Patent Grant
7228386
1. Field of the Invention
This invention is related to the field of caching and, more particularly, to mechanisms for disabling a portion of a cache.
2. Description of the Related Art
Processors typically provide a set of registers which may be used by programs as a high speed, private storage for operands. The operands stored in registers may be frequently accessed variables, or may be intermediate results in a complex, multi-instruction calculation. Unfortunately, for many tasks, the number of registers provided in the processor may be too few to hold all of the operands of interest. In such cases, many of the frequently accessed variables and/or intermediate results are written to and read from memory locations during execution of the program.
The memory may have a significantly higher latency than the registers, limiting the speed at which the program may be executed as compared to the speed that may be achieved if all operands were in registers. Processors and/or the computer systems including the processors may provide caches to alleviate the memory latency. Generally, a cache is a relatively small, high speed memory which may store copies of data corresponding to various recently-accessed memory locations. Generally, cache storage is allocated and deallocated in units of cache lines (a group of bytes from contiguous memory locations). In other words, the cache may include multiple entries, and each entry may include storage for a cache line of bytes. If requested data for an access is not in the cache (a “miss”), an entry is allocated for the cache line including the requested data and the cache line is filled into the allocated entry. In some cases, a portion of the cache line (often called a “sector”) may be valid while other portions are invalid. However, the entire cache entry is allocated for the cache line if one or more of the sectors are valid.
While caches may reduce the average latency for memory accesses, the latency for any particular access is generally not deterministic. If the requested data is stored in the cache (a “hit”), then the latency is low. If the requested data is a miss in the cache, on the other hand, then the latency is high. Accesses may miss the cache for a variety of reasons. For example, an access misses the cache if the cache line including the requested data has not been accessed before. A current access may miss even if the cache line has been previously accessed, if other cache lines which map to the same entry (or set of entries, depending on the cache structure) are accessed in the interval between the previous access and the current access. In multi-master coherent environments, coherency activity due to accesses by other masters may cause cache lines to be invalidated in the cache. These various causes of cache misses may be hard to predict, and thus managing the cache to provide deterministic, low latency access to certain operands may be difficult to achieve. Accordingly, a mechanism for providing deterministic, low latency access to operands is desired.
The problems outlined above are in large part solved by a cache in accordance with the present disclosure. The cache may be programmed to disable one or more entries from allocation for storing memory data (e.g. in response to a memory transaction which misses the cache). Furthermore, the cache may be programmed to select which entries of the cache are disabled from allocation. Since the disabled entries are not allocated to store memory data, the data stored in the entries at the time the cache is programmed to disable the entries may remain in the cache. The programmable disable of entries may have a variety of uses.
For example, in one embodiment, the disabled entries may be used as a private scratch pad memory. In one specific implementation, the cache also provides for direct access to entries in response to direct access transactions. Thus, the cache may provide a private scratch pad memory by disabling the one or more entries and using the direct access transactions to directly access the disabled entries for scratch pad reads and writes. Since the entries are disabled for use in the fill and eviction of memory data, these activities may not affect the storage of scratch pad data in the disabled entries. In one embodiment, the direct access transactions are specified via a direct access address space separate from the memory address space. Thus, coherency activity in the memory space may not affect the data in the direct access address space. Scratch pad (direct access) read and write transactions may experience a deterministic, low latency (the latency of a cache hit). For example, operands which are frequently accessed and/or intermediate results may be stored in the scratch pad memory.
Additionally, since the disabling of entries is programmable, the size of the scratch pad memory may be varied to suit the performance needs of the programs being executed. If no scratch pad memory is needed, the entire cache may be used to cache memory data. If scratch pad memory is needed, the number of entries may be selected to provide enough scratch pad memory while leaving the remainder of the cache available to cache memory data. In one implementation, the cache is set associative and one or more ways of the cache may be disabled. Accordingly, a way in which a manufacturing defect is detected may be disabled, and the remaining ways may still be used to store memory data. However, if the defect affects only some of the entries within the disabled way, the remaining entries may be used as scratch pad memory.
While the programmable disable of entries may be used for scratch pad operations, other uses are contemplated as well. For example, the programmable disable of entries may be used for testing associative embodiments. Each of the entries which may be eligible to store data for a given transaction may be disabled except for one entry to be tested. The given transaction may be performed, loading test data into the test entry. The test data may then be read to ensure that the test entry is functioning properly.
Furthermore, the programmable disable of entries may be used to lock memory data into the cache. For example, only the entries which are to be locked may be enabled, memory transactions accessing the data to be locked into the cache may be performed (causing the data to be loaded into the entries, assuming the data misses in the cache), then the entries may be disabled (and other entries enabled). Since the entries are disabled from allocation in response to subsequent cache misses, the data may be locked into the disabled entries. Memory reads and writes to the data may be detected as hits in the cache.
