Method for Compiling and Executing Programs for Reprogrammable Instruction Set Accelerator

The present application is related to co-pending U.S. patent application entitled REPROGRAMMABLE INSTRUCTION SET ACCELERATOR USING A PLURALITY OF PROGRAMMABLE EXECUTION UNITS AND AN INSTRUCTION PAGE TABLE, invented by Stephen M. Trimberger, filed on the same day as the present application (application Ser. No. 080/483,240).

1. Field of the Invention

The present invention relates to techniques to improve the speed of microprocessors using reprogrammable hardware; and more particularly to the use of reprogrammable execution units in parallel with predefined execution units in a data processing system.

2. Description of Related Art

General purpose computers based on current microprocessors have a single predefined instruction set. The instruction set is devised to improve the speed of a large number of typical applications, given a limited amount of logic with which to implement the instructions. General purpose processors include so called complex instruction set computers (CISC) which have sophisticated instruction sets designed for commercial settings, and reduced instruction set computers (RISC), where the basic instruction set is kept to a minimum that gives good performance over a broad range of applications.

In the interest of good overall performance for a given amount of logic, general purpose processors using both CISC and RISC approaches leave off instructions that may be beneficial for some problems. Therefore, the left off instructions must be replaced with a sequence of the predefined instructions, so that these special problems take longer to solve. Some types of instructions that are commonly left off general purpose processors include floating point arithmetic, graphics manipulations, and bit field extraction used in encryption, data compression and image processing.

If the need for a given instruction is great enough in some applications, some users may benefit from a special purpose instruction set accelerator. The instruction set accelerator intercepts the instructions and interprets them in place of the general purpose processor. This has been done with floating point arithmetic (for example, the Intel 387 class of floating point processors) and for graphics operations. This solution is only cost effective if many or most computer users need the additional speed for those operations so that the cost of developing the special purpose hardware accelerator for a specific instruction is shared. Although most computer users can use fast graphics, very few need, for example, fast encryption. Thus, special purpose instruction set accelerators may not be developed for the encryption algorithm, even though great improvements in performance could be achieved.

Special purpose processors have been built to provide very high speed solutions to compute intensive problems, such as encryption and image processing. These processors replace not a single instruction, but whole programs. Because only a few people need these special processors, they are expensive but provide a huge improvement in performance. Instructions to a special purpose processor are typically sent by commands from the host general purpose processor, however, not as instructions that the special purpose processor intercepts. Thus, special interface software is required to access the special purpose processor.

Computer users are also faced with the prospect of numerous instruction set accelerators and special purpose processors in their computers, one for each special application they may do. This adds to size, weight and expense of computers. Further, even the commonly used special operations are not needed always, so most of the hardware accelerators will be unused at any one time.

An alternative technique for reconfiguring a general purpose processor involves a use of writable microstores. One method of building a general purpose processor implements instructions by emulating them with microcode. Microcode comprises instructions that control flow through a set of functional units in the microprocessor. Each instruction on the general purpose processor is emulated by several microinstructions. The general purpose processor has a microcontroller that reads the microcode from the microstore, and uses the instruction value to determine where in the microstore to execute and what to do to perform the logic for the instruction.

Typically, a manufacturer stores microcode in read only memory. However, some microprogrammed computers have been built with a writable microstore. In these machines, a user can write a program that emulates a new instruction set. However, these systems require an embedded microcontroller, and deal with a fixed set of functional units. Also, the microinstruction fetching technique divides each instruction into a number of instructions, rather than replacing a slow instruction with a fast one. Thus, these systems have limited use for improving performance of systems that need special purpose instructions.

One prior art approach to improving performance of general purpose processors involves the use of field programmable gate array (FPGA) logic configured as a co-processor attached to the same bus as the host processor. See for example U.S. Pat. No. 5,361,373 to Gilson. This approach involves capturing entire sub-routines, detecting when the host CPU enters the captured sub-routine, and then taking over execution of the programmed function in the FPGA hardware. When the function completes, the FPGA returns control to the host CPU itself. However, this approach requires complex coordination with the host CPU, including maintaining CPU state and the like while the field programmable co-processor executes the sub-routine. The cost of overhead of the process, such as maintaining and restoring CPU state via dump and un-dump operations on the CPU, limits application of the field programmable co-processor to relatively complex sub-routines.

Such reprogrammable hardware accelerators, like dedicated special purpose processors before them, are targeted at huge speed improvements in large scale operations. Therefore, their applicability is limited. They tend to be large and complicated, and only help with a limited number of special problems. Further, interfacing to such devices is non-standard, because they do not interpret instructions on the microprocessor bus.

Accordingly, it is desirable to provide a technique for using reprogrammable logic to accelerate special instructions for a general purpose processor which is practical to implement, and provides a significant performance improvement over prior art systems. This will provide the ability to a user to reprogram the host processor such that it includes an instruction set based on defined and programmed instructions optimized for the users particular applications.

In addition, it is desirable to provide tools for generating computer programs using programmed instructions, and to develop programmed instructions for specific applications.

