JPH09128245A - Compiling processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate an object while compiling time is suppressed to within a range corresponding to the reference situation of arraying data in a loop processing by providing a loop conversion processing part rewriting a description on a multiplex loop in a program into that corresponding to judgment by means of a loop inversion judgment means. SOLUTION: A source program analysis part 10 inputs the given source program 1 and generates an intermediate language program 1a expressed by an intermediate language for compiler based on the program. Then, the loop inversion judgment means in the loop conversion processing part 20 judges whether the circulating directions of inner loops belonging to outer loops are alternately inverted or not is judged at every repetition of the outer loops in the multiplex loops in the intermediate language program 1a. Then, a judgedresult is rewritten. Thus, the object can be generated while compiling time in the loop processing is suppressed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a compilation processing system, and more particularly to a compilation processing system for generating an object having higher execution performance while suppressing an increase in compilation time as much as possible.

[0002]

2. Description of the Related Art Conventionally, "blocking algorithm for controlling data movement in numerical computation programming" has been used as one of the compiling processes for reducing cache contention that occurs when controlling the execution of objects generated by a compiler. Hikaru Samukawa / IPSJ Journal Vo
l.33, No.10, Oct.1992) are known. This compile process converts the loop described in the source program into multiple loops and localizes the array data access range in the inner loop as much as possible to reduce cache contention and improve the execution performance of objects. ing.

[0003]

The compilation process of the prior art described above is a process effective when a special array calculation such as array multiplication or inner product calculation is assumed to be executed in advance. That is, when applying the above-mentioned compilation processing to a source program in which calculation using a general array is described, a large number of algorithm conversion patterns are assumed in advance, and the execution performance by the objects generated based on each pattern is assumed. The problem is that the compile time required to execute the compile process increases in combination because it is necessary to calculate the improvement effect of the above and extract only one algorithm conversion pattern that is expected to have the maximum improvement effect. There was a point.

An object of the present invention is to solve the above problems and compile a source program in which a calculation using a general array is described within a range according to the reference status of array data in loop processing. While saving time
An object of the present invention is to provide a compile processing system capable of generating an object having higher execution performance than before.

[0005]

In order to achieve the above object, a compilation processing system of the present invention analyzes a program in which a desired algorithm is described, and performs overall control of various processing operations necessary for realizing the algorithm. In the compile processing system for generating an object for performing the above, a loop conversion processing unit is provided.

The loop conversion processing section has a loop reverse rotation judging means and a loop reverse rotation rewriting means. Then, the loop reversal determination means alternates the circulation direction of the inner loop belonging to the outer loop for each iteration of the outer loop in the multiple loop having at least a double nest structure described in the program. Determine whether to reverse. Further, the loop reverse rewriting means rewrites the description regarding the multiple loops in the program to that according to the judgment by the loop reverse judging means. Instead of the determination by the loop inversion determination means, it is possible to allow the user to specify whether or not the circulation direction of the inner loop belonging to the outer loop is alternately inverted for each iteration of the outer loop.

Here, "reverse the circulation direction of the loop" means that the original process of the corresponding loop is to increment the value of the loop control variable by the increment value [K] from the initial value [M] to the final value [N]. Up to the initial value [M
+ K × [(N−M) ÷ K]] is changed to the final value [M] to be executed. Also,
"[X]" represents the maximum integer value that does not exceed the real value [x]. For example, initial value [M: 5], final value [N: 2]
5] and the increment value [K: 4], the value of the loop control variable is (5 → 9 → 13 → 17 → 21 → 2) in the original process.
Necessary processing is performed while changing as in 5). On the other hand, if "the loop circulation direction is reversed," the value of the loop control variable is changed to (25 → 21 → 17 → 13 → 9 →
Necessary processing is performed while changing as in 5).

The operation based on the above configuration will be briefly described below.

