JPH0594470A - Vectrization system - Google Patents

Abstract

PURPOSE:To prepare an objective program which can be executed at a high speed on a vector computer by vectrizing a loop including a recurrence formula in a specific form. CONSTITUTION:As for the vector computer having a vector command of the recurrence formula in an x(i)=f (x(i-1)) form, in a compiler 42 which inputs an original program 41 described in a high-level language. and prepares an objective program 48, a loop structure selecting means 44 selects the loop including the substitution sentence in the x(i)=f (x(i-n)) form while (n) is the multiple of an increment(incr), among the loop structures of an intial value(init), terminal value(term) and increment(incr). A loop structure transforming means 45 transforms the selected loop structure into a dual loop structure constituted of an ouside loop of an initial value (0), terminal value(n/incr), increment (1), and loop control variable(i), and an inside loop of an initial value (init + (i)), terminal value(term) and increment(n). A vectrizing means 46 vectrizes the inside loop of the transformed dual loop.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a compiler for inputting a source program written in a high-level language to generate an object program for a vector computer, and more particularly to a vector instruction sequence for executing a loop structure at high speed. Vectorization method.

[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. Therefore, in a compiler that generates a target program for a vector computer by translating a source program written in a high-level language such as FORTRAN language, an instruction sequence that analyzes the source program and uses vector instructions as much as possible. To generate. This is called vectorization. Vectorization is usually performed on the loop structure.

In general, whether or not a loop structure can be vectorized depends on whether or not the definition reference relation of an array appearing in the loop structure is preserved even when vectorized. It is known that the definition reference relationship is suitable for vectorization if any of the following four conditions is satisfied.

The array is not defined anywhere in the loop. All definition references to an element of the array are closed in the same iteration of the loop. When an element of the array is defined or defined and referenced in different iterations of the loop, the order of appearance of the definition or reference in the source program and the order of appearance in the iteration of the loop are the same. There is. When an element of the array is defined in a loop, there is no overlap with elements of the same array that are defined or referenced elsewhere in the loop of the defined array.

A definition reference relationship of array elements that does not satisfy the condition is called a backward dependency relationship.

Therefore, conventionally, it has been determined whether or not the loop structure can be vectorized in light of the above four conditions, and the loop structure that is determined to be vectorizable is vectorized.

[0007]

Therefore, for example, in the loop described in the FORTRAN language shown in FIG. 2, since the arrays A and B are also defined, the condition is not satisfied. The array element B (n) is defined in the assignment statement a of FIG. 2 at the nth iteration and is referenced in the assignment statement b of FIG. 2 at the n-10th iteration, so the condition is not satisfied. The array element B (n) is referred to by the assignment statement a in FIG. 2 at the nth iteration, and is defined by the assignment statement b in FIG. 2 at the n-10th iteration. The appearance order in and the appearance order in the iteration of the loop are opposite, and the condition is not satisfied. The array B has 11th to 110th defined elements, and the 1st to 100th referenced elements. Therefore, 1
The condition is not satisfied because the 1st to 100th elements are defined and referenced. Since all of the conditions are not satisfied, the conventional technique has a problem that the loop shown in FIG. 2 cannot be vectorized.

Further, for example, the FORTRA shown in FIG.
Even in a loop written in the N language, the i-th element of the array A is defined in the i-th iteration and referenced in the (i + 1) -th iteration in the assignment statement c in FIG. In principle, vectorization is not possible because none of them are satisfied.

However, a recurrence formula in which the value of each element of the array depends on the value of the immediately preceding element of the array as in the example of FIG. Many of them are equipped with special instructions that allow vector processing of recurrence formulas. Therefore, a compiler that generates an object program for a vector computer equipped with such recurrence type vector instructions also recognizes the recurrence type pattern in the loop and vectorizes the loop as shown in FIG. ..

However, in the loop described in the FORTRAN language including the assignment statement d shown in FIG. 6, each element of the array A is not the previous element, but three elements before the array A as shown in FIG. It depends on the value of the element. Therefore,
Since it does not apply to the recurrence pattern, the conventional technique has a problem that the loop shown in FIG. 6 cannot be vectorized.

Therefore, an object of the present invention is to provide a vectorization system capable of vectorizing loops as shown in FIGS. 2 and 6, which could not be vectorized by the prior art.

That is, for example, the loop of FIG.
When it is divided into 0, 11 to 20, ..., 91 to 100 iterations, there is no overlap between the defined element and the referenced element of the array B between the respective divided iterations, so they are separated. Can be vectorized to. The vectorization method of the present invention makes it possible to generate a target program with higher execution performance by automatically performing such a modification and vectorizing a loop that could not be vectorized by the conventional technology. To do.

