US 5,889,996 AGrant
Accelerator for Interpretive Environments
Issue Date:1999-03-30
•31 Claims
•5 Drawing Sheets
Abstract
An apparatus and method for accelerating interpreters, interpretive environments, and the like optimizes the use of caches closest to a processor. An instruction set implementing a virtual machine (interpreter, interpretive environment) is written to fit each instruction at an individual cache line's address in the processor cache. The processor cache may be loaded with the instruction set in a compiled, linked, loaded image. After loading the processor cache, the cache is pinned, locked, disabled from flushing the contents or replacing the contents of any cache line. Faster loading of the processor cache may be achieved by flushing the processor cache and running an application containing all of the instructions of the virtual machine instruction set. Level-1 processor caches integrated into central processing units, particularly instruction caches or code caches are ideally suited to implementation of the invention. Examples include Intel's Pentium.TM. class products and Motorola's Power PC Processors.
Metadata
Assignee
- Novell Inc.
Inventor
- Phillip M. Adams
Application Information
Application Number:US 7670559
Filing Date:1996-12-16
Priority Date:1996-12-16
Art Unit:274
Classifications
IPC:
G06F 930
Field of Search:
395711705;708113;118;138;119;120;121;122;123;140;125;133-135;137;144;159
Patent Drawings (5 sheets)
Description
Background
1. The Field of the Invention
The invention relates generally to computers hosting interpreted languages and emulators, and more specifically to accelerators for emulators and interpreters such as Windows.TM., JAVA, Visual Basic, and other virtual machine environments executable by processors having access to caches.
2. The Background Art
Operations executed by a processor of a computer proceed in a synchronization dictated by a system clock. Accordingly one characteristic of a processor is a clock speed. For example, a clock speed may be 33 megahertz, indicating that 33 million cycles per second occur in the controlling clock.
A processor may execute one instruction per clock cycle, less than one instruction per clock cycle, or more than one instruction per clock cycle. Multiple execution units, such as are contained in a Pentium.TM. processor, may be operated simultaneously. Accordingly, this simultaneous operation of multiple execution units, arithmetic logic units (ALU), may provide more than a single instruction execution during a single clock cycle.
In general, processing proceeds according to a clock's speed. Operations occur only as the clock advances from cycle to cycle. That is, operations occur as the clock cycles. In any computer, any number of processors may exist. Each processor may have its own clock. Thus, an arithmetic logic unit (ALU) may have a clock operating at one speed, while a bus interface unit may operate at another speed. Likewise, a bus itself may have a bus controller that operates at its own clock speed.
Whenever any operation occurs, a request for interaction is made by an element of a computer. Then, a transfer of information, setup of input/output devices, and setup of the state of any interfacing devices, must all occur.
Each controller of any hardware must operate within the speed or at the speed dictated by its clock. Thus, clock speed of a central processing unit does not dictate the speed of any operation of a device not totally controlled by that processor.
These devices must all interface with one another. The slowest speed will limit the performance of all interfacing elements. Moreover, each device must be placed in the state required to comply with a request passed between elements. Any device that requires another device to wait while some higher priority activity occurs, may delay an entire process.
For example, a request for an instruction or data within a hard drive, or even a main, random-access memory, associated with a computer, must negotiate across a main system bus. A central processing unit has a clock operating at one speed. The bus has a controller with a clock that may operate at another speed. The memory device has a memory management unit that may operate at another speed.
Further to the example, a Pentium.TM. processor having a clock speed of 100 megahertz may be connected to peripheral devices or main memory by an industry standard architecture (ISA) bus. The ISA bus has a specified clock speed of 8 megahertz. Thus, any time the Pentium.TM. processor operating at 100 megahertz requests data from the memory device, the request passes to the opposite side of the ISA bus. The data may not be processed or delivered at a speed greater than that of the bus at 8 megahertz. Moreover, a bus typically gives low priority to the central processing unit. In order to avoid underruns and overruns, the input/output devices receive priority over the processor. Thus, the 100 megahertz processor may be "put on hold" by the bus while other peripheral devices have their requests filled.
Any time a processor must access any device beyond its own hardware pins, the hardware interface to the computer outside the processor proper, the required task cannot be accomplished within one clock count of the processor. As a practical matter, a task is not usually completed in less than several clock counts of the processor. Due to other priorities and the speeds of other devices, as well as the need to adjust or obtain the state configurations of interfacing devices, many clock counts of a processor may occur before a task is completed as required.
Associated with every hardware interface between hardware components, elements, and the like (anything outside an individual integrated chip), a hardware handshake must occur for any communication. A handshake, including a request and an acknowledgement, must occur in addition to a transfer of actual data or signals. Handshake protocols may actually involve several, even many, clock counts for the request alone, the acknowledgement alone, and for passing the data itself. Moreover, a transmission may be interrupted by a transaction having a higher priority. Thus, communicating over hardware interfaces is relatively time consuming for any processor. Hardware interfacing may greatly reduce or eliminate the benefits of a high-speed processor.
To alleviate the need to communicate across hardware interfaces during routine processing, modern computer architectures have included caches. In general, processors benefit from maintaining as close to themselves as possible all instructions, data, and clock control. This proximity reduces the need for interfaces, the number of interfaces, the interface complexity, and thus, the time required for compliance with any instruction or necessary execution. Thus, caches have been moved closer and closer to the processor.
Memory caches are common. Such a cache is created within a dedicated portion of a memory device. These are different, however, from caches dedicated to a processor.
The INTEL 386.TM. processor contains an optional external cache connected to the processor through a cache controller chip. The INTEL 486.TM. contains an internal 8 kilobyte cache on the central processing unit itself. Within the chip containing the processor, is integrated a cache. This cache is dedicated to both code and data accesses.
The 486.TM. also supports another cache (a level-2 cache, as opposed to the primary or level-1 cache just described above). Access to the level-2 cache is through an external cache controller chip, similar to that of the 386.TM.. In each case, for both the 386.TM. and 486.TM. processors, the external cache controller is itself positioned on a side of the processor's internal bus (CPU bus) opposite that of the processor.
The Pentium.TM. processors contain a level-1 (primary) data cache as well as a level-1 code cache. Thus, code and data are segregated, cached separately. The Pentium.TM. processors continue to support an external, level-2 cache across a CPU bus.