The present invention provides a technique providing a reprogrammable instruction set accelerator (RISA). The RISA can be programmed with reprogrammable logic to do small scale data manipulation and computation, just like the instruction set accelerators currently in use. Furthermore, one or more RISAs can be tightly coupled to instruction and data paths in parallel with the predefined execution units in microprocessors. This tight coupling provides fast and efficient execution of new instructions and a significant improvement in performance compared to prior art approaches. The RISA provides the capability for users to program instructions that may be difficult to do in the general purpose processor, and not in wide enough use to warrant a hardware accelerator. Further, the reprogrammable instruction set accelerator can be reprogrammed with different instructions at different times, saving space and cost in the computer.

The instruction set accelerator may be reprogrammed for each program that runs on the computer, or a manufacturer may select a few instructions and ship them stored with the computer. A program may reprogram the reprogrammable instruction set accelerator several times during the program to speed up different parts of the program. The logic space in the reprogrammable instruction set accelerator may be allocated by the computer system, and instruction sets swapped, using operations similar to overlays, virtual memory or caching.

Accordingly, the present invention can be characterized as a data processor which comprises a defined execution unit coupled to internal buses of the processor for execution of a predefined set of instructions, combined with one or more programmable execution units coupled to the internal buses for execution of programmed instructions. The programmable execution units comprise in one aspect of the invention, field programmable gate arrays having respective configuration stores, and resources for accessing the configuration stores to program the programmable execution unit.

In one aspect of the invention, an instruction register is included in the data processor which holds a current instruction for execution, and is coupled to an instruction data path to supply the instruction to the defined execution unit and to the programmable instruction unit in parallel, through appropriate decoding resources.

The processor may include instruction fetching resources and other logic which are responsive to condition codes. A condition code register is connected to receive condition codes from both the defined execution unit and from the programmable execution unit.

In addition, because the programmable execution unit may be reprogrammed after manufacture, the timing for execution of a programmed instruction may not be well predictable. Thus, the programmable execution unit includes logic to signal the instruction fetching resources to provide a next instruction when execution of the programmed instruction is done.

The programmable execution unit may comprise a configuration store, and resources for accessing the configuration store to program the programmable execution unit are provided, which according to one alternative utilize the internal buses of the data processor. Thus, the programmable execution unit may be reconfigured under control of the defined execution unit. Alternatively, the programmable execution unit may include a configuration port which is independent of the internal buses of the data processing system, which allows access to the configuration store for reprogramming the programmable execution unit through a separate port.

As mentioned above, the processor includes an instruction register which holds instructions in sequence for execution, and an instruction data path which supplies the instruction to the defined execution unit and the one or more separately configurable programmable execution units. Instruction management logic is included according to one aspect of the invention which is coupled with the instruction register. The instruction management logic includes a store, configured like a cache page table store, which stores identifiers for the programmed instructions in the respective sets of programmed instructions for which the configurable programmable execution units are configured. Matching logic is included which matches the incoming instructions in the instruction register with instructions in the store to determine when a programmed instruction miss occurs in the instruction sequence. Thus, if the incoming instruction does not match with instructions in the store, then a miss is indicated. In case of a miss, the instruction management logic issues a signal which causes the processor to stop, and to reconfigure the programmable execution unit to execute the instruction. According to one alternative, the instruction management logic includes instruction mapping logic to supply a pointer to a location storing configuration data for the programmable execution units needed for the incoming instruction. The configuration data is stored on chip, or stored external to the chip in system memory in alternative approaches.

When the processor includes more than one programmable execution unit, the store which holds identifiers for the sets of instructions configured in the system, will also include an indication of which of the plurality of separately configurable programmable execution units is configured for the specific incoming instruction. This indicator is supplied to the decoder, or otherwise coupled with the instruction path to properly route the control data.

The instructions according to one aspect of the invention will have a prespecified format, including an opcode field specifying an operation by one of the defined and programmable execution units, and plurality of address fields specifying addresses of operand data and result data. The pre-specified format according to one alternative may include a defined/programmed flag, specifying a defined or programmed instruction. Thus, the decoder will be responsive to the flag to enable or disable the programmable execution unit for the purposes of access to the internal buses and register files on the device. According to another alternative, the pre-specified format may include an immediate data field, such that programmed instructions use the opcode field to identify the instruction as a programmed instruction, and the immediate data field to identify a programmed operation. A third alternative instruction format includes both an opcode for the defined execution unit and an opcode for the programmable execution unit.

Accordingly, the present invention provides a new method for executing a computer program which includes a particular function. The method includes providing a defined instruction execution unit and programmable instruction execution unit in parallel with the defined instruction execution unit. The programmable instruction execution unit is programmed to execute at least a portion of the particular function in response to a programmed instruction. A sequence of instructions is supplied including the defined instructions and the programmed instruction. The defined instructions are executed in the sequence in the defined instruction execution unit and the programmed instruction is executed in the programmable instruction execution unit. The programmable instruction execution unit can be reprogrammed when the user changes from one application to the next using a configuration port for the programmable instruction execution unit.

