JPH05324346A - Vectorization system - Google Patents

Abstract

PURPOSE:To generate an object program which can execute without a problem and without being affected by a loop length at the time of execution even when the loop length of a loop which needs operation arrangement at the time of vectorization is not identified at the time of compiling. CONSTITUTION:A loop structure selection means 5 selects loop structure where the loop length of the loop needing operation arrangement at the time of vectorization is not identified based on the results of the processing of a loop structure analyzation means 3 and a vectorization means 4. A loop structure deformation means 6 generates a third intermediate text 13 which likely processes by securing the operation arrangement of the number of elements for the loop length if the operation arrangement of such number of the elements can be secured at the time of execution and can likely process by the operation arrangement of the maximum number of the elements, which can be secured if the operation arrangement of such number of the elements can not be secured from a second intermediate text 11. An object program generation means 7 generates the object program 8 from the third intermediate program 13.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vectorization method in a compiler which inputs a source program written in a high level language and generates an object program for a vector computer.

[0002]

2. Description of the Related Art In a computer having a vector operation instruction (vector computer), an operation between a plurality of regularly arranged data strings (vector data) can be executed at a high speed by a vector instruction. In a compiler that translates a source program written in a high-level language such as FORTRAN language to generate an object program for a vector computer, an instruction sequence that analyzes the source program and processes it using vector instructions as much as possible is used. To generate. This is called vectorization. Vectorization is usually performed on the loop structure.

By the way, when performing vectorization, a work array may be required.

For example, in a vector computer of a system in which all results of vector operations are stored in a memory, F shown in FIG.
When vectorizing a loop written in the ORTRAN language, the intermediate result of C (I) * D (I) is stored for the expression B (I) + C (I) * D (I) on the right side of the assignment statement. It is necessary to prepare a working array for

On the other hand, in a vector computer in which the result of vector operation is stored in a vector register, since the intermediate result can be stored in the vector register, it is not necessary to use the working array as in the above example. However, in the loop structure, statements that can be vector processed and statements that cannot be vector processed are mixed. Therefore, divide the loop structure into vector processable parts and vector processable parts, and only the vector processable parts. A work array may be required when the vectorization method is used.

For example, in a loop described in FORTRAN language as shown in FIG. 3, a variable S is defined by an assignment statement that can be vector processed, and a WRI that cannot be vector processed.
It is referenced in the TE statement. Since the loop cannot be divided as it is, the entire loop cannot be vectorized. However, by assigning the work array w to the variable S, the loop can be divided as shown in FIG. 4, so that the loop including the assignment statement is vectorized.

[0007]

However, in the loop described in the FORTRAN language shown in FIGS. 2 and 3, since the loop end value is the variable M, the size of the work array w cannot be known until execution time. .. Therefore, in the conventional technology, a method of generating a target program that secures a necessary work area at a place needed at the time of execution, and the size of a work array required at the time of compilation are assumed, and The method of statically allocating the area was taken.

However, in the object program generated by these methods, if the necessary area cannot be secured at the time of execution or the size is larger than the size assumed at the time of compilation, there is a problem that the execution cannot be continued any more. there were.

The present invention has been made in view of the above circumstances, and an object of the present invention is to execute a loop at the time of execution even if the loop needs a work array whose length is unknown at the time of execution. An object of the present invention is to provide a vectorization method for generating an object program which can be executed without any limitation without being limited in length.

[0010]

In order to achieve the above object, the present invention provides a vectorizing method of a compiler for inputting a source program written in a high-level language to generate an object program for a vector computer. Equipped with such means.