One should understand that the expression "bus", hereinabove, refers to the processor bus, rather than the system bus. For example, the main system bus connects a processor to the main memory. However, the cache controllers and caches on a processor, or external to the processor but simply located across a processor's internal bus interface unit, do not rely on the main system bus.
A cache has some fixed amount of memory. A code cache will contain certain executable instructions, a data cache will contain data, and a non-segregated cache may contain both. The memory of any type of cache is typically subdivided into cache lines. For example, a typical cache line may contain 32 bytes of information. Thus, a cache line contains a standard number of bytes in which space may be stored a copy of certain information obtained from a main memory device.
Associated with each cache line is a tag. The tag binds a physical address and a logical address corresponding to the contents of an associated cache line.
The physical and logical addresses contained in the tag associated with a cache line may correspond to a physical location in the main memory device, and a logical position within an application respectively.
Caches associated with a processor are transparent, even hidden, with respect to a user and an application. Each cache has an associated controller. In operation, a cache controller effectively "short circuits" a request from a processor to a memory unit. That is, if a particular address is referenced, and that address exists in a tag associated with the contents of a cache line in a cache, the cache controller will fulfill the request for the instruction out of the cache line containing it. The request is thus fulfilled transparently to the processor. However, the effect of a cache is to eliminate, as much as possible, communication through hardware interfaces as described above. Thus, a cache may greatly improve the processing speed of applications running on processors.
Tags may also have associated therewith two numbers referred to as "use bits." The use bits may typically represent a simple count of use. This count may be useful to the cache controller in determining which cache lines are the least recently used. Accordingly, a cache controller may refer to the least recently used (LRU) count to determine which cache lines have been referenced the least number of times.
Incidently, but significantly, with respect to the invention, some cache controllers may churn a cache. That is, if an insignificant number of bits is contained in the LRU or use bits, then a counter may be improperly reset to zero due to count "wrap-around" during high use. Thus, highly-used cache lines may actually be swapped out, churning the cache and dramatically decreasing efficiency.
Several difficulties exist with caches. A cache controller has a general purpose function to service address requests generally. For example, a virtual machine may be implemented in some limited number of instructions. In operating such a virtual machine, a computer processor has an underlying native language in which the virtual machine instructions are written. The virtual machine instructions will be requested repeatedly. The virtual machine instructions are accessed relatively slowly if they are treated simply as another general purpose instruction being retrieved periodically into the cache.
Many processors pipeline instructions. Two problems may occur with pipelining. The first is flushing a pipeline as a result of a branch. The other is stalling due to requested data not arriving within a next clock count in sequence. That is, whenever a cache "miss" occurs, a request has been made to the cache, but the cache cannot respond because the information is not resident. Misses may occur repeatedly over extensive numbers of clock counts while a cache controller accesses a main memory device to load the requested instructions or data. Misses decimate the efficiency of processors. Meanwhile, even with branch prediction methods, a pipeline may flush several instructions with a resulting loss of processing performance.
Brief Summary and Objects of the Invention
In view of the foregoing, it is a primary object of the present invention to provide an accelerator for increasing the execution speed of interpretive environments.
It is another object of the invention to provide programmatic control of persistence of executables stored in a code cache.
It is another object of the invention to provide a set of virtual machine instructions forming a virtual machine, in which each of the compiled or assembled, linked, and loaded native code segments implementing a virtual machine instruction is sized to fit within a standard number of cache lines in a cache.
It is another object of the invention to provide a heuristic determination for pinning the contents of a cache programmatically by a processor.
It is another object of the invention to provide a main memory device containing data structures adaptable to be cached, pinned, and thereafter executed by a processor, without requiring repeated access to the main memory device for retrieval of any instruction.
It is another object of the invention to provide a virtual machine containing an instruction set sized to fit completely within a cache, to be addressable directly by cache line, a unique cache line number corresponding to each machine-native code segment implementing a single virtual machine instruction.
It is another object of the invention to provide a device having programmatic control over the content and persistence of the contents of a cache, particularly a code cache, and more particularly a level-1 code cache, especially a level-1 code cache integrated into a central processing unit.
It is another object of the invention to provide a method to accelerate execution of an interpretive environment by copying instructions of an instruction set into the code cache and pinning those instructions for the duration of the use by the processor of any instructions in the set, in order to increase the speed of processing the virtual machine instructions, eliminate cache misses, optimize pipelining within the processor, while minimizing supporting calculations such as those for addressing and the like.
It is another object of the invention to provide heuristic determination of when to pin a cache, particularly a code cache, based on a cost function of some performance parameter, such as frequency of use, infrequency of use, size, and inconvenience of reloading a particular instruction to be cached.
Consistent with the foregoing objects, and in accordance with the invention as embodied and broadly described herein, an apparatus and method are disclosed in one embodiment of the present invention as including a central processing unit (CPU) having an operably associated processor cache, preferably a level-1 cache. The level-1 cache is closest to the actual processor in the CPU.
The cache may actually be integrated into the CPU. The processor may be programmed to install a full set of virtual machine instructions (VMI) in the cache. The VMI set may then be pinned to eliminate cache "misses" by the processor.
Much interest has been focused over decades on virtual machines. Nevertheless, the slow performance (compared to native code processing) of virtual machines has largely counter-balanced the platform-independent benefits associated therewith.
However, specific knowledge may exist with respect to a particular environment. To take better advantage of interpreted environments generally, such as virtual machines, and cache technology, an apparatus and method in accordance with the invention may rely on this knowledge of the execution environment for a virtual machine in order to optimize the use of both the virtual machine instructions and the cache. Knowing in advance that certain instructions will definitely be required, much faster execution speeds may be obtained by maintaining those instructions in a low level cache as close to the processor as possible.
For example, in one embodiment, an apparatus and method in accordance with the invention may "programmatically control" the contents of the cache. The cache may be loaded with a full set of virtual machine instructions, properly compiled or assembled, linked, and loaded.
The set may incorporate in a length not to exceed a standardized specified number of cache lines, the executable, machine-language implementation of each command or instruction provided in an interpretative environment. The set, fit to the total available cache lines, may define a virtual machine (the entire interpreter). The set may be pinned, after being loaded into a previously evacuated cache.
Loading may be accomplished by running a simple application having no particular meaning, but containing all of the VMIs at least once. Knowing that the cache will respond as designed, one may thus load all of the native code segments implementing the VMIs automatically into the cache in the fastest mode possible, controlled by the cache controller. Yet, the entire process is prompted by programmatic instructions, knowingly applied.
