JP2021009443A - Information processing apparatus and compiler program - Google Patents

Abstract

To provide an information processing apparatus and a compiler program that enable high-speed division of loops included in a source code.SOLUTION: Dependency information indicating a dependency relationship between instructions included in a loop of an intermediate language generated from a source code is generated, a first matrix is generated by converting the generated dependency information into a matrix, the instructions included in the loop are assigned to a plurality of groups based on a degree of dependency among the instructions calculated from the generated first matrix, and the loop is divided for each assigned group.SELECTED DRAWING: Figure 19

Description

The present invention relates to an information processing device and a compiler program.

For example, a compiler program (hereinafter, also simply referred to as a compiler) used for HPC (High Performance Computing) or the like performs optimization processing for improving processing performance such as processing speed when compiling source code. Specifically, the compiler performs loop splitting, for example, to split a loop included in the source code (hereinafter, also referred to as a loop to be split) into a plurality of loops.

This makes it possible for the compiler to suppress the occurrence of optimization-inhibiting factors caused by, for example, lack of hardware resources. Further, the compiler can suppress a decrease in cache efficiency, for example (see, for example, Patent Documents 1 and 2).

JP-A-2002-123563 JP-A-2009-104422

Here, a compiler as described above analyzes, for example, the dependency between each instruction included in the loop to be divided, and distributes each instruction to a plurality of divided loops based on the analysis result. Split the target loop.

However, the number of instruction combinations for which dependency analysis needs to be analyzed increases according to the number of instructions contained in the loop to be divided. Therefore, if the number of instructions contained in the loop to be split is enormous, the compiler will take a lot of time to split the loop, and the source code cannot be compiled efficiently.

Therefore, in one aspect, it is an object of the present invention to provide an information processing apparatus and a compiler program capable of dividing a loop included in a source code at high speed.

In one aspect of the embodiment, the first matrix is formed by generating dependency information indicating the dependency between each instruction included in the loop of the intermediate language generated from the source code and converting the generated dependency information into a matrix. A group distribution unit that distributes each instruction included in the loop to a plurality of groups based on the degree of dependence between the information generation unit to be generated and each instruction calculated from the generated first matrix, and the plurality of distribution units. Each group has a loop dividing portion for dividing the loop.

According to one aspect, it is possible to divide the loop included in the source code at high speed.

FIG. 1 is a diagram illustrating a configuration of the information processing system 10. FIG. 2 is a flowchart illustrating a compilation process performed by the information processing apparatus 1. FIG. 3 is a diagram illustrating a hardware configuration of the information processing device 1. FIG. 4 is a block diagram of the function of the information processing device 1. FIG. 5 is a flowchart illustrating an outline of the process of S2. FIG. 6 is a flowchart illustrating the details of the process of S2. FIG. 7 is a flowchart illustrating the details of the process of S2. FIG. 8 is a flowchart illustrating the details of the process of S2. FIG. 9 is a specific example for explaining the contents of the intermediate language 22. FIG. 10 is a diagram illustrating a specific example of the dependency information 131. FIG. 11 is a diagram illustrating a specific example of the dependency graph 131a. FIG. 12 is a diagram illustrating a specific example of the first matrix 132. FIG. 13 is a diagram illustrating a specific example of the second matrix 133. FIG. 14 is a diagram illustrating a specific example of the dependency information 131. FIG. 15 is a diagram illustrating a specific example of the first matrix 132. FIG. 16 is a diagram illustrating a specific example of the second matrix 133. FIG. 17 is a diagram illustrating a specific example of the dependency information 131. FIG. 18 is a diagram illustrating a specific example of the first matrix 132. FIG. 19 is a diagram illustrating a specific example of the dependency graph 131a. FIG. 20 is a specific example for explaining the contents of the split loop. FIG. 21 is a flowchart illustrating the process of S2 in the second embodiment. FIG. 22 is a flowchart illustrating the process of S2 in the second embodiment. FIG. 23 is a flowchart illustrating the process of S2 in the second embodiment. FIG. 24 is a diagram illustrating a specific example of the dependency information 131. FIG. 25 is a diagram illustrating a specific example of the first matrix 132. FIG. 26 is a diagram illustrating a specific example of the first matrix 132. FIG. 27 is a diagram illustrating a specific example of the dependency graph 131a. FIG. 28 is a specific example for explaining the contents of the split loop.

[Information processing system configuration]
First, the configuration of the information processing system 10 will be described. FIG. 1 is a diagram illustrating a configuration of the information processing system 10.

As shown in FIG. 1, the information processing system 10 includes, for example, an information processing device 1 composed of one or more physical machines, a storage unit 130 provided inside or outside the information processing device 1, and an operation terminal 5. Including. The operation terminal 5 is, for example, a PC (Personal Computer) used by a worker who compiles the source code, and is connected to the information processing device 1 via a network NW.

When the information processing device 1 (compiler operating in the information processing device 1) comes to the timing of starting compilation (hereinafter, also referred to as compilation start timing), for example, the source code 21 stored in the storage unit 130 is acquired. , The intermediate language 22 is generated by performing a process of compiling the acquired source code 21 (hereinafter, also referred to as a compilation process), and further, an object code 23 is generated from the generated intermediate language 22. The compilation start timing may be, for example, the timing at which the operator gives an instruction to start the compilation process via the operation terminal 5.

Further, when the timing for executing the object code 23 (hereinafter, also referred to as the code execution timing) is reached, the information processing device 1 acquires the object code 23 stored in the storage unit 130, and the object generated by the compilation process. A process for executing the code 23 (hereinafter, also referred to as a code execution process) is performed. Hereinafter, the compilation process and the code execution process performed by the information processing apparatus 1 will be described.

[Compilation process by information processing device]
First, the compilation process performed by the information processing apparatus 1 will be described. FIG. 2 is a flowchart illustrating a compilation process performed by the information processing apparatus 1.

As shown in FIG. 2, the information processing apparatus 1 generates the intermediate language 22 by performing lexical analysis and syntactic analysis of the source code 21 (S1). Then, the information processing device 1 stores, for example, the generated intermediate language 22 in the information storage area 130.

After that, the information processing apparatus 1 optimizes the intermediate language 22 generated in the process of S1 (S2). Specifically, the information processing device 1 performs processing such as loop splitting for each of the loops included in the intermediate language 22.

Subsequently, the information processing apparatus 1 generates the object code 23 from the intermediate language 22 optimized in S1, for example (S3). Then, the information processing device 1 stores, for example, the generated object code 23 in the storage unit 130.

Here, when performing the processing of S2 described with reference to FIG. 2, the information processing apparatus 1 analyzes, for example, the dependency relationship between each instruction included in the loop, and divides each instruction into a plurality of units based on the analysis result. By allocating to each loop, the loop to be analyzed is divided.

However, the number of instruction combinations for which dependency analysis needs to be analyzed increases according to the number of instructions contained in the loop to be divided. Therefore, for example, when the number of instructions included in the loop to be divided is enormous, the information processing apparatus 1 requires a lot of time for loop division, and the source code can be compiled efficiently. It disappears.

Therefore, the information processing device 1 in the present embodiment generates and generates information (hereinafter, also referred to as dependency information) indicating the dependency relationship between each instruction included in the loop of the intermediate language 22 generated from the source code 21. A matrix (hereinafter, also referred to as a first matrix) is generated by transforming the dependency information.

