JP2003323304A - Method and device for speculative task generation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and device which can perform a speculative execution only when reduction of execution time can be expected in order to reduce the overhead occurring when the speculative execution fails over a wide range. <P>SOLUTION: A basic block is segmented by a basic data extraction part 11 from a compiler 2, a data preparation finishing point, an execution determination edge, a non-execution determination edge, and the amount of data for every basic block are obtained at a data dependence analyzing part 12 and a control dependence analyzing part 13. A parallel task selection part 14 seeks a link of the basic blocks having a desirable characteristics for parallel processing, and determines the link as a parallel task. Among the parallel tasks, a parallel task in which (1) the number of commands between the data preparation finishing point and the execution determination edge in a certain basic block group is greater than a predetermine number, (2) the number of commands between the data preparation finishing point and the non-execution determination edge is less than a predetermined number, or (3) the amount of data generated within the range of a certain basic block group and used outside the range is less than a predetermined number, is made to be a speculative task. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention, in a system in which a program is divided into a plurality of threads and executed in parallel by a plurality of processors, starts speculative task generation before the execution of a certain thread is decided. A method and apparatus. In recent years, a system having a plurality of processors in a system and capable of simultaneously executing a plurality of threads has been put into practical use. In such a system, it is essential for efficient use of the system to execute processes or threads at the same time as many as the number of processors. On the other hand, in order to obtain the result of a certain program in a short time, the program is divided into a plurality of threads and executed in parallel to increase the speed. In this case, in the process of decomposing a program into a plurality of threads, it is important to generate a sufficient number of threads that can be executed in parallel, which is equal to the number of processors. However, due to the nature of the program, if you try to create threads only by calculations that are found to be necessary, there are cases where a sufficient number of threads cannot be created. in this case,
A speculative execution method has been proposed in which a program is executed at high speed by creating a thread (speculative thread) that starts execution before it is determined whether it is needed or not, and at the same time, the effective utilization rate of the system is improved. . The present invention relates to a method and apparatus for generating a speculative task in such a speculative execution method.

[0002]

2. Description of the Related Art Conventional speculative execution has realized a method of starting execution of a branch destination instruction before a result of a conditional branch instruction is determined based on prediction of a conditional branch direction by hardware. These are intended to improve the execution speed by speculatively executing the predicted branch direction instruction. However, since only a narrow range can be analyzed, even if a certain instruction becomes executable, it is difficult to execute it, or an instruction (thread) to be speculatively executed cannot be extracted from a wide range. Further, since speculative execution is basically performed for all branch instructions, there is a problem that speculative execution is performed even when it is not effective. Further, as a technique for performing speculation over a wider range, a method for performing speculation at a task level has been proposed (Japanese Patent Laid-Open No. 200-200200).
1-75802 Publication, Paper 1: Yamana, Yasue, Ishii, Muraoka, "Pre-evaluation method between macro tasks in parallel processing system", IEICE Transactions DI, Vol.J77-DI, N
o.5, pp.343-353,1994.5, Paper 2: Yamana, Sato, Kodama,
Sakane, Sakai, Yamaguchi, "Distributed control method of speculative execution between macro tasks in parallel computer EM-4", IPSJ Journal, Vol.36, No.7, pp.1578-1588, 1995.7).

[0003]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
In the method disclosed in Japanese Patent No. 5802, since the value of a variable whose value has not yet been determined is predicted and used, the overhead at the time of speculation failure may increase. Furthermore, what kind of speculative task is generated in which case is not specified. The above paper 1 proposes a speculative task generation method based on control dependence and data dependence between basic blocks. In this method, the basic tasks are actively speculated by copying the basic tasks so that the data dependencies between the basic tasks become unique without depending on the control. obtain. The present invention has been made in view of the above circumstances, and the present invention performs speculation execution over a wide range and only when shortening of the execution time can be expected when speculation is successful, and the overhead at the time of speculation failure is reduced. It is an object of the present invention to provide a speculative task generation method and apparatus capable of performing the speculative task.

