JPH07141328A - Compiling device - Google Patents

Abstract

PURPOSE:To provide a compiling device which can vectorize a loop part consisting of a conditional loop instruction sentence of a form where the repetitive execution frequency is not specified in regard to the compiling processing of a computer program. CONSTITUTION:This compiling device includes a translation processing part 3 which generates an object program 4 based on the analyzing result of an analysis processing part 2. The part 2 is provided with a control variable deciding part 30 which decides a loop control variable among those variables of a conditional expression of a conditional loop instruction sentence, an updated value setting part 31 which decides the updated value of the control variable, an initial value setting part 32 which calculates the initial value of the control variable, a final value setting part 33 which decides the final value of the control variable out of the conditional expression of the conditional loop instruction sentence, an executing frequency setting part 34 which decides the loop executing frequency based on the initial value, the final value and the updated value of the control variable, and a loop reconstructing part 35 which reconstructs such a loop that is repetitively executed by the decided frequency.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention does not explicitly indicate the number of repetitions in the process of translating a source program of a computer into a target program consisting of a vector execution part for executing vector instructions and a scalar execution part for executing scalar instructions. The present invention relates to a compiling device that enables vectorization of a loop part by a conditional loop statement of a format.

[0002]

2. Description of the Related Art As known to those skilled in the art as a common sense technique, a vector processing device executes an operation on data in a vector register and stores the operation result in a vector register. A vector instruction that performs processing such as storing is executed at high speed by arithmetic control of an advanced pipeline system.

When a compiler translates a source program written in a certain programming language into a target program that can be executed by a required computer, the source program is analyzed and efficiently executed by the vector processing device as described above. It is well known to perform a so-called vectorization process in which a part of a process to be performed is identified and translated into an object program that executes a vector instruction to be a vector execution part.

In this case, a portion where vectorization is identified as inappropriate and the so-called scalar instruction is executed by a computer is called a scalar execution portion. The main target of vectorization is the part described as the iterative process for the sequence data in the source program, but the typical description of the iterative process in the source program is DO in the well-known FORTRAN programming language.
There is a so-called DO loop description with a statement.

Regarding the vector execution part, the reason why the high speed performance of the vector processing device cannot be obtained unless the necessary vector data is loaded into the vector register in advance and the vector operation is executed as described above. Therefore, the program part to be vectorized is the number of times of repeated execution (hereinafter referred to as the number of rotations) that determines the vector length of the array data to be processed in the above-mentioned iterative processing.
Must be confirmed before the first execution of the iterative process.

In the case of the above-mentioned DO command statement, as is well known, by the description such as "DO 15 I = 1,10,3" in the example of FIG. 4A, "the program up to label 15 is repeatedly executed. Then, the initial value of the control variable I is set to 1, and the increment value 3 is added to the control variable I for each repeated execution so that the value of the control variable I becomes the final value 10
If it exceeds, the execution ends. "

Therefore, the rotation speed in this case is as follows: rotation speed = (final price−initial value + incremental value) / incremental value = (10-1
It can be confirmed as +3) / 3 = 4.

The initial value, the final value, and the increment value may be variables, and if any one of them is specified by a variable,
The value of the rotation speed cannot be calculated during vectorization processing, but if the contents of the program are described correctly, DO
Since the values of these variables must be determined by the time the command statement is executed, the number of rotations can be determined by the execution before the first execution of the iteration is started.

Therefore, the analysis processing unit of the compiling device develops a DO loop as shown in FIG. 4A as shown in FIG. 4B as a processing flow chart. In the figure, "rotation speed setting" is a part that gives an initial value to a control variable and sets the rotation speed in a variable I'representing the rotation speed. The rotation speed is a constant determined at compile time as described above, or It is a calculation formula for calculation at the time of execution.

