JP5169322B2 - Variable optimization apparatus, variable optimization program, compiler, variable optimization method, and compilation method - Google Patents

Description

  The present invention relates to a compiler and a variable optimization apparatus, and more particularly, a compiler that generates intermediate code from a source program and generates an object program from the intermediate code, and a variable optimization apparatus used for such a compiler. About. The present invention also relates to a compilation method, a variable optimization method, and a program.

  In recent years, the operating speed of CPUs (processors) has been increasing due to improvements in hardware and the like. On the other hand, the performance of memory access has not been improved compared to the increase in CPU speed. For this reason, a program with little memory access can greatly improve the performance, but a program with many memory accesses has a small range of performance improvement. When generating a program, how to improve the performance of a program with many memory accesses is an issue.

  The number of memory accesses can be reduced by performing loop transformation such as loop fusion, and by deleting the common expression to combine accesses to the same memory existing in a plurality of loops into one. By performing such processing in the compiler and reducing the number of memory accesses in the execution program, it is possible to improve performance in terms of software.

In a program such as Fortran, the lower limit / upper limit of an array referred to by an array formula may be declared using a variable. In that case, since the shape of the sequence cannot be determined at the time of translation, the sequence formula may not be sufficiently optimized. An example of a technique for compiling by replacing variables with constants is described in Patent Document 1. In Patent Document 1, a variable that has been defined once is selected, and the variable is replaced with a constant.
JP-A-2-118732

  By the way, programming languages such as Fortran, which are widely used in scientific and engineering calculations, can calculate for each element of an array by describing the calculation of the array and the array, and the expression that assigns the calculation result in one sentence. An array expression is defined. FIG. 23 shows an example of a source program. A (m1: m2) in Formula 1-1 represents a partial array that is a part of the array including elements a (m1) to a (m2) of the array a. m1 is called the lower limit, and m2 is called the upper limit. The dimension of the array represents the number of elements of the array to be accessed. For example, in the case of a (1: 5), since the five elements a (1), a (2), a (3), a (4), and a (5) are included, the size of the array is 5. . In the expression 1-1, for elements from m1 to m2, an array b and an array c are expressed as a (m1) = b (m1) + c (m1) and a (m1 + 1) = b (m1 + 1) + c (m1 + 1). This means that the addition result for each element is assigned to each element of the array a. Since the calculation is performed for each element of the array, it is normal to describe only arrays of the same shape in one array formula. On the other hand, Formula 1-3 is the entire sequence. Expression 1-3 represents substituting each element of the partial array of c for each element of x.

  The compiler internally generates a loop having an initial value and a final value calculated based on the lower limit and upper limit of the array appearing in the array expression. The compiler optimizes the generated loop, fuses the same number of loops, and deletes common expressions to reduce access to the same array elements in the loop and speed up the program. Plan. In order to fuse a plurality of loops in this optimization process, the number of loop iterations calculated based on the initial value and the final value of the loop must be the same during translation. However, in programs such as Fortran, the lower and upper bounds of an array referenced in an array expression may be declared using variables. In this case, the shape of the array cannot be determined during translation. In some cases, optimization could not be performed sufficiently. This will be described below.

  FIG. 24 shows intermediate code generated from the source program of FIG. The array formulas of Formula 1-1 to Formula 1-3 and Formula 1-5 included in the source program of FIG. 23 are converted into four loops (from do to enddo) included in FIG. 24 by the intermediate code generation unit. The When converting to a loop, if the array is an entire array, the initial and final values of the loop are determined based on the lower and upper limits in the declaration expression, and in the case of a partial array, based on the lower and upper limits of the reference location. To determine an initial value and a closing price. For example, since the array formula 1-1 is an operation of the partial array (m1: m2) and the size is m2-m1 + 1, and the array formula 1-3 includes the partial array c (m1: m2). The initial value of the loop is set to 0 and the closing value m2-m1.

  In the conventional method, using the fact that the dimensions and dimensions of all the arrays referenced in one array formula are the same, based on the lower limit and upper limit of one array referenced in the array formula, The initial value and the closing price were determined. For example, with regard to the array formula 1-3, the loop count is determined as m2-m1 + 1 by paying attention only to the number of dimensions and dimensions of the array referred to in the array formula 1-3. As described above, in the conventional method, the initial value and the final value of the loop are not determined based on the information on the lower limit / upper limit of all the arrays referenced in the array formula. If so, multiple sequence loops could not be fused between multiple sequence formulas. As a result, access to the same array element in the loop cannot be consolidated by deleting the common expression, and memory access cannot be reduced.

  For example, in FIG. 24, the loop count of the array formulas 1-1 and 1-3 is set to m2-m1 + 1 times, and the loop count of the array formula 1-2 is set to n2-n1 + 1 times. Here, paying attention to the array formula 1-3, it can be seen that n2-n1 and m2-m1 are equal since the dimensions of the entire array x and the dimensions of the partial array c (m1: m2) are the same. If this information can be applied to other array formulas, the loop count n2-n1 + 1 in the array formula 1-2 can be replaced with m2-m1 + 1, and three loops can be combined into one. is there. However, in the prior art, since such an analysis is not performed, the compiler cannot recognize that m2-m1 and n2-n1 have the same value, and cannot compile a loop.

  Regarding replacement of a variable with a constant, Patent Document 1 selects a variable that has been defined once and performs replacement from the variable to the constant. However, in Patent Document 1, a variable that is not defined in the procedure cannot be replaced with a constant. In addition, only variables are replaced with constants, and variables and expressions are not analyzed. For this reason, when the number of loops is defined using different variables or expressions, it is impossible to recognize that the number of loops is the same and fuse the loops.

