Memory Pattern Analysis Tool for Use in Optimizing Computer Program Code

The present invention pertains to the field of computer software. More particularly, the present invention relates to a memory pattern analysis tool for assisting in the optimization of computer program code.

An important aspect of designing computer software is optimization. Optimization generally refers to the process of improving computer program code by modifying the code to run faster, to use memory more efficiently, or both. Optimization is generally performed using two basic approaches: 1) code restructuring, and 2) data placement and restructuring. Several well-known code restructuring methods exist for improving the memory hierarchy performance of software, such as scalar replacement, loop interchange, loop distribution, loop fusion, loop blocking, unroll-and-jam, strip-mine-and preload, and strip-mine-and-interchange. These techniques can sometimes be combined to achieve an even higher level of performance. In addition, several well-known data placement and restructuring techniques exist for improving memory hierarchy performance, such as data alignment in COMMON block, padding and aligning structures, placing scalar variables on the same cache block, and array merging.

Existing technology for optimizing software has certain disadvantages, however. For example, a basic problem in optimizing software is determining which of the above optimization methods to apply to a given portion of code. Certain compilers are available which will automatically select and apply various loop transformations in order to restructure program code. However, in so doing, the compilers often determine and apply only certain transformations that are possible for a portion of code, not the transformations that are best suited, or even appropriate, for that portion of code. Consequently, the resulting optimization may be less than optimal. One reason for this deficiency is that the compilers do not have information describing the dynamic memory utilization of the program. Cache simulators have been developed which generate certain dynamic information, such as the number of cache hits and misses. However, cache simulators are limited to cache performance statistics, and the problem remains of determining which optimizations to apply. Further, the compilers do not focus on certain critical portions of the code that have a substantial impact on memory hierarchy performance, such as memory references which occur within loops. Hence, what is needed is a tool which overcomes these deficiencies, for assisting in the optimization of software.

A method of assisting optimization of computer program code including at least one memory reference is provided. Performance data relating to the memory reference is computed during execution of the program code. The performance data is applied to a number of sets of conditions, with each set of conditions corresponding to one of a number of optimization techniques. After determining if each set of conditions is satisfied, an output is generated, for each set of conditions that is satisfied, indicating that the corresponding optimization technique may be beneficial if applied to the program code.

Other features of the present invention will be apparent from the accompanying drawings and from the detailed description which follows.

The present invention is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:

FIG. 1 is a block diagram of a computer system in which the present invention is implemented.

FIG. 2 is a block diagram of a memory pattern analysis tool.

FIG. 3A is a flowchart showing steps for instrumenting, compiling, and linking a program to a memory pattern analysis tool.

FIG. 3B is a flowchart illustrating the step of executing an instrumented program.

A method of assisting optimization of computer program code is described. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be evident, however, to one skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to avoid unnecessarily obscuring the present invention.

The present invention includes a memory pattern analysis tool that indicates which of a number of possible memory optimizations may be helpful if applied to a particular sequence of program code. The present invention is particularly useful in optimizing programs written in languages that are amenable to large numbers of loops, such as FORTRAN or C; however, the present invention is not limited to such languages.

The present invention includes method steps which, in the preferred embodiment, are embodied in computer software. These steps may be performed by a general-purpose processor programmed with instructions that cause the processor to perform the necessary steps. In alternative embodiments, the necessary steps may be performed by specific hardware components that contain hard-wired logic for performing the steps or any combination of programmed general-purpose computer components and customed hardware components.

FIG. 1 illustrates a computer system 1 in which the present invention is implemented. The computer system 1 includes a central processing unit (CPU) 10 capable of performing instructions according to the present invention. The CPU 10 is coupled to a main memory 12 and a cache memory 14 by a system bus 28. A number of peripheral components 18-21 are coupled to a peripheral component interconnect (PCI) bus 26. The PCI bus 26 is connected to the system bus 28 by a bus bridge 16. The peripheral components 18-21 may include any of a number of commonly used input/output devices, such as a display monitor, a keyboard, a mouse or other cursor control device, and a magnetic disk drive, CD-ROM drive, or other mass storage device. In the preferred embodiment, instructions according to the present invention may be stored in memory 12 or on a magnetic disk, CD-ROM, or other mass storage device accommodated by one of the peripheral components 18-21, or any combination of these devices.

