JP3225940B2 - Program optimization method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method and an apparatus for optimizing a program, and more particularly, to a method and an apparatus for optimizing a program including a loop having a loop invariant in a compiler for generating a target program from a source program.

[0002]

2. Description of the Related Art FIG. 8 shows an example of optimization of a loop-invariant expression in a conventional compiler. In the conventional compiler, a loop invariant expression detecting step 72 detects a loop invariant expression, that is, an expression composed of a loop invariant variable and a constant, and a loop invariant expression moving step 73 causes such an expression to be calculated before the loop. Move to. In that case, the result of the calculation is assigned to a temporary variable, and the calculation of the original expression in the loop is replaced with a reference to the temporary variable.

[0003] Here, a loop invariant variable is a variable whose value is not changed in a loop. For example, in the example of FIG. 9, there is no assignment for the variable x in the loop. Therefore, if the value of x is 1 immediately before the loop,
The value of x remains at 1 until the execution of the loop ends.
That is, x is a loop invariant variable. On the other hand, the variable y is substituted by the statement y = i in the loop, and the value is changed during the execution of the loop. Therefore, y is not a loop invariant variable.

When the input given to the program changes, the value held in x may change immediately before the loop, so it is called a loop-invariant "variable". Is not changed.
For example, if a constant is substituted for x just before the loop, the value is always constant for any input,
In such a case, optimization can be easily performed by a conventional optimization method.

That is, a loop invariant variable is a variable whose value is guaranteed not to change only during the execution of a loop.

As an example, consider the loop shown in FIG. In this loop, x is an invariant variable and line 2
The equations shown in 2 and 23 are loop invariant equations.

When the conventional program optimization method is applied to the loop shown in FIG. 10, the loop shown in FIG.
The row indicated by 4 is generated. Line 44 shows the invariant moved before loop 43. The calculation result is assigned to temporary variables t1 and t2.

According to such program optimization,
When the loop of FIG. 10 is executed, the calculations of the invariant expressions 22 and 23 must be performed each time the loop is executed. However, in the optimized program of FIG.
Is calculated only once, thus shortening the execution time.

On the other hand, Japanese Patent Application Laid-Open No. 63-546 discloses an example of using profile information in program optimization.
Japanese Patent Application Publication No. 46-46, which discloses an optimization such as vectorization using the number of times of execution of a loop and the satisfaction ratio of a conditional expression as profile information.

[0010]

In the conventional method for optimizing program information described with reference to FIG. 8, when a variable including a variable whose value is not determined at the time of compiler is included in a loop, calculation of an expression including the variable is performed. It is possible to move before the loop, but it cannot create a more efficient loop.

Further, efficient optimization cannot be performed only with profile information such as the number of times of execution of a loop.

An object of the present invention is to provide a method and an apparatus which can obtain a target program more efficiently than ever in optimizing a program including a loop having a loop invariant variable.

Another object of the present invention is to provide a recording medium on which a program (software) for realizing a compiler is recorded.

[0014]

According to a first aspect of the present invention, there is provided a method for optimizing a target program in a language processing program (compiler) for generating a target program from a source program. Detecting a variable only for a loop having a loop invariant including a variable whose value does not change in the loop; a profile including at least a value at the time of execution of the variable and the number of times of execution of the loop. and dynamic information measurement step of collecting information, the fruit of the variables in the profile information
When the number of executions when the value at the time of execution is a specific constant is large and the size of the loop is equal to or smaller than the specific size, a step of determining to duplicate the loop; and Generating a loop in which the use places of the variables in the loop are replaced by the constants,
When the value of the variable is the constant, the generated loop; otherwise, a loop replication step of inserting a control statement that executes the original loop; and And various optimization steps for performing various optimizations.

According to another aspect of the present invention, there is provided an object program optimizing apparatus for a compiler for generating an object program from a source program, the apparatus comprising a loop invariant including a variable whose value does not change in a loop. Only for a loop having an expression, a loop invariant variable detecting means for detecting the variable, a dynamic information measuring means for collecting profile information including at least a value at the time of execution of the variable and the number of times of execution of the loop, The case where the runtime value of the variable in the profile information is a specific constant.
When the number of lines is large and the size of the loop is equal to or smaller than a specific size, a copy permission / non-permission determining unit that determines that the loop is to be replicated, and when it is determined that the loop is to be replicated, a use location of the variable in the loop. Is generated, and a control statement that executes the generated loop when the value of the variable is the constant, and otherwise executes the original loop is inserted. Loop optimizing means for performing various optimizations on the duplicated loop.

