JPH07225753A - Compiler system - Google Patents

Abstract

PURPOSE:To provide a compiler system for deciding the validity of vectorization and executing the vectorization even when there is pointer indirect quotation concerning the processing of a compiler at a computer. CONSTITUTION:A source program is analyzed by an analysis part 1, a data dependence relation setting part 3 performs data dependence relation setting processing, an optimization executing part 4 executes optimization while referring to a data dependence relation table, a code generating part 5 generates an object program prepared by vectorizing an vectorization enabled part, a data quotation classification part 2 discriminates each of data quotation concerning the analyzed result of the source program, defines each of data quotation of the pointer indirect quotation, quotation concerning an external variable and quotation of a variable processed with an address as A classification and defines the other quotation as B classification, and the data dependence relation setting part 3 performs the data dependence relation setting processing only for all the data quotation included in the B classification and the analysis of only a pair of data quotation included in the A classification is unconditionally made impossible.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a compiler device for translating a source program of a computer and generating a target program having a vectorized portion.

[0002]

2. Description of the Related Art As is widely known, a programming language called C has a pointer function, and a variable declared as a pointer is used to set the value of the variable. A so-called pointer indirect reference for quoting data whose address is can be described in the source program.

FIG. 2 shows an example of a C language program including pointer indirect reference. In FIG. 2 (a), a line defines a function below the next line, indicating that the function is func1 (). .

Is a declaration of double type data, a line is a declaration of int type data, and the double type data includes array data a and pointer (indicated by *) p. The line reads the value from the terminal as a decimal number (indicated by "% d") and assigns it to variable j (& j indicates the address of variable j). The line finds the address of a [j] and sets the pointer. Indicates that the value is assigned to p.

A line describes a loop in a program for
The statement indicates that the assignment statement in the subsequent line is repeated while the variable i is incremented by 1 from 0 to 9. In the row, p [i] indicating the assignment destination is a pointer indirect reference, and indicates the i-th element of the array whose head element is the data pointed to by the pointer p.

As is well known, the calculation part of an array by a loop, which is generally composed of for statements, etc., is a main target for generating a vectorized object program for processing by a vector processing device. .

Therefore, the compiler having the vectorization function, as an optimization process, first, at least one of the data citations in the loop is a data citation that defines a value.
The data dependence is examined for all combinations of two data citations for the same sequence data, and a data dependence relationship table is created in which values indicating the data dependence for each of those data citation pairs are set.

Next, referring to the data dependency table,
If there is no data dependency or the program can be changed so as to avoid the influence, change the program and generate the vectorized object program.

However, conventionally, when the pointer indirect reference is included, the following situation may occur, so that the optimization is unconditionally not performed. That is,
For the data citation pair of the loop, all the values set in the data dependency table are unconditionally set to "?" To indicate that the data dependency cannot be analyzed. Therefore, the optimization processing that refers to the data dependency table is performed. Then, we have no choice but to judge that it cannot be vectorized.

The reason for such treatment is as follows. For example, in the example of FIG. 2A, assume that the values of the variables j and a immediately before the execution of the for statement are as follows. j = 2 a = (1, 2, 3, 4, 5, 6, 7, 8, 9,10,11,12,13,1
4, .....) If it is executed sequentially without vectorization, a after executing the for statement is a = (1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1 , 2,13,1
4, .....) is obvious when following the program.

However, when vectorized, the sequence of values of a before execution is read into the vector register of the vector processing device all at once, and it is sent to the pipeline type arithmetic unit for arithmetic operation, and then sequentially output from the pipeline. For example, the results obtained are sequentially stored in different areas of the vector register and are written to the area of a at the same time. Therefore, the value of a after execution is a = (1, 2, 1, 2, 3, 4, 5, 6, 7, 8, 9,10,13,1
4, .....), which is different from the result of a above.

[0012] As described above, vectorization causes a movement different from that intended in the description of the program, resulting in a different calculation result. Therefore, if such a result is obtained, vectorization should not be performed.

Next, FIG. 2 (b) is another example using the pointer indirect reference. Since the sentence is similar to that of (a), the individual description will be omitted. Since the indirect reference is used for the reference and the variable b of the indirect reference is also used only for the reference in the line, vectorization does not result in the result like the example of (b) due to the pointer indirect reference. .