This "programmatic control," in lieu of general purpose control, of a cache, especially a code cache, may completely eliminate cache "misses. " This greatly enhances the effective operating speed of an interpreted or interpretive environment. Effective execution speeds have been increased by the invention to 50 times those of conventional operation of interpreters.
Brief Description of the Drawings
The foregoing and other objects and features of the present invention will become more fully apparent from the following description and appended claims, taken in conjunction with the accompanying drawings. Understanding that these drawings depict only typical embodiments of the invention and are, therefore, not to be considered limiting of its scope, the invention will be described with additional specificity and detail through use of the accompanying drawings in which:
FIG. 1 is a schematic block diagram of an apparatus in accordance with the invention;
FIG. 2 is a schematic block diagram showing implementation details for one embodiment of the apparatus of FIG. 1;
FIG. 3 is a schematic block diagram of executable modules and data structures consistent with one implementation of an apparatus and method in accordance with the invention;
FIG. 4 is a schematic block diagram of a method in accordance with the invention; and
FIG. 5 is a schematic block diagram of registers used in one currently preferred embodiment of the invention for addressing.
Detailed Description of the Preferred Embodiments
It will be readily understood that the components of the present invention, as generally described and illustrated in the Figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following more detailed description of the embodiments of the system and method of the present invention, as represented in FIGS. 1 through 5, is not intended to limit the scope of the invention, as claimed, but it is merely representative of the presently preferred embodiments of the invention.
The presently preferred embodiments of the invention will be best understood by reference to the drawings, wherein like parts are designated by like numerals throughout.
Those of ordinary skill in the art will, of course, appreciate that various modifications to the details illustrated in the schematic diagrams of FIGS. 1-5 may easily be made without departing from the essential characteristics of the invention. Thus, the following description is intended only as an example, and simply illustrates one presently preferred embodiment consistent with the invention as claimed herein.
Referring now to FIGS. 1-3, and more particularly, an apparatus 10 may include a node 11 (client 11, computer 11) containing a processor 12 or CPU 12. The CPU 12 may be operably connected to a memory device 14. A memory device 14 may include one or more devices such as a hard drive or non-volatile storage device 16, a read-only memory 18 (ROM) and a random access (and usually volatile) memory 20 (RAM).
The apparatus 10 may include an input device 22 for receiving inputs from a user or another device. Similarly, an output device 24 may be provided within the node 11, or accessible within the apparatus 10. A network card 26 (interface card) or port 28 may be provided for connecting to outside devices, such as the network 30.
Internally, a bus 32 (system bus 32) may operably interconnect the processor 12, memory devices 14, input devices 22, output devices 24, network card 26 and port 28. The bus 32 may be thought of as a data carrier. As such, the bus 32 may be embodied in numerous configurations. Wire, fiber optic line, wireless electromagnetic communications by visible light, infrared, and radio frequencies may likewise be implemented as appropriate for the bus 32 and the network 30.
Input devices 22 may include one or more physical embodiments. For example, a keyboard 34 may be used for interaction with the user, as may a mouse 36. A touch screen 38, a telephone 39, or simply a telephone line 39, may be used for communication with other devices, with a user, or the like. Similarly, a scanner 40 may be used to receive graphical inputs which may or may not be translated to other character formats. A hard drive 41 or other memory device 14 may be used as an input device whether resident within the node 11 or some other node 52 (e.g., 52a, 52b, etc.) on the network 30, or from another network 50.
Output devices 24 may likewise include one or more physical hardware units. For example, in general, the port 28 may be used to accept inputs and send outputs from the node 11. Nevertheless, a monitor 42 may provide outputs to a user for feedback during a process, or for assisting two-way communication between the processor 12 and a user. A printer 44 or a hard drive 46 may be used for outputting information as output devices 24.
In general, a network 30 to which a node 11 connects may, in turn, be connected through a router 48 to another network 50. In general, two nodes 11, 52 may be on a network 30, adjoining networks 30, 50, or may be separated by multiple routers 48 and multiple networks 50 as individual nodes 11, 52 on an internetwork. The individual nodes 52 (e.g. 52a, 52b, 52c, 52d) may have various communication capabilities.
In certain embodiments, a minimum of logical capability may be available in any node 52. Note that any of the individual nodes 52a-52d may be referred to, as may all together, as a node 52.
A network 30 may include one or more servers 54. Servers may be used to manage, store, communicate, transfer, access, update, and the like, any number of files for a network 30. Typically, a server 54 may be accessed by all nodes 11, 52 on a network 30. Nevertheless, other special functions, including communications, applications, and the like may be implemented by an individual server 54 or multiple servers 54.
In general, a node 11 may need to communicate over a network 30 with a server 54, a router 48, or nodes 52. Similarly, a node 11 may need to communicate over another network (50) in an internetwork connection with some remote node 52. Likewise, individual components 12-46 may need to communicate data with one another. A communication link may exist, in general, between any pair of devices.
Referring now to FIG. 2, a processor 12 may include several internal elements. Connected to the bus 32, a bus interface unit 56 handles the bus protocols enabling the processor 12 to communicate to other devices over the bus 32. For example, the instructions or data received from a ROM 18 or data read from or written to the RAM 20 may pass through the bus interface unit 56.
In some processors, a processor cache (e.g. cache 58, 64), such as a level-1 cache 58 may be integrated into the processor 12. In specific embodiments of processors 12, such as the Pentium.TM. and Pentium.TM. Pro processors, as well as the Power PC.TM. by Motorola, the level-1 cache 58 may be optionally subdivided into an instruction cache 60 and a data cache 62. level-1 cache 58 is not required in a processor 12. Moreover, segregation of the instruction cache 60 from the data cache 62 is not required. However, a level-1 cache 58 provides rapid access to instructions and data without resort to the main memory 18, 20 (RAM 20). Thus, the processor 12 need not access (cross) the bus interface unit 56 to obtain cached instructions and data.
Certain processors 12 maintain an external cache 64. The external cache 64 is identified as a level-2 cache in FIG. 2. Nevertheless, the level-2 cache 64 may be a level-1 cache if no level-1 cache 58 is present on the processor 12 directly. Similarly, the external cache 64 may or may not be segregated between an instruction cache 66 and a data cache 68. Any suitable processor cache may be used.
