JP3235462B2 - while loop structure conversion device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a wh for converting a while loop which is difficult to calculate the number of repetitions described in C language into a for loop which can easily calculate the number of repetitions.
The present invention relates to an ile loop structure conversion device.

[0002]

2. Description of the Related Art It has been conventionally performed to increase the processing speed by performing parallelization such as vectorization on a loop structure. By the way, when parallelizing a loop, it is necessary to obtain the number of loop repetitions at the entry point of the loop. Conventionally, the number of loop repetitions is obtained as follows.

[0003] For example, the loop is a D in the FORTRAN language.
In the case of an O loop, there are control variables for the loop, and
Since the initial value, the closing value, and the increment value of the control variable are described, the number of repetitions of the loop can be easily obtained by performing the calculation of the number of repetitions = (closing value−initial value + increment value) / increment value. .

In the case of a C language for loop which can be described without using a clear loop control variable, the first expression (initialization expression) and the second expression (repetition condition expression) given to the for loop , The control variable of the loop and its initial value, final value, and increment value are obtained from the three expressions (increment expression), and the same operation as described above is performed based on them to obtain the number of loop iterations. (See, for example,
No. 303499).

[0005]

On the other hand, in a while loop written in C language, there are no expressions corresponding to the first and third expressions of the for loop, and only a repetition condition expression corresponding to the second expression is provided. Existing. Therefore, in the while loop, the control variables and their initial values, closing values, and increment values cannot be determined as in the for loop, and the number of loop iterations cannot be obtained. Therefore, it has been conventionally difficult to perform parallelization such as vectorization on a while loop.

Accordingly, an object of the present invention is to convert a while loop written in C language into a for loop capable of obtaining the number of loop iterations, thereby providing a while loop.
An object of the present invention is to enable parallelization such as vectorization for the e-loop.

[0007]

In order to achieve the above object, the present invention provides a loop structure analysis unit for extracting a while loop from a source file and a while loop extracted by the loop structure analysis unit by repeating an initialization expression. A for loop including a conditional expression, an incremental expression, an update expression for a control variable in the while loop, and a corresponding update expression, wherein the control variable of the while loop is held as the initialization expression immediately before entering the loop. And a final value determined by the repetition condition expression of the while loop as the repetition condition expression, wherein the control variable of the for loop is an initial value of the control variable of the for loop different from the control variable of the while loop. And an expression for determining whether or not the while loop has been reached.
Contains multiple update expressions for control variables
Increases the control variables in each of the multiple update expressions
Or a formula for increasing or decreasing the control variable of the for loop by a total value of the decreasing values,
An update expression corresponding to the control variable in the e-loop and an update expression corresponding to the control variable of the while loop as an induction variable of the control variable of the for loop as an update expression
and a loop structure conversion unit for converting into an or loop.

In the above configuration, the loop structure analysis unit extracts a while loop from the source file, and the loop structure conversion unit converts the while loop into an initialization expression, a repetition condition expression, an increment expression, and a while loop. Is converted into a for loop including an update expression for the control variable and a corresponding update expression.

Further, according to the present invention, even when parallelization such as vectorization cannot be performed on a converted for loop, the processing can be performed at higher speed than when the original while loop is directly executed as a scalar. In order to achieve this, the loop structure conversion unit is configured to incorporate a loop invariant in the while loop immediately before the for loop.

[0010] In the above configuration, the loop structure conversion unit incorporates the in-loop invariant in the while loop immediately before the for loop.

[0011]

Next, an embodiment of the present invention will be described in detail with reference to FIG.

The loop structure analyzer 11 in the loop structure converter 1 reads the source file 2 and searches for a description of a while loop 21. When the description of the while loop 21 is found, the loop structure analysis unit 11 passes the description to the repetition condition determination unit 12 as a conversion target. The description other than the while loop is written to the intermediate file 3 as it is.

The repetition condition determination unit 12 tests whether the repetition condition expression 211 of the while loop 21 includes a comparison identifier, that is, whether the repetition condition expression 211 matches the pattern of the repetition condition expression shown in FIG. The expressions on the left and right sides of the conditional expression are composed of operations between variables and constants, and must not include functions. If the conditional expression 211 is
If the pattern does not match the pattern shown in (1) or contains a function, the while loop 21 is removed from the conversion target and written out to the intermediate file 3 in the original form.

