JP3658771B2 - Compiler device - Google Patents

Description

[0001]
[Industrial application fields]
The present invention relates to a compiler apparatus that performs processing of a compiler that translates a computer source program and generates a target program having a vectorized portion.
[0002]
[Prior art and problems to be solved by the invention]
As is widely known, a programming language called C language has a pointer function and uses a variable declared as a pointer to quote data whose value is the address of the variable, so-called pointer indirect quoting. Can be described in the source program.
[0003]
FIG. 2 shows an example of a C language program including indirect pointer citation. In FIG. 2 (a), line (1) defines a function in the following line (2) and below, and the function is func1 (). It shows that.
[0004]
(3) is a declaration of double type data, (4) is a declaration of int type data, and double type data includes array data a and pointer (indicated by *) p.
Line (5) reads the value from the terminal as a decimal number (indicated by "% d") and assigns it to variable j (indicating the address of variable j by & j), line (6): a [j] Is obtained and assigned to the pointer p.
[0005]
Line (7) is a for statement describing a loop of the program, and indicates that the variable (i) is incremented by 1 from 0 to 9 and the assignment statement in line (8) is repeated.
{Circle over (8)} p [i] indicating the assignment destination in the line indicates pointer indirect citation, and indicates the i-th element of the array having the data indicated by the pointer p as the first element.
[0006]
As is well known, the calculation part of an array by a loop generally composed of a for sentence or the like is a main object for generating a target program vectorized so as to be processed by a vector processing device.
[0007]
Therefore, a compiler having a 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, and for all combinations of two data citations for the same array data, The data dependency relationship is examined, and a data dependency relationship table in which a value indicating the data dependency relationship for each data citation pair is set is created.
[0008]
Next, with reference to the data dependency relationship table, if there is no data dependency or the program can be changed to avoid the influence, the data dependency relationship table is changed to generate a vectorized object program.
[0009]
However, in the case where conventional pointer indirect citation is included, there is a possibility that the following situation may occur, so optimization is not unconditionally 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 relationship cannot be analyzed, and accordingly, the optimum data reference table is referred to. In the conversion process, it must be determined that vectorization is impossible.
[0010]
The reason for this treatment is as follows. For example, in the example of FIG. 2A, assume that the values of variables j and a immediately before the for statement is executed are as follows.
j = 2
a = (1, 2, 3, 4, 5, 6, 7, 8, 9,10,11,12,13,14, .....)
If executed sequentially without vectorization, a after executing the for statement is
a = (1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2,13,14 .....)
It will be clear if you follow the program.
[0011]
However, when vectorization is performed, the sequence of values a before execution is read into the vector register of the vector processing unit all at once, sent to a pipeline type arithmetic unit, and the result sequentially output from the pipeline is obtained. For example, since the data is sequentially stored in different areas of the vector register and written to the area of a at the same time, the value of a after execution is
a = (1, 2, 1, 2, 3, 4, 5, 6, 7, 8, 9,10,13,14, .....)
Which is different from the result of a.
[0012]
As described above, since vectorization moves differently from what is intended in the description of the program, and as a result, the calculation result is different, it should not be vectorized when such a result is obtained.
[0013]
Next, FIG. 2 (b) is another example using pointer indirect citation, and since it is a sentence similar to (a), individual explanation is omitted. In this example, the pointer is on line (7). Since indirect citation is used for reference and the variable b of the indirect citation destination is also used only for reference in line (8), the result as in the example of (b) above that is caused by pointer indirect citation even when vectorized Does not occur.
[0014]
That is, in the case of the example of FIG. 2B, it is desired that an efficient target program can be generated by vectorization even if there is pointer indirect citation.
An object of the present invention is to provide a compiler apparatus in which the vectorization range is expanded by efficiently selecting a case where it is possible to relatively easily determine that vectorization is possible even if there is a pointer indirect citation.
[0015]
[Means for Solving the Problems]
FIG. 1 is a block diagram showing the configuration of the present invention.