Then, the information processing apparatus 1 distributes each instruction included in the loop into a plurality of groups based on the degree of dependence between the instructions calculated from the generated first matrix, and divides the loop into each of the distributed groups.

That is, the information processing apparatus 1 in the present embodiment analyzes the dependency relationship for each combination of instructions included in the loop to be divided by performing an operation using the first matrix generated from the dependency information. Instead, determine the distribution destination of each command.

As a result, the information processing apparatus 1 can efficiently perform loop splitting, and the time required for compiling the source code 21 can be shortened.

Further, the information processing device 1 specifies a loop splitting method in which the dependency between the instructions included in the different split loops is the sparsest by determining the distribution destination of each instruction included in the loop to be split by calculation. It becomes possible to do. Therefore, the information processing apparatus 1 can also shorten the execution time of the object code 23 generated from the source code 21 by dividing the loop according to the specified method.

[Hardware configuration of information processing system]
Next, the hardware configuration of the information processing system 10 will be described. FIG. 3 is a diagram illustrating a hardware configuration of the information processing device 1.

As shown in FIG. 3, the information processing device 1 includes a CPU 101 which is a processor, a memory 102, an external interface (I / O unit) 103, and a storage medium 104. The parts are connected to each other via the bus 105.

The storage medium 104 has, for example, a program storage area (not shown) for storing a program 110 (compiler) for performing a compilation process. Further, the storage medium 104 has, for example, a storage unit 130 (hereinafter, also referred to as an information storage area 130) for storing information used when performing a compilation process. The storage medium 104 may be, for example, an HDD (Hard Disk Drive) or an SSD (Sock State Drive).

The CPU 101 executes the program 110 loaded from the storage medium 104 into the memory 102 to perform a compilation process.

Further, the external interface 103 communicates with, for example, the operation terminal 5.

[Information processing system functions]
Next, the function of the information processing system 10 will be described. FIG. 4 is a block diagram of the function of the information processing device 1.

As shown in FIG. 4, the information processing device 1 organically cooperates with hardware such as the CPU 101 and the memory 102 and the program 110 to form a division determination unit 111, an information generation unit 112, and information. Various functions including the management unit 113, the group distribution unit 114, the loop splitting unit 115, and the proximity determination unit 116 are realized.

Further, the information processing apparatus 1 stores, for example, the dependency information 131, the first matrix 132, and the second matrix 133 in the information storage area 130, as shown in FIG.

The division determination unit 111 determines, for example, whether or not to perform loop division for each loop included in the intermediate language 22 generated from the source code 21. Specifically, for example, when the intermediate language 22 contains a loop in which the number of registers and the number of streams used at the time of execution exceeds the number of registers and the number of streams of the CPU 101, the division determination unit 111 performs loop division for the loop. Make a judgment to that effect.

The information generation unit 112 generates dependency information 131 indicating the dependency relationship between each instruction included in the loop to be divided. Specifically, the information generation unit 112 generates the dependency information 131 corresponding to the loop when the division determination unit 111 determines that the loop split is performed. Further, the information generation unit 112 generates the first matrix 132 by converting the generated dependency information 131 into a matrix.

The information management unit 113 stores, for example, the dependency information 131 and the first matrix 132 generated by the information generation unit 112 in the information storage area 130.

The group distribution unit 114 distributes each instruction included in the loop to be divided into a plurality of groups based on the degree of dependence between the instructions calculated from the first matrix 132 generated by the information generation unit 112.

The loop splitting unit 115 performs loop splitting on the loop to be divided for each group distributed by the group sorting unit 114.

The proximity determination unit 116 determines, for example, whether or not a plurality of instructions for accessing each data stored at adjacent addresses in the memory 102 are included in the loop to be divided.

Then, for example, when the proximity determination unit 116 determines that a plurality of instructions for accessing each data stored at adjacent addresses in the memory 102 are included in the loop to be divided, the information generation unit 112 Generates dependency information 131 corresponding to the dependency relationship between each instruction included in the loop to be divided and the positional relationship in the memory 102 of the data accessed by each instruction. The explanation of the second matrix 133 will be described later.

[Outline of the first embodiment]
Next, the outline of the first embodiment will be described. Specifically, the outline of the process of S2 described with reference to FIG. 2 will be described. FIG. 5 is a flowchart illustrating an outline of the process of S2.

The information generation unit 112 of the information processing device 1 generates dependency information 131 indicating the dependency relationship between each instruction included in the loop of the intermediate language 22 generated from the source code 21 (S11).

Then, the information generation unit 112 generates the first matrix 132 by converting the dependency information 131 generated in the process of S11 into a matrix (S12).

Subsequently, the group distribution unit 114 of the information processing apparatus 1 sets a plurality of instructions included in the loop to be divided based on the degree of dependence between the instructions calculated from the first matrix 132 generated in the process of S12. Allocate to groups (S13).

After that, the loop splitting unit 115 of the information processing apparatus divides the loop to be split for each group sorted in the process of S13 (S14).

As a result, the information processing apparatus 1 can efficiently perform loop splitting, and the time required for compiling the source code 21 can be shortened.

Further, the information processing device 1 specifies a loop splitting method in which the dependency between the instructions included in the different split loops is the sparsest by determining the distribution destination of each instruction included in the loop to be split by calculation. It becomes possible to do. Therefore, the information processing apparatus 1 can also shorten the execution time of the object code 23 generated from the source code 21 by dividing the loop according to the specified method.

[Details of the first embodiment]
Next, the details of the first embodiment will be described. 6 to 9 are flowcharts illustrating the details of the process of S2 described with reference to FIG. 9 to 20 are views for explaining the details of the process of S2.

As shown in FIG. 6, the division determination unit 111 of the information processing device 1 determines whether or not the intermediate language 22 stored in the information storage area 130 includes a loop to be divided (S21). Hereinafter, specific examples of the intermediate language 22 will be described.

[Specific examples of intermediate languages]
FIG. 9 is a specific example for explaining the contents of the intermediate language 22. Specifically, FIG. 9 is a specific example for explaining the intermediate language 22 when the source code 21 is represented by an assembler instruction.

The intermediate language 22 shown in FIG. 9 includes "mov reg_i, 1" which is an instruction indicating that the variable reg_i is set to 1 as an initial value, and "LABEL1:" which is a label indicating the start position of the loop.

Further, the intermediate language 22 shown in FIG. 9 is an instruction indicating that the value stored in the reg_ith position of the array A is set in the variable reg1, which is an instruction "load reg1, mem" A (reg_i) "" and the array B. Includes "load reg2, mem" B (reg_i) "" which is an instruction indicating that the value stored in the reg_i th position of is set in the variable reg2.

Further, the intermediate language 22 shown in FIG. 9 is an instruction indicating that the value calculated by adding the value set in the variable reg1 and the value set in the variable reg2 is set in the variable reg3. It includes "add reg3, reg1, reg2" and "load reg4, mem" C (reg_i) "" which is an instruction indicating that the value stored in the reg_ith position of the array C is set in the variable reg4.

Further, the intermediate language 22 shown in FIG. 9 is an instruction indicating that the value stored in the reg_i position of the array D is set in the variable reg5, which is an instruction "load reg5, mem" D (reg_i) "" and the array C Includes "load reg6, mem" C (reg_i + 10) "" which is an instruction indicating that the data stored in the reg_i + 10th position of the above is set in the variable reg6.