  The present invention provides a compiler, a variable optimization device, a method, and a program that can solve the above-described problems of the prior art and can perform loop fusion even when the number of loops is defined using different variables and expressions. The purpose is to do.

The variable optimization apparatus of the present invention
To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting unit for detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detection unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detection position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimensional formula replacement unit that replaces the dimensional formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes;
It is characterized by providing .

The compiler of the present invention
Computer
To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, wherein the intermediate code, sequence type detecting unit detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detecting unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detected position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimension formula replacement unit that replaces the dimension formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes,
A code generation unit for generating an object program based on the intermediate code loop-optimized by the optimization unit;
It is made to function as .

The variable optimization method of the present invention includes:
Computer, to read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting step of detecting a plurality of sequence type,
For each of the plurality of arrays each calculated by the plurality of array formulas detected in the array formula detection step, the computer expresses a dimension formula representing the dimensions of the array as a variable representing the upper limit and a variable representing the lower limit. calculated using the dimensions of the first array dimension expression is calculated using the first variable, a dimension of the second sequence size equation using the second variable is calculated, it is A variable relationship determination step for determining a relational expression representing a relation of the same dimensions when the first array and the second array are operated with the same array formula;
The computer is configured to calculate the first array and the second array at a position before the array expression detection position detected from the intermediate code, and the first variable or the second array Variable definition detection that detects a pre-defined expression that defines a value of a variable and detects a post-defined expression that defines the value of the first variable or the second variable at a position after the position of detecting the array expression Steps,
Based on the relational expression determined in the variable relation determining step, the computer detects a pre-defined expression in the variable definition detecting step among a plurality of array dimensional expressions respectively calculated in the plurality of array formulas. Dimensional expression of the array calculated by the array expression detected at the position from the position to the position where the post-definition expression is detected at the variable definition detection step, calculated using the first variable A dimension formula replacing step for replacing the dimension formula with a dimension formula calculated using the second variable;
The plurality of loops from a plurality of loop processes repeatedly executed according to each of the plurality of array formulas, wherein the computer calculates an array of dimensions represented by the same dimension formula as the dimension formula replaced in the dimension formula replacing step. An optimization step for performing a loop optimization that deletes a common arithmetic expression common to the processes and fuses the plurality of loop processes;
It is characterized by having .

The compiling method of the present invention
Computer, to read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting step of detecting a plurality of sequence type,
For each of the plurality of arrays each calculated by the plurality of array formulas detected in the array formula detection step, the computer expresses a dimension formula representing the dimensions of the array as a variable representing the upper limit and a variable representing the lower limit. calculated using the dimensions of the first array dimension expression is calculated using the first variable, a dimension of the second sequence size equation using the second variable is calculated, it is A variable relationship determination step for determining a relational expression representing a relation of the same dimensions when the first array and the second array are operated with the same array formula;
The computer is configured to calculate the first array and the second array at a position before the array expression detection position detected from the intermediate code, and the first variable or the second array Variable definition detection that detects a pre-defined expression that defines a value of a variable and detects a post-defined expression that defines the value of the first variable or the second variable at a position after the position of detecting the array expression Steps,
Based on the relational expression determined in the variable relation determining step, the computer detects a pre-defined expression in the variable definition detecting step among a plurality of array dimensional expressions respectively calculated in the plurality of array formulas. Dimensional expression of the array calculated by the array expression detected at the position from the position to the position where the post-definition expression is detected at the variable definition detection step, calculated using the first variable A dimension formula replacing step for replacing the dimension formula with a dimension formula calculated using the second variable;
The plurality of loops from a plurality of loop processes repeatedly executed according to each of the plurality of array formulas, wherein the computer calculates an array of dimensions represented by the same dimension formula as the dimension formula replaced in the dimension formula replacing step. An optimization step for performing a loop optimization that deletes a common arithmetic expression common to the processes and fuses the plurality of loop processes;
A code generation step in which the computer generates an object program based on the intermediate code loop-optimized in the optimization step;
It is characterized by having .

The variable optimization program of the present invention is:
Computer
To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, wherein the intermediate code, sequence type detecting unit detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detecting unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detected position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimension formula replacement unit that replaces the dimension formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes,
It is made to function as .

  In the compiler, variable optimization device, method, and program of the present invention, loop fusion is possible even when the number of loops is defined using different variables and expressions.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 shows a compiler according to the first embodiment of the present invention. The compiler 10 includes an intermediate code generation unit 11, a variable optimization unit 12, an optimization unit 13, and a code generation unit 14. The compiler translates a program such as Fortran and generates a load module. The function of each part of the compiler can be realized by executing a predetermined program on a computer.

  The source program 31 is a source of a program written in a predetermined programming language such as Fortran. The source program 31 is stored in a memory or the like. The intermediate code generation unit 11 reads the source program 31 and generates an intermediate code 32 based on the source program 31. The generated intermediate code 32 is stored in a memory or the like.

  The variable optimization unit 12 reads the intermediate code 32, performs variable optimization processing in the intermediate code 32, and generates an intermediate code 36. The generated intermediate code 36 is stored in a memory or the like. The optimization unit 13 performs vectorization, parallelization, loop optimization, common expression deletion, and the like on the intermediate code 36. The processing performed by the optimizing unit 13 includes at least loop optimization for grouping loops having the same number of loops and deletion of common expressions. The optimization unit 13 stores the intermediate code 37 as a result of the loop optimization or the like in a memory or the like.