The figure shows the configuration of a compiler apparatus, in which an analysis unit 1 analyzes a source program described in a programming language having a pointer function, and a data dependency setting unit 3 shows a program portion forming a loop according to the analysis result. Data dependency is analyzed for all combinations in which at least one of the two data citation combinations extracted from the data citation is a data citation that defines a value, and each combination having a data dependency and data dependency For each combination whose relationship cannot be analyzed, a data dependency setting process for setting the display of each combination and a predetermined data dependency value in the data dependency table is performed, and the optimization executing unit 4 displays the data dependency table. With reference to this, the possibility of vectorization is identified, and if necessary, the code generation unit 5 performs optimization necessary for vectorization. According to a compiler apparatus for generating an object program obtained by vectorizing the vectorization moiety, provided data reference classification section 2.
[0016]
The data citation classification unit 2 discriminates each data citation with respect to the analysis result of the source program, and each of the data citations is one of a pointer indirect citation, a citation about an external variable, and a citation of a variable whose address is taken. If the data citation does not belong to the A class, the data citation is set to the B class.
[0017]
The data dependency setting unit 3 performs the data dependency setting processing only for all data citations of the B classification classified by the data citation classification unit 2 and combines two data citations extracted from all the data citations of the A classification For each combination in which at least one of the data citations is a data citation that defines a value, the data dependency value indicating unconditional analysis is set in the data dependency table.
[0018]
[Action]
When it is clear that the pointer indirect citation has no data dependency, the compiler apparatus of the present invention analyzes whether or not vectorization is possible, and performs vectorization if possible.
[0019]
【Example】
FIG. 3 is a diagram illustrating an example of the processing flow of the data citation classification unit 2 in FIG. 1. Data citations are sequentially examined for vectorization candidate portions such as for loops based on the analysis result of the analysis unit 1. Step 10 retrieves one data citation.
[0020]
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 p is a pointer). The data is classified as category A in 14.
[0021]
If it is not a pointer indirect citation, it is determined at processing step 12 whether the data citation is an external variable (ie, a variable defined outside a function definition, etc.). Quote. This is because an external address such as the one described above may have an external address (described later).
[0022]
If none of the above, then step 13 identifies whether the data citation is addressed.
Here, taking an address means that the array variable of the data reference describes that the address is obtained and the address value is assigned to the pointer. In the example of the program, the variable that is the target of the & operator is a typical example, but other cases where the address is implicitly taken into account in the specification of the programming language are also considered.
[0023]
As a result of the above identification, if the address is taken, the data citation of the classification A is determined in the processing step 14, and if none of the above three conditions is satisfied, the data citation is determined as the classification B in the processing step 15. Has been done for all data citations or is identified in processing step 16 and repeated.
[0024]
Regarding the data citation in the program example of FIG. 2 (b), the classification result by the above processing will be described. In this example, p [i], a [i], b [i] and a Since there are [i + 100], they are taken out and examined sequentially.
[0025]
As a result, p [i] corresponds to the condition of indirect pointer citation, a [i] does not correspond to any of the three conditions, and b [i] is neither a pointer indirect citation nor an external variable. ([5] line in Fig. 2 (b)) is applicable, and a [i + 100] does not correspond to any condition, so p [i] and b [i] are classified as A, a [i] and a [i + 100] are classified as B.
[0026]
FIG. 4 is a diagram illustrating an example of the processing flow of the data dependency setting unit 3 in FIG. 1. According to the classification result of the data citation classification unit 2, data for each pair of two data citations is classified for each classification. It is assumed that the dependency relations are sequentially examined. First, a pair of data citations are extracted from the processing step 20 for the classification A.
[0027]
If both data citations in the pair extracted in processing step 21 are both references (both are not data citations that define values), there is no data dependency on this data citation pair, so no processing is necessary. To do.
[0028]
If at least one of the pairs is a data citation that defines a value, in processing step 22, the data citation of the pair and “?” (Unanalyzable) are set in the data dependency table as the data dependency value.
[0029]
The above processing is repeated while identifying in the processing step 23 so that the above processing is performed for all two combinations of the data citation classified into the classification A. When the processing of the classification A is completed, the classification B Process.
[0030]
In the case of the program of the above example, the data citation of the classification A was p [i] and b [i]. Therefore, when this pair is processed, since both are data references, there is no data dependency. To be judged.
