JP5477206B2 - Compiling device and compiling program - Google Patents

Description

  The present application relates to a compiling device and a compiling program.

  There is a technique called “profiler” as a technique for quickly identifying a part that is a bottleneck of the execution performance of a program and finding a method for improvement. This profiler is a technique for acquiring PA (Profile Analysis) information (performance analysis information, hardware information) representing behavior during program execution, such as the number of function calls, the number of cycles, and the number of cache misses, during program execution. . This profiler is roughly divided into the following two types.

  -At the time of execution, the general PA information is acquired by interruption at regular intervals.

  A function call for acquiring PA information is manually added to the source program, and detailed PA information is acquired for a part of the source program.

  The latter method is selected to acquire detailed PA information, but a loop or processing block that is not expressed in the source program may be generated by optimization processing at the time of compilation. In this case, even if a PA information acquisition function is added to the source program, there is a possibility that detailed PA information in a new loop or processing block generated at the time of compilation cannot be acquired.

JP 9-62544 A Japanese Patent Laid-Open No. 2000-194466 JP 2004-110824 A

  In one aspect, the present invention is to generate an object program that can acquire PA information for a loop or a processing block that is newly generated during compilation.

  The compiling device of the present disclosure includes an acquisition unit, a first conversion unit, a first setting unit, an optimization unit, a second setting unit, a function embedding unit, and a second conversion unit.

  Here, the acquisition unit acquires the PA position file and the source program. The PA position file is a file in which a PA range representing a range on the program is described, which acquires PA information representing a behavior at the time of program execution. The source program is a program created using a program language.

  The first conversion means converts the source program acquired by the acquisition means into a first program described in an intermediate language.

  The first setting means reads the PA range from the PA position file acquired by the acquisition means, and sets the PA range on the first program.

  The optimization unit performs an optimization process on the first program in which the PA range is set by the first setting unit, and generates a second program that takes over the PA range set on the first program. .

  The second setting means checks whether or not the number of loops or processing blocks in the PA range on the second program has increased as compared with the number of loops or processing blocks in the PA range on the first program. Then, when the number of loops or processing blocks in the PA range on the second program has increased, the second setting means sets each loop or processing block in the PA range on the second program. Set a new PA range.

  The function embedding means embeds a PA information acquisition function for acquiring PA information in the second program for each PA range set in the second program.

  Further, the second conversion means converts the second program in which the PA information acquisition function is embedded into an object program executable on the computer.

  Further, the compile program disclosed herein is a program that is executed in an arithmetic processing device that compiles a program and constructs the compiling device of the present subject in the arithmetic processing device.

  According to the compiling apparatus and the compiling program of the present disclosure, it is possible to acquire detailed PA information for a new loop or processing block generated at the time of compiling. Therefore, detailed analysis of program execution performance becomes possible.

It is the figure which showed the example by which SIMD-ization is performed in an optimization process and a some loop is produced | generated as a result. It is the figure which showed the example by which the loop distribution was performed by the optimization process. 1 is an external view of a computer that operates as an embodiment of a compiling device. FIG. 4 is a hardware configuration diagram of a computer whose appearance is shown in FIG. 3. It is function explanatory drawing of the compiling apparatus as embodiment. It is a figure which shows the description content of PA position file. It is a figure which shows optimization information. It is the figure which showed the image of the process which acquires a source program and converts into the 1st program by an intermediate language. It is a flowchart which shows the process in the PA position file analysis part 13 of FIG. It is a figure which shows the image of the process in PA position file analysis part. It is an image figure which shows the 2nd program produced | generated by the optimization process. It is a figure which shows the first half part of the flowchart which shows the process of PA acquisition function insertion part 15 of FIG. It is a figure which shows the latter half part of the flowchart which shows the process of PA acquisition function insertion part 15 of FIG. It is an image figure of the 2nd program of the stage by which the increase in the number of loops was recognized. It is an image figure which shows the state by which the new PA range was set on the 2nd program. It is an image figure which shows the state by which the PA information acquisition function was embedded in the 2nd program. It is an image figure when the same content as the PA acquisition function on the intermediate language image of FIG. 6 is written down on the source program. It is the figure which showed the source code image when "PA range setting location designation information" other than "all" is designated. It is a figure which shows the example of a display screen at the time of compilation execution. It is the figure which showed the display image of PA information obtained by execution of the object program embedded with the PA information acquisition function stored in the execution file.

  Hereinafter, a case where the number of loops and processing blocks increases due to optimization processing at the time of compilation will be exemplified, and then the embodiment will be described.

  FIG. 1 is a diagram illustrating an example in which SIMD processing is performed in the optimization process, and as a result, a plurality of loops are generated. The optimization process is performed on an intermediate language suitable for handling in the compiling apparatus. In FIG. 1, for the sake of simplicity, regardless of whether the program is a source program or a program using the intermediate language. Shown in source program format.

  “Do i = 1,100” on line (1) in the source program of FIG. 1A means that the following processing is performed.