The control variable extraction unit 13 searches for a control variable of the loop from the repetition condition expression 211 which is determined by the repetition condition determination unit 12 to match the pattern of the repetition condition. The control variable must monotonically increase or decrease by a constant value each time the loop is repeated. Conversely, variables other than control variables must always keep a constant value in the loop. Therefore, iterative conditional expression 2
When only one of the variables appearing in the expressions on the left and right sides in FIG. 11 is updated in the loop, it may be a control variable.

The control variable extracting unit 13 checks whether or not each of the variables appearing in the conditional expression 211 is updated in the loop. Only one variable is updated in the loop. The variable is extracted as a control variable only when it is not updated in the loop. If the control variables cannot be extracted, the while loop 21 is removed from the loop to be converted, and written in the intermediate file 3 in its original form.

Next, the increment type determining unit 14 checks the control variable updating expression 212 extracted by the control variable extracting unit 13 and tests whether the updating expression 212 is an incremental type. In order to determine whether the update expression 212 is an incremental expression, the incremental expression determining unit 14 determines whether the update expression 212
Test to see if it matches the incremental pattern as shown in The invariant expression in the loop that appears in the increment expression is composed of operations of variables and constants, but must always have a constant value in the loop. Thus, if a variable exists in an invariant in a loop, this variable must not be updated in the loop. Also, the increment expression must be executed every time the loop is repeated. Therefore, the increment expression must not be an executable statement under the conditional statement. The incremental type determination unit 14 also performs a test under this condition.

One of the features of the present invention is that even if an update expression of a control variable appears at a plurality of places in a while loop, f
Although it is possible to convert to or loop,
In this case, all of the update expressions [1] to [n] for the control variables must satisfy the above conditions. If there is at least one control variable update expression that does not satisfy the above condition, the while loop 21 is excluded from the conversion target and is written out to the intermediate file 3 as it is.

The loop structure conversion unit 15 includes a repetition condition expression 211 that the repetition condition determination unit 12 determines to match the pattern of the repetition condition expression, and a control variable extraction unit 13
Is converted into the for loop 31, the closing price calculation formula 313, and the increment calculation formula 314, based on the control variables extracted by the formula (1) and the update formula 212 determined by the increment formula determination unit 14 to be the increment formula. And writes it to intermediate file 3.

The description of the for loop in C language is FO
The degree of freedom of description is higher than that of the DO loop of the RTRAN language, and various loops can be described other than those in which the loop is repeated depending on the value range of the control variable such as the DO loop. However, in such a loop, the number of iterations of the loop becomes unknown, and optimization such as vectorization and parallelization cannot be performed. So, for loop 31
A loop is formed using control variables in the same manner as the DO loop,
Make the initial value, closing price, and increment value clear.

As a control variable used in the converted for loop 31, a control variable of a for loop different from the original while loop is used. The control variable used in the original while loop is the induction variable of the control variable of the for loop (the variable that increases and decreases in proportion to the increase or decrease of the control variable of the loop at each iteration of the loop, and the primary variable of the control variable It is expressed by a formula).

The first expression 3111, the second expression 3112, and the third expression 3113 given to the for loop are as follows.

The first expression 3111 is an initialization expression for giving an initial value to a control variable of the for loop, and is executed only once at the entrance of the loop.

The second expression 3112 is a for loop repetition conditional expression, which is evaluated each time the loop is repeated, and terminates when the evaluation result becomes false. Here, fo
Gives an expression that evaluates whether the control variable of the r loop has reached the closing price.

The third expression 3113 is an increment expression of a for loop, and increases or decreases a control variable of the for loop by a constant value every time the loop is repeated.

Further, each formula will be described in detail.

The first expression (initialization expression) 3111:

This is an expression for giving an initial value to the control variable of the for loop. The initial value is the value held in the control variable of the while loop immediately before entering the loop in the original while loop. . That is, the first expression 31
11 is as shown in (A) below.

Control variable of for loop = control variable of while loop (A)

The second expression (repetition condition expression) 3112:

The while loop iteration conditional expression 211
The control variable in the middle is moved to the left side of the conditional expression, variables and constants that take a certain value in the other loops are moved to the right side,
It is deformed as shown in (B) below.

