JP3795055B1 - Value prediction apparatus, multiprocessor system, and value prediction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a value prediction apparatus and method for performing prediction with reference to a prediction table and updating the prediction table, and requesting continuous prediction for the same entry before the prediction result is confirmed and the prediction table is updated. The purpose is to make a correct prediction when is presented.
SOLUTION: A prediction history holding unit 5 that holds a history of a prediction for which prediction processing has not been completed is compared with an identifier included in the prediction history and an identifier given at the start of the prediction processing. Counting means 6 for counting the number of times the identifiers match, and the prediction circuit 1 sends the output value of the prediction circuit to the update circuit 2, which uses the output value of the prediction circuit to state the prediction table Is updated, and the state of the prediction table is returned to the prediction circuit. The prediction circuit performs the prediction process again using the state of the virtually updated prediction table, and the number of times that both identifiers match. The operation of the prediction circuit and the update circuit is repeated according to the above.
[Selection] Figure 2

Description

  The present invention predicts a data value by referring to a prediction table based on an identifier (usually abbreviated as ID) of an entry (event) input in response to a prediction request, and updates the prediction table The present invention relates to a value prediction apparatus having a function to

  On the other hand, the present invention refers to a multiprocessor system that performs value speculation execution in units of threads using the value prediction device, and refers to a prediction table based on an identifier of an entry that is input in response to a prediction request. The present invention relates to a value prediction method for predicting data values.

  In general, in order to improve the performance of a computer system, in a data processing apparatus such as a processor, a value prediction apparatus (value predictor) having a function of predicting corresponding data values in advance according to the contents of an input entry. Also called). This value prediction apparatus is generally configured to predict a data value by referring to a prediction table, and to update the prediction table when the success or failure of the prediction is confirmed. The value predicting device having such a structure is usually manufactured by a semiconductor integrated circuit such as an LSI (Large Scale Integrated Circuit) together with a data processing device.

  FIG. 1 is a block diagram showing a configuration of a conventional value prediction apparatus. However, here, the configuration of a conventional value prediction apparatus 100 used in a data processing apparatus such as a processor is shown in a simplified manner.

  The conventional value prediction apparatus 100 in FIG. 1 basically refers to the prediction table 30 based on the identifier of the entry (ID in FIG. 1) input in response to a prediction request, and sets the value of the corresponding data. The prediction circuit 1 that performs the prediction, and the update circuit 2 that updates the value of the corresponding data in the prediction table 30 with the output value of the prediction circuit 1 as the prediction value PV when the prediction success or failure of the prediction circuit 1 is confirmed. And.

  Here, the prediction table 30 is normally built in the prediction table holding unit 3 such as a RAM (Random Access Memory). When it is determined whether or not the prediction by the prediction circuit 1 is successful and the prediction process is completed, an update value RV corresponding to the prediction value PV output from the prediction circuit 1 is sent to the update circuit 2, and this update is performed. The value of the corresponding data in the prediction table 30 is updated by the update process of the circuit 2 (that is, the value of the corresponding data is rewritten by the update value RV).

  Furthermore, the value prediction apparatus 100 in FIG. 1 includes a hash circuit 4 that generates an index for referring to the prediction table 30 based on an identifier (ID) of an input entry. In the hash circuit 4, an index indicating which address in the prediction table 30 stores the data related to the entry is generated based on the identifier of the input entry.

  In the value prediction device 100 of FIG. 1, the functions of the prediction circuit 1 and the update circuit 2 are usually realized by a CPU (Central Processing Unit) 12 that performs various controls of the computer system. Further, the CPU 12 is provided with a storage unit 14 such as a ROM (Read-Only Memory) or a RAM. The storage unit 14 stores in advance various programs including programs for executing the functions of the prediction circuit 1 and the update circuit 2. Note that a ROM or RAM built in the CPU 12 can be used as the ROM or RAM.

  Further, the value prediction apparatus 100 of FIG. 1 is provided with a CPU (Central Processing Unit) 12 that performs various controls of the computer system including the value prediction apparatus 100. Further, the CPU 12 is provided with a storage unit 14 such as a ROM (Read-Only Memory) or a RAM. The storage unit 14 stores in advance programs for performing various controls of the computer system. Note that a ROM or RAM built in the CPU 12 can be used as the ROM or RAM.

  The prediction circuit 1 and the update circuit 2 are usually configured by a hardware circuit including an arithmetic circuit and a register. However, the functions of the prediction circuit 1 and the update circuit 2 can be realized by the CPU 12. In this case, a program for executing the functions of the prediction circuit 1 and the update circuit 2 as described above is stored in the storage unit 14. By reading the program from the storage unit 14 and executing it by the CPU 12, the functions of the prediction circuit 1 and the update circuit 2 are realized by software.

  As described above, the conventional value prediction apparatus 100 of FIG. 1 performs prediction of the value of the corresponding data by referring to the prediction table 30 based on the identifier of the input entry, and the prediction result is confirmed and predicted. When the above process is completed, the output value of the prediction circuit 1 is set as the predicted value PV (that is, the update value RV), and the value of the corresponding data in the prediction table 30 is updated.

  Therefore, the prediction of the data value is performed by referring to the prediction table 30 based on the identifier of the input entry, and before the prediction result is confirmed and the prediction table 30 is updated, the same entry is continuously recorded. On the other hand, when the prediction is to be performed, a situation occurs in which the value of the corresponding data in the prediction table 30 is not updated. In this case, since the value of the data in the prediction table 30 is not an appropriate value after the update, a problem that correct prediction cannot be performed on the same entry occurs.