Further, for example, the processing actually performed by the loop structure shown in FIG.
It is equivalent to the recurrence formula operation for three sets of element sequences of 7, 10, 2, 5, 8, 8, 11, 3, 6, 9, and 12, and these can be vectorized. The vectorization method of the present invention makes it possible to generate a target program with higher execution performance by automatically performing such a modification and vectorizing a loop that could not be vectorized by the conventional technology. To do.

[0014]

The vectorization method of the present invention that enables vectorization of a loop as shown in FIG. 2 inputs a source program written in a high-level language and generates an object program for a vector computer. In the compiler, the loop structure in the source program is recognized, the first intermediate text for the loop structure is generated, and the initial value init, the final value term, and the increment value inc of the loop are generated.
The loop structure analysis means for generating a loop information table storing the information of r, the first intermediate text of the loop structure obtained by the loop structure analysis means and the loop information table are input, and A definition / reference relationship analysis means for analyzing the definition / reference relationship and generating a definition reference table storing information on the definition / reference relationship of the array,
When the definition reference table obtained by the definition reference relationship analysis means is input and there is an array having a backward dependency relationship in the loop structure, the number of repetition times in which each element of the array is referenced in the loop structure and A loop structure selecting unit that obtains a difference n between defined iterations, selects a loop structure in which the value of the difference n is constant in the loop structure, and generates a loop deformation information table storing n, and the loop. For the loop structure selected by the structure selecting means, the first intermediate text and the loop information table generated by the loop structure analyzing means, and the loop deformation information table generated by the loop structure selecting means are input to input the loop structure. Is an initial value init, a final value term, and an increment value incr ×
i of the outer loop with the index variable i of n and the original loop structure
To a min (i + (n−1) × incr, term) inner loop that executes a loop, and a loop structure deforming means for generating a second intermediate text deformed into a double loop structure, and the loop structure deforming means. Second generated
A vectorizing means for inputting the intermediate text and transforming the inner loop structure of the double loop into a third intermediate text for vector processing, and a third intermediate text generated by the vectorizing means for inputting a series of vector instructions and And an instruction generating means for generating a scalar instruction sequence.

The vectorization method of the present invention that enables vectorization of a loop as shown in FIG. 6 inputs a source program written in a high-level language, and x (i) = f (x (i-
In a compiler for generating a target program for a vector computer having a vector instruction for processing a recurrence formula of the form 1)), a loop structure in a source program is recognized, and a first intermediate text for the loop structure is generated. Initial value init, final value term of the loop,
The loop structure analysis means for generating the loop information table storing the information of the increment value incr, the first intermediate text of the loop structure obtained by the loop structure analysis means and the loop information table are input, and x (i) = f (x (i-
n)), and n is the increment value incr
And a loop structure selecting means for generating a loop deformation information table storing n and a loop structure selected by the loop structure selecting means are generated by the loop structure analyzing means. First
The intermediate text and the loop information table and the loop transformation information table generated by the loop structure selection means are input, and the loop structure is set to an initial value 0 and a final value n / inc.
r, an outer loop having an index variable i with an increment value of 1 and an initial value i
A loop structure modification means for generating a second intermediate text modified into a double loop structure consisting of nit + i, a final value term, and an inner loop with an increment value n, and the second intermediate text generated by the loop structure modification means are input. A vectorization means for transforming the inner loop structure of the double loop into a third intermediate text for vector processing, and a third intermediate text generated by the vectorization means are input to generate a series of vector instructions and a scalar instruction sequence. And an instruction generating means for performing the operation.

[0016]