Further, the "decision section I" in FIG. 4 (b) is a section for making a decision to pass the execution when the rotation speed is 0 or less, and the "in-loop calculation" is for the in-loop calculation section in (a). Contents, "Judgment part
"II" is a part for performing a control for repeating the in-loop calculation until the value of the variable I'is decremented by 1 every time the in-loop calculation is executed and I'becomes 0 as a result. Note that in certain cases where there is a reference to the variable I in the loop, it is necessary to add the update of I (I = I + 3 in this example) after the in-loop calculation of the flow chart.

The translation processing unit of the compiling apparatus identifies whether the calculation portion in the loop is suitable for vectorization, and when vectorizing, the translation processing unit performs vectorization with the rotation speed as the processing vector length, and the target program. Generates the vector execution part of.

However, unlike the DO statement, which does not explicitly indicate the element that determines the number of revolutions, it specifies the period during which the logical value of the arbitrarily specified expression is true as the period during which the loop is repeatedly executed. There is a statement (which is a conditional loop statement).

The C language while command statement is a typical example of the conditional loop command statement as described above, and is described as in the example of FIG. 5A, and the expression in parentheses following the command word while. The logical value of is evaluated for each execution, and if it is true, the program described in curly braces is repeated, and if the value of the expression is false, execution is terminated. specify.

Therefore, the analysis processing unit analyzes the program of the example of FIG. 5 (a) and develops it as shown in the flow chart of the process of FIG. 5 (b). The expression "i <10" in the above and "i ++" in the loop and executed immediately after the calculation (as is well known, in the C language, i ++
Is a post-increment, i = i +
1 to indicate that the FORTRAN D
This is the same as specifying "DO ... i = 0,9,1" in the O command statement, and the rotation speed is (9-0 + 1) / 1 = 10.

Next, terms used in the following description will be defined. Expression A is the expression in parentheses following the command word while
("I <10" in the above example). The expression B is an expression (“i ++” in the above example) for updating a certain variable described in the expression A and having a role corresponding to the control variable of the DO command statement during loop execution.

However, in general, from the syntax of the while statement,
Expression B is not required to be included in the while statement. Further, even if there is an update like the expression B, the increment / decrement value of the update is not always specified to be a constant value during the loop rotation.

That is, the while statement has the original specifications:
Unlike the DO statement, which is not a type that explicitly defines the number of rotations, the conventional C language or other compiler allows the vectorization described above. Is not subject to vectorization.

It is an object of the present invention to provide a compiling device capable of vectorizing a loop based on a conditional loop instruction statement in which the number of times of repeated execution is not specified as described above.

[0019]

FIG. 1 is a block diagram showing the configuration of the present invention. The figure shows the configuration of the compiling device, and an analysis processing unit 2 for analyzing a source program 1 including a conditional loop statement for repeatedly executing a loop while a logical value of an arbitrary conditional expression is true, and an analysis processing unit 2 The compiling device includes the translation processing unit 3 that generates the object program 4 on the basis of the analysis result and performs vectorization processing for a loop whose execution count is fixed.

The analysis processing unit 2 selects the control variable from the control variable determining unit 30 that determines the control variable of the loop from the variables that form the conditional expression of the conditional loop statement and the control that is repeatedly executed from the instructions. The update value setting unit 31 for obtaining the update value of the control variable determined by the variable determination unit 30, the initial value setting unit 32 for obtaining the initial value of the control variable, and the final value of the control variable from the conditional expressions of the conditional loop statement. Execute to determine the number of loop executions from the final value setting unit 33 to be obtained, the initial value obtained by the initial value setting unit 32, the final price obtained by the final price setting unit 33, and the update value obtained by the update value setting unit 31 Number of times setting unit 34 and execution number setting unit
And a loop reconfiguration unit 35 that reconfigures the loop so that the loop is repeated the number of times of execution determined by 34.

[0021]

With the compiling device of the present invention, it is discriminated whether or not the number of rotations of the repetitive execution program can be fixed before the start of execution of the conditional loop command statement such as the while command statement. Since the same expansion as in the case of the DO statement is performed, it is possible to generate the vectorized target program from the source program of the conditional loop statement by performing the same vectorization processing as before on the expanded program. .