[0004]

In order to reduce the probability of speculative failure and to reduce the overhead at the time of failure, the present invention executes speculative execution after all the data required by the task to be speculatively executed is prepared. In the speculative execution method to be started, the above problem is solved as follows. As shown in FIG. 1, a source code or an intermediate language expression thereof is received from a compiler 2, a basic block cutout unit 11 cuts out a basic block, and a control relation therebetween is constructed. Then, the data dependence analysis unit 12 and the control dependence analysis unit 13 obtain the data shown in (1) below for each basic block. Then, the parallel task selection unit 14 selects a parallel task and a speculative task as described in (2) below. The parallel code generator 15 generates a code or the like required for parallel execution. (1) Obtain the following data for each basic block.・ Data preparation point: When all the data to be used in the basic block are prepared ・ Data usage point: When the data generated in the basic block is used or redefined ・ Execution confirmation edge: Execution of the basic block When the basic block is not executed ・ The output data amount: The amount of data generated by the basic block and used by other blocks (2) From the data, find the connection of basic blocks that satisfy any or all of the following properties (i), and use them as a parallel task, and select one that satisfies any or all of the following properties (ii). It is a speculative task. (i) Desired property for parallel execution (a) The number of instructions c1 from the data ready point of a certain basic block group to the basic block group and the number of instructions c2 from the end of the basic block group to the data use point Sum (c1
+ C2) is large. (b) c1 + c2 or c2 is larger than the number of instructions in the connection of the basic block groups. (c) The number of instructions d in the connection of the basic block groups in the program is larger than the predetermined value. (ii) Properties desirable for speculative execution (a) The number of instructions a between a data ready point and an execution confirmed edge of a certain basic block group is larger than a predetermined value. (b) The number of instructions b between the data preparation completion point of a basic block group and the non-execution confirmed edge is less than a predetermined value. (c) For a certain basic block group, the amount of data generated in the range and used outside the range e (OU
TPUT data amount) is less than a predetermined value. (3) The speculative task generated as described above starts at the data preparation completion point, transitions from the speculative state to the normal execution state at the execution confirmation edge, and the data generated in the task is transferred to other tasks. Accessible from. When the control is transferred to the non-execution speculative edge, the speculative task is canceled. In the present invention, a task satisfying the above property (i) is defined as a parallel task, and a task satisfying the above property (ii) is defined as a speculative task. Speculative execution can be performed only when shortening can be expected, and speculative execution at the task level with a small overhead at the time of speculative failure can be performed, and the execution time can be shortened.

FIG. 2 shows an operation at the time of successful speculation and an operation at the time of unsuccessful speculation in executing a speculative task. FIG. 10A shows the operation at the time of successful speculation. The speculative task is activated at the data preparation completion point described above, and the speculative task is executed in parallel with the normal task. Then, when the execution confirmation edge is passed, the output is globalized (the writing of the Output data from the work area to the regular area) is performed by the "Globalization OK notification of Output". When the task is completed, "output global
Then, the data can be used by the next task. The operation at the time of speculative failure is shown in FIG. 7B. When the execution confirmation edge is passed, the activated speculative task is canceled.

Here, as shown in (i) above, the desirable characteristics of parallel execution are as follows. (a) The sum of the number of instructions c1 from the data ready point to the basic block and the number of instructions c2 from the basic block to the data use point and data redefinition point is large. (b) The above c1 + c2 is substantially equal to or larger than the instruction number d of the basic block. (c) The number of instructions d in the basic block is large. That is, as is clear from FIG. 2, by making the tasks that satisfy the above conditions (a) and (b) into parallel tasks, the tasks that use the data generated by the tasks can perform processing without waiting. The processing delay can be reduced. Execution time can be expected to be shortened by setting the tasks that satisfy (c) as parallel tasks. Further, as shown in (ii) above, desirable properties for speculative execution are as follows. (a) The number of instructions from the data ready point to the execution confirmation edge is larger than a. (b) The number of instructions b from each data preparation completion point to the non-execution confirmation edge is small. (c) The amount e of data (Output Data) generated in the basic block and used in other areas is small. As is clear from FIG. 2, by making the condition that satisfies the above condition (a) a speculative task, the processing can be sped up by the number of instructions a and the execution time can be expected to be shortened. In addition, by making the speculative task satisfying the condition of (2) above, it is possible to reduce the amount of wasted processing even if speculative execution fails, and further, the above condition (3)
By making the speculative task satisfying the above condition, the storage area of the data generated by the task can be reduced even if the speculative execution fails, the post-processing can be simplified, and the overhead can be reduced.