  The code generation unit 14 reads the intermediate code 37 optimized by the optimization unit 13 and generates an object program 38 based on the intermediate code 37. The generated object program 38 is stored in a memory or a holding storage device. As the intermediate code generation unit 11, the optimization unit 13, and the code generation unit 14, existing intermediate code generation means, optimization means, and code generation means can be used, respectively.

  The variable optimization unit 12 includes declaration information / array expression detection means 21, reference array detection means 22, variable relationship determination means 23, variable definition detection means 24, and variable replacement means 25. The declaration information / array expression detection means 21 searches the intermediate code 32 to detect array declaration information and array expressions, and registers them in the declaration information table and array expression table 33. The reference sequence detection means 22 checks the sequence referred to in each sequence formula with respect to the sequence formula registered in the sequence formula table, and for each sequence, the type of sequence (full sequence / partial sequence), the lower limit of the sequence And the upper limit are registered in the reference sequence table 34. The upper limit and lower limit of the array are the lower limit and the upper limit registered in the declaration information table if the array is an entire array, and the lower limit and the upper limit at the reference location if the array is a partial array.

  The variable relationship determining means 23 refers to the reference array table 34, calculates the dimensions of the array based on the lower limit and the upper limit of the array, and utilizes the fact that the dimensions of the arrays in the same array formula are the same. To determine the values of variables and formulas used to calculate dimensions. If the value is not determined, the relationship with other variables and expressions is determined. For example, when a sequence having a lower limit / upper limit of n1 / n2 and a sequence having a lower limit / upper limit of m1 / m2 are referred to in the same sequence formula, the fact that both dimensions are the same, The relational expression between them is determined as n2-n1 = m2-m1. The variable relationship determining unit 23 registers a relational expression representing the determined value or relationship and a determination expression that is an array expression that has determined the relation in the fixed variable table 35.

  The variable definition detection unit 24 searches the intermediate code 32 before and after the determination formula registered in the fixed variable table 35 as a starting point, and the relational expression registered in the fixed variable table 35 is the lower limit or upper limit of the partial array. Check if it refers to a variable that appears. If they are referenced, search for the expression that defines the variable referenced by the relational expression before and after the decision expression, and define the variable that defines the variable before the decision expression (pre-defined expression) and the variable after it. And an expression (post-definition expression) to be detected. The variable definition detecting unit 24 registers the detected pre-defined formula and post-defined formula in the fixed variable table 35. When a plurality of pre-defined formulas and post-defined formulas are detected, a formula that may be executed in the execution path closest to the decision formula is registered. If there is no expression to be defined, NULL is registered in the pre-defined expression or post-defined expression of the definite variable table 35. When only the variables used in the declaration expressions of all the arrays are referred to, NULL is registered in the pre-defined expression and the post-defined expression of the defined variable table 35.

  The variable replacement means 25 searches the intermediate code 32 and refers to the right side or the left side (hereinafter, described as the left side) of the relational expression registered in the defined variable table 35 between the pre-defined formula and the post-defined formula. A place to be detected is detected, and this is replaced with the right side of the relational expression. In the search, when the pre-defined formula is NULL, the search is performed from the beginning to the post-defined formula, and when the post-defined formula is NULL, the search is performed from the pre-defined formula to the end. For example, if the relational expression is n2-n1 = m2-m1, if n2-n1 exists between the pre-defined expression and the post-defined expression in the intermediate code 32, it is replaced with m2-m1. To do. The variable replacement unit 25 outputs the replacement result as an intermediate code 36.

  The intermediate code 36 generated by the variable optimizing unit 12 is subjected to optimization such as loop optimization and common expression deletion in the optimizing unit 13, and memory access is reduced compared to the intermediate code 36. Intermediate code 37 is obtained. By generating the object program 38 from the intermediate code 37 in the code generation unit 14, the object program is generated from the intermediate code that has been optimized by the optimization unit 13 with respect to the intermediate code 32 generated by the intermediate code generation unit 11. As compared with the case of generating the program, a program with reduced memory access can be obtained.

  FIG. 2 shows an operation procedure of the variable optimization unit 12. The declaration information / array expression detecting means 21 searches the intermediate code 32, detects the array declaration information and the array expression, and registers them in the declaration information table and the array expression table 33 (step S1). 3 and 4 show data examples of the declaration information table and the array formula table, respectively. By detecting the declaration information of the array from the intermediate code 32 corresponding to the source program 31 shown in FIG. 23, the arrays a, b, c, d, e, x, y, and z are registered in the declaration information table ( FIG. 3). Further, by detecting an array formula from the intermediate code 32, array formulas corresponding to the array formulas 1-1 to 1-3 and 1-5 of FIG. 23 are registered in the array formula table (FIG. 4).

  The reference sequence detection means 22 creates a reference sequence table 34 for each sequence formula registered in the sequence formula table, and the array name and sequence of the sequence referenced by each sequence formula are stored in each reference sequence table. And the type of array (full array / partial array) are registered (step S2). For the shape (lower limit: upper limit), if the array is an entire array, the lower limit and the upper limit registered in the declaration information table are registered. If the array is a partial array, the lower limit and the upper limit at the reference location are registered. 5A to 5D show data examples of the reference sequence table. Referring to the array formula table shown in FIG. 4, there are four array formulas (array formulas 1-1 to 1-3 and 1-5) in the intermediate code 32, so a reference sequence table is created corresponding to these. (FIGS. 5A to 5D).