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A compile device for translating a source program written in C language will be described as an example.

Here, terms used in the following description will be explained. The "value definition" for a variable means an operation of setting a value to the variable or updating the value of the variable, as is commonly used by those skilled in the art.

"Regressive calculation" for variables means
As is commonly used by those skilled in the art, it means an operation of performing a certain operation on the value of the variable and setting the operation result as a new value of the variable.

A simple and typical example of a recursive operation is "i =
i + 1 "and so on. Returning to the description of the apparatus, the analysis processing unit 2 in FIG. 1 analyzes the while statement and, if the description of the while statement and the repeated execution program specified from it is grammatically correct, those are described, for example, in FIG. It is expanded as shown in b) and processed by the control variable determination unit 30, the update value setting unit 31, the initial value setting unit 32, the final value setting unit 33, the execution count setting unit 34, and the loop reconfiguration unit 35.

Therefore, the control variable determination unit 30 analyzes the expression A of the while statement and selects one control variable candidate, and then defines the value of the candidate variable only once in the repeated execution program. The value definition is always executed each time it is executed, and it is checked whether the value definition is recursive addition or subtraction, and the above value definition conditions (hereinafter referred to as the condition of formula B) are satisfied. If so, the candidate variable is set as a control variable.

In the C language, the definition of the value may be included in the A expression. That is, for example, the expression A is "++ i <10"
As is well known, each time the logical value of the condition "i <10" of the repeated execution period is evaluated, the value of "i = i + 1" is defined before the evaluation. In the case of such a description, it is necessary to check the condition of the expression B including the definition part of this value.

The above-mentioned processing for determining the control variables is performed by sequentially replacing the candidate variables in the expression A if necessary. If there are a plurality of control variables or if there are a plurality of control variables, the loop of this while statement is vectorized. No, and this analysis is stopped.

When one control variable can be determined, the update value setting unit 31 then updates the control variable determined by the control variable determination unit 30 with the addition / subtraction value of the recursive addition or subtraction described above. Ask as. In addition, below, this value is called an increment value.

Next, the initial value setting unit 32 first sets the value of the control variable in the expression A as the initial value. Also, the closing price setting unit 33
A variable or a value that is an element that analyzes the expression and gives the logical value of the expression A to the control variable is set as the final value.

Regarding the initial value of the control variable and the increment value and the final value represented by the variable, if necessary, the analysis result of the analysis processing unit 2 is checked, and the variable used as the control variable is set in the loop to be processed. The initial value is obtained by finding the value that the user has immediately before entering. If the value is immediately set, the value is acquired, and if the value is not set until execution, the variable remains as it is.

When the expression A includes the expression B as described above, there are pre-increment, pre-decrement, etc., and post-increment, post-decrement as shown in FIG. 2 (a). The initial value is changed accordingly.

In the case where the expression A has a form such that the closing price is not displayed explicitly, as in the case of "i", the omitted closing price is determined according to the language specifications. That is, since the logical value of i is true (= 1) when i is non-zero and false (= 0) when i is 0, “i” is the same as “i ≠ 0” in the case of expression A. , And the closing price is 0.

For the increment value, if the value is not determined, various checks based on the positive / negative of the increment value, which will be described later, cannot be performed. This analysis is judged to be outside, and this analysis is stopped.

An option designated by the user is provided, and when the option is designated, even if the value of the increment value is not determined until the time of execution, it is processed as having a correct sign. You may

Further, the final price setting unit 33 and the update value setting unit 31 are
For the above closing value and increment value, it is checked whether or not there is a value definition in the repetitive execution program, and if there is a value definition in this program, it is not possible to confirm the rotation speed before execution. Since it is possible or simply not possible, this loop is not vectorizable, and this analysis is stopped.

When the initial value, the increment value, and the final value are determined as described above, the execution number setting unit 34 uses them to set the number of rotations of the loop of the repeatedly executed program to the variable representing the number of rotations (see the above-mentioned figure). 4 (b), which corresponds to the step of "rotation number setting") is generated.