[0007]

FIG. 1 is a diagram showing the configuration of a speculative task generator according to an embodiment of the present invention. In the figure, 1 is a source program, 2 is a compiler, and 3 is a code generated by the compiler. Basic block cutout part 1
1 receives a source code or its intermediate language expression from a compiler 2, cuts out a basic block, and builds a control relationship between them. The data dependence analysis unit 12 inputs Input Data, O for each of the cut-out basic blocks.
Output Data, find data ready point. The control dependence analysis unit 13 obtains an execution confirmed edge and a non-execution confirmed edge for each of the basic blocks. As will be described later, the parallel task selection unit 14 uses (1) the number of instructions from the data ready point to the basic block, (2) the data ready point,
The number of instructions up to the execution confirmation edge, (3) the number of instructions from the data ready point to the non-execution confirmed edge, (4) the number of instructions from the end of the basic block to the data use point, data redefinition point, 5) A task to be executed in parallel or speculatively is obtained based on the number of instructions in the basic block and (6) the amount of data (Output Data) generated in the basic block and used in other areas. The parallel code generation unit 15 generates the code (scheduling task, synchronization wait, etc.) necessary for parallel execution. As will be described later, the speculative task generated as described above is activated at the data preparation completion point, transitions from the speculative state to the normal execution state at the execution confirmation edge, and the data generated in the task Make it accessible to tasks.

FIG. 3 is a conceptual diagram showing the relationship between the data preparation completion point, the execution confirmed edge, the non-execution confirmed edge and the like.
As shown in the figure, the data preparation completion point is the final instruction that determines the Input Data of a certain task, and is determined for each execution confirmation edge. The execution-determined edge is a set of edges where it is first determined that a task is executed, and the non-execution-determined edge is a set of edges where it is initially determined that a task is not executed. The speculative task is
The task is started at the data preparation completion point, and when the non-execution confirmation edge is reached, the tax is canceled. Also, at the execution confirmation edge, transition from the speculative state to the normal execution state,
Output Da generated by executing this task
Other tasks can access ta.

The processing of each unit shown in FIG. 1 will be described below. (1) Process for obtaining data ready point FIG. 4 shows an algorithm for obtaining a data ready point.
5 to 6 are diagrams for explaining the processing by the above algorithm. The circles shown in FIGS. 5 to 6 indicate basic blocks, and FIG. 7 shows the meaning of each color applied to each basic block. Further, FIG. 8 shows an operation for obtaining the data preparation completion point. The process of obtaining the data preparation completion point by the data dependence analysis unit 12 will be described below. Here, the process of obtaining the data preparation completion point of the basic block F shown in FIGS. 5 to 6 will be described, but the same process is performed for each basic block. Input of basic block F
Data output points are basic blocks A, B, and C. In the following description, the basic block B1 and the basic block B2 in the algorithm of FIG.
It corresponds to basic blocks A, B, and C.

(I) Pre-processing The following processing is performed for each basic block. (1-1) Inpu of the basic block F in FIG.
The flow graph is sequentially traced from all the basic blocks at the t Data output point, and the traceable basic block is marked with X. As a result, the mark X is added to each basic block as shown in FIG. (1-2) Follow the flow graph from the basic block F in reverse order, and mark Y for each basic block. As a result, the mark Y is attached to each basic block as shown in FIG. (1-3) Paint the basic block with both mark X and mark Y in blue, and the others in white. As a result, the basic blocks A, B, C, D, E and F are painted blue as shown in FIG. (1-4) Paint the basic block having the Input Data of the basic block F black. As a result, the basic blocks A, B and C are painted black as shown in FIG.

(Ii) Processing body In the processing below, it is assumed that there is no white block and is ignored. For each non-white basic block, while the color changes, the information is propagated back to the control flow, and the operation shown in FIG. 8 is performed from the color of the basic block of itself and the color of the succeeding block. Then, the basic block for which the preparation completion point is obtained (in this case, the basic block F)
When the processing is completed for the basic blocks other than, the basic blocks of black and yellow are set as the preparation completion points for the basic block F. When the above process is performed on the basic block A, the basic block A is black and there is no blue basic block in the subsequent blocks (B) other than white.
As shown in, the basic block A is red.