Execution, normally associated with a processor 12, is actually most closely related to a fetch/decode unit 70, an execute unit 72, and a writeback unit 74. Likewise, associated with each cache 58, 64, is typically an inherent, integrated, hardware controller. The cache controller may be thought of as control logic built into the cache hardware.
When the fetch unit 71a issues a request for an instruction, the request goes to the bus interface unit 56. The level-1 cache 58 makes a determination whether or not the request can be satisfied by data or instructions identified with the logical address requested from cached data and instructions.
If an instruction cannot be provided by the level-1 cache 58, the level-2 cache 64 may respond to the request. If the desired item (data or instruction) is not present in either the level-1 cache 58 or the level-2 cache 64, then the main memory 18, 20 may respond with the desired item. Once the request has been fulfilled by the fastest unit 58, 64, 20, 18 to respond with the desired item, the request is completed, and no other devices will respond.
Main memory may include the ROM 18, the RAM 20, or both. Nevertheless, many computers boot up using the contents of the ROM 18 and thereafter use the RAM 20 for temporary storage of data associated with applications and the operating system. Whenever "main memory" is mentioned, it is contemplated that it may include any combination of the ROM 18 and RAM 20.
Once an instruction is retrieved for the fetch unit 71a, the instruction is passed to the decode unit 71b. The fetch 71a and decode 71b are typically highly integrated, and perform in an overlapped fashion. Accordingly, a fetch/decode unit 70 is typical.
As a practical matter, the decode unit 71b may identify a current instruction to be executed. Identification may involve identification of what type of instruction, what type of addressing, what registers will be involved, and the like. The presence of the instruction in an instruction register, may itself stimulate execution on the next clock count.
Once identification of an instruction is completed by the decode unit 71b, an execute unit 72 may immediately process the instruction through low-level, control-loop hardware. For example, sequencers, registers, and arithmetic logic units may be included in an execute unit 72.
Each instruction as it is fetched, decoded, executed, and the like, may require interaction between an individual processing unit 70, 72, 74 and a register pool 76. The registers 76 (register pool 76) are hidden from programmers and applications. Nevertheless, the hardware architecture of the processor 12 provides a hardware logic governing interaction between the units 70, 72, 74 and between the registers 76 and the units, 70, 72, 74.
Upon completion of execution of an instruction, a writeback unit 74 may provide an output. Accordingly, the output may be passed to the bus interface unit 56 to be stored as appropriate. As a practical matter, a result may be stored in a cache 58 of a level-1 variety or in a level-2 cache 64. In either event, a writeback unit 74 will typically write through to the main memory 18, 20 an image of the result.
Modern processors 12, particularly the Pentium.TM. processors, use a technique called pipelining. Pipelining passes an instruction through each of the fetch/decode/execute steps undergone by that instruction as quickly as possible. An individual instruction is not passed completely through all of its processing steps before the next instruction in order is begun.
For example, a first instruction may be fetched, and on the next clock count another instruction may be fetched while the first instruction is being decoded. Thus, a certain parallel, although slightly offset in sequence, processing occurs for instructions.
An advantage of a method and apparatus in accordance with the invention is that more instructions may be pipelined. That is, prediction routines have been built into hardware in the Pentium.TM. class of processors 12. However, prediction is problematic. Inasmuch as a branch may occur, within approximately every five machine code instructions on average, the pipeline of instructions will be in error periodically. Depending on the sophistication of a prediction methodology, one or more instructions in a pipeline may be flushed after entering a pipeline at the fetch unit 71a.
Referring now to FIG. 3, a virtual machine 90 or an instruction set 90 implementing a virtual machine 90 on a processor 12 is illustrated schematically. Relationships are illustrated for caching 80 or a cache system 80 for storing loaded and executable instructions 106 (e.g. 106a) corresponding to virtual machine instructions 91 (e.g. 91a) of a virtual machine 90 or virtual machine instruction set 90.
A virtual machine 90 may be built upon any available programming environment. Such virtual machines 90 may sometimes be referred to as interpreters, or interpreted systems. Alternatively, virtual machines 90 are sometimes referred to as emulators, wherein a set of instructions 91a-n may be hosted on a processor 12 of one type to mimic or emulate the functional characteristics of a processor 12 in a hardware device of any other type.
An application may be written to run on or in an environment created for a first hardware device. After the application is fully developed and operational, the application may then be "ported" to another machine. Porting may simply include writing a virtual machine 90 for the second hardware platform. Alternatively, an application may be developed in the native language of a first machine, and a single set 90 of virtual machine instructions 91a-n may be created to emulate the first machine on a second machine. A virtual machine 90 is sometimes referred to as an emulation layer. Thus, an emulation layer or virtual machine 90 may provide an environment so that an application may be platform-independent. A JAVA interpreter, for example, performs such a function.
An executable 82 loaded into main memory 18, 20 contains the original images of the contents of the cache system 80. A building system 84 that may be thought of as an apparatus, modules running on an apparatus, or a system of steps to be performed by an apparatus, is responsible to build contents to be loaded into the executable 82.
A builder 86 may be tasked with building and loading an executable image 100 of a virtual machine 90. Similarly, a builder 88 may build an executable image 130 of the instructions 106 implementing an application written in the virtual machine instructions 91 constituting the virtual machine 90. In general, the executable 130 or executable image 130 may represent any application ready to be executed by the execute unit 72 of the processor 12. One embodiment of an executable 130 or an image 130 may be an application written specifically to prompt a high speed loading as described with respect to FIG. 4 below.
A virtual machine 90 or a set 90 of virtual machine instructions 91a-n may contain an individual instruction (e.g. 91a, 91b, 91n) corresponding to each specific, unique function that must be accommodated by the virtual machine 90. The virtual machine instruction 91n, for example, provides the ability to terminate execution.
In FIG. 3, the builder 86 may include source code 90, virtual machine source code 90. The source code 90 may be assembled or compiled by an assembler 92 or compiler 92, as appropriate. The virtual machine may operate adequately, whether dependent on assembly or compilation. The assembler 92 or compiler 92 operates for native code. Native code, may be thought of as code executable directly on a processor 12 in the apparatus 10.