(1) Recognizing the loop structure in the source program, generating a first intermediate text for the loop structure, and initializing the initial value init and the final value te of the loop structure.
rm, increment value incr, loop length m, and loop structure analysis means for generating a loop information table storing exponent variables (2) The first intermediate text of the loop structure obtained by the loop structure analysis means is input, and the loop structure is entered. Is converted into a second intermediate text to be executed by vector processing, and a work array information table storing work array information necessary for execution by vector processing is generated. (3) The loop structure Loop structure selecting means for inputting the loop information table generated by the analyzing means and the working array information table generated by the vectorizing means and selecting a loop structure in which the size of the required working array is not determined at compile time (4) For the loop structure selected by the loop structure selection means, the second structure generated by the vectorization means The interworking text, the work sequence information table, and the loop information table generated by the loop structure analysis means are input to secure the work sequence having the maximum number of elements that can be secured with the loop length m or less before the loop structure. An intermediate text for performing the process of setting the secured number of elements in the variable n is inserted, and the loop structure has an initial value 0, a final value term-incr, an increment value n × incr and an outer loop having an index variable i and an initial value i.
nit + i, final price min {init + i + (n-1) ×
incr, term}, with an increment value incr, it is transformed into a double loop structure consisting of an inner loop of the exponent variable I, and the definition and reference of the I-th element of the working array appearing in the loop (I
-Init-i) / incr + 1 definition to the 1st element,
A loop structure modifying means for generating a third intermediate text replaced with a reference, and (5) a target program generating means for generating a target program by inputting the third intermediate text generated by the loop structure modifying means

Further, in another configuration, the following means is provided in place of the loop structure deforming means and the object program generating means of the above (4) and (5).

(4 ') The second intermediate text generated by the vectorizing means, the work sequence information table, and the loop information table generated by the loop structure analyzing means are input and selected by the loop structure selecting means. In the loop structure, where L is a constant appropriately determined by the compiler, an outer loop having an index variable i with an initial value of 0, a final value of term-incr, and an increment value of L × incr, an initial value init + i, and a final value min {init + i + (L-1 ) × incr, te
rm} and the increment value incr are transformed into a double loop structure consisting of an inner loop of the exponent variable I, and the definition and reference of the working array element appearing in the loop are (I-init-i) / i.
Loop structure modification means for generating a third intermediate text, which is replaced with the definition and reference to the ncr + 1th element (5 ′) The third intermediate text generated by the loop structure modification means and the work generated by the vectorization means Program generating means for inputting a sequence information table for use and generating a target program in which a work sequence is allocated with a size L

[0014]

In the vectorization method of the present invention, when the compiler inputs a source program written in a high-level language and generates an object program for a vector computer, the loop structure analysis means recognizes the loop structure in the source program. Then, the first intermediate text for the loop structure is generated, and the loop information table storing the initial value init, the final value term, the increment value incr, the loop length m, and the exponent variable of the loop structure is generated, and then the vectorization means. Inputs the first intermediate text of the loop structure obtained by the loop structure analysis means, converts the loop structure into the second intermediate text to be executed by vector processing, and
A work sequence information table storing work sequence information necessary for execution in vector processing is generated, and then the loop structure selection means uses the loop information table generated by the loop structure analysis means and the vectorization means. Input the generated work array information table and select a loop structure in which the size of the required work array is not determined at compile time.

In the configuration including the means (4) and (5), if the work array having the number of elements corresponding to the loop length can be secured at the time of execution, the work array having such a number of elements is secured to perform the processing. In order to ensure that a work array with such a number of elements cannot be secured and the work array with the maximum number of elements that can be secured is secured so that processing can continue, For the loop structure selected by the selecting means, the second intermediate text generated by the vectorizing means, the work sequence information table, and the loop information table generated by the loop structure analyzing means are input, and the loop is inserted before the loop structure. An intermediate text that secures a work array with the maximum number of elements that can be secured with a length of m or less and sets the secured number of elements in a variable n is inserted. Period value of 0, the closing price term-incr,
An outer loop having an index variable i with an increment value n × incr, an initial value init + i, and a final value min {init + i + (n-
1) × incr, term}, with an increment value incr, it is transformed into a double loop structure consisting of an inner loop of the exponent variable I, and the definition and reference of the I-th element of the working array appearing in the loop are defined as (I-init- i) / incr + 1 Generates a third intermediate text which is replaced with the definition and reference to the 1st element, and the target program generating means inputs the third intermediate text generated by the loop structure transforming means to generate a target program. ..