Similarly, when the above process is performed on the basic block B, the basic block B is black, and a part (D) of the succeeding blocks (B, C) other than white has a blue basic block. As shown in (g), the basic block B is red. The basic block D is colored yellow. Further, the basic blocks C, D and E do not change because the succeeding blocks are all blue. Further, since the basic block F satisfies the condition of if (basic block B3! = ...) in the above algorithm and has no change, the basic block B is colored red. The basic block D is yellow. Here, the while statement of the algorithm is executed once, but since no change has occurred, the processing ends here. As a result of the above processing, the basic block C becomes black and the basic block D becomes yellow as shown in FIG. 6 (i), so that the basic blocks C and D become the preparation completion points.

FIG. 9 shows a processing flow for obtaining the data preparation completion point. Note that steps S1 to S6 in the figure are the "(i) preprocessing", and steps S7 to S11 are the "(ii) processing body". In step S1, it is checked whether there is a basic block for which a data preparation completion point has not been obtained. If there is no basic block that has not been obtained, the process ends. If there is a basic block for which the data preparation completion point has not been obtained, one basic block for which the data preparation completion point has not been obtained is selected in step S2. Step S
In FIG. 3, as shown in FIG. 5B, the flow graph is sequentially traced from all the basic blocks at the Input Data output points of the selected basic block, and the traced basic block is marked with X. In step S4, the flow graph is traced in reverse order from the basic block selected as shown in FIG. Then, in step S5, as shown in FIG. 5 (d), the basic blocks attached with the marks X and Y are painted blue, and the others are painted white. Further, in step S6, the basic block having the Input Data of the basic block selected as shown in FIG. 6E is painted black.

Next, in step S7, it is checked whether or not there is a color change in the basic block as a result of the above processing. If there is no color change, the process returns to step S1 to check whether there is a basic block for which the data preparation completion point has not been obtained. If there is no basic block, the process ends, and if there is, the above steps S1 to S6.
Process. If there is a color change, it is checked in step S8 whether there is a basic block that is not white and has not been processed. If not, the process returns to step S7 to perform the above process. If there is a basic block that is not white and has not been processed, in step S9, one basic block that has not been processed is selected, and it is checked whether it is the same as the basic block for which the data preparation completion point is obtained. If so, the process returns to step S8. If the basic blocks for which the data preparation completion point is obtained are not the same, the processing based on FIG. 8 is performed as shown in FIGS. 6 (f) and 6 (g), and the process returns to step S8. By performing the above processing, the preparation completion point can be obtained as shown in FIG.

(2) Process for obtaining execution-determined edge and non-execution-determined edge FIG. 10 shows an algorithm for obtaining the execution-determined edge and the non-execution-determined edge. Further, FIG. 11 is a diagram for explaining the processing by the above algorithm. The process of obtaining the execution-determined edge and the non-execution-determined edge by the control dependence analysis unit 13 will be described below. Here, the process of obtaining the execution-determined edges (plurality) and the non-execution-determined edges (plurality) with respect to the data preparation completion point of the basic block F will be described, but the same process is performed for each basic block. The data preparation completion points of the basic block F are the basic blocks C and D as described above. In the following description, the basic block B1 and the basic block B2 in the algorithm of FIG. 10 correspond to the basic blocks C and D, respectively. (i) Pre-processing (1-1) The flow graph is sequentially traced from the basic blocks C and D at the respective data preparation completion points of the basic block F, and the traced basic block is marked with X. (1-2) Follow the flow graph from basic block F in reverse order, and mark Y. (1-3) Paint the basic block with the mark X and the mark Y together in blue, and the others in white. As a result, as shown in FIG. 11, the basic blocks E and F are painted blue.