By native code is indicated the processor-specific instructions 91 that may be executed directly by a processor 12. By directly is not necessarily meant that the native code is always written in binary ones and zeros. Native code 106 may be written in a language to be assembled 92 or compiled 92 into object code 94 and to be eventually linked 96 into an executable 100 loaded for execution. Executables 100 may then be loaded 99 into a memory device 20, 18 for ready execution on or by an execute unit 72 of a processor 12. An executable 100 stored in a non-volatile storage device 16 may sometimes be referred to as an executable file. Once properly loaded 99 into the main memory 18, 20 associated with a processor 12 an executable 100 may be executed by a processor 12.
The assembler 92 or compiler 92 provides object code 94 in native code instructions. The object code 94 may be linked to library routines or the like by a linker 96. The linker 96 may provide all other supporting instructions necessary to run the object code 94. Thus, the linker 96 provides, as output, executable code 98. As a practical matter, the executable code 98 will be run directly from main memory 18, 20 as a loaded executable 100. Thus, a loader 99 may load the executable code 98 into main memory 18, 20 as the loaded code 100.
Code segments 106a-n are written in native code. When any code segment 106a-n (e.g. 106a, 106b, 106c, 106n) is executed, the result is the desired output from the corresponding virtual machine instruction 91a-n (e.g. 91a, 91b, 91c, 91n, respectively). Virtual machine instructions 91a-n identify every available function that may be performed by the virtual machine 90. The instructions 106a-n illustrate segments 106a-n, implementations in native code, executably the hardware, processor 12, that must produce the result associated with each individual virtual machine instruction 91a-n.
Each of the code segments 106a-n contains a FETCH instruction 108 DECODE instruction 110 and JUMP instruction 112. The instructions 108-112 promote pipelining. Thus, the subject of each of the respective instructions decode 110, fetch 108, and JUMP 112 correspond to the very next instruction, the second next instruction, and the third next instruction, respectively, following an instruction 91a-n being executed and corresponding to a code segment 106a-n in question.
A virtual machine instruction set 90 should include a HALT instruction 91n. Thus, a virtual machine instruction 91n within the virtual machine 90 will contain a segment 106n of native code indicating to the processor 12 the fetching and decoding process for instructions used in all applications. The last virtual machine instruction 91a-n contained within a loaded application 130 is a HALT instruction 91n (106n).
In FIG. 3, the loaded executable 100 may be stored in a block 114 separated by block boundaries 116. In the Pentium.TM. class of processors, each block 114 contains 32 bytes of data. The instruction set 90 or virtual machine 90 contains no more than 256 virtual machine instructions 91a-n. Accordingly, the code segments 106a-n, when compiled, linked, and loaded, may each be loaded by the loader 99 to begin at a block boundary 116, in one currently preferred embodiment. Thus, the number of blocks 114 and the size of each block 114 may be configured to correspond to a cache line 140 in the cache 60. Thus, an image of a code segment 106a-n, compiled, linked, and loaded for each virtual machine instruction 91a-n, exists in a single cache line 140. Likewise, every such virtual machine instruction 91a-n and its native code segment 106a-n has an addressable, tagged, cache line 140 available in the 256 cache lines.
In addition to the builder 86, a builder 88 may build any virtual machine application 120. In FIG. 3, the process of building an application 120 is illustrated. For example, a mock application may be constructed for the exclusive purposes of high-speed loading of the code segments 106 into the cache lines 140. In the embodiment shown, virtual machine source language code 120 or source code 120 may be written to contain instructions 91 arranged in any particular order. In general, instructions 91 are used by a programmer in any suitable order to provide and execute an application 120.
In an embodiment of an apparatus and method in accordance with the invention, the source code 120 may simply contain each of the virtual machine instructions 91 in the virtual machine language. The source code 120 may be assembled or compiled by an assembler 122 or compiler 122 depending on whether the language is an assembled or a compiled language.
The assembler 122 or compiler 122 generates (emits, outputs) virtual machine code. The output of the assembler 122 or compiler 122 is object code 124. The object code 124 may be linked by a linker 126 to produce an executable code 128. The executable code 128 may be loaded by a loader 129 into main memory 18, 20 as the loaded executable 130.
The loaded executable 130 is still loaded virtual machine code. Thus, an application developed in the virtual machine language must be run on a virtual machine. The virtual machine 90 is stored in the cache 60. The cache 60 may actually be thought of as any processor cache, but the closest cache to a processor 12, is capable of the fastest performance.
The loaded executable 130 is comprised of assembled or compiled, linked, and loaded, virtual machine instructions 132. A main memory device 20 is byte addressable. Each of the virtual machine instructions 132 begins at an address 134. Thus, each virtual machine instruction 132 may be of any suitable length required. Nevertheless, a virtual machine address zero 135 may be identified by a pointer as the zero position in the virtual machine 130. Each subsequent address 134 may thus be identified as an offset from the virtual machine zero 135. A last instruction 136 should be effective to provide an exit from the loaded executable 130. Typically, loaded executables 130 are executed in the order they are stored in the memory device 20.
The cache 60 has associated therewith a tag table 142. For each cache line 140, an appropriate tag line 144 exists (e.g. 144a, 144b, 144c). Associated with each tag line 144, is a logical address 146 corresponding to the address 134 of the cache line 140 in question. Likewise, a physical address 148 in a tag line 144 corresponds to an address 116 or block boundary 116 at which the code 114 is stored in the main memory 18, 20. A control field 144c may contain symbols or parameters identifying access rights, and the like for each cache line 140.
Thus, in general, a loaded executable 130 (application 130) has a logical address 134 associated with each virtual machine instruction 132. The logical address 134 associated with the beginning of an instruction 132 is bound by the tag table 142 to the physical address 116 associated with the executable code 100 associated with the corresponding code segment 106 whose compiled, linked, and loaded image is stored at the respective cache line 140 associated with the tag line 144 binding the logical address 134, 146 to the physical address 116, 148.
Referring to FIG. 4, a method 160 is described and illustrated schematically. The method 160 locks or pins a cache after loading the native code implementation of individual virtual machine instructions into the cache.
A disable 162 may be executed by the processor to disable interrupts from being serviced. The disable 162 provides temporary isolation for the cache 60, enabling completion of the process 160 or method 160. The cache 60 is next flushed 164 typically with write-back, which causes "dirty" cache data to be written back to main memory 18, 20. Thus, in the control field 150 may be a byte indicating that each cache line 140 is available. Thus, the processor 12 need not thereafter execute the multiple steps to remove the contents of any cache line 140 in preparation for loading new contents.