Broadly speaking, a cache is contemplated. The cache includes a memory and a control circuit coupled to the memory. The memory has a plurality of entries, each of the plurality of entries being capable of storing a cache line of data. The control circuit is programmably configurable to disable allocation of one or more entries of the plurality of entries for storing data corresponding to memory transactions. The one or more entries are programmably selectable from the plurality of entries.
Additionally, a system is contemplated comprising a processor, a memory controller, and a cache. The processor is configured to initiate transactions. Coupled to receive the transactions, the memory controller is capable of being coupled to a memory. The cache, also coupled to receive the transactions, includes a plurality of entries. Each of the plurality of entries is capable of storing a cache line of data. The cache is programmably configurable to disable allocation of one or more entries of the plurality of entries to store data corresponding to the transactions which target a location within the memory. The one or more entries are programmably selectable from the plurality of entries.
Moreover, a method is contemplated. One or more entries in a cache are programmably disabled from allocation to store data corresponding to memory transactions. The one or more entries are programmably selectable from a plurality of entries in the cache.
Other objects and advantages of the invention will become apparent upon reading the following detailed description and upon reference to the accompanying drawings in which:
While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the drawings and detailed description thereto are not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
Turning now to
Generally, L2 cache 14 may support the programmable disabling of one or more entries from allocation in response to memory transactions which miss L2 cache 14. Furthermore, L2 cache 14 may support the selection of which of the cache entries are disabled. For example, L2 cache 14 may be set associative and may support the disabling of one or more ways of the cache. Alternatively, L2 cache 14 may support disabling of one or more indexes of one or more ways of the cache. L2 cache 14 may include one or more registers, accessible via a configuration transaction, which may be programmed with information indicating which entries of L2 cache 14 are disabled from allocation in response to memory transactions.
The disabled entries may be used as a scratch pad memory by processors 12A-12B. A scratch pad memory is a private memory used by the processors, similar to the fashion in which registers operate. For example, in an embodiment in which one or more of the ways may be disabled, a scratch pad may be established by disabling the one or more ways (and enabling the remaining ways). Since the selected ways are disabled (and thus the data stored therein may not be dislodged due to allocation of entries for memory transactions which miss), the scratch pad data may be unaffected by normal cache replacement activities. Thus, the scratch pad data may remain stored in L2 cache 14 even as other data is filled into and evicted from the remaining ways. Direct access transactions, described below, may be used to access the scratch pad data.
Additionally, since the entries which are disabled are programmable, the size of the scratch pad memory may be varied to suit the performance needs of the programs being executed. If no scratch pad memory is needed, the entire L2 cache 14 may be used to cache memory data. If scratch pad memory is needed, the number of entries may be selected to provide enough scratch pad memory while leaving the remainder of the L2 cache 14 available to cache memory data.
The programmable disable of selected entries in the cache may have a variety of uses in addition to the scratch pad use described above. For example, the programmable disable may be used to facilitate testing. To test a particular entry, other entries which could be allocated for data corresponding to a test transaction may be disabled. The test transaction may be performed, and the data is stored into the particular entry. The test data may then be read to ensure it was stored in the correct entry. The other entries could be tested in a similar manner. As another example, the programmable disable may be used for general locking of data into L2 cache 14 to provide fairly deterministic, low latency access to certain memory data locked into the L2 cache.
Processors 12A-12B and/or the I/O bridges/interfaces may initiate transactions on the bus 24. Transactions may include read transactions (transfers of data to the transaction initiator) and write transactions (transfers of data from the transaction initiator). Transactions may further include various coherency commands (e.g. an invalidate command) which may or may not involve a transfer of data.
System 10 supports various types of transactions on bus 24. Memory transactions are transactions which target a memory location. Additionally, system 10 may support direct access transactions to L2 cache 14. As used herein, the term “direct access transaction” refers to a transaction which targets an entry in L2 cache 14 and explicitly specifies that entry. If the direct access transaction is a read, the data in the specified entry is returned by L2 cache 14. If the direct access transaction is a write, the data provided in the write transaction is stored in the entry. System 10 may further support I/O transactions and configuration transactions on bus 24.
An addressable range of system 10 is defined by the size of the addresses which may be transmitted on bus 24. The addressable range may be divided into several address spaces including a memory address space, a direct access address space, an I/O address space, and various configuration address spaces. In this embodiment, the address space which includes the address of a transaction may identify the type of the transaction. Thus, an address within the memory address space indicates that the transaction is a memory transaction and the address identifies a targeted memory location in memory 26 (and thus memory controller 16 and L2 cache 14 may respond to the transaction). An address within the I/O memory mapped address space indicates that the transaction is an I/O transaction and the address targets an I/O device on one of the I/O interfaces. An address within the configuration address space indicates that the transaction is a configuration transaction and the address targets various configuration registers in devices within system 10.