Thus, the reprogrammable instruction set accelerator may be programmed through internal processor data paths or through separate, dedicated programming paths initiated by an instruction from the general purpose, defined execution unit. Selecting the instructions to emulate can be done manually, by inspecting the instructions or procedures to be executed, and crafting new instructions to implement chosen functionalities. A compiler may be used to automate the addition of programmed instructions to programs it compiles to improve their speed.

Instruction selection may also be done automatically, by profiling the program to see how frequently various procedures or lines of code are used, then replacing them with programmed instructions. This requires software to compile from the programming language to logic gates. This capability now exists with high level synthesis from VHDL, VERILOG or other languages for specifying logic. Given such capability, a high level synthesis system that takes a high level programming language such as "C" as input and generates logic for the RISA is provided.

Extraction of instructions may also be done on the fly, by profiling analysis of the instructions to the general purpose processor, then compiling from those instructions to logic in the reprogrammable instruction set accelerator. This profiling may be done before hand, or may be done during execution. In the latter case, the computer learns which instructions are frequently executed and optimizes them as it runs.

Alternatively, many instructions for the RISA can be defined in advance and the compiler may use all or a subset of these additional instructions when compiling. The needed instructions are loaded into the RISA when the program runs.

Thus, the present invention can also be characterized as a method for translating a computer program into executable code that includes providing a library of defined instructions executable by a defined execution unit, and of programmed instructions executable by a configurable execution unit, when configured for execution of the programmed instructions. The program is compiled using the library to produce an executable version of the program using both defined and programmed instructions. The step of compiling includes in one approach generating a preliminary object code using the library, identifying programmed instructions in the preliminary object code according to a criterion relevant to efficiency of use of resources needed to configure the configurable execution unit, such as how often they are used and the like, and replacing the identified programmed instructions with defined instruction segments to produce the executable version. Thus, when configurable resources are limited, the programmed instructions which are seldom used can be replaced by segments of defined instructions to conserve configurable resources.

According to another approach, the library includes a plurality of sets of programmed instructions and the step of compiling includes generating preliminary object code versions using respective sets of instructions in the library, and selecting one of the preliminary object code versions as the executable version. According to this approach, the most efficient version is selected, in terms of criterion desired for a particular application, such as conservation of configurable resources, speed of execution, or the like.

The invention can also be characterized as a method for translating a computer program to executable code which comprises providing a library of instructions including instructions executable by a defined execution unit, and generating preliminary object code using the library of instructions. Using the preliminary object code based on defined instructions, sequences of instructions are identified in the preliminary object code to be replaced by a programmed instruction. The identified sequence is replaced in the preliminary object code with a programmed instruction to produce the executable version. Thus, after the step of identifying a sequence of instructions, a programmed instruction is synthesized, and configuration data for configuring the configurable execution unit is generated in order to create a synthesized programmed instruction. The synthesized programmed instruction is tailored to the specific identified sequence of instructions for optimum performance.

In addition, the step of identifying a sequence of instructions may include generating a profile of the preliminary object code and selecting based on the profile a commonly used sequence of instructions. By replacing a commonly used sequence of instructions within the program by a single programmed instruction, performance of the overall program can be significantly increased.

The library of instructions used for generating the preliminary object code according to this aspect may include both defined instructions, and pre-configured programmable instructions. The technique of identifying and replacing commonly used sequences can thus optimize the code, even when both defined and programmed instructions are used in the sequence being replaced.

Another technique for identifying the sequence of instructions for replacement may be implemented automatically. Thus for instance, a microprocessor may include an instruction cache in the instruction fetching path. The contents of the instruction cache are monitored to identify a sequence of instructions which resides in the cache for a significant amount of time. When such a sequence is detected, configuration data for the programmable execution unit is generated to create a programmed instruction to replace the identified sequence. When the programmable execution unit is configured, the identified sequence of instructions in the instruction cache can be replaced by the programmed instruction.

Accordingly, the present invention provides a technique for improving the performance and flexibility of general purpose processors based on the use of reprogrammable logic techniques. The invention provides greater performance improvements and more flexibility than prior art attempts to optimize instruction execution in general purpose processors.

Other aspects and advantages the present invention can be seen upon review of the drawings, the detailed description and the claims which follow.

FIG. 1 is a simplified block diagram of data processing system utilizing the reprogrammable instruction set accelerator (RISA) according to the present invention.

FIG. 2 is a schematic diagram of a integrated circuit microprocessor including a defined execution unit and a reprogrammable execution unit according to the present invention.

FIG. 3 illustrates one example instruction format for use according to the present invention.

FIG. 4 illustrates an alternative example instruction format for use according to the present invention.

FIG. 5 illustrates another alternative example instruction format for use according to the present invention.

FIG. 6 provides a schematic block diagram of a microprocessor according to the present invention with a plurality of RISA units, and including a page table storing identifiers of programmed instructions for which the RISA units are configured.

