic links at index i_B.m. The address to the loader-dependent table is determined at runtime upon method invocation. Typically, the loader-dependent part of the runtime representation of a method includes a pointer to the loader-dependent table of resolved symbolic links for the method. The loader-dependent table is shared between the loader-dependent part of the runtime representation of all methods of a class type. Upon invocation of a method, the runtime loads from the method's loader-dependent representation the table of resolved symbolic links and caches its reference in a location that can be efficiently accessed (e.g., in a register or on an execution stack).

[0145] In one embodiment, guarding against uninitialized classes and unresolved links can be done by generating instructions that implement barriers for class initialization and link resolution. Barriers are best implemented by storing a distinguishable value at entries of the constant pool caches, and testing for the value immediately after an entry is read, since almost all uses of the barriers coincide with reading an entry of the constant pool cache. Thus, barriers can be implemented as a test of the value retrieved from the constant pool cache.

[0146] For example, the following Table 4 summarizes the changes made to the code generator of the dynamic compiler of a particular implementation of the JVM (namely, the Java HotSpot™ virtual machine) to make the code generator produce loader re-entrant code, in accordance with one embodiment of the present invention.

[0147]

TABLE 4
		Loader-
		dependent
Bytecode	Barrier	information	loader re-entrant code
new	CIB	Instance size,	ld [LcpCache +
		address of	offset_to_size],Rsize
		instanceKlass	brnz,a Rsize initialized
			/*barrier*/
			ld [LcpCache +
			offset_to_klass], Rklass
getstatic/	CIB	Address of	id [LcpCache +
putstatic		static variable	offset_to_var
			address],Rvar
			brnz, a Rvar initialized
			...
invokestatic	CIB	entry point	ld [LcpCache +
			offset_to_call_address],
			Rentry_point
			call Rentry_point
			...
getfield/	LRB	offset to	ld [LcpCache +
putfield		instance	offset_to_field_offset],
		variable	Roffset
			brnz, a Roffset resolved
			...
invokevirtual	LRB	vtable index	ld [LcpCache +
			offset_to_method_offset],
			Rvtable_offset
			brgez, a Rvtable_offset
			resolved

[0148] The first column lists bytecodes whose implementation is loader dependent. The second column indicates the type of barrier required for the bytecode. The third column lists the loader-dependent information required to implement the bytecode. The fourth column provides an example of the loader re-entrant code emitted by the compiler for a Scalable Processor Architecture (SPARC) processor. As shown in Table 4, the code generated for a loader-dependent bytecode follows the same pattern: (1) obtain the loader-dependent information from the loader-dependent table of resolved symbolic links (in our example, the runtime guarantees that the address to that table is cached in register LcpCache); (2) test the value obtained from the table against a distinguishable value indicating that the symbolic link hasn't been resolved (on the SPARC, the distinguishable value is chosen such that a test and a branch can be performed with a single branch-on-register value instruction); and (3) exploit the information obtained from the table of resolved symbolic links (e.g., in the case of a getstatic bytecode, use the address obtained from the table in a load instruction to load the value of a static variable of a class).

[0149] The case of static method invocations needs further explanation. Loader re-entrance introduces two problems with respect to invocation of static methods. First, the class loaders sharing the loader re-entrant code may link to different implementations of a method. For example, assume a class A such that one of its method invokes a static method of a class B, and two loaders L1 and L2 such that <A, L1>˜<A, L2>. There is a possibility that B^L1≈B^L2 is false (i.e., B^L1 and B^L2 are not the same type and cannot share their runtime representation). Therefore, the code generated by the compiler needs to obtain the appropriate address of where to dispatch the static method invocation for the current class loader. Second, since the compiled code is shared across class loaders, there is a possibility that class B is not initialized for one class loader.

[0150] To address both problems, as illustrated in FIG. 12, an invocation of a static method fetches a target of a call from a constant pool cache of a caller. The value of the target is set according to a class initialization state of the class that defines the called static method, such that the call routes execution to a different place for different states. This solution avoids paying for a systematic test of the class initialization state by exploiting a registered indirect call to route execution to the appropriate code for handling static method invocation.

[0151] FIG. 12 is a simplified block diagram of compiled call site 820 for the static method using the constant pool cache, in accordance with one embodiment of the present invention. An entry of the constant pool cache 810 for the static method can be in one of three states 800, depending on an initialization state of a class whose static method is being invoked. When entry of the constant pool cache 810 used by the call site is in the NOT INITIALIZED state, call address field 805 of the constant pool cache entry contains the address of runtime handler for class initialization barrier 825. Runtime handler 825 tests the initialization state of the class of the static method being invoked. Because class initialization does not update all constant pool cache entries 810 that refer to static methods of the class, but just updates the constant pool cache entry that triggered class initialization, the state 800 of the class may be different from the one recorded in the constant pool cache entry. That is, the class may be in any of the three NOT INITIALIZED, BEING INITIALIZED, or INITIALIZED states 800.

[0152] If class is INITIALIZED, call address field 805 is set to an entry point of the method 815 (e.g., either a first instruction of the compiled code for the static method in one embodiment or the interpreter's code for entering the static method in another embodiment). If the class is BEING INITIALIZED, the call address field 805 is set to an address of an initializing stub 835 whereby the initializing stub is used to synchronize class initialization. The call address 805 of the initializing stub 835 is obtained from the method descriptor. The initializing stub 835 tests whether the caller's thread is the initializer. If not, the initializing stub 835 calls the runtime which will block the caller until the class is INITIALIZED. Otherwise, the initializing stub calls the entry point of the method 815.

[0153] On the other hand, if the class is NOT INITIALIZED, then two sub-cases are considered: if class initialization requires executing some code (i.e., the class declares one or more static initializers), the runtime creates the initializing stub 835 for the BEING INITIALIZED state, records its call address 805 in the method descriptor (such that calls from any location can find the initializing stub 835), updates the constant pool cache entry 810 of the caller with the call address of the initializing stub, and then proceeds to initializing the class. Once the class is initialized, the runtime sets the constant pool cache entry 810 used by the caller to the INITIALIZED state by updating the constant pool cache entry with the entry point 815 of the method's compiled code.