(Ii) Basic block C of data preparation completion point of basic block F of the processing main body,
The following processing is performed for each of D. Follow the basic blocks (C, D) following the blue basic block in order, and mark the once traced blocks so that they do not go more than once. Then, if the subsequent block is white, the edge that goes to the white basic block is added to the non-execution determined edge for the data ready point (C, D) of the basic block F. If the traced edge is the control-dependent edge of the basic block F (the edge that is first determined to go to the basic block F), that edge is the execution-determined edge for the data ready point (C, D) of the basic block F. And For example, regarding the basic block C in FIG. 11, since C-F is an edge that is first determined to go to the basic block F, it is a control-dependent edge and an execution-determined edge. With the above processing, as shown in FIG. 11, for the basic block C, the execution confirmed edge becomes C-F and the non-execution confirmed edge becomes C-W. For the basic block D, the execution-determined edges are DF and EF, and the non-execution-determined edges are DY and EU.

FIG. 12 shows a processing flow for obtaining the execution confirmed edge and the non-execution confirmed edge. Note that steps S1 to S6 in the figure are the "(i) preprocessing", and steps S7 to S11 are the "(ii) processing body". In step S1, it is checked whether there is a basic block for which an execution-determined edge has not been obtained. If there is no basic block that has not been obtained, the process ends. If there is a basic block for which the execution-determined edge is not found, in step S2,
Select one basic block for which the execution-determined edge has not been obtained. In step S3, the flow graph is sequentially traced from the basic blocks of all the data ready points of the selected basic block, and the traceable basic block is marked with X. In step S4, the flow graph is traced in reverse order from the selected basic block, and a mark Y is attached.
Then, in step S5, the basic blocks attached to the marks X and Y are painted blue, and the others are painted white.

Next, in step S6, it is checked whether or not there is an unprocessed data preparation completion point, and if there is no data preparation completion point, the process proceeds to step S1. If so, in step S7, one basic block at the data preparation completion point that has not been processed is selected. In step S8, the subsequent basic block of the selected basic block is sequentially traced to the blue basic block, and if there is a basic block having a white basic block, the edge to the white basic block is unexecuted. Add to the edge. Then, in step S9, the blocks following the selected basic block are sequentially followed by the blue basic block, and if the edge traced is the control-dependent edge of the selected basic block, that edge is set to this data ready point. Add to execution confirmed edge. Then, the process returns to step S6 to repeat the above process. By performing the above processing, it is possible to obtain the execution confirmed edge and the non-execution confirmed edge as shown in FIG.

(3) Processing for Finding Tasks to be Executed in Parallel and Speculatively FIG. 13 shows an algorithm for finding tasks to be executed in parallel and speculatively executed. FIG. 14 is a diagram for explaining the fusion of basic blocks and basic block groups. Hereinafter, a process of obtaining a task to be executed in parallel or speculatively by the parallel task selection unit 14 will be described. (i) For each basic block, find the basic data of the following basic block. Here, the following data is the basic data of the basic block. -A: the number of instructions from each data preparation completion point to its execution confirmation edge-b: the number of instructions from each data preparation completion point to its non-execution confirmation edge-c1: From the data preparation completion point to its basic block Number of instructions If there are multiple paths in the control flow,
The calculation is performed with an appropriate path (for example, if the shortest path length or a loop is passed 10 times). C2: the number of instructions from the basic block to the data use point and data redefinition point, d: the number of instructions of the basic block, e: data generated in the basic block and used in other areas (Output Data) ) Amount

Further, the properties satisfying the following conditions are assumed to be desirable for parallel execution and desirable for speculative execution. (1) Properties desirable for parallel execution-c1 + c2 is large-d is large-c1 + c2 is approximately equal to or greater than d, or c2 is large (2) Properties desirable for speculative execution-a is large-b is small- e is small

(Ii) In order to try to generate a task with a certain basic block at the head, the following basic blocks are set for all subsequent basic blocks of the task with each basic block in the descending order of c1 as a starting point of the task. Perform processing. Case1: In the control flow, the following basic block B2 has the same definite edge as the basic block B1: When the succeeding basic block B2 having the same definite edge as the basic block B1 is fused with each other (the basic block (B1 and B2), the data ready point, the data use point, the data redefinition point are obtained, the above a to e are obtained, and if the desired properties of the parallel tasks are improved, they are fused. For example, in FIG. 14A, the fixed execution edges to the basic block A and the basic block B are both the edges from the basic block Z to the basic block A. In this case, fusion of block A and block B is attempted, and if the above conditions are satisfied, basic block A and basic block B
Fuse.