In addition, "load reg7, mem" E (reg_i) "", which is an instruction indicating that the value stored in the reg_ith position of the array E is set in the variable reg7, and the value set in the variable reg3 and the variable reg4. Includes "add reg8, reg3, reg4" which is an instruction indicating that the value calculated by adding the value set in is set in the variable reg8.

Further, the intermediate language 22 shown in FIG. 9 is an instruction indicating that the value calculated by adding the value set in the variable reg4 and the value set in the variable reg8 is set in the variable reg9. Adding the value set in the variable reg9 to the value calculated by multiplying "add reg9, reg4, reg8" by the value set in the variable reg3 and the value set in the variable reg5. Includes "madd reg10, reg3, reg5, reg9" which is an instruction indicating that the value calculated by the above is set in the variable reg10.

Further, in the intermediate language 22 shown in FIG. 9, the value set in the variable reg10 is added to the value calculated by multiplying the value set in the variable reg5 and the value set in the variable reg6. Add the value set in the variable reg7 and the value set in the variable reg11 to the command "madd reg11, reg5, reg6, reg10" indicating that the value calculated by this is set in the variable reg11. It includes "add reg12, reg7, reg11" which is an instruction indicating that the value calculated by the above is set in the variable reg12.

Further, the intermediate language 22 shown in FIG. 9 is an instruction indicating that the value set in the variable reg12 is stored in the reg_ith position of the array F, which is an instruction "store mem" F (reg_i) ", reg12" and the variable reg_i. It includes "add reg_i, reg_i, 1" which is an instruction indicating that the value calculated by adding 1 to the value set in is set in the variable reg_i.

Further, the intermediate language 22 shown in FIG. 9 has "cmp reg_i, 100" which is an instruction indicating to compare the value set in the variable reg_i with 100, and the value set in the variable reg_i is 100 or less. In the case of, it branches to "LABEL1:" indicating the start position of the loop, and when the value set in the variable reg_i exceeds 100, it is an instruction indicating the end of the loop "bleicc, LABAL1". Including.

Hereinafter, the number described at the left end of each instruction in the intermediate language 22 shown in FIG. 9 is also referred to as an instruction number of each instruction. Further, hereinafter, each of the instructions whose instruction numbers are "1" to "13" will also be referred to as instruction 1 to instruction 13.

Returning to FIG. 6, when the intermediate language 22 stored in the information storage area 130 includes a loop to be divided (YES in S22), the information generation unit 112 of the information processing apparatus 1 is divided by the process of S21. Dependency information 131 indicating the dependency relationship between each instruction included in the loop determined to be the target loop is generated (S23). Hereinafter, a specific example of the dependency information 131 will be described.

[Specific example of dependency information]
FIG. 10 is a diagram illustrating a specific example of the dependency information 131.

The dependency information 131 shown in FIG. 10 is a "data dependency list" which is a list of "instruction numbers" in which instruction numbers of each instruction included in the intermediate language 22 are stored and instruction numbers of instructions having a dependency relationship with each instruction. , And a "group name" in which the name of the group (split loop) including each instruction is stored is included as an item.

Specifically, in the intermediate language 22 described with reference to FIG. 9, the variable reg1 whose value is set by the instruction 1 is referred to only by the instruction 3. Therefore, for example, as shown in FIG. 10, the information generation unit 112 stores "3" in the "data dependency list" of the information whose "instruction number" is "1".

Further, in the intermediate language 22 described with reference to FIG. 9, the variable reg2 whose value is set by the instruction 2 is referred to only by the instruction 3. Therefore, for example, as shown in FIG. 10, the information generation unit 112 stores "3" in the "data dependency list" of the information whose "instruction number" is "2".

Further, in the intermediate language 22 described with reference to FIG. 9, the instruction 3 refers to the variable reg1 whose value is set by the instruction 1 and the variable reg2 whose value is set by the instruction 2. Further, in the intermediate language 22 described with reference to FIG. 9, the variable reg3 whose value is set by the instruction 3 is referred to by the instruction 8 and the instruction 10. Therefore, for example, as shown in FIG. 10, the information generation unit 112 adds "1", "2", "8", and "10" to the "data dependency list" of the information whose "instruction number" is "3". Remember.

Further, the information generation unit 112 stores, for example, the instruction number of each instruction as the initial value of the "group name" corresponding to each instruction, as shown in FIG. The description of other information included in FIG. 10 will be omitted.

The information generation unit 112 may generate the dependency graph 131a corresponding to the content included in the dependency information 131 in the process of S23. Hereinafter, a specific example of the dependency graph 131a will be described.

[Specific example of dependency graph (1)]
FIG. 11 is a diagram illustrating a specific example of the dependency graph 131a. Hereinafter, an edge showing a bidirectional relationship is also referred to as a bidirectional edge, and an edge showing a unidirectional relationship is also referred to as a unidirectional edge.

Specifically, in the dependency graph 131a shown in FIG. 11, between the node corresponding to the instruction 1 and the node corresponding to the instruction 3, the node corresponding to the instruction 2 and the node corresponding to the instruction 3 are connected to the instruction 3. Between the corresponding node and the node corresponding to the instruction 8, between the node corresponding to the instruction 3 and the node corresponding to the instruction 10, and between the node corresponding to the instruction 4 and the node corresponding to the instruction 8. Bidirectional edges are set for each. Further, in the dependency graph 131a shown in FIG. 11, between the node corresponding to the instruction 4 and the node corresponding to the instruction 9, the node corresponding to the instruction 5 and the node corresponding to the instruction 10 correspond to the instruction 5. Between the node and the node corresponding to the instruction 11, between the node corresponding to the instruction 6 and the node corresponding to the instruction 11, and between the node corresponding to the instruction 7 and the node corresponding to the instruction 12. Bidirectional edges are set. Further, in the dependency graph 131a shown in FIG. 11, the node corresponding to the instruction 8 and the node corresponding to the instruction 9 correspond to the instruction 10, and the node corresponding to the instruction 9 and the node corresponding to the instruction 10 correspond to the instruction 10. Between the node and the node corresponding to the instruction 11, between the node corresponding to the instruction 11 and the node corresponding to the instruction 12, and between the node corresponding to the instruction 12 and the node corresponding to the instruction 13. Bidirectional edges are set. That is, 15 bidirectional edges are set in the dependency graph 131a shown in FIG.

In the dependency information 131 described with reference to FIG. 10, for example, "3" is stored in the "data dependency list" of the information in which the "instruction number" is "1", and the "instruction number" is "3". "1" is stored in the "data dependency list" of certain information. Therefore, for example, as shown in FIG. 11, the information generation unit 112 sets a bidirectional edge between the node corresponding to the instruction 1 and the node corresponding to the instruction 3.

Further, in the dependency information 131 described with reference to FIG. 10, for example, "3" is stored in the "data dependency list" of the information in which the "instruction number" is "2", and the "instruction number" is "3". "2" is stored in the "data dependency list" of the information. Therefore, for example, as shown in FIG. 11, the information generation unit 112 sets a bidirectional edge between the node corresponding to the instruction 2 and the node corresponding to the instruction 3.