[0154] In one embodiment, the explicit test on a value of a vtable index on virtual method invocations (as shown in Table 4) can be eliminated using a similar mechanism: vtables are prefixed at a known negative offset with a pointer to a fake method descriptor, and constant pool cache entries for virtual methods are initialized with that negative offset. Thus, if a link to the method has not been resolved, the code for dispatching method invocation retrieves the fake method descriptor. This fake method descriptor is formatted as a normal method descriptor except that its entry point refers to a code stub that calls a runtime's method resolution mechanism. Once the method is resolved, the vtable index in the constant pool cache is updated with that of the resolved method. This trick eliminates the cost of the LRB, such that method invocation only pays for the indirection.

[0155] Several simple optimizations can further reduce the overhead of loader re-entrance. First, in one embodiment, barriers and indirections via the constant pool cache are unnecessary for compiled codes that refer to symbols of the current class and any of its super-classes. For instance, the code generated for a getfield for a field of an instance of the current class does not include a barrier, nor does it load an offset from the constant pool cache. Instead, the compiler simply treats the offset as a constant and uses the offset directly in the load instruction that loads the field.

[0156] The reason is that all symbolic links to the current class that do not refer to loader-private data (e.g., static variables) are constants across class loaders that share the representation of the current class (and therefore, that share the compiled code of its methods) since they all define the class from the same class file. The sharing conditions, as described above, also guarantee that all super-classes of a class C are either the same or are shared across loaders that share class C. Hence, symbolic links to the super-classes of class C remain constant across loaders that share C. The compiler can apply this optimization whenever it can statically determine that a symbol refers to a field or a method of the current class or any of its super-classes. Similarly, if a static type of the receiver at a call site is a sub-class of the current class and an invoked method m overrides a method of the current class or of one of its super-classes, then the vtable index of the method m is constant across defining loaders of the current class, and both the indirection and the LRB are unnecessary in this case. Nevertheless, there are still cases where an indirection needs to be maintained even though an operation refers to the current class. For example, indirection needs to be maintained when accessing the static variables of the current class and when allocating an object of the current class because the reference to the current class needs to be obtained based on the current loader.

[0157] Another embodiment to eliminate the cost of software LRB is to exploit a hardware detected fault, if available. For instance, instead of testing explicitly for a particular offset value, an LRB can be implemented by initializing offsets to instance variable in the constant pool caches with a misaligned offset, such that an aligned load instruction results in a hardware fault. The JVM then catches a corresponding signal, and execution is routed to the runtime to resolve the missing link. In one exemplary embodiment, this strategy eliminates all branch instructions from loader-dependent sections of the compiled code on a SPARC processor.

[0158] One source of overhead introduced by loader re-entrance is tracking the constant pool cache of the current method. However, this overhead can be avoided for compiled methods that do not use any symbolic links to other classes (or whose symbolic link can be determined to be constants across all loaders). These compiled methods eliminate one or two extra loads during method invocation, depending on the path to obtain the constant pool cache of the invoked method.

[0159] Some implementations of the JVM may use a form of inline caches to accelerate virtual method invocation, in accordance with one embodiment of the present invention. Loader re-entrance poses no particular problem to using the inline caches except that loader re-entrance may reduce their effectiveness when sharing is exploited. For example, let us consider a scenario where a call site in a method A.m invokes a method B.m, and two loaders L1 and L2 that each defines A such that they share its runtime representation (i.e., <A, L1>˜<A, L2>). Let us further assume that all the targets of all invocations from both loaders is the class B (i.e., B^L1 for invocations from <A, L1>, and B^L2 for invocations from <A, L2>. If B^L1=B^L2, then the inline cache will remain monomorphic and will be 100% efficient. However, if B^L1˜B^L2, then the inline cache becomes polymorphic, although invocations to methods of either class dispatch to the same shared compiled code. This illustrates how call sites that are monomorphic in absence of sharing become polymorphic when compiled code is made sharable.

[0160] To summarize, to produce loader re-entrant code, the dynamic compiler: (1) generates instructions at either a call site or on method entry, depending on the type of call, to keep track of the current constant pool cache; (2) generates instructions to fetch a value of a symbol from the constant pool cache of the current method if the symbol does not refer to the current class or any of its super-class; and (3) generates instructions that implement CIB and LRB whenever the compiler cannot determine if a symbol has been already resolved or a class already initialized. The present invention exploits the fact that barriers are needed when loader-dependent data is needed to efficiently implement barriers as a test of the value of data retrieved from the loader-dependent table. When the data retrieved from loader-dependent table is an address of code to execute next, the test can be omitted.

IV. Sharing of Dynamically Compiled Code—Clonable Conditional Loader Reentrant Code

[0161] As discussed above, making compiled code loader re-entrant may introduce additional overheads (e.g., due to barriers, indirections, and the tracking of the constant pool cache across method invocations) to the dynamic compiler. In general, dynamic compilation exploits the runtime knowledge of resolved links to remove the overhead of dynamic linking. For instance, the dynamic compiler can determine the offset to a field of an object at compile-time and generate a simple load instruction that does not require using at runtime any meta-information (such as the runtime constant pool cache). Such optimizations are not possible with loader re-entrant code because a level of indirection is required wherever a symbolic link to another class is used. For example, loading the field of an object requires determining at runtime what is the current class loader and then finding out what the offset to the field is in the context of that class loader.

[0162] A different approach to reduce the overhead of dynamic compilations while avoiding the overheads associated with loader re-entrant code is to maintain a copy of the compiled code private to each class loader and optimized for a particular class loader. However, instead of compiling from scratch the method for each class loader that defines it, the compiled code is produced by cloning (i.e., copying) an existing version of the compiled code, and modifying the cloned code's class loader dependent part. This makes generation of compiled code for a method substantially faster as steps for parsing bytecodes, building an intermediate representation, performing optimization of the intermediate representation, and generating code are avoided. However, the first compilation of a method from its bytecodes needs to record information specifying where class loader dependencies are made (e.g., identifying which sequences of instructions within the compiled code are class loader-dependent and the type of dependencies for each such sequences).