In the vectorization method of the present invention that enables vectorization of a loop as shown in FIG. 2, when a source program written in a high-level language is input and an object program for a vector computer is generated, a loop structure analysis is performed. The means recognizes a loop structure in the source program, generates a first intermediate text for the loop structure, and generates a loop information table storing information on an initial value init, a final value term, and an increment value incr of the loop. The definition reference relation analysis means inputs the first intermediate text and the loop information table of the loop structure obtained by the loop structure analysis means, analyzes the definition / reference relation of the array appearing in the loop structure, and defines the array. A definition reference table that stores information on reference relationships is generated, and the loop structure selection means obtains the definition reference relationship analysis means. Enter the definition look-up table, which is,
When there is a sequence having a backward dependency in the loop structure, the difference n between the repetition times defined as the repetition times in which each element of the sequence is referred to in the loop structure is calculated, and the difference n
A loop structure whose value is constant in the loop structure is selected, a loop deformation information table storing n is generated, and the loop structure deforming means analyzes the loop structure for the loop structure selected by the loop structure selecting means. The first intermediate text and loop information table generated by the means and the loop deformation information table generated by the loop structure selecting means are input, and the loop structure is set to the initial value in.
index variable i of it, closing price term, and increment value incr × n
From the outer loop with and the original loop structure i to min (i
+ (N−1) × incr, term) to generate a second intermediate text transformed into a double loop structure consisting of an inner loop that executes iterations, and the vectorization means generates the second intermediate text, and the vectorization means generates the second intermediate text. Two intermediate texts are input, the inner loop structure of the double loop is transformed into a third intermediate text for vector processing, and the instruction generation means inputs the third intermediate text generated by the vectorization means and a series of vector instructions. And a scalar instruction sequence are generated.

Further, in the vectorization method of the present invention that enables vectorization of a loop as shown in FIG. 6, a source program written in a high-level language is input and x (i) = f
When generating an object program for a vector computer having a vector instruction for processing a recurrence formula of the form (x (i-1)), the loop structure analysis means recognizes the loop structure in the source program, and the loop structure Generate the first intermediate text for the initial value of the loop init,
A loop information table storing information on the final value term and the increment value incr is generated, and the loop structure selecting means inputs the first intermediate text and the loop information table of the loop structure obtained by the loop structure analyzing means, and x ( i) =
A loop structure information table is created by selecting a loop structure that includes an expression of the form f (x (i−n)) and n is a multiple of the increment value incr, and generates a loop deformation information table storing n. Input the first intermediate text and the loop information table generated by the loop structure analysis means for the loop structure selected by the loop structure selection means, and the loop deformation information table generated by the loop structure selection means, The loop structure has an initial value of 0,
Generate a second intermediate text transformed into a double loop structure consisting of an outer loop having a final value n / incr and an increment value of 1 and an index variable i and an initial loop init + i, a final value term, and an inner loop of an increment value n, and vectorizing The means inputs the second intermediate text generated by the loop structure transforming means, and 2
The inner loop structure of the heavy loop is transformed into a third intermediate text for vector processing, and the instruction generating means inputs the third intermediate text generated by the vectorizing means and generates a series of vector instructions and a scalar instruction sequence.

[0018]

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

Referring to FIG. 1, a compiler 2 to which an embodiment of the vectorization method of the present invention is applied includes a loop structure analyzing means 3, a definition reference relation analyzing means 4, a loop structure selecting means 5, and a loop structure. The transformation program 6, the vectorization device 7, and the command generation device 8 are provided, and the source program 1 written in a high-level language is input to generate a target program 9 for a vector computer.

In FIG. 1, 10, 14, 15,
Reference numerals 11, 12, and 13 respectively represent the first, second, and third intermediate texts, the loop information table, the definition reference table, and the loop modification information table generated by the compiler 2 during the processing.

The function and operation of each part of this embodiment will be described below.

The compiler 2 inputs the source program 1 and generates a target program 9.

At this time, the loop structure analysis means 3 recognizes the loop structure appearing in the source program 1 and generates the first intermediate text 10 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 11 storing the increment value incr is generated.

Next, the definition / reference relationship analysis means 4 inputs the first intermediate text 10 and the loop information table 11 generated by the loop structure analysis means 3, and determines the definition / reference relationships of all sequences appearing in the loop structure. Analysis is performed to obtain information on the position of each array that is defined and referenced in the source program and the number of times that the definition is referenced repeatedly. Definition reference table 1
Store in 2.

Next, the loop structure selecting means 5 inputs the definition reference table 12 generated by the definition / reference relationship analyzing means 4 and checks whether or not there is a definition reference of an array having a backward dependency in the loop structure. If there is a certain array having a backward dependency relationship, the difference between the number of times that each element of the array is defined and the number of times that the element is referenced is taken, and this is n. Then, if n has a constant value in the loop structure, the loop structure is made deformable, n is stored in the loop deformation information table 13, and control is passed to the loop structure deforming means 6.

Next, the loop structure modification means 6 inputs the first intermediate text 10 and the loop information table 11 generated by the loop structure analysis means 3 and the loop modification information table 13 generated by the loop structure selection means 5. , An initial value init, a final value term, an increment value incr × n outer loop, and an initial value i, a final value MIN (i + (n−1) × in
cr, term) and the inner loop of the increment value incr to generate a second intermediate text 14 transformed into a double loop structure. Where i is the index variable of the outer loop, MIN
(N, m) is a function that returns the smaller one of n and m.

