JPS63106865A - Vectorizing method for loop including recurrent arithmetic calculation - Google Patents

Abstract

PURPOSE:To shorten an execution time, by performing a vector processing when possibility of vectorization exists by deciding it just before executing a loop, even when it is not decided whether the vectorization is possible or not at time of compilation. CONSTITUTION:A function which finds a vectorization possible condition at the time of detecting the loop including recurrent arithmetic calculation in a conventional vector complier, and the function which generates an object program which includes two kinds of codes, a vectorized code and a non- vectorized code, and selects either code by deciding the vectorization possibility at time of execution, are added. And a vector compiler providing with such vectorization system is stored in a main memory device 1, and the vectorization is executed by a central processor 3. As a result, source programs stored in large capacity memory devices 5 and 6 are converted to object programs, and all of the loops possible to be vectorized out of the loops including the recursive arithmetic calculation are executed with vectorized codes. In such as way, the execution time can be shortened.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a method for vectorizing a loop including recursive operations using a vector compiler.

2. Description of the Related Art Conventional vector compilers vectorize loops that include recursive operations when information as to whether or not they can be vectorized is given in advance by a programmer.

Problems to be Solved by the Invention However, there has been a problem in that loops containing recursive operations cannot be vectorized, for which it cannot be determined whether vectorization is possible until execution time. For example, assume that a DO loop including the following recursive operation is to be compiled.

(1) Do 10 1=L, M, N10
A (I)=A (I+K) (2) Do 10 I=1,10010
A (I+M)=A (I+N) In the example (1), if the signs of N and K are the same, and in the example (2), if M≦N, it is not a recursive operation, so vectorization is possible. However, with conventional vector compilers, these conditions must be known at the time of compilation to vectorize.

Therefore, even if it is not possible to determine whether vectorization is possible at compile time, the present invention determines whether vectorization is possible immediately before execution of the loop and performs vector processing if possible. It is something.

Means for Solving the Problems The technical means of the present invention for solving the above problems is to add a function to a conventional vector compiler to obtain a vectorizable condition when a loop including a recursive operation is detected, and a vector compiler. A function that generates an objective program that contains two types of code: vectorized code and non-vectorized code, and can determine the rittleability at runtime and select which code to execute. It is an addition.

The case where vectorization is possible here refers to the case where no calculation is performed using the value of the array variable changed during the loop. Furthermore, if the recursive expression includes variables that change during the loop other than loop control variables, it is determined that vectorization is not possible.

action The effect of this technical means is as follows.

If a loop containing a recursive operation exists in the source program, the above target program will be generated, so if that part is determined to be vectorizable at runtime, the vectorized code will be selected and executed. will reduce execution time.

EXAMPLE Hereinafter, an example of the present invention will be described with reference to the drawings.

FIG. 1 shows an example of a hardware configuration for implementing the present invention. In FIG. 1, 1 is the main memory;
2 is a storage control unit, 3 is a central processing unit, 4 is a channel processor, and 5 and 6 are large storage units. 1 above j61
A vector compiler having the vectorization method of the present invention is stored in the device, and is executed by the central processing unit No. 3. As a result, the source program stored in the large storage device No. 5 is converted into a target program and stored in the large storage device No. 6.

FIG. 2 is a functional block diagram of a vector compiler having the backing method of the present invention. The source interpretation section is
A source program is input from the Daiki 1q device shown in Figure 1/5, and the sentences are interpreted and converted into intermediate code. The storage allocation unit performs address allocation for various data in the program. The loop detection section detects loops in the program. The recursive operation loop vectorization unit extracts a loop including a recursive operation from the program and determines vectorizable conditions. The target program output section converts the intermediate code into the target program and outputs it to the large storage device shown in Figure 1/6. As for the loop including the above, a target program having a processing flow as shown in FIG. 4 is generated.

FIG. 3 shows the flow of processing in the recursive calculation rube vectorization section. First, a recursive operation in a loop is detected. Specifically, this detects a statement that may perform an operation using the value of an array variable changed during loop execution, and can be realized using conventional techniques.

After detecting a recursive operation, the conditions under which the recursive operation can be converted into a vector are determined. This can be obtained by calculating the following equation for all combinations of subscripts of the array variables on the left and right sides, and then taking the logical product.

((Right side array variable subscript) - (Left side array variable subscript)) × (
(Loop upper limit) - (Loop lower limit)) ≧ 0 However, if a variable other than the loop control variable included in the subscript of the array variable changes during loop execution, that is, if such a variable is assigned within the loop. If so, it shall be considered that vectorization is not possible. In the target program output section, for the loop for which this vectorizable condition was found,
A target program is generated that includes both vectorized code and non-vectorized code, and can selectively execute either code at the time of execution based on the determination result of the vectorizable condition. The process flow of this purpose program is explained in the fourth section.
As shown in the figure.

As an example, consider a loop that includes a recursive operation as shown below.

Do 10 I=1.5 10 A H+M)=A (I+N) In this case, the vectorizable condition is ((I +N)-(1+M)) x (5-1)
Since it is obtained by ≧0, it becomes N2M. Therefore, the processing flow regarding this part of the target program to be generated is the same as that of the vectorizable condition in FIG. 4 replaced by N2M. When this is executed, if N2M is satisfied immediately before executing this recursive operation, the vectorized code will be executed.

Effects of the Invention As mentioned above, the present invention has the effect that execution time can be significantly reduced because all loops that include recursive operations that can be vectorized are executed using vectorized code. .

[Brief explanation of the drawing]

FIG. 1 is a hardware configuration diagram for implementing the present invention,
Fig. 2 is a functional block diagram of a vector compiler using the vectorization method of the present invention, Fig. 3 is a flowchart of processing in the recursive operation loop vectorization section, and Fig. 4 is a flowchart of the loop portion including the recursive operation of the target program. It is a flowchart of processing. ■...Main storage device, 2...Storage control unit, 3...Central processing unit, 4.Old channel processor, 5゜6... Large storage device. Name of agent: Patent attorney Toshio Nakao ('JC')
) IR group

Claims (1)

[Claims] It is equipped with a means to detect vectorizable conditions by detecting loops including recursive operations in a source program, and includes both vectorized code and non-vectorized code, and determines vectorizability at runtime. A method for vectorizing a loop including a recursive operation, characterized by generating a target program in which code can be selectively executed.