According to still another aspect of the present invention, there is provided a recording medium recording a program for executing each step of the program optimizing method.

[0017]

FIG. 1 is a functional block diagram of a program optimizing device according to the present invention. This program optimizing device includes a loop invariant variable detecting means 1, a dynamic information measuring means 2, a duplication determining means 3, a loop duplicating means 4, and various optimizing means 5. The function of each means will be described together with the description of the program optimization method.

FIG. 2 is a flowchart for explaining the program optimizing method. Steps 12, 13, 15, 16, and 17 in this flowchart are shown in FIG.
Correspond to the functions of the means 1, 2, 3, 4, and 5.

FIG. 3 is an example of a loop to be optimized according to the present invention written in C language. This is the same as the example shown in FIG.

In a compiler for generating a target program 18 from a source program 11, a loop invariant variable detection step 12 detects a variable whose value is not changed in a loop, and a dynamic information measurement step 13
When the program is executed, information (profile information 14) such as the value actually set to the variable and the number of times of execution of the loop is collected. Next, it is determined whether or not the loop is to be duplicated in the determination step 15 using the profile information. Since there is a problem that the size of the program increases due to the duplication of the loop, when determining whether to duplicate, consider the profile information,
Only those with a large effect are targeted for optimization in the present invention.

If the loop invariant variable is often a specific value, a loop duplication step 16 creates a loop in which the use of the variable in the loop is replaced by the constant, and at the same time, the value of the variable is When the value is the specific value, a duplicated loop is added, and otherwise, a control statement for executing the original loop is added at the same time. Then, optimization is applied by various optimization steps 17,
The target program 18 is obtained.

The dynamic information measurement in step 13 will be described in detail. Although there are various methods for measuring dynamic information, two representative methods will be described.

First, in the first dynamic information measurement method, as shown in FIG. 4, the loop invariant variable detected in step 12 of detecting the loop invariant variable is immediately before or immediately after the loop as shown in FIG. Add code that writes values sequentially to a file, compile and execute the program to which the code was added (this is different from the final target program 18). As a result, for example, FIG.
Is executed 12 times (note that this is not the number of iterations of the loop itself), and the file contains 1
0,9,10,10,2,10,10,10,10,1
Assuming that 0, 1, and 10 are written, the value assigned to this variable is 9 in the case of 10 in the execution,
It can be seen that the cases of 9, 2, and 1 were each performed once.

The second dynamic information measurement method differs from the first method in that the loop invariant variable detection step detects the loop invariant variable and compiles the program as it is to generate a target program. This is different from the final objective program 18). In the target program, an address of an area allocated on a memory is detected for a loop invariant variable. Thereafter, a CPU having a function of checking a value stored in a memory area of a specified address (address of a loop invariable variable) at a specified timing (immediately before execution of a loop) and collecting statistics.
By executing the target program using a simulator, statistical information (profile information) of values stored in the memory area is obtained.

Using the profile information obtained by the first and second methods, it is determined in step 16 whether or not the loop is to be duplicated. A typical method for setting the criterion is as follows. explain.

First, a first method is a method of allowing a user to specify a criterion. The compiler that implements the present invention provides the user with the distribution of the loop invariant variables and their values (information about how many times the whole is, and which value was there and how many times) and the size of the loop (the final execution program (Since there is a disadvantage that the size increases), the decision is left to the user, and the loop is copied only when instructed by the user.

As a second method, when the value of the loop invariant variable is equal to or more than a certain threshold value (for example, 90%) of the total number of times, the value of the loop invariable variable is a specific value, and the size of the loop is within the specific size. This is a method in which the setting is made so that the loop is duplicated only when it is.

Next, various optimization steps 17 will be described. What is executed in the various optimization steps is optimization of a general compiler, but in order to effectively implement the present invention, it is assumed that at least the following optimization items are included. (1) Constant calculation A loop invariant whose result is determined at compile time, such as a calculation between constants, is replaced with the value of the calculation result. (2) Movement of loop invariant expression The value of a loop invariant expression consisting of only a loop invariant variable and a constant does not change even if it is calculated at any point during the execution of the loop. As it is redundant as it is calculated each time the loop is repeated, the calculation of the expression is performed only once immediately before the loop, the result is stored in a temporary variable, and the temporary variable is referenced in the loop. To change. This is called loop-invariant movement and is a widely used optimization method.