Further, in the dependency information 131 described with reference to FIG. 10, for example, "8" is stored in the "data dependency list" of the information in which the "instruction number" is "3", and the "instruction number" is "8". "3" is stored in the "data dependency list" of the information. Therefore, for example, as shown in FIG. 11, the information generation unit 112 sets a bidirectional edge between the node corresponding to the instruction 3 and the node corresponding to the instruction 8. The description of other information included in FIG. 11 will be omitted.

Returning to FIG. 6, the information generation unit 112 generates the first matrix 132 by converting the dependency information 131 generated in the process of S23 into a matrix (S24). Hereinafter, a specific example of the first matrix 132 will be described.

[Specific example of the first matrix]
FIG. 12 is a diagram illustrating a specific example of the first matrix 132.

In the first matrix 132 shown in FIG. 12, the rows corresponding to each of "1" to "13" are the rows corresponding to each of the instructions 1 to 13, and correspond to each of "1" to "13". The columns to be used are the columns corresponding to each of the instructions 1 to 13. Further, each element (each column) of the first matrix 132 shown in FIG. 12 has a value of "1" indicating that there is a dependency between the instruction corresponding to the row and the instruction corresponding to the column. Alternatively, "0", which is a value indicating that there is no dependency between the instruction corresponding to the row and the instruction corresponding to the column, is stored.

Specifically, in the dependency information 131 shown in FIG. 10, "3" is stored in the "data dependency list" of the information in which the "instruction number" is "1". Therefore, as shown in FIG. 12, the information generation unit 112 stores, for example, "1" in the column included in the column corresponding to the instruction 3 among the columns included in the row corresponding to the instruction 1.

Further, in the dependency information 131 shown in FIG. 10, "3" is stored in the "data dependency list" of the information in which the "instruction number" is "2". Therefore, as shown in FIG. 12, the information generation unit 112 stores, for example, "1" in the column included in the column corresponding to the instruction 3 among the columns included in the row corresponding to the instruction 2.

Further, in the dependency information 131 shown in FIG. 10, "1", "2", "8" and "10" are stored in the "data dependency list" of the information whose "instruction number" is "3". There is. Therefore, as shown in FIG. 12, the information generation unit 112 is included in, for example, a column included in the column corresponding to the instruction 1 and a column corresponding to the instruction 2 among the columns included in the row corresponding to the instruction 3. "1" is stored in each of the column, the column included in the column corresponding to the instruction 8, and the column included in the column corresponding to the instruction 10. The description of other information included in FIG. 12 will be omitted.

Returning to FIG. 6, the group distribution unit 114 of the information processing apparatus 1 generates a second matrix 133 indicating the degree of dependence between each instruction from the first matrix 132 generated in the process of S24 (S25). Hereinafter, a specific example of the second matrix will be described.

[Specific example of the second matrix]
FIG. 13 is a diagram illustrating a specific example of the second matrix 133.

In the second matrix 133 shown in FIG. 13, the rows corresponding to each of "1" to "13" are the rows corresponding to each of the instructions 1 to 13, and correspond to each of "1" to "13". The columns to be used are the columns corresponding to each of the instructions 1 to 13. Further, in each element (each column) of the second matrix 133 shown in FIG. 13, the degree of dependence between the instruction corresponding to the row and the instruction corresponding to the column is stored.

Here, the degree of dependence between the instructions may be, for example, a clustering index value calculated by the Newman algorithm. In this case, the group distribution unit 114 calculates the clustering index value ΔQ by using, for example, the following equation 1.

ΔQ = e _ij + e _ji -2a _i a _j = 2 (e _ij- a _i a _j ) (Equation 1)

In Equation 1 above, the variable _ai is the number of unidirectional edges between the node corresponding to the instruction i and the node for another instruction among the unidirectional edges included in the dependency graph 131a described with reference to FIG. The variable a _j indicates the ratio, and the variable a _j indicates the ratio of the number of unidirectional edges between the node corresponding to the instruction j and the node for other instructions among the unidirectional edges included in the dependency graph 131a described with reference to FIG. Shown. Further, in Equation 1, the variable e _ij is the number of unidirectional edges between the node corresponding to the instruction i and the node corresponding to the instruction j among the unidirectional edges included in the dependency graph 131a described with reference to FIG. Indicates the ratio of.

Specifically, the state shown by the dependency graph 131a described with reference to FIG. 11 is a state in which 15 bidirectional edges are included, which is the same as a state in which 30 unidirectional edges are included. Further, in the dependency graph 131a described with reference to FIG. 11, the number of unidirectional edges between the node corresponding to the instruction 12 and the other node, and the unidirectional edge between the node corresponding to the instruction 13 and the other node. The number and the number of unidirectional edges between the node corresponding to the instruction 12 and the node corresponding to the instruction 13 are 3, 1, and 1, respectively. Therefore, for example, when the instruction i is the instruction 12 and the instruction j is the instruction 13, the group distribution unit 114 calculates “0.060” as the clustering index value ΔQ as in the following equation (2). To do.

2 * ((1/30)-(3/30) * (1/30)) = 0.06 ... (Equation 2)

Therefore, as shown in FIG. 13, the group distribution unit 114 stores, for example, "0.060" in the column included in the column corresponding to the instruction 13 among the columns included in the row corresponding to the instruction 12. ..

The clustering index value ΔQ when the instruction i is the instruction 13 and the instruction j is 12, is the same as the clustering index value ΔQ when the instruction i is the instruction 12 and the instruction j is 13. Therefore, for example, when the group distribution unit 114 calculates the clustering index value ΔQ when the instruction i is the instruction 12 and the instruction j is 13, the instruction i is the instruction 13 and the instruction j is The clustering index value ΔQ in the case of 12 may not be calculated. Then, in this case, as shown in FIG. 13, the group distribution unit 114 calculates the clustering index value ΔQ in the column included in the column corresponding to the instruction 12 among the columns included in the row corresponding to the instruction 13. It may store "0.000" which is a value indicating that the above is not performed. The description of other information included in FIG. 13 will be omitted.

Returning to FIG. 7, the group distribution unit 114 determines the distribution destination of each instruction corresponding to the maximum value among the values of the elements of the second matrix 133 generated in the process of S25 to the same group (S31).

Specifically, in the second matrix 133 described with reference to FIG. 13, among the columns included in the row corresponding to the instruction 12, the column included in the column corresponding to the instruction 13 and the column included in the row corresponding to the instruction 7. Among them, the maximum value "0.060" is set in each of the columns included in the column corresponding to the instruction 12. Therefore, the group distribution unit 114 specifies the combination of the instruction 12 and the instruction 13 or the combination of the instruction 7 and the instruction 12 among the combinations of the instructions corresponding to each element of the second matrix 133 described with reference to FIG. Then, for example, when the combination of the instruction 12 and the instruction 13 is specified, the group distribution unit 114 determines that the instruction 12 and the instruction 13 are distributed to the same group.

Then, the group distribution unit 114 determines the distribution destination in the process of S31 among the rows in the second matrix 133 generated in the process of S25 or the row in the second matrix 133 regenerated in the process of S33 (process described later). A plurality of rows corresponding to each instruction are converted into a single row having the sum of the elements of the same column in the plurality of rows as an element, and reproduced by the processing of the second matrix 133 or S33 generated in the processing of S25. Among the columns in the second matrix 133 formed, a single column having a plurality of columns corresponding to each instruction whose distribution destination is determined in the process of S31 and the sum of the elements of the same row in the plurality of columns as elements. The first matrix 132 is regenerated by converting to (S32).