The present invention includes a method for identifying which of a number of possible optimization techniques may be beneficial for a given sequence of computer program code. It is assumed for purposes of this description that a program to be optimized contains a number of loops, and that a number of instructions within those loops contain memory references. A memory reference is defined herein as a reference made within an instruction to a location in memory, such that a load, store, or modify operation is performed within the CPU 10 when the instruction is executed. As will be described below in detail, the program to be optimized is initially instrumented by a compiler to cause certain predetermined information to be generated pertaining to each dynamic instance of a memory reference during execution of the program. Upon execution of the program, the trace is generated. From the trace, the loops in the program are identified. For each loop, a number of dynamic performance parameters are computed relating to the memory references occurring within that loop. These parameters include dynamic memory usage information and cache statistics. For each loop, these performance parameters are then applied to a number of sets of conditions. Each set of conditions is associated with a particular known memory optimization technique. A set of conditions forms the criteria for determining whether the corresponding optimization technique will likely be useful if applied to the program. An indication is generated, for each set of conditions that is satisfied, that the optimization corresponding to that set of conditions may be a beneficial. In addition, for each loop in the program, a list of the memory references contained within the loop is generated along with a listing of the computed performance parameters for each memory reference.

FIG. 2 provides a block diagram of the memory pattern analysis tool 32. The tool 32 includes a cache simulator 34 that is logically coupled to a symbolic reference analyzer 36. In the preferred embodiment, the tool 32 is embodied in the form of a library file which is linked to the program to be optimized prior to execution of the program. During execution of the program, the tool 32 receives as input a memory reference trace generated as the program is executed and outputs memory optimization suggestions and the performance parameters mentioned above.

Associated with the program to be optimized is a compiler for converting the program to object code. Referring now to FIG. 3A, the compiler first detects the memory references which are present in the source code (step 301). Next, the compiler instruments (i.e., inserts additional procedure code into) the source code (step 302) to configure the source code to generate a memory reference trace during execution. Specifically, the compiler configures the program to generate, during execution, a trace comprising one record for every dynamic instance of a memory reference. Each record includes the following information: 1) the instruction address of the instruction containing the memory reference; 2) the instruction type of the instruction containing the memory reference (i.e., load, store, or modify); 3) the memory address being accessed as a result of the memory reference; and 4) the location in the source code of the memory reference (file name, line number, and column number in which the memory reference appears). After the program is instrumented, the program is compiled (step 303) and linked with the tool 32, which is a library file (step 304), to produce an executable program. The program is then executed (step 305) in conjunction with a data set, causing both the memory reference trace to be generated and the tool 32 to be invoked as the program is executed.

FIG. 3B is a flowchart illustrating the step of executing the program (step 305) in greater detail. First the individual loops in the program are detected from the memory reference trace using a loop detection algorithm (step 310). In the preferred embodiment, loop nests of up to two levels will be detected. Various loop detection algorithms are known and are suitable for this task.

After detection of the loops, a number of performance parameters are computed for each loop in the program (step 311). These performance parameters include cache statistics and other parameters which will be described below. The cache statistics are generated by the cache simulator 34 portion of the tool 32, while the other parameters are computed by the symbolic reference analyzer 36 portion. Various cache simulators are currently known and can be used to generate the cache statistics. An example of a suitable cache simulator is the CPROF program, developed at the University of Wisconsin.

A number of sets of conditions are provided, with each set of conditions being associated with a different known optimization technique. Each set of conditions represents criteria for determining whether the corresponding optimization technique will likely be useful if applied to the program. For each loop, the performance parameters are applied to each set of conditions (step 312). The tool 32 outputs a list including the memory references contained within the loop, certain ones of the performance parameters associated with each memory reference in the loop, and any optimizations for which the corresponding set of conditions has been satisfied (step 313).