The formula for calculating the number of revolutions can be classified as shown in FIG. 3 (a) depending on the conditional expression of the expression A, the addition / subtraction of the expression B, and whether the increment value is positive or negative. If the conditional expression is “≠ (not equal)”, and if the increment value is other than ± 1, it is necessary to consider the possibility that the value of the control variable will be updated by jumping over the final value. In order to simplify the above, in the present embodiment, when the conditional expression is ≠, only "± 1" is allowed as the increment value, and in the case of other increment values, this analysis is stopped.

Further, when the conditional expression includes the magnitude judgment (<,>), the state of the loop can be identified by the combination of the conditions shown in FIG. 3 (b), and the loop may be a permanently rotating loop. In the case of an "erroneous loop", the loop is regarded as non-vectorizable and the analysis is stopped. When the option of the increment value is specified, this judgment is omitted and the loop is considered to be correct.

As is clear, even in the case of an erroneous loop under the condition of FIG. 3 (b), there is a program included in the loop that causes the loop to jump out under some condition,
In some cases, this program cannot be made an error because it is written assuming that the execution will be forcibly stopped from the outside.

As described above, when vectorization is possible, if all three values of the initial value, the increment value, and the final value are fixed, the rotational speed is calculated from them and the value is calculated according to FIG. 3 (a). Generate a statement that assigns to the variable of rotation speed as a constant.

If any of the three values is a variable whose value is not determined until execution, a calculation formula for calculating the rotation speed is generated according to FIG. 3A, and the calculation result of the calculation formula is generated. Generate a statement configured to assign to the variable for the number of revolutions.

The loop reconstructing unit 35 prepends a statement of any of the forms thus generated in the portion of the repetitive execution program as a step of setting the rotation number, so that while
The loop of the imperative sentence is converted into a program expanded into the same configuration as the case of the DO imperative sentence, and the analysis result of the analysis processing unit 2 to be passed to the translation processing unit 3 is set. FIG. 5C is a diagram showing an example in which the program of FIG. 5A is converted by the above processing.

The translation processing unit 3 is a program in which the processing content of the loop calculation portion processes vector data by the normal vectorization processing similar to the case of the DO statement for the expanded program, and is suitable for vectorization. If it is suitable for vectorization, the target program of the vector execution part for processing the data of the vector length based on the rotation number indicated by the expansion program is generated.

FIG. 6 shows a while statement loop by the control variable determination unit 30, the update value setting unit 31, the initial value setting unit 32, the final value setting unit 33, the execution number setting unit 34 and the loop reconstructing unit 35 of the present invention. It is a figure showing an example of the flow of processing of analysis and expansion,
First, in processing step 10, the A expression of the while statement is analyzed,
Decide the variable to be a control variable candidate.

In processing step 11, for each candidate variable, a sentence for updating the variable, that is, a candidate of the expression B, is searched from the repeated execution program, and processing step 12
, And sift them, and the update is performed only once, and the candidates of the formula B, which is the definition of the recursive value, are left.

If there is only one B-expression candidate identified in the processing step 13, the following processing is continued by setting it as the B-expression. If there are no B-expression candidates or there are a plurality of B-expression candidates, this processing is executed. Abort.

When the processing is stopped, the analysis processing unit 2 is notified, for example, and the analysis processing unit 2 excludes the while statement loop from the vectorization target as in the other portions. It is assumed that the information received from the computer is passed to the translation processing unit 3 as it is, and the same applies when the processing is stopped below.

When the formula B is determined, the increment value is determined from the formula in the processing step 14, and it is discriminated whether or not the value of the increment value is actually determined in the processing step 15 (for this purpose, the analysis processing unit 2 is inquired). , If the value is fixed, get that value),
If the value is not determined until the time of execution, the processing is stopped (in this example, the increment value option is not provided).