  First, i = 1 is set, and the process up to “end do” on the (3) line is executed.

  After that, i is incremented by 1 and the process until “end do” is repeated until i = 100.

  In addition, “a (i) = 3” in the (2) th line means that 3 is assigned to a (i) in the array a.

  Furthermore, “end do” in the (3) line means that the loop is at the end position of “do i = 1,100” in the (1) line.

  Here, a simple loop shown in FIG. 1A will be described as an example.

  FIG. 1B shows a program created by optimization processing in compilation based on the program of FIG.

  In FIG. 1B, the “if” statement on the (1) line is the line before the “else” on the (5) line (up to the (4) line) when the array a is on the 8-byte boundary. ) Is executed. Otherwise, the process after “else” on the (5) line is executed.

  Here, it is assumed that the memory access is in units of 8 bytes and the access boundary is determined by hardware. However, this condition itself has no significant meaning here, and detailed description thereof is omitted.

  In (2) line “do i = 1,99,2”, i = 1 is set first and the process up to (4) line “end do” is executed, and then i is incremented by two. On the other hand, (4) the process up to the line “end do” is repeated until i = 99.

  Also, “a (i: i + 1) = 3” in the (3) th line means that 3 is substituted for both a (i) and a (i + 1).

  It is SIMD (Single Instruction) that a plurality of data is processed by one instruction (here, a (i) = 3 processing and a (i + 1) = 3 processing) as in the processing of the (3) line. This is called “Multiple Data”. According to this SIMD, high-speed processing can be expected.

  (4) “end do” in the line means that the end position of the loop starts with “do i = 1,99,2” in line (2).

  (5) If the line does not satisfy the example in the “if” statement on line (1) (here, the condition that “array a is on an 8-byte boundary”), this “else” statement This means that the process is performed.

  (6) For “do i = 1,100,1” in the row, first, i = 1 is set, and the processing until “end do” in the (8) row is executed, and then i is incremented by 1 However, (8) means that the process up to the line is repeated until i = 100.

  (7) “a (i) = 3” in the line means that 3 is assigned to a (i) in the array a.

  (8) “end do” in the line means the end position of the loop starting from “do i = 1,100,1” in the (6) line.

  (9) “end if” on the line means that the end position of the process starts with the “if” statement on the (1) line.

  That is, in FIG. 1B, when the array a is on the 8-byte boundary, the processing speed is increased by SIMD in the (3) th row, otherwise simple sequential processing is performed. Represents a program.

  In the case of the program shown in FIG. 1B, the block surrounded by if and endif shown in FIG. 1B corresponds to the loop shown in FIG. In the case of FIG. 1A, a simple loop is formed, but in the case of FIG. 1B, loops are generated under the if and else subordinates of the if-else statement. That is, in FIG. 1B, there are two loops (a) and (b) in the block (c).

  When a function for acquiring PA information related to the loop of FIG. 1A is embedded in the source program of FIG. 1A, simply, PA information for the entire block of FIG. Thus, PA information cannot be obtained for the loops (a) and (b) alone. Even if the PA information of the block (c) is acquired, it may be unclear whether the loop (a) or the loop (b) was executed, and the analysis of the program execution performance may be insufficient. .

  In the case of an embodiment to be described later, it is possible to acquire separate PA information for each of the loops (a) and (b), and a detailed analysis of the program execution performance can be performed.

  Here, an example has been described in which loops are generated under the if and else subordinates of an if-else statement by optimization processing, but depending on the description of the program, a processing block may be generated. The actual program is usually a more complicated description, but the processing block will be explained as follows, for example, in the example of FIG. That is, instead of the loop of (b) shown in FIG. 2B, three processes “a (1) = 3”, “a (2) = 3”, and “a (3) = 3” are sequentially performed. There is a possibility that a program written to be executed at the same time is generated. In such a case, a series of three processes “a (1) = 3”, “a (2) = 3”, and “a (3) = 3” is regarded as one processing block. For such a processing block, it is possible to acquire independent PA information for the processing block as in the case of the loop.

  FIG. 2 is a diagram illustrating an example in which loop distribution is performed by optimization processing.

  FIG. 2A shows a source program, and “do i = 1, n” in the (1) line means that the following processing is performed.

  First, i = 1 is set, and the processing up to (end do) in the (4) th line is executed.

  Thereafter, i is incremented by 1 until i = n is reached, and the process until “end do” in the (4) th line is repeated.

  Further, a (i) = i in the (2) th row in FIG. 2A means that i is substituted for a (i) in the array a.

  B (i) = b (i−1) + a (i) in the third row is obtained by adding the value of b (i−1) and the value of a (i) in the array b to b This means that it is assigned to (i).

  Furthermore, “end do” in the (4) th line means the end position of the loop starting from “do i = 1, n” in the (1) th line.

  FIG. 2B shows a program created based on the source program of FIG. 2A by the optimization process.

  “Do i = 1, n” in the (1) line in FIG. 2B is the same as in the first line in FIG.