Specifically, the group distribution unit 114 updates the dependency information 131 based on the determination result in the process of S31, for example. Then, the group distribution unit 114 regenerates the first matrix 132 by referring to the updated dependency information 131, for example. Hereinafter, a specific example of the processing of S32 will be described.

[Specific example of processing of S32]
14 and 15 are diagrams for explaining a specific example of the process of S32.

For example, when the instructions 12 and the instruction 13 are determined to be distributed to the same group in the process of S31, the group distribution unit 114 has the "instruction number" of "13" as shown in FIG. Dependency information 131 is updated so that the value stored in the "group name" of the information is updated to "12" which is the value stored in the "group name" of the information whose "instruction number" is "12". To do.

After that, as shown in FIG. 15, for example, the group distribution unit 114 sets a value set in a column included in the line corresponding to the instruction 12 in a column included in the line corresponding to the instruction 12 and an instruction 13 The value calculated by adding the value set in the column included in the row corresponding to is added for each same column is stored. Further, as shown in FIG. 15, for example, the group distribution unit 114 has a value set in a column included in the column corresponding to the instruction 12 and a value set in the column included in the column corresponding to the instruction 12, and the instruction 13 The value calculated by adding the value set in the column included in the column corresponding to is added for each row is stored.

Specifically, in the first matrix 132 described with reference to FIG. 12, for example, among the columns included in the row corresponding to the instruction 12, "1" is stored in the column included in the column corresponding to the instruction 7. Therefore, among the columns included in the row corresponding to the instruction 13, "0" is stored in the column included in the column corresponding to the instruction 7. Therefore, as shown in FIG. 15, the group distribution unit 114 has, for example, "1" and "0" in the columns included in the column corresponding to the instruction 7 among the columns included in the row corresponding to the instruction 12. Memorize "1" which is the sum of.

Further, in the first matrix 132 described with reference to FIG. 12, for example, among the columns included in the row corresponding to the instruction 11, "1" is stored in the column included in the column corresponding to the instruction 12. Among the columns included in the row corresponding to the instruction 11, "0" is stored in the column included in the column corresponding to the instruction 13. Therefore, as shown in FIG. 15, the group distribution unit 114 has, for example, "1" and "0" in the columns included in the column corresponding to the instruction 12 among the columns included in the row corresponding to the instruction 11. Memorize "1" which is the sum of.

Further, in the first matrix 132 described with reference to FIG. 12, for example, among the columns included in the row corresponding to the instruction 12, "0" is stored in the column included in the column corresponding to the instruction 12. Of the columns included in the row corresponding to the instruction 12, "1" is stored in the column included in the column corresponding to the instruction 13, and among the columns included in the row corresponding to the instruction 13, the instruction 12 "1" is stored in the column corresponding to the column corresponding to, and "0" is stored in the column included in the column corresponding to the instruction 13 among the columns included in the row corresponding to the instruction 13. It is remembered. Therefore, as shown in FIG. 15, the group distribution unit 114 has, for example, "0" and "1" in the columns included in the column corresponding to the instruction 12 among the columns included in the row corresponding to the instruction 12. , "2" which is the sum of "1" and "0" is stored. The description of other information included in FIG. 15 will be omitted.

Returning to FIG. 7, the group distribution unit 114 regenerates the second matrix 133 from the first matrix 132 regenerated in the process of S32 (S33).

Specifically, for example, the group distribution unit 114 generates (regenerates) the second matrix 133 shown in FIG. 16 by performing the same processing as the processing of S25 on the first matrix 132 described with reference to FIG. To do.

Then, the group distribution unit 114 determines whether or not the number of distribution destination groups of each instruction included in the loop determined to be the division target in the process of S21 has reached a predetermined number or less (S34).

Specifically, values (12 types of values) from "1" to "12" are stored in the "group name" of the dependency information 131 described with reference to FIG. Therefore, in this case, the group distribution unit 114 specifies "12" as the number of distribution destination groups of each instruction included in the loop determined to be the division target in the process of S21. Then, for example, when the predetermined number in the process of S34 is "2", the group distribution unit 114 has the number of groups to which each instruction is distributed included in the loop determined to be the division target in the process of S21. It is determined that the number has not reached the predetermined number or less.

On the other hand, only "1" and "5" (two kinds of values) are stored in the "group name" of the dependency information 131 shown in FIG. Therefore, in this case, the group distribution unit 114 specifies "2" as the number of distribution destination groups of each instruction included in the loop determined to be the division target in the process of S21. Then, for example, when the predetermined number in the process of S34 is "2", the group distribution unit 114 has the number of groups to which each instruction is distributed included in the loop determined to be the division target in the process of S21. It is determined that the number has reached a predetermined number or less.

The group distribution unit 114 generates, for example, the first matrix 132 shown in FIG. 18 in response to the generation of the dependency information 131 shown in FIG. Further, the group distribution unit 114 generates, for example, the dependency graph 131a shown in FIG. 19 in response to the generation of the dependency information 131 shown in FIG. Hereinafter, a specific example of the dependency graph 131a shown in FIG. 19 will be described.

[Specific example of dependency graph (2)]
In the dependency graph 131a shown in FIG. 19, the node group corresponding to each of instruction 1, instruction 2, instruction 3, instruction 4, instruction 8 and instruction 9, and instruction 5, instruction 6, instruction 7, instruction 10, instruction 11, The node group corresponding to each of the instruction 12 and the instruction 13 is included in a different group.

Then, in the dependency graph 131a shown in FIG. 19, between the group including the node corresponding to the instruction 1 and the group including the node corresponding to the instruction 5, the node corresponding to the instruction 3 and the node corresponding to the instruction 10 are connected. An edge between the two, and an edge between the node corresponding to the instruction 9 and the node corresponding to the instruction 10 are set.

That is, in the dependency graph 131a shown in FIG. 19, instruction 1, instruction 2, instruction 3, instruction 4, instruction 8, and instruction 9 are included in one of the split loops, and instruction 5, instruction 6, instruction 7, and instruction 10 are included. When loop splitting is performed so that the instructions 11, instruction 12, and instruction 13 are included in the other division loop, the number of edges between the instructions included in each of the different division loops can be suppressed to two. It is shown that.

Returning to FIG. 9, when it is determined that the number of distribution destination groups of each instruction included in the loop determined to be the division target in the process of S21 has reached a predetermined number or less (YES in S41), the information processing device 1 The loop splitting unit 115 of the above performs loop splitting of the loop determined to be the split target in the processing of S21 according to the contents of the second matrix 133 generated in the processing of S33 (S42).

After that, the information processing device 1 ends the processing of S2. When it is determined that the intermediate language 22 stored in the information storage area 130 does not include the loop to be divided, the information processing device 1 similarly ends the process of S2 (NO in S22).

On the other hand, when it is determined that the number of distribution destination groups of each instruction included in the loop determined to be the division target in the processing of S21 has not reached a predetermined number or less (NO in S41), the group distribution unit 114 Performs the processing after S32 again. Hereinafter, a specific example of the split loop after performing the loop split will be described.

[Specific example of split loop]
FIG. 20 is a specific example for explaining the contents of the split loop. FIG. 20A is a specific example for explaining one of the divided loops, and FIG. 20B is a specific example for explaining the other of the divided loops.