  In other words, in the conventional value prediction apparatus, in a period between the timing of performing the prediction on the entry input in response to the first prediction request and the timing of updating the prediction table after the prediction result is confirmed, Even if a second or more prediction request is made for the same entry, correct prediction cannot be performed. Therefore, in the conventional value predicting apparatus, it is necessary to complete the update operation by updating the prediction table after executing the prediction operation with reference to the prediction table for the input entry. Even if a continuous prediction request for the same entry is presented between the update operation and the update operation, it cannot be appropriately handled.

  When predicting the value of data using a conventional value prediction device, the time from when the prediction is made with reference to the prediction table until the value of the data in the prediction table is updated is relatively short, or prediction In the case where the value of data in the prediction table does not change significantly by updating the table, the above problem does not become so large. However, in particular, when performing complex prediction in a relatively coarse unit, such as when performing value speculation by thread unit value prediction, the prediction table is updated after performing prediction with reference to the prediction table. The above time becomes longer and the above problems become serious.

  Here, for reference, the following patent documents 1 to 7 related to the conventional value predicting apparatus as described above are presented as prior art documents. In particular, Patent Documents 5 to 8 exemplify thread speculative execution apparatuses that use value prediction in units of threads by a value prediction apparatus.

  In Patent Document 1, when branch prediction is performed, prediction is performed by actually executing a branch instruction using a prediction data value calculated by a data value predictor, instead of performing prediction using a history of past branches. A processor for performing is disclosed.

  In Patent Document 2, a field for instructing whether or not to perform value prediction is provided in a part of an instruction, and an instruction for applying value prediction only to an instruction whose performance can be improved by value prediction based on this field or A processor has been disclosed in which an instruction cache is prepared to suppress performance degradation due to unnecessary value prediction and to improve prediction accuracy when performing value prediction.

  In Patent Literature 3, when a processor executes a memory operation instruction, it refers to a history of success / failure results of speculative execution in the past, predicts whether data speculative execution is successful, and if success is predicted, If the memory operation instruction is speculatively executed with respect to data dependency in a non-program order and a failure is predicted, the failure probability of the data dependency speculative execution is reduced by executing the memory operation instruction deterministically in the program order, A data-dependent speculative execution control device that improves the execution performance of a program is disclosed. In such a data-dependent speculative execution control device, when predicting the success or failure of data speculative execution, the target address of the load instruction is assigned to a specific entry in the speculative execution success / failure history table with duplication, for example, by a hash function circuit. I have to.

  In Patent Document 4, in a branch prediction method using PHT (Pattern History Table), a tagged PHT unit is divided into a plurality of units, and entry interference is avoided while using as little memory capacity as possible to perform branch prediction. A branch prediction apparatus that improves accuracy is disclosed.

  Patent Document 5 includes a processor that speculatively executes a task level instruction based on a predicted value by selective value prediction in a multiprocessor system, and a verification device that is connected to the processor and performs verification of value prediction. A speculative execution device is disclosed in which the execution state of a processor is monitored by a verification request from the processor and an interrupt is generated in the processor if the prediction is not correct.

  In Patent Document 6, there is a control unit for executing a group of instructions belonging to a plurality of threads in parallel in a single instruction processing device, and the dependency of generation and use of data belonging to the same thread by a thread dividing unit The instruction pair is recognized and divided into a plurality of threads, and the data value is predicted by the value prediction unit before the data value is actually generated. The control unit is based on the value predicted by the value prediction unit. By executing the instruction using the data as a separate thread and speculatively executing by the instruction execution means, it is possible to execute in parallel a pair of instructions belonging to the same thread and having a dependency on generation and use of data. There is disclosed a speculative execution type processor that can be used.

  In Patent Document 7, when a program is divided into a plurality of threads and these threads are allocated to a plurality of processors and executed speculatively, at least one in a speculative execution state is executed by a thread manager that controls the operation of each processor. A speculative multi-thread processing device is disclosed in which a plurality of execution paths in a program are speculatively executed at the same time using a certain processor.

  In Patent Document 8, a thread execution unit on a parallel processing device having a plurality of thread execution units is effectively used, and both branch side and non-branch side are speculatively executed by an instruction after a conditional branch instruction. A parallel processing device is disclosed that allows execution to be continued without substantially deviating from branch prediction by canceling execution on the wrong side when a branch instruction is determined.

  However, in any of Patent Documents 1 to 8, the same entry is entered before prediction of the data value is performed by referring to the prediction table for the input entry, and the prediction result is confirmed and the prediction table is updated. No mention is made as to what to do when a request for continuous prediction is presented. Therefore, in any of Patent Documents 1 to 8, a problem similar to that of the conventional value prediction apparatus as described above occurs.

JP 2000-132390 A JP 2002-312162 A JP 2002-351657 A Japanese Patent Laid-Open No. 2003-5956 JP 2001-75802 A JP 2001-356904 A JP 2000-47887 A Japanese Patent Laid-Open No. 11-96005

  The present invention has been made in view of the above problems, and predicts data values for an input entry, and continuously predicts the same entry before the prediction result is confirmed and the prediction table is updated. It is an object of the present invention to provide a value predicting apparatus, a multiprocessor system, and a value predicting method capable of performing correct prediction when the above request is presented.