Next, the processing for optimizing the loop in FIG.
This will be specifically described with reference to the flowchart of FIG.

In the loop shown in FIG. 3, the variable x is a loop invariant whose value is not changed in the loop.
3 is a loop invariant. Such a loop invariant variable x
Is detected in step 12.

In the dynamic information measurement step 13, it is assumed that the value of x is 15 in most cases when the program including this loop is executed to collect the profile information 14. Then, it is assumed that it is determined in the determination step 15 that copying is to be performed.

By the step 16 of duplicating the loop, FIG.
The program is changed as shown in First half loop 4
2 is a part executed when x is 15, and the latter half of the loop 43 is a part executed in other cases. Here, when the first half of the loop 42 is executed, it is determined that x is 15; therefore, x in the first half of the loop is 1
Replaced as 5.

Next, in step 17 of various optimizations, the above-described constant calculation and movement of the loop invariant are applied. By the constant calculation, if (1) in the first half loop 42 is obtained.
5> 10) is always satisfied, so the condition is deleted.
5 * 2 and 15/5 are 30, 3 respectively. In addition, the movement of the loop invariant causes the invariant x *
2, x / 5 are moved before the loop as t1 = x * 2, t2 = x / 5. And the calculation of these invariants is
It is replaced by a temporary reference to variables t1 and t2. As a result, the target program shown in FIG. 6 is obtained.

This objective program comprises a first half loop 32 in which the efficiency is improved and a second half loop 33 in which the loop invariant is moved. According to such a target program, in many cases, the more efficient loop 32 is executed, so that the overall program execution speed is improved.

[0035]

FIG. 7 is a flowchart showing an embodiment of the present invention.

In step 51, the source program 1
1 is read and changed to an internal representation that can be handled by the compiler.

Next, at step 52, it is detected that the loop invariant variable is x.

In step 53, as described with reference to FIG. 4, a code for sequentially writing the value of the variable to a file is inserted immediately before the loop into the detected loop invariant variable.

In step 54, the program to which the code has been added is compiled, and the dynamic information measurement purpose program 19 is created from the internal representation.

In step 55, the target program is executed by the simulator to obtain the profile information 14.

In step 56, it is determined from the obtained dynamic information measurement whether the value C of the loop invariant variable is equal to or more than N% of the total number of times and the loop size is equal to or less than S bytes.

If "Yes" in the step 56, the loop is duplicated in a step 57. Various optimizations including the above-described constant calculation (step 58), movement of the loop invariant expression (step 59), and other optimization (step 60) are performed on the duplicated loop. As another example of the optimization, a process of allocating variables stored in a memory to a resist that can be accessed at high speed is included.

Finally, in step 61, the final target program 18 is created from the internal representation.

In the case of "No" in step 56, the process proceeds to step 59 without duplicating the loop, the loop invariant expression is moved, and the object shown in FIG. The program is obtained.

The program optimizing apparatus and method of the above-described embodiments and examples are actually executed also by a program recorded on a recording medium.

[0046]

According to the present invention, a loop that is more efficient than the original loop is generated, and the more efficient loop is executed in many cases. improves. This is because by using profile information, even for a loop that includes a variable whose value is not determined at compile time, it is possible to generate an efficient loop in which the variable is replaced with an appropriate constant. It is.

In general, the collection of variable profile information (information of the values of execution values) tends to be enormous in both the data amount and the collection time. However, the profile information of the present invention has both the data amount and the collection time. It can be collected in a practical range. This means that the collection of variable profile information is limited to loop-invariant variable values within the loop,
This is because the value of each of those variables need only be collected once immediately before (or immediately after) the loop.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a program optimization device of the present invention.

FIG. 2 is a flowchart for explaining a program optimization method.

FIG. 3 is an example of a loop to be optimized according to the present invention written in C language.

FIG. 4 is a diagram illustrating an example in which a code for sequentially writing the value of a variable to a file immediately before a loop is added.