An address within the direct access address space indicates that the transaction is a direct access transaction and the address specifies the entry of L2 cache 14 to be directly accessed by the direct access transaction. For memory transactions, the entry (if any) of L2 cache 14 accessed to store or return data for the memory transaction is determined by comparing the tags of cache lines stored in the entry to the address of the memory transaction. On the other hand, direct access transactions explicitly specify the entry to be accessed. L2 cache 14 may access the specified entry without regard to the tag comparisons normally used to select the accessed entry.
While the above description uses address spaces to determine transaction type, other embodiments may determine transaction type in other ways. For example, command type encodings could be provided on control signals transmitted during the address portion of the transaction.
In one embodiment, direct access transactions may be used to access the disabled entries in L2 cache 14. Thus, in an embodiment for using a portion of L2 cache 14 as a scratch pad, the scratch pad may be established by disabling the one or more entries to be used as a scratch pad. Processors 12A-12B may perform direct access transactions to read and write scratch pad entries (including initializing the scratch pad with data, if desired). Since the normal process of filling memory transaction data into cache lines of L2 cache 14 does not affect the disabled entries, the scratch pad data is not evicted from L2 cache 14. Furthermore, since the direct access transactions are not in the memory address space, coherency activity in the memory address space does not affect the transactions (nor the data stored in the scratch pad entries, if they were written using direct access transactions). Accordingly, a deterministic latency (the latency of an L2 cache hit) may be achieved for scratch pad accesses.
It is noted that the programmable disable of cache ways may be used to disable a way or ways that have a manufacturing defect. In this manner, failure due to data corruption in the defective entries may be avoided while still permitting use of the non-defective ways. However, the defect may affect only some of the entries within the disabled way. Accordingly, functional entries within the disabled way may be used as scratch pad memory. More efficient use of the non-defective entries may thereby be achieved.
The scratch pad memory may be initialized in any desired fashion. For example, the random data that may be in the scratch pad memory when the entries are disabled may be used. In another alternative, direct access write transactions may be used to initialize the scratch pad memory locations.
Processors 12A-12B may be designed to any instruction set architecture, and may execute programs written to that instruction set architecture. Exemplary instruction set architectures may include the MIPS instruction set architecture (including the MIPS-3D and MIPS MDMX application specific extensions), the IA-32 or IA-64 instruction set architectures developed by Intel Corp., the PowerPC instruction set architecture, the Alpha instruction set architecture, the ARM instruction set architecture, or any other instruction set architecture.
L2 cache 14 is a high speed cache memory. L2 cache 14 is referred to as “L2” since processors 12A-12B may employ internal level 1 (“L1”) caches. If L1 caches are not included in processors 12A-12B, L2 cache 14 may be an L1 cache. Furthermore, if multiple levels of caching are included in processors 12A-12B, L2 cache 14 may be a lower level cache than L2. L2 cache may employ any organization, including direct mapped, set associative, and fully associative organizations. In one particular implementation, L2 cache 14 may be a 512 kilobyte, 4 way set associative cache having 32 byte cache lines. A set associative cache is a cache arranged into multiple sets, each set comprising two or more entries. A portion of the address (the “index”) is used to select one of the sets (i.e. each encoding of the index selects a different set). The entries in the selected set are eligible to store the cache line accessed by the address. Each of the entries within the set is referred to as a “way” of the set. The portion of the address remaining after removing the index (and the offset within the cache line) is referred to as the “tag”, and is stored in each entry to identify the cache line in that entry. The stored tags are compared to the corresponding tag portion of the address of a memory transaction to determine if the memory transaction hits or misses in the cache, and is used to select the way in which the hit is detected (if a hit is detected).
Memory controller 16 is configured to access memory 26 in response to memory transactions received on bus 24. Memory controller 16 receives a hit signal from L2 cache 14, and if a hit is detected in L2 cache 14 for a memory transaction, memory controller 16 does not respond to that memory transaction. If a miss is detected by L2 cache 14, or the memory transaction is non-cacheable, memory controller 16 may access memory 26 to perform the read or write operation. Memory controller 16 may be designed to access any of a variety of types of memory. For example, memory controller 16 may be designed for synchronous dynamic random access memory (SDRAM), and more particularly double data rate (DDR) SDRAM. Alternatively, memory controller 16 may be designed for DRAM, Rambus DRAM (RDRAM), SRAM, or any other suitable memory device.