During execution of the program, the tool 32 computes the performance parameters for each loop and stores these parameters in a data structure. The data structure has the following information and organization:

The terms set forth in the data structure above are defined as follows:

A "sub-loop" is a loop contained within a loop.

A "nearby loop" is a loop that is located within a predetermined number of lines of source code of another loop.

A "list of references" is a list of the memory references contained within a loop.

"n" is the number of times a loop is iterated.

"min" is the minimum memory address that is made from a reference point (i.e., from the point in the program at which a reference occurs).

"max" is the maximum memory address that is made from a reference point.

"x.sub.-- stride" is the change in memory address per inner loop iteration, divided by the reference size. The reference size is the number of bytes of memory occupied by the reference, which depends upon the data type of the reference. Dividing by the reference size puts x.sub.-- stride in terms of array elements, as opposed to memory address space.

"x.sub.-- is.sub.-- constant.sub.-- stride" is a Boolean variable which is TRUE only if x.sub.-- stride is a constant.

"y.sub.-- stride" is the change in memory address per outer loop iteration, divided by the reference size. As with X.sub.-- stride, the reference size is the number of bytes of memory occupied by the reference, which depends upon the data type of the reference. Dividing by the reference size puts x.sub.-- stride in terms of array elements, as opposed to memory address space.

"y.sub.-- is.sub.-- constant.sub.-- stride" is a Boolean variable which is TRUE only if y.sub.-- stride is constant.

"x.sub.-- length" is the largest memory address difference of a reference in a loop invocation, divided by the reference size (to put X-length in terms of array elements).

"#cache.sub.-- misses" is the number of cache misses. This parameter is computed buy the cache simulator 34.

"#cache.sub.-- hits" is the number of cache hits. This parameter is computed buy the cache simulator 34.

"#non.sub.-- aligned.sub.-- accesses" is the number of non-aligned accesses that are made from the reference.

A "relation" is two references that appear in the same loop and have made the same change in x.sub.-- stride.

A "source reference" is a reference of a relation which is read in performing a memory operation.

A "destination reference" is a reference of a relation which is written in performing a memory operation.

"same.sub.-- x.sub.-- stride.sub.-- occurrences" is the number of times that two references have made the same change in x.sub.-- stride.

"x.sub.-- difference" is the constant (if defined) difference between memory addresses of two references.

In the preferred embodiment, conditions are provided for the following code restructuring optimizations: 1) scalar replacement, 2) loop interchange, 3) loop distribution, 4) loop fusion 5) loop blocking, 6) unroll-and-jam, 7) strip-mine-and-preload, and 8) strip-mine-and-interchange. In certain cases, a combination of one or more of the preceding transformations may be recommended. Loops having rectangular, triangular, and trapezoidal-shaped iteration space will be focused on for evaluating loop transformations. In addition, conditions are provided corresponding to the following data placement and restructuring optimizations: 1) data alignment and COMMON block, 2) padding and aligning structures, and 3) array merging.

Two scenarios are considered for scalar replacement: 1) loops having loop-dependent dependencies, and 2) loops having loop-independent dependencies. The following code is an example of a loop-dependent dependency: ##EQU1##

The following set of conditions applies to loop-dependent dependencies:

(1) There are no non-constant stride references in the loop, and

(2) There are references in the loop such that:

a) x.sub.-- stride=1, x.sub.-- stride'=1,

b) x.sub.-- difference/opr.sub.-- size is a small constant <>0, and

c) y.sub.-- stride=y.sub.-- stride'.

In the above set of conditions, and in each of the following sets of conditions, the following clarifications apply:

x.sub.-- stride and y-stride apply to a given reference in a loop, while x.sub.-- stride' and y-stride' apply to a different reference in the loop;

opr.sub.-- size refers to the operand size; and

"a small constant" depends on the number of registers that are available and, in the preferred embodiment, is a constant less than or equal to four.