[0163] Producing a new version of the compiled code for a method already compiled is then just a matter of cloning an existing compiled code, using the information recorded during the first compilation from the method's bytecodes to locate the class loader-dependent sequences of instructions, and modifying the compiled code according to the context of the class loader on behalf of which the clone is being produced. Since compiled code generation is amortized faster, switching from bytecode interpretation to compiled code execution can take place earlier for methods that have been already compiled once for one loader.

[0164] In another embodiment, the information recorded to support code cloning can be further exploited by observing that the loader-dependent information derived from symbolic links may be constant across loaders that share the runtime representation of the class that defines the compiled method. For example, a symbolic link to an instance variable of a class B from a class A is constant across two loaders L1 and L2 such that <A, L1>˜<A, L2> if B^L1≅B^L2. If this is the only symbolic link used in a method m of A refer to instance variables of B, the compiled code for m can be shared between the class types <A, L1> and <A, L2> (note however that a different compiled code for m is used for any other class type <A, L> such that B^L1≅B^L is false). The compiled code for m shared between <A, L1> and <A, L2> is not loader reentrant, but conditionally loader reentrant. The difference is that loader reentrant code can be shared by two loaders defining the same class no matter how each loader resolves symbolic link to other classes, whereas, conditionally loader reentrant code exploits resolved links. In other words, conditionally loader reentrant code includes sequences of instructions that inline data obtained from a resolved link (e.g., an offset, or an address in memory) and should therefore be shared between loaders that resolve symbols to data of equal value.

[0165] FIG. 13 is a simplified block diagram of data structures to support both code cloning and conditional loader reentrance, in accordance with one embodiment of the present invention. A shared runtime representation of a method 930 includes bytecodes 950 of the method and a head of a list of compiled code for the method, one of which was obtained from compiling the bytecodes of the method, the others from cloning this first compiled code or any of its clone. The method shared runtime representation 930 also includes a dependency table 900 that contains the information recorded during compilation from the method's bytecodes 950 and used to support code cloning and conditional loader reentrance. The information recorded in the dependency table 900 are loader-independent and are used by the compiler to compute compiled code dependencies for a particular loader. Each compiled code of the linked list of compiled code can be used by one loader (e.g., compiled code 905 is used by loader L1) or more (e.g., compiled code 910 is used by loaders L3 and L2). Each loader-private method representation 940 includes a pointer to one of the compiled code of shared method object 930.

[0166] An exemplary format of the dependency table 900 is depicted in FIG. 13. The dependency table 900 is organized in three sections: a header 915, a dependent class summary section 920, and a dependencies section 925 that contains all the dependencies. Header 915 comprises a summary of the dependencies and the number of dependent class. The number of dependent class allows a size of the dependent classes summary section 920 to be determined. In one embodiment, the dependent class summary section 920 is made of a class pool index and a summary of the dependencies on the corresponding class. The summary identifies the types of dependencies to the class. The class pool index enables a class type to be identified in a loader-independent way such that the corresponding class type for a given loader is retrieved from the class pool of that loader's private representation of the method. In another embodiment, each dependencies section 925 comprises a bytecode position relative to the first bytecode instruction of bytecodes 950 and a program counter (pc) offset relative to the first instruction of the compiled code. Both are constant across all the class loaders that share a runtime representation of the class. The bytecode position allows bytecode instruction 955 to be retrieved from the method bytecodes 950. Constant pool cache index 945 is an operand of bytecode instruction 955, and the constant pool cache index identifies a symbolic reference whose resolved value is exploited by the compiled code generated for the bytecode instruction.

[0167] Additionally, dependencies section 925 are grouped per class, and each group is ordered similarly to the order used in dependent class summary section 920. This organization allows easy navigation through dependencies and allows skipping at once all the dependencies to a class when no modifications to compiled code that depends on this class is required.

[0168] The compiled code of a method can be shared between multiple class loaders (that already share the runtime representation of the class that defines the method) if the compiled code has no dependency, or if the class loaders have exactly the same dependencies. For example, the method may refer to symbols of classes defined by the boot class loader. In this case, the compiled code for the method can obviously be shared between the class loaders since symbolic links to methods and variables of boot classes will refer to exactly the same object in all class loaders.

[0169] FIG. 14 is an exemplary code that shows how sharing of compiled code can be permitted even if the classes referred to by the compiled code are not the same, in accordance with one embodiment of the present invention. FIG. 14 shows two classes A and B and two class loaders L1 and L2 that define each class A such that <A, L1 >˜<A, L2>. Although <A, L1 >≠<A, L2>, the compiled code produced for a method getx 1400 of class A can be shared between class types <A, L1> and <A, L2> since the one dependency of the compiled code is the offset to an instance variable x 1425 of class A, which is guaranteed by the sharing conditions, as described above, to be the same for both class loaders. Such symbolic link references from a class to itself do not need to be recorded in a dependency table. In contrast, the compiled code produced for a method getx 1405 cannot be shared between the class loaders L1 and L2 as the method getx 1405 depends on the address of a static variable X 1420, which differs for each defining class loader L1 and L2. The case for a method foo 1410 is more subtle: if C^L1≅C^L2, then the compiled code can be shared since z resolves to the same offset for both class loaders, either because they are the same class type or because sharing conditions guarantee this. Otherwise, the method cannot be shared. Similarly, the compiled code for a method bar 1415 can be shared between L1 and L2 if B^L1=B^L2.

[0170] The following Table 5 shows the type of dependency and when code modification is required to adapt a clone of a compiled code to a new class loader, in accordance with one embodiment of the present invention.