  In step S2, arrays a, b, and c are set for expression 1-1 (FIG. 5A), and arrays c, d, and e are set for expression 1-2 (FIG. 5B), C and x are registered in the reference sequence table for Expression 1-3 (FIG. 5C), and y and z are stored for Expression 1-5 (FIG. 5D). In addition, the fact that the array is the entire array is registered for the array x in Formula 1-3, and the fact that it is a partial array is registered for the other arrays. For x, which is the entire array, the shape (n1: n2) registered in the declaration information table is registered in the reference array table 34, and for other arrays, the shape of the reference location is registered in the reference array table 34. For example, referring to FIG. 23, since the reference location of the array a in Expression 1-1 is m1 to m2, (m1: m2) is registered as the shape of the array a in the reference array table for Expression 1-1. To do.

  The variable relationship determining means 23 refers to the reference sequence table 34, calculates the size from the lower limit and the upper limit of the array in each array formula, and utilizes the fact that the dimensions of each array in the same array formula are the same. Then, the values of variables and expressions to be referred to in the dimension calculation are determined. For example, in the same reference sequence table, when there are an array whose shape is defined by two variables and an array defined by two constants, using the fact that the dimensions of the arrays are the same, Variables can be made constant from the relationship. When the value is not determined, the relationship with other variables and expressions is determined. For example, in the same reference sequence table, when there is an array whose shape is defined by two variables and another array whose shape is defined by two different variables, the fact that both dimensions are the same is used. Thus, the relational expression between the four variables can be derived. The variable relationship determining unit 23 registers a relational expression representing the determined value or relationship and a determination expression that is an array expression that has determined the relation in the fixed variable table 35 (step S3).

  FIG. 6 shows an example of data in the definite variable table 35. Referring to the reference sequence tables (FIGS. 5A, 5B, and 5D) corresponding to the expressions 1-1, 1-2, and 1-5, in these, variables (m1, m2, or n1 having the same shape) , N2), there is no value or relational expression that can be determined from these. Referring to the reference sequence table corresponding to Equation 1-3 (FIG. 5C), the shape of the array x is (n1: n2), and the shape (reference location) of the array c is (m1: m2). Yes. Using the fact that both shapes are equal, a relational expression of n1-n2 = m1-m2 is obtained. The variable relationship determining unit 23 registers the obtained relational expression “n1-n2 = m1-m2” and the array formula “1-3” for which the relational expression is determined in the fixed variable table 35.

  The variable definition detection unit 24 searches the intermediate code 32 before and after the determination formula registered in the definite variable table 35, and refers to the variable in which the relational expression determined by the determination formula appears in the lower limit / upper limit of the partial array. Judge whether or not. If it is referenced, search for an expression that defines a variable referenced by the relational expression before and after the decision expression. The variable definition detection means 24 registers the expression that defines the variable before the decision formula obtained in the search as a pre-defined formula in the fixed variable table 35, and the formula that defines the variable after the decision formula is the post-defined formula. Is registered in the definite variable table 35 (step S4). When a plurality of pre-defined formulas or post-defined formulas are searched in the search, the formula that may be executed closest to the decision formula in the execution path is registered. If there are no pre-defined formula and post-defined formula, NULL is registered in the definite variable table 35. If the relational expression determined by the determination formula does not refer to a variable appearing at the lower limit / upper limit of the partial array, NULL is registered in the defined variable table 35.

  Whether or not the relational expression “n2-n1 = m2-m1” determined by the determination formula “1-3” refers to a variable appearing at the lower limit / upper limit of the partial sequence in the source program shown in FIG. When examined, the relational expressions refer to variables appearing at the lower limit / upper limit of the partial array in the array formulas 1-1, 1-2, and 1-5. When an expression that defines a variable to be referred to by a relational expression is searched before and after the determination expression, the value of n2 is defined by Expression 1-4 after the determination expression. Does not exist to define before the determinant. Therefore, the variable definition detection unit 24 registers Formula 1-4 in the post-definition formula of the definite variable table 35 and registers NULL in the pre-definition formula.

  If there is an expression that defines a variable of the relational expression before the decision expression, it is unclear whether the relation defined by the relational expression can be used as it is before the previous definition expression. Similarly, if there is an expression that defines a variable of a relational expression after the determination formula, it is unclear whether the relation defined by the relational expression can be applied after the definition formula. Therefore, the range in which the relational expression can be used is a range sandwiched between the pre-defined formula and the post-defined formula in the intermediate code 32. In FIG. 23, in the array formulas 1-1, 1-2, and 1-5, the relational expression refers to a variable that appears at the lower limit / upper limit of the partial array. Since it is after the post-defining formula, the relational formula can be applied to the array formulas 1-1 and 1-2.

  The variable replacement unit 25 searches the intermediate code 32 while referring to the fixed variable table 35 and the reference array table 34, and the relational expression registered in the fixed variable table 35 between the pre-defined formula and the post-defined formula. A part corresponding to the left side is detected, and the part is replaced with the right side (step S5). If a value has been determined, replace it with that value. If the pre-defined formula is NULL, the search is performed in the range from the beginning of the intermediate code to the post-defined formula, and if the post-defined formula is NULL, the search is performed in the range from the pre-defined formula to the end of the intermediate code. The variable replacement means 25 outputs the intermediate code 36 after the replacement.

  FIG. 7 shows the intermediate code 36 after the variable optimization unit 12 performs variable optimization. The intermediate code 32 before variable optimization is as shown in FIG. 24, and the array formulas 1-1 to 1-3 are independent loops. The number of loops is m2-m1 + 1 times, n2-n1 + 1 times, and m2-m1 + 1 times. In step S5, the relational expression “n2-n1 = m2-m1” is used to replace the portion represented by n2-n1 with m2-m1 in the intermediate code 32 in the range from the beginning to Expression 1-4. . Thereby, the intermediate code 36 shown in FIG. 7 is obtained.

