JP3953697B2 - Compiler and recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a compiler and a recording medium for optimizing the number of expansions when a source program is vectorized.
[0002]
[Prior art]
Conventionally, when a source program is vectorized, an unrolling process has been performed to increase the speed by reducing the number of loop iterations. The number of developments in this conventional unrolling process was determined as follows.
[0003]
(1) The number of expansions specified in the compiler directive.
(2) If the number of iterations of the outer loop is clearly known, that value is taken as the number of expansions.
[0004]
(3) In addition to the above, if the number of instructions after expansion does not exceed an allowable range (for example, a range where there is no register shortage), n-fold expansion is performed.
[0005]
[Problems to be solved by the invention]
With the above conventional number of expansions, it is good if the number of operations is large, but there is a problem that an object that does not use registers effectively is generated in a loop with a small number of operations.
[0006]
There is also a problem that there is a risk that the amount of objects becomes large due to an increase in the number of unrolling.
In order to solve these problems, the present invention performs an unrolling process optimized based on the number of loop iterations when vectorizing a source program, deletes duplicate memory accesses, and executes an object program. The purpose is to improve performance.
[0007]
[Means for Solving the Problems]
Means for solving the problem will be described with reference to FIG.
In FIG. 1, a source program 1 is a target program for optimizing the number of expansions when vectorizing.
[0008]
The compiler 2 receives the source program 1 and generates an executable object program 10, and is composed of an optimization unit 4 and the like here.
[0009]
The optimization unit 4 performs optimization by vectorizing the source program 1, and is composed of a vectorization unit 5, a vector optimization unit 6, and the like here.
[0010]
The vector optimizing means 6 optimizes the number of expansions when vectorizing. Here, the vector optimizing means 6 comprises an unrolling means 7 and a memory access deleting means 8.
[0011]
The unrolling means 7 expands with the optimal number of expansions when vectorizing.
The memory access deleting means 8 changes the overlapping memory access after the expansion with the optimum expansion number into an inter-register transfer instruction and deletes it.
[0012]
Next, the operation will be described.
The vectorizing means 5 vectorizes the source program 1, and the unrolling means 7 analyzes the source program 1 and virtually expands it based on the number of registers obtained from the number of loop iterations detected and the maximum number of registers necessary for the operation. The number of virtual expansions is calculated, the number of expansions is calculated based on the data dependency of the vectorized program and expanded, and the memory access deletion means 8 is a memory duplicated from the expanded vector instruction sequence. When there is an access instruction, it is changed to a register-to-register move instruction.
[0013]
Further, when the vectorization means 5 vectorizes the source program 1 and the unrolling means 7 analyzes the source program and the number of loop iterations detected is unknown, it is obtained from the vector dimension of the array used in the loop. The number of virtual expansions is calculated based on the number of registered registers and the maximum number of registers required for the operation, and the number of expansions is calculated based on the data dependency of the vectorized program for the calculated number of virtual expansions. A vectorized program is expanded based on the number, and the memory access deleting means 8 changes to a register-to-register move instruction when there is an overlapping memory access instruction from the expanded vector instruction sequence.
[0014]
At these times, the maximum number of expansions is set by default, and the number of expansions is calculated within a range not exceeding the maximum value.
Therefore, the execution performance of the object program 10 is improved by performing an unrolling process optimized based on the number of loop iterations when vectorizing the source program 1 and deleting redundant memory accesses. Is possible.
[0015]
DETAILED DESCRIPTION OF THE INVENTION
Next, embodiments and operations of the present invention will be described in detail sequentially with reference to FIGS.
[0016]
FIG. 1 shows a system configuration diagram of the present invention.
In FIG. 1, a source program 1 is a program to be subjected to optimization of the number of expansions when vectorization is performed, and is a program in which a loop exists, for example, as shown in FIG.
[0017]
The compiler 2 inputs the source program 1 and generates an object program 10 in an executable format. Here, the compiler 2 includes a source program analysis means 3, an optimization means 4, a code generation means 9, and the like. Is.
[0018]
The source program analysis means 3 performs morphological analysis, syntax analysis, etc. on the source program 1 to generate an intermediate language. Here, in practice, vectorization and unrolling processing, which will be described later, are performed on the basis of intermediate language to which information obtained through morphological analysis and syntax analysis is added. The description will be made assuming that vectorization and unrolling processing are performed.
[0019]
The optimization unit 4 optimizes the source program 1 and is composed of a vectorization unit 5 and a vector optimization unit 6 in this example.
[0020]
The vectorization means 5 vectorizes the source program 1, for example, a source program 1 shown in FIG. 3A (to be described later) is a program shown in FIG. 3B (a program using a vector instruction that operates on a vector computer). ).
[0021]
The vector optimizing means 6 optimizes the vectorized program, and here is composed of an unrolling means 7 and a memory access deleting means 8.
[0022]
The unrolling means 7 expands the vectorized program with the optimal number of expansions to reduce the number of loop iterations.
The memory access deleting means 8 changes a memory access that is duplicated in a program after being developed with the optimum number of expansions into an inter-register transfer instruction and deletes it.
[0023]
The code generation means 9 converts the optimized program into an executable object program 10.
Next, the operation of the configuration of FIG. 1 will be described in detail according to the order of the flowchart of FIG.
[0024]
FIG. 2 shows a flowchart for explaining the operation of the present invention.
In FIG. 2, S1 determines whether there is a compiler directive designation. In the case of YES, the process proceeds to the processing after S 6 in accordance with the specification. On the other hand, if NO, the process proceeds to S2.
[0025]
In S2, it is determined whether the number of loop repetitions is unknown. This is because the number of loops of the source program 1 is unknown, for example, it is determined that the number of loops is found in the source program 1 in FIG. 3A described later, and the number of loops in the source program 1 in FIG. Is determined to be unknown. In the case of YES, the process proceeds to S 3. In the case of NO, the process proceeds to S 9.
[0026]
S 9, since the number of iterations of the loop is determined not unknown in S2 NO, the
Number of virtual expansions = (number of registers estimated from the number of iterations in the loop) / (maximum number of necessary registers)
Ask for. For example, in the program after vectorization of FIG. 3B described later, in the case of the program, (the number of registers estimated from the number of repetitions in the loop) = 256, and the maximum number of necessary registers = 1. Therefore, although the number of virtual deployments is 256, the maximum number of virtual deployments is determined to be 4 because the maximum is set to 4 by default in consideration of various performances of the vector computer to be executed.
[0027]
S 10, the deletion of the memory access is possible or not. For example, as shown in (c-1) of FIG. 3 to be described later, there is an overlapping memory access in the case of expansion (in this case, quadruple expansion), and the memory access is deleted and changed between registers. Determine if it can be replaced with a move command. If YES, the process proceeds to S 11. On the other hand, in the case of NO, the process proceeds to S 6.
[0028]
S 11, since proved to memory accessible delete YES in S 10, the official number deployment is corrected view data dependencies. For example, when the vectorized program of FIG. 3 to be described later (b) it is determined from the double expansion since it is possible to remove a memory access instruction, as it is formally expand the number of virtual deployment number 4 decided in S 9 To do. Then, the process proceeds to S 6.
[0029]
S 6 performs the expansion process of the loop unrolling. This is formally expanded number determined in S 9, where on the basis of 4, for example, in FIG. 3 (b) deploying quadruple as shown in the FIG. 3 (c-1).
[0030]
S 7, the deletion of the memory access is possible or not. In the case of YES, the process proceeds to S 8. If NO, the process ends.
S 8 since proved possible to remove the memory access by a YES S 7, and delete processing of the memory access. For example, the overlapping memory access in the program after quadruple expansion of (c-1) in FIG. 3 is replaced with an inter-register movement instruction (for example, VMOVE instruction), the overlapping memory access is deleted, and (c− Modify the program shown in 2).
[0031]
By the procedure of S 8 from S 11, S 6 from NO, S 9 of the above S1 of NO, S2, if the number of iterations of the loop of the program is found, the number of iterations of the loop performing multiple developed for inner loop It is possible to improve the execution performance of the object program 10 by reducing the memory access and replacing the duplicated memory access with the VMOVE instruction to minimize the memory access.