[0171]

TABLE 5
	Conditions to Leave	What to Change if
Type of Symbolic link	Code Unchanged	Condition is False
instance variable	< C, Lr>≃<C, Lo>	offset in load/store
		instruction
static variable	< C, Lr >=< C, Lo >	static variable address in
		immediate value register
		load
class	< C, Lr >=< C, Lo >	class address and
		instance size in
		immediate value register
		load
static method	< C, Lr >=< C, Lo >	address of method entry
		point in call instruction
dynamically bound	< C, Lr >≃<C, Lo>	vtable index in load
method		instruction
statically bound method	< C, Lr >=< C, Lo >	address of method entry
		point in call instruction

[0172] Table 5 lists for each type of dependency the conditions for leaving a compiled code unmodified, and what modifications are required otherwise. Lr is the class loader requesting the clone and Lo is an owner of the original compiled code. When requested to produce a clone of the compiled code of a method, the compiler iterates over the dependency table of the method and determines what modifications are needed. A copy of the compiled code is made the first time the need for a modification is determined. If no changes are required, no copy is made and a requester shares the compiled code with its current owners. To this end, the compiler adds the method of the requester to the set of owners of the compiled code, and the compiled entry point of the requester's method is set to that of the compiled code. Additionally, if inline caches are used to implement virtual method invocation, the modification to the clone consists of resetting its inline caches instead of change the vtable index.

[0173] FIG. 15 is a simplified flowchart diagram of a compiler iterating over a dependency table of a method to determine if the compiled code of a method m can be shared between two defining loaders of a class Cm, in accordance with one embodiment of the present invention. Lr is a class loader requesting the compilation of the method m of class type <Cm, Lr> and Lo is an owner of a compiled code of the method m of class type <Cm, Lo> such that <Cm, Lr>˜<Cm, Lo>.

[0174] Starting in operation 1500, the compiler starts with setting a result variable can_share to true. Then, for every class C in the class dependencies summary section of the dependency table, the compiler tests if C^Lo (operation 1510) and C^Lr (operation 1520) are resolved by their respective loaders. If C^Lo and C^Lr cannot be resolved, the compiled code cannot be shared since information is missing to determine whether the dependencies are the same, and the compiler move on to operation 1570 to set the result variable can_share to false. Otherwise, in operation 1530, the compiler tests if the two class type C^Lr and C^Lo are the same. The test is done by comparing references to the loader-private runtime representation of class types C^Lr and C^Lo obtained from class pools of <Cm, Lr> and <Cm, Lo>, respectively, at the index recorded at the entry in the class dependencies summary for class C. If the test succeeds, the compiler can move to the next class in the class dependencies summary section in operation 1580 if more classes are available. If the class types are not the same, the compiler further determines if sharing is possible in operation 1540.

[0175] In operation 1540, the compiler first tests if the compiled code has dependencies that requires C^Lr and C^Lo to be the same class type. If C^Lr and C^Lo are required to be the same class type, then the compiled code cannot be shared (previous step has already determined that they were different), and the compiler moves on to set the can_share result variable to false in operation 1570. Otherwise, the compiler continues to operation 1550 to determine if C^Lr and C^Lo share the same runtime representation. The determination is done by comparing the reference to the shared runtime representation of class C stored in the loader-private representation of each class type C^Lr and C^Lo which are obtained from the constant pool cache of <Cm, Lr> and <Cm, Lo>, respectively, as already described above. If the shared runtime representation of both class types is the same, the compiler continues on to the next entry of the class dependencies summary section in operation 1580 if another entry is available. Else, in operation 1560, the compiler further tests if the compiled code has dependencies that requires C^Lr˜C^Lo. If the compiled code has dependencies that requires C^Lr˜C^Lo, then the compile code cannot be shared and the compiler moves on to set the can_share result variable to false in operation 1570. Otherwise, the compiler continues on to operation 1580.

[0176] If the compiler has determined that the compiled code of method m for <Cm, Lo> cannot be used for method m of <Cm, Lr>, the compiler may determine if sharing is possible with another compiled code for method m if another one is available. It is possible that performing the determination for all the available compiled code for method m may be too expensive. In this case, the compiler may decide to clone one of the available compiled code before the compiler has performed the determination on all the available compiled code.

[0177] How does the compiler choose what compiled code to clone when more than one is available? In one embodiment, the compiler chooses the compiled code that is the most popular, i.e., the one shared by the largest number of class loaders, on the ground that it represents the common case. In another embodiment, the compiler iterates over each compiled code and takes the compiled code that requires the smallest number of changes. In order to bound the time it takes for choosing the compiled code to clone, the number of compiled code iterated over may be limited. In still another embodiment, the compiler chooses the most recently cloned compiled code.

[0178] FIG. 16 is a simplified flowchart diagram of a compiler iterating over a dependency table of a method m to produce a clone of one of its compiled code, in accordance with one embodiment of the present invention. Lr is a defining loader of class Cm for which the compiler produces the clone and Lo is a defining loader of Cm which owns the compiled code of m that is being cloned, such that <Cm, Lr>˜<Cm, Lo>.

[0179] For every class C in the class dependencies summary section of the dependency table, the compiler tests if C^Lr and C^Lo have been resolved in operations 1600, 1610, and 1650. If neither are resolved, no change to the sequences of instructions that depends on class C are necessary as all use of C in the compiled code is made of call to the runtime to resolve C, and the compiler continues to operation 1660. If only one of C^Lr and C^Lo is resolved, the compiler moves to operation 1640 where the compiler changes all sequences of instructions that depend on C into call to the runtime to resolve C^Lr. This means that even if C^Lr is already resolved, the compiler still changes sequences of instructions that depends on C into calls to the runtime. In another embodiment, this case may be dealt differently by changing the sequence of instructions according to Table 5, assuming that C^Lr=C^Lo. Thereafter, the compiler continues to operation 1660.