That is, in the case of the example of FIG. 2B, it is desired that even if there is pointer indirect reference, it is possible to generate an efficient target program by vectorizing. The present invention efficiently selects a case where it is relatively easy to determine that vectorization is possible even if there is pointer indirect citation,
A compiler device having an expanded vectorization range is intended.

[0015]

FIG. 1 is a block diagram showing the configuration of the present invention. The figure shows the configuration of the compiler device. The analysis unit 1 analyzes a source program written in a programming language having a pointer function, and the data dependency relationship setting unit 3 analyzes the program portion forming a loop according to the analysis result. Of the combinations of the two data citations extracted from the data citations, the data dependency relationship is analyzed for at least one of the data citations for which at least one data citation is a data citation defining a value, and each combination having a data dependency relationship,
For each combination in which the data dependency relationship cannot be analyzed, a data dependency relationship setting process for setting the display of each combination and a predetermined data dependency relationship value in the data dependency relationship table is performed, and the optimization execution unit 4 causes the data dependency relationship By referring to the relation table, it is possible to identify whether or not vectorization is possible, perform the necessary optimization when vectorization is possible, and the code generation unit 5 follows the optimization result to create a target program that vectorizes the vectorizable part. It is a compiler device for generating and includes a data citation classifying unit 2.

The data citation classification unit 2 determines each data citation in the analysis result of the source program, and each data citation is a pointer indirect citation, an external variable citation,
And the data citation is classified as A on the condition that it is one of the citations of the variable whose address is taken, and the data citation is classified as B when it does not belong to the A classification.

The data dependency relationship setting unit 3 determines only all data citations of the B class classified by the data citation classifying unit 2.
The data dependency setting processing is performed, and all the data citations of the A classification are extracted from all the data citations of at least one of the combinations of the two data citations, and at least one of the data citations is a data citation that defines a value. The data dependency value indicating that analysis is impossible is set in the data dependency table.

[0018]

With the compiler apparatus according to the present invention, it is analyzed whether or not vector indirect pointer indirect reference can be vectorized if it is clear that there is no data dependency, and vectorization is performed if possible.

[0019]

EXAMPLE FIG. 3 is a diagram showing an example of the processing flow of the data citation classifying section 2 of FIG.
It is assumed that data citations are sequentially checked for vectorization candidate portions such as for loops, and one data citation is taken out in processing step 10.

In processing step 11, it is identified whether the data citation is a pointer indirect citation (that is, a citation made in the form of p [i] as described above, where the pointer is p), and it is a pointer indirect citation. For example, in processing step 14, the data of category A is cited.

If it is not a pointer indirect reference, it is identified in processing step 12 whether the data reference is an external variable (that is, a variable defined outside the function definition, etc.), and if it is an external variable, it is classified in processing step 14. Data from A is used. This is because it is possible that the address is taken externally (described later) with the above external variables.

If none of the above, processing step 13
Identifies whether the data citation is addressed. Here, taking an address means that it is described that an array variable for data reference is asked for the address and the address value is substituted into a pointer, and therefore the above description is given. The variable that is the target of the & operator is a typical example in the above program example, but other cases where addresses are implicitly taken due to the specifications of the programming language shall also be considered.

As a result of the above identification, if the address is obtained, it is classified as the data reference of the classification A in the processing step 14, and the above 3
If none of the conditions is satisfied, the processing step 15 sets the data citation as the classification B, and the above classification processing is repeated in the processing step 16 by discriminating whether or not all the data citations have been performed.

Regarding the data citation of the program example of FIG. 2 (b), the classification result by the above processing will be described. In this example, p [i], a [i],
Since there are b [i] and a [i + 100], they are sequentially taken out and examined.

As a result, p [i] corresponds to the condition of pointer indirect reference, a [i] does not correspond to any of the three conditions, and b [i] is neither pointer indirect reference nor external variable. , The address is taken (line in Figure 2 (b)), which corresponds to the condition, a [i + 10
Since 0] does not correspond to any condition, p [i] and b [i] are classified as A, and a [i] and a [i + 100] are classified as B.

FIG. 4 is a diagram showing an example of the processing flow of the data dependency relationship setting section 3 of FIG.
According to the result of classification, the data dependence is sequentially examined for each pair of two data citations in each classification. First, in processing step 20 for the class A, a pair of data citations are extracted.

If both data citations of the pair retrieved in processing step 21 are both references (neither is a data citation that defines a value), there is no data dependency relationship for this data citation pair, so processing is required. There is no one.

