CPU with cache memory

This invention relates to the field of central processing units (CPUs) and more particularly to CPUs supporting cache memory.

As the need for higher performance computer systems increases, the design of central processing units (CPUs) incorporated within such systems are becoming more and more sophisticated. In order to achieve high performance CPUs utilize internal (L1) or external (L2) cache memory which are architected to be in closer proximity to the CPU relative to the system memory. Caches are high speed storage which exist between the CPU and memory. Typically the first time data or instructions are accessed from main memory a copy is placed into the cache. Should this data or instruction be accessed again it can be retrieved faster due to the higher speed of the cache over memory. However, typically the amount of storage contained in the cache is much smaller than the amount of storage represented by main memory. Therefore, cache coherency techniques exist to determine how data and instructions are maintained within the cache. These techniques are based on maintaining a 'high hit rate' (getting the instructions and data from the cache as frequently as possible).

The cache memory, which may be Static Random Access Memory (SRAM), increases CPU performance by storing the data most recently used by the CPU. Because of its proximity to the CPU and its high speed, cache memory allows CPU to access its needed data much faster without initiating a lengthy memory cycle for getting data from the system memory.

Because of the cost associated with SRAM devices, cache memory has a finite size. Today, a typical L1 cache size is between 8 Kilobytes to 64 Kilobytes while L2 cache sizes range from 128 Kilobytes to 1 Megabyte. Generally, computer systems and CPUs are designed to force out, i.e. cast out, old cache data when transfer of new cache data causes size limitation of the cache size to be exceeded.

One of the more popular CPU platforms is the X86 family of processors produced by Intel Corporation, some of which offer CPUs having an L1 cache memory. For example, some of Intel's 486 processors include 8 Kilobytes of internal L1 cache, whilst Intel's Pentium Pro (R) processor has 16K L1 cache and 512K L2 cache. However, much of the software available for the X86 family of processors makes use of certain instruction types, such as Repeat (REP) instructions, that cause useless data to purge useful cache data. Purging of useful CPU data degrades CPU performance since such data must be replenished by lengthy system memory cycles generated by the CPU. REP instructions are used to manipulate blocks of data. The purpose of the manipulation can be to move the data from one area of memory to another, to compare a block of data with another block of data to search a block of data for a specific pattern or similar operation. The advantage of REP instructions is that a block of data can manipulated with a signal instruction.

Unfortunately a REP instruction will tend to 'purge' the cache. This is to say, when the repeat operation is moving a block of data that block of data is also placed into the cache thereby removing that data which is already in the cache. In many cases although the data is placed in the cache, the CPU has no further need for it. However other data which the CPU may have had a need for is removed and must be accessed from memory with the associated lower performance. Therefore, a repetitive string search instruction may cause the L1 cache memory to be filled with useless string data at the cost of loss of useful cache data. Therefore it is desirabble to improve CPU performance when certain instructions are executed by the CPU.

Accordingly, the invention provides a central processing unit (CPU) for a computer system having a cache storage device comprising: <ul id="ul0001" list-style="none" compact="compact"><li>an instruction processing unit for processing CPU instructions including an instruction decoder for decoding said instructions for determining an instruction type;</li><li>an instruction type detector for detecting cacheable or non-cacheable instruction types; and</li><li>an execution unit for executing said instructions; said execution unit being responsive to the instruction type detector for outputting a cachability indicator signal indicating whether data associated with a detected instruction type is cacheable.</li></ul>

It is preferred that said instruction type detector detects repeat type instructions, and said execution unit is responsive to detection of said repeat type instruction for outputting said cachability indicator signal to indicate that the data associated with the repeat type instruction is not cacheable.

The preferred embodiment includes a storage buffer for storing data associated with said repeat instruction type, typically comprising two registers. The cache storage device is coupled to the CPU via a data bus having a predetermined data bus width, and said registers have a register width substantially equal to an integer multiple (eg two) of the data bus width.

The preferred embodiment further includes an internal cache controller responsive to said cachability indicator signal for determining cachability of data associated with the instruction type into said internal cache.

A CPU such as described above can be used in a computer system also comprising system memory and a memory controller for controlling system memory. Preferably said memory controller is responsive to said cachability indicator signal for determining cachability of data associated with the instruction type into said internal cache. In the preferred embodiment, the computer system further comprises an external cache.