[0180] If both C^Lr and C^Lo are resolved, the compiler continues to operation 1630 to test if the two class types C^Lr and C^Lo are the same. The test is done by comparing references to the loader-private runtime representation of class types C^Lr and C^Lo obtained from class pools of <Cm, Lr> and <Cm, Lo>, respectively, at the index recorded at the entry in the class dependencies summary for class C. If the test succeeds, no modification is needed because all symbolic links to class C refer to the same class in both loaders Lr and Lo and the compiler move to operation 1660. If the test fails, the part of the dependencies section that contains all the dependencies to C is iterated over, and for each dependency, the corresponding sequence of instructions within the compiled code is modified according to Table 5, as illustrated in operation 1630 (depending on how the conditions of Table 5 evaluate for a given dependency, the corresponding sequence of instruction may be left unchanged). When all the dependencies on C have been iterated over, the compiler moves to operation 1660, which determines if there is another class in the class dependencies summary section to process. If there is another class, the compiler starts from operation 1600 again. Otherwise, the compiler ends and the clone of the compiled code is ready for use by method m of <Cm, Lr>.

[0181] Finally, the present invention can be a combination of all the strategies discussed above, in accordance with another embodiment of the present invention, whereby compiled code can be made either loader re-entrant, in which case a single version of compiled code is maintained for a method, or conditionally loader re-entrant, in which case the compiled code may be shared if dependencies across loader are constants, or cloned if the dependencies across the loader are not constants.

[0182] As can be appreciated, the present invention describes how an implementation of the JVM can share the runtime representation of classes, including the bytecodes of classes, across multiple defining loaders. Sharing is achieved by separating the part of the runtime representation of a class that depends on loader-specific data (e.g., resolved links, static variables, etc.) from the part of the runtime representation that is loader independent. Furthermore, byte code interpretation is made loader re-entrant by adding link resolution and class initialization barriers as well as efficient access to loader dependent parts of the runtime representation of the class.

[0183] The aspects of the present invention are numerous. Most importantly, the amount of main-memory consumed by executing software programs is reduced by allowing the runtime representation of dynamically loaded software components to be shared when certain sharing conditions have been met. In this manner, beneficially, the embodiments of the present invention may avoid dealing with situations wherein object layouts or virtual tables are different. Another aspect of the present invention is that interpretation of bytecode is made class loader re-entrant by re-organizing the runtime representation of classes and adding class initialization and link resolution barriers. In this manner, the bytecode of methods, one of the most space-consuming part of the runtime representation of a class, can be shared across class loaders, while loader dependent part, of the runtime representation of classes can be efficiently accessed. Still another aspect of the present invention is that sharing imposes almost no space or processing overhead when classes are defined by a single class loader. Yet another aspect is the capability of the present invention to factor out the cost of loading a class across multiple loaders defining the class.

[0184] Yet another aspect is that by avoiding the costs associated with executing a software program each time, the impact on both the start-up time of the Java program as well as the overall performance of the Java program can be significant. In this manner, sharing the main memory runtime representation of classes can reduce both the footprint of Java programs as well as the runtime costs.

[0185] The present invention also describes how the dynamic compiler of a JVM capable of sharing the runtime representation of classes across a plurality of loaders can produce loader-reentrant code. In addition to the benefits already brought by the sharing of the runtime representation of classes across loaders, compilation costs for a class defined by multiple loaders are paid only once (as opposed to once per loader), since the loader reentrant code thus produced can be shared unconditionally across the loaders. Furthermore, because the code is loader-reentrant, it can be used immediately, therefore eliminating interpretation costs. This should result in better overall performance for applications that repeatedly use class loaders (e.g., web-server, application server, applet container, etc.). Further, as only one copy of compiled code is maintained for the same method defined by multiple class loaders, memory footprint is saved and, thus, makes the application scaling better with the number of class loader used.

V. Inline Caching in Presence of Code Sharing Between Class Loaders

[0186] Sharing of compiled code across class types may make inline caching ineffective in some cases, and the embodiments described herein provide additions to inline caching in order to take such cases of sharing into account. As will be explained in more detail below, the improvements to inline caching include mechanisms that take into account the potential sharing, by multiple class loaders, of the target of a method invocation. The improvements are related to implementations of the JVM that are capable of sharing the runtime representation of classes, including bytecodes, and under some conditions, compiled code produced by the dynamic compiler from bytecodes, across multiple class loaders of the same or multiple programs.

[0187] A straightforward and popular implementation of method invocation in object-oriented programming languages with single-inheritance (such as the Java programming language or C#) uses a virtual method table (or vtable for short). Typically, the address of the vtable for a given class instance can be retrieved with a single load from a header stored in each class instance. Entries of the vtable hold information to locate the address of the executable code for a method. In some implementation supporting mix-mode execution (i.e., both bytecode interpretation and compiled code execution), a method may have multiple entry points. Each entry point comprises specialized code to handle particular cases of who issued the call (i.e., call issued from the interpreter or from compiled code), and what form does the called code has (is the called code interpreted or compiled). Typically, the addresses of the sequences of instructions that implement the entry points to a method are recorded in the data structure describing the method, which is herein referred to as a method object. Given such an organization of data structures, a straightforward implementation is to store in each vtable entries a reference to a method object.

[0188] Thus, a table-driven method call can result in three loads (loading of the receiver's class descriptor from its header, loading of the method object from the vtable embedded in the class, and loading of the method's entry point for this call from the method object) and one register indirect call. Both loads and register indirect calls are expensive on deeply pipelined modern processor architectures.

[0189] Inline caching seeks to eliminate some loads from method invocations and to replace the register indirect calls with much cheaper direct calls by specializing call sites that repeatedly dispatch to the same method of the same class type. To this end, inline caching relies on the observation that many call sites are monomorphic in nature (i.e., never issue a call to more than one different receiver class type). In this case, the level of indirection provided by the vtable, and which helps supporting polymorphism efficiently, is not necessary. However, one cannot just generate a call to the address of the target of the method invocation at compile-time, even if compilation takes place at runtime. First, because although a call site may be monomorphic, the type of the call site's target may not be known until runtime upon the first use of the call site. Thus, the setting of the direct address to the invoke method's entry point had to be postponed until the first method invocation from the call site. Second, although statistics have shown that most call sites are monomorphic, it is impossible in general to tell in advance which call sites are strictly monomorphic. Further, a call site might be monomorphic for a given period, and then becomes polymorphic. Polymorphic call sites may dispatch to methods of different class type during execution of a program, and such dispatch may be detected and adjusted at runtime from an initially monomorphic call site to a polymorphic call site.