FIG. 7 is a diagram of a computer program for generating object code using both programmed and fixed instructions according to the present invention.

FIG. 8 is a diagram of a computer program used for compiling a computer program with synthesized programmed instructions dynamically created to improve performance of the program.

A detailed description of the preferred embodiments of the present invention is provided with reference to the figures, in which FIG. 1 illustrates data processing system using the reprogrammable instruction set accelerator according to the present invention. As shown in the figure, the data processing system includes a host CPU generally 10 coupled to the system bus 11. Also coupled to the buses are instruction memory 12, data memory 13, user interface 14, possibly a disk drive 15, and other processing resources generally 16. The host CPU is made up of the basic microprocessor (MP) components generally 20, and a reprogrammable instruction set accelerator (RISA) generally 21. The basic microprocessor includes internal buses 22 and 23, also called buses A and B, respectively. A defined instruction, fixed execution unit 24 (or ALU) is coupled to buses 22 and 23, and supplies a result through a multiplexer 25 to bus 22. A register file 33 is accessible across the buses 22 and 23. An instruction path schematically represented by the instruction register IR 26 supplies an instruction to the defined instruction execution unit 24. Also, the defined instruction execution unit 24 generates condition codes which are supplied through multiplexer 27 to a condition code register 28 which is used by the processor as well known in the art in instruction sequencing and the like.

In parallel with the general purpose microprocessor is the reprogrammable instruction set accelerator RISA 21. The RISA 21 comprises a field programmable gate array 30, which includes a configuration store 31. The field programmable gate array 30 is coupled to the internal buses 22 and 23, and supplies a result through multiplexer 25 to bus 22. The instruction path through instruction register 26 supplies an instruction to the field programmable gate array 30 in substantially the same manner and timing as it supplies instructions to the defined execution unit 24. The field programmable gate array 30 also supplies conditions codes through multiplexer 27 to the condition code register 28.

A configuration store 31 is coupled with the field programmable gate array 30. The configuration store 31 may accessible through a dedicated port generally 35, or by means of the internal buses 22 and 23 for dynamically reprogramming in a field programmable gate array 30.

In the embodiment illustrated in FIG. 1, the RISA 21 is implemented using field programmable gate array logic. The field programmable gate array logic may take a variety of forms, such as the dynamically reconfigurable architecture described in our co-pending U.S. patent entitled A PROGRAMMABLE LOGIC DEVICE WHICH STORES MORE THAN ONE CONFIGURATION AND MEANS FOR SWITCHING CONFIGURATIONS, filed Apr. 20, 1994, U.S. Pat. No. 5,426,378, issued on Jun. 20, 1995, invented by Randy T. Ong, which was owned at the time of invention and is currently owned by the same assignee as the present application, and which is incorporated by reference as if fully set forth herein. Alternative programmable logic structures may be utilized. For instance, the RAM based configuration store of typical FPGA designs may be replaced using reprogrammable non-volatile stores such as EEPROM. Also, the configuration store may be programmable during manufacture rather than dynamically. Thus, the manufacturer may use more permanent programming technology such as, anti-fuses or the like to configure a new instruction into a previously defined instruction set.

As shown in FIG. 1, the field programmable gate array is used as an execution unit which is reprogrammable, and connected in parallel with the defined execution unit 24. The system expects the field programmable gate array 30 to return results in a same manner as the standard defined instruction execution unit. The field programmable gate array 30 in the embodiment shown uses the same write back path as the defined execution unit. However, a separate write back path may be used for the FPGA 30 if desired.

The RISA 21 includes an optional data port or ports 36, 37 dedicated for use by the RISA 30. The data port 36 is coupled to the system memory 12, 13 across bus 11. The data port 37 is coupled to an external data source, such as a video data source, an audio data source or memory isolated from the system bus 11.

The field programmable gate array 30, executes instructions that take an amount of time which is not predictable prior to configuration, and do not match well with the pipeline speed of the fixed execution unit 24. Thus, logic is included to hold the processor during execution of a programmed instruction, and to signal the processor when the programmed instruction is complete.

In addition, one embodiment of the RISA 21 operates in an overlapped fashion with the defined instruction execution unit 24 for some operations, taking advantage of the parallel execution units in the system for greater performance.

As shown in FIG. 1, in one aspect of the invention, a compiler specially designed for the RISA based system is included in the instruction memory of the data processing system. Also, an instruction library including programmed instructions for the RISA and defined instructions for the fixed execution unit is stored in the data memory in support of the compiler. This software, and the supporting libraries, enable on the fly reconfiguration of the RISA, or compiling of programs to take maximum advantage of the RISA, as described in more detailed below with reference to FIGS. 6-8.

Some examples of the kinds of operations that may be implemented using the reprogrammable instruction set accelerator include the following:

bit rotation/field extraction, used in instruction emulation or encryption or decryption;

on-bit counting;

polynomial evaluation;

spreadsheet resolution, such as an instruction that calculates each cell in the spreadsheet to be resolved;

searching in a document;

spell-checking;

database access routines;

procedure invoke/return operations;

programming language interpreters;

emulation of another processor; and

context switching for multi-processing.