  FIG. 8 shows an intermediate code 37 obtained by performing loop optimization and common expression deletion on the intermediate code 36. The intermediate code 37 shown in FIG. 7 is obtained by performing the loop optimization and the deletion of the common expression in the optimization unit 13 to obtain the intermediate code 37 with reduced memory access (FIG. 8). That is, the intermediate code 37 in which the array formulas 1-1 to 1-3 are described in one loop is obtained. Thereafter, an object program 38 is generated by the code generation unit 14 for the intermediate code 37, so that a memory access is reduced and a high-speed program is obtained.

  FIG. 9 shows the memory access and operation to be executed. As a comparative example, FIG. 10 shows memory access and calculation executed when an object program is generated from the intermediate code shown in FIG. In FIGS. 9 and 10, a portion indicated by an arrow represents a memory access. When the object program is generated from the intermediate code of FIG. 24, the number of memory accesses is 5 (m2-m1 + 1) +3 (n2-n1 + 1) +2 (10-n1 + 1) times. Since m2−m1 = n2−n1, this can be rewritten as 8 (m2−m1 + 1) +2 (10−n1 + 1). On the other hand, the number of memory accesses when the object program is generated from the intermediate code shown in FIG. 8 is 7 (m2−m1 + 1) +2 (10−n1 + 1) times. Therefore, by performing variable optimization processing, the number of memory accesses can be reduced by (m2−m1 + 1) times, and the program execution can be speeded up by that amount.

  In the present embodiment, the variable relationship determination unit 23 defines the lower limit / upper limit of each array referred to in the array formula using the fact that the dimensions of the arrays constituting the same array formula are the same. Determine the value of a variable or expression, or determine the relational expression between variables. When the value is determined, the value of the variable used in the array declaration in the procedure can be determined if it is an entire array, and the value of the variable or expression that determines the reference location can be determined if it is a partial array. When the value is not determined, the relationship with other variables and expressions is determined. For the partial array, it is analyzed whether or not the determined value or relationship is changed by another expression, and a range in which the determined value or relationship can be applied is obtained. Thereafter, the variable substitution means 25 converts the variable or expression into a constant according to the determined value. Also, using the determined relational expression, the variable or expression is replaced with the same variable or the same expression.

  The compiler internally generates a loop having an initial value and a final value calculated based on the lower limit and upper limit of the array appearing in the array expression. Whether or not the number of loops is the same since the initial value and the final value are defined in different variables by performing variable substitution in the variable substitution means 25 and making the variables constant and unifying the variables. It becomes possible to determine that the number of loops of a plurality of unknown loops is the same number of loops. As a result, loop fusion is possible and common expressions can be deleted. By deleting common expressions and grouping accesses to the same array element in a loop, the number of memory accesses can be reduced. Thereby, the program execution time can be shortened.

  Next, a second embodiment of the present invention will be described. FIG. 11 shows a compiler according to the second embodiment of the present invention. In the compiler 10 a of this embodiment, the variable optimization unit 15 includes declaration information / array expression detection means 21, reference array / continuous appearance array expression detection means 26, dimension comparison means 27, and branch addition means 28. In the first embodiment, when the variable optimizing unit 12 (FIG. 1) determines a variable or expression value that is referred to in the calculation of dimensions of each dimension at the time of translation, or a variable or expression having the same value, Replaced the original expression with the determined variable or expression. In the present embodiment, when a variable or expression value that is referred to in the calculation of dimensions of each dimension at the time of translation, or a variable or expression having the same value is not determined, a code that selects a code to be executed at the time of execution is generated.

  The declaration information / array expression detection means 21 searches the intermediate code 32, detects the upper and lower limits of the array declaration, and the array expression, and registers them in the declaration information table and the array expression table 33. The reference sequence / sequentially appearing sequence type detection means 26 is configured to select the sequence referenced in each sequence type, the type of sequence (total sequence / partial sequence), the lower limit of the sequence, and the sequence type registered in the sequence type table. The upper limit is checked and registered in the reference sequence table 39. Further, it is checked whether or not the array formula continues immediately before and after the array formula. If the array formula continues immediately before or immediately after, the array formula is registered in the reference sequence table 39.

  The size comparison unit 27 refers to the reference sequence table 39 to check whether or not it is possible to determine that the size of the sequence referred to in the continuous sequence formula is the same during translation. If there is a continuous array formula that cannot be determined to have the same size at the time of translation, the array formula is registered in the optimization candidate sequence list 40. The branch addition means 28 generates an intermediate code 36 in which codes for selecting whether the dimensions are the same or not are added to the intermediate code 32 for the array formulas registered in the optimization candidate sequence list 40. To do. For the code executed when the dimensions are the same, the initial value and closing value of the other loop are replaced with the initial value and closing value of the loop generated for the first array expression.

  FIG. 12 shows an operation procedure of the variable optimization unit 15 of this embodiment. The declaration information / array expression detecting means 21 searches the intermediate code 32, detects the array declaration information and the array expression, and registers them in the declaration information table and the array expression table 33 (step S11). FIG. 13 shows an example of a source program. 14 and 15 show data examples of the declaration information table and the array formula table 33. FIG. By detecting array declaration information from the intermediate code 32 corresponding to the source program 31 shown in FIG. 13, the arrays a, b, c, d, e, and x are registered in the declaration information table (FIG. 14). Further, by detecting the array formula from the intermediate code 32, the array formula corresponding to the array formulas 9-1 to 9-3 in FIG. 13 is registered in the array formula table (FIG. 15).