A while loop comparison condition invariant in a loop (B)

Here, since the invariant in the loop always takes a constant value inside the loop, it is inefficient if it is calculated every time the loop is repeated. Therefore, when a variable is included in the invariant expression in the loop, the value taken by the invariant expression is set as the closing value, and the closing value calculation formula 313 is shown in the following (C), and this formula is calculated in advance before the for loop. Keep it.

Closing value = invariant expression in loop (C)

As a result, the repetition conditional expression is rewritten as shown in the following (D).

Control variable of while loop Comparison conditioner Close value (D)

Further, the control variable of the while loop replaced with the control variable of the for loop is given by the following equation (31).
It is assumed to be 12. That is, the second expression 3112 is as shown in the following (E).

Control Variable of For Loop Comparison Conditioner Closing Price (E)

The third expression (incremental expression) 3113:

First, the update expression 212 of the control variable in the original while loop is rewritten into an update expression 312 in the for loop using the control variable of the for loop.
Now, it is assumed that there are n update expressions 212 in the original while loop, and the incremental expression determination unit 14 has found that they have the following pattern.

Update expression [1]: control variable of while loop = control variable of while loop + invariant expression in loop [1] Update expression [2]: control variable of while loop = control variable of while loop + invariant expression in loop [2]: Update expression [n]: control variable of while loop = control variable of while loop + invariant expression in loop [n]

These equations can be rewritten as follows using the control variables of the for loop.

Update expression [1]: control variable of while loop = control variable of for loop + invariant expression in loop [1] Update expression [2]: control variable of while loop = control variable of for loop + invariant expression in loop [1] + loop invariant [2]: update formula [n]: while loop control variable = for loop control variable + loop invariant [1] + loop invariant [2] +... +
Invariant in loop [n]

Here, each invariant expression in the loop is an expression composed of invariable variables in the loop other than control variables and operations between constants, and always takes a constant value in the loop. Therefore,
The value taken by the invariable expression in the loop of each incremental expression is defined as an incremental value, and n
Increment value [1] = Invariant expression in loop [1] Increment value [2] = Increment value [1] + Invariant expression in loop [2] Increment value [3] = Increment Value [2] + Invariant expression in loop [3]: Increment value [n] = Increment value [n-1] + Invariant expression in loop [n], and these expressions are calculated in advance before the for loop. deep. By rewriting the update expression using the increment value, the update expression 312 of the control variable of the original while loop in the for loop can be obtained.

Update equation [1]: control variable of while loop = control variable of for loop + increment value [1] Update equation [2]: control variable of while loop = control variable of for loop + increment value [2] Update Formula [3]: control variable of while loop = control variable of for loop + increment value [3]: update formula [n]: control variable of while loop = control variable of for loop + increment value [n]

The above is a modification of the update expression of the control variable of the original while loop. Next, the increment expression (third expression) 3113 for the control variable of the for loop is as follows.
It is expressed as the following (F) using the increment value [n].

The control variable of the for loop = the control variable of for + the increment value [n] (F)

As described above, while loop 21 is for
Loop 31, closing price calculation formula 313, increment value calculation formula 31
4 is converted.

Next, an embodiment of the present invention will be described.

In FIG. 1, the loop structure analysis unit 11
The description of the while loop is searched from the read source file 2. FIG. 4 is a diagram showing an example of the while loop 4 found by the loop structure analysis unit 11. Note that w
The description other than the file loop is stored in the intermediate file 3 as it is.
Is written out.

The repetition condition determining unit 12 finds the while loop 4 shown in FIG.
It is tested whether or not the repetition conditional expression 41 in FIG. 2 matches the pattern of the repetition conditional expression shown in FIG. The repetition condition expression 41 of this example matches the pattern of FIG.

The control variable extraction unit 13 calculates the repetition condition 4
From among the variables appearing in 1, search for a control variable. In the case of this example, the variables appearing in the repetition conditional expression 41 are i, j, and k. If only one of them is updated in the loop, the variable may be a control variable. Here, the control variable extraction unit 1
3 extracts the variable i as a control variable because the variable i is only updated in the loop and the variables j and k are not updated.