[0190] Inline caching addresses all these issues as follows. A call site constantly evolves during program execution and can be in one of the three following states: unresolved, monomorphic, megamorphic. FIG. 17 is a simplified block diagram illustrating the evolution of a call site, in accordance with one embodiment of the present invention. When in the unresolved state 1704, the call site routes execution to a runtime function that computes the target of the call and overwrites the call site with instructions that implements a monomorphic call site 1702. A monomorphic call site 1702 first sets a register to the address of the cached receiver class type of the invoked method, and then calls directly the address of an unverified entry point of the method selected for that cached receiver class type.

[0191] The unverified entry point of a compiled method checks if the class type of the actual receiver of the method is the same as the class type cached at the call site. If they differ, execution is routed to an inline cache miss handler. Otherwise, the code is executed.

[0192] The inline cache miss handler can transform the monomorphic call site 1702 into a different kind of call site, called megamorphic, by replacing the instruction calling the unverified entry point with an instruction calling a stub that implements the more expensive table-driven method dispatch.

[0193] Polymorphic inline caching is a refinement of the above technique that supports dispatching to a small number of target without the recourse to the dispatch table. Instead, a sequence of comparisons with multiple class types is performed in an out-of-line stub when a call site has a small number of targets. In other words, the call instruction at the call site routes execution to a small stub that comprises instructions for testing the class type of receiver of the invoked method with multiple class types of past receivers of the invoked method. However, polymorphic works well if the number of targets of a call site is small.

[0194] Implementation of the Java Virtual Machine (JVM) can scale better by sharing the runtime representation of classes, including the bytecodes and, under some conditions, dynamically compiled code of methods between multiple class loaders (of the same or multiple) programs. As discussed above in Section II, it has been shown that the sharing of most of the runtime representation of classes (excluding compiled code) across multiple class loaders of the same or of different programs is possible if three conditions hold between the defined class types. Thus, each class type is implemented as a small loader-dependent class descriptor and a larger loader-independent runtime representation that may be used by multiple class types. A typical implementation includes in each loader-dependent class descriptor a reference to its loader-independent runtime representation.

[0195] Further, sharing of the compiled code of a method defined by class types that already share their runtime representation is also possible under some conditions. This is the case, for instance, when the method is loader-reentrant (i.e., the method does not include any sequence of instruction that depends on loader-specific data, such as the address to a class's static variable) or when the method is loader-reentrant (i.e., the method includes sequences of instructions that are specific to a class loader but that are identical across multiple (but not all) defining loaders of class type that share their runtime representation).

[0196] Sharing compiled code across class loaders offers two main benefits: sharing avoids recompiling a method from bytecodes, and sharing makes compiled code available almost immediately, thus avoiding interpretation costs (in implementation of the JVM that mix bytecode interpretation and dynamic compilation, interpretation is used until dynamic compilation becomes cost-effective to compile bytecode into native code).

[0197] However, sharing of compiled code may make inline caching ineffective and may result in worse overall performance than if the compiled code is not shared. To understand why, reference is made to the exemplary code depicted in the following Table 6.

[0198]

	TABLE 6
	class B {	class A {
	C c;	public int fg(B b, int k){
	Random rgen;	return b.g(k) * 4;
	int z;	}
	public int g(int j) {	public int fh(B b, int k){
	return c.i + j;	return b.h(k) * 4;
	}	}
	public int h(int k){	private int j;
	return z*rgen.nextInt( ) % k;	...
	}	}
	}

[0199] Table 6 shows two classes A and B, wherein A defines a method fg (respectively fh) invoking a method g (respectively h) of class B, and two class loaders L1 and L2 such the following class types are defined <A, L1>, <B, L1>, <A, L2>, and <B, L2>. Lets further assume that <A, L1> and <A, L2> and, respectively, <B, L1> and <B, L2> can share their runtime representation (denoted <A, L1>˜<A, L2> and <B, L1 >˜<B, L2>). Lets further assume that the compiled code of methods B.g and B.h can be shared across loaders L1 and L2. Let also assume that the call site in method f to invoke method g is monomorphic.

[0200] If <A, L1> and <A, L2> use each a distinct copy of the compiled code for fh, all invocations to b.h from <A, L1>, respectively, <A, L2> will go through a monomorphic inline cache configured for <B, L1>, <B, L2>, respectively, regardless of whether compiled code for b.h is shared between <B, L1> and <B, L2>, and thus achieving best performance of virtual method invocations. However, if <A, L1> and <A, L2> share the same compiled code for method fh, then invocation through the inline cache of fh's compiled code has now two receiver class types, namely <B, L1> and <B, L2>, even if the call routes to the same compiled code for b.h when this one is shared between <B, L1> and <B, L2>. In this case, the sharing of fh's compiled call is detrimental to performance as sharing results in turning a naturally monomorphic inline cache into a megamorphic, table-driven call site.

[0201] Sharing of the compiled code of a method across class loaders may be unconditional, or conditional to some dependencies. When sharing of compiled code is unconditional, the compiled code is said to be loader reentrant. When sharing of compiled code is conditional to some dependencies, or not possible at all, the compiled code is said to be loader dependent. To illustrate this, let us consider again the class B shown in Table 6. Let us assume that two class loaders L1 and L2 define, respectively, class types <B, L1> and <B, L2>. Method h is loader reentrant since its definition does not rely on any data specific to either L1 or L2. The condition for sharing of the runtime representation of B between defining loaders L1 and L2 implies that the location of instance variable z is the same in instances of <B, L1> and <B, L2>. Further, class Random is defined by the primordial class loader and is the same across all class loaders (including in some implementation of the JVM that supports multitasking).