Further, in the configuration provided with the means (4 ') and (5'), the work array having the number of elements corresponding to the constant L determined by the compiler is statically secured, and the processing is continued with that number of elements. In order to be able to do so, the loop structure modification means inputs the second intermediate text generated by the vectorization means, the work sequence information table and the loop information table generated by the loop structure analysis means, and the loop structure selection means. The loop structure selected in is set to the initial value 0 and the final value term.
-Incr, an outer loop having an index variable i with an increment value L x incr, an initial value init + i, and a final value min {init
+ I + (L-1) × incr, term}, increment value in
It is transformed into a double loop structure consisting of the inner loop of the exponent variable I with cr, and the definition and reference of the working array element appearing in the loop is further defined as the definition and reference to the (I-init-i) / incr + 1st element. The replaced third intermediate text is generated, and the object program generating means inputs the third intermediate text generated by the loop structure transforming means and the work array information table generated by the vectorizing means, and operates with the size L. Generates the target program with the allocated arrays.

[0017]

Embodiments of the present invention will now be described in detail with reference to the drawings.

Referring to FIG. 1, one embodiment of a compiler 2 to which the vectorization method of the present invention is applied is a loop structure analysis means 3, a vectorization means 4, a loop structure selection means 5, and a loop structure modification means. 6 and a target program generating means 7, which inputs the source program 1 and generates a target program 8.

In FIG. 1, 9, 11, 13, 1
Reference numerals 0 and 12 are the first, second, and third intermediate texts, the loop information table, and the work array information table, which are generated by the compiler 2 in the process of processing.

The function and operation of each unit will be described below.

The compiler 2 inputs the source program 1 described in a high level language and generates a target program 8.

At this time, the loop structure analyzing means 3 recognizes the loop structure appearing in the source program 1 and generates the first intermediate text 9 for the loop structure, and at the same time, the initial value init, the final value term of the loop structure,
The loop information table 10 storing the increment value incr, the loop length m obtained from them, and the index variable j = I of the loop structure is generated.

For example, in the case of the loop structure shown in FIG. 3, the loop structure analyzing means 3 has an initial value init = 1 and a final value term =.
With M and the increment value incr = 1, the loop information table 10 with the loop length m = M and the index variable j = I is generated.

Next, the vectorizing means 4 inputs the first intermediate text 9 of the loop structure and transforms the loop structure into the second intermediate text 11 which is executed by vector processing. Also,
Work array information table 12 that stores information such as size and type of work arrays required for vectorization
To generate.

For example, in the case of the loop structure shown in FIG. 3, the vectorization means 4 replaces the variable S with the working array w to divide the loop structure into two loop structures, as shown in FIG. Second intermediate text 11 like vector processing
To generate. Further, for the work array w, the work array information table 12 storing information that the variable before replacement is S and the type is a real number type is generated.

Next, the loop structure selection means 5 inputs the work sequence information table 12 generated by the vectorization means 4 and the loop information table 10 generated by the loop structure analysis means 3, and a work sequence is required. In addition, since the loop length is not a constant, the size of the working array is not determined unless it is executed at the time of execution.

In the example of FIG. 3, since the loop length is the variable M, the loop length is unknown at the time of compilation, and a work array for replacing the variable S is required, the loop structure selecting means 5 processes the loop structure. Selected as a target.

Next, the loop structure modification means 6 generates the third intermediate text 13 by modifying the intermediate text of the loop structure selected by the loop structure selection means 5 as follows.

First, the second generated by the vectorization means 4
The intermediate text 11 and the loop information table 10 generated by the loop structure analysis unit 3 are input, a work array having the maximum number of elements that can be ensured with a loop length m or less is secured immediately before the loop structure, and the secured elements are secured. Insert an intermediate text that performs the process of setting the number to the variable n.

Next, the loop structure is defined as an outer loop having an initial value 0, a final value term-incr, an increment value n × incr, and an index variable i, and an initial value init + i and a final value min.
{Init + i + (n-1) × incr, term},
2 with an inner loop of exponent variable I with increment value incr
It transforms into a heavy loop structure.

Further, the definition and reference of the I-th element of the working array appearing in the loop structure is (I-init-i) / i.
Replace with the definition and reference to the ncr + 1 element.