The execute steps 166 correspond to execution by the processor 12 of individual instructions 132 in a loaded application 130. Upon fetching for execution 166 each instruction 132, the processor 12 places a request for the instruction 132 next in order in the loaded application 130.
The cache controller for the cache 60 first reviews the contents of the tag table 142 to determine whether or not the desired instruction is present in the cache 60. Having been flushed, the cache 60 has no instructions initially. Accordingly, with each execute 166, a new instruction 132 is loaded from the main memory 18, 20 into the cache 60 at some appropriate cache line 140. Immediately after loading into the cache 60, each instruction 132 in order is executed by the processor 12. However, at this point, any output is ignored. The execution 166 is simply a by-product of "fooling" the cache into loading all the instructions 132 as rapidly as possible, as pre-programmed into the hardware.
In one embodiment of an apparatus and method in accordance with the invention, a loaded application 130 contains every instruction 132 required to form a complete set of instructions for a virtual machine. The instructions 132 are actually code segments 106 implementing a virtual machine instruction 91 in the native code of the processor 12. No output is needed from the initial application 130 run during the method 160.
In one currently preferred embodiment of an apparatus and method in accordance with the invention, the virtual machine instruction set 100 is written so that each block 114 contains a single instruction 91. Moreover, the instruction set 90 is written to occupy exactly the number of cache lines 140 available in the cache 60.
In certain embodiments, an individual instruction 91 may occupy more than a single cache line 140. For example, some caches may have a 16 byte line length. Thus, a 32 byte length for an instruction 91 may require two cache lines 140. In one presently preferred embodiment, a number of cache lines 140 may correspond exactly to the number of blocks 114 required to hold all of the instructions 91, such that each instruction 91 may be addressed by referring to a unique cache line 140.
Thus, upon completion of execution of an initial application 130 configured for loading the cache 60, no output may be provided. However, the cache 60 with its controller operating normally, loads every instruction 91 referenced by the application 130. Therefore, each cache line 140 contains a code segment 106 or native code segment 106 implementing a virtual machine instruction 91. Each cache line 140 contains the code segment 106 corresponding to a virtual machine instruction 91 in a cache 60 having a line length of 32 bytes.
After the executions 166 of the virtual machine instructions 132 of the application 130 designed for the loading of virtual machine instruction code 106 into the cache 60, a disable 168 may disable the cache 60. The effect of the disable 168 is to pin the contents of each cache line 140. Pinning (locking) indicates that the cache controller is disabled from replacing the contents of any cache line 140.
Nevertheless, the cache 60 continues to operate normally, otherwise. Thus, the controller of the cache 60 will continue to refer to the tag table 142 to determine whether or not an address 146, 148 requested is present. In the case of a virtual machine 90, every instruction 91 will be present in the cache 60, if designed optimally in accordance with the invention. Thus, the tag table 142 will always contain the code 106 associated with any address 146, 148 representing any virtual machine instruction 91.
Less than a full set of instructions 91 may be loaded into a cache 60. Alternatively, for a cache 60 having more cache lines 140 than needed for storing a virtual machine 90 in its entirety, unused cache lines 140 may be devoted to other code, loaded in a similar way, prior to pinning. Code may be selected according to recency of use, cost/benefit analysis of use, or cost/benefit analysis of retrieval from main memory 18, 20.
The cache 60 is used by way of example. The virtual machine 90 will operate fastest by using the cache 60 closest to the fetch/decode unit 70. Alternatively, another cache 64 may be used. Thus, everything describing the cache 60 may be applied to the cache 66 or the cache 64 so far as loading and pinning of the cache 60 are concerned. The enable 170 may re-enable the interrupts so that the processor 12 may resume normal operations.
Referring to FIG. 5, an efficient fetch/decode/JUMP algorithm may begin with an XOR of the contents of a register EAX 180 against itself. The effect of the XOR is to zero out the contents of the EAX register 180. The contents of register EAX 180 may represent a pointer. Following this clearing operation, a MOVE instruction (MOV) may move the contents of a memory location corresponding to a pointer (next logical instruction number) and identified by the label or logical instruction number stored in a register EBX 190 into the register AL 186.
The register AL 186 is the lower eight bits of the AX register 182. The AX register 182 is the lower 16 bits of a 32 bit EAX register 180. The upper eight bits of the AX register 182 constitute the AH register 184. The AL 186 or lower register 186 thus receives the contents of a memory location corresponding to a current instruction 91 being pointed at by the contents of the EBX 190 register.
Following the MOVE instruction, a SHIFT instruction may shift left by five bits (effectively a multiplication by a value of 32) the contents of the EAX register 180. Since the EAX register 180 was zeroed out, and only the AL register was filled, a shift left of the AL register 186 multiplies its value by 32. This shift left is effectively a decoding of the instruction that was fetched by the MOVE instruction.
Continuing with the procedure, a JUMP instruction may be implemented to position EAX in the set of virtual machine instructions. Note that each virtual machine instruction 91 in the complete set 90, when loaded, is written within the same number of bytes (32 bytes for the native code segment implementing the virtual machine instruction). The code segment 106 for each instruction 91 begins at a block boundary 116 and at the beginning of a cache line 140. Thus, a virtual machine instruction number multiplied by 32 will step through each of the native code segments 106. Thus, a JUMP to EAX constitutes a direct addressing of the native code segment 106 required to implement a particular virtual machine instruction 91.
Other mechanisms exist to address memory 20. For example, vector tables are commonly used. However, such mechanisms require certain calculations to occur in order to execute a JUMP. Moreover, memory access is required in order to complete the determination of a value in a vector table. Thus, the processor 12 must request access to the main memory 18, 20 in order to fulfill the request for a vector table entry. Accessing main memory and other operations requiring requests to be managed by the bus 32 may increase access times by more than orders of magnitude. The simple arithmetic logic unit operation of a JUMP in the preferred embodiment, is much more efficient than the vector table approach that imposes a memory reference on top of a simple JUMP ALU operation.
Different types of caching implementations may exist in hardware. Three common types of cache architectures are direct-mapped, fully-associative, and a set-associative. Cache technology is described in detail in Computer Architecture: A Quantitative Approach by John L. Hennessy and David A. Patterson published in 1990 by Morgan Kaufman Publishers, Inc. of San Mateo, Calif. (See Chapter 8).