FIG. 2 provides a more detailed block diagram of an integrated circuit microprocessor which includes the RISA according to the present invention. Those skilled in the art will recognize the basic components of the microprocessor. Thus the figure is intended to represent wide variety of microprocessor architectures, including complex instruction set processors, reduced instruction set processors, and other data processing architectures.

Thus, the system includes a execution unit 100 which is optimized for a predefined instruction set. The execution unit 100 is coupled to internal buses 101 and 102 for receiving operand data at its ports A and B, and supplying result data on its port Y back to one of the internal buses, e.g. bus 101. A register file 103, and input/output registers 104 are also coupled to the buses 101 and 102, and act as sources for operands, and locations to store result data. The registers accessible by the executing unit 100 include private registers 140 which are dedicated for use by the defined execution unit 100, and not by the programmable execution unit (FPGA RISA 120 described below). In an optional embodiment, the private registers 140 are directly coupled with the defined execution unit 100 as indicated by line 141.

In another embodiment, internal bus 101 is broken into two sections at point 150, thereby allowing independent operation of execution unit 100 and RISA 120. In yet another embodiment, internal bus 101 is broken into two sections at point 151, thereby allowing independent outputs with shared inputs.

Also, coupled to the internal buses 101 and 102 is an instruction address register 105, and other resources in the instruction address fetching path. The instruction address register 105 includes basic incrementing logic 106 for sequencing through a sequence of instruction addresses for an instruction memory off chip. Also, an instruction control state machine 107 is coupled to the instruction address register 105 for managing the instruction stream as known in the art. The instruction control state machine 107 is also coupled to other components on the chip as appropriate.

A condition code register 108 supplies condition codes to the instruction control state machine 107 involved in the instruction sequencing decisions, and also as indicated by arrow 109 to other processing resources in the system that suits the particular implementation. Condition codes are generated on line 110 by the defined execution unit 100 as known in the art.

An instruction register 111 receives instructions generated in response to instruction address register 105. Instructions are supplied from the instruction register 111 to decoder resources 112. The decoder resources 112 supply control signals, generally represented by arrow 113, throughout the device to control accessing of the register files, bus timing, and other functions on the processor. The decoder also generates an opcode on line 114 which is supplied at the instruction input I on the defined execution unit 100.

External data, instruction, and address ports 115 are included as known in the art, for managing flow of data, instructions and addresses into and out of the chip. The external data, instruction, and address ports 115 are coupled to the I/O registers 104, the instruction address register 105, and the instruction register 111.

According to the present invention, a reprogrammable instruction set accelerator is included on the chip. Thus, a RISA field programmable gate array (RISA FPGA) 120 is coupled to the internal buses 101 and 102, to receive operand data at ports A and B and supply result data at port Y to and from the buses. The opcode from line 114 is supplied at an instruction port I on the RISA FPGA 120. Condition codes generated by the RISA FPGA 120 are supplied on line 121 to the condition code register 108. These condition codes may use the same spaces in the condition code register 108 as the defined execution unit 100, use separate bits within the same register 108, or use a different register altogether coupled to operate in parallel with the condition code register 108 for the defined instructions.

The RISA FPGA 120 is also connected across line 122 to an external configuration port 123. The external configuration port 123 may take the variety of configurations. For instance, the external configuration port 123 can be implemented as fast DMA port or as a serial port with appropriate control signals for loading a data stream into a configuration store of the FPGA 120 independent of the internal buses 101 and 102. In an alternative, the RISA FPGA 120 is coupled with the internal buses 101 and 102, such as represented by line 124. In this case, configuration of the RISA FPGA 120 is managed by the defined execution unit 100 using the buses 101 and 102 and the I/O registers 104 of the basic architecture, without using additional pins on the chip.

According to the present invention, the RISA FPGA 120 also includes a program done port PD, and a configuration done port CD. The program done port is coupled to logic within the FPGA 120 to generate a signal on line 125, used to indicate when execution of a programmed instruction in the RISA FPGA 120 is completed. This signal is supplied to the instruction control state machine 107 and to other timing resources in the system, to hold the processor while a programmed instruction is executing. Because programmed instructions may require more time to execute than a defined instruction in the optimized execution unit 100, the pipeline must be held to accommodate the longer execution stage. The system may be improved or optimized with additional pipeline stages and the like to improve performance as necessary. Alternatively, the instruction state machine may ignore the PD signal so that the execution unit 100 and the RISA 120 may operate simultaneously.

The signal on the configuration done port CD is used by the RISA FPGA 120 to delineate a configuration process in the RISA FPGA 120. This port is coupled to the condition code register or the instruction control state machine as necessary, as indicated by line 126. The configuration done signal is used during reconfiguration of the RISA FPGA. In one state, the system is held while the RISA FPGA loads configuration data. The other state, the system is signaled that the RISA FPGA is ready for execution. This system may be configured such that during reconfiguration of the RISA FPGA 120, the defined execution unit 100 is capable of continuing execution as long as an opcode for the RISA is not encountered. Alternatively, the system may be held up completely to allow for reconfiguration of the RISA FPGA 120.