Next, the increment type determination unit 14 tests whether the update expressions 421, 422, and 423 of the variable i are increment types. These update expressions must conform to the incremental expression pattern shown in FIG. The invariant expression in the loop that appears in the update expression is composed of operations of variables and constants, but must always have a constant value in the loop. Update expressions 421, 422, appearing at three places in the loop
In 423, there are two variables m and n other than i, but these two variables m and n are not updated inside the loop and always take a constant value. Therefore, all three update expressions 421, 422, and 423 match the pattern of the incremental expression in FIG. 3, and are the incremental expressions.

The loop structure conversion unit 15 is a part of the while structure shown in FIG.
e loop 4 is converted into for loop 5 in FIG.
(3) Conversion into increment formulas 541, 542, 543. In the for loop 5, a control variable for the for loop different from the control variable i of the while loop 4 is used.
Here, it is assumed that the control variable for the for loop is INDX. First formula 511 and second formula 5 given to for loop 5
The twelfth and third expressions 513 are respectively derived as follows.

The first expression (initialization expression) 511:

An expression for initializing the control variable INDX of the for loop, the variable i at the entry of the original while loop
Is given to the control variable INDX. Therefore, the first expression 511 of the for loop is as shown in the following (G).

INDX = i (G)

The second expression (repetition condition expression) 512:

The repetitive conditional expression 41 of the while loop 4 in FIG. 4 is 3 + k> j + i, and the control variable i is shifted to the left side of the repetitive conditional expression.
When variables and constants having constant values in other loops are collected on the right side, i <3 + k−j. Here, 3 + k−
Since j is invariant in the loop, it can be driven out of the loop. Therefore, if the closing price is TERM,
The closing price calculation formula 53 is TERM = 3 + k−j. This equation is calculated immediately before entering the for loop. Further, using this TERM, the repetition conditional expression is set as i <TERM, and the control variable i of the while loop is replaced with the control variable INDX of the for loop, so that INDX <TERM is obtained. This is the second expression 512 of the for loop 5.

The third expression (incremental expression) 513:

First, the update expressions 421, 422, and 423 of the control variables of the original while loop 4 are rewritten using the control variables INDX of the for loop.

In the original while loop 4, in the order of execution, the update equation [1]... I = i + n + 1 (421) The update equation [2]... I = i + m (422) The update equation [3]. There are three update expressions 421, 422, 423 of i + 1... (423). When these are rewritten using the control variable INDX of the for loop, the following is obtained.

Update equation [1] i = INDX + n + 1 Update equation [2] i = INDX + n + 1 + m Update equation [3] i = INDX + n + 1 + m + 1

Here, the invariable expressions in the loop of each incremental expression are summarized as an incremental value. INCR1, IN
Assuming CR2 and INCR3, the increment value [1]... INCR1 = n + 1 (541) The increment value [2]... INCR2 = INCR1 + m (542) The increment value [3]... INCR3 = INCR2 + 1 (543) These equations 541, 542, 543 are for
Calculate just before the loop.

When the update formula is rewritten using these increment values, the update formula [1]... I = INDX + INCR1 (521) The update formula [2]... I = INDX + INCR2 (522) The update formula [3]. INDX + INCR3 (523) is obtained. These update expressions are the update expressions 52 of the control variable i of the original while loop in the for loop 5.
1,522,523.

Next, the third expression of the for loop 5 (incremental expression)
Where the increment value is: INDX = INDX + INCR3 (513) using the increment value INCR3. As described above, the while loop 4 in FIG. 4 is converted into the for loop 5 in FIG.

By providing the intermediate file 3 including the converted loop to an existing vectorizing or parallelizing compiler, the loop can be vectorized or parallelized.

[0067]

As described above, according to the present invention,
The following effects are obtained.

The first effect is that the while loop for which it was difficult to obtain the number of repetitions can be subjected to parallelization such as vectorization. The reason is f
A loop structure conversion unit that converts a while loop having only an expression corresponding to an or-loop repetition conditional expression into a for loop including an initialization expression, a repetition conditional expression, and an increment expression and capable of obtaining the number of repetitions by a conventional technique. Is provided.

A second effect is that even a while loop in which an update expression for updating a control variable is incorporated at a plurality of positions in the loop can be subjected to parallelization such as vectorization. The reason is that the for loop
An update expression for a control variable during a while loop
Is included, each of the above update formulas
Sum of the values that increase or decrease the control variable in
Expression to increase or decrease the control variable of the for loop only
Because it was.