  “A (i) = i” in the (2) line in FIG. 2B is the same as the (2) line in FIG.

  Further, “do i = 1, n” in the (4) line in FIG. 2B is the same as “do i = 1, n” in the (1) line. However, the end position of the loop starting from the (4) th line is “end do” on the (6) th line.

  Further, “b (i) = b (i−1) + a (i)” in the (5) line in FIG. 2B is the same as the (3) line in FIG. In other words, in the case of FIG. 2, the loop composed of the entire area from line (1) to line (4) in FIG. 2A corresponds to the entire area indicated by (c) in FIG. To do. In the case of FIG. 2 (A), it is composed of a simple loop, but in the case of FIG. 2 (B), loop distribution is performed, and (a) and (b) are included in the range shown in (c). There are two loops.

  Here, a simple source program is used as an example. For example, in the case of a loop that performs complicated arithmetic processing in which a register for arithmetic is insufficient, the loop is separated into a plurality as shown in FIG. In some cases, efficient calculation processing can be performed. If it is determined that the loop distribution by the optimization process is more efficient, the loop distribution as shown in FIG. 2B may be performed.

  The case of FIG. 2 is the same as the case of FIG. 1, and when a function for obtaining PA information is embedded in the source program of FIG. 2A, it is simply shown in FIG. 2B (c). The PA information for the entire range is acquired. In this case, it is not possible to obtain PA information for the loops (a) and (b) alone. Therefore, when the execution performance in the range of (c) is poor, there is a possibility that detailed analysis will be insufficient to determine whether there is a problem in the loop of (a) or a problem in the loop of (b).

  Therefore, in the following, an embodiment in which independent PA information can be acquired also for the loops (a) and (b) in FIG. 1 (B) and FIG. 2 (B) will be described.

  FIG. 3 is an external view of a computer that operates as an embodiment of the compiling device, and FIG. 4 is a hardware configuration diagram of the computer whose external appearance is shown in FIG.

  A computer 100 illustrated in FIG. 3 includes a main body 110 that incorporates a CPU (Control Processing Unit), a memory, a hard disk device, and the like which will be described later. The computer 100 also has an image display unit 120 that displays an image on the display screen 121 in accordance with an instruction from the main body unit 110. The computer 100 further includes a keyboard 130 for inputting instructions, character information, and the like in the computer 100 in response to key operations by the user. Further, the computer 100 includes a mouse 140 that performs a user's operation to move a cursor displayed on the display screen 121 and to input an instruction by designating an icon or the like displayed on the display screen 121.

  The main body 110 has a CD / DVD loading slot 111 into which a CD (Compact Disk) and a DVD (Digital Versatile Disk) (hereinafter referred to as CD / DVD without distinguishing between CD and DVD). A CD / DVD drive for driving a CD / DVD loaded from the CD / DVD loading slot 111 is built in the main body 110.

  The main body 110 further includes a CPU (Control Processing Unit) 112, a memory 113, a hard disk device 114, and a network interface 115, as shown in FIG. FIG. 5 also shows the above-described CD / DVD drive 116 that drives the CD / DVD 117.

  A bus 150 connects these elements to each other and the image display unit 120, the keyboard 130, and the mouse 140 whose appearance is shown in FIG. 4.

  Here, the hard disk device 114 stores various programs and data. The program stored in the hard disk device 114 is read out and expanded in the memory 113, and the program expanded on the memory 113 is executed by the CPU 112. The network interface 115 is an element that communicates with the outside via a LAN (Local Area Network), the Internet, or the like.

  Here, the hard disk device 114 stores a compile program for creating an object program based on a source program (see, for example, FIGS. 1A and 2A).

  The source program is stored in the CD / DVD 117 and read by the CD / DVD drive 116. Alternatively, it may be received via the network interface 115. When a compile program stored in the hard disk device 114 is expanded in the memory 113 and executed by the CPU 112, the computer 100 operates as an example of a compile device. When the computer 100 operates as a compiling device, an object program is created based on the source program acquired as described above. The created object program is once stored in the hard disk device 114 and then written to the CD / DVD by the CD / DVD drive 116 or transmitted via the network interface 115. In this way, the object program is delivered to the execution destination computer (not shown) and executed on the computer.

  FIG. 5 is a functional explanatory diagram of the compiling device as the embodiment.

  The compile device 10 shown in FIG. 5 represents functions realized in the computer by executing the compile program in the computer shown in FIGS.

  The compiling device 10 in FIG. 5 is a device that takes in a source program 1 created in a program language and executes compiling to convert the source program 1 into an object program 2 that can be executed on a computer that is scheduled to be executed. Here, the computer on which the object program 2 is executed does not prevent the computer used for realizing the functions of the compiling apparatus 10 from being the same computer, but may be another computer.

  Here, in the case of the present embodiment, the compiling device 10 takes in the PA location file 3 in addition to the source program 1 and embeds a function for acquiring PA information based on the description of the PA location file 3. 2 is generated. In the present embodiment, a function for obtaining “number of cycles”, “number of instructions”, “cache miss”, etc. within the section specified by the PA range is embedded as PA information.