  In order to solve the above problem, the value prediction apparatus of the present invention refers to data already stored in the prediction table based on the identifier of the entry every time an entry to be predicted is input, A prediction circuit that predicts the value of the data, and an update circuit that updates the value of the corresponding data in the prediction table using the output value of the prediction circuit as a prediction value when the prediction process by the prediction circuit is completed And a prediction history holding means for temporarily holding a prediction history that is currently in progress and is not completed, an identifier included in the prediction history, and start of the prediction processing Counting means for comparing the identifiers given at times and counting the number of times both identifiers match each other, and when there is a number of times that both identifiers match, the prediction circuit The output value of the prediction circuit obtained by performing processing is sent to the update circuit, and the update circuit virtually updates the state of the prediction table using the output value of the prediction circuit, and virtually updates The predicted state of the prediction table is returned to the prediction circuit, and the prediction circuit performs a prediction process again using the virtually updated state of the prediction table, and determines the number of times that both the identifiers match. Accordingly, the operation of the prediction circuit and the update circuit is repeated without actually updating the prediction table, and the output value of the prediction circuit obtained thereby is used as the final prediction value.

  Preferably, in the value predicting device of the present invention, when there is no number of times that the two identifiers match, the prediction circuit performs the prediction process only once, and obtains the output value of the prediction circuit obtained thereby. The final predicted value is used.

  Preferably, in the value predicting apparatus of the present invention, when the number of times that both the identifiers match is 2 or more, the prediction table updated virtually while the operations of the prediction circuit and the update circuit are repeated. A buffer circuit for temporarily holding the state is provided.

  Preferably, in a multiprocessor system having the value predicting apparatus of the present invention and a plurality of processors, a program of a single thread is divided into a plurality of threads, and value speculation is performed by the plurality of processors by value prediction in units of threads. When the plurality of threads to be processed by the plurality of processors are input as the entry to the value prediction device, the order in which the plurality of threads are input to the value prediction device is switched. It is possible to perform correct value prediction for each of the above threads.

  On the other hand, the value prediction method of the present invention predicts the value of the data by referring to the data already stored in the prediction table based on the identifier of the input entry, and the prediction process is completed. When updating the value of the corresponding data in the prediction table using the value obtained as a result of the prediction process as a prediction value at the time of the prediction, the prediction history that is currently in progress and has not been completed is temporarily stored. And the step of comparing the identifiers included in the prediction history with the identifiers given at the start of the prediction process, counting the number of times both identifiers match each other, and both the identifiers Are present, the state of the prediction table is virtually updated using the value obtained as a result of the prediction process, and the state of the prediction table virtually updated is updated. The prediction process is repeated and the prediction process is repeated without actually updating the prediction table according to the number of times that both the identifiers match and the step of performing the prediction process again. And a step.

  In summary, according to the present invention, after predicting the data value based on the identifier (ID) for the input entry, the prediction process is currently in progress and the corresponding data value in the prediction table is updated. If not, the previous prediction history is temporarily stored in the prediction history register, etc., and the identifier included in the prediction history is compared with the identifier given at the start of the prediction process. Counts the number of matches.

  Further, when there is a number of times that both of the above identifiers match, the state of the prediction table is virtually updated using a value obtained as a prediction processing result, and the state of the prediction table virtually updated is used. The prediction process is performed again, and the prediction process is repeated without actually updating the prediction table according to the number of times the above two identifiers match, and the correct prediction is realized using the resulting value as the final predicted value. Like to do.

  Therefore, according to the present invention, after predicting the value of data for an input entry, a request for consecutive predictions for the same entry in a state where the previous prediction is in progress and the prediction table is not updated. Is presented based on the final predicted value by repeatedly processing the previous prediction for the same entry using the state of the prediction table virtually updated with the history of the previous prediction. The correct prediction can be made.

  In particular, in a multiprocessor system including the value prediction device of the present invention, when performing value speculation execution by value prediction in a thread unit, a plurality of prediction table states virtually updated by the value prediction device of the present invention are used. By executing the threads simultaneously, correct value prediction can be performed for each thread even if the order in which the plurality of threads are input to the value prediction device is changed.

  Hereinafter, the configuration and operation of the preferred embodiments of the present invention will be described with reference to the accompanying drawings (FIGS. 2 to 8).

  FIG. 2 is a block diagram showing the configuration of the value prediction apparatus according to the first example of the present invention. However, here, the configuration of the value predicting apparatus 10 according to the first embodiment of the present invention used in a data processing apparatus such as a processor is shown in a simplified manner. Hereinafter, the same components as those described above are denoted by the same reference numerals.

  As in the case of the above-described conventional value prediction apparatus (see FIG. 1), the value prediction apparatus 10 according to the first embodiment of FIG. 2 uses the identifier (ID) of the entry input in response to the prediction request. Based on the prediction circuit 1 that predicts the value of the corresponding data with reference to the prediction table 30 and the success or failure of the prediction by the prediction circuit 1 is determined, the output value of the prediction circuit 1 is used as the prediction value PV. And an update circuit 2 that updates the value of the corresponding data in the prediction table 30. Here, after the prediction of the value of the corresponding data with reference to the prediction table 30, if the success or failure of the prediction by the prediction circuit 1 is confirmed, the update value RV becomes the same value as the prediction value PV.

  Also in the value prediction apparatus 10 according to the first example of FIG. 2, the prediction table 30 is built in the prediction table holding unit 3 such as a RAM. When it is determined whether or not the prediction by the prediction circuit 1 is successful and the prediction process is completed, an update value RV corresponding to the prediction value PV output from the prediction circuit 1 is sent to the update circuit 2, and this update is performed. The value of the corresponding data in the prediction table 30 is updated by the update process of the circuit 2.