The split loop shown in FIG. 20A includes instruction 1, instruction 2, instruction 3, instruction 4, instruction 8 and instruction 9 among the intermediate languages described in FIG.

Then, the split loop shown in FIG. 20 (A) is an instruction indicating that the value set in the variable reg3 is stored in the reg_ith position of the array tp_array1 after the instruction 1 or the like, "store tp_array1 (reg_i), reg3". , And "store tp_array2 (reg_i), reg9", which is an instruction indicating that the value set in the variable reg9 is stored in the reg_ith position of the array tp_array2.

On the other hand, the split loop shown in FIG. 20B includes instruction 5, instruction 6, instruction 7, instruction 10, instruction 11, instruction 12, and instruction 13 among the intermediate languages described in FIG.

Then, the split loop shown in FIG. 20 (B) is an instruction indicating that the value stored in the reg_i th of the array tm_array1 is set in the variable reg3 before the instruction 5 and the like, "load reg3, tp_array1 (reg_i). ) ”, And“ load reg9, tp_array2 (reg_i) ”which is an instruction indicating that the value stored in the rig_i th of the array tp_array2 is set in the variable reg9.

That is, as described with reference to FIG. 19, the value set in the variable reg3 (value corresponding to the instruction 3) and the value set in the variable reg9 (value corresponding to the instruction 9) are shown in FIG. 20B. It is referred to in instruction 10 included in the split loop shown in. Therefore, the split loop shown in FIG. 20A includes an instruction to store the value set in the variable reg3 and the value set in the variable reg9 in a temporary array. Further, the division loop shown in FIG. 20B includes an instruction for fetching a value stored in a temporary array.

As described above, the information processing apparatus 1 in the present embodiment generates the dependency information 131 indicating the dependency relationship between each instruction included in the loop of the intermediate language 22 generated from the source code 21, and generates the dependency information 131. The first matrix 132 is generated by converting to a matrix.

Then, the information processing device 1 distributes each instruction included in the loop into a plurality of groups based on the degree of dependence between the instructions calculated from the generated first matrix 132, and divides the loop into each of the distributed groups. ..

That is, the information processing device 1 in the present embodiment performs an operation using the first matrix 132 generated from the dependency information 131, thereby analyzing the dependency relationship for each combination of instructions included in the loop to be divided. To determine the distribution destination of each command without performing.

As a result, the information processing apparatus 1 can efficiently perform loop splitting, and the time required for compiling the source code 21 can be shortened.

Further, the information processing device 1 specifies a loop splitting method in which the dependency between the instructions included in the different split loops is the sparsest by determining the distribution destination of each instruction included in the loop to be split by calculation. It becomes possible to do. Therefore, the information processing apparatus 1 can also shorten the execution time of the object code 23 generated from the source code 21 by dividing the loop according to the specified method.

When the information processing device 1 determines in the processing of S21 and S22 that the intermediate language 22 stored in the information storage area 130 includes a plurality of loops to be divided, the processing after S23 is to be divided. It may be performed for each loop.

Further, the information processing apparatus 1 may use an algorithm other than the Newsman algorithm (for example, an algorithm such as a label propagation method or a K-means method) when calculating the clustering index value in the processing in S25.

[Second Embodiment]
Next, the details of the second embodiment will be described. 21 to 23 are flowcharts illustrating the process of S2 in the second embodiment. In addition, FIGS. 24 to 28 are diagrams for explaining the details of the processing of S2 in the second embodiment.

The compilation process in the second embodiment is different from the compilation process in the first embodiment, and the loop is divided by referring to the positional relationship of the data accessed by each instruction in the memory 102.

As shown in FIG. 21, the division determination unit 111 determines whether or not the intermediate language 22 stored in the information storage area 130 includes a loop to be divided (S51).

Then, when it is determined that the intermediate language 22 stored in the information storage area 130 includes the loop to be divided (YES in S52), the proximity determination unit 116 of the information processing device 1 is, for example, in the memory 102. In (S53), it is determined whether or not a plurality of instructions for accessing each data stored in adjacent addresses are included in the loop to be divided (S53).

Specifically, the proximity determination unit 116 determines, for example, whether or not a plurality of instructions for accessing each data stored in the same array are included in the same loop to be divided.

That is, when the object code 23 is executed, for example, when the first data of the first array is accessed with the execution of the first instruction, the CPU 101 performs each data of the first array stored in the memory 102. Data of a predetermined size including the above is temporarily stored in a cache memory (not shown), and the first data stored in the cache memory is accessed.

Then, for example, when access to the second data in the first array occurs due to the execution of the second instruction different from the first instruction, the CPU 101 determines that each data in the first array is still stored in the cache memory. , Access the second data stored in the cache memory. On the other hand, when each data in the first array has already been expelled from the cache memory due to the occurrence of access to other data, the CPU 101 has a predetermined size including each data in the first array stored in the memory 102. Data is stored in the cache memory again, and the second data stored in the cache memory is accessed.

Therefore, for example, when the first instruction and the second instruction as described above are included in the loop to be divided, the information processing apparatus 1 loops so that the first instruction and the second instruction are included in the same division loop. The division is performed so that the execution timing of the first instruction and the execution timing of the second instruction are brought close to each other.

As a result, the information processing apparatus 1 can suppress the probability that the first data is expelled from the cache memory during the execution of the loop to be divided. Therefore, the information processing device 1 can reduce the frequency of occurrence of the process of re-storing the first data in the cache memory, and can shorten the execution time of the object code 23.

Therefore, in the process of S53, the proximity determination unit 116 identifies a plurality of instructions for accessing each data included in the same array from the instructions included in the loop to be divided, for example.

Note that the proximity determination unit 116 may specify, for example, a plurality of instructions for accessing data within a range corresponding to a predetermined rotation speed among the data included in a certain array.

Returning to FIG. 21, the information generation unit 112 shows the dependency relationship between each instruction included in the loop determined to be the division target in the process of S51, and the positional relationship of the data accessed by each instruction in the memory 102. Dependency information 131 is generated (S54).

[Specific example of dependency information]
FIG. 24 is a diagram illustrating a specific example of the dependency information 131.

The dependency information 131 shown in FIG. 24 refers to the instruction numbers of a plurality of instructions for accessing each data stored at adjacent addresses in the memory 102, in addition to the items included in the dependency information 131 described with reference to FIG. It has a "cache sharing dependency list" which is a list as an item.

Specifically, in the intermediate language 22 described with reference to FIG. 9, the data stored in the array C is referred to in each of the instruction 4 and the instruction 6. Therefore, for example, as shown in FIG. 24, the information generation unit 112 stores "6" in the "cache sharing dependency list" of the information whose "instruction number" is "4", and the "instruction number" is "6". "4" is stored in the "cache sharing dependency list" of the information. Further, in this case, the information generation unit 112 does not have any information in the "cache sharing dependency list" of the information whose "instruction number" is other than "4" and "6", for example, as shown in FIG. Memorize the "-" indicating.

Returning to FIG. 21, the information generation unit 112 generates the first matrix 132 by converting the dependency information 131 generated in the process of S54 into a matrix (S55). Hereinafter, a specific example of the first matrix 132 will be described.

[Specific example of the first matrix]
FIG. 25 is a diagram illustrating a specific example of the first matrix 132.