  The compiling device 10 includes a source program / PA position file acquisition unit 11, a syntax analysis unit 12, a PA position file analysis unit 13, an optimization unit 14, a PA acquisition function insertion unit 15, a code generation unit 16, and the like. Have Of these, the PA acquisition function insertion unit 15 includes a PA range setting unit 151 and a function embedding unit 152.

  The source program / PA location file acquisition unit 11 is an example of an acquisition unit, and takes the source program 1 and the PA location file 2 into the compiling device 10. The source program 1 is a program created in a program language such as a Fortran language, a C language, or a C ++ language. Further, the PA position file 3 is a file in which a PA range representing a range on the program is described for obtaining PA information representing a behavior at the time of executing the program. The PA position file 3 also includes PA range setting location designation information for designating the new PA range setting location when a new PA range is set by the PA range setting unit 151 described later, along with the PA range. It has been described.

FIG. 6 is a diagram showing the description contents of the PA location file in the present embodiment. On the first line, “PA range setting location designation information” is described after the “*” mark. This PA range setting location designation information includes
"All"
"Entity-loop-only"
“Main-loop-only”
"Mod-loop-only"
There are four types.

  These will be described with reference to FIGS. 1B and 2B.

  “All” means that PA information is acquired for all the loops of (a), (b), and (c) of FIG. 1 (B) and FIG. 2 (B).

  “Entity-only” means that PA information is acquired only for the loop of (c), and it is not necessary to acquire independent PA information for the individual loops of (a) and (b). Yes.

  “Main-loop-only” acquires PA information only for the first loop of child loops, that is, the loop of (a), and acquisition of PA information for the loops of (b) and (c) It means that it is unnecessary.

  Furthermore, “mod-loop-only” obtains PA information for each loop (the loop of (b) in the case of FIG. 1B and FIG. 2B) except the first loop among the child loops. This means that it is not necessary to acquire PA information for the loops (a) and (c).

  The second line of the description contents in the PA position file means the following.

  “Source file name” is the name of the source file in which the source program for which PA information is to be acquired is described, and “module name” is the name of the module for acquiring PA information in the source file. Further, the “starting line number and ending line number” are a starting line number and an ending line number representing the start position and the end position, respectively, of the PA acquisition range in the module. A range between the start position and the end position is referred to as “PA range” here.

  In FIG. 6, the PA position file is described in two lines. However, the same description as the second line can be made in the third and subsequent lines, and a plurality of PA ranges can be specified in one PA position file. is there.

  Returning to FIG.

  The syntax analysis unit 12 of the compiling device 10 illustrated in FIG. 5 is an example of a first conversion unit, and the source program acquired by the source program / PA position file acquisition unit 11 is an intermediate suitable for handling by the compiling device 10. Convert to the first program written in the language.

  The PA position file analysis unit 13 is an example of a first setting unit. That is, the PA position file analysis unit 13 reads a PA range defined by the start position of the PA acquisition range and the end position from the PA position file, and sets the PA range (start position and end position) as described above. 1 is set on the program.

  The optimization unit 14 is an example of an optimization unit. The optimization unit 14 performs the optimization process on the first program in which the PA range is set, and takes over the PA range set on the first program. Generate the program. The optimization processing here refers to processing for changing the first program so that the program is executed more efficiently, such as SIMD conversion and loop distribution shown in FIGS.

  The contents of the optimization process by the optimization unit 14 are left as optimization information 141.

  FIG. 7 is a diagram showing optimization information.

  Here, the position where the program is changed by the optimization process is specified by the “source file name”, “module name”, and “target line number”. The “optimization type” identifies the type of program change performed by the optimization process. The “optimization type” includes, for example, “SIMD” (see FIG. 1), “loop distribution” (see FIG. 2), and the like.

  The PA range setting unit 151 that constitutes the PA acquisition function insertion unit 15 shown in FIG. 5 is an example of a second setting unit, and is within the PA range on the second program when compared with the first program. Check for an increase in the number of loops or processing blocks. If it has increased, a new PA range is set for each loop or processing block in the PA range on the second program. In setting the new PA range, a new PA range is set based on “PA range setting location designation information” described in the PA position file 3.

  The function embedding unit 152 constituting the PA acquisition function inserting unit 15 is an example of a function embedding unit, and a PA information acquisition function for acquiring PA information is added to the second program. Embed each PA range set to.

  Furthermore, the code generation unit 16 is an example of a second conversion unit, and converts the second program in which the PA information acquisition function is embedded into an object program 2 that can be executed on a computer. The source program 1 may be divided into a plurality of programs such as a main routine and a subroutine. At this time, the code generation unit 16 creates an object program for the plurality of programs. Thereafter, the linker 17 integrates a plurality of object programs, and finally creates an executable object program.

  Hereinafter, the operation of each part of the compiling device 11 shown in FIG. 5 will be further described with examples of the source program and the PA location file.