In the loop conversion processing section that constitutes the compilation processing system of the present invention, the loop inversion determination means is
Regarding the specified multiple loop including the inner loop L, the circulation direction of the inner loop L can be reversed based on the definition reference relation beyond the loop of the variables and array data described in the inner loop L. First, it is determined whether or not there is. When the determination result indicating that the inner loop L is reversible is obtained, the reference situation of the array data in the inner loop L is further analyzed, so that the iteration direction of the outer loop advances once and then the circulation direction of the inner loop L advances. , The number of data stored in the cache can be reused at the last round of the inner loop L executed immediately before, and the round direction of the inner loop L is reversed based on this. The magnitude of the effect obtained by this is determined. That is, when the loop reversal determination means obtains the determination result that the circulation direction of the inner loop L can be reversed and the effect obtained by reversing the circulation direction of the inner loop L is sufficiently large, It is determined that the circulation direction of the inner loop L is alternately reversed for each iteration of the outer loop.

When the user or the like instructs that the circulation direction of the corresponding inner loop L is alternately reversed by inserting a directive statement in the source program or by specifying an option at the time of compilation, the loop reversal determination means is The determination processing by the loop reverse rotation determination means described above is omitted, and only the determination result indicating that the circulation direction of the inner loop L is alternately reversed for each iteration of the outer loop is output.

As described above, in the compilation processing system of the present invention, the calculation using a general array is performed by alternately reversing the circulation direction of the inner loop belonging to the outer loop for each iteration of the outer loop in the multiple loop. Also for the described source program, while suppressing the increase in compile time within the range according to the reference status of array data in loop processing, cache conflicts that frequently occurred in the past are reduced, and objects with high execution performance are generated. can do. In addition, when executing the compile process, by specifying the inversion of the cyclic direction at the time of iterating the inner loop described above, and performing the compile process based on this, an object with high execution performance is generated while further suppressing the increase in compile time. can do.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of a compilation processing system according to the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of a compilation processing system according to the present invention. In the figure, 1 is a source program in which a desired algorithm is specifically described in a programming language, 1a is an intermediate language for a compiler, and an intermediate language program in which an algorithm having the same contents as in the source program 1 is described, 2 is described later. Another intermediate language program generated from the intermediate language program 1a by rewriting is 1b, loop information in the source program 1, 1c is data flow information in the source program 1, 3 is an object generated from the intermediate language program 2, 10 is Source program analysis part, 2
Reference numeral 0 is a loop conversion processing unit, and 30 is an object generation processing unit.

In FIG. 1, the source program analysis unit 1
0 inputs the given source program 1 and generates an intermediate language program 1a expressed in an intermediate language for a compiler based on the input source program 1. Further, the obtained intermediate language program 1a is analyzed, and the nest structure of all the loops described in the source program 1 and the loop information 1b regarding the initial value, the final value and the increment value of the loop control variable in each loop are obtained. , Data flow information 1c relating to the definition reference relationship that crosses the loop of each variable and array data described in each loop. Since many known examples of the processing in the source program analysis unit 10 are known, detailed description thereof will be omitted. Then, the loop conversion processing unit 20
Uses the intermediate language program 1a, loop information 1b, and data flow information 1c obtained as described above as input data to generate an intermediate language program 2 whose loop structure has been converted so as to reduce cache contention as described later. Finally, the object generation processing unit 30 performs the optimization process and the command word generation process using the intermediate language program 2 as input data to generate the object 3.

FIG. 2 is a diagram showing a specific processing flow in the loop conversion processing section 20 in FIG. In the figure, the inside of the broken line 21 indicates the processing of the “loop reverse rotation determination means” and the broken line 22.
The inside corresponds to the processing of the "loop reverse rewriting means". The loop conversion processing unit 20 sequentially executes the processes of the loop inversion determining unit 21 and the loop inversion rewriting unit 22 for each of the multiple loops in the intermediate language program 1a to convert the loop structure so that cache contention is reduced. I do.

The loop reverse rotation judging means 21 is an inner loop L.
For the designated multiple loop including the following, the following [possibility determination A] (Step 2 surrounded by the one-dot chain line in FIG. 2)
10a to 216) and [Effect determination B] (step 219 in FIG. 2) are executed. Then, in [possibility determination A], it is determined that the traveling direction of the inner loop L can be reversed, and in [effect determination B], it is determined that the effect obtained by reversing the traveling direction of the inner loop L is sufficiently large. Only when this is done, it is determined that the circulation direction of the inner loop L is alternately reversed for each iteration of the outer loop.

[Acceptability determination A]: Steps 210a to 21
6 For all variables and array data described in the inner loop L, if there is no definition reference relationship such as flow dependency, output dependency, or inverse dependency that exceeds the frame of loop L, It is determined that the traveling direction can be reversed. If even one of the above-mentioned dependencies exists, it is determined that it is impossible to reverse the circulation direction of the loop L.