Hence, if the above conditions are satisfied, then the tool 34 will generate an output indicating that scalar replacement may be a beneficial transformation for the subject loop.

The following is an example of scalar replacement with loop-independent dependencies. ##EQU2##

If the following conditions are satisfied in the case of loop-independent dependences, then the tool 34 will generate an output indicating that scalar replacement may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references in the loop, and

(2) There are references in the loop such that:

a) x.sub.-- stride'=x.sub.-- stride=0,

b) y.sub.-- stride'=y.sub.-- stride <>0, and

c) x.sub.-- difference/opr.sub.-- size=y.sub.-- stride * a small constant.

The following is an example of loop interchange and scalar replacement: ##EQU3##

If the following conditions are satisfied, then the tool 34 will generate an output indicating that loop interchange followed by scalar replacement may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride accesses in the loop, and

(2) There is at least one reference such that:

a) x.sub.-- stride=small constant <>0, and

b) y.sub.-- stride=0,

and

(3) There is no reference in the loop such that:

a) x.sub.-- stride=1, and

a) y.sub.-- stride=n

The following is an example of unroll-and-jam and scalar replacement: ##EQU4##

If the following conditions are satisfied, then the tool 34 will generate an output indicating that a combination of unroll-and-jam and scalar replacement may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references, in the loop, and

(2) There are pair-wise references such that:

a) x.sub.-- stride=x.sub.-- stride',

b) y.sub.-- stride=y.sub.-- stride'=a constant <>0, and

c) x.sub.-- difference/opr.sub.-- size=(y.sub.-- stride * a small constant) <>0

If the following conditions are satisfied, then the tool 34 will generate an output indicating that loop interchange may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references in the loop, and

(2) There is at least one reference such that:

a) #cache.sub.-- misses is high

b) y.sub.-- stride=1,

c) x.sub.-- stride=constant,

d) x.sub.-- stride <>0, and

e) x.sub.-- stride <>1.

If the following condition is satisfied, then the tool 34 will generate an output indicating that loop distribution may be a beneficial transformation for the subject loop: the loop has a number of non-interfering references that have high conflict misses (i.e., a miss ratio greater than 10%).

References "interfere" when the accessing of an array element of one reference causes a value of the other reference to be overwritten. One way of determining the existence of interfering references is to determine whether there is overlap between the two references, i.e., whether the highest address accessed by one reference is lower than the lowest address accessed by the other reference. However, it should be appreciated that overlap is only an indication of the possibility of interfering references and not an absolute test. A example of a function for computing interfering references based on overlap is as follows:

where:

ref1 is a first reference in a loop,

ref2 is a second reference in the loop,

ref1.min is the lowest address accessed by the first reference,

ref1.max is the highest address accessed by the first reference,

ref2.min is the lowest address accessed by the second reference, and

ref2.max is the highest address accessed by the second reference.

It should be appreciated that other functions for computing interfering references may also be used.

The following is an example of loop fusion:

For loop fusion, any nearby loops to the subject loop are detected before applying the conditions. If the following conditions are satisfied, then the tool 34 will generate an output indicating that loop fusion may be a beneficial transformation for the subject loop:

(1) the subject loop and the nearby loop are of equal length (n is the same for the two loops), and

(2) there are references in different loops that exhibit locality, i.e.:

a) x.sub.-- stride'=x.sub.-- stride <>0

b) y.sub.-- stride'=y.sub.-- stride, and

c) overlap(ref1, ref2),

where "overlap" is defined above with respect to loop fusion.