  FIG. 8 is a diagram showing an image of processing for acquiring a source program and converting it into a first program in an intermediate language.

  FIG. 8A shows an example of a source program, FIG. 8B shows an example of a PA position file, and FIG. 8C shows an example of a first program in an intermediate language.

  “Source program (a.f90)” described in FIG. 8A means that the name of the source file in which the source program is described is “a.f90”.

  In addition, “subroutine sub1 (a, b, n)” on the (1) line of this source program is a subroutine that is named “sub1” on and after the (1) line and has a, b, n as arguments. It represents that.

  Also, “integer :: n” on the (2) line represents that n is an integer.

  In addition, “real, dimension (1000) :: a, b” on the (3) line indicates that each of a and b is a real number and each is an array having 1000 elements. Yes.

  Further, the (4) to (7) lines are the same as the (1) to (4) lines in FIG. 2A, respectively, and a duplicate description is omitted here.

  Furthermore, “end subroutine sub1” on line (8) represents the end position of the subroutine named “sub1”.

  In addition, “PA location file (page.txt)” in FIG. 8B means that the file name of this PA location file is “page.txt”.

  Also, “* all” on the first line in the PA position file in FIG. 8B indicates that “PA range setting location designation information” described with reference to FIG. 6 is “all”. .

  Further, “a.f90”, “sub1”, “4”, and “7” in the second line in the PA position file of FIG. 8B have source file names “a.f90” (FIG. 8A )), The module name is “sub1”, and the PA range is from the fourth line to the seventh line.

  Here, it is assumed that the source program / PA location file acquisition unit 11 of the compiling apparatus 10 illustrated in FIG. 5 has acquired the source program of FIG. 8A and the PA location file of FIG. 8B.

  At this time, the syntax analysis unit 12 in FIG. 5 creates a first program described in an intermediate language whose image is shown in FIG. 8C based on the source program shown in FIG. .

  The “basic processing block (Line1)”, “basic processing block (Line2)”, “basic processing block (Line3), and“ basic processing block (Line8) ”in FIG. 8C are the source programs of FIG. Are the parts in which the instructions on the (1) line, (2) line, (3) line, and (8) line are converted into an intermediate language, respectively, and the “loop” in FIG. “(Line4-7)” is a portion in which a loop from the (4) line to the (7) line in FIG. 8A is described in an intermediate language. The “loop (line 4-7)” includes a line 4 block, a line 5 and 6 block, a counter block for repeating Lines 5 and 6 a specified number of times, and a line 7 block. The block of Line 4 corresponds to the (4) line of FIG. 8A, and the blocks of Line 5 and 6 correspond to the (5) and (6) lines of FIG. The counter block is a block for repeating the processing of Lines 5 and 6 while counting i from 1 to n. The block of Line 7 is a block corresponding to the (7) line in FIG. The arrows from the counter block to the Line 5 and 6 blocks indicate repetition of the processing. Further, the arrow from the Line 4 block to the Line 7 block satisfies the loop processing condition between the (4) line and the (7) line in FIG. 8A, for example, n = 0 or n <0. In the case of not doing so, it represents skipping from the (4) line to the (7) line. The “loop (line 4-7)” in FIG. 8C is one of “processing blocks” composed of a plurality of basic processing blocks that execute one set of processing.

  FIG. 9 is a flowchart showing processing in the PA position file analysis unit 13 of FIG.

  FIG. 10 is a diagram showing an image of processing in the PA position file analysis unit 13.

  In the PA position file analysis unit 13, as shown in FIG. 9, first, from the PA position file (see FIG. 8B), “source file name” “module name” “PA range (start position) And end position) "are read (step S11). Then, the first program described in the intermediate language (see FIG. 8C) is traced, and the line number designated as the PA range start position and end position is searched (step S12). An attribute value representing the start position and end position is set in association with the line number found by this search (step S13). Here, the number of processing blocks and loops existing between the start position and the end position is also counted and set.

  This process will be described in the continuation of the example shown in FIG. 8. As shown in FIG. 10, the attribute value “PA (1) start / all / blockNum: 1” is set in the Line4 block, and the Line7 block is set. An attribute value “PA (1) end / all” is set.

  “PA (1) start” among the attribute values set in the block of Line 4 represents the start position of the PA range of the serial number (1). “All” indicates that “PA range setting location designation information” described with reference to FIG. 6 is “all”. Further, “blockNum: 1” represents that the number of processing blocks or loops in the PA range is 1.

  “PA (1) end” among the attribute values set in the block of Line 7 represents the end position of the PA range of the serial number (1). In “all”, “all” in the attribute values of the block of Line 4 is repeated.

  In step S14 of FIG. 9, it is determined whether or not the settings for all the PA ranges have been completed from the PA position file. If there is an unset PA range, the process returns to step S11 to determine the unset PA range. The process is repeated.

  Here, the PA position file shown in FIG. 8B is composed of two lines, and the PA range is only one described in the second line. However, as described above, by describing the same description as the second line in the third and subsequent lines, a plurality of PA ranges can be set with one PA position file. Step S14 in FIG. 9 is a step that takes into consideration that a plurality of PA ranges may be set.