[0202] By comparison, method g is conditionally loader re-entrant as the method depends on how both loader L1 and L2 resolve class C. Sharing of the compiled code of g between L1 and L2 is allowed if C^L1≅C^L2. That is, class C with initiating loader L1 is either the same class type as or satisfy the sharing conditions with class C with initiating loader L2. If this is the case, then compiled code for g can be shared between <B, L1> and <B, L2>.

[0203] As discussed previously, sharing of compiled code can lead to worse performance than when using distinct copy in part because sharing degrades the efficiency of inline caching. Polymorphic inline caching can mitigate this performance problem by turning the inline cache into a multi-line cache. However Polymorphic inline caching's performance degrades with the number of targets and can improve performance for a small number of different targets.

[0204] When the compiled code is loader-reentrant, the compiled code is guaranteed to be re-usable unconditionally by all class types sharing the same runtime representation. Thus, it is possible to avoid testing individually each class type and to test instead if the class type of the receiver has the same shared runtime representation as the class type that first resolved the call site. This is referred to as a shared monomorphic inline cache. The shared monomorphic inline cache is implemented by caching the reference to the shared runtime representation of a class type at a call site, and then testing in the unverified entry point of the invoked method if the shared runtime representation of the receiver's class type equals the cached one.

[0205] The shared monomorphic inline cache incurs an extra-load to obtain the reference to the shared runtime representation of a class from the class descriptor of the receiver. On the other hand, the shared monomorphic inline cache avoids the costly register-indirect call of table-driven method dispatch, and can handle an arbitrarily large number of receiver's class types that share their runtime representation.

[0206] This exemplary embodiment may be used when more that one class types that share the same runtime representation define a loader-reentrant method. When the call site to a loader reentrant method is monomorphic, the standard monomorphic inline caching can be used. However, handling a cache miss from a call site in the monomorphic state may result in setting the inline cache to a shared monomorphic state.

[0207] Accordingly, loader-reentrant methods can be compiled with three different entry points, namely a verified entry point, an unverified entry point for call sites in the monomorphic state, and a shared unverified entry point for call sites in the shared monomorphic state. FIG. 18 is a simplified block diagram illustrating the different entry points to the compiled code of a loader-reentrant method, in accordance with one embodiment of the present invention. As shown in FIG. 18, loader-reentrant method 1850 is compiled with verified entry point 1852, unverified entry point for monomorphic call sites 1854, and shared unverified entry point for shared monomorphic call sites 1856. The verified entry point 1852 bypasses any type checks related to inline caching and is used for table-driven calls from polymorphic call sites. The unverified entry point for monomorphic call sites 1854 is used when the call site invokes a method with a receiver of the same class type and performs a test for comparing the receiver class to the call site's cached receiver. The shared unverified entry point for shared monomorphic call sites 1856 is used when the call site invokes a method with receivers of possibly different class types that all share their runtime representation. The shared unverified entry point performs a test comparing the receiver class type's reference to its shared runtime representation with the one cached at the call site.

[0208] The original inline caching mechanism requires the verified and unverified entry points described above. FIG. 19 is a simplified block diagram illustrating how the standard inline caching mechanisms can be modified to simply support sharing of code of loader-reentrant method across class loaders, in accordance with one embodiment of the present invention. Initially, the inline cache associated with a call site is empty. In this embodiment, inline caching operates as in the original technique: the call site routes execution directly into the runtime for resolving the target of this first invocation from the call site. The call site is then overwritten with a direct call to the unverified entry point of the code of the resolved method, and the reference of the resolved type is written in the inline cache at the call site. Subsequent calls jump directly to the unverified entry point of the code and perform a type check to verify that the receiver's type equals the cached type. If the test succeeds, the execution branches to the code's verified entry point. If the test fails, a cache miss handler is called.

[0209] If the method is not loader-reentrant, the cache miss handler simply turns the call site into a megamorphic call site 1802 that use table-driven dispatch. Otherwise, the cache miss handler first determines if the receiver's type shares its runtime representation with that of the cached receiver. This can be done simply by comparing the references of the two class types to their respective shared runtime representations. If they are the same, then the call site can be changed to a shared monomorphic call site 1804. Otherwise, the call site is turned into a table-driven dispatch call.

[0210] Calls from a call site associated with a shared monomorphic inline cache jumps to the shared unverified entry point of the code of the loader-reentrant method. The shared unverified entry point compares the reference of the shared representation of the receiver's class type with the cached reference to a shared representation of class type. If they are the same, the execution branches to the verified entry point of the method's code, otherwise a second level cache miss handler is called to change the call site into a table-driven call.

[0211] The following Table 7 shows an exemplary embodiment of the entry points of loader reentrant methods on a SPARC V9 processor:

[0212]

	TABLE 7
	unverified_entry_point_1:
	ld [ %recv + offset_to_class ], %class
	cmp %Ginline_cache, %class
	be verified_entry_point
	nop
	call ——cache_miss_handler1 nop
	unverified_entry_point_2:
	ld [ %recv + offset_to_class ], %class
	ld [ %class + offset_to_shared_rep],
	%shared_rep
	cmp %Ginline_cache, %shared_rep
	be verified_entry_point
	nop
	call ——cache_miss_handler2 nop
	verified_entry_point:

[0213] As shown in Table 7, the shared monomorphic inline cache adds a single load to the original inline caching mechanism.