  Further, the value prediction apparatus 10 according to the first embodiment of FIG. 2 refers to the prediction table 30 based on the identifier of the input entry as in the case of the conventional value prediction apparatus (see FIG. 1) described above. A hash circuit 4 is provided for generating an index for this purpose. In the hash circuit 4, an index indicating which address in the prediction table 30 stores the data related to the entry is generated based on the identifier of the input entry.

  Further, the value predicting device 10 in FIG. 2 is provided with a CPU 12 that performs various controls of the computer system including the predicting device 10 as in the case of the above-described conventional value predicting device (see FIG. 1). Further, the CPU 12 is provided with a storage unit 14 such as a ROM or a RAM. The storage unit 14 stores in advance programs for performing various controls of the computer system. Note that a ROM or RAM built in the CPU 12 can be used as the ROM or RAM.

  The prediction circuit 1 and the update circuit 2 are configured by a hardware circuit including an arithmetic circuit, a register, and the like, as in the case of the conventional value prediction apparatus (see FIG. 1) described above. However, the functions of the prediction circuit 1 and the update circuit 2 can be realized by the CPU 12. In this case, a program for executing the functions of the prediction circuit 1 and the update circuit 2 as described above is stored in the storage unit 14. By reading the program from the storage unit 14 and executing it by the CPU 12, the functions of the prediction circuit 1 and the update circuit 2 are realized by software.

  Furthermore, the value prediction apparatus 10 according to the first embodiment of FIG. 2 includes a prediction history register 5 that temporarily holds a prediction history that is currently in progress and is not completed, and a prediction history. And a counting circuit 6 that compares the identifier given at the start of the prediction process and counts the number of times that both identifiers match each other. The prediction history register 5 and the counting circuit 6 correspond to the prediction history holding unit and the counting unit of the present invention, respectively. An additional circuit 15 including the prediction history register 5 and the counting circuit 6 is provided in the value prediction apparatus 10 as a hardware circuit that represents the feature of the present invention. The prediction history register 5 adds an identifier of an entry to be predicted at the start of the prediction process and deletes it at the end of the prediction process.

  Here, the operation of calculating the predicted value of the data value in response to the prediction request in the value prediction apparatus 10 of FIG. 2 will be schematically described. When the prediction circuit 1 calculates a predicted value of the data value, the hash circuit 4 generates a hash value (index) based on the identifier (ID) of the entry, and a prediction table referred to based on the hash value The prediction circuit 1 calculates the predicted value of the data value using the entries in 30.

  Here, the number of predictions obtained by the prediction history register 5 and the counting circuit 6 (that is, the number of times that the identifier included in the prediction history matches the identifier given at the start of the prediction process) is the same. The prediction circuit 1 and the update circuit 2 are circulated as many times as the number of predictions for which the process of the identifier of the entry has not been completed. In this case, the update circuit 2 uses the prediction value output from the prediction circuit 1 as an update value, and sends the state of the prediction table to be updated to the prediction circuit 1 without updating the prediction table 30. Thereby, in a state where the prediction table has not been updated, the state of the table to be updated later by the preceding prediction is virtually and temporarily generated, and accurate prediction is realized.

  When all or part of the prediction currently in progress is canceled for some reason, including a prediction failure, the identifier of the canceled prediction entry is deleted from the prediction history register 5. At this time, the prediction table 30 is not changed. With the above processing, the state of the prediction table 30 and the prediction history register 5 follows the cancellation of the execution of prediction, and the value prediction device 10 operates correctly.

  Next, operations of the prediction circuit 1 and the update circuit 2 in the value prediction apparatus 10 of FIG. 2 will be described in more detail. In the prediction table 30, if only the prediction that has already been processed has been updated and there is a preceding prediction that has not been processed yet, the state of the prediction table updated by this preceding prediction is new. It is necessary to make a prediction of a new request. Here, when all preceding predictions are successful, the predicted value and the updated value take the same value. For this reason, about the preceding prediction, the state of the prediction table updated by the prediction can be produced | generated by assuming the prediction value as an update value. Even when there are two or more preceding predictions, the state of the target prediction table can be generated by repeating this process a plurality of times.

  When there is no preceding prediction, that is, when the number of times the identifier included in the prediction history and the identifier given at the start of the prediction process match each other is 0, a conventional value prediction device (see FIG. 1) As in the case of), the prediction circuit 1 is operated only once to perform a new prediction process. In this case, the output value of the prediction circuit 1 obtained by the above processing becomes the predicted value.

  When the preceding prediction exists, that is, when the number of times the identifier included in the prediction history and the identifier given at the start of the prediction process match each other is 1 or more, the prediction circuit 1 The output value of the prediction circuit 1 obtained by performing only the prediction process is sent to the update circuit 2. The update circuit 2 virtually updates the prediction table using the output value of the prediction circuit 1 as the update value and updates the prediction table without reflecting the update result in the prediction table 30. The state of the prediction table is returned to the prediction circuit 1. The state of the virtually updated prediction table is temporarily stored in the RAM 11. In the value prediction apparatus 10 of FIG. 2, the prediction circuit 1 is provided with the RAM 11, but the prediction circuit 1 may be provided with a built-in RAM. The prediction circuit 1 operates again using the virtually updated state of the prediction table, and performs prediction processing. The operations of the prediction circuit 1 and the update circuit 2 are repeated as many times as necessary for this operation (that is, the number of times that the identifier included in the prediction history and the identifier given at the start of the prediction process match each other). The output value of the prediction circuit 1 obtained after that is the final predicted value.