The following is an example of loop blocking:

If the following conditions are satisfied, then the tool 34 will generate an output indicating that loop blocking may be a beneficial transformation for the subject loop:

There is at least one reference for which:

a) x.sub.-- stride=0, and

b) y.sub.-- stride=n * a small constant <>0

or,

1) There are no non-constant stride references, and

2) There is at least one reference such that:

a) x.sub.-- stride=1,

b) y.sub.-- stride=(n * a small constant),

c) x.sub.-- length is large,

d) #cache.sub.-- miss rate is high,

e) x.sub.-- stride'=x.sub.-- stride,

f) y.sub.-- stride'=y.sub.-- stride,

g) x.sub.-- difference/opr.sub.-- size=y.sub.-- stride * a small constant <>0, and

h) overlap(ref1, ref2).

If the following conditions are satisfied, then the tool 34 will generate an output indicating that loop unroll-and-jam may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references in the loop, and

(2) There is a reference such that:

a) x.sub.-- stride=1,

b) y.sub.-- stride=0, and

c) the reference is in a loop nest of depth>1,

or,

1) There are no non-constant stride references in the loop, and

2) there are pair-wise references such that:

a) x.sub.-- difference/opr.sub.-- size=y.sub.-- stride * a small constant <>0,

b) x.sub.-- stride'=x.sub.-- stride, and

c) y.sub.-- stride'=y.sub.-- stride=(n * a small constant).

If the following conditions are satisfied, then the tool 34 will generate an output indicating that strip-mine-and-preload may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references in the loop,

(2) There is a small number of references (e.g., less than or equal to four), and

(3) for all references in the loop:

a) x.sub.-- stride=1, and

b) x.sub.-- length is large (i.e., larger than the size of the cache).

If the following conditions are satisfied, then the tool 34 will generate an output indicating that strip-mine-and-interchange may be a beneficial transformation for the subject loop:

(1) There are no non-constant stride references in the loop,

(2) There is a small number of references in the loop, and

(3) For all references:

a) x.sub.-- stride=1,

b) x.sub.-- length is large (i.e., larger than the size of the cache), and

c) y.sub.-- stride=0 or n.

If the following conditions are satisfied, then the tool 34 will generate an output indicating that data alignment in COMMON block may be beneficial:

There are pair-wise references that:

a) x.sub.-- difference <>0, and

b) remainder ((x.sub.-- difference/opr.sub.-- size)/CACHE.sub.-- SIZE)=0

or small constant,

c) x.sub.-- stride=x.sub.-- stride'=1,

d) y.sub.-- stride=y.sub.-- stride',

e) x.sub.-- length is large,

f) x.sub.-- length' is large, and

g) there is a high cache miss ratio in the loop.

If the following condition is satisfied, then the tool 34 will generate an output indicating that padding and aligning structures may be beneficial: There are references that have large numbers of non-aligned accesses. What constitutes a "high number" may be determined according to the user's needs.

If the following conditions are satisfied, then the tool 34 will generate an output indicating that array merging may be beneficial:

In an inner loop, there are pair-wise references that:

(1) the last change in x.sub.-- stride does not equal 1, and

(2) the last change in x.sub.-- stride equals the last change in x.sub.-- stride',

(3) rel->same.sub.-- x.sub.-- stride.sub.-- occurrences/total>0.8, where total=#cache.sub.-- hits+#cache.sub.-- misses, and

(4) x.sub.-- stride <>0, x.sub.-- stride <>1.

For each loop, after all sets of conditions have been applied, the tool 32 outputs a list including the memory references contained within the loop, certain ones of the performance parameters associated with each memory reference in the loop, such as cache statistics, and any optimizations for which the corresponding set of conditions has been satisfied. In one embodiment, the output generated for each loop has the following information and format:

›list of optimization techniques for which conditions have been met.!

Thus, an exemplary output might appear as follows:

Note that, although only certain performance parameters such as cache statistics are output for each reference in the exemplary embodiment, all performance parameters or different selected performance parameters may be output in other embodiments, according to the user's needs.

Thus, a method of assisting optimization of computer program code has been described. Although the present invention has been described with reference to specific exemplary embodiments, it will be evident that various modifications and changes may be made to these embodiments without departing from the broader spirit and scope of the invention as set forth in the claims. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense.