JP2533938B2 - Vector arithmetic processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Outline] In a program translation process, the present invention relates to a processing method of a vector operation expression described in a program, and particularly to a processing method when a vector having a different array shape is included in one expression. Semantic analysis unit that analyzes the input program and finds the vector arithmetic expression when the arithmetic operations include vector arithmetics with different vector shapes so that they can be processed without nesting. The vector arithmetic expression found by the department is expanded by basic vector arithmetic, and an intermediate text is created by inserting a vector start phrase and a vector end phrase at the beginning and end of the vector control range of the expanded vector arithmetic formula, respectively. The LJNDA language conversion unit, which is different from the LJNDA language conversion unit described above, is used in the vector arithmetic expression to be processed. When a vector operation of a vector shape is detected, it is managed as a nest as a separate vector control range, and then the vector control range of each managed nest is sequentially searched from the deepest nest, before the most parent vector operation expression. It has a configuration to move to and delete the nest.

[Industrial applications]

The present invention relates to a processing method of a vector operation expression described in a program in a program translation processing, and more particularly to a processing method when vectors of different array shapes are included in one expression.

[Conventional technology]

Up to FORTRAN 77, vector operations were not allowed in the language specifications. Therefore, it was not necessary to consider the vector operation processing.

In FORTRAN 8X, vector operations are allowed due to the language specifications. The language specification defines that the operands have the same shape in the following cases. However, if the other is a scalar, the shape is adjusted to the vector side.

(I) Operand of built-in operation (ii) Right and left sides of assignment (iii) Argument of element quoting function Here, the shape is determined by the dimension of the vector operand and the number of elements in that dimension. For example, the shape (d1, d2,
…, Dn) is n-dimensional and the number of elements in each dimension is d1, d
Represents a vector that is 2, ... dn.

That is, this language specification requires that, in one expression (including assignment), the shape of each operand that constitutes it is basically the same. It means that.

A conventional vector operation processing method for a program written in FORTRAN 8X having such a language specification will be described below.

This conventional vector operation processing method detects the vector operation part in the semantic analysis part of the program translation (compilation) process, defines the range in the LINDA language conversion part, and expands the expression into the element operation form. Create a message and pass it to the optimization processing unit as interface information. This LINDA
The VCR represents the control range of vector operations with the same shape, and the start is indicated by the vector start word OVCR and the end is indicated by the vector end word CVCR. In the optimization processing section, LIN
Vector control range VC sandwiched between OVCR and CVCR in DA
R is taken out and expanded by looping.

 FIG. 9 shows a specific processing example according to this conventional method.

In FIG. 9, 1 is a source program, 2 is an intermediate text (LINDA language), 3 is a vector control book (CVR book), 4 is a semantic analysis unit, 5 is a LINDA language conversion unit, and 6 is an optimization processing unit. .

Source program 1 shows an example in which a vector operation expression that can be described in FORTRAN 8X is described. Here, "REAL * 4" in the declaration section indicates that the array element is a real number of 4 bytes, and "A (100)" indicates that the array A has 100 array elements. Similarly, arrays B, C, and X each have 100 array elements. Further, the vector operation expression X = A + B * C (1) is an operation expression for each array A, B, C, X declared in the declaration section. In this embodiment, the loop expansion is performed according to the procedure described later.

Intermediate text (LINDA language) 2 is source program 1
The vector operation expression (1) described therein is rewritten and expanded at a vector operation level as shown in the figure. Here, "OVCR" represents the vector start expression of the vector operation expression, and "CVCR" represents the vector end expression of the vector operation expression. The vector operation sandwiched between these OVCR and CVCR is the vector operation expression "X = A + B + * C" described in the source program 1 developed at the vector operation level. These Vt1 and Vt2 are intermediate variables.

The vector control book (VCR book) 3 stores pointers that point to the start and end of vector control of the vector operation expression in the intermediate text 2 generated by rewriting the vector operation expression.

The semantic analysis unit 4 analyzes the input expression and extracts a vector operation part.

The LINDA word conversion unit 5 rewrites the vector operation part extracted from the expression into a vector, and inserts a vector start word (OVCR word) and a vector end word (CVCR word) at the beginning and end of this rewritten vector operation. The generated intermediate text (LINDA language) 2 is generated.

The optimization processing unit 6 creates a vector control book (VCR book) 3 in which pointers pointing to the vector start words and the vector end words of the generated intermediate text 2 are stored, and the rewritten vector operation remains unchanged. It takes out what can be taken out and optimizes it at the vector operation level, then expands the loop.

[Problems to be Solved by the Invention]

By the way, there may be an exception (when the shape is different in the formula) as shown in FIG. 10 and FIG. 11 in the vector operation.

The exception shown in Fig. 10 is when the user function IFUN is included in the arithmetic expression. At this time, the result of IFUN is a scalar, but this argument (ii) is a vector having the shape of [5,5,5], and looking at the assignment statement (i), the shape is [10,10]. .

The exception shown in Fig. 11 is when there is an array constructor. The array constructor is always a one-dimensional array in the language specification, but it is supposed that it may appear in different shapes.

The following problems occur in the processing of the program including the exception as shown in FIGS. 10 and 11.

1) In the above-described conventional vector operation processing method, the start and end of vector operation are managed by the VCR book. Where one vector control range VCR
Within, the condition is that the shapes are the same (because the shape determines the number of loop expansions of vector operations). Therefore, Figs. 10 and 11
The exceptions shown in the figure cannot be described in one VCR.

2) Therefore, from the flow of formula processing, it is natural to nest vector operations of different shapes with VCR, as shown in Fig. 12, but for the following reasons, VCR is nested. To have a problem.

The meaning of the program may change (for example, in the case of counting passages for each loop).

Even if the meaning of the program does not change, there is a performance problem. In other words, if you nest VCRs, the number of loops will be multiplied by each, so it will take time to process.

It is an object of the present invention to enable processing when vector operations having different vector shapes are included in the vector operations without nesting.

[Means for solving the problem]

In order to solve the above-mentioned problems in the conventional method, the present invention uses a vector control range VCR within one vector control range VCR when performing the process of designating the start and end of the vector control range VCR in the LINDA language conversion unit of the program translation process. When a vector operation having a different vector shape starts, that is, when a nest appears, it is moved to the beginning of expression processing to eliminate the nest.

FIG. 1 is a block diagram showing the principle of the present invention, 4 is a semantic analysis unit, and 5 is a LINDA speech conversion unit, which correspond to the conventional system shown in FIG. Reference numeral 7 is a VCR nest management book, which is an extension of the vector control book (VCR book) in FIG. 9, and is created and chained for each VCR of the nest.

The function of the configuration will be described according to the processing flow shown in the semantic analysis unit 4 and the LINDA language conversion unit 5 in FIG.

Then, it is determined whether or not the expression includes a vector operation, and the processing is divided into and.

Is the processing when the vector operation is not included, and outputs a word for the expression processing as in a normal expression.

Is a process including a vector operation, creates a VCR nest management book, and performs LINDA language conversion processing. The structure of the VCR nest management book is shown as 7-1 and 7-2 in FIGS. 2 to 5, and the details will be described later in the embodiment.

Then, the vector start word OVCR indicating the start of the vector operation of the parent expression is output.

Now, the shape change of the vector is detected, and the process is divided into and.

Then, because there is a change in shape, the vector start word OVCR indicating the start of the vector control range VCR that becomes a nest is output,
Then run

 Then, the words for the expression processing are output.

Then, when the end of the operation of the same shape is detected, the process is executed if it is completed, and the process returns to if not completed.