  When the preceding prediction is completed, as in the case of the conventional value prediction device (see FIG. 1), an update value corresponding to the final prediction value is used by the update circuit 2 to correspond in the prediction table 30. The data value is updated. On the other hand, if the preceding prediction is canceled for some reason, the prediction circuit 1, the update circuit 2, and the prediction table 30 are not affected.

  As already described, in the value prediction apparatus 10 of FIG. 2, the prediction history register 5 and the counting circuit 6 are configured by hardware circuits. However, the functions of the prediction history register 5 and the counting circuit 6 as well as the functions of the prediction circuit 1 and the update circuit 2 can be realized by software of the CPU 12. The storage unit 14 such as a ROM or a RAM can store in advance programs for executing the functions of the prediction circuit 1, the update circuit 2, the prediction history register 5, and the counting circuit 6.

  For example, the program stored in the storage unit 14 includes a step of temporarily holding a prediction history that is currently in progress and a prediction process, an identifier included in the prediction history, and a prediction process. A step of comparing the identifiers given at the start, counting the number of times both identifiers match each other, and if there is a number of times both identifiers match, a prediction table using a value obtained as a prediction processing result The state is updated virtually, and the prediction table is actually updated according to the number of times that both of the identifiers coincide with the step of performing the prediction process again using the virtually updated state of the prediction table Without repeating the prediction process, and using the value obtained thereby as the final predicted value.

  Further, in the first embodiment shown in FIG. 2, when the CPU 12 is operated using a computer-readable storage medium (or recording medium), a storage medium (for example, holding the contents of the program as described above) It is preferable to prepare a hard disk H) in the external storage device 16 such as a disk device. The storage medium of the first embodiment is not limited to the above-described one, but a floppy disk, an MO (Magneto-Optical Disk), a CD-R (Compact Disk-Recordable), a CD-ROM (Compact). It can be provided in the form of various storage media such as portable media such as Disk Read-only Memory) and other fixed media.

  According to the first embodiment, after the prediction of the data value for the input entry, the continuous prediction for the same entry is performed in the state where the preceding prediction is in progress and the prediction table is not updated. Even when a request is presented, it is possible to make a correct prediction based on the final predicted value by repeatedly performing the process of the preceding prediction using the state of the prediction table virtually updated by the history of the preceding prediction. Will be able to do.

  FIG. 3 is a circuit block diagram showing a specific example of the prediction history register and the counting circuit of FIG. Here, a specific circuit configuration example of the prediction history register 5 and the counting circuit 6 in the additional circuit 15 in the value prediction apparatus 10 of FIG. 2 is illustrated.

  As shown in FIG. 3, the prediction history register 5 (see FIG. 2) has a structure of a queue register 50 that holds a history of preceding predictions that have not been processed yet in a serial format. Each time a request for prediction is presented, the queue register 50 adds an identifier used for referring to the prediction table (the ID of the entry to be predicted sent from the hash circuit 4) to the end. On the other hand, the prediction process is completed and the value of the data in the prediction table is updated. At the same time, the identifier corresponding to this data is deleted from the head of the queue register 50, and the entire queue register is shifted. In the queue register 50, there are prediction identifiers in order of time that have not yet been completed since the prediction process is started and the prediction table has not been updated.

  On the other hand, the counting circuit 6 (see FIG. 2) includes a comparator 60 and a counting counter 61. For each entry in the prediction history register, the comparator 60 compares the identifier (ID) included in the prediction history so far with the identifier (ID) given at the time of the prediction request, and the valid in the entry. For the correct entry, the number of times that both identifiers match each other is output. The count counter 61 counts and accumulates the number of times that both identifiers output from the comparator 60 match each other. The number of times accumulated in this way indicates the number of predictions that are being processed for the same identifier.

  When all or a part of the prediction currently in progress is cancelled, the corresponding identifier is deleted from the prediction history register 30 (see FIG. 2) for the canceled prediction, and the entry is invalidated.

  FIG. 4 is a time chart for explaining the temporal relationship between prediction and update. In FIG. 4A, in a conventional value predicting apparatus (for example, see FIG. 1), the time (t) between the time when the input entry is predicted and the time when the prediction table is updated is shown. In FIG. 4B, in the value prediction apparatus of the method of the present invention (see, for example, FIG. 2), the time when the input entry is predicted, and the time when the prediction table is updated The relationship between and for time (t) is shown.

  For the value prediction apparatus of the conventional method as shown in FIG. 4A, the prediction table is updated when the update operation is completed after the prediction operation is executed with reference to the prediction table for the input entry. And the order of the prediction operation and the update operation cannot be changed.

  On the other hand, in the value prediction apparatus of the method of the present invention as shown in FIG. 4B, the first prediction operation is performed after referring to the prediction table for the input entry and executing the first prediction operation. When a prediction request for the same entry is presented in a state where the prediction operation is in progress and the prediction table has not been updated (for example, three prediction requests earlier than the prediction table update time are presented) Even if it is performed), the process of the preceding prediction is repeated by using the state of the prediction table virtually updated by the history of the preceding prediction (for example, by repeating three times) to obtain the final result. A correct prediction can be performed based on the predicted value.

  FIG. 5 is a block diagram showing the configuration of the value prediction apparatus according to the second example of the present invention. However, here, the configuration of the value predicting apparatus 10a according to the second embodiment of the present invention used in a data processing apparatus such as a processor is shown in a simplified manner.