[0032]
Next, a case where the number of loop repetitions is unknown will be described below.
In FIG. 2, S 3 is determined to be unclear how many times the loop is repeated when S 2 is YES.
Number of virtual expansions = (number of registers estimated from the number of elements appearing in the vector dimension subscript of the array used in the loop) / (maximum number of necessary registers)
Ask for. For example, in the program after vectorization of FIG. 4B described later, in the case of the program, (the number of registers estimated from the number of repetitions in the loop) = 256, (the maximum value of the necessary number of registers) = 1. .3, it is determined that the number of virtual deployments = 256 / 1.3 = 196.
[0033]
S 4 estimates the maximum number of expansion from the data dependencies. For example, in the case of the program shown in FIG. 4 (b), deletion of memory access cannot be expected even after double expansion or more, so the maximum expansion number = 2 is determined.
[0034]
S 5 is officially number expansion is determined in a range not exceeding the maximum number of deployment. Here, for example, in the case of FIG. 4B, as described above, since the maximum number of expansions = 2, the number of formal expansions not exceeding this is determined to be two.
[0035]
S 6 performs the expansion process of the loop unrolling. This, S 5 formal development number determined in, here on the basis of 2, for example, to deploy double as shown in shown in FIG. 4 (c) the (b) of FIG.
[0036]
S 7, the deletion of the memory access is possible or not. In the case of YES, the process proceeds to S 8. If NO, the process ends.
S 8 since proved possible to remove the memory access by a YES S 7, and delete processing of the memory access.
[0037]
More S1 of NO, S2 YES of, by the procedure of S 8 from S 3, if the number of iterations of the program loop is unknown, the number of elements appearing in the index of the vector dimension of the array that are used in the interior of the loop, An object program that determines the number of expansions based on the maximum number of necessary registers, performs multiple expansion, reduces the number of loop iterations, replaces duplicate memory access with a VMOVE instruction, and minimizes memory access. 10 execution performance can be improved.
[0038]
FIG. 3 is an explanatory diagram of the present invention (part 1, identification of the number of loop iterations). Here, a program for determining the number of loop iterations will be described below.
FIG. 3A shows an example of the source program 1. The source program 1 is a program composed of a double loop as shown in the figure.
[0039]
FIG. 3B shows an example of a program before unrolling development. This is a program after vectorizing the source program 1 in FIG. 3A, and is a single development before unrolling development as shown.
[0040]
FIG. 3C shows a program after unrolling development. Here, the program before unrolling expansion in FIG. 3B is quadruple expanded to create the program of (c-1), and then the duplicate memory access in the program of (c-1) The instruction is deleted and replaced with a register-to-register move instruction (VMOVE instruction) to reduce the memory access to the minimum necessary, and the program (c-2) is created.
[0041]
(C-1) in FIG. 3 shows an example of a program in which the program in FIG. 3 (b) is expanded four times (the number of expansions 4 determined in S3 to S5 in FIG. 2 described above). In the case of quadruple expansion, another loop is generated for the remainder of the division of the loop divided by 4 (see the DO loop at the end of (c-1) in FIG. 3). . Here, as shown in the figure, since there are three overlapping memory accesses, these overlapping memory accesses are deleted and replaced with a VMMOVE instruction, and the program shown in FIG. Has improved.
[0042]
(C-2) in FIG. 3 shows an example of a program when memory access is deleted. This is a modification in which one of the overlapping memory accesses in the program of (c-1) in FIG. 3 is deleted and the other is transferred between registers by the VMOVE instruction for processing.
[0043]
The determination of the number of expansions in the case of the program of FIG. 3 will be described.
In the program of FIG. 3A, the number of loop iterations is known, and the number of registers estimated from the number of loop iterations is 256 here, and the maximum number of necessary registers is 1, so the number of virtual expansions = 256. / 1 = 256, but since the normal maximum virtual deployment number = 4 is set by default, the virtual deployment number = 4 is determined. Next, from the viewpoint of whether or not memory access can be deleted, from the viewpoint of data dependency, it is found that memory access can be deleted from double expansion, so the official expansion number is determined to be 4 as it is. When the determined quadruple expansion is performed, a portion divisible by 4 is set as the upper loop (four-fold expansion loop) of (c-1) in FIG. 3, and the remaining portion of FIG. 3 (c-1). Let it be the lower loop (the loop that has been unfolded once). Further, if the overlapping memory access is replaced with a register-to-register move instruction (VMOVE instruction), the program of (c-2) in FIG. 3 can be generated.