FIG. 5 is a diagram showing a copied program.

FIG. 6 is a diagram showing an obtained target program.

FIG. 7 is a flowchart showing an embodiment.

FIG. 8 is a diagram illustrating an example of optimization of a loop-invariant expression in a conventional compiler.

FIG. 9 is a diagram for explaining a loop invariant variable.

FIG. 10 is a diagram illustrating an example of a loop.

FIG. 11 is a diagram showing a program generated by a conventional program optimization method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Loop invariant variable detection means 2 Dynamic information measurement means 3 Duplication possibility determination means 4 Loop duplication means 5 Various optimization means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-10-40112 (JP, A) JP-A-8-87417 (JP, A) JP-A-9-319587 (JP, A) JP-A-9-99 330233 (JP, A) JP-A-2-176938 (JP, A) P. Chang et al., "Using Profile Information to Assist Class Code Optimizedions", Software-Practice and Experience, 1991, Vol. 21, No. 12, p. 1301-1321 (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 9/45

Claims (6)

(57) [Claims] 1. A method for optimizing a target program in a language processing program for generating a target program from a source program, the method comprising only a loop having a loop invariant including a variable whose value does not change in the loop. a step of detecting the variable, and the value of the execution time of the variables, and dynamic information measurement step of collecting at least contains profile information and execution count of the loop, the run-time of said variable in said profile information If the value of is a specific constant, the number of executions is large, and if the size of the loop is equal to or smaller than a specific size, a step of determining to duplicate the loop; and Generating a loop in which the use places of the variables in the loop are replaced with the constants,
At the same time, when the value of the variable is the constant, the generated loop, otherwise, a loop copying step of inserting a control statement to execute the original loop, the copied step, A program optimization method, comprising: various optimization steps for performing various optimizations on a loop. 2. Collection of the profile information is performed by adding a code for sequentially writing the value of the variable to a file immediately before or immediately after the loop, and compiling and executing the program to which the code is added. 2. The method according to claim 1, further comprising the steps of: 3. The method according to claim 1, wherein the profile information is collected by compiling a source program in which the variable is detected as it is to generate a target program, and in the target program, the variable is allocated to the variable in a memory. Executing the target program using a simulator having a function of detecting an address of an area, checking a value stored in a memory area at a specified address at a specified timing, and obtaining statistics, the 2. The method according to claim 1, wherein the method is performed by obtaining profile information of a value stored in the program. 4. The various optimizing steps include: a constant calculation for replacing a loop invariant whose result is determined at compile time with a value of a calculation result; and a calculation of a loop invariant only once immediately before a loop and temporarily storing the result. 4. The method according to claim 1, further comprising: moving a loop invariant expression stored in a variable and changing the temporary variable in the loop so as to refer to the temporary variable.
The program optimization method according to any one of the above. 5. An object program optimizing apparatus for a language processing program for generating an object program from a source program, wherein the variable is used only for a loop having a loop invariant including a variable whose value does not change in the loop. a loop invariant variable detection means for detecting a value of the execution time of the variables, and dynamic information measuring means for collecting at least contains profile information and execution count of the loop, the execution of the variables in the profile information If the value of the time is a specific constant, the number of executions is large, and if the size of the loop is equal to or smaller than a specific size, a copy permission / non-permission determining unit that determines to copy the loop; and Generates a loop in which the use places of the variables in the loop are replaced with the constants,
At the same time, if the value of the variable is the constant, the generated loop is inserted; otherwise, a loop duplicating unit that inserts a control statement that executes the original loop; and A program optimizing device, comprising: various optimizing means for performing various optimizing operations on a program. 6. When optimizing a target program when generating a target program from a source program, the variable is detected only for a loop having a loop invariant including a variable whose value does not change in the loop. steps and the value of the execution time of the variables, and dynamic information measurement step of collecting at least contains profile information and execution count of the loop, the runtime value specific constant of the variable in the profile information If the number of executions is large and the size of the loop is equal to or smaller than a specific size, a step of judging to duplicate the loop; and Generate a loop in which the used places are replaced with the constants,
At the same time, when the value of the variable is the constant, the generated loop, otherwise, a loop copying step of inserting a control statement to execute the original loop, the copied step, A recording medium on which is recorded a program for executing various optimization steps for performing various optimizations on the loop, and