Thus in the preferred embodiment, a central processing unit is coupled to a cache memory device via a data bus having a predetermined width. The CPU includes an instruction processing unit for processing CPU instructions having an instruction type field. The instruction processing unit includes an instruction type detector for detecting cacheable or non-cacheable instruction types. The CPU also includes an execution unit for executing the instructions. The execution unit is responsive to the instruction type detected by the instruction type detector for providing a CPU output signal indicating whether the data associated with a detected instruction type is cacheable.

A preferred embodiment of the invention will now be described by way of example only with reference to the following drawings: <ul id="ul0002" list-style="none" compact="compact"><li>Figure 1, is a block diagram of an exemplary computer system;</li><li>Figure 2, is a block diagram of a CPU in accordance with the present invention; and</li><li>Figure 3 is a diagram of an instruction format executed by the CPU of Figure 2.</li></ul>

Referring to Figure 1, this shows a block diagram of a computer system 100, such as an IBM Personal Computer system. The computer system 100 may operate as a stand-alone, desk-top work station or it may be part of a larger computer network operating either as a client or server station.

The computer system 100 comprises two processing sides, a host processing side 110 which performs high level processing functions, such as those specified by operating system and applications, and a local processing side 120 which performs peripheral functions, such as providing storage media interface or communication interface for the computer system 100. The host processing side 110 and local processing side 120 are coupled to each other via an input/output (I/O) expansion bus 130. The I/O expansion bus 130 is compliant with one of many suitable I/O expansion bus protocols that allow a local interface controller to take control of the bus for performing a specified function, such as transferring blocks of data across the I/O expansion bus 130. Examples of such bus protocols include those defined by PCI or MicroChannel specifications.

The host processing side 110 may include a single central processing unit 103 or (in a multi- processing system) a plurality of central processing units for controlling the overall function of the computer system 100. The Central processing unit 103 may be one of many commercially available Central processing units such as the Pentium Processor (R) offered by Intel Corporation, or a PowerPC(R) processor offered by IBM Corporation.

As is well known, the Central processing unit 103 initially operates by executing firmware instructions stored in a Read-Only-Memory (ROM) 101. The firmware instructions, also known as Basic I/O System or BIOS may include instructions for detecting attached devices and instructions for a power on self test (POST). Generally, upon completion of the firmware instructions, the Central processing unit 103 executes operating system instructions by loading the OS from a local side storage device 114 into system memory 107. The transfer of data between the Central processing unit 103, the ROM 101 and the system memory 107 is controlled by a memory controller 109, whilst transfer of data between the host processing side 110 and the I/O expansion bus 130 is controlled by an I/O bridge interface 111. In order to enhance processor performance the Central processing unit 103 may include an internal cache (L1 cache) 104 and an optional external cache (L2 cache) 105 for storing most recently used data and instructions thus reducing the need for initiating long system memory access cycles. The cache memory generally comprises a high speed static random access memory (SRAM) which is positioned in close proximity to the Central processing unit 103. The data transfers in and out of the L1 and L2 caches are controlled by a well known cache controller 106. The cache controller stores and fetches data in a multiple of the data bus width referred to as cache lines. The I/O bridge interface 111 is designed to interface between the memory controller 109, the Central processing unit 103 via the external cache controller 106, and the I/O expansion bus 130. The Central processing unit 103 interfaces with the I/O expansion bus 130 and consequently with the local processing side 120 via a suitable I/O bus bridge interface 111.

As explained later in detail the storage device 114 interfaces with the I/O expansion bus 130 via local interface controller 113. One of ordinary skill in the art will appreciate that the depicted host processing side 110 is merely exemplary, and that many variations on system architecture, uniprocessing or multi processing, etc may be provided to implement the functional requirement of the host processing side 110.

On the local processing side 120, the computer system 100 may include a plurality of local interface controllers 113, such as Small Computer System Interface (SCSI) controllers, which are coupled to various peripheral devices 115. As such, storage devices, communication devices, printing devices, networking devices, imaging devices, etc. may be added to complement system functionality and features. For example, the computer system 100 may be utilized as a server station having a Fast Wide SCSI local interface controller with a Redundant Array of Inexpensive Drives (RAID) as storage devices.

Referring to Figure 2, a block diagram of an exemplary Central processing unit 200 is shown. The CPU 200 includes an instruction processing unit 201 which receives instructions from an internal L1 instruction cache device 205. The L1 instruction cache device 205 stores a plurality of instructions which are fetched for processing by a instruction fetch controller 203.

As is well known, the L1 instruction cache device 205 receives and stores instructions over a CPU bus 230 comprising a data bus 235, an address bus 237 and a control bus 239. The CPU local bus is coupled to and interfaces with a CPU interface unit 223 for interfacing the CPU with external components such as the L2 external cache and the system memory shown in Figure 1. The data bus 235 and address bus 237 have a predefined width comprising a predefined number of bits. The data and address bus width characterizes data transfer capability and the size of addressable area of system memory by the CPU. Accordingly, data bus 235 may for example comprise a 16-bit, 32-bit, or a 64-bit data width while the address bus typically has an address width in the range from 16 to 64 bits. The instruction processing unit 201 among other things determines the type of the instruction and appropriately passes it to one or more execution units 221. As is well known super scalar CPUs may include a plurality of execution units for executing multiple tasks simultaneously. Depending on the instruction, the execution units may for example perform a mathematical or a logical operation on operand data provided by the instruction. The execution units receive needed data from an L1 data cache 207 which, in similar fashion to the L1 instruction cache, stores data received over the CPU bus or data received from the E-unit 217. The CPU also includes a set of general registers identified as AX-EX for storing in a well known manner instruction operand, temporary data and indexes as dictated by the instruction 300. The instruction processing unit upon receipt of the instruction 300 processes it to determine the instruction type by decoding. Instruction decoding is performed by the instruction decoder 209. Upon execution of the instruction, the E-unit appropriately outputs signals on the data bus 235, address bus 237 and/or control bus 239 over the CPU interface unit 223.

Referring to Figure 3, a pictorial diagram of an exemplary instruction 300 is shown. The illustrated instruction 300 is a Complex Instruct Set Command (CISC) instruction having a format compatible with the CISC instruction format used in the X86 processor platform offered by the Intel Corporation. The instruction 300 includes a prefix portion 301, an opcode portion 303 defining the logical or mathematical command to be performed by the execution unit 221, an addressing mode specifier field 305, an address displacement field 307 and an immediate data field 309. As is well known CISC instructions have a variable length and some opcodes may not utilize all the instruction fields. Intel CISC instruction format and programming is very well known in the art and is more fully defined in the Intel i486 Microprocessor Programmer's Reference Manual.

In accordance with the present invention, the instruction processing unit 201 includes an instruction type detector 211 for detecting whether an instruction type preceding an opcode is a repeatable type. For example, if the prefix portion 301 indicates a repeat operation, the instruction type detector 211 is adapted to detect such repeat instruction type. The E-unit 217 is responsive to a detection by the instruction type detector 211 of a repeatable instruction type for generating a signal 215 indicating that the data associated with the detected instruction is not to be placed into the cache. The Does Not Cache (DNC) indicator signal 215 is outputted from the CPU via the bus interface unit 223. The CPU 103 further includes a set of buffers 219 which are used as a temporary storage area to perform operations on the data associated with the detected repeat instructions. In the preferred embodiment, the set of buffers 219 comprises two such buffers, each of which has the width of two cache lines. That is, the width of each buffer of the buffer set 219 is equal to twice the size of the cache line. This buffer size ensures that all data involved in such repetitive move instructions is properly stored.

The DNC signal is used with the fetching or storing of data or instructions which are cacheable (not memory mapped I/O or such) and simply indicates to all caches to handle the data in the following manner:<img file="EP0779581A2_D0001.tif" />

Repeat (REP) instructions are a class of instructions whereby the instruction is automatically repeated a pre-determined number of times.

The number of times the instruction is to be re-executed is determined at the time the instruction is executed by the content of a specific general purpose register (GPR) within the CPU. This specific GPR is referred to as the count register (CREG).

