JPH0497484A - Vectorization processing system for compiler - Google Patents

Abstract

PURPOSE:To perform the vectorization of an outside loop by performing loop switching after replacing the designation of the initial value and the end value of a control variable in an inside loop in a duplex loop by a condition statement. CONSTITUTION:It is recognized that the vectorization of the control variable in the outside loop can be performed though that of the control variable in the inside loop is impossible, and a repeat condition conversion part 6 called by a vectorization part 4 inserts codes to find the minimum value of the initial value and the maximum value of the end value of the control variable in the inside loop first. Thence, the designation of the initial value and the end value in the inside loop are replaced by the minimum value and the maximum value, respectively, and also, the code in the loop is rewritten to an intermediate text by storing in an appropriate condition block. After that, a loop switching part 7 called by the vectorization part 4 performs the switching of the loop on the duplex loop, and generates the intermediate text. Finally, the vectorization part 4 performs the vectorization of the inside loop by using a mask vector, and generates the intermediate text. Thereby, the vectorization of the outside loop can be performed.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to vectorization processing of a compiler in order to increase the vectorization rate of a target program and shorten the execution time in a computer system capable of executing vector operations. It is related to the method.

[Conventional technology]

In a computer system that can perform vector operations,
Vector operations are extremely fast compared to scalar operations. In order to maximize the capabilities of such a so-called vector computer, it is an important requirement to increase the proportion of the portion of the program that is executed by vector operations, that is, the vectorization rate. Therefore, it is desirable that the factors that make vectorization impossible are as small as possible.

Due to this background, when loops are nested twice, the control variables of the inner loop (such as the Do variable in the FORTRAN language) cannot be vectorized, but the control variables of the outer loop cannot be vectorized. Something can happen. In this case, a method is adopted in which the inner loop and outer loop are swapped and the swapped inner loop is vectorized.

[Problem to be solved by the invention]

However, if the range of possible values of the inner loop's control variable depends on the repetition of the outer loop, it is generally difficult to determine the initial and final values of the inner loop's control variable after replacement, so the range of possible values is extremely limited. Loop swapping was not possible except in simple cases.

[Means to solve the problem]

As described above, an object of the present invention is to specify the initial value and final value of the control variable of the inner loop for a double loop in which the value range of the control variable of the inner loop depends on the repetition of the outer loop. The purpose is to vectorize the outer loop by replacing it with a conditional statement and then swapping the loop.

In order to achieve the above object, the vectorization processing method of the compiler in the present invention analyzes the definition citation relationships of variables and arrays for each loop existing in the program, and generates a vectorized target program. The above loop runs twice, and the control variables of the inner loop cannot be vectorized due to definition/quotation.
For the control variable of the outer loop, if the vector is north, and the initial value and final value of the control variable of the inner loop depend on the repetition of the outer loop, ■ the minimum value of the initial value of the control variable of the inner loop. (If the control variable of the inner loop decreases, the maximum value) and the maximum value of the final value (However, if the control variable of the inner loop decreases, the minimum value or less is the same)
A means of inserting the code that asks for just before the double loop,
■Set the initial and final values of the control variables of the inner loop to the above minimum values,
A means of replacing it with the maximum value, and a means of inserting a conditional statement to store the entire code inside the inner loop in a conditional block that is executed only when the control variable of the inner loop is within the range of the original initial value and final value. and ■ a means for exchanging the inner loop and the outer loop,
The feature is that the double loop is vectorized by loop swapping.

〔Example〕

This will be explained below with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the figure, 1 is the compiler, 2 is the compiler control unit, 3 is the syntax analysis unit, 4 is the vectorization unit, 5 is the target program generation unit, 6 is the repetition condition conversion unit, 7 is the loop replacement unit, and 8 is the source program , 9 indicate the target programs, respectively.

The syntax analysis unit 3 reads the source program 8, performs syntax analysis, and generates intermediate text.

The vectorization unit 4 determines the possibility of vectorization for each loop in the intermediate text, and for loops that can be vectorized, rewrites that portion into vectorized intermediate text. The target program generation unit 5 generates a target program from the intermediate text.

As an example, consider a case where a loop as shown in FIG. 2 exists in a source program.

The syntax analysis unit 3 generates intermediate text as shown in FIG. The vectorization unit 4 first defines the array A in the loop.
Analyze citation relationships. As a result, it is found that the control variable J of the inner loop cannot be vectorized, but the control variable I of the outer loop is vectorized.

Next, the vectorization unit 4 calls the repetition condition conversion unit 6. The repetition condition conversion unit 6 first inserts a code for finding the minimum initial value and maximum final value of the control variables of the inner loop. Next, the initial value and final value specifications of the inner loop are replaced with the above minimum and maximum values, respectively, and the code within the loop is placed in an appropriate condition block and rewritten into intermediate text as shown in Figure 4.

Next, the vectorization unit 4 calls the loop replacement unit 7. For the double loop in FIG. 4, the loop switching unit 7
Intermediate text as shown in FIG. 5 is generated by performing loop replacement.

Finally, the vectorization unit 4 vectorizes the inner loop shown in FIG. 5 using a mask vector to generate the intermediate text shown in FIG. 6.

The present invention provides a compiler with the above function for the innermost double loop in triple or more loops and the innermost double loop after appropriate loop replacement for triple or more loops. Needless to say, it also includes a vectorization processing method.

〔Effect of the invention〕

As explained above, the vectorization processing method of the present invention vectorizes loops that were not vectorized by conventional compilers, thereby reducing the execution time of the target program.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of source code used to explain the operation of the embodiment of the present invention, FIGS. 3, 4, 5, FIG. 6 is a diagram showing an example of intermediate text used to explain the operation of an embodiment of the present invention. 1 is a compiler, 2 is a control unit, 3 is a syntax analysis unit, 4 is a vectorization unit, 5 is a target program generation unit, 6 is a repetition condition conversion unit, 7 is a loop replacement unit, 8 is a source program, and 9 is a target program It is.

Claims (1)

[Claims] In the compiler's vectorization processing method, which analyzes the definitions and citation relationships of variables and arrays for each loop that exists in a program written in a high-level language and generates a vectorized target program, the loops are nested twice. , due to definitions and citations, the control variables of the inner loop cannot be vectorized, but the control variables of the outer loop can be vectorized, and the initial and final values of the control variables of the inner loop depend on the iterations of the outer loop. , the minimum initial value (however, the maximum value if the control variable of the inner loop decreases) and the maximum value of the final value (however, the minimum value if the control variable of the inner loop decreases) of the control variable of the inner loop. A method for inserting a code that calculates the value (the same value and below) immediately before the double loop, a means for replacing the initial value and final value of the control variable of the inner loop with the above minimum and maximum values, and a means for replacing the control variable of the inner loop with the original initial value. A means for inserting a conditional statement for storing the entire code inside the inner loop in a conditional block that is executed only when the value/end value is within the range, and a means for exchanging the inner loop and the outer loop. A compiler vectorization processing method characterized by vectorizing loops by exchanging them.