If at least one of the pairs is a data reference that defines a value, in step 22, the data reference of the pair and "?" (Not analyzable) as the value of the data dependency are set in the data dependency table. Set.

The above-mentioned processing is repeated while performing the above-described processing while performing the identification in the processing step 23 so that all the two combinations of the data citations classified into the classification A are identified. Class B is processed.

In the case of the program of the above example, since the data citations of classification A were p [i] and b [i], when this pair is processed, both are data references. It is judged that there is no.

For category B, the data dependency relationship for each pair of two data citations is processed in the same manner as the conventional data dependency analysis processing, and the data dependency relationship for each required pair is calculated. Set to.

That is, if a pair of data citations of classification B is taken out in processing step 24, at least one of the pairs retrieved in processing steps 25 and 26 defines a value, and if they are data citations of the same sequence data, data dependence It is assumed that the analysis of the relationship is required, and that the other pairs do not need to be processed because there is no data dependency relationship.

The analysis of the data dependency is performed in the processing step 27.
Then, if it is determined that there is a data dependency by performing necessary calculation between the subscripts of both data citations and determining that there is a data dependency, the data dependency table of the pair and the data dependency of the data dependency are displayed in processing step 28. Set the value of the calculation result.

For example, a [i] and a which are classified into class B in the above example
For the pair [i + 100], the subscripts i and i + 100 in both data citations
For and, the difference between the two is taken to obtain -100, and the result is divided by 1 in the for loop and 1 in the loop variable, resulting in -1
00 is calculated as the calculated value, and this is set as the set value of the data dependency.

This value is to be vectorized by the optimizing unit 4 of FIG. 1 which makes a decision as to whether or not vectorization is possible by a usual conventional process, and performs an appropriate optimization if necessary. It is used as judgment data to enable it.

The above-described processing from processing step 24 to processing step 28 is repeated while identifying in processing step 29 so that all the two combinations of data citations classified into classification B are performed. Ends the process.

The optimization execution section 4 refers to the data dependency table created above to perform the optimization process. The details are omitted because the processing is the same as the conventional one. However, the data dependency “?” Is not unconditionally set in the data dependency table unlike the conventional one.
If it is a for loop, the calculated data dependency value "-100" is set only for the pair of a [i] and a [i + 100], so effective optimization processing is performed based on it. Be implemented.

[0038]

As is apparent from the above description, according to the present invention, in the optimization processing by vectorization in compiling a computer program, it is clear that pointer indirect reference also has no data dependency relationship. Since vectorization is analyzed and vectorization is performed if possible, there is a remarkable industrial effect that the vectorization rate can be improved and an efficient target program can be generated.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of the present invention.

FIG. 2 is a diagram illustrating a program example.

FIG. 3 is a flowchart of the process of the present invention (No. 1)

FIG. 4 is a flowchart of the process of the present invention (No. 2)

[Explanation of symbols]

 1 analysis unit 2 data citation classification unit 3 data dependency setting unit 4 optimization execution unit 5 code generation unit 10-16, 20-29 processing steps

Claims (1)

[Claims] 1. An analysis unit (1) analyzes a source program written in a programming language having a pointer function, and a data dependency setting unit (3) stores data of a program portion forming a loop according to the analysis result. Of the combinations of the two data citations extracted from the citation, the data dependency is analyzed for all the combinations in which at least one of the data citations is a data citation that defines a value, and each combination having the data dependency and the data dependency For each combination that cannot be analyzed, the data dependence setting process that sets the display of each combination and the predetermined data dependence value in the data dependence table is performed, and the optimization execution unit (4) See
A compiler device that identifies whether vectorization is possible, performs the necessary optimization when vectorization is possible, and the code generation unit (5) generates an object program that vectorizes the vectorizable part according to the optimization result. Therefore, a data citation classification unit (2) is provided, and the data citation classification unit (2) discriminates each data citation in the analysis result of the source program, and each data citation regards pointer indirect citation and external variable. Data citation and the citation of the variable whose address is taken, the data citation is classified as A, and the data citation not belonging to A is classified as B, and the data dependence setting unit (3 ) Is the data citation classification unit (2)
The data for which the data dependency setting process is performed only for all the data citations of the B category that have been classified, and data is extracted from all the data citations of the A category, and at least one of the data citation combinations defines a value. A compiler apparatus, which is configured to set the data dependency relationship value indicating unanalyzable unconditionally for each of the quoted combinations in the data dependency relationship table.