Therefore, in the case of the loop structure shown in FIG.
As shown in FIG. 5, the outer loop has an initial value of 0 and a final value of M.
-1, the increment value is n, the index variable is i, the initial value of the inner loop is 1 + i, the final value is MIN (i + n, M), the increment value is 1, and the index variable is transformed into a double loop structure of I, and ,
The definition and reference of the work array element w (I) is w {(I-1-
i) / 1 + 1} = w (I-i) is replaced with the definition and reference, and before the loop, the working array w having the maximum number of elements that can be secured with the loop length M or less is secured and the number of elements is The third intermediate text 13 in which the subroutine getwork (M, w, n) to be set to n is inserted is generated.

Finally, the object program generating means 7 is the third intermediate text 13 generated by the loop structure transforming means 6.
To generate a series of vector instructions and scalar instructions to generate the object program 8.

When the target program 8 generated as described above is executed by the vector computer, M is determined.
Then, if the work array w having the determined number of M elements can be secured, the processing is performed using such a work array w. Further, even if only the work array w having the number of elements smaller than M can be secured, the processing efficiency is somewhat lowered, but the processing is continued without any trouble by using the work array w having the secured number of elements. It

Although the working array is dynamically secured at the time of execution in the above embodiment, it may be statically secured in advance.

FIG. 6 is a block diagram of a compiler to which such a vectorization method is applied. The same reference numerals as those in FIG. 1 denote the same parts, 6'is a loop structure transforming means, and 7'is a target program generating means. .. The embodiment shown in FIG. 6 is different from the embodiment shown in FIG. 1 in the points of the loop structure modifying means 6'and the object program generating means 7 ', and therefore only those points will be described below.

The loop structure transforming means 6 ′ inputs the second intermediate text 11 generated by the vectorizing means 3, the work sequence information table 12 and the loop information table 10 generated by the loop structure analyzing means 3 and loops. The structure has an initial value 0, a final value term-incr, and an increment value L × inc.
r, the outer loop of the index variable i, and the initial value init + i,
Final value min {init + i + (L-1) × incr, t
erm}, the increment value incr, and the inner loop of the exponent variable I, which is transformed into a double loop structure, and the definition and reference of the work array element appearing in the loop structure are (I-init-
i) / incr + 1 Replace with the definition and reference to the 1st element.

Here, L is a constant appropriately determined by the compiler 2, and a value equal to the size of the vector register length is preferably used.

The object program generating means 7'is also the third intermediate text 13 generated by the loop structure transforming means 6 '.
Then, the working array information table 12 generated by the vectorizing means 4 is input, and a target program in which the working array is allocated with the size L is generated.

In a vector computer having a vector register, the number of elements that can be processed by one vector instruction is limited to the size of the vector register. Therefore, it is sufficient to secure the size of the working array by the length of the vector register. Further, since the vector register length is generally about several tens to several hundreds, it is possible to eliminate the overhead of securing an area at the time of execution when the number of elements of the working array is fixedly allocated by the vector register length.

The embodiment of FIG. 6 focuses on such a point. For example, in the loop structure shown in FIG.
7, the initial value of the outer loop is 0, the final value is M−1, the increment value is 10, the index variable is i, the initial value of the inner loop is 1 + i, and the final value is MIN (i + 10,
M), the increment value is 1, and the index variable is transformed into a double loop structure with I. Further, the definition and reference of the work array element w (I) are replaced with the definition and reference of w (Ii). The working array w has a size of 10 according to DIMENSION w (10).
Is statically assigned in.

[0042]

As described above, according to the present invention, the work array having the maximum number of elements that can be secured within the loop length at the time of execution can be secured so that the processing can be continued, or statically secured in advance. Since a target program that can continue processing with a work array with the specified element length L is generated, even if a loop that needs to know the loop length at run time requires a work array, the loop length at run time is There is an effect that it is possible to create a target program that can be continuously executed without any limitation.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a compiler to which an embodiment of a vectorization system of the present invention is applied.

FIG. 2 is a diagram showing an example of a loop structure described in FORTRAN language.

FIG. 3 is a diagram showing an example of another loop structure described in FORTRAN language.