That is, first, as initialization of work variables, "unjudged" is set in the reference flags corresponding to all the variables and array data in the loop L in step 210a, and each variable and array is initialized in step 210b. Set the number "0" to the reuse count counter and the non-reuse count counter for the data. Step 211
Is an end determination unit for determining whether or not there is a reference flag corresponding to the variable and the array data described in the loop L with a setting value of “undetermined” [possibility determination A]. If there is no “undetermined” (step 21)
1 = NO), in [Effect determination B] of step 219,
If there is an "undetermined" item (step 211 = Y)
ES) branches the processing from step 212 onward.

In step 212, one of the variables and array data described in the loop L, whose reference flag setting value is "undetermined" (x), is taken out. Then, in step 213, when the iterative execution of the outer loop (Outer_L) described one level higher than the inner loop L proceeds once, whether the variable or array data x extracted in step 212 is reused or not. To judge. If the result of this determination is that it will be reused (step 213 = YES), the flow branches to step 214a to increment (+1) the value of the reuse count counter. If it is determined that the non-reuse counter is not reused (step 213 = NO), the process branches to step 214b to increment (+1) the value of the non-reuse counting counter. After the processing in step 214a or 214b, in step 215, it is determined whether or not the variable or array data x has an interdependence relationship with another variable or array data beyond the frame of the loop L. Then, when it is determined that there is an interdependence relationship that exceeds the frame of the loop L (step 215 = YES), it is impossible to reverse the circulation direction of the loop L, so [effect determination B] of step 219. And processing of loop reverse rewriting means 2
2 is skipped, and the process related to the reverse rotation of loop L is completed. Further, when it is determined that there is no interdependence beyond the frame of the loop L (step 215 = NO), “determined” is set in the reference flag corresponding to the variable or the array data x in step 216. Then, the process returns to step 211 again, and the above [possibility determination A] is repeated for another variable or array data x.

[Effect Judgment B]: Step 219 By analyzing the reference situation of the variable and array data in the inner loop L, the circulation direction of the inner loop L is reversed after the outer loop Outer_L has been executed once. , The number of execution patterns that may reuse the data stored in the cache during the final loop of the inner loop L executed immediately before, and the number of execution patterns that are not likely to be reused. If the number of execution patterns of the former is larger than the number of execution patterns of the latter, the inner loop L
It is determined that the effect obtained by reversing the traveling direction of is sufficiently large. When the number of execution patterns of the former is smaller than the number of execution patterns of the latter, it is determined that there is no effect obtained by reversing the circulation direction of the inner loop L.

In other words, in this embodiment, the data finally stored in the cache is set to the numerical value finally set in the above-mentioned reuse counter by the execution of [possibility determination A] in steps 210a to 216 described above. To the number of execution patterns that may be used, to the number of execution patterns where the numerical value finally set in the above-mentioned non-reuse counting counter is not likely to reuse the data stored in the cache, Equivalent to. Therefore, in step 219, it is determined whether the set value of the reuse count counter is larger than the set value of the non-reuse count counter, and if the set value of the reuse count counter is larger than the set value of the non-reuse count counter. At (step 219 = YES), the process branches to the process 22 of the loop reverse rewriting means. If the set value of the reuse counting counter is smaller than or equal to the set value of the non-reusing counting counter (step 219 = NO), the process 22 of the loop reverse rewriting unit is skipped and the process related to the reverse of the loop L is performed. To finish.

However, due to the insertion of directives in the source program or the specification of options during compilation,
When the user has previously instructed to reverse the circulation direction of the corresponding inner loop L, the loop reversal determination means 21 omits the [approvability determination A] and [effect determination B], and the inner loop is repeated for each iteration of the outer loop. Only the determination result indicating that the circulation direction of the loop L is alternately reversed is output.

The loop reverse rotation rewriting means 22 determines the outer loop Ou according to the determination result of the loop reverse rotation determining means 21.
A new loop (Rev_L) in which the traveling direction of the inner loop L is reversed is created for each of the multiple loops in which the traveling direction of the inner loop L belonging to the outer loop is to be alternately reversed for each iteration of ter_L (step 221). ). Then, the original inner loop L is set as the first inner loop L ₁ , the new loop Rev_L is set as the second inner loop L _2, and the loop control variable j for counting the number of iterations of the outer loop Outer_L (initial value: 1) is newly generated, and the first inner loop L ₁ is executed when the value of j is an odd number, and the second inner loop L ₂ is executed when the value of j is an even number. , The original inner loop L described in the outer loop Outer_L is the first inner loop L ₁ and the second inner loop L ₂
Replace with

FIG. 3 is a diagram showing an example of a program before and after rewriting by the loop conversion processing unit 20 in FIG.
3A shows the original program before rewriting (pre-conversion program), and FIG. 3B shows the program after rewriting (post-conversion program). In addition,
Although the source program is described in FIG. 3 for the sake of explanation, in the actual processing performed in this embodiment, the intermediate program 1a generated from the source program in FIG. Based on
The loop conversion processing unit 20 is assumed to generate the intermediate language program 2 corresponding to the source program of FIG.

In FIG. 3A, the pre-conversion program is composed of multiple loops having a triple nest structure. The loop process (upper loop Outer_L) located at the highest nest is the process within the broken line frame 32 that is repeatedly executed according to the loop control variable K31. The loop process (middle loop L) located in the middle nest is the process within the broken line frame 34 that is repeatedly executed according to the loop control variable J33. The loop process (lower loop Inner_L) located in the lowest nest is the process within the broken line frame 36 that is repeatedly executed according to the loop control variable I35. Further, the loop control variables of all of the above-mentioned upper loop Outer_L, middle loop L, and lower loop Inner_L are initial value: 1 final value: 1000 and increment value: 1.

Among the above-described loop processes in the pre-conversion program, in the lower loop Inner_L, the array data X to be referenced or updated clearly has an interdependence relationship beyond the loop frame. Therefore, as a result of the above [possibility determination A], it is excluded from the target of rewriting by the loop conversion processing unit 20. On the other hand, in the middle loop L, the variable (B) is referred to and updated in addition to the above-mentioned array data, but all of them have only the definition reference relation within the frame of the loop. Therefore, as a result of the above [possibility determination A], it is determined that it is possible to reverse the traveling direction. Furthermore, after the iteration of the upper loop Outer_L (outer loop) advances once, the middle loop L (inner loop)
It is expected that the data stored in the cache memory will be reused at the last loop of the intermediate loop L executed immediately before by reversing the circulating direction of. Therefore, as a result of the above-mentioned [effect determination B], it is determined that the effect obtained by reversing the circulation direction of the middle loop L is sufficiently large, and it is a target of rewriting by the loop conversion processing unit 20. As described above, in the pre-conversion program of FIG. 3A, the part of the middle loop L is the target of rewriting by the loop conversion processing unit 20.

In FIG. 3B, the converted program is composed of multiple loops having a triple nest structure. In the figure, the loop process (upper loop Outer_L) located in the highest nest is the loop control variable K31a.
This is the processing within the broken line frames 32a and 32b that is repeatedly executed according to (initial value: 1, final value: 1000, increment value: 2). The loop processing located in the middle nest is the broken line frame 3
The process in the first middle loop L ₁ executed in 2a and the process in the second middle loop L ₂ executed in the broken line frame 32b. The loop process located in the lowest nest is the first lower loop In executed within the broken line frame 34a.
The process in ner_L _{1 and} the process in the second lower loop Inner_L ₂ executed in the broken line frame 34b.

Here, the processing by the first middle loop L ₁ has the same contents as the processing by the middle loop L shown in FIG. 3A, and the loop control variable J33a (initial value: 1, final value: The processing in the broken line frame 34a is repeatedly executed according to 1000, increment value: 1). The process by the second middle loop L ₂ is a process corresponding to the process by the loop Rev_L in which the circulation direction of the middle loop L shown in FIG. 3A is reversed, and the loop control variable J33b (initial value: 1000 , The final value: 1, and the increment value: -1), the processing in the broken line frame 34b is repeatedly executed. Further, the processing by the first lower loop Inner_L ₁ and the second lower loop Inner_L ₂ are both shown in FIG.
The contents are the same as the processing by the lower loop Inner_L shown in FIG. That is, the processing by the first lower loop Inner_L _{1 is performed} by the loop control variable I35a (initial value: 1, final value: 1000,
The processing in the broken line frame 36a is repeatedly executed according to the increment value: 1), and the processing by the second lower loop Inner_L ₂ is the loop control variable I35b (initial value: 1, final value: 1000, increment value:
According to 1), the processing within the broken line frame 36b is repeatedly executed.

The calculation result obtained by the pre-conversion program shown in FIG. 3A and the calculation result obtained by the post-conversion program shown in FIG. 3B are the same. The difference between the two programs is the processing by the intermediate loop that is repeatedly executed according to the loop control variable J. That is, in the case of the pre-conversion program of FIG. 3A, the values of the loop control variables J and I are sequentially incremented as the middle loop L and the lower loop Inner_L are executed, and when the final values reach 1000 respectively, In the loop Inner_L, the calculation using the following array data group is performed.

[Sequence data group] G (1000) A1 (1000,1000) X (999,1000) A2 (1000,1000) X (1001,1000) A3 (1000,1000) X (1000,1000) A4 ( 1000,1000) X (1000,1000) X (1000,1000) or more.

Subsequently, when the loop control variable K in the upper loop Outer_L is incremented, the values of the loop control variables I and J are set to the initial value: 1 again. Therefore, at the beginning of the loop processing by the lower loop Inner_L,
The calculation using the following array data group is performed.

[Sequence Data Group] G (1) A1 (1,1) X (0,1) A2 (1,1) X (2,1) A3 (1,1) X (1,1) A4 ( 1,1) X (1,1) X (1,1) or more.

In the above example, it is considered that there are very few cases where each value of the array data group that has not been referenced until recently is read out to the cache memory, and in most cases, after a cache miss, the hard disk device It will be newly read from a swap file or the like provided in the auxiliary storage device such as.

On the other hand, in the case of the program after conversion shown in FIG. 3B, the first middle loop in the broken line frame 32a.
With the execution of L ₁ and the first lower loop Inner_L ₁ , the values of the loop control variables J and I are incremented, reaching the final values of 1000, respectively, and the array data is output in the _first lower loop Inner_L ₁ as described above. After the calculation using the group is performed, the processing of the first middle loop L ₁ and the first lower loop Inner_L ₁ in the broken line frame 32a ends, and the broken line frame 3
The processing of the second middle loop L ₂ and the second lower loop Inner_L ₂ in 2b is started. At this time, the value of the loop control variable I of the second lower loop Inner_L ₂ is set to the initial value: 1 as in the pre-conversion program, but the value of the loop control variable J of the second middle loop L ₂ is set. Is set to an initial value: 1000, which is equal to the final value when the processing in the broken line frame 32a ends. Therefore, at the beginning of the loop processing by the second lower loop Inner_L _2, the calculation using the following array data group is performed.

[Sequence data group] G (1000) A1 (1,1000) X (0,1000) A2 (1,1000) X (2,1000) A3 (1,1000) X (1,1000) A4 ( 1,1000) X (1,1000) X (1,1000) or more.

Of the above array data group, G (1000) is used in the operation in the _first lower loop Inner_L ₁ executed immediately before, so it is certain that it is read in the cache memory. is there. Also, in a general two-dimensional array variable, X (0,1000) X (1,1000) X (2,1000): X (998,1000) X (999,1000) X (1000,1000), etc. As described above, a group of array data whose second-dimensional subscripts have the same value are secured in continuous areas on the auxiliary storage device. Therefore, for example, A1 (1,1000) in the sequence data group
Has a high probability of being read on the cache memory together with A1 (1000,1000) used in the _first lower loop Inner_L ₁ executed immediately before. Therefore, when referring to the array data at the beginning of the loop processing by the second lower loop Inner_L ₂ , the possibility of a cache hit in which desired data is obtained from the cache memory is improved.

The processing of the second middle loop L ₂ and the second lower loop Inner_L ₂ in the broken line frame 32b is completed, and the first middle loop L ₁ and the first lower loop in the broken line frame 32a are completed. The same applies when the processing of the loop Inner_L ₁ is started. That is, the value of the loop control variable J is decremented and the value of the loop control variable I is incremented with the execution of the second middle loop L ₂ and the second lower loop Inner_L ₂ , and the final values are up to 1 and 1000, respectively. When it reaches, in the second lower loop Inner_L ₂ , the operation using the following array data group is performed. Then, the second middle loop L ₂ and the second lower loop In in the broken line frame 32b.
The processing of ner_L ₂ ends.

[Sequence Data Group] G (1) A1 (1000,1) X (999,1) A2 (1000,1) X (1001,1) A3 (1000,1) X (1000,1) A4 ( 1000,1) X (1000,1) X (1000,1) or more.

Then, after the loop control variable K in the upper loop Outer_L is incremented twice, the processing of the first middle loop L ₁ and the first lower loop Inner_L ₁ in the broken line frame 32a is started again. When it comes to the first
The values of the loop control variable I of the middle loop L ₁ and the loop control variable J of the first lower loop Inner_L ₁ are both set to an initial value: 1. Therefore, the first lower loop Inner_
At the beginning of the loop processing by L _1, the operation using the above-mentioned array data group is performed. Therefore, the broken line frame 32a
First middle loop L ₁ and first lower loop in
The process of Inner_L ₁ is completed, and the second line in the broken line frame 32b
For the same reason as when the processing of the middle loop L ₂ and the second lower loop Inner_L ₂ is started, the possibility of a cache hit in which desired data is obtained from the cache memory is improved.

According to the above-described embodiment, by alternately reversing the circulation direction of the inner loop belonging to the outer loop for each iteration of the outer loop in the multiple loop,
Even for a source program in which calculation using a general array is described, the cache contention that frequently occurred in the past has been reduced while suppressing the increase in compile time within the range according to the reference status of array data in loop processing. , It is possible to generate an object having high execution performance. In addition, when executing the compile process, by specifying the inversion of the cyclic direction at the time of iterating the inner loop described above, and performing the compile process based on this, an object with high execution performance is generated while further suppressing the increase in compile time. can do.

[0041]

As described in detail above, according to the compile processing system of the present invention, the cyclic direction of the inner loop belonging to the outer loop is alternately reversed for each iteration of the outer loop in the multiple loop, and thus the general Even for source programs in which calculations using various arrays are described, cache contention, which has frequently occurred in the past, is reduced while suppressing increase in compile time within the range depending on the reference status of array data in loop processing. An effect that an object having execution performance can be generated is obtained. In addition, when executing the compile process, by specifying the inversion of the cyclic direction at the time of iterating the inner loop described above, and performing the compile process based on this, an object with high execution performance is generated while further suppressing the increase in compile time. The effect of being able to do is obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an exemplary embodiment of a compilation processing system of the present invention.

FIG. 2 is a diagram showing a specific processing flow in a loop conversion processing unit in FIG.

3 is a diagram showing an example of a program before and after rewriting by a loop conversion processing unit in FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 source program 1a, 2 intermediate language program 1b loop information 1c data flow information 3 object 10 source program analysis unit 20 loop conversion processing unit 21 loop reversal determination unit 22 loop reverse rewriting unit 30 object generation processing unit

Claims (3)

[Claims] 1. A compile processing system that analyzes a program in which a desired algorithm is described and generates an object for performing overall control of various processing operations required to realize the algorithm, and at least a program described in the program. For each of the multiple loops having a double nest structure, loop reversal determination means for determining whether to alternately reverse the circulation direction of the inner loop belonging to the outer loop for each iteration of the outer loop in the multiple loop, A compile processing system characterized in that a loop conversion processing unit having a loop reversal rewriting means for rewriting the description relating to the multiple loops in the program to the one according to the judgment by the loop reversal judging means is provided. 2. For each of the multiple loops described in the program, instead of the determination by the loop reversal determination means, the circulation direction of the inner loop belonging to the outer loop is alternately reversed for each iteration of the outer loop. The compile processing system according to claim 1, wherein whether or not to specify is specified. 3. For each of the multiple loops in which the circulation direction of the inner loop belonging to the outer loop is to be alternately reversed for each iteration of the outer loop, the loop reversal rewriting unit is configured such that the number of cycles of the outer loop is an odd number. And a second inner loop whose traveling direction is opposite to that of the first inner loop, which is executed when the number of cycles of the outer loop is an even number. 3. The compilation processing system according to claim 1, wherein the description regarding the multiple loop is rewritten.