JP2000194566A - System and method for compilation and storage medium stored with compiler - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the overhead of a wait time in program execution and to improve the execution performance of a program by leveling load distribution to respective loops which are loop-parallelized. SOLUTION: A branch pattern analyzing means 6 analyzes the branch pattern of a condition statement in a loop by referring to branch history information obtained by the execution of an object program, and a loop parallelizing means 7 parallelizes loops with optimum load distribution according to the analysis result; and the branch pattern analyzing means 6 analyzes the branch pattern having changed owing to the parallelization again and a loop optimizing means 8 generates optimum codes for the respective loops divided by the parallelization according to the analysis result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a compiling system, a compiling method, and a recording medium storing a compiler for generating an object program for efficiently executing a loop including a conditional branch instruction.

[0002]

2. Description of the Related Art A conventional compiling system is disclosed in, for example, Japanese Patent Application Laid-Open No. 10-240573. This compilation system operates under the source program to be compiled and program control.
A compiling device for compiling the source program as an input, an object program generated by the compilation, a number of repetitions for each execution of each loop process for repeating a predetermined instruction stream obtained by executing the object program a plurality of times, and A branch history file in which a branch history of a conditional branch instruction in the loop processing is recorded.

[0003] In the compiling apparatus, the branch history information collection code generating means generates the loop execution count,
Generates an object program in which code for collecting the number of repetitions for each execution of the loop and the branch history of conditional statements in the loop is embedded, and executes the object program to display the branch history of each conditional statement in the loop as branch history information. Generates the optimal code for loop processing by collecting the collected branch history information into a file, analyzing the branch pattern information by the branch pattern analysis means using the collected branch history information as input, and recompiling the same program based on the analysis result. I am trying to do it.

[0004]

However, in such a conventional compiling device, since no consideration is given to the parallelization in which the load is distributed, the branch pattern obtained from the branch history file is not taken into account. However, there has been a problem that the analysis result of (1) cannot be used for parallelization, and the analysis result of the branch pattern cannot be used for load distribution.
Further, since the branch pattern analysis means could not obtain information on the parallelization and optimization performed on the loop, if the parallelization and optimization such that the branch pattern changes for the loop, There is a problem that the branch pattern cannot be used in the optimization for the loop.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problem, and has an advantage of equalizing the load distribution of each loop subjected to parallel processing.
As a result, it is possible to reduce the overhead of the waiting time at the time of executing the program and to improve the execution performance of the program, and it is possible to generate an optimum code for each of the parallelized loops, thereby realizing further higher speed of the program. It is an object of the present invention to obtain a recording medium on which a compiling system, a compiling method, and a compiler are recorded.

[0006]

According to one aspect of the present invention, there is provided a compiling system for compiling an object program from a source program and executing the object program for each execution of a loop. A branch history information file in which the number of repetitions and the branch history of the conditional statement in the loop are recorded;
A branch provided in the compiling device and for embedding, in the object program, a code for collecting information on the number of repetitions for each execution of a loop and the branch history of a conditional statement in the loop to create the branch history information file. History information collection code generation means, and causes the branch pattern analysis means to analyze a branch pattern for a loop by using the branch history information and the parallelization information from the branch history information file as inputs. Loop parallelizing means for performing the above-mentioned parallelization by performing optimal load distribution on a target loop based on the analysis result of the branch pattern obtained by the analyzing means; The result of the branch pattern analysis means re-analyzing the branch pattern for the loop in which , The loop optimization means, in which so as to generate an optimal code for the loop.

According to a second aspect of the present invention, there is provided a compile method for generating a code for collecting information on a branch history of a conditional statement in a loop by compiling a source program by a branch history information collection code generating means. The branch pattern analysis unit analyzes the branch pattern of the conditional statement in the loop by referring to the branch history information obtained by executing the object program, and, based on the analysis result, the loop parallelization unit The loop is subjected to parallel processing with optimum load distribution, and the branch pattern changed by the parallel processing is analyzed again by the branch pattern analysis unit. According to the analysis result, the loop optimization unit Generate optimal code for each loop divided by parallelization It is obtained by way.

According to a third aspect of the present invention, there is provided a recording medium storing a compiler according to a third aspect of the present invention, wherein a code for collecting information on a branch history of a conditional statement in a loop is embedded in an object program generated by compiling a source program. A branch pattern analysis process for analyzing the branch pattern of the conditional statement in the loop by referring to the information collection code generation process and the branch history information obtained by executing this object program, and executing the loop based on the analysis result. A loop parallelization process for performing a parallelization process with an optimum load distribution for the loop and a branch pattern changed by being further parallelized are analyzed once again, and according to the analysis result, each of the segments divided by the parallelization is analyzed. That allows a computer to perform loop optimization processing that generates optimal code for a particular loop Is a record of the ram.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a compiling system of the present invention. In FIG. 1, reference numeral 1 denotes a source program to be compiled, 2 denotes a compiling device which receives the source program 1 as an input, and 3 denotes a compiling device. The generated object program 4 is a branch history information file that records the number of repetitions for each execution of the loop obtained by executing the object program 3 and the branch history of conditional statements in the loop. Further, in the compiling device 2, the code 5 embeds a code for collecting information on the number of repetitions for each execution of the loop and the branch history of the conditional statement in the loop into the object program 3 in order to create the branch history information file 4. It is a branch history information collection code generation means.

Reference numeral 6 denotes a branch pattern analyzing means for analyzing a branch pattern for a target loop by inputting information from the branch history information file 4 and information from a loop parallelizing means to be described later. Based on the analysis result of the branch pattern obtained by the above, the above-mentioned loop parallelizing means for performing parallelization with optimum load distribution for the target loop, and the loop parallelizing means 8 are parallelized by the loop parallelizing means 7. This is a loop optimizing unit that generates an optimal code for the loop based on the result of re-analyzing the branch pattern by the branch pattern analyzing unit 6 for the loop in which the branch pattern has changed.

Next, the operation will be described in detail with reference to the contents of the source program 1 shown in FIGS. Here, in the parallelization of information, an explanation will be given by taking as an example a multiprocessor with ten processors capable of neglecting the communication overhead caused by data distribution and capable of performing vector operations. The source program 1 in FIG. 2 is a program that substitutes 1.0 for A (I) only when X (I) is not 0.0 for array A having 1000 elements. First, when the object program 3 is generated from the source program 1 of FIG. 2 by the compiling device 2, the code for collecting the branch history information of the conditional statement in the loop is converted into the object program 3 by the branch history information collection code generation means 5. Embed.

The information to be obtained by the code for collecting the branch history information embedded in the object program 3 includes which loop in the source program 1, which execution, and which loop This is information on whether the conditional statement was true or false.
By executing the object program 3 in which the code for collecting the branch history information as described above is embedded, the branch history information file 4 in which the branch history of the conditional statement in the loop is recorded is obtained. When the same source program 1 is recompiled, the branch pattern analysis unit 6 analyzes the branch pattern of the conditional statement in the loop by using the branch history information file 4 as an input, and based on the analysis result, the loop parallelizing unit. 7 performs parallelization with optimum load distribution for the loop. The branch pattern analysis unit 6 re-analyzes the changed branch pattern in the changed loop based on the information of the parallelization performed by the loop parallelization unit 7, and the loop optimization unit 8 performs the analysis based on the analysis result. An object program 3 in which an optimal code has been generated for a given loop is generated.

For example, the loop in the source program 1 of FIG. 2 is executed only once during the execution of the program, the number of repetitions of the loop is 1,000, and all the conditional statements in the loop are true when the variable I is 1 to 500. Yes, and all subsequent iterations are false. By the first compilation and execution of the source program 1, the loop of the source program 1 is executed once and the loop is
A branch history information file 4 is generated which stores information such that the condition statement in the loop is true when the conditional statement is repeated from 1 to 500 and false when the conditional statement is from 501 to 1000.

When the source program 1 is recompiled, the branch pattern analysis unit 6 analyzes the branch pattern by using the branch information file 4 as an input, and the loop parallelizing unit 7 performs parallel analysis based on the analysis result. Parallelization is performed so that the true rates of the conditional statements in each loop divided by the conversion become equal. In the above case, 10
When block distribution is performed such that the 0th iteration is performed on the first processor, 5 of the 10 parallelized loops do not have any conditional statements in the loop that become true, and no processing is performed. The loop is repeated without performing the above, and the load distribution becomes uneven. Therefore, the first iteration selects the first processor, the second iteration selects the second processor, and performs parallelization.

Then, based on the information parallelized by the loop parallelizing means 7, the branch pattern analyzing means 6 re-analyzes the branch pattern for the loop in which the branch pattern has changed, and based on the analysis result. , The loop optimizing means 8 generates a code that is optimal for the target loop. In this case, the iteration of each parallelized loop is 100
The analysis result shows that the first 50 repetitions of the conditional statement in the loop are true, and the remaining 50 repetitions are false. Performance improvement can be expected by performing vector operation rather than scalar operation. Therefore, the object program 3 in which the vector operation instruction with the mask is generated is generated.

As a result, the true ratio in each of the parallelized loops becomes equal, so that parallelization can be performed so that the load distribution becomes uniform, and the overhead of the waiting time during program execution due to the uneven load distribution can be reduced. Can be reduced.
Also, by performing the branch pattern analysis again on each of the parallelized loops, an optimal code can be generated for each of the parallelized loops. It is possible to increase performance.

Next, another application example of the present invention will be described.
With respect to the source program 1 in FIG. 2, the loop in the source program 1 is executed once during execution, the number of loop iterations is 1,000, the conditional statements in the loop are true from the first to fifth times, and the sixth time. Consider a case where false is repeated 10 times from the first to the 100th time. In such a case, when 10 processors are subjected to cyclic distribution such that the first iteration is the first processor and the second iteration is the second processor, 10 parallelized loops are obtained. Of the five, the conditional statement never becomes true in the loop, the loop is repeated without performing any processing, and the load distribution becomes uneven. For this reason 10
Parallelization on the processor becomes less efficient.

However, the branch pattern analyzing means 6 in the present invention analyzes the branch pattern recorded in the branch information file 4, and the loop parallelizing means 7 uses the analysis result to execute the first to 100th iterations. Since the parallelization is performed by the block distribution such as the first processor, the load distribution for the ten processors becomes equal. Further, the branch pattern analysis means 6 analyzes the branch pattern again based on the information parallelized by the loop parallelization means 7, and based on the analysis result, the loop optimization means 8 separates Of the 100 repetitions of this loop, only the first five are true, so that a code for a scalar operation or a compression / decompression instruction is generated. This can reduce the overhead of the waiting time at the time of program execution due to uneven load distribution and the useless overhead generated when using the vector operation with the mask, so that the execution performance of the program can be improved.

[0019]

As described above, according to the present invention, the code for collecting the information on the branch history of the conditional statement in the loop is converted into the object obtained by compiling the source program by the branch history information collection code generating means. It is embedded in a program, refers to the branch history information obtained by executing this object program, analyzes the branch pattern of the conditional statement in the loop by the branch pattern analysis means, and based on the analysis result, the loop parallelization means The parallel processing is performed by performing the optimal load distribution on the loop, and the branch pattern changed by the parallel processing is analyzed again by the branch pattern analysis means. According to the analysis result, the loop optimization means Optimal code is generated for each loop divided by the parallelization. Therefore, the load distribution of each loop subjected to the loop parallel processing is equalized, thereby reducing the overhead of the waiting time at the time of executing the program and improving the execution performance of the program. It is possible to generate an optimal code, thereby obtaining an effect that the program can be further speeded up.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a compilation system according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing contents of a soap program in FIG. 1;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Source program 2 Compiling device 3 Object program 4 Branch history information file 5 Branch history information collection code generation means 6 Branch pattern analysis means 7 Loop parallelization means 8 Loop optimization means

Claims (3)

[Claims] A compiling device for generating an object program from a source program; a branch history information file in which the number of repetitions for each execution of a loop and a branch history of a conditional statement in the loop are recorded by executing the object program; A branch provided in the compiling apparatus and for embedding, in the object program, a code for collecting information on the number of repetitions for each execution of a loop and the branch history of a conditional statement in the loop in order to create the branch history information file. History information collection code generation means; branch pattern analysis means for analyzing a branch pattern for a loop by using the branch history information and parallelization information from the branch history information file as input; branches obtained by the branch pattern analysis means Target rules based on the pattern analysis results Loop parallelizing means for performing the above-mentioned parallelization by performing optimal load distribution on the loop, and the branch pattern analysis means re-performs the branch pattern for a loop whose branch pattern has been changed by the parallelization by the loop parallelizing means. A compilation system comprising: a loop optimizing unit that generates an optimum code for a loop based on a result of the analysis. 2. A code for collecting information on a branch history of a conditional statement in a loop is embedded in an object program generated by compiling a source program by a branch history information collection code generating means, and is obtained by executing the object program. With reference to the obtained branch history information, the branch pattern analysis means analyzes the branch pattern of the conditional statement in the loop, and based on the analysis result, performs optimal load distribution for the loop by the loop parallelizing means. The parallelization process is performed, and the branch pattern changed by the parallelization is analyzed again by the branch pattern analysis unit. According to the analysis result, each of the loops divided by the parallelization is determined by the loop optimization unit. A compilation method characterized by generating optimal code by using the compiler. 3. A branch history information collection code generating process for embedding a code for collecting information on a branch history of a conditional statement in a loop into an object program generated by compiling a source program, and the execution of the object program. Pattern analysis processing for analyzing the branch pattern of a conditional statement in a loop by referring to the obtained branch history information, and parallelizing processing that optimally distributes the load to the loop based on the analysis result. A loop parallelizing process and a loop optimizing process for re-analyzing a branch pattern changed by being further parallelized and generating an optimum code for each of the loops divided by the parallelization according to the analysis result. And a program for causing a computer to perform the following.