High speed I/O interface 18 may be an interface to a high speed I/O interconnect. For example, high speed I/O interface 18 may implement the Lightning Data Transport (LDT) I/O fabric developed by Advanced Micro Devices, Inc. Other high speed interfaces may be alternatively used.
I/O bridge 20 is used to link one or more I/O interfaces (e.g. I/O interfaces 22A-22B) to bus 24. I/O bridge 20 may serve to reduce the electrical loading on bus 24 if more than one I/O interface 22A-22B is bridged by I/O bridge 20. Generally, I/O bridge 20 performs transactions on bus 24 on behalf of I/O interfaces 22A-22B and relays transactions targeted at an I/O interface 22A-22B from bus 24 to that I/O interface 22A-22B. I/O interfaces 22A-22B may be lower bandwidth, higher latency interfaces. For example, I/O interfaces 22A-22B may include one or more serial interfaces, Personal Computer Memory Card International Association (PCMCIA) interfaces, Ethernet interfaces (e.g. media access control level interfaces), Peripheral Component Interconnect is (PCI) interfaces, etc.
Bus 24 may have any suitable protocol and construction. According to one implementation, bus 24 may be a split transaction bus. The implementation may include an address bus supporting up to 40 bits of addressable range and a data bus capable of transmitting one cache line per clock cycle (e.g. 32 bytes). Other widths for either the address bus or data bus are possible and contemplated. The bus may also include transactions to support maintaining memory coherency (e.g. an invalidate command). The bus may use any suitable signalling technique (e.g. differential or non-differential signalling).
It is noted that system 10 (and more particularly processors 12A-12B, L2 cache 14, memory controller 16, I/O interfaces 18 and 22A-22B, I/O bridge 20, and bus 24 may be integrated onto a single integrated circuit as a system on a chip configuration. In another configuration, memory 26 may be integrated as well. Alternatively, one or more of the components may be implemented as separate integrated circuits, or all components may be separate integrated circuits, as desired. Any level of integration may be used.
As used herein, a transaction “targets” a location or device if the location or device is the provider of data for the transaction (for a read transaction) or receiver of data for the transaction (for a write transaction). Viewed in another way, a transaction may target a location or device if the address of the transaction is mapped to that location or device.
Turning now to
Generally, tags memory 30 stores the tag information for each entry in L2 cache 14, and data memory 32 stores the cache line of data for each entry in L2 cache 14. Thus, an entry comprises a tag memory storage location and a data memory storage location. The tag memory storage location stores the tag for the entry (and possibly other information, such as validity and dirty information). For example, in one implementation, a tag for an entry includes the tag portion of the address (e.g. bits 39:17 in the illustrated embodiment), a valid bit indicating the validity or lack of validity of the entry, and a dirty bit indicating the dirty or clean state of the cache line. A cache line is dirty if at least one byte of the cache line has been modified in L2 cache 14 and the modification has not been written to memory 26. The data memory storage location stores the cache line of data in the entry.
During memory transactions, the index portion of the input address (e.g. bits 16:5 in the illustrated embodiment) is provided to tags memory 30 and data memory 32. Each memory selects a set of storage locations in response to the index. Tags memory 30 outputs the tag from each selected storage location to comparators 34A-34D, which compare the tags to the tag portion of the input address. If the tag compares equally, and the entry is valid, the corresponding comparator 34A-34D may signal a hit to control circuit 36. Control circuit 36 may assert the hit signal to memory controller 16 if any of the comparators 34A-34D indicates a hit, and may determine a way selection from the output of comparators 34A-34D as well. Each comparator 34A-34D is comparing the tag from a different way of L2 cache 14, and thus the comparator output indicating a hit is an indication of the hitting way. Control circuit 36 provides the way selection to tags memory 30 and data memory 32.
Data memory 32 provides data from the storage location at the selected index and way, or receives data into that storage location, depending upon whether the memory transaction is a read or write transaction. Control circuit 36 receives the control information corresponding to the transaction, and generates control signals for data memory 32 and tags memory 30. The control signals may include, for example, an enable signal and a read/write signal to indicate whether the memory is to read or write.
Tags memory 30 may be updated in response to the transaction as well. For example, if the transaction is a write, the dirty bit in the hitting entry may be updated.
Additionally, if the transaction is a miss in L2 cache 14 and is cacheable, L2 cache 14 may select a way for eviction to receive a line fill of the missing line. The evicted way may be provided as the way selection, and the valid bit in the evicted entry may be cleared to invalidate the cache line. The cache line may also be provided from data memory 32 for writing to memory 26 if the cache line is dirty. It is noted that tag updates and evictions may not occur in the same clock cycle that L2 cache 14 is accessed to determine the hit/miss of the input address, in some embodiments.