In the preferred embodiment, the instructions which can be repeated are the following:

o Compare String (CMPS)

Compares the string starting at the memory address pointed to by the source register to the string starting at the memory address pointed to by the destination register. Depending on the specific op-code of the instruction, the compare can be on a byte, word, or double word basis. For the REP operation the source and destination address are incremented and the next sequential byte, word or double word are compared. This operation is repeated the number of times indicated by the CREG.

o Scan String (SCAS)

Compares the contents of the source register to the string starting at the memory address pointed to by the destination register. Depending on the specific op-code of the instruction, the compare can be on a byte, word, or double word basis. For the REP operation the destination address is incremented and the next sequential byte, word or double word is compared. This operation is repeated the number of times indicated by the CREG.

o Load String (LODS)

Loads the destination register with the string starting at the memory address pointed to by the source register. Depending on the specific op-code of the instruction, the load can be on a byte, word, or double word basis. For the REP operation the source address is incremented, and the next sequential byte, word or double word is loaded. This operation is repeated the number of times indicated by the CREG.

o Store String (STDS)

Stores the contents of source register at the memory address pointer to by the destination register. Depending on the specific op-code of the instruction, the store can be on a byte, word, or double word basis. For the REP operation the destination address is incremented, and the contents of the source register are stored at the next sequential byte, word or double word address of memory. This operation is repeated the number of times indicated by the CREG.

o OUT String (OUTS)

Stores the contents of the memory location pointed to by the source register to the I/O port address pointed to by the destination register. Depending on the specific op-code of the instruction, the store can be on a byte, word, or double word basis. For the REP operation the source address is incremented, and the next sequential byte, word or double word of memory is stored to the I/O port address pointed to by the destination register. The operation is repeated the number of times indicated by the CREG.

o IN String (INS)

Loads the data from the I/O port address pointed to by the source register to the memory location pointed to by the destination register. Depending on the specific op-code of the instruction, the store can be on a byte, word, or double word basis. For the REP operation the destination address is incremented, and the next sequential byte, word or double word of memory is loaded from the I/O port address. The operation is repeated the number of times indicated by the CREG.

o MOVE String (MOVS)

Moves the data at the memory location pointed to by the contents of the source register to the memory location pointed to by the destination register. Depending on the specific op-code of the instruction, the store can be on a byte, word, or double word basis. For the REP operation the source and destination address are incremented, and the next sequential byte, word or double word of memory is moved. The operation is repeated the number of times indicated by the CREG.

A pseudo code implementation is presented below for each of the above defined REP instructions:

Initially, for all of the following operations, the CREG has been loaded with the number of times the repeat operation is to be executed, at some time prior to the execution of the instruction which is to be repeated. For all the following operations the two buffers 219 are referenced as A buffer and B buffer.<img file="EP0779581A2_D0002.tif" /><img file="EP0779581A2_D0003.tif" /><img file="EP0779581A2_D0004.tif" /><img file="EP0779581A2_D0005.tif" />

Decrement the CREG IF not zero repeat the operation When the CREG is equal to zero the REP MOVS is completed at that time a data store of the B buffer is issued by the execution unit. simultaneously the DNC signal is set. When the store is complete, the DNC signal is reset.

It may be appreciated that this approach eliminates the problem of cache purging during repeat instructions by preventing the repeat instructions from impacting the cache (or caches for systems having multiple levels of CPU caches). The Does Not Cache (DNC) signal is unique in that when data is transferred, the data is to be treated as if it were cacheable. Data is transferred a cache line at a time. However when the DNC signal is set caches will not load or store data into the cache consistent with maintaining data coherency within the system.

As described above, this is accomplished this by using a new signal referred to as Does Not Cache (DNC) . DNC is an indicator to the caches that this reference is to be handled as a normal cache access, except that the data associated with this request is not to be placed into the cache. This is different from a request which is non-cacheable. Non-cacheable data is data which should not be placed into a cache. Memory mapped I/O is an example of such data. Non-cacheable data is typically handled differently by the system. Typically when non-cacheable data is fetched it is not fetched as a cache line. Also non-cacheable data is typically handled differently as regards to cache snooping. In the present embodiment, all repeat instruction types are designated as non-cacheable. That is, the data associated with such instruction type is non-cacheable. This designation has been made to avoid casting out valuable cache data and replacing it with useless data involved in a repetitive move instruction. Instead the set of buffers 219 are used as temporary storage space for storing the data associated with the repetitive command type. As such, each buffer is designed to have a width equal to a multiple of the data width of the CPU bus.