[0044]
FIG. 4 is an explanatory diagram of the present invention (No. 2, loop repetition count unknown). Here, a program whose loop repetition count is unknown will be described below.
FIG. 4A shows an example of the source program 1. The source program 1 is a program composed of a double loop as shown in the figure.
[0045]
FIG. 4B shows an example of a program before unrolling development. This is a program after vectorizing the source program 1 in FIG. 4A, and is a single development before unrolling development as shown.
[0046]
FIG. 4C shows a program after unrolling development. Here, the program before unrolling development of FIG. 4B is created by double development.
The determination of the number of expansions in the case of the program of FIG. 4 will be described.
[0047]
Since the number of loop iterations is unknown in the program of FIG. 4A, the number of registers estimated from the number of elements appearing in the vector dimension subscript of the array used in the loop is 256, which is necessary. Since the maximum value of the number of registers is 1.3, the number of virtual expansions = 256 / 1.3 = 196. If the maximum number of expansions is estimated from the data dependency, the deletion of memory access cannot be expected even if the number of expansions is two or more, so the maximum number of expansions is determined as two. Since the number of formal expansions is determined within a range not exceeding the maximum number of expansions, in this case, the number of formal expansions is determined to be 2 and double expansion is performed as shown in the program of FIG. 4C.
[0048]
【The invention's effect】
As described above, according to the present invention, when the source program 1 is vectorized, an unrolling process optimized based on the number of loop iterations and the like is performed, and a redundant memory access is deleted. Therefore, the execution performance of the object program 10 can be improved.
[Brief description of the drawings]
FIG. 1 is a system configuration diagram of the present invention.
FIG. 2 is a flowchart explaining the operation of the present invention.
FIG. 3 is an explanatory diagram (part 1) of the present invention.
FIG. 4 is an explanatory diagram (part 2) of the present invention.
[Explanation of symbols]
1: source program 2: compiler 3: source program analysis means 4: optimization means 5: vectorization means 6: vector optimization means 7: unrolling means 8: memory access deletion means 9: code generation means 10: object program

Claims (4)

In a compiler that optimizes the number of expansions when vectorizing a source program,
Means for vectorizing the source program;
Means for determining the number of iterations of the loop detection, or by detecting the number of repetitions of the loop unknown analyzes the source program,
If the number of loop iterations is detected , calculate the number of virtual expansions based on the number of registers obtained from the detected number of loop iterations and the maximum number of registers required for the operation , while Means for calculating the number of virtual expansions based on the number of registers obtained from the vector dimension of the array used in the loop and the maximum number of registers necessary for the operation when it is determined that the number of times is unknown ;
Means for calculating the number of expansions based on the data dependency of the vectorized program for the calculated number of virtual expansions;
A compiler comprising means for developing the vectorized program based on the calculated number of expansions. 2. The compiler according to claim 1, further comprising means for changing to a register-to-register move instruction when there is a duplicate memory access instruction after the program is expanded .   3. The compiler according to claim 1, wherein a maximum value of the expansion number is set as a default, and the expansion number is calculated within a range not exceeding the maximum value. On the computer,
Means for vectorizing the input source program;
Means for determining the number of iterations of the loop detection, or by detecting the number of repetitions of the loop unknown analyzes the source program,
If the number of loop iterations is detected , calculate the number of virtual expansions based on the number of registers obtained from the detected number of loop iterations and the maximum number of registers required for the operation , while Means for calculating the number of virtual expansions based on the number of registers obtained from the vector dimension of the array used in the loop and the maximum number of registers necessary for the operation when it is determined that the number of times is unknown ;
Means for calculating the number of expansions based on the data dependency of the vectorized program for the calculated number of virtual expansions;
A computer-readable recording medium recording a program that functions as means for developing the vectorized program based on the calculated number of expansions.

Images