Then, the vector end word CVCR is output when the vector control range VCR of the same shape ends. Then run

Then, it is judged whether it is a CVCR of a nested VCR or a CVCR of a parent expression. In the former case, is executed, and in the latter case, it ends.

Then, the process of moving from the deepest nesting VCR to the front of the parent expression in order is performed, the nesting is resolved, and the process returns.

In this way, create the LINDA word of the intermediate text,
It is passed to an optimization processing unit (not shown), and is expanded into a loop processing format for each vector control range VCR.

[Action]

According to the vector operation processing method of the present invention shown in FIG. 1, when one vector operation expression (referred to as a parent expression) includes vectors of different shapes, that part is outside the parent expression. Since it is moved to, the nest is not created and efficient processing is possible.

For example, in the vector arithmetic expression A = B + F ₁ (C + F ₂ (D + E)), when the vector B and the vector C and the vector C and the vectors D and E have different shapes, the arithmetic expression F ₂ (D + E), F ₁ (C + F ₂ ) is taken out in order and arranged before the parent expression A = B + F ₁ .

F ₂ (D + E) F ₁ (C + F ₂ ) A = B + F ₁ As a result, nest processing of operations having vectors of different shapes is eliminated, and the optimization processing unit may perform normal loop expansion for each operation. .

In the processing of moving the operation parts of different vector shapes included as one operation expression nest structure by the LINDA language conversion of the present invention to the outside of the parent expression, the operation parts of each nest are treated as an independent vector control range VCR. Respectively V
This can be done reliably by performing nest management and controlling the processing process with the CR nest management book.

Also, moving the nested VCR outside the parent expression in this way does not change the meaning of the program. This is ensured by the VCR result value of the inner nest being invariant to the outer VCR being always closed from the inner side. Furthermore, since this movement processing follows the flow of expression processing, the overhead of program translation time can be reduced.

Example Next, an example of the present invention will be described.

2 to 5 show the LINDA speech conversion processing process according to the present invention in a concrete example. In these figures, 7-1 and 7-2 are VCR nest management books, and the programs to be processed are those shown in FIG.

The LINDA word is a CONT word that indicates a break, VSHAPE that indicates a shape
The word, the OVCR word indicating the beginning of the VCR, the VLOAD word indicating the expression processing of the vector operation, and the CVCR word indicating the end of the VCR are included.

The VCR nest management book includes CONT words, OVCR words, and CVCR words in LINDA.
CONT Voting, OVCR Voting, CVCR Voting, and Parent Chained Voting, which commands the parent or nest VCR Nest Management Book, and VCR of the Inner Nest It has a child chain instruction to instruct the nest management book.

Figure 2 shows the VCR nest management book 7-1 and the LINDA language when the analysis processing of the LINDA language conversion is performed from the left to B of the vector arithmetic expression A = B + IFUN (C + D) ....
The shape is [10,10]. CVCR words and parent and child V
Since there is no CR nest management book, VCR nest management book 7
The corresponding entry of -1 is blank.

Fig. 3 shows that the parsing of the argument of the function IFUN of the expression started and the shape changed from [10,10] to [5,5,5].
The VCR is started and the child VCR nest management book 7-2 is created. And VCR Nest Management Book 7-2 and VCR Nest Management Book 7
Between -1 and -1, a chain consisting of a parent chain and a child chain is set up.

Figure 4 shows the state of the VCR nest management book and LINDA language immediately after the analysis of the function IFUN of the expression is completed and the CVCR language is output.

FIG. 5 shows the state of the VCR nest management book and the LINDA language after the analysis of the function IFUN of the expression is completed and the VCR is moved.

CONT managed by VCR Nest Management Book 7-2
The words to the CVCR words are moved before the CONT words managed by the VCR nest management book 7-1. As a result, the VCR nest management book 7-2 is released.

FIG. 6 is an example flow of LINDA language conversion processing. Next, each procedure will be described.