Next, in the processing step 16, the expression A is analyzed, and if necessary, the value of the variable is obtained by referring to the analysis result of the analysis processing section 2 for the program up to this loop, thereby obtaining the initial value. Get the value and closing price. Note that FIG. 2 is referred to in determining the initial value, and the rule is followed.

In processing step 17, it is searched for the definition of the value in the repetitive execution program for the variable used as the final value and the increment value, and if it is identified in processing step 18 that the definition of the result value exists, Since the number of rotations cannot be confirmed as shown in, the processing is stopped.

If there is no definition of the above value, processing step
In step 19, the correctness of the loop is judged according to the table of FIG. 3 (b), and if the conditional expression is ≠, it is discriminated whether the increment value is other than ± 1, and if either of the conditions is not satisfied, the processing is stopped. If the processing is possible, the rotational speed calculation formula is determined from the table of FIG. 3A in processing step 20 and the program is converted.

The conversion of the program is performed by first processing step 21.
Then, the determination unit I of the expansion program (see FIG. 5B) received from the analysis processing unit 2 is moved to the loop outside the loop. In processing step 22, a program for determining the value of the number of revolutions according to the formula for calculating the number of revolutions and setting it in the variable i'of the number of revolutions is generated and inserted as a number of revolutions setting step before the determination unit I.

Next, in processing step 23, the judgment condition for passing the repetitive execution program of the judgment unit I is set so that the value of the variable i'is 0.
Change to branch below. In the processing step 24, the original unconditional branch step is changed to the step of the judging section II, and the judging section II subtracts 1 from the variable i ′ of the rotation speed, and the result is 0.
If not, branch to the beginning of the calculation in the loop.

By the above conversion, the expansion program (b) of the while statement loop of FIG. 5A is converted as shown in FIG. 5C, and the expansion in the case of the DO loop (FIG. 4B) ), The translation processing unit 3 can perform vectorization processing on this expansion program by the same processing as in the case of the conventional DO loop.

[0056]

As is apparent from the above description, according to the present invention, in the compiling device for performing the vectorization processing, the conditional loop instruction statement of the format which is not included in the conventional vectorization and in which the number of times of repeated execution is not explicitly stated is provided. There is also a remarkable industrial effect that vectorization is also possible for the loop part due to.

[Brief description of drawings]

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

FIG. 2 is a diagram illustrating an initial value when a value of a control variable is defined in expression A.

FIG. 3 is a diagram illustrating a rotational speed calculation formula and the like.

FIG. 4 is a diagram illustrating an example of a DO loop.

FIG. 5 is a diagram illustrating an example of a while loop.

FIG. 6 is a flow chart of processing of the present invention.

[Explanation of symbols]

 1 source program 2 analysis processing unit 3 translation processing unit 4 target program 10 to 24 processing steps 30 control variable determination unit 31 update value setting unit 32 initial value setting unit 33 final value setting unit 34 execution count setting unit 35 loop reconfiguration unit

Claims (1)

[Claims] 1. An analysis processing unit (2) for analyzing a source program (1) including a conditional loop statement for repeatedly executing a loop while a logical value of an arbitrary conditional expression is true; A compiling device comprising a translation processing unit (3) for generating a target program based on an analysis result and performing vectorization processing for a loop whose execution count is fixed, wherein the analysis processing unit (2) The control variable determination unit (30) that determines the control variable of the loop from the variables that make up the conditional expression of the loop statement, and the control variable determination unit that is selected from the commands that are repeatedly executed.
An update value setting part (31) that obtains the update value of the control variable determined in (30), an initial value setting part (32) that obtains the initial value of the control variable, and control from the conditional expression of the conditional loop statement. The closing price setting part (33) to find the closing price of the variable, the initial value found in the initial value setting part (32), the closing price found in the closing price setting part (33), and the updated value found in the updated value setting part (31) From, the execution count setting unit that determines the execution count of the loop (34)
And a loop reconfiguration unit (3) that reconfigures the loop so that the loop is repeated for the number of executions determined by the execution count setting unit (34).
5) A compiling device characterized by having and.