  The reference sequence / continuous appearance sequence type detection means 26 creates a reference sequence table 39 for each sequence type registered in the sequence type table, and stores the reference sequence table 39 in each reference sequence table. The array name, array shape, and array type (full array / partial array) are registered. For each array formula, when there is an array formula immediately before or after the array formula, the array formula is registered in the reference sequence table 39 as the sequence formula immediately before or after (step S12). When registering the array shape (lower limit: upper limit), if the array is an entire array, register the lower limit and upper limit registered in the declaration information table, and if it is a partial array, register the lower limit and upper limit at the reference location. To do.

  FIGS. 16A to 16C show data examples of the reference sequence table. Referring to the array formula table shown in FIG. 15, since there are three array formulas (array formulas 9-1 to 9-3) in the intermediate code, the reference sequence / continuous appearance array formula detection means 26 corresponds to these. Thus, a reference sequence table is created (FIGS. 16A to 16C). Thereafter, arrays a, b, and c are assigned to 9-1 (FIG. 16 (a)), and arrays c, d, and e are assigned to expression 9-2 (FIG. 16 (b)) and expressions 9-3. C and x are registered in the reference sequence table (FIG. 16C). The array types are all partial arrays. As the shape, the shape of the reference location of each array is registered in the reference array table 39. For example, referring to FIG. 13, since the reference location of the array a in Expression 9-1 is m1 to m2, (m1: m2) is registered as the shape of the array a in the reference array table for Expression 9-1. To do.

  Referring to FIG. 13, there is no other array formula immediately before the array formula 9-1, and immediately after the array formula 9-2. The reference sequence / sequentially appearing sequence expression detection means 26 registers “none” indicating that there is no sequence expression in “just before” of the reference sequence table corresponding to the array formula 9-1, and immediately after that, the array formula 9− 2 is registered. As for the array formula 9-2, since there is the array formula 9-1 immediately before and there is the array formula 9-3 immediately after, the array formula 9-1 is registered “just before” in the reference sequence table, and “immediately after”. The array formula 9-3 is registered in. Since the array formula 9-3 has the array formula 9-2 immediately before and there is no array formula immediately after it, the array formula 9-2 is registered “immediately before” in the reference sequence table, and “one” is set to “none”. sign up.

  The dimension comparison means 27 refers to the reference sequence table 39 and detects a location where the array formulas are continuous. Thereafter, for successive sequence formulas, it is examined whether or not the dimensions of the sequences referred to in these sequence formulas can be determined to be the same during translation. If there is a continuous array formula that cannot be determined to have the same dimension during translation, the array formula is registered in the optimization candidate sequence list 40 (step S13). Referring to FIGS. 16A to 16C, by examining “immediately before” and “immediately” of each array formula, it is detected that array formulas 9-1 to 9-3 are continuous array formulas. After detecting the consecutive array formulas, it is determined whether or not it is possible to determine that the dimensions of the arrays referred to in the continuous array formulas 9-1 to 9-3 are the same by referring to the shape of each array in the reference sequence table 39. Investigate.

  FIG. 17 shows a data example of the optimization candidate sequence expression list 40. Referring to FIG. 16A, the size of the array referred to in the array formula 9-1 is m2-m1 + 1. Referring to FIG. 16B, the dimension of the array referred to in the array formula 9-2 is n2-n1 + 1. In this case, since it is unclear whether or not the size of the sequence to be referenced is the same between the array formula 9-1 and the array formula 9-2, the size comparison unit 27 includes the optimization candidate sequence formula list 40, The array formulas 9-1 and 9-2 are registered (FIG. 17). Similarly, when it is determined whether or not it is possible to determine that the size of the array to be referred to is the same in the array formula 9-2 and the array formula 9-3, the size of the array to be referred to in the array formula 9-3 is m2- Since it is m1 + 1, it cannot be judged that it is the same. The dimension comparison unit 27 additionally registers the sequence formula 9-3 in the optimization candidate sequence formula list 40.

  The branch adding means 28 adds an intermediate code 36 to the intermediate code 32 by adding, to the intermediate code 32, a code for selecting whether the dimensions are the same or not for the array formulas registered in the optimization candidate sequence list 40. Is generated (step S14). The intermediate code 36 generated by the branch addition means 28 becomes the intermediate code output from the variable optimization unit 15. 18 and 19 show examples of intermediate code data before and after addition by the branch addition means 28. FIG. The intermediate code 32 shown in FIG. 18 is an intermediate code corresponding to the source program shown in FIG. 13 generated by the intermediate code generator 11. The branch addition means 28 generates a code to be executed from the intermediate code 32 when the dimensions of the arrays referred to in the array formulas 9-1 to 9-3 are the same, that is, m2-m1 = n2-n1 is established. Then, an intermediate code 36 to which this is added is generated (FIG. 19).

  The intermediate code 36 shown in FIG. 19 includes a code part (between if and else) executed when m2-m1 = n2-n1 is established and a code part (between else and endif) executed when m2−m1 = n2−n1 is not established. ). The code portion executed when the dimensions are the same is obtained by replacing the initial value and the end value of the loop corresponding to each array expression in the intermediate code 32 (FIG. 18) as the generation source with the array expression 9-1 which is the first array expression. It is replaced with the initial value and the closing price of the corresponding loop. That is, in FIG. 18, the initial value and the closing price of the loop corresponding to the array formula 9-2 that had the initial value = 0 and the closing price n2−n1 are the initial value and the closing price of the loop corresponding to the array formula 9-1, respectively. 0 and m2-m1 are replaced. The contents of the code portion executed when the dimensions are not the same are the same as those shown in FIG.

  The optimization unit 13 optimizes the intermediate code 36 (FIG. 19) output from the variable optimization unit 15. FIG. 20 shows the intermediate code 37 after optimization. In the optimized intermediate code 37, the loop is combined into one in the code portion that is executed when m2-m1 = n2-n1 holds. On the other hand, the code portion executed when m2-m1 = n2-n1 is not established remains the same as the intermediate code 32 shown in FIG. 18 because loop fusion cannot be performed.

