The invention described herein relates to the field of microprocessor architecture. More particularly, the invention relates to the scheduling of multiple load operations within a load cycle of a pipelined microprocessor.
Microprocessor performance may be increased by enabling multiple cache load operations to be executed within the same cycle. One method of increasing the load bandwidth of a microprocessor is to support additional cache ports that may be accessed in parallel. However, supporting additional cache ports within cache memories, such as a Level 1 (L1) cache can be expensive in terms of die area and cycle time.
Other techniques to increase load bandwidth include interleaving, replication, time-division multiplexing, and line buffering. Interleaving involves dividing a cache into a number of sub-banks and using low-order address bits to access the banks. However, interleaving requires more die area for crossbar switching to direct loads and retired stores to the proper cache bank.
Replication involves emulating an N-port cache by replicating an M-port data cache array N/M times. While replication eliminates the bank conflict problem of interleaving, it may be expensive in terms of die area. Furthermore, while replication addresses the load-bandwidth problem, it exacerbates the store-bandwidth problem since store traffic must be broadcast to all of the replicated arrays simultaneously in order to ensure that each array has an updated copy of the data.
Time-division multiplexing involves emulating an N-port data cache by decreasing the cycle time of an M-port array by a factor of N/M. However, time-division multiplexing is difficult and expensive to implement and scale to higher frequencies.
Line buffering involves adding a small line-buffer that holds cache lines recently read from the cache by load operations. Subsequent loads may obtain data from this buffer, which can be multi-ported due to its small size. However, line buffering is complex and expensive in terms of cycle times, because loads that miss in the buffer must go to the cache, thereby increasing latency.
The features and advantages will become apparent from the following detailed description in which:
A method and apparatus for increasing load bandwidth is described, intended to enable increased load operation performance without substantially increasing die area or power within a pipelined microprocessor architecture.
The present invention is described by way of embodiments, which exemplify the features of the present invention. The invention may be implemented in hardware or software or a combination thereof. In particular, hardware circuits described herein may be implemented using standard complementary metal-oxide semiconductor (CMOS) technology, or other technologies not specifically described. Integrated circuits described herein, such as a microprocessor, cache, or other device, may be implemented as monolithic integrated circuits or on separate die.
Load and store operations may be stored within a load buffer 115 and store buffer 120, respectively, before accessing system memory. A load operation may retrieve load data faster than accessing the data from the load buffer by attempting to retrieve the load data from the L1 data cache. However, more current versions of load data may exist within the fast store-forwarding buffer that will be updated in system memory or the L1 data cache. Therefore, load operations may be satisfied faster by attempting to retrieve the load data from the fast store-forwarding buffer, rather than wait for the most current version of the load data to be updated within the L1 data cache or system memory.
The fast store forwarding buffer, the L1 data cache, the load buffer, and the store buffer may be implemented using circuitry well known to those of ordinary skill in the art, including Static Random Access Memory (SRAM) cells using standard complementary metal-oxide-semiconductor (CMOS) technologies. However, other memory circuit architectures and/or semiconductor technologies may be used. Similarly, the schedule may be implemented as a semiconductor logic circuit using standard CMOS devices. Alternatively, some or all of the scheduler may implemented on a machine-readable medium in which instructions are stored, which when executed by a processor perform operations to schedule load operations within a microprocessor.
In order to access the fast store-forwarding buffer and the L1 data cache, an address generation unit 125 generates a partial address in order to facilitate fast access to the L1 data cache. In order to access the load buffer or store buffer, a full address is generated by the address generation unit, which may take more time to generate and decode than the partial address generated to access the L1 data cache. Furthermore, L1 data cache typically supports faster access speeds than other memory structures within a computer system. Therefore, a time savings can be realized by attempting to retrieve load data from the L1 data cache instead of other memory structures, such as the store or load buffers or system memory.
However, failed attempts to retrieve load data from either the fast store forwarding buffer or L1 data cache may need to be rescheduled and re-issued in order to satisfy a load operation. Similarly, any operations that depend on the failed load operation must also be rescheduled, as the data upon which they depend has not been successfully retrieved.