  When setting of attribute values for all PA ranges is completed (step S14 in FIG. 9), the first program (see FIG. 10) in which those attribute values are set is passed to the optimization unit 14 shown in FIG. Step S15).

  In the optimization unit 14 in FIG. 5, optimization processing is performed on the first program received from the PA position file analysis unit 13. Here, it is assumed that loop distribution (see FIG. 2) has been performed as a result of the optimization processing.

  At this time, the optimization unit 14 generates optimization information 141 having the source file name “a.f90”, the module name “sub1”, the target line number “4, 7”, and the optimization type “loop distribution” ( FIG. 5 and FIG. 7).

  FIG. 11 is an image diagram showing a second program generated by the optimization process.

  There are two “loops (lines 4-7)”. The upper loop is a loop in charge of the processing of the (5) line in FIG. 8A, and the lower loop is a loop in charge of the processing of the (6) line in FIG. 8A. In addition, “L-dst” is described in both the upper loop and the lower loop. This “L-dst” indicates optimization information 141 (see FIG. 5) regarding the loop. In addition, the second program takes over the attribute value related to the PA range of the first program (see FIG. 10) before optimization. The attribute value related to the start position of the PA range is associated with the Line 4 block in the upper loop of the two loops. The attribute value relating to the end position of the PA range is associated with the Line 7 block of the lower loop of the two loops.

  At this stage, one PA range (herein referred to as “PA range <1>”) is set from the beginning of the upper loop to the end of the lower loop.

  FIGS. 12 and 13 are diagrams showing the first half and the second half of the flowchart showing the processing of the PA acquisition function insertion unit 15 shown in FIG. 5, respectively.

  The PA acquisition function insertion unit 15 performs the processes shown in FIGS. 12 and 13 for the second program subjected to the optimization process.

  Here, first, the optimization information 141 (see FIG. 5) is examined to determine whether or not the optimization processing related to the loop has been performed (step S21). When the optimization process related to the loop is performed, it is determined whether or not the loop subjected to the optimization process is a loop included in the PA range (step S22). When the optimization process related to the loop is not performed (step S21), or when the optimization process related to the loop is performed, the optimization process is not performed regarding the loop in the PA range (step S22). Proceed to step S33 of FIG. Step S33 will be described later.

  If the optimization process is being performed for the loop in the PA range (step S22), the process proceeds to step S23, the process flow in the PA range is followed, and the number of processing blocks and loops is counted. When the number has not increased (step S24), the process proceeds to step S33.

  FIG. 14 is an image diagram of the second program at the stage where the increase in the number of loops is recognized.

  Here, the number of loops that originally did not exist has increased to two, and when PA range <1> is set as the parent PA range, two loops are placed under the parent PA range <1>. It is recognized that there are child PA ranges <1.1> and <1.2>.

  When it is recognized that the number of loops or processing blocks has increased (FIG. 12, step S24), the process proceeds to step S25, and a new PA range is set on the second program.

  FIG. 15 is an image diagram showing a state in which a new PA range is set on the second program.

  Here, “PA (1, 1.1) start / all” is set as an attribute value related to the PA range in the block of Line 4 in the upper loop of the two upper and lower loops. The upper loop Line 7 block has “PA (1.1) end / all”, and the lower loop Line 4 block has “PA (1.2) start / all”. An attribute value of “PA (1, 1.2) end / all” is set in the block of Line7.

  Among the attribute values set in the Line 4 block of the upper loop, “PA (1, 1.1) start” is both the parent PA range <1> and the child PA range <1.1>. Represents the start position of the PA range. Since “all” has already been described, description thereof is omitted here.

  “PA (1.1) end” among the attribute values set in the Line 7 block of the upper loop represents the end position of the child PA range <1.1>.

  Of the attribute values set in the Line 4 block of the lower loop, “PA (1.2) start” represents the start position of the child PA range <1.2>.

  Further, “PA (1, 1.2) end” among the attribute values set in the Line 7 block of the lower loop is the parent PA range <1> and the child PA range <1.2>. Represents the end position of both PA ranges.

  Here, in this way, the parent PA range <1> and the two child PA ranges <1.1> and <1.2> are set so that the parent-child relationship can be understood.

  In step S26 of FIG. 12, “PA range setting location designation information” is read from the PA position file. As described above, the “PA range setting location designation information” includes “all-only”, “main-loop-only”, and “mod-loop-only” in addition to “all” illustrated here. There are a total of four types.

  In step S27, it is determined whether or not the read “PA range setting location designation information” is “entry-only”, and if it is “entry-only”, the attribute value related to the child PA range is deleted. (Step S28).

  This is because “entity-only” does not require individual PA information for the child PA range, and is information for instructing to collectively acquire PA information for the parent PA range.

  If the “PA range setting location designation information” is “main-loop-only”, the attribute values related to the second and subsequent child PA ranges of the child PA ranges are deleted (step S30). Here, the reason for expressing the second or later is that there may be three or more loops or processing blocks in the parent PA range.