  FIG. 21 shows memory access and calculation when the intermediate code shown in FIG. 20 is executed. When the object program generated based on the optimized intermediate code 37 is executed, the number of memory accesses when m2-m1 = n2-n1 is satisfied at the time of execution is 7 (m2-m1 + 1) times. Further, the number of memory accesses when m2−m1 = n2−n1 is not satisfied at the time of execution is 5 (m2−m1 + 1) +3 (n2−n1 + 1) times.

  When the intermediate code 32 shown in FIG. 18 is optimized by the optimization unit 13 without passing through the variable optimization unit 15, the code executed when m2-m1 = n2-n1 in the intermediate code 36 is not satisfied. Like the part, there is no place where the loop is fused. Therefore, when the object program generated based on the intermediate code 32 is executed, the number of memory accesses is 5 (m2−m1 + 1) +3 (n2−n1 + 1). When the values of m2-m1 and n2-n1 are equal, the number of memory accesses can be expressed as 8 (m2-m1 + 1) times.

  FIG. 22 shows a comparison table of the number of memory accesses. When the object program based on the intermediate code 36 is executed, the number of memory accesses is the same, that is, the size is different 7 times (m2−m1 + 1) when m2−m1 = n2−n1 holds, that is, m2. When -m1 = n2-n1 is not satisfied, 5 (m2-m1 + 1) +3 (n2-n1 + 1) is obtained. On the other hand, when the object program based on the intermediate code 32 is executed, the number of memory accesses is when m2−m1 = n2−n1 is satisfied, and when 8 (m2−m1 + 1) and m2−m1 = n2−n1 are not satisfied. This is 5 (m2-m1 + 1) +3 (n2-n1 + 1) times. In this way, the variable optimization unit 15 adds a code that is executed when the dimensions of the arrays referred to by the continuous array formula are the same, so that the number of memory accesses when the dimensions are the same at the time of execution is (m2−m1 + 1). ) Times can be reduced.

  In this embodiment, when it is unclear whether or not the shapes of the array formulas that appear successively match at the time of translation, the code that is executed when the dimensions are the same is added to the intermediate code, and the dimensions are equal at the time of execution The code executed when the dimensions are the same and the code executed when the dimensions are different are called. The code portion executed when the dimensions are the same can be loop-fused and the common expression can be deleted by the optimization unit 13. Therefore, at the time of execution, the number of memory accesses when the dimensions of consecutive array formulas are equal can be reduced, and the execution speed of the program can be improved.

  As described above, the present invention has been described based on the preferred embodiment. However, the compiler, the variable optimization device, the method, and the program of the present invention are not limited to the above-described embodiment. Those in which various modifications and changes have been made to the configuration are also included in the scope of the present invention.

The block diagram which shows the compiler of 1st Embodiment of this invention. The flowchart which shows the operation | movement procedure of a variable optimization part. The figure which shows the example of data of a declaration information table. The figure which shows the example of data of an arrangement | sequence formula table. (A)-(d) is a figure which shows the example of data of a reference sequence table. The figure which shows the example of data of a definite variable table. The figure which shows the intermediate code after optimization by the variable optimization part. The figure which shows the intermediate code after the optimization by the optimization part. The figure which shows the memory access in the intermediate code shown in FIG. The figure which shows the memory access in the intermediate code | symbol which does not pass through the optimization by a variable optimization part. The block diagram which shows the compiler of 2nd Embodiment of this invention. The flowchart which shows the operation | movement procedure of the variable optimization part in 2nd Embodiment. The figure which shows the example of a source program. The figure which shows the example of data of a declaration information table. The figure which shows the example of data of an arrangement | sequence formula table. (A)-(c) is a figure which shows the example of data of a reference sequence table. The figure which shows the example of data of an optimization candidate arrangement | sequence list. The figure which shows the intermediate code corresponding to the source program shown in FIG. The figure which shows the intermediate code after optimization by the variable optimization part. The figure which shows the intermediate code after the optimization by the optimization part. The figure which shows the memory access in the intermediate code shown in FIG. The figure which shows the memory access frequency before and behind variable optimization. The figure which shows an example of a source program. The figure which shows the intermediate code produced | generated from the source program of FIG.

Explanation of symbols

10: Compiler 11: Intermediate code generation unit 12, 15: Variable optimization unit 13: Optimization unit 14: Code generation unit 21: Declaration information / array expression detection unit 22: Reference sequence detection unit 23: Variable relationship determination unit 24: Variable definition detection means 25: variable replacement means 26: reference sequence / continuous appearance array expression detection means 27: dimension comparison means 28: branch addition means 31: source programs 32, 36, 37: intermediate code 33: declaration information table, array expression Tables 34 and 39: Reference sequence table 35: Deterministic variable table 38: Object program 40: Optimization candidate sequence expression list

Claims (6)