In an apparatus and method in accordance with the invention, any type of cache 60 may be used. In one currently preferred embodiment, a two-way set associative cache 60 may be used.
In a direct-mapped cache 60, several blocks or lines 140 exist. A cache line 140 may contain some selected number of bytes, as determined by the hardware. Typical cache lines 140 have a length of 16 or 32 bytes. Likewise, each cache structure will have some number of addressable lines. An eight bit addressing scheme provides 256 cache lines in a cache.
Each byte of memory within a memory device 14, including read/write types as well as read-only types, especially a main random access memory device 20, is directly addressable. One common caching scheme for a direct mapped cache architecture may map a memory device 20 to cache lines 140 by block. The memory's addressable space may be subdivided into blocks, each of the same size as a cache line. For example, an entire random access memory 20 may be subdivided into 32-byte blocks for potential caching.
A significant feature of a direct-mapped cache is that every block of memory within the source memory device 20 has a specific cache line 140 to which it will be cached any time it is cached. In one scheme, the least significant bits in an address corresponding to a block within a memory device may be truncated to the same size as the address of a cache line 140. Thus, every block of memory 20 is assigned to a cache line 140 having the same least significant bit address.
In a fully-associative, caching architecture, no binding need exist between any particular block of memory in the memory device, and any cache line a priori. Allocation of a cache line 140 space to a particular block of memory 20 is made as needed according to some addressing scheme. Typical schemes may include random replacement. That is, a particular cache line 140 may simply be selected at random to receive an incoming block to be cached.
Alternative schemes may include a least-recently-used (LRU) algorithm. In a least-recently-used (LRU) scheme, a count of accesses may be maintained in association with each cache line 140. The cache line 140 that has been least recently accessed by the processor 12 may be selected to have its contents replaced by a incoming block from the memory device 20.
A set-associative architecture subdivides an associative cache into some number of associative caches. For example, all the lines 140 of a cache 60 may typically be divided into groups of two, four, eight, or sixteen, called "ways." Referring to the number of these ways or subcaches within the overall cache 60, as n, this subdivision has created an n-way set-associative cache 60.
Mapping of block-frame addresses from a main memory device 20 to a cache line 140 uses the associative principle. That is, each way includes an nth fraction of all the available cache lines 140 from the overall cache 60. Each block from the main memory device 20 is mapped to one of the ways. However, that block may actually be sent to any of the cache lines 140 within an individual way according to some available scheme. Either the LRU or the random method may be used to place a block into an individual cache line 140 within a way.
For example, a main memory address may be mapped to a way by a MODULO operation on the main memory address by the number of ways. The MODULO result then provides the number of a "way" to which the memory block may be allocated. An allocation algorithm may then allocate the memory block to a particular cache line 140 within an individual way.
Another cache may be used, with less effective results. Loading and pinning may also be done using test instructions, although more time-consuming. Instead of test instructions, the proposed method flushes the cache, running a simple application 130 containing every VMI 91 of a desired set 90 to be loaded. Before disabling the cache 60, the method may use the cache's internal programming, built into the fundamental hardware architecture, to provide a high-speed load. Disabling permits access to the cache 60, but not replacement, completing an effective pinning operation.
In one currently preferred embodiment, the closest cache to the processor is used. For example, in the pentium.TM. processor, the level-1 code cache may be used. In other embodiments, an external cache, or a level-1 integrated (not segregated between code and data) cache may be used.
Pinning is particularly advantageous once an environment, or rather the executable instructions constituting an environment, have been programmed in a form that fits the entire instruction set into an individual cache, with one instruction corresponding to one cache line. Benefits derived from this method of architecting and pinning the virtual machine are several.
For example, no cache line 140, during execution of a virtual machine 90, need ever be reloaded from main memory 18, 20. In addition to the time delay associated with having to access the bus 32, access times within memory devices 14 themselves vary. Typically, a cache access time is an order of magnitude less than the access time for a main memory location. Reloading a cache line 140 is likewise a time-consuming operation.
Here, every branch destination (the object of a JUMP) within the virtual machine 90 may be located at a fixed cache line position. Thus, no penalty is created for address generation within the cache 60 itself. Rather, each cache line 140 may be addressed directly as the address of the instruction 91 being requested.
That is, typically, a cache controller must manage an addressing algorithm that first searches for a requested reference within the cache. If the reference is not present, then the cache controller requests over the bus 32 from main memory the reference. The address generation, management, and accessing functions of the cache controller are dramatically simplified since every desired address is known to be in the cache for all code references.
Many modern processors such as the pentium.TM. series by INTEL contain hardware supporting branch prediction. That is, when a branch operation is to be executed, the processor predicts the destination (destination of a JUMP) to which the branch will transfer control. With a pinned cache containing the entire instruction set 90 of the virtual machine 90, all branch destinations are known. Every instruction has a cache line 140 associated therewith which will never vary. Not only does this correspondence not vary within a single execution of the virtual machine, but may actually be permanent for all loadings of the virtual machine.
Likewise, a branch prediction table is typically updated along with cache line replacement operations. Since the cache lines 140 need never be replaced while the virtual machine is loaded into the cache, and pinned, the branch prediction table becomes static. Inasmuch as the prediction table becomes static, its entries do not change. Moreover, every referenced code instruction is guaranteed to be in the cache. Therefore, any benefits available to a branch prediction algorithm are virtually guaranteed for an apparatus and method operating in accordance with the invention. Flushes of the pipelined instructions now approach a theoretical minimum.
In the pentium.TM. processor by INTEL, two arithmetic logic units (ALUs) correspond to a `U` pipeline and a `V` pipeline. Each arithmetic logic unit (ALU) may execute an instruction with each clock count. However, if two instructions must occur in sequence, then one pipeline may be idled. Thus, the `V` pipeline may be idled during any clock count that requires two instructions to be executed in sequence rather than in parallel.
Typical optimal programming on pentium.TM. processors may achieve 17 to 20 percent pairing between instructions. By pairing is meant that instructions are being executed in both the `U` and `V` pipelines. Here that occurs about 17 to 20 percent of the time in a pentium.TM. processor.
Due to the careful architecture of the instruction set, as well as pinning the instruction set, a method and apparatus in accordance with the invention may routinely obtain 60 percent utilization of the `V` (secondary) pipeline. The selection and ordering of the virtual machine instructions have been implemented to optimize pairing of instructions through the pipelines.