The L1 data cache may support a limited number of cache ports as cache ports may be costly in terms of die area, cycle time, and power consumption. Store forwarding buffer ports, on the other hand, typically do not require as much die area as L1 data cache ports. Therefore, multiple store forwarding buffer ports may be implemented without incurring as much die cost as when implementing multiple L1 cache ports. The architecture illustrated in
Alternatively, two loads per cycle may be supported by the microprocessor architecture of
In one embodiment, a prediction is made as to whether a load will be satisfied by accessing the store-forwarding buffer or the L1 data cache. The loads are then scheduled and issued accordingly. If a load operation is predicted to be satisfied by accessing the store-forwarding buffer, the accuracy of the prediction may be verified by checking to see whether a store corresponding to the target address of the predicted load operation is later stored within the store buffer. If the prediction was correct, a predictor 201 may be updated in order to better predict the likelihood that the same load operation will be satisfied by accessing the store-forwarding buffer in the future. Likewise, if the prediction was incorrect, the predictor may be updated accordingly.
Multiple ports may be implemented relatively inexpensively with regard to cycle time, power consumption, and die area in the store-forwarding buffer. In one embodiment, two store-forwarding buffer ports 205 are implemented whereas only one L1 data cache port 210 is implemented such that multiple load operations may be scheduled and issued without incurring excessive die area, power consumption, and cycle time cost.
While one embodiment of the present invention has been described in which two store-forwarding buffer ports are implemented, more than two store-forwarding buffer ports may be implemented in other embodiments. Likewise, although one L1 data cache port has been implemented in one embodiment of the present invention, more than one L1 data cache port may be implemented, depending upon the cost and power consumption requirements of the computer system in which the present invention is used.
Various prediction algorithms may be implemented in order to predict whether a load operation will be satisfied by accessing the store forwarding buffer or the L1 data cache. In one embodiment, an N-bit saturating bimodal counter may be implemented as a prediction algorithm. An N-bit saturated bimodal counter may be realized by a two-bit counter which increments or decrements between four states, each corresponding to a different strength of a prediction for a given load operation.
For example, if it is determined that a load was predicted correctly, the predictor may increment the two-bit counter to indicate that the load operation is now a “strongly predicted” load. Alternatively, if it is determined that the load operation was incorrectly predicted to be satisfied by accessing the store-forwarding buffer, the counter may be decremented to indicate that the load is a “weakly predicted” load.
Other prediction algorithms may be used, including a static prediction bit in the compiler or a “wait table” predictor, such as that used in the ALPHA 21264 microprocessor.
Regardless of the particular prediction algorithm used, the predictor may be implemented either as a semiconductor logic circuit using standard CMOS semiconductor devices, or other semiconductor technology. Alternatively, some or all of the predictor may implemented on a machine-readable medium in which instructions are stored, which when executed by a processor perform operations to predict whether load operations within a microprocessor will be satisfied by accessing a fast store-forwarding buffer.
In one embodiment, the store forwarding buffer and address generation unit are dual ported to allow all loads to check the store forwarding buffer, including loads that are not predicted to be satisfied by accessing the store forwarding buffer. Loads not predicted to be satisfied by accessing the store-forwarding buffer may, therefore, access load data without incurring replay.
Embodiments of the present invention, either in whole or in part, may be implemented using in hardware or by a machine-readable medium containing instructions, which when executed by a processor, perform operations corresponding to the present invention. Accordingly, some aspects of the invention may be implemented in hardware while others are implemented in software.
While this invention has been described with reference to illustrative embodiments, this description is not intended to be construed in a limiting sense. Various modifications of the illustrative embodiments, as well as other embodiments, which are apparent to persons skilled in the art to which the invention pertains are deemed to lie within the spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
5210838 | Jensen | May 1993 | A |
5822788 | Kahn et al. | Oct 1998 | A |
6055629 | Kulkarni et al. | Apr 2000 | A |
6141747 | Witt | Oct 2000 | A |
Number | Date | Country | |
---|---|---|---|
20030065908 A1 | Apr 2003 | US |