  Furthermore, when the “PA range setting location designation information” is “mod-loop-only” (step S31), the attribute value related to the PA range of the first child of the child PA ranges is deleted (step S31). S32).

  In the specific example shown here (see FIG. 15), since “all” is designated, all of steps S27, S29, and S31 are “no”, and the attribute values of the child PA range are all. It remains without being deleted.

  Of the processes shown in FIGS. 12 and 13, the process up to step S32 corresponds to the process of the PA range setting unit 151 in FIG. 5, and the process after step S33 is the process of the function embedding unit 152 in FIG. It corresponds to.

  In step S33, PA information acquisition function insertion processing is performed. Here, an attribute value related to the PA range is referred to, and a function for acquiring PA information is embedded in the second program for each PA range.

  FIG. 16 is an image diagram showing a state in which the PA information acquisition function is embedded in the second program.

  Of the two upper and lower loops, “Call PA_start (1.1)” is embedded in the Line 4 block of the upper loop, and “Call PA_end (1.1)” is embedded in the Line 7 block of the upper loop. It is.

  Also, “Call PA_start (1.2)” is embedded in the Line 4 block of the lower loop, and “Call PA_end (1.2)” is embedded in the Line 7 block of the lower loop. .

  “Call PA_start (1.1)” embedded in the block of Line 4 in the upper loop is an instruction for calling a function for starting the acquisition of PA information. In addition, “Call PA_end (1.1)” embedded in the Line 7 block of the upper loop is an instruction for calling a function for executing the stop of acquisition of PA information started by “Call PA_start (1.1)”. It is.

  Similarly, “Call PA_start (1.2)” embedded in the Line 4 block of the lower loop and “Call PA_end (1.2)” embedded in the Line 7 block are respectively , An instruction for calling each function for starting the acquisition of PA information and ending the acquisition of PA information. By attaching the numbers <1.1> and <1.2> to the two PA ranges, it is related so that the PA range obtained by adding the two PA ranges is the parent PA range <1>. Yes.

  In step S33 of FIG. 13, for example, when the PA information acquisition function is embedded as shown in FIG. 16, the second program in which the PA acquisition function is embedded is passed to the code generation unit 16 (see FIG. 5).

  The code generation unit 16 receives the second program in which the PA acquisition information is embedded, and generates an object program executable by the computer. In addition, when there are a plurality of programs such as a main routine and a subroutine, the code generation unit 16 and the linker 17 also combine the plurality of programs to finally generate an executable object program.

  FIG. 17 is an image diagram when the same content as the PA acquisition function on the intermediate language image of FIG. 16 is written on the source program.

  Here, the loop is distributed to two loops, that is, the loop of the (3) line to the (5) line and the loop of the (8) line to the (10) line. Here, an instruction for calling a function for executing the PA information acquisition start of the parent PA range (1) is described in the (1) line. Further, in the (2) line, an instruction for calling a function for executing the PA information acquisition start of the child PA range (1.1) is described. Further, the (6) line describes an instruction for calling a function for stopping the acquisition of the PA information of the child PA range (1.1). Further, the (7) line describes an instruction for calling a function for starting the acquisition of PA information of the child PA range (1.2). Further, an instruction for calling a function for stopping acquisition of PA information of the child PA range (1.2) is described in the (11) line, and a PA in the parent PA range (1) is described in the (12) line. An instruction for calling a function for stopping the acquisition of information is described.

  Here, in FIG. 16, the PA information acquisition function related to the parent PA range is not explicitly embedded. As described above, the parent-child relationship is known, and the parent PA information is automatically obtained as the sum of the PA information of the two children.

  On the other hand, in the case of FIG. 17, it is assumed that the programmer writes down, the programmer performs the loop distribution by himself, and also embeds the PA acquisition function by himself.

  Thus, conventionally, the parent-child relationship of the PA range is not automatically recognized, and the PA range <1>, PA range <1.1>, and PA range <1.2> are set as separate PA ranges, respectively. Be recognized. Therefore, it is necessary for the programmer to describe PA information acquisition functions for all PA ranges.

  FIG. 18 is a diagram showing a source code image when “PA range setting location designation information” other than “all” is designated.

  FIG. 18A shows a case where “entry-only” is designated, and a PA information acquisition function is embedded so as to include the entire plurality of loops.

  FIG. 18B shows a case where “main-loop-only” is designated, and the PA information acquisition function is embedded so as to include only the first loop.

  FIG. 18C shows a case where “mod-loop-only” is designated, and the PA information acquisition function is embedded so as to include only the second loop. Although only two loops are shown here, when there are three or more loops, in the “mod-loop-only”, the PA information is independently set for each of the second and subsequent loops. PA information acquisition function is embedded.

  FIG. 19 is a diagram showing an example of a display screen at the time of compiling.