Each element (each column) of the first matrix 132 shown in FIG. 25 has a value of "1" indicating that there is a dependency between the instruction corresponding to the row and the instruction corresponding to the column, in the row. "5", which is a value indicating that the positions of the data to be accessed by the corresponding instruction and the instruction corresponding to the column are close to each other in the memory 102, or the instruction corresponding to the row and the instruction corresponding to the column. There is no dependency between the two, and "0", which is a value indicating that the positions in the memory 102 of the data accessed by each of the instruction corresponding to the row and the instruction corresponding to the column are not close to each other, is stored. Will be done.

Specifically, in the dependency information 131 described with reference to FIG. 24, "6" is stored in the "cache sharing dependency list" of the information in which the "instruction number" is "4", and the "instruction number" is "4". "4" is stored in the "cache sharing dependency list" of the information which is "6". Therefore, as shown in FIG. 25, the information generation unit 112 stores, for example, "5" in the column included in the column corresponding to the instruction 6 among the columns included in the row corresponding to the instruction 4. Further, as shown in FIG. 25, the information generation unit 112 stores, for example, "5" in the column included in the column corresponding to the instruction 4 among the columns included in the row corresponding to the instruction 6.

That is, the effect on the execution time of the object code 23 due to the increase in the number of cache misses is greater than the effect on the execution time of the object code 23 due to the inclusion of a plurality of dependent instructions in different split loops. It can be judged that it is large. Therefore, as shown in FIG. 25, the information generation unit 112 has, for example, that the values indicating that the positions of the data accessed by each instruction in the memory 102 are close to each other have a dependency relationship between the instructions. The first matrix 132 is generated so as to be larger than the value indicating.

Returning to FIG. 21, the group distribution unit 114 generates a second matrix 133 indicating the degree of dependence between each instruction from the first matrix 132 generated in the process of S55 (S56).

Then, as shown in FIG. 22, the group distribution unit 114 determines the distribution destination of a plurality of instructions corresponding to the maximum values among the values of the elements of the second matrix 133 generated in the process of S56 to the same group. (S61).

Subsequently, the group distribution unit 114 corresponds to each instruction whose distribution destination is determined by the processing of S61 among the rows in the second matrix 133 generated by the processing of S56 or the second matrix 133 regenerated by the processing of S63. The second row is converted into a single row whose elements are the sum of the elements of the same column in the plurality of rows, and regenerated by the processing of the second matrix 133 or S63 generated by the processing of S56. Of the columns in the matrix 133, a plurality of columns corresponding to each instruction whose distribution destination is determined in the process of S61 are converted into a single column having the sum of the elements of the same row in the plurality of columns as elements. Regenerates the first matrix 132 (S62).

Further, the group distribution unit 114 regenerates the second matrix 133 from the first matrix 132 regenerated in the process of S62 (S63).

After that, the group distribution unit 114 determines whether or not the number of distribution destination groups of each instruction included in the loop determined to be the division target in the process of S51 has reached a predetermined number or less (S64).

Specifically, for example, when the predetermined number in the processing of S64 is "2" and the first matrix 132 shown in FIG. 26 is generated, the group distribution unit 114 is a division target in the processing of S51. It is determined that the number of distribution destination groups of each instruction included in the loop determined to be present has reached a predetermined number or less.

As shown in FIG. 27, in the dependency graph 131a corresponding to the first matrix 132 shown in FIG. 26, each instruction (instruction 4 and instruction 6) that refers to each data stored in the same array is distributed to the same group. Has been done.

Subsequently, when it is determined that the number of the distribution destination groups of each instruction included in the loop determined to be the division target in the processing of S51 has reached a predetermined number or less (YES in S71), the loop division unit 115 determines. According to the contents of the second matrix 133 generated in the process of S63, the loop split of the loop determined to be the split target in the process of S51 is performed (S72).

Then, the information processing device 1 ends the processing of S2. When it is determined that the intermediate language 22 stored in the information storage area 130 does not include the loop to be divided, the information processing device 1 also ends the processing of S2 (NO in S52).

On the other hand, when it is determined that the number of the distribution destination groups of each instruction included in the loop determined to be the division target in the processing of S51 has not reached the predetermined number or less (NO in S71), the group distribution unit 114 Performs the processing after S62 again. Hereinafter, a specific example of the split loop after performing the loop split will be described.

[Specific example of split loop]
FIG. 28 is a specific example for explaining the contents of the split loop. FIG. 28 (A) is a specific example for explaining one of the divided loops, and FIG. 28 (B) is a specific example for explaining the other of the divided loops.

The split loop shown in FIG. 28A includes instruction 1, instruction 2, instruction 3, instruction 4, instruction 6, instruction 8 and instruction 9 among the intermediate languages described in FIG.

Further, the split loop shown in FIG. 28 (A) is an instruction indicating that the value set in the variable reg3 is stored in the reg_ith position of the array tp_array1 after the instruction 1 or the like, "store tp_array1 (reg_i), reg3". , Which is an instruction indicating that the value set in the variable reg6 is stored in the reg_ith position of the array tp_array2, and the value set in the variable reg9 is set in the reg_i of the array tp_array3. It includes "store tp_array3 (reg_i), reg9" which is an instruction indicating to store in the third position.

On the other hand, the split loop shown in FIG. 28B includes instruction 5, instruction 7, instruction 10, instruction 11, instruction 12, and instruction 13 among the intermediate languages described in FIG.

Further, the split loop shown in FIG. 28 (B) is an instruction indicating that the value stored in the reg_i th of the array tp_array1 is set in the variable reg3 before the instruction 5 and the like, "load reg3, tp_array1 (reg_i). ) ”,“ Load reg6, tp_array2 (reg_i) ”, which is an instruction to set the value stored in the reg_ith position of the array tp_array2 in the variable reg6, and the value stored in the rig_ith position of the array tp_ary3 It includes "load reg9, tp_array3 (reg_i)" which is an instruction indicating that the variable reg9 is set.

That is, the number of temporary sequences in the split loop shown in FIG. 28 is larger than that in the split loop described in FIG. However, in the division loop shown in FIG. 28, each instruction (instruction 4 and instruction 6) that refers to each data stored in the same array is included in the same division loop.

As a result, the information processing device 1 can further shorten the execution time of the object code 23.

The above embodiments can be summarized as follows.

(Appendix 1)
An information generation unit that generates a first matrix by generating dependency information indicating the dependency between each instruction included in the loop of the intermediate language generated from the source code and converting the generated dependency information into a matrix.
A group distribution unit that distributes each instruction included in the loop into a plurality of groups based on the degree of dependence between each instruction calculated from the generated first matrix.
Each of the plurality of sorted groups has a loop splitting unit for splitting the loop.
An information processing device characterized by this.

(Appendix 2)
In Appendix 1,
The information generation unit
For each combination of instructions included in the loop, it is determined whether or not the dependency information indicates that there is a dependency between the instructions included in each combination.
Among the combinations of the instructions, the element corresponding to the combination in which the dependency information indicates that there is a dependency between the instructions included in each combination is set as the first value, and each combination of the combinations of the instructions has The first matrix is generated by setting the element corresponding to the combination in which the dependency information does not indicate that there is a dependency between the included instructions as the second value.
An information processing device characterized by this.