In an optional embodiment, special working registers 130 are included which are more closely coupled with the RISA FPGA 120 than the standard register file 103. These working registers are dedicated for use by the RISA FPGA 120 by the decoder 112, or in one alternative, as indicated by line 131, access to them is explicitly controlled by the RISA FPGA 120 as necessary. In an alternative, the registers are accessible under control of both the decoding resources 112 and the FPGA resources 120.

Accordingly, the present invention provides the ability to execute a predefined set of instructions with optimized hardware, such as the execution unit 100, while allowing a user to reconfigure the system to execute a programmed instruction using programmable execution unit resources, such as the RISA FPGA 120.

Instruction formats used to accommodate the defined instructions as well as the programmed instructions are illustrated in FIGS. 3, 4, and 5. In FIG. 3, an instruction format, such as used for instructions stored in the instruction register 111 is illustrated for a defined instruction 200, and for a programmed instruction 201. The instruction format includes a defined/programmed (D/P) bit, an opcode, a plurality of addresses, such as source addresses for ports A and B, and a result address for the output Y, and an immediate data field. Thus, for a defined instruction, the D/P field 202 indicates a defined instruction, the opcode field includes a standard opcode such as an "add" instruction in field 203. The source of operand A is identified as register R3 in field 204. The source of operand B is identified as register R4 in field 205. The destination for result Y is specified as register R5 in field 206. Immediate data involved in the operation may or may not be stored in the final field 207.

A programmed instruction is illustrated at 201. In the defined/programmed flag stores the programmed instruction flag in field 208. An opcode defined for the field programmable gate array is included in field 209. There may be a plurality of such opcodes FPGAOP(N) configured for the RISA as suits a particular application. The addresses for operand A and operand B are stored in fields 210 and 211. The address for the result Y is included in field 212. Immediate data involved in the operation, if any, is stored in field 213.

An alternative instruction format is shown in FIG. 4. In FIG. 4, the defined/programmed flag is not used. Rather, the defined instruction 250 includes an opcode field, storing a standard opcode such as "add" in field 251. The addresses for operands A and B are stored in fields 252 and 253. The address of result Y is stored in field 254, and immediate data is stored if necessary in field 255. A programmed instruction is shown at generally reference number 256. It includes a defined opcode FPGAOP in field 257. This signals the decoder that a programmed function is specified in the immediate data field. The addressees for operands A and B are stored in fields in 258 and 259. The address in result Y is stored in field 260. The specific programmed opcode is stored in field 261, which is normally immediate data.

FIG. 5 shows a third alternative instruction format, which includes both a defined instruction opcode field and a programmed instruction opcode field. The format includes a defined opcode field 270, followed by three register address fields 271, 272, and 273 for operands A and B and result Y respectively. Next, a programmed opcode field 274 is included followed by a register address field 275 for operand C. Finally, an immediate data field 276 is included. The decoder 112 is responsive to the instructions in the format of FIG. 5 to select one set of registers for the defined opcode in field 270, such as registers defined in fields 271, 272, and 273 and a separate register for the programmed opcode in field 274, such as the register in field 275. This separate register may reside in the working register file 130. Immediate data in field 276 may or may not be necessary. According to this instruction format, both the defined opcode and the programmed opcode are decoded and executed in parallel, or partially in parallel, for improved performance.

Thus, the decoder 112 in the system of FIG. 2 is configured to respond to instructions in the formats of FIG. 3, FIG. 4, or FIG. 5 as appropriate to manage resources on the microprocessor to execute both the defined instructions and the programmed instructions.

FIG. 6 illustrates additional features of a microprocessor implementing the present invention. In particular, it illustrates an architecture using more than one RISA, and more details concerning the instruction sequencing and RISA instruction handling according to other aspects of the invention. Thus, FIG. 6 illustrates the internal data pipeline 300 of the microprocessor schematically. This internal pipeline includes the features shown in FIG. 2, and other features as known in the art The internal pipeline 300 includes the fixed logic unit 301, which executes the defined instructions. A first programmable execution unit 302 labeled RISA A, is included, and a second programmable execution unit 303 labeled RISA B is included. The RISA A 302 and RISA B 303 are coupled in parallel with the fixed ALU 301 in the manners described above with reference to FIG. 2. The first and second RISAs 302, 303 may be implemented using a single field programmable gate array structure, so long as it is large enough to support the parallel connections and has sufficient I/O resources. Alternatively, more than one separate FPGA module may be used.