Way enable register 42 stores an indication of which ways are enabled for allocation for memory transactions and which ways are disabled. Thus, when control circuit 36 is selecting a way for eviction in response to a cache miss, the disabled ways are not selected. In this manner, the data in the disabled ways is not evicted, and the entries in the disabled ways are not allocated to store data corresponding to the memory transactions which miss. The disabled ways may be accessed using memory transactions which hit the data in the disabled ways or using direct access transactions, depending on the embodiment. Direct access transactions are described in more detail below.
L2 cache 14 may employ any eviction/replacement strategy. For example, if the least-recently used (LRU) strategy is used, the disabled ways in each set may be prevented from being marked as least recently used (e.g. they may be forced to most recently used). Thus, an enabled way is always marked as least recently used and is selected for eviction. If a random replacement strategy is used, the random selector may be managed so as not to generate the disabled ways. For example, if a counter is used, the counter may be incremented after each access or each replacement. The counter may be modified to skip the disabled ways when being incremented. More elaborate random generators (e.g. a linear feed back shift register, or LFSR) may similarly be prevented from generating a way selection for the disabled ways.
During direct access transactions, control circuit 36 receives an assertion of the direct access signal from decoder 38. Decoder 38 decodes the address of the transaction to detect that the address is in the direct access address space, and asserts the direct access signal if the address is in the direct access address space. For example, in the illustrated embodiment, the direct access address space may be the addresses having a predetermined combination of address bits 39:27. In one particular implementation, the combination of bits 39:27 may be (in hexadecimal) 00D and bit 27 may be 0. It is noted hat the address space may be made smaller by including more of the most significant address bits to decoder 38 (e.g. additional bits below bit position 27) or larger by decoding fewer bits. Furthermore, the selection of an address range for the direct access address space is arbitrary and may be located anywhere in the addressable range, as desired.
In response to the direct access signal, control circuit 36 ignores the hit signals from comparators 34A-34D. The direct access transaction is treated as a cache hit. The entry to be accessed (read or written) is explicitly identified by the direct access transaction. For example, in the present embodiment, address bits other than those used to identify the direct access address space are used to supply the index and way of the entry. More particularly in the illustrated embodiment, the same address bits used to index the cache in memory transactions are used to supply the index in a direct access transaction. In this manner, additional hardware to provide the index to tags memory 30 and data memory 32 for direct access transactions may be avoided. The way is supplied in other address bits (e.g. more significant address bits than the index). In the illustrated embodiment, the more significant address bits contiguous to the index bits are used to convey the way selection (e.g. address bits 18:17 provided to control circuit 36). Control circuit 36 provides the way selection indicated by the direct access transaction as the way selection to tags memory 30 and data memory 32.
It is noted that, while direct access transactions are illustrated as providing the index and way in the address in the present embodiment, other embodiments are contemplated. For example, the index and way may be provided on control signals separate from the address. Alternatively, each entry in L2 cache 14 may be assigned an entry number, and the entry number may be communicated by direct access transactions (e.g. as part of the address or on control signals). Still further, direct mapped implementations may use only an index (or entry number) and fully associative embodiments may use only a way (or entry number). Any method for explicitly communicating the selected entry may be used in various implementations of a direct access transaction.
Since the hit signals from comparators 34A-34D are ignored by control circuit 36 for direct access transactions, the tag of the selected entry need not match the tag portion of the input address for direct access transactions. If the direct access transaction is a read, the data in the selected entry is provided regardless of any tag match or mismatch. If the direct access transaction is a write, the data provided in the direct access transaction is written to the selected entry, overwriting the data currently stored in the entry. For direct access write transactions, control circuit 36 may update the tag in the selected entry with the tag portion of the input address. In this manner, if the entry was previously storing valid memory transaction data, the entry will be a miss for memory transactions affecting that memory transaction data. Additionally, bits in the tag portion of the address may be used to specify the desired state of the valid and dirty bits in the tag. These bits may be written to the valid and dirty bits in the tag of the specified entry. Additionally, since the direct access transactions are treated as cache hits, there may be no eviction of a cache line in response to direct access transactions.
Control circuit 36 may assert the hit signal to memory controller 16 in response to direct access transactions, if desired. Alternatively, since the illustrated embodiment employs a separate address space for memory transactions and direct access transactions, memory controller 16 may ignore direct access transactions on bus 24.