The third effect is that while is temporarily converted to
The le loop contains an executable statement that is not suitable for parallelization such as vectorization in its executable statement. Even if it is converted into a for loop, vectorization or the like cannot be performed because of the above-mentioned executable statement. Even if it ca n’t be done,
This is to be executed at a higher speed than executing the original while loop with slurs. The reason is that the original while
The invariant expression in the loop in the loop is incorporated immediately before the for loop, so that the repetition conditional expression and the increment expression in the for loop become simpler than the original while loop, and for example, Must be executed by a scalar even if the original while
It is faster than executing the e-loop with slurs.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a diagram showing a pattern of a repetition conditional expression.

FIG. 3 is a diagram showing an incremental pattern.

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

FIG. 5 is a diagram illustrating an example of a for loop after conversion.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Loop structure conversion apparatus 11 ... Loop structure analysis part 12 ... Repetition condition judgment part 13 ... Control variable extraction part 14 ... Incremental expression judgment part 15 ... Loop structure conversion part 2 .... Source file 21 ... while loop 211 ... Repetition condition expression 212 ... Update formula 3 ... Intermediate file 31 ... For loop 3111 ... First formula (Initial formula) 3112 ... Second formula (Repeat condition formula) 3113 ... Third formula (Increment formula) 312 ... Update formula 313 ... Final formula 314 ... calculation formula of increment value 4 ... while loop 41 ... repetition condition formula 421,422,423 ... update formula 5 ... for loop 511 ... first formula (initialization formula) 512 ... second formula (repetition condition formula) 513 ... Third formula (incremental formula) 521, 522, 523 ... update formula 53 ... calculation formula of closing price 541, 542, 543 ... calculation formula of closing price

Continuation of the front page (56) References JP-A-7-141328 (JP, A) JP-A-5-303499 (JP, A) JP-A-5-189244 (JP, A) JP-A-4-215132 (JP) A). Allen et al., "Compiling C for Vectorization, Parallelization, and Inline Expansion", ACM SIGPLAN Notices, 1988, Vol. 23, No. 7, pp. 241-249 Proc. Of the IPSJ 45th Annual Conference, late 1991, p. 5-51 to 5-52 (58) Field surveyed (Int. Cl. ⁷ , DB name) G06F 9/45

Claims (3)

(57) [Claims] A loop structure analysis unit for extracting a while loop from a source file; and a while loop extracted by the loop structure analysis unit for an initialization expression, a repetition condition expression, an increment expression, and a control variable in the while loop. A for loop including an update expression and a corresponding update expression, wherein the wh
an expression for setting a value held immediately before the control variable of the ile loop enters the loop to an initial value of the control variable of the for loop different from the control variable of the while loop;
The repetition conditional expression includes an expression for determining whether the control variable of the for loop has reached a final value determined by the repetition conditional expression of the while loop , and the control variable is updated during the while loop as the increment expression.
If multiple expressions are included, each of the multiple update expressions
Sum of the values that increase or decrease the control variable in
And a control variable of the while loop is defined as an induction variable of the control variable of the for loop including an expression for increasing or decreasing the control variable of the for loop as an update formula corresponding to the control variable of the while loop. And a loop structure conversion unit for converting into a for loop including an expression. 2. The method according to claim 1, wherein the loop structure conversion unit includes
2. The while loop structure conversion device according to claim 1, further comprising a configuration in which an invariant in a loop in the e loop is incorporated immediately before the for loop. 3. The whi extracted by the loop structure analysis unit.
a repetition condition determining unit that determines whether or not a comparison condition is included in a repetition condition expression in the le loop; and the repetition condition expression that is determined by the repetition condition determination unit to include the comparison condition is a control variable extracting unit for extracting a while loop control variable; and an incremental expression determining unit for determining whether all update expressions for the while loop control variable extracted by the control variable extracting unit match the pattern of the incremental expression. Wherein the loop structure conversion unit is configured to determine whether all of the update expressions match the incremental expression pattern in the incremental expression determination unit.
3. The while loop structure conversion device according to claim 2, further comprising a configuration for processing only the le loop.