Case2: In the control flow, when the succeeding basic block B2 has an edge in the basic block group fused so far as the execution-determined edge, it returns to the preceding basic block B3 of the basic block B2 and the basic block B3. , A data ready point, a data use point, and a data redefinition point in the case where all the basic blocks that are control-dependent on an edge between the subsequent basic blocks are fused, the a to e are obtained, and the parallel processing is performed. If the desired properties of the task improve, merge. FIG. 14B shows an example of the above fusion. In FIG. 14B, basic block A
And B are already fused. Here, when trying to fuse the basic blocks C, the execution-determined edge BC of the basic block C is different from the execution-determined edge ZA of the fused basic blocks A and B. In this case, the preceding basic block B of the basic block C and its subsequent basic blocks C,
The fusion of the basic blocks C, D, and E that are control-dependent on the edge (BC, BE) between E and E is attempted, and if the above conditions are satisfied, the basic blocks C, D, and E are fused. . (iii) In the finally fused basic blocks, those having a value having a desirable property for the parallel task or more are regarded as parallel tasks. Further, among them, a speculation task is one having a value having a desirable property for speculative execution or more.

FIG. 15 shows a processing flow for obtaining a parallel task and a speculative execution task. First, in step S1, the c1 is obtained, and the basic blocks are arranged in descending order of c1. In step S2, it is checked whether there is any basic block that has not been checked yet, and if not, the process ends. Also, if you have a basic block that you have not yet examined,
In step S3, one basic block is selected from the beginning and used as the starting point of the task. In step S4, it is checked whether or not the task has a subsequent basic block that has not been checked. If not, go to step S10. Step S10
Then, in the fused basic blocks, those having a value having a desirable property for the parallel task or more are regarded as the parallel task. Among them, those having a value having a desirable property for speculative execution or more are regarded as speculative tasks, and the process returns to step S2. In step S4, if there is a subsequent basic block that has not been checked in the task, the process proceeds to step S5, and a subsequent basic block that has not been checked is selected. In step S6,
It is checked whether the succeeding basic block has the same execution confirmed edge as the basic block of the seed (determination as to the case 1). If they are equal, go to step S8, and if they are not equal, go to step S7.

In step S8, the data ready point, the data use point, and the data redefinition point in the case where the subsequent block is fused with the task are obtained, and the above a to e are obtained, and if the desirable properties for the parallel task are improved. To fuse. Further, in step S7, when the subsequent basic block is merged into the task, it is checked whether or not the edge in the basic block group has the definite edge (determination as to the case 2). If it has a fixed edge, it goes to step S9, and if it doesn't, it goes back to step S4. In step S9, the data ready point, the data use point, and the data redefinition point in the case of returning to one preceding basic block and merging all the basic blocks that are control-dependent on the edge between it and the following basic block And the above a to e
Are obtained, and if the desired properties of the parallel task are improved, they are fused, and the process returns to step S4. By performing the above processing, it is possible to generate a speculative task as well as a parallel task. (4) Parallel code generation process When the parallel task and the speculative task are generated as described above, the parallel code generation unit 15 generates codes such as a scheduling task and a synchronization wait code required to execute the tasks in parallel. To generate. It should be noted that this process can be realized using a conventional technique.

The coded parallel task and speculative task generated as described above are allocated to each processor element of the parallel computer system, and parallel execution and speculative execution are performed. FIG. 16 is a diagram schematically showing the operation when the speculative task is executed. In the figure, when a speculative task start signal from the task start point generates a task and speculative execution is started, and control is transferred to the uncertain execution edge, speculative execution is canceled and the task ends. In addition, the Globalization signal is notified (in this case, since speculative execution has failed, data writing is not performed). In addition, the execution confirmation edge causes a transition from the speculative state to the normal execution state, and the O generated by the task is generated.
The output data is written, and the Global signal is notified. As a result, the data generated in the speculative task can be accessed by other tasks.