Decoder 40 is provided to detect a write to way enable register 42. Way enable register 42 may be assigned an address within a configuration address space, for example. More particularly, in one embodiment, way enable register 42 may be assigned an address range. The most significant address bits which define the range may be decoded by decoder 40 (e.g. bits 39:12 in the illustrated embodiment) to provide a write enable to way enable register 42. The remaining address' bits (e.g. bits 11:8 in the illustrated embodiment) may be used to provide the value to store into way enable register 42. For example, the illustrated embodiment uses a bit for each way. The bit may be set to indicate that the way is enabled and clear to indicate that the way is disabled (or vice versa). In this manner, any number of ways may be enabled or disabled and the ways may be individually enabled and disabled. Thus, a large degree of flexibility in the enabling and disabling of ways may be provided. More particularly, if all the bits are set, then all ways are enabled. Other encodings disable various ways. While the above description detects a write to way enable register 42 via the address and draws the value to be written from the address, other embodiments are possible. For example, the data may be provided as data in the write transaction, on separate control signals from the address, etc. Any suitable manner for communicating a value to way enable register 42 may be used.
It is noted that the embodiment illustrated in
It is noted that, while tags memory 30 and data memory 32 are illustrated separately in
It is noted that, while the embodiment illustrated in
It is noted that
As used herein, the term “register” refers to any clocked storage device. For example, a register may be a flip-flop (or “flop”), a latch, a series of latches, a clocked random access memory (RAM) cell, or any other desirable storage.
Turning now to
Memory transaction address 50 includes an offset field 54, an index field 56, and a tag field 58. Offset field 54 includes the bits defining the offset within a cache line. Index field 56 includes the bits used to index L2 cache 14 (e.g. bits 16:5 in the embodiment illustrated in
Direct access transaction address 52 includes offset field 54 and index field 56, similar to memory transaction address 50. Additionally, direct access transaction address 52 includes a way field 60 identifying the way to be accessed in response to the direct access transaction (e.g. bits 18:17 in the embodiment illustrated in
It is noted that the bits comprising offset field 54 may not actually be transmitted on one embodiment of bus 24. Instead, byte enables may be transmitted indicating which bytes are accessed by the transaction. The byte enable for the byte identified by offset field 54 is set, and additional byte enables based on the size of the transaction may be set.
Turning now to
If the read transaction is a direct access transaction (decision block 70), L2 cache 14 provides the data from the selected entry in response to the read transaction (block 72).
The selected entry is identified by the explicit index and way provided by the direct access transaction. Whether or not the tag portion of the address in the selected entry matches the address of the read transaction does not affect the forwarding of data, and no cache eviction may occur.
If the read transaction is a memory transaction (“no” leg of decision block 70), L2 cache 14 determines if the address hits (decision block 74). If the address hits, the data from the hitting entry is provided in response to the read transaction (block 76). If the read transaction is a miss, L2 cache 14 selects an entry for eviction of the cache line stored therein, to be replaced by the cache line accessed by the read transaction (block 78). Entries in disabled ways are inhibited from being selected.
Turning now to
If the write transaction is a direct access transaction (decision block 80), L2 cache 14 stores the data included in the write transaction into the selected entry (block 82). The selected entry is identified by the explicit index and way provided by the direct access transaction. Whether or not the tag portion of the address in the selected entry matches the address of the write transaction does not affect the updating of the selected entry. Furthermore, if valid data is stored in the entry, that data is overwritten (even if the tag does not match the address of the write transaction). Additionally, the tag of the selected entry may be updated with the corresponding portion of the address of the write transaction (block 84). In this manner, the entry may not be affected by coherency activity in the memory address space.
If the write transaction is a memory transaction (“no” leg of decision block 80), L2 cache 14 determines if the address hits (decision block 86). If the address hits, the data included in the write transaction is written to the hitting entry (block 88). If the write transaction is a miss, L2 cache 14 selects an entry for eviction of the cache line stored therein, to be replaced by the cache line accessed by the write transaction (block 90). Entries in disabled ways are inhibited from being selected.
While the embodiment illustrated via
In
In
More particularly, the disabled entries may be the entries within a disabled way, as indicated by way enable register 42 and at a disabled index, as indicated by index enable register 100. In the example, the upper half of the indexes are controlled by one bit of index enable registers 100 and the lower half of the indexes are controlled by the other bit. Thus, the upper half of the indexes are enabled, and the lower half of the indexes are disabled. Additionally, in the example, way 3 is disabled. Thus, the entries within the lower half of the indexes within way 3 are disabled and may be used as scratch pad. All indexes of ways 0, 1, and 2 are enabled and the upper half of the indexes of way 3 are enabled for storing memory transaction data. In other embodiments, the index disable may be more granular (e.g. there may be more bits in index enable registers 100 corresponding to more groups of indexes).