[0031]
For classification B, the data dependency for each pair of two data citations is processed in the same manner as the conventional data dependency analysis processing, and the data dependency is set in the data dependency table for each required pair. .
[0032]
That is, if a pair of classification B data citations is extracted in processing step 24 and at least one of the pairs extracted in processing steps 25 and 26 defines a value, and the data citation of the same sequence data is analyzed, the data dependency is analyzed. It is assumed that there is no data dependency for the other pairs and no processing is required.
[0033]
The data dependency analysis is performed in processing step 27 by performing necessary calculations between the subscripts of both data citations, and if it is determined that there is a data dependency, the data dependency table is displayed in processing step 28. Set the data citation and the value of the data dependency calculation result.
[0034]
For example, for the pair of a [i] and a [i + 100] that are classified as B in the above example, the difference between the subscripts i and i + 100 of both data citations is taken to obtain -100, The result is divided by the increment of for loop 1 and the coefficient 1 of the loop variable to obtain -100 as a calculated value, and this is set as the data dependency setting value.
[0035]
This value is determined later by the optimization execution unit 4 of FIG. 1 by using the usual conventional processing to determine whether or not vectorization is possible, and if necessary, it is optimized and vectorized. It is used as judgment data.
[0036]
In order to perform the above processing from processing step 24 to processing step 28 for all two combinations of data citations classified into classification B, the processing is repeated while identifying in processing step 29, and processing of classification B is performed. End.
[0037]
The optimization execution unit 4 performs optimization processing with reference to the data dependency table created as described above.
Since the processing is the same as in the prior art, the details are omitted, but the data dependence “?” Is not unconditionally set in the data dependence table as in the past. If so, the calculated data dependency value “-100” is set only for the pair of a [i] and a [i + 100], and effective optimization processing is performed based on the data dependency value “−100”. The
[0038]
【The invention's effect】
As is clear from the above description, according to the present invention, in the optimization process by vectorization in the compilation of a computer program, if it is clear that there is no data dependency for pointer indirect citation, whether or not vectorization is possible Is analyzed, and vectorization is performed when possible. Therefore, there is a remarkable industrial effect that the vectorization rate can be improved and an efficient target program can be generated.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of the present invention. FIG. 2 is a diagram for explaining an example of a program. FIG. 3 is a flowchart of processing according to the present invention (part 1).
FIG. 4 is a flowchart of the processing of the present invention (part 2).
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Analysis part 2 Data quotation classification part 3 Data dependence setting part 4 Optimization execution part 5 Code generation part
10-16, 20-29 processing steps

Claims (1)

The analysis unit analyzes a source program written in a programming language having a pointer function,
The data dependency setting unit , for each combination in which at least one data citation is a data citation that defines a value among the two data citation combinations extracted from the data citation of the program part forming a loop according to the analysis result Data dependency that analyzes data dependency and sets display of each combination and predetermined data dependency value in data dependency table for each combination that has data dependency and combination that cannot be analyzed Perform the setting process,
The optimization execution unit refers to the data dependency table to identify whether or not the analysis result can be vectorized .
According code generator is the vector of a result, a compiler apparatus for generating an object program obtained by vectorizing the vectorization moiety,
A data citation classification unit is provided, the data citation classification unit determines each data citation for the analysis result of the source program analyzed by the analysis unit , and each data citation is a pointer indirect citation, a citation for an external variable, and On the condition that it is one of the citations of the variable whose address is taken, the data citation is classified as A classification, and the case where it does not belong to A classification is classified as B classification,
The data dependency setting unit further only for all data citations classified B classification of the data cited classification unit performs the data dependency setting process, taken from all the data quoted in the A classification out the two data reference For each combination in which at least one of the data citations is a data citation defining a value, the data dependency value indicating unconditional analysis is set in the data dependency table. A compiler apparatus characterized by that.

Images