For example, in the loop structure shown in FIG. 2, the initial value is 1, the final value is 100, the increment value is 1 and n is 10. Therefore, as shown in FIG. 100, the increment value is 10, the initial value of the inner loop is the index variable L of the outer loop, and the final value is MIN (L + (10-1)
X1,100), that is, MIN (L + 9,100), and an increment value of 1, and is transformed into a double loop structure. The inner loop of the double loop structure can be vectorized because the condition is satisfied.

Next, in the vectorization means 7, the second intermediate text 14 transformed by the loop structure transformation means 6 is input, and the inner loop structure of the double loop is subjected to the third vector processing.
It is transformed into the intermediate text 15.

Finally, the instruction generation means 8 inputs the third intermediate text 15 generated by the vectorization means 7 to generate a series of vector instructions and scalar instructions, and outputs the object program 9.

Referring to FIG. 4, one embodiment of a compiler 42 to which another embodiment of the vectorization method of the present invention is applied is a loop structure analyzing means 43 and a loop structure selecting means 4.
4, loop structure deformation means 45, and vectorization means 46
And a command generation means 47, which inputs a source program 41 written in a high-level language and generates a target program 48 for a vector computer.

In FIG. 4, 49, 52, 53,
Reference numerals 50 and 51 respectively represent the first, second, and third intermediate texts, the loop information table, and the loop modification information table generated by the compiler 42 during the processing.

The vector computer which is the target for generating the target program is x (i) = x (i-1) + y
It is assumed that a recursive vector instruction of the form (i) is provided.

The function and operation of each part of this embodiment will be described below.

The compiler 42 inputs the source program 41 and generates a target program 48.

At this time, the loop structure analysis means 43 recognizes the loop structure appearing in the source program 41, generates the first intermediate text 49 for the loop structure, and also sets the initial value init and the final value ter of the loop structure.
Loop information table 50 storing m and increment value incr
To generate.

For example, in the loop structure of FIG. 6, a loop information table 50 having an initial value 3, a final value 12, and an increment value 1
Is generated.

Next, the loop structure selecting means 44 inputs the first intermediate text 49 and the loop information table 50 generated by the loop structure analyzing means 43, and x (i) = x (im) in the loop structure. Check if an expression of the form + y (i) appears, and if so, check if m is a multiple of the increment value incr of the loop structure. If m is a multiple of incr, the loop structure can be vectorized by deformation, m is stored in the loop deformation information table 51, and control is passed to the next loop structure deforming unit 45.

For example, in the loop structure of FIG. 6, the assignment statement d conforms to the form x (i) = x (i−n) + y (i), and m = 3 = 3 × 1. Since the increment value is a multiple of 1 and the condition is satisfied, 3 is stored in the loop deformation information table 51.

Next, the loop structure modification means 45 inputs the first intermediate text 49 created by the loop structure analysis means 43 and the loop modification information table 51 created by the loop structure selection means 44, and sets the loop structure as an initial value. The second intermediate text 52 transformed into a double loop structure consisting of an outer loop of the control variable i with 0, the final value m / incr−1, and the increment value 1 and an inner loop with the initial value incr + i, the final value term, and the increment value n. To generate.

For example, since the loop structure shown in FIG. 6 has an initial value 3, a final value 12, an increment value 1, and m = 3, the initial value 0 and the final value 3 / 1-1 = 2 as shown in FIG. , An outer loop of the increment value 1 and the control variable L, and an inner loop having an initial value of 1 + L, a final value of 12, and an increment value of 3 are transformed into a double loop structure.

Next, the vectorizing means 46 inputs the second intermediate text 52 transformed by the loop structure transforming means 45 and transforms the inner loop structure of the double loop into the third intermediate text 53 of vector processing.

Finally, the instruction generation means 47 inputs the third intermediate text 53 generated by the vectorization means 46 to generate a series of vector instructions and scalar instructions, and outputs the object program 48.

[0043]

The vectorization method of the present invention described above has the following effects.

Since the loop structure as shown in FIG. 2 in which there is a definition reference dependency that extends over a backward dependency, the target program with a higher execution speed can be generated.

For a vector computer having a vector instruction for processing a recurrence formula of the form x (i) = f (x (i-1)), x (i) = f (in the loop structure of the increment value incr x
(I−n)) can be vectorized even when an expression of a form in which n is a multiple of incr is included, and a target program with higher performance can be generated.