Index enable register 100 may be included in an embodiment of L2 cache 14 similar to the embodiment shown in
Turning next to
A write to way enable register 42 is performed to enable the ways to be used to lock the data into the cache (referred to as the “selected ways”) and to disable the remaining ways (referred to as the “remaining ways”—block 110). There may be one or more selected ways, and one or more remaining ways.
One or more memory transactions are performed (block 112). Each memory transaction targets data to be loaded into the selected ways. Since the remaining ways are disabled, an entry in one of the selected ways is allocated to store the data corresponding to each memory transaction (assuming the memory transactions miss L2 cache 14). Either read or write transactions may be performed, as desired. Generally, data may be loaded into each entry of each selected way, or into a number of entries as may be desired.
Another write to way enable register 42 is performed to disable the selected ways and enable the remaining ways (block 114). Thus, the data loaded into the selected ways may be locked into the selected ways by disabling allocation of entries within the selected ways for subsequent cache misses. The remaining ways may be allocated for subsequent cache misses, since they are enabled. The data in the selected ways may be accessed by performing additional memory transactions to the addresses corresponding to the data (block 116). These additional memory transactions may hit in entries within the disabled ways, and the data may be provided from (for a read) or updated into (for a write) the hitting entries.
While the above discussion uses selected ways and remaining ways, other embodiments may use any other groupings of entries supported by the programmable disabling function. For example, embodiments employing index register 100, a combination of index register 100 and way enable register 42, or other methods for indicating the disabled entries may be used.
Turning next to
Generally, the database of system 10 carried on carrier medium 300 may be a database which can be read by a program and used, directly or indirectly, to fabricate the hardware comprising system 10. For example, the database may be a behavioral-level description or register-transfer level (RTL) description of the hardware functionality in a high level design language (HDL) such as Verilog or VHDL. The description may be read by a synthesis tool which may synthesize the description to produce a netlist comprising a list of gates in a synthesis library. The netlist comprises a set of gates which also represent the functionality of the hardware comprising system 10. The netlist may then be placed and routed to produce a data set describing geometric shapes to be applied to masks. The masks may then be used in various semiconductor fabrication steps to produce a semiconductor circuit or circuits corresponding to system 10. Alternatively, the database on carrier medium 300 may be the netlist (with or without the synthesis library) or the data set, as desired.
While carrier medium 300 carries a representation of system 10, other embodiments may carry a representation of any portion of system 10, as desired, including one or more of L2 cache 14, memory controller 16, portions of L2 cache 14 (e.g. way enable register 42, index enable register 100, decoder 40, control circuit 36, and/or other elements illustrated in
Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.
This application is a continuation of and claims priority to U.S. patent application having an application Ser. No. 09/633,683; filed Aug. 7, 2000 now U.S. Pat. No. 6,848,024, which application is hereby incorporated by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 3693165 | Reiley et al. | Sep 1972 | A |
| 4044338 | Wolf | Aug 1977 | A |
| 4453212 | Gaither et al. | Jun 1984 | A |
| 4511994 | Webb et al. | Apr 1985 | A |
| 4575792 | Keeley | Mar 1986 | A |
| 4633440 | Pakulski | Dec 1986 | A |