  The value predicting device 10a according to the second embodiment of FIG. 5 has almost the same configuration as the value predicting device 10 (see FIG. 2) according to the first embodiment described above, but the update circuit 2 has a buffer. The point where the circuit 20 is provided is different from the value prediction apparatus 10 according to the first embodiment described above.

  In the value prediction apparatus 10a according to the second embodiment of FIG. 5, when the prediction value of the data value is calculated by the prediction circuit 1 in response to a prediction request, the value prediction apparatus 10 according to the first embodiment described above. As in the case of the above, the hash value is generated by the hash circuit 4 based on the identifier (ID) of the entry, and the prediction circuit 1 uses the entry in the prediction table 30 referred to based on this hash value. The predicted value of the data value is calculated.

  Here, when the number of preceding predictions is 1 or more, that is, when the number of times the identifier included in the prediction history and the identifier given at the start of the prediction process match each other is 1 or more, the prediction circuit 1 sends the output value of the prediction circuit 1 obtained by performing the prediction process the same number of times to the update circuit 2. The update circuit 2 virtually updates the prediction table using the output value of the prediction circuit 1 as the update value and updates the prediction table without reflecting the update result in the prediction table 30. The state of the prediction table is returned to the prediction circuit 1. The prediction circuit 1 operates again using the virtually updated state of the prediction table, and performs prediction processing. The output value of the prediction circuit 1 obtained after repeating the operations of the prediction circuit 1 and the update circuit 2 as many times as necessary for this operation becomes the final prediction value.

  In particular, when the number of preceding predictions is 2 or more, it is necessary to temporarily hold the state of the virtually updated prediction table during the repeated operation of the prediction circuit 1 and the update circuit 2. Therefore, in the value predicting apparatus 10a according to the second embodiment of FIG. 5, the buffer circuit that temporarily holds the state of the virtually updated prediction table during the repetitive operations of the prediction circuit 1 and the update circuit 2. 20 is provided in the update circuit 2.

  According to the value predicting apparatus according to the second embodiment of FIG. 5, the value obtained by the first operation in the above repetitive operations is stored in the buffer circuit 20 and the second and subsequent operations are performed. In this case, the value can be read from the buffer circuit 20 and used effectively, so that the efficiency in calculating the predicted value by the repeated operation of the prediction circuit 1 and the update circuit 2 is improved.

FIG. 6 is a flowchart for explaining the processing flow of the prediction operation and the update operation according to the embodiment of the present invention. Here, when the prediction circuit 1, the update circuit 2, the prediction history register 5, and the hardware circuit of the counting circuit 6 are operated by the control of the CPU 12 (for example, see FIG. 2), the prediction operation and the update operation of the present invention are executed. The processing flow will be described sequentially.
FIG. 6 is a flowchart for explaining the processing flow of the prediction operation and the update operation according to the embodiment of the present invention. Here, the processing flow when the CPU 12 (see, for example, FIG. 2) is operated to execute the prediction operation and the update operation of the present invention will be sequentially described.

  When predicting the data value by referring to the prediction table based on the identifier of the input entry, first, in step S1, the prediction history that is currently in progress and is not completed is calculated as the prediction history. It keeps temporarily in the register.

  Next, in step S2, for each entry in the prediction history register, a comparison is made between the identifier (ID) included in the prediction history held so far and the identifier (ID) given at the start of the prediction process. Like to do.

  Further, in step S3, it is determined whether or not there is a number of times that the number of preceding predictions, that is, the identifier included in the prediction history and the identifier given at the start of the prediction process match each other. Yes.

  Furthermore, when it is determined in step S4 that there is no preceding prediction, that is, it is determined that the number of times that the identifier included in the prediction history and the identifier given at the start of the prediction process match each other is 0. In such a case, as in the case of a conventional value prediction apparatus (for example, see FIG. 1), the prediction circuit is operated only once to perform a new prediction process.

  When it is determined that the preceding prediction exists, that is, when it is determined that the number of times that the identifier included in the prediction history and the identifier given at the start of the prediction process match each other is 1 or more In step S5, the prediction circuit sends the output value of the prediction circuit obtained by performing the prediction process the same number of times to the update circuit.

  Further, in step S6, the update circuit virtually updates the prediction table using the output value of the prediction circuit as the update value, and returns the virtually updated state of the prediction table to the prediction circuit.

  Further, in step S7, the prediction circuit operates again using the virtually updated state of the prediction table to perform prediction processing. The output value of the prediction circuit obtained after repeating the operations of the prediction circuit and the update circuit as many times as necessary for this operation becomes the final prediction value.

  In the flowchart of FIG. 6, the processing flow in the case where the prediction circuit 1, the update circuit 2, the prediction history register 5, and the counting circuit 6 are configured by hardware circuits is described. Even when the functions of the prediction history register 5 and the counting circuit 6 are realized by the CPU 12, it is possible to perform a processing flow almost the same as the above-described processing flow by software (a program stored in the storage unit 14). is there.

  FIG. 7 is a block diagram showing a configuration of an application example of the value prediction apparatus of the present invention, and FIG. 8 is a schematic diagram showing a situation in which a change occurs between prediction and update in FIG. Here, an example is shown in which the value prediction apparatus 10 (see, for example, FIG. 2) of the present invention is applied to a multiprocessor system 80 that performs thread speculation execution when performing value speculation execution by thread unit value prediction. ing.