[Brief description of drawings]

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

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

FIG. 3 is an explanatory diagram of a loop structure in which the loop structure of FIG. 2 is modified by the vectorization method of the embodiment of FIG.

FIG. 4 is a configuration diagram of a compiler to which another embodiment of the present invention is applied.

FIG. 5 is a diagram showing an example of a loop structure including a recurrence expression written in the FORTRAN language.

FIG. 6 is x (i) = f described in FORTRAN language.
It is a figure which shows the example of the loop structure containing the assignment statement of the form of (x (i-n)).

7 is an explanatory diagram of a dependency relationship of each element of the array A in the loop structure of FIG.

8 is an explanatory diagram of a loop structure obtained by modifying the loop structure shown in FIG. 6 by the vectorization method of the embodiment of FIG.

[Explanation of symbols]

 1, 41 ... Primitive program 2, 42 ... Compiler 3, 43 ... Loop structure analysis means 4 ... Definition reference relationship analysis means 5, 44 ... Loop structure selection means 6, 45 ... Loop structure transformation means 7, 46 ... Vectorization means 8 , 47 ... Instruction generating means 9, 48 ... Objective program 10, 49 ... First intermediate text 11, 50 ... Loop information table 12 ... Definition reference table 13, 51 ... Loop transformation information table 14, 52 ... Second intermediate text 15, 53 ... third intermediate text

Claims (2)

[Claims] 1. A compiler, which inputs a source program written in a high-level language and generates an object program for a vector computer, recognizes a loop structure in the source program and generates a first intermediate text for the loop structure. A loop structure analysis means for generating a loop information table storing information of an initial value init, a final value term, and an increment value incr of the loop, and a first intermediate text and a loop of the loop structure obtained by the loop structure analysis means. A definition / reference relation analysis means for inputting an information table, analyzing the definition / reference relation of an array appearing in a loop structure, and generating a definition reference table storing information on the definition / reference relation of the array, and the definition / reference relation analysis means Enter the definition reference table obtained in step 1 above and enter the definition reference table When there is a column, a difference n between repetition times defined as repetition times in which each element of the array is referred to in the loop structure is obtained, and a value of the difference n is constant within the loop structure. Loop structure selecting means for selecting and generating a loop deformation information table storing n, and a first intermediate text and loop information table generated by the loop structure analyzing means for the loop structure selected by the loop structure selecting means And the loop deformation information table generated by the loop structure selecting means, and the loop structure is defined as an outer loop having an index variable i having an initial value init, a final value term, and an increment value incr × n, and the original loop structure. I to min (i + (n-1) × incr, t
erm) and an inner loop that executes iteration 2
A loop structure modification means for generating a second intermediate text modified into a double loop structure, and a second intermediate text generated by the loop structure modification means are input to perform a vector process on the inner loop structure of the double loop. A vectorization method characterized by including vectorization means for transforming into intermediate text, and instruction generation means for inputting the third intermediate text generated by the vectorization means to generate a series of vector instructions and a scalar instruction sequence. .. 2. A target program for a vector computer having a vector instruction for inputting a source program written in a high-level language and processing a recurrence formula of the form x (i) = f (x (i-1)). A compiler for generating recognizes a loop structure in a source program, generates a first intermediate text for the loop structure, and stores a loop information table storing information on an initial value init, a final value term, and an increment value incr of the loop. Input the loop structure analysis means to be generated, the first intermediate text of the loop structure obtained by the loop structure analysis means and the loop information table, and x
Contains an expression of the form (i) = f (x (i−n)), and n
Loop structure selecting means for selecting a loop structure in which is a multiple of the increment value incr and generating a loop deformation information table storing n, and the loop structure analysis for the loop structure selected by the loop structure selecting means. The first intermediate text and the loop information table generated by the means and the loop deformation information table generated by the loop structure selecting means are input, and the loop structure is set to an initial value 0, a final value n / incr, and an increment value 1
The outer loop with index variable i and the initial value init + i,
A loop structure modification means for generating a second intermediate text modified into a double loop structure consisting of an inner loop having a final value term and an increment value n, and a second intermediate text generated by the loop structure modification means are input, and 2 Vectorizing means for transforming the inner loop structure of the heavy loop into a third intermediate text for vector processing, and an instruction for inputting the third intermediate text generated by the vectorizing means to generate a series of vector instructions and a scalar instruction sequence A vectorization method comprising: a generation means.