[0214] In implementations of the JVM that use mixed-mode execution (bytecode interpretation and execution of native code produced by dynamic compilation), the target of method invocations from a compiled call site may be an interpreted method. Sharing of bytecodes across class loaders can be achieved by making the bytecode interpreter loader-reentrant, as discussed above. In this embodiment, sharing of bytecodes is allowed regardless of loader-dependencies. Hence, when a target of a call is interpreted, the cache miss handler for the monomorphic state can set an inline cache to the shared monomorphic state if the receiver type and the cached receiver type share the same runtime representation, regardless of whether the target is loader-dependent. When compiled code is produced for the interpreted method for at least one of its receiving type, the state of the inline cache is changed if the compiled code is loader-dependent. Two straightforward options are to (i) reset the inline cache to the unresolved state, or (ii) set the inline cache to the megamorphic state.

[0215] FIG. 20 is a flowchart diagram of an implementation of the unverified entry point for calls from an inline cache set in the monomorphic state, in accordance with one embodiment of the present invention. Starting in operation 1902, an unique identifier of a class type of the receiver of the method invocation is retrieved. A check is conducted in operation 1904 to determine whether the unique identifier of the class type of the receiver equals to an unique identifier of a class type cached at a call site. If they are equal, then operation 1914 branches to verified entry point of the method. Else, another check is conducted in operation 1906 to determine whether a shared representation of a class type of the receiver equals to a shared runtime representation of the class type cached at the call site. If they are not equal, then the call address at the call site is changed to a stub containing virtual table driven method call, and execution is routed to an invoked method in operation 1916. Otherwise, another check is conducted in operation 1908 to determine whether the invoked method is class loader-reentrant.

[0216] As shown in FIG. 20, if the invoked method is not class loader reentrant, then operation 1916 discussed above is executed. However, if the invoked method is class loader-reentrant, then cached unique identifier of class type is changed with the unique identifier of the shared representation of the cached class type in operation 1910. Subsequently, in operation 1912, the call address at the call site is changed to an address of shared unverified entry point for calls from shared inline cache. Finally, operation 1914 branches to verified entry point of the method.

[0217] FIG. 21 is a flowchart diagram of an implementation of the shared unverified entry point for calls from an inline cache set in the shared monomorphic state, in accordance with one embodiment of the present invention. Starting in operation 2002, an unique identifier of a class type of the receiver of the method invocation is retrieved. Subsequently, in operation 2004, the unique identifier of a shared representation is retrieved using the unique identifier of the class type of the receiver of the method invocation. A check is then conducted in operation 2006 to determine whether the shared runtime representation of the class type of the receiver is equal to a shared runtime representation cached at a call site. If they are equal, then operation 2008 branches to verified entry point of the method's code. Else, the call address is changed in operation 2010 at a call site to a stub containing virtual table driven method call, and execution is routed to the invoked method.

[0218] To summarize, the invention improves the efficiency of inline caching when code can be shared across class loaders as follows:

• • [0219] 1. An extra state, called shared monomorphic, is added to the number of possible state of an inline cache. The shared monomorphic state is used for inline caches at call sites that target a loader-reentrant method with more than one type of receivers, and such that all the receiver types are defined by different loaders that can share the same runtime representation of class type.
  • [0220] 2. Compiled code of loader reentrant methods are supplemented with an additional entry point, called the shared unverified entry point, that is used by a call site in the shared monomorphic state. The shared unverified entry point verifies that the receiver shares its runtime representation with the cached receiver. If the runtime representation is shared, then the compiled code of the method is executed. Otherwise, execution is routed to a runtime function that handles the inline cache miss.
  • [0221] 3. The handler for inline cache misses when the cache is in the monomorphic state is supplemented with logic to determine if the class type of the receiver of the call shares its runtime representation with the current cached receiver type. If the runtime representation is shared, the cache is set in the shared monomorphic state if either of the following holds: (i) the target has no compiled code, or (ii) the target has loader-reentrant compiled code. In all other cases, the inline cache is set in the megamorphic state.
  • [0222] 4. An additional handler is necessary to handle cache misses from inline cache in the shared monomorphic state. The handler sets the cache to the megamorphic state.

[0223] Although specific reference is made to terminology defined by Sun Microsystems, Inc., it should be understood that any name could be used for such terms, so long as the desired functionality is achieved. For instance, reference is made to the JVM and class loaders, but in a broad sense, these terms simply define the underlying environment. As such, it must be understood that the system of the present invention can be implemented in any appropriate environment so long as the runtime representation of program components can be shared across a plurality of component loaders. Accordingly, any reference to a particular standard should be viewed only as exemplary and focus should be placed on the claimed functional operation.

[0224] With the above embodiments in mind, it should be understood that, the invention may be practiced with other computer system configurations including hand-held devices, microprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframe computers and the like. Furthermore, the invention may employ various computer-implemented operations involving data stored in computer systems. Such operations require physical manipulation of physical quantities. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared, and otherwise manipulated. Further, the manipulations performed are often referred to in terms, such as producing, identifying, determining, or comparing.

[0225] Any of the operations described herein that form part of the invention are useful machine operations. The invention also relates to a device or an apparatus for performing these operations. The apparatus may be specially constructed for the required purposes, or it may be a general-purpose computer selectively activated or configured by a computer program stored in the computer. In particular, various general-purpose machines may be used with computer programs written in accordance with the teachings herein, or it may be more convenient to construct a more specialized apparatus to perform the required operations.

[0226] The invention can also be embodied as computer readable code on a computer readable medium. The computer readable medium is any data storage device that can store data which thereafter be read by a computer system. Examples of the computer readable medium include hard drives, network attached storage (NAS), read-only memory, random-access memory, CD-ROMs, CD-Rs, CD-RWs, magnetic tapes, and other optical and non-optical data storage devices. The computer readable medium can also be distributed over a network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion. Furthermore, although the present invention primarily targets the Java programming language, it can be used for other programming languages with similar features (e.g., SmallTalk, any object oriented programming language with dynamic loading, etc.).

[0227] Although the foregoing invention has been described in some detail for purposes of clarity of understanding, it will be apparent that certain changes and modifications may be practiced within the scope of the appended claims. Accordingly, the present embodiments are to be considered as illustrative and not restrictive, and the invention is not to be limited to the details given herein, but may be modified within the scope and equivalents of the appended claims.