  The multiprocessor system 80 of FIG. 7 includes a value prediction apparatus 10 (see, for example, FIG. 2) of the present invention and a plurality of processors (for example, a first processor 7-1 to an nth processor 7-n (n is 2). The above-described arbitrary positive integer))) and a thread control unit 8 that performs thread control for executing thread speculative execution are provided.

  Here, since the value prediction apparatus 10 has the same configuration as the value prediction apparatus (see FIG. 2) according to the first embodiment described above, the re-explanation regarding the value prediction apparatus 10 is omitted here. .

  In the multiprocessor system 80 of FIG. 7, a plurality of processors (first processor 7-1 to n-th processor 7-n) are arranged on a single semiconductor integrated circuit chip, and a single thread program is executed. Is divided into a plurality of threads TH (see FIG. 8), and thread speculation is executed in parallel by a plurality of processors by the thread control unit 8 in the chip.

  In the chip of this semiconductor integrated circuit, the value predicting device 10 of the present invention is mounted as a thread level value predicting device. In the prediction of the thread level value, the value used in the thread is predicted at the start of the thread, the value speculation execution is performed, and the prediction table is updated when the success or failure of the result of the value speculation execution is determined. Here, each thread is executed in parallel, and the time until the success or failure of each thread is predicted becomes longer than the instruction level value prediction.

  Therefore, as shown in the schematic diagram of FIG. 8, when parallel value speculative execution is performed with each iteration of the loop as a thread TH with respect to the loop portion of the program, a plurality of threads having the same identifier are simultaneously executed, and the prediction table Before is updated, a request for value prediction for threads having the same identifier occurs.

  When prediction requests are generated in this order, the conventional value prediction apparatus (see FIG. 1) cannot cope with it, so that it is impossible to perform value prediction regarding threads. However, by using the value predicting apparatus 10 of the present invention (see, for example, FIG. 2), the correct value for each thread can be obtained even in a situation where parallel value speculative execution is performed for a plurality of threads TH as shown in FIG. It is possible to make a prediction.

  More specifically, in the multiprocessor system 80 of FIG. 7, when a plurality of threads to be processed by a plurality of processors are input as entries to the value prediction apparatus 10, the prediction and update and time (on the right side of FIG. t), as shown in the relationship with t), even if the order of prediction and update of the input thread is switched, the preceding prediction is made using the state of the prediction table virtually updated by the history of the preceding prediction. By repeating this process, accurate value prediction for each thread is realized.

(Supplementary Note 1) Each time an entry to be predicted is input, a prediction circuit that refers to data already stored in a prediction table based on an identifier of the entry and predicts a value of the data; In a value prediction device having an update circuit that updates the value of the corresponding data in the prediction table using the output value of the prediction circuit as a prediction value when the prediction processing by the prediction circuit is completed,
Prediction history holding means for temporarily holding a prediction history that is currently in progress and is not completed,
A counter for comparing the identifier included in the prediction history with the identifier given at the start of the prediction process, and counting the number of times both identifiers match each other;
When there is a number of times that both the identifiers match, the prediction circuit sends the output value of the prediction circuit obtained by performing the prediction process to the update circuit, and the update circuit The state of the prediction table is virtually updated using an output value, the state of the virtually updated prediction table is returned to the prediction circuit, and the prediction circuit is virtually updated with the prediction table Perform the prediction process again using the state of
The operation of the prediction circuit and the update circuit is repeated without actually updating the prediction table according to the number of times the two identifiers match, and the output value of the prediction circuit obtained thereby is the final predicted value. A value predicting device characterized by that.

  (Supplementary Note 2) When there is no number of times that the two identifiers match, the prediction circuit performs the prediction process only once, and the output value of the prediction circuit obtained thereby is used as a final prediction value. The value predicting device according to supplementary note 1, wherein:

  (Supplementary Note 3) When the number of times the two identifiers match is 2 or more, the state of the virtually updated prediction table is temporarily held while the operations of the prediction circuit and the update circuit are repeated. The value predicting apparatus according to appendix 1, wherein a buffer circuit is provided.

  (Supplementary note 4) When a prediction process that is currently in progress and is not completed is canceled, an identifier corresponding to the prediction to be canceled is deleted from the prediction history, and an entry having the identifier is invalidated. The value predicting device according to supplementary note 1, characterized by:

(Additional remark 5) In the multiprocessor system which has a value prediction apparatus as described in any one of additional remarks 1 to 4, and a plurality of processors,
When dividing a program of a single thread into a plurality of threads and performing value speculation execution by value prediction on a thread basis by the plurality of processors, the plurality of threads to be processed by the plurality of processors are the entries. Even when the order in which the plurality of threads are input to the value prediction device is switched when input to the value prediction device, it is possible to perform correct value prediction for each of the threads. A multiprocessor system.

(Supplementary Note 6) Based on the identifier of the input entry, the value of the data is predicted with reference to the data already stored in the prediction table, and when the prediction process is completed, the prediction In the value prediction method for updating the value of the corresponding data in the prediction table using the value obtained as the processing result as the predicted value,
Temporarily holding a history of predictions that are currently in progress and not complete,
Comparing the identifier included in the prediction history with the identifier given at the start of the prediction process, and counting the number of times both identifiers match each other;
When there is a number of times that both the identifiers match, the state of the prediction table is virtually updated using the value obtained as a result of the prediction process, and the state of the prediction table virtually updated is used. The step of performing the prediction process again,
Repeating the prediction process without actually updating the prediction table according to the number of times the two identifiers match, and setting a value obtained thereby as a final predicted value. Value prediction method.