FIG. 4 is a diagram showing an example of a loop structure obtained by dividing the loop structure of FIG. 3 using a work array w.

5 is a diagram showing an example in which the loop structure of FIG. 3 is modified by the vectorization method of the embodiment of FIG.

FIG. 6 is a configuration diagram of a compiler to which another embodiment of the vectorization method of the present invention is applied.

7 is a diagram showing an example in which the loop structure of FIG. 3 is modified by the vectorization method of the embodiment of FIG.

[Explanation of symbols]

 1 ... Primitive program 2 ... Compiler 3 ... Loop structure analyzing means 4 ... Vectorizing means 5 ... Loop structure selecting means 6, 6 '... Loop structure modifying means 7, 7' ... Target program generating means 8 ... Target program 9 ... First Intermediate text 10 ... Loop information table 11 ... Second intermediate text 12 ... Work sequence information text 13 ... Third intermediate text

Claims (2)

[Claims] 1. In a vectorization method of a compiler for inputting a source program written in a high-level language and generating an object program for a vector computer, a loop structure in the source program is recognized, and a first intermediate text for the loop structure. And the initial value init, the final value term, and the increment value inc of the loop structure.
Loop structure analysis means for generating a loop information table storing r, loop length m, and exponential variables, and a first intermediate text of the loop structure obtained by the loop structure analysis means are input, and the loop structure is executed by vector processing. And a vectorization means for generating a work sequence information table storing work sequence information necessary for execution in vector processing, as well as converting to a second intermediate text. Loop structure selecting means for inputting the loop information table and the working array information table generated by the vectorizing means, and selecting a loop structure in which the size of the required working array is not determined at compile time, and the loop structure selecting means For the loop structure selected by means,
The second intermediate text generated by the vectorizing means, the work sequence information table, and the loop information table generated by the loop structure analyzing means are input, and the maximum loop length m or less that can be secured before the loop structure is obtained. An intermediate text for securing a working array of the number of elements and setting the secured number of elements in a variable n is inserted, and a loop structure is set to an initial value 0, a final value term-incr, and an increment value n × i.
Outer loop with index variable i in ncr and initial value init
+ I, final price min {init + i + (n-1) × inc
r, term} and the increment value incr are transformed into a double loop structure consisting of an inner loop of the exponent variable I, and the definition and reference of the I-th element of the working array appearing in the loop are defined as (I-in
it-i) / incr + 1 definition of the first element, loop structure transformation means for generating a third intermediate text replaced with a reference, and the third intermediate text generated by the loop structure transformation means are input to execute the target program. A vectorization method comprising: a target program generating means for generating. 2. In a vectorization method of a compiler for inputting a source program written in a high-level language and generating an object program for a vector computer, a loop structure in the source program is recognized, and a first intermediate text for the loop structure is recognized. And the initial value init, the final value term, and the increment value inc of the loop structure.
Loop structure analysis means for generating a loop information table storing r, loop length m, and exponential variables, and a first intermediate text of the loop structure obtained by the loop structure analysis means are input, and the loop structure is executed by vector processing. And a vectorization means for generating a work sequence information table storing work sequence information necessary for execution in vector processing, as well as converting to a second intermediate text. Loop structure selecting means for inputting the loop information table and the work array information table generated by the vectorizing means, and selecting a loop structure in which the size of the necessary work array is not determined at compile time, and the vectorizing means. Generated by the second intermediate text and the work sequence information table generated by Enter the loop information table, a loop structure, which is selected by the loop structure selection means, the compiler constants suitably determined as L, the initial value 0, the closing price term-
incr, an outer loop having an index variable i with an increment value L × incr, an initial value init + i, and a final value min {init +
i + (L-1) × incr, term}, increment value inc
The definition of the working array element which is transformed into a double loop structure consisting of the inner loop of the exponent variable I by r and appears in the loop,
A loop structure modification means for generating a third intermediate text in which a reference is replaced with a definition to the (I-init-i) / incr + 1th element and a reference, and a third intermediate text generated by the loop structure modification means and the above A vectorization method, comprising: a target program generation means for inputting a work array information table generated by the vectorization means and generating a target program in which a work array is allocated with a size L.