FIG. 6 illustrates certain aspects of the instruction path for one embodiment of the present invention. Thus, the instruction fetching resources include an instruction sequencer 310, which may include an instruction cache 311 for holding pre-fetched and commonly used instructions. The instruction sequencer 310 supplies a current instruction to instruction register 312. The instruction register 312 supplies the current instruction on line 313 to a decoder 314. Also, the instruction, or a subset of the instruction, which may be hashed, is supplied on line 315 to a RISA instruction page table 316. Also, line 315 supplies the identifier of the instruction to page table management logic 317. The RISA instruction page table 316 is used for detecting a RISA instruction miss, which occurs when the instruction in the instruction register 312 is a programmed instruction, but is not listed in the table 316. If it is not listed in the table 316, then the page table management 307 will issue a signal represented by line 318 which causes the execution of the instruction to halt, and initiates a configuration process. The configuration process results in configuring one of the programmable instruction execution units to execute the instruction which caused the miss, and in updating the page table with the newly configured instruction. Also on the integrated circuit, a memory 320 is included for storing a RISA instruction map table. In the event of a RISA instruction miss, the page table management 317 will forward the identifier of the instruction on line 321 to the RISA instruction map table in memory 320. This table stores a pointer represented by line 322 which indicates a location in a configuration store of data for configuring one of the programmable execution units to execute the instruction subject of the miss. The configuration store is an external configuration store 323 in the example shown in the figure. The external configuration store is coupled through configuration resources represented by arrow 324 to the internal programmable execution units 302 and 303, and used to configure one of them to execute the instruction.

In alternative systems, the RISA instruction map table is not on the integrated circuit. However, performance may be improved by incorporating this table on chip. Also, alternative systems include the configuration store 323, or a subset of it, on chip.

According to the embodiment shown in FIG. 6 where there are a plurality of programmable execution units RISA A, RISA B as shown in the figure, the RISA instruction page table 316 includes a field for each instruction in the table which indicates which of the plurality of programmable execution units is configured to execute the instruction. Thus, the field 330 stores the value A or B for each instruction in the table as illustrated in FIG. 6. This indicator from field 330 is supplied on line 331 to the decoder 314 which decodes the incoming instruction on line 313 appropriately. The decoder 314 supplies control signals and op codes for executing fixed instructions on line 340, supplies control signals and op codes for instructions executed by the programmable execution unit RISA A 302 on line 341, and supplies the control signals and op codes for RISA B 303 on line 342, as appropriate.

In the system of FIG. 6, as mentioned above, an instruction cache 311 is included. Sequences of fixed instructions or combinations of fixed instructions and programmed instructions which dwell in the instruction cache 311, are excellent candidates for new programmed instructions. Thus, as discussed in more detail below, these sequences of instructions may be replaced by a programmed instruction with optimum data flow and inner loop fit for the microprocessor using a programmed instruction to replace the sequence. The programmed instruction is then executed upon detection of an access to the start of the sequence in the cache, or the program can be re-compiled to include the new programmed instruction. As these instructions are generated, the benefit of each instruction should be scored in order to manage the growth of the RISA instruction list. As this list expands, the configurable resources in the programmable execution units can be exhausted. Thus, management of the selection of RISA instructions is advisable using techniques such as those described below with reference to FIGS. 7 and 8.

Using the architecture of FIG. 6, the RISA A and RISA B can be set up so that one runs while the other is being configured. Also, there may be separately loadable RISA blocks within a single field programmable gate array. Thus a field programmable gate array may be partially reprogrammable and able to configure during operation of another segment. In addition, reconfiguration of the RISAs may be initiated on command from a running program, replacing other instructions. The RISA instruction page table 316 provides information about which instructions the programmable execution units contain. Alternatively, the RISAs A and B are configured to store this information, and selectively respond to op codes supplied in parallel to both.

As mentioned above, when a RISA instruction occurs, the processor checks the page table so see if the desired instruction is present. If it is present, the instruction proceeds. If it is not present, the program is stopped while one of the programmable execution units is reloaded or incrementally loaded with the new instruction. The location of the programmed instructions may be stored in a special instruction map as shown in the figure, loaded at any time, but before the instruction is requested. The RISA instruction miss may be handled like a cache page fault in other storage systems, allowing other programs to run during the instruction miss processing.

The technique used for loading instructions in the page table, and configuring the programmable execution units, may take a variety of forms. For instance, RISA instructions may be collected into sets of instructions which have unique applications (spread sheet applications, floating point applications, image processing applications, or the like). The page table is then loaded with the sets of instructions in an overlay technique to manage contents of the RISA instruction page table. Alternatively, a least recently used method or other cache management type method can be used for managing which instruction to replace when new ones are added.

FIGS. 7 and 8 illustrate computer programs which can be used to translate a high level instruction based program to final object code using both RISA and fixed instructions.

FIG. 7 illustrates a process which begins with a high level language program 400. This program is supplied to a compiler 401 which compiles the program using a library which includes in this example a plurality of sets of RISA instructions 402-A, 402-B and a set of fixed instructions 403. The compiler 401 generates preliminary object code 404 using the library of instructions, including the RISA instructions selected by the compiler. The set of RISA instructions selected by the compiler may vary from program to program. Thus, some programs may require more RISA instructions than others which cause them to consume the configurable resources available in the system for configuring RISA instructions. Thus, according to the method in FIG. 7, the preliminary object code is supplied to a module which selects which RISA instructions to be kept in the final object code, and replaces the other RISA instructions with sets of fixed instructions. Thus, seldom used RISA instructions in the preliminary object code are identified in the module 405. This is an iterative process as illustrated by line 406, executed until the set of RISA instructions fits within the available configurable resources for the RISA. A final object code 407 is produced according to this technique.