It is determined whether or not the expression to be processed now includes a vector operation. The preceding semantic analysis part (phase)
It can be judged by the information from.

Create a VCR nest management book. Details of this processing will be described later with reference to FIG.

 Find the vector operation part.

 Outputs the vector start word (OVCR word).

If the expression does not contain a vector operation, output the corresponding expression processing statement.

′ If the expression includes a vector operation, output the corresponding word of the expression processing. This process is the same as.

 It is determined whether or not the expression processing is completed.

 Performs output processing of vector end word (CVCR word).

 Details of this processing will be described later with reference to FIG.

The process of creating the VCR nest management book will be described with reference to FIG.

If the shape change of the expression is checked and the shape change is detected, the VCR nest management book for the expression (part) currently being processed is acquired. The obtained VCR nest management list is pointed from the base point, but if the parent VCR nest management list exists as shown in Figure 3, the innermost (uppermost) It is ordered from the child chain instruction of the VCR nest management book.

Finally, the CONT message is output, and C of the acquired VCR nest management book is acquired.
Let the ONT say your vote.

Next, the vector end word (CVCR word) output processing will be described with reference to FIG.

First, when the equation processing is completed (Fig. 6), the CVCR word is output. The output CVCR words are ordered from the VCR nest management book.

Here, if the VCR nest management book for the expression currently being processed is the management book of the base point, the OVCR statement register and the CVCR statement catalog are cleared from that, while the VCR nest management book is not the base material. In that case, the words from the CONT word vote to the CVCR word vote are moved to the base point or immediately before the CONT word that is pointed to by the COND word vote in the VCR nest management book of the immediately upper level (the fifth point).
See the example in the figure). Then, clear the CONT words, OVCR words, and CVCR words from the original VCR nest management list.

〔The invention's effect〕

According to the present invention, the corresponding processing in the case where a vector operation expression includes a function reference or an array constructor having different vector shapes can be realized by partially improving the LINDA language conversion unit (phase) of the compiler. it can. Further, in this case, other phases such as the optimization processing unit (phase) are not affected, which is convenient when the system is expanded in the future.

[Brief description of drawings]

FIG. 1 is a block diagram of the principle of the present invention, FIGS. 2 to 5 are explanatory diagrams of specific examples (No. 1 to No. 4) of the LINDA speech conversion processing process according to the present invention, and FIG. 6 is a LINDA speech conversion processing. Fig. 7 is a flow chart of an embodiment of processing, Fig. 7 is a VC in Fig. 6.
Fig. 8 is a flow chart of the R nest management book creation process, Fig. 8 is a flow diagram of the vector end word (CVCR word) output process in Fig. 6, and Fig. 9 is an explanatory diagram of a processing example by the conventional vector operation processing method. Figure 10 is an illustration of exceptions with differently shaped vectors, Figure 11 is an illustration of exceptions with different array constructors, and Figure 12 is an illustration of the LINDA language output when VCRs are nested. In Fig. 1: 4: Semantic analysis part 5: LINDA language conversion part 7: VCR nest management book

Claims (1)

(57) [Claims] 1. A semantic analysis unit (4) for analyzing a inputted program to find a vector arithmetic expression in a vector arithmetic processing method for expanding a vector arithmetic during translation processing of a program, and a semantic analysis unit (4). The vector operation expression found by is expanded by basic vector operation, and the vector start expression (OVCR expression) and vector end expression (CVCR expression) are added at the beginning and end of the vector control range (VCR) of the expanded vector operation expression. ) Intermediate text with and (in LINDA)
And a LINDA word conversion unit (5) for creating a vector control range (VCR) that sequentially detects different vector control ranges (VCR) when vector operations of different vector shapes are detected in the vector operation expression to be processed. ) As a nest, and then move the vector control range (VCR) of each managed nest sequentially from the deepest of the nest to the front of the parent vector operation expression, and perform processing to eliminate the nest. Vector processing method characterized by.