(Appendix 3)
In Appendix 1,
The group distribution unit distributes each instruction included in the loop to a plurality of groups so that the dependency relationship existing between the instructions included in the different groups is reduced.
An information processing device characterized by this.

(Appendix 4)
In Appendix 1,
The group distribution section
From the first matrix, a second matrix showing the degree of dependence between each instruction is generated.
Based on the generated second matrix, each instruction included in the loop is distributed to the plurality of groups.
An information processing device characterized by this.

(Appendix 5)
In Appendix 4,
The group distribution section
For each combination of instructions included in the loop, the degree of dependence between the instructions included in each combination is calculated.
The second matrix is generated by using each of the calculated degrees of dependence as an element.
An information processing device characterized by this.

(Appendix 6)
In Appendix 5,
The group distribution unit calculates the degree of dependence by using the Newman algorithm.
An information processing device characterized by this.

(Appendix 7)
In Appendix 5,
The group distribution section
Among the values of the elements of the second matrix, the distribution destinations of the plurality of instructions corresponding to the maximum values are determined in the same group.
The plurality of rows corresponding to the plurality of instructions in the second matrix are converted into a single row having the sum of the elements of the same column in the plurality of rows as elements, and the plurality of rows in the second matrix. The first matrix is regenerated by converting the plurality of columns corresponding to the instruction of to a single column having the sum of the elements of the same row in the plurality of columns as elements.
The second matrix is regenerated from the regenerated first matrix, and the second matrix is regenerated.
Until the number of groups determined as the distribution destination of each instruction included in the loop becomes a predetermined number or less, the process of determining, the process of regenerating the first matrix, and the process of regenerating the second matrix are performed. repeat,
An information processing device characterized by this.

(Appendix 8)
In Appendix 2, further
It has a proximity determination unit that determines whether or not a plurality of instructions for accessing each data stored in adjacent storage areas are included in the loop.
The information generation unit generates the first matrix based on the determination result of whether or not the plurality of instructions are included in the loop.
An information processing device characterized by this.

(Appendix 9)
In Appendix 8,
When the information generation unit determines that the plurality of instructions are included in the loop, the information generation unit sets the element corresponding to the plurality of instructions to a third value larger than the first value. The first matrix is generated.
An information processing device characterized by this.

(Appendix 10)
In Appendix 8,
Each data stored in the adjacent storage area is the same array.
An information processing device characterized by this.

(Appendix 11)
Generates dependency information that shows the dependency between each instruction included in the intermediate language loop generated from the source code.
The first matrix is generated by converting the generated dependency information into a matrix.
Based on the degree of dependence between each instruction calculated from the generated first matrix, each instruction included in the loop is divided into a plurality of groups.
The loop is divided for each of the plurality of sorted groups.
A compiler program characterized by having a computer perform processing.

(Appendix 12)
In Appendix 11,
In the process of generating the first matrix,
For each combination of instructions included in the loop, it is determined whether or not the dependency information indicates that there is a dependency between the instructions included in each combination.
Among the combinations of the instructions, the element corresponding to the combination in which the dependency information indicates that there is a dependency between the instructions included in each combination is set as the first value, and each combination of the combinations of the instructions has The first matrix is generated by setting the element corresponding to the combination in which the dependency information does not indicate that there is a dependency between the included instructions as the second value.
A compiler program characterized by that.

(Appendix 13)
In Appendix 11,
In the process of distributing to a plurality of groups, each instruction included in the loop is distributed to a plurality of groups so that the dependency existing between the instructions included in the different groups is reduced.
A compiler program characterized by that.

(Appendix 14)
In Appendix 11,
In the process of distributing to a plurality of groups,
From the first matrix, a second matrix showing the degree of dependence between each instruction is generated.
Based on the generated second matrix, each instruction included in the loop is distributed to the plurality of groups.
A compiler program characterized by that.

1: Information processing device 5: Operation terminal 21: Source code 22: Intermediate language 23: Object code 130: Storage unit NW: Network

Claims (11)

An information generation unit that generates a first matrix by generating dependency information indicating the dependency between each instruction included in the loop of the intermediate language generated from the source code and converting the generated dependency information into a matrix.
A group distribution unit that distributes each instruction included in the loop into a plurality of groups based on the degree of dependence between each instruction calculated from the generated first matrix.
Each of the plurality of sorted groups has a loop splitting unit for splitting the loop.
An information processing device characterized by this. In claim 1,
The information generation unit
For each combination of instructions included in the loop, it is determined whether or not the dependency information indicates that there is a dependency between the instructions included in each combination.
Among the combinations of the instructions, the element corresponding to the combination in which the dependency information indicates that there is a dependency between the instructions included in each combination is set as the first value, and each combination of the combinations of the instructions has The first matrix is generated by setting the element corresponding to the combination in which the dependency information does not indicate that there is a dependency between the included instructions as the second value.
An information processing device characterized by this. In claim 1,
The group distribution unit distributes each instruction included in the loop to a plurality of groups so that the dependency relationship existing between the instructions included in the different groups is reduced.
An information processing device characterized by this. In claim 1,
The group distribution section
From the first matrix, a second matrix showing the degree of dependence between each instruction is generated.
Based on the generated second matrix, each instruction included in the loop is distributed to the plurality of groups.
An information processing device characterized by this. In claim 4,
The group distribution section
For each combination of instructions included in the loop, the degree of dependence between the instructions included in each combination is calculated.
The second matrix is generated by using each of the calculated degrees of dependence as an element.
An information processing device characterized by this. In claim 5,
The group distribution unit calculates the degree of dependence by using the Newman algorithm.
An information processing device characterized by this. In claim 5,
The group distribution section
Among the values of the elements of the second matrix, the distribution destinations of the plurality of instructions corresponding to the maximum values are determined in the same group.
The plurality of rows corresponding to the plurality of instructions in the second matrix are converted into a single row having the sum of the elements of the same column in the plurality of rows as elements, and the plurality of rows in the second matrix. The first matrix is regenerated by converting the plurality of columns corresponding to the instruction of to a single column having the sum of the elements of the same row in the plurality of columns as elements.
The second matrix is regenerated from the regenerated first matrix, and the second matrix is regenerated.
Until the number of groups determined as the distribution destination of each instruction included in the loop becomes a predetermined number or less, the process of determining, the process of regenerating the first matrix, and the process of regenerating the second matrix are performed. repeat,
An information processing device characterized by this. In claim 2, further
It has a proximity determination unit that determines whether or not a plurality of instructions for accessing each data stored in adjacent storage areas are included in the loop.
The information generation unit generates the first matrix based on the determination result of whether or not the plurality of instructions are included in the loop.
An information processing device characterized by this. In claim 8.
When the information generation unit determines that the plurality of instructions are included in the loop, the information generation unit sets the element corresponding to the plurality of instructions to a third value larger than the first value. The first matrix is generated.
An information processing device characterized by this. In claim 8.
Each data stored in the adjacent storage area is the same array.
An information processing device characterized by this. Generates dependency information that shows the dependency between each instruction included in the intermediate language loop generated from the source code.
The first matrix is generated by converting the generated dependency information into a matrix.
Based on the degree of dependence between each instruction calculated from the generated first matrix, each instruction included in the loop is divided into a plurality of groups.
The loop is divided for each of the plurality of sorted groups.
A compiler program characterized by having a computer perform processing.

Images