  In the first line, “frt a.f90 -pa_range“ parrange. “txt” -o a.exe ”is input by an operation of the keyboard 130 (see FIG. 3) by the operator. Here, “frt” is a command for compiling a source program described in the Fortran language. “A.f90” is the name of the source file in which the source program is described (see FIG. 8A). Further, -pa_range “place.txt” indicates that the file name of the PA location file to be associated with the source file “a.f90” is “place.txt” (see FIG. 8B). Further, “-o a.exe” indicates that the name of the execution file in which the final object program is stored is “a.exe”.

  In other words, the first line embeds a PA information acquisition function in the source program stored in the file “a.f90” described in the Fortran language while referring to the PA location file “parrange.txt”. It represents an operator command instructing generation of an object program. The first line further instructs to store the generated object program in the file “a.exe”.

  The second and subsequent lines are information that is sequentially displayed according to the progress of execution of compilation by the compiling device. By displaying up to “execution file completion (a.exe)” on the last line, it is understood that the compilation is completed.

  FIG. 20 is a diagram showing a display image of PA information obtained by executing an object program embedded with a PA information acquisition function, stored in an execution file.

  The first line is PA information related to the parent PA range <1>. The PA information of the parent PA range <1> is the sum of the PA information of the child PA range <1.1> of the second row and the PA information of the child PA range <1.2> of the third row.

  “Interval” indicates the PA range. “4-7 / block1 (a.f90-sub1)” on the first line is a. This means that the PA information is related to the processing blocks from the fourth line to the seventh line of f90 / sub1. The same applies to the second and third lines. “Cycle”, “instructions”, and “cachemiss” represent the number of cycles, the number of instructions, and the number of cache misses when each processing block is executed.

  As described above, in this embodiment, even when the number of loops or processing blocks is increased by the optimization processing, detailed PA information about each loop or each processing block can be acquired. Therefore, detailed analysis of program execution performance becomes possible.

DESCRIPTION OF SYMBOLS 1 Source program 2 Object program 3 PA position file 10 Compile apparatus 11 Source program / PA position file acquisition part 12 Syntax analysis part 13 PA position file analysis part 14 Optimization part 15 PA acquisition function insertion part 16 Code generation part 17 Linker 100 Computer 110 Main body 111 CD / DVD loading slot 112 CPU
113 Memory 114 Hard Disk Device 115 Network Interface 116 CD / DVD Drive 117 CD / DVD
DESCRIPTION OF SYMBOLS 120 Image display part 130 Keyboard 121 Display screen 140 Mouse 141 Optimization information 150 Bus 151 PA range setting part 152 Function embedding part

Claims (4)

An acquisition means for acquiring a PA position file describing a PA range representing a range on the program for obtaining PA information representing a behavior at the time of program execution, and a source program created using a program language;
First conversion means for converting the source program acquired by the acquisition means into a first program described in an intermediate language;
First setting means for reading the PA range from the PA position file and setting the PA range on the first program;
Optimizing the first program in which the PA range is set by the first setting means to generate a second program that takes over the PA range set on the first program And
Check if there is an increase in the number of loops or processing blocks in the PA range on the second program as compared to the number of loops or processing blocks in the PA range on the first program; When the number of loops or processing blocks in the PA range on the second program has increased, a new PA range is set for each loop or processing block in the PA range on the second program A second setting means;
Function embedding means for embedding a PA information acquisition function for acquiring the PA information in the second program for each PA range set in the second program;
A compiling apparatus comprising: a second conversion unit configured to convert the second program in which the PA information acquisition function is embedded into an object program executable on a computer. The PA position file describes PA range setting location designation information for designating the new PA range setting location when the new PA range is set by the second setting means together with the PA range. ,
The compiling apparatus according to claim 1, wherein the second setting means sets the new PA range on the second program based on the PA range setting location designation information. It is executed in an arithmetic processing unit that executes a program or compiles, and in the arithmetic processing unit,
An acquisition means for acquiring a PA position file describing a PA range representing a range on the program for obtaining PA information representing a behavior at the time of program execution, and a source program created using a program language;
First conversion means for converting the source program acquired by the acquisition means into a first program described in an intermediate language;
First setting means for reading the PA range from the PA position file and setting the PA range on the first program;
Optimizing the first program in which the PA range is set by the first setting means to generate a second program that takes over the PA range set on the first program And
Check if there is an increase in the number of loops or processing blocks in the PA range on the second program as compared to the number of loops or processing blocks in the PA range on the first program; When the number of loops or processing blocks in the PA range on the second program has increased, a new PA range is set for each loop or processing block in the PA range on the second program A second setting means;
Function embedding means for embedding a PA information acquisition function for acquiring the PA information in the second program for each PA range set in the second program;
A compiling program comprising: a compiling device having second converting means for converting the second program in which the PA information acquisition function is embedded into an object program executable on a computer. The PA position file describes PA range setting location designation information for designating the new PA range location when the second setting means sets a new PA range together with the PA range,
4. The compiled program according to claim 3, wherein the second setting means sets the new PA range on the second program based on the PA range setting location designation information.

Images