(Supplementary Note 7) The value prediction method further includes:
The value prediction according to claim 6, further comprising a step of performing the prediction process only once when the number of times that both the identifiers match does not exist, and setting a value obtained thereby as a final prediction value Method.

(Supplementary note 8) Based on the identifier of the input entry, the value of the data is predicted with reference to the data already stored in the prediction table, and when the prediction process is completed, the prediction When updating the value of the corresponding data in the prediction table using the value obtained as the processing result as the predicted value,
A temporary history of forecasts that are currently in progress and are not complete,
Comparing the identifier included in the prediction history with the identifier given at the start of the prediction process, counting the number of times both identifiers match each other,
When there is a number of times that both the identifiers match, the state of the prediction table is virtually updated using the value obtained as a result of the prediction process, and the state of the prediction table virtually updated is used. And run the forecast again
A program for executing the process of repeating the prediction process without actually updating the prediction table according to the number of times the two identifiers match, and making a value obtained thereby as a final predicted value A computer-readable storage medium characterized by being stored.

  In the present invention, when a plurality of threads are executed in parallel by a plurality of processors and value speculation execution is performed by value prediction in units of threads, even if the order of prediction and update of input threads is switched, each thread Therefore, the present invention can be applied to a multi-processor system of a thread speculative execution type that can perform correct value prediction.

It is a block diagram which shows the structure of the conventional value prediction apparatus. It is a block diagram which shows the structure of the value prediction apparatus which concerns on 1st Example of this invention. FIG. 3 is a circuit block diagram illustrating a specific example of a prediction history register and a counting circuit in FIG. 2. It is a time chart for demonstrating the temporal relationship between prediction and update. It is a block diagram which shows the structure of the value prediction apparatus which concerns on 2nd Example of this invention. It is a flowchart for demonstrating the processing flow of the prediction operation | movement and update operation | movement by the Example of this invention. It is a block diagram which shows the structure of the example of application of the value prediction apparatus of this invention. FIG. 8 is a schematic diagram illustrating a situation in which a change occurs between prediction and update in FIG. 7.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Prediction circuit 2 Update circuit 3 Prediction table 4 Hash circuit 5 Prediction history register 6 Count circuit 7-1 to 7-n Multiple processors 8 Thread control unit 10 Value prediction device 10a Value prediction device 11 RAM
12 CPU
DESCRIPTION OF SYMBOLS 14 Memory | storage part 15 Additional circuit 16 External memory | storage device 20 Buffer circuit 30 Prediction table holding | maintenance part 50 Queue register 60 Comparator 61 Count counter 80 Multiprocessor system 100 Value prediction apparatus

Claims (5)

A prediction circuit that predicts a value of the data by referring to data already stored in a prediction table based on the identifier of the entry each time an entry to be predicted is input, and a prediction by the prediction circuit In the value prediction device having an update circuit that updates the value of the corresponding data in the prediction table using the output value of the prediction circuit as the prediction value at the time when the processing of is completed,
Prediction history holding means for temporarily holding a prediction history that is currently in progress and is not completed,
A counter for comparing the identifier included in the prediction history with the identifier given at the start of the prediction process, and counting the number of times both identifiers match each other;
When there is a number of times that both the identifiers match, the prediction circuit sends the output value of the prediction circuit obtained by performing the prediction process to the update circuit, and the update circuit The state of the prediction table is virtually updated using an output value, the state of the virtually updated prediction table is returned to the prediction circuit, and the prediction circuit is virtually updated with the prediction table Perform the prediction process again using the state of
The operation of the prediction circuit and the update circuit is repeated without actually updating the prediction table according to the number of times the two identifiers match, and the output value of the prediction circuit obtained thereby is the final predicted value. A value predicting device characterized by that.   When there is no number of times that the two identifiers match, the prediction circuit performs the prediction process only once, and an output value of the prediction circuit obtained thereby is used as a final prediction value, The value prediction apparatus according to claim 1.   When the number of coincidence of both identifiers is 2 or more, a buffer circuit is provided that temporarily holds the state of the virtually updated prediction table while the operations of the prediction circuit and the update circuit are repeated. The value prediction apparatus according to claim 1, wherein: In the multiprocessor system which has a value prediction device according to any one of claims 1 to 3, and a plurality of processors,
When dividing a program of a single thread into a plurality of threads and performing value speculation execution by value prediction on a thread basis by the plurality of processors, the plurality of threads to be processed by the plurality of processors are the entries. Even when the order in which the plurality of threads are input to the value prediction device is switched when input to the value prediction device, it is possible to perform correct value prediction for each of the threads. A multiprocessor system. The value of the data is predicted by referring to the data already stored in the prediction table based on the identifier of the input entry, and obtained as the prediction processing result when the prediction processing is completed In a value prediction method for updating a value of corresponding data in the prediction table using a value as a predicted value,
Temporarily holding a history of predictions that are currently in progress and not complete,
Comparing the identifier included in the prediction history with the identifier given at the start of the prediction process, and counting the number of times both identifiers match each other;
When there is a number of times that both the identifiers match, the state of the prediction table is virtually updated using the value obtained as a result of the prediction process, and the state of the prediction table virtually updated is used. The step of performing the prediction process again,
Repeating the prediction process without actually updating the prediction table according to the number of times the two identifiers match, and setting a value obtained thereby as a final predicted value. Value prediction method.

Images