(Supplementary Note 1) A method of generating a speculative task of a program, wherein the time point when all the data used in the basic block are collected is analyzed from the data dependence of each basic block in the program, and the time point is analyzed. As a data preparation completion point, the time at which the execution of the basic block is confirmed with respect to the data preparation completion point is analyzed, and the time is defined as an execution confirmation edge, and between the data preparation completion point and the execution confirmation edge of a certain basic block group. A speculative task generation method in which the basic block group is treated as a speculative task when the number of instructions in the above is greater than a predetermined value. (Supplementary Note 2) A method of generating a speculative task of a program, wherein when the amount of data generated in a certain basic block group and used outside the certain range is less than a predetermined value, the basic block group Is a speculative task, and a speculative task generation method. (Supplementary note 3) A method of generating a speculative task of a program, analyzing from the data dependency of each basic block in the program the time when all the data used in that basic block are complete, and completing the data preparation The point at which it is determined that the basic block is not executed is analyzed with respect to the data ready point, and the point is set as the non-execution definite edge, and the data ready point of a certain basic block group and the non-execution definite edge. A speculative task generation method, characterized in that when the number of instructions in between is less than a predetermined value, the basic block group is treated as a speculative task. (Supplementary Note 4) From the data dependency of each basic block in the program, the time point when the data generated in each basic block is used is analyzed, and the time point is set as the data use point, and from the data preparation completion point of a certain basic block group, When the sum of the number of instructions up to the basic block group and the number of instructions from the end of the basic block group to the data use point is large, the basic block group is treated as a speculative task. Appendix 3 Speculative task generation method. (Supplementary Note 5) From the data dependency of each basic block in the program, the time when the data generated in each basic block is used is analyzed, and the time is set as the data use point, and from the data preparation completion point of a certain basic block group, The sum of the number of instructions to the basic block group and the number of instructions from the end of the basic block group to the data use point, or the number of instructions from the end of the basic block group to the data use point is the connection of the basic block groups. 5. The speculative task generation method of appendices 1, 2, 3 or 4 wherein the basic block group is set as a speculative task when the number of instructions is larger than the number of instructions. (Additional remark 6) When the number of instructions in the connection of the basic block groups in the program is larger than a predetermined value, the basic block group is treated as a speculative task.
2, 3, 4 or the speculative task generation method of Appendix 5. (Supplementary note 7) A speculative task generation device for a program, a means for cutting out a basic block from a program, and a data dependency of each basic block in the program are used to analyze the time when all the data used in the basic block are complete. A means for obtaining a data preparation completion point, a means for analyzing the time when the execution of the basic block is determined with respect to the data preparation completion point, and obtaining an execution confirmation edge, and a data preparation completion point and execution confirmation for a certain basic block group. A speculative task generation device, comprising: means for determining whether the number of instructions between edges is larger than a predetermined value and selecting the basic block group as a speculative task when the instruction count is larger than the predetermined value. (Supplementary Note 8) A speculative task generation program for generating a speculative task of a program, in which the data to be used in the basic block is all collected from the data dependency of each basic block in the target program. Processing for obtaining a data preparation completion point, processing for analyzing the time when the execution of the basic block is determined for the data preparation completion point, and setting the execution time as the execution confirmation edge; A speculative task generation program for causing a computer to execute a process of selecting a basic block group as a speculative task when the number of instructions between a data preparation completion point of a block group and an execution confirmed edge is larger than a predetermined value. (Supplementary Note 9) A speculative task generation program for generating a speculative task of a program, wherein the program obtains the amount of data generated in a range of a certain basic block group and used outside the range. A speculative task generation program for causing a computer to execute a process and a process of selecting the basic block group as a speculative task when the amount of data is smaller than a predetermined value. (Supplementary Note 10) A speculative task generation program for generating a speculative task of a program, wherein the program determines the time point at which all the data used in the basic block are complete from the data dependency of each basic block in the program. A process of analyzing and obtaining a data ready point, a process of analyzing a time point at which the basic block is determined not to be executed for the data ready point, and setting the time point as a non-execution definite edge, and a certain basic block A speculative task generation program for causing a computer to execute a process in which the basic block group is a speculative task when the number of instructions between the data preparation completion point of the group and the non-execution confirmed edge is less than a predetermined value.