| 4654778 | Chiesa et al. | Mar 1987 | A |
| 4807115 | Torng | Feb 1989 | A |
| 4833642 | Ooi | May 1989 | A |
| 4858105 | Kuriyama et al. | Aug 1989 | A |
| 4907278 | Cacinati et al. | Mar 1990 | A |
| 4996641 | Talgam et al. | Feb 1991 | A |
| 5019971 | Lefsky et al. | May 1991 | A |
| 5113514 | Albonesi et al. | May 1992 | A |
| 5125083 | Fite et al. | Jun 1992 | A |
| 5163142 | Mageau | Nov 1992 | A |
| 5193163 | Sanders et al. | Mar 1993 | A |
| 5226126 | McFarland et al. | Jul 1993 | A |
| 5226130 | Favor et al. | Jul 1993 | A |
| 5241663 | Rohwer | Aug 1993 | A |
| 5303362 | Butts, Jr. et al. | Apr 1994 | A |
| 5317716 | Liu | May 1994 | A |
| 5325504 | Tipley et al. | Jun 1994 | A |
| 5353425 | Malamy et al. | Oct 1994 | A |
| 5369753 | Tipley | Nov 1994 | A |
| 5377345 | Chang et al. | Dec 1994 | A |
| 5410669 | Biggs et al. | Apr 1995 | A |
| 5414827 | Lin | May 1995 | A |
| 5416783 | Broseghini et al. | May 1995 | A |
| 5432918 | Stamm | Jul 1995 | A |
| 5450551 | Amini et al. | Sep 1995 | A |
| 5471598 | Quattromani et al. | Nov 1995 | A |
| 5487162 | Tanaka et al. | Jan 1996 | A |
| 5493667 | Huck et al. | Feb 1996 | A |
| 5510934 | Brennan et al. | Apr 1996 | A |
| 5524212 | Somani et al. | Jun 1996 | A |
| 5526510 | Akkary et al. | Jun 1996 | A |
| 5539878 | Kikinis | Jul 1996 | A |
| 5551001 | Cohen et al. | Aug 1996 | A |
| 5557763 | Senter et al. | Sep 1996 | A |
| 5564034 | Miyake | Oct 1996 | A |
| 5584014 | Nayfeh et al. | Dec 1996 | A |
| 5586253 | Green et al. | Dec 1996 | A |
| 5586293 | Baron et al. | Dec 1996 | A |
| 5588126 | Abramson et al. | Dec 1996 | A |
| 5592679 | Yung | Jan 1997 | A |
| 5638537 | Yamada et al. | Jun 1997 | A |
| 5644752 | Cohen et al. | Jul 1997 | A |
| 5668815 | Gittinger et al. | Sep 1997 | A |
| 5668972 | Liu et al. | Sep 1997 | A |
| 5671444 | Akkary et al. | Sep 1997 | A |
| 5748640 | Jiang et al. | May 1998 | A |
| 5752261 | Cochcroft, Jr. | May 1998 | A |
| 5761712 | Tran et al. | Jun 1998 | A |
| 5768555 | Tran et al. | Jun 1998 | A |
| 5784588 | Leung | Jul 1998 | A |
| 5787490 | Ozawa | Jul 1998 | A |
| 5802338 | Rechtschaffen et al. | Sep 1998 | A |
| 5809528 | Miller et al. | Sep 1998 | A |
| 5809531 | Brabandt | Sep 1998 | A |
| 5875465 | Kilpatrick et al. | Feb 1999 | A |
| 5897651 | Cheong et al. | Apr 1999 | A |
| 5913224 | MacDonald | Jun 1999 | A |
| 5974507 | Arimilli et al. | Oct 1999 | A |
| 5974508 | Maheshwari | Oct 1999 | A |
| 5983321 | Tran et al. | Nov 1999 | A |
| 6016533 | Tran | Jan 2000 | A |
| 6026461 | Baxter et al. | Feb 2000 | A |
| 6115792 | Tran | Sep 2000 | A |
| 6148370 | Kobayashi | Nov 2000 | A |
| 6161167 | Witt | Dec 2000 | A |
| 6185657 | Moyer | Feb 2001 | B1 |
| 6185703 | Guddat et al. | Feb 2001 | B1 |
| 6223256 | Gaither | Apr 2001 | B1 |
| 6237083 | Favor | May 2001 | B1 |
| 6240432 | Chuang et al. | May 2001 | B1 |
| 6240532 | Cho | May 2001 | B1 |
| 6263082 | Ishimoto et al. | Jul 2001 | B1 |
| 6269427 | Kuttanna et al. | Jul 2001 | B1 |
| 6279083 | MacDonald | Aug 2001 | B1 |
| 6295608 | Parkes et al. | Sep 2001 | B1 |
| 6351789 | Green | Feb 2002 | B1 |
| 6405287 | Lesartre | Jun 2002 | B1 |
| 6425056 | Meyer | Jul 2002 | B2 |
| 6430655 | Courtright et al. | Aug 2002 | B1 |
| 6434671 | Chung | Aug 2002 | B2 |
| 6591347 | Tischler et al. | Jul 2003 | B2 |
| 6848024 | Rowlands et al. | Jan 2005 | B1 |
| Number | Date | Country |
|---|---|---|
| 0 061 570 | Oct 1982 | EP |
| 0 259 095 | Mar 1988 | EP |
| 325 420 | Jul 1989 | EP |
| 0 381 471 | Aug 1990 | EP |
| 0 436 092 | Jul 1991 | EP |
| 0 459 232 | Dec 1991 | EP |
| 0 459 233 | Dec 1991 | EP |
| 0 569 221 | Nov 1993 | EP |
| 0 687 979 | Dec 1995 | EP |
| 997 821 | May 2000 | EP |
| 03010901.1 | Sep 2003 | EP |
| 2 214 336 | Aug 1989 | GB |
| 2 263 985 | Aug 1993 | GB |
| 2 263 987 | Aug 1993 | GB |
| 2 281 422 | Mar 1995 | GB |
| Number | Date | Country | |
|---|---|---|---|
| 20050044325 A1 | Feb 2005 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 09633683 | Aug 2000 | US |
| Child | 10950275 | US |