According to an alternative embodiment also shown in FIG. 7, the compiler 401 is run using a single set of RISA instructions for each pass. Thus, a first preliminary object code is generated using the set 402-A of RISA instructions. This first preliminary object code is optimized to fit the available configurable resources using the module 405 and a final object code 407 based on the set 402-A is produced. This final object code is saved using a find best set module 408. Next, the compiler is executed using the set 402-B of RISA instructions. The final object code produced at block 407 for the second set 402-B of RISA instructions is then compared with the earlier produced object code to find the best fit for the system in terms of speed of execution and use of configurable resources in the RISA. This process is iterative, with the find best set module 408 signaling the compiler 401 to create new object code files for each of the sets 402-A and 402-B of RISA instructions in the system.

FIG. 8 illustrates a system which automatically synthesizes RISA instructions for a computer program. Thus, the system begins with a high level language program at block 500. A library is included which includes RISA instructions 501-A, 501-B, and additional sets of RISA instructions if desired. The library also includes a set of fixed instructions 502. The program 500 is supplied to a compiler 503 which compiles the program 500 to produce preliminary object code 504 which may include RISA instructions from one or both of sets 501-A and 501-B. The preliminary object code 504 is profiled in block 505, such as by executing an instruction like the "prof" UNIX instruction which generates profile data for a computer program. The execution profile generated by block 505, identifies commonly used sequences of fixed instructions or combinations of fixed and RISA instructions. These commonly used sequences are supplied to a synthesis block 506 which synthesizes a new programmed instruction for commonly used sequences. The synthesized RISA instruction is merged with the balance of the object code as indicated across line 507 in a merge module 508. The merged object code is applied as a final object code 509 for use by the processor. The synthesis block 506 generates gate level configuration data for the programmable execution unit in the system for which the compiler 503 is designed.

The synthesis and merge technique shown in FIG. 8 can also be used dynamically in association with the sequences of instructions which dwell on the instruction cache 311 of FIG. 6.

Thus, techniques for deciding which instructions to configure into the RISA are provided. For instance, the program can be designed to run on the fixed execution unit only, and measurements made of which instructions or instruction sequences are most common. As mentioned above, this measurement technique is called profiling. These most commonly used sequences, which may be bounded by size to manage development of the RISAs, are collected into separate groups. Each group is optimized to form a single RISA instruction that performs the whole task. This optimization uses high level synthesis techniques to optimize the data path. Alternatively, the compiler can supply several different sets of instructions. The program is compiled several times, one for each set of instructions. The best program is then selected for use in the system.

According to another alternative method, the compiler has many predefined RISA instructions and compiles the program using them all. After compilation, the list of RISA instructions used in the program is kept. If the RISA instructions in the list of used RISA instructions do not fit into the configurable resources of the RISA, one or more of the RISA instructions is converted back to CPU instructions. For instance, a library of CPU instruction segments that implement each of the RISA instructions is kept. A segment from the library is merged with the preliminary object code in place of the RISA instruction. Alternatively, profiling can be used to find the RISA instructions which are used least often. These least used RISA instructions are converted back into fixed instructions, or to a combination of fixed instructions and simpler RISA instructions, until the used RISA instructions fit within the available configurable resources.

The compiler relies on these dynamic strategies for optimizing configurable resources and use of RISA instructions, to load the programmable execution units with instructions appropriate to the current task.

Accordingly, the present invention applies field programmable gate array technology to improving the performance of general purpose processors. By providing one or more reprogrammable instruction set accelerators on chip, optimized microprocessors can be improved to handle special programmed instructions. The special programmed instructions can implemented by configuring the RISA during manufacture, or dynamically before or during execution of a program. Dynamic reconfiguration is suitable so long as the reconfiguration overhead does not use more time than is saved by including the special programmed function in the instruction sequence. However, the ability to optimize execution of special functions in a high performance microprocessor integrated circuit can have significant impact on performance in many environments, even though the particular instruction added to the system may vary widely from user to user.

The reconfiguration of the RISA can be automated, using tools commonly found in current logic synthesis and compiling technologies. Thus, programs to be executed can be analyzed using sophisticated synthesis and compiling software to identify functions that could be improved by configuration of the FPGA RISA. Once identified, a configuration can be created and stored in memory. At the appropriate time during execution of a program, the configuration can be loaded into the configuration store on chip, and performance of execution of the program vastly improved.

The foregoing description of preferred embodiments of the invention has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Obviously, many modifications and variations will be apparent to practitioners skilled in this art. It is intended that the scope of the invention be defined by the following claims and their equivalents.