To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting unit for detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detection unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detection position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimensional formula replacement unit that replaces the dimensional formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes;
A variable optimization device comprising: Among the plurality of sequence formulas detected by the sequence formula detection unit, is it determined that the dimensions of the plurality of sequences calculated by each of the plurality of sequence formulas processed in succession are the same during translation? A size discriminating unit for discriminating whether or not based on the dimensional formula calculated by the variable relationship determining unit;
When the variable relationship determining unit determines that the dimensions of the plurality of arrays are the same at the time of translation, it is determined that the dimensions of the plurality of arrays are not determined at the time of translation. And a non-identical dimension process executed when the dimensions of the plurality of arrays are not the same, and a branch addition unit for adding a process to branch the process to,
The optimization unit, according to each of the plurality of array formulas that are processed in succession, a loop optimization of a plurality of loop processing that is repeatedly executed as the same size processing,
The variable optimization apparatus according to claim 1, wherein: Computer
To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, wherein the intermediate code, sequence type detecting unit detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detecting unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detected position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimension formula replacement unit that replaces the dimension formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes,
Variable optimization program to function as Computer, to read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting step of detecting a plurality of sequence type,
For each of the plurality of arrays each calculated by the plurality of array formulas detected in the array formula detection step, the computer expresses a dimension formula representing the dimensions of the array as a variable representing the upper limit and a variable representing the lower limit. calculated using the dimensions of the first array dimension expression is calculated using the first variable, a dimension of the second sequence size equation using the second variable is calculated, it is A variable relationship determination step for determining a relational expression representing a relation of the same dimensions when the first array and the second array are operated with the same array formula;
The computer is configured to calculate the first array and the second array at a position before the array expression detection position detected from the intermediate code, and the first variable or the second array Variable definition detection that detects a pre-defined expression that defines a value of a variable and detects a post-defined expression that defines the value of the first variable or the second variable at a position after the position of detecting the array expression Steps,
Based on the relational expression determined in the variable relation determining step, the computer detects a pre-defined expression in the variable definition detecting step among a plurality of array dimensional expressions respectively calculated in the plurality of array formulas. Dimensional expression of the array calculated by the array expression detected at the position from the position to the position where the post-definition expression is detected at the variable definition detection step, calculated using the first variable A dimension formula replacing step for replacing the dimension formula with a dimension formula calculated using the second variable;
The plurality of loops from a plurality of loop processes repeatedly executed according to each of the plurality of array formulas, wherein the computer calculates an array of dimensions represented by the same dimension formula as the dimension formula replaced in the dimension formula replacing step. An optimization step for performing a loop optimization that deletes a common arithmetic expression common to the processes and fuses the plurality of loop processes;
A variable optimization method having : Computer
To read out the intermediate code containing the sequence formula for calculating a plurality of sequences, wherein the intermediate code, sequence type detecting unit detecting a plurality of sequence type,
For each of the plurality of arrays calculated by the plurality of array formulas detected by the array formula detection unit, a dimension formula representing the dimensions of the array is calculated using a variable representing the upper limit and a variable representing the lower limit of the array. The dimension of the first array in which the dimensional expression is calculated using the first variable and the dimension of the second array in which the dimensional expression is calculated using the second variable are: A variable relationship determining unit that determines a relational expression representing a relation of the same dimensions when the array and the second array are operated with the same array formula;
The array formula for calculating the first array and the second array is a position before the array expression detection position detected from the intermediate code, and the value of the first variable or the second variable is A variable definition detecting unit that detects a pre-defined formula that detects a pre-defined formula that defines a value of the first variable or the second variable at a position after the sequence formula detected position;
Based on the relational expression determined by the variable relation determination unit, from the position where the pre-defined expression is detected by the variable definition detection part among the dimensional expressions of the plurality of arrays respectively calculated by the plurality of array formulas , The dimension formula of the array calculated by the array formula detected at the position up to the position where the post-definition formula is detected by the variable definition detection unit, and the dimension formula calculated using the first variable, A dimension formula replacement unit that replaces the dimension formula calculated using the second variable;
In accordance with each of the plurality of array formulas for calculating the array of dimensions represented by the same dimension formula as the dimension formula replaced by the dimension formula replacement unit, each of the loop processes that are repeatedly executed is common to the plurality of loop processes. An optimization unit that performs a loop optimization that deletes the common arithmetic expression and fuses the plurality of loop processes,
A code generation unit for generating an object program based on the intermediate code loop-optimized by the optimization unit;
Compiler to function as Computer, to read out the intermediate code containing the sequence formula for calculating a plurality of sequences, from the intermediate code, and sequence type detecting step of detecting a plurality of sequence type,
For each of the plurality of arrays each calculated by the plurality of array formulas detected in the array formula detection step, the computer expresses a dimension formula representing the dimensions of the array as a variable representing the upper limit and a variable representing the lower limit. calculated using the dimensions of the first array dimension expression is calculated using the first variable, a dimension of the second sequence size equation using the second variable is calculated, it is A variable relationship determination step for determining a relational expression representing a relation of the same dimensions when the first array and the second array are operated with the same array formula;
The computer is configured to calculate the first array and the second array at a position before the array expression detection position detected from the intermediate code, and the first variable or the second array Variable definition detection that detects a pre-defined expression that defines a value of a variable and detects a post-defined expression that defines the value of the first variable or the second variable at a position after the position of detecting the array expression Steps,
Based on the relational expression determined in the variable relation determining step, the computer detects a pre-defined expression in the variable definition detecting step among a plurality of array dimensional expressions respectively calculated in the plurality of array formulas. Dimensional expression of the array calculated by the array expression detected at the position from the position to the position where the post-definition expression is detected at the variable definition detection step, calculated using the first variable A dimension formula replacing step for replacing the dimension formula with a dimension formula calculated using the second variable;
The plurality of loops from a plurality of loop processes repeatedly executed according to each of the plurality of array formulas, wherein the computer calculates an array of dimensions represented by the same dimension formula as the dimension formula replaced in the dimension formula replacing step. An optimization step for performing a loop optimization that deletes a common arithmetic expression common to the processes and fuses the plurality of loop processes;
A code generation step in which the computer generates an object program based on the intermediate code loop-optimized in the optimization step;
Compile method having

Images