The present invention may be embodied in other specific forms without departing from its spirit or essential characteristics. The described embodiments are to be considered in all respects only as illustrative, and not restrictive. The scope of the invention is, therefore, indicated by the appended claims, rather than by the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.
Claims
What is claimed and desired to be secured by United States Letters Patent is:
1. An apparatus for accelerating execution in an interpretive environment, the apparatus comprising: a processor for executing executables; a memory device operably connected to the processor and containing the executables to be executed by the processor; a processor cache operably connected to the memory device to receive and persistently store the executables, the processor programmed to control the persistence of the executables within the processor cache by pinning the executables within the processor cache.
2. The apparatus of claim 1, wherein the processor cache and processor are contained within a central processing unit.
3. The apparatus of claim 1, wherein the processor is programmed to control the persistence of the executables in the processor cache.
4. The apparatus of claim 3, wherein the memory device contains a pinning executable effective to program the processor to pin the processor cache to control the persistence of the executables.
5. The apparatus of claim 4, wherein the executables are interpreter executables effective to provide an interpretive environment.
6. The apparatus of claim 5, wherein the interpretive environment provides a software virtual machine.
7. The apparatus of claim 3, wherein the processor is programmed to determine heuristically the contents pinned in the processor cache.
8. The apparatus of claim 7, wherein the processor is programmed with a heuristic based on a frequency of use of the executables.
9. The apparatus of claim 8, wherein the processor allocates a priority of executables for pinning in the processor cache in accordance with the frequency of use of each executable.
10. The apparatus of claim 1, wherein the executables are effective to provide an interpreter executed by the processor.
11. A memory device containing data structures for supporting accelerated execution of applications in an interpretive environment, the applications being executed by a processor having a processor cache associated therewith and the processor being operably connected to the memory device, the data structures comprising: an interpreter comprising instructions corresponding to a virtual machine effective to be executed by the processor; a pinning executable loadable into the processor and effective to control persistance of the instructions in the processor cache through pinning the instructions in the processor cache.
12. The memory device of claim 11, wherein the cached instructions comprise a complete instruction set constituting the virtual machine.
13. The memory device of claim 11, wherein the cached instructions are selected from a complete instruction set constituting the virtual machine.
14. The memory device of claim 13, wherein the pinning executable further comprises a selection executable effective to select the cached instructions from the complete instruction set.
15. The memory device of claim 14, wherein an instruction is selected, to be included in the cached instructions, in accordance with a cost function of a performance parameter.
16. The memory device of claim 15, wherein the performance parameter includes a frequency of execution of the instruction.
17. The memory device of claim 15, wherein the performance parameter includes a size corresponding to the instruction.
18. The memory device of claim 11, wherein the processor cache and the processor are both contained within a central processing unit.
19. The memory device of claim 11, wherein the processor is further programmed to control persistence of the instructions in the processor cache.
20. The memory device of claim 19, wherein the data structures further include a heuristic effective to determine which instructions are included as contents pinned in the processor cache.
21. The memory device of claim 11, wherein the processor cache is a level-1 processor cache.
22. A method to accelerate execution in an interpretive environment, the method comprising: providing a memory device storing an interpreter comprising instructions and storing a pinning executable loadable by a processor; providing a processor operably connected to the memory device to load the pinning executable and execute the instructions; providing a processor cache operably associated with the processor; loading the instructions to the processor cache; pinning of the instructions in the processor cache by the pinning executable to control the persistence thereof.
23. The method of claim 22, wherein the interpreter constitutes a virtual machine.
24. The method of claim 22, wherein the instructions comprise a complete instruction set corresponding to the entire interpreter.
25. The method of claim 22 further comprising copying other instructions that are not part of the interpreter, to the processor cache before pinning the instructions.
26. The method of claim 22, wherein the processor cache is a level-1 processor cache.
27. The method of claim 22 further comprising controlling by the processor of the pinning of the processor cache.
28. The method of claim 27 further comprising heuristically determining how to control the pinning of the processor cache.
29. The method of claim 28, wherein heuristically determining comprises selecting the instructions to place in the processor cache in accordance with a cost function of a performance parameter.
30. The method of claim 29, wherein the performance parameter reflects a frequency of use.
31. The method of claim 22 wherein copying further comprises running by the processor an application containing every instruction of a virtual machine instruction set, and wherein pinning further comprises disabling the cache controller against replacing the contents of cache lines in the cache.
Patent Citations (10)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US4277826 | 1981-07-01 | Collins et al. | |
| US4583166 | 1986-04-01 | Hartung et al. | |
| US4811215 | 1989-03-01 | Smith | |
| US5023776 | 1991-06-01 | Gregor | |
| US5202993 | 1993-04-01 | Tarsy et al. | |
| US5249286 | 1993-09-01 | Alpert et al. | |
| US5371872 | 1994-12-01 | Larsen et al. | |
| US5471591 | 1995-11-01 | Edmondson et al. | |
| US5652889 | 1997-07-01 | Sites | |
| US5781792 | 1998-07-01 | Asghar et al. |
Non-Patent Literature (6)
- TDB 0795.0013 "Instruction Cache Block Touch Retro-Fitted onto Microprocessor", IBM Technical Disclosure Bulletin, Vo. 38 No. 07 Jul. 1995, pp. 53-56.
- TDB 0890.0027 "Preemptible Cache Line Prefetch Algorithm and Implementation", IBM Technical Disclosure Bulletin, vol. 33 No. 3B Aug. 1990, pp. 371-373.
- TDB 1294.0035 "Dual On-Chip Instruction Cache Organization in High Speed Processors", IBM Technical Disclosure Bulletin, vol. 37 No. 12 Dec. 1994, pp. 213-214.
- HP600123.0108, "Architecture of the Series 700 Bus", Hewlett Packard manual excerpt, SPI Database of Software Technologies, .COPYRGT.1995 Software Patent Institute, three pages.
- rubin 001.0001, "An Experience Teaching a Graduate Course in Cryptography and Abstract and 1 Introduction", 1996 Aviel D. Rubin excerpt, SPI Database of Software Technologies, .COPYRGT.1996 Software Patent Institute, two pages.
- rubin 001.0018, "An Experience Teaching a Graduate Course in Cryptography, pp. 7 to 10", 1996 Aviel D. Rubin excerpt, SPI Database of Software Technologies, .COPYRGT.1996 Software Patent Institute, two pages.