[0027]

As described above, according to the present invention, the speculative execution is performed over a wide range and only when the execution time can be expected to be shortened when the speculation succeeds, and the overhead at the speculative failure is small. Speculative execution of can be performed. Therefore, the execution time can be shortened.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a speculative task generation unit according to an embodiment of the present invention.

FIG. 2 is a diagram showing an operation at the time of successful speculation and an operation at the time of unsuccessful speculation.

FIG. 3 is a conceptual diagram showing a relationship such as a data preparation completion point, an execution confirmed edge, and a non-execution confirmed edge.

FIG. 4 is a diagram showing an algorithm for obtaining a data preparation completion point.

FIG. 5 is a diagram (1) illustrating a process of obtaining a data preparation completion point.

FIG. 6 is a diagram (2) illustrating a process of obtaining a data preparation completion point.

FIG. 7 is a diagram illustrating the meaning of each color applied to a basic block.

FIG. 8 is a diagram showing an operation of obtaining a data preparation completion point.

FIG. 9 is a diagram showing a processing flow for obtaining a data preparation completion point.

FIG. 10 is a diagram showing an algorithm for obtaining an execution confirmed edge and a non-execution confirmed edge.

FIG. 11 is a diagram illustrating a process of obtaining an execution confirmed edge and a non-execution confirmed edge.

FIG. 12 is a diagram showing a processing flow for obtaining an execution confirmed edge and a non-execution confirmed edge.

FIG. 13 is a diagram showing an algorithm for obtaining a task to be executed in parallel or speculatively.

FIG. 14 is a diagram for explaining fusion of basic blocks and basic block groups.

FIG. 15 is a diagram showing a processing flow for obtaining a parallel task and a speculative execution task.

FIG. 16 is a diagram schematically showing an operation at the time of executing a speculative task.

[Explanation of symbols]

1 source program 2 compiler 3 codes 11 Basic block cutting section 12 Data dependency analysis unit 13 Control dependence analysis unit 14 Parallel task selection unit 15 Parallel code generator

Claims (5)

[Claims] 1. A method for generating a speculative task of a program, comprising:
The time when all the data to be used in the basic block are collected is analyzed, the time is defined as a data preparation completion point, and the time when the execution of the basic block is determined with respect to the data preparation completion point is analyzed. A speculative task generation method, wherein a speculative task is defined as an execution-determined edge, and when the number of instructions between a data preparation completion point of a basic block group and the execution-determined edge is greater than a predetermined value, the basic block group is treated as a speculative task. 2. A method of generating a speculative task of a program, wherein when a certain basic block group has a data amount that is generated in a range and is used outside the range is less than a predetermined value, the basic block A speculative task generation method characterized in that a group is a speculative task. 3. A method for generating a speculative task of a program, wherein the data dependency of each basic block in the program is:
The time when all the data to be used in the basic block are collected is analyzed, the time is defined as the data ready point, and the time when the basic block is determined not to be executed is analyzed for the data ready point, and the time is analyzed. Is a non-execution definite edge, and when the number of instructions between a data preparation completion point of a certain basic block group and the non-execution definite edge is less than a predetermined value, the speculative task is characterized by that basic block group being a speculative task. Generation method. 4. A program speculative task generation device for cutting out a basic block from a program, and data dependence of each basic block in the program,
A means for analyzing the time when all the data to be used in the basic block are obtained to obtain a data preparation completion point, and a time when the execution of the basic block is confirmed with respect to the data preparation completion point, and obtaining an execution confirmation edge And a means for determining whether the number of instructions between a data preparation completion point and an execution confirmed edge of a certain basic block group is larger than a predetermined value, and selecting the basic block group as a speculative task when it is larger than the predetermined value. A speculative task generation device having: 5. A speculative task generation program for generating a speculative task of a program, wherein all the data used in the basic block is based on the data dependency of each basic block in the target program. There is a process of analyzing the aligned time points to obtain a data preparation completion point, and a process of analyzing the time point when the execution of the basic block is confirmed with respect to the data preparation completion point and setting the time point as an execution confirmation edge. A speculative task generation program characterized by causing a computer to execute a process of selecting a basic block group as a speculative task when the number of instructions between the data preparation completion point and the execution confirmed edge of the basic block group is larger than a predetermined value. .