Methods and apparatus for dynamic prediction by software

This application is a continuation of U.S. patent application Ser. No. 11/344,403, filed on Jan. 31, 2006, which claims the benefit of the filing date of U.S. Provisional Patent Application No. 60/650,157, filed Feb. 4, 2005, the entire disclosures of which are hereby incorporated herein by reference.

The present invention relates generally to the field of computing systems, and methods for improving instruction execution, for example, in reducing branch instruction delays in pipelined processors.

Computer processors often use instruction buffers to speed up the execution of programs. Such a buffer, also referred to as an “instruction cache,” allows the processor to queue the next several instructions in a pipeline while the processor is simultaneously executing another instruction. Thus, when the processor finishes executing an instruction the next instruction in the cache is available and ready for execution.

Many modern computing systems utilize a processor having a pipelined architecture to increase instruction throughput which use one or more instruction caches implemented by hardware or firmware.

Pipelining of instructions in an instruction cache may not be effective, however, when it comes to conditional jumps. When a conditional jump is encountered, the next set of instructions to be executed will typically be either the instructions immediately following the conditional jump instruction in sequence, which is currently stored in the instruction cache, or a set of instructions at a different address, which may not be stored in the cache. If the next instruction to be executed is not located at an address within the instruction cache, the processor will be effectively paused (e.g., by executing “NOP” instructions) for a number of clock cycles while the necessary instructions are loaded into the instruction cache.

Accordingly, when a conditional branch or jump is made, the processor is likely to have to wait a number of clock cycles while a new set of instructions are retrieved. This branch instruction delay is also known as a “branch penalty.” A branch penalty may be shorter when branching to an instruction already within the cache and longer when the instruction must be loaded into the cache.

As an illustrative example of a branch penalty, consider the following set of processor instructions

In this sample set of processor instructions, program execution will jump from the “Branch COND, L 1 ” instruction to the “Inst C 0 ” instruction at label L 1 if COND is TRUE, or non-zero. Otherwise, program execution will proceed to the “Inst B 0 ” instruction first. If the instructions at the L 1 label are not in the processor's instruction buffer at the time the “Branch COND, L 1 ” instruction is executed, the processor will have to read the instructions into the buffer, during which time no program instructions are executed. The processor clock cycles wasted by a processor awaiting instructions to be read into its instruction buffer are a measure of the branch penalty. To further illustrate a branch penalty using this scenario, consider the above set of processor instructions at execution time when the branch is taken and the instructions at L 1 are not present in the processor's instruction buffer:

For simplicity, this example assumes a single clock cycle per instruction, branch or NOP. Also, the number of clock cycles needed to provide the instruction “Inst C 0 ” is only for exemplary purposes, i.e., the “Inst C 0 ” instruction may be available to the processor before or after clock cycle 18 . In the above example, although the “Branch to L 1 ” instruction was executed at clock cycle 2 , the instruction “Inst C 0 ” at L 1 was not available to the processor in its instruction buffer until clock cycle 18 . The branch penalty in this example is thus 16 clock cycles.

Several methods have been developed to minimize or eliminate the branch penalty. These methods include both hardware and software approaches. Hardware methods have included the development of processor instruction pipeline architectures that attempt to predict whether an upcoming branch in an instruction set will be taken, and pre-fetch or pre-load the necessary instructions into the processor's instruction buffer. In theory, pipelined processors can execute one instruction per machine clock cycle when a well-ordered, sequential instruction stream is executed. This may be accomplished even though each instruction itself may implicate or require a number of separate micro-instructions to be effectuated.

In one pipeline architecture approach, a fixed algorithm is employed to predict if an instruction branch will be taken. This approach has a drawback in that the fixed algorithm is not changeable, and thus cannot be optimized for each program executed on the processor.

Another hardware approach uses a branch history table (“BHT”) to predict when a branch may be taken. A BHT may be in the form of a table of bitmaps wherein each entry corresponds to a branch instruction for the executing program, and each bit represents a single branch or no-branch decision. Some BHT's provide only a single bit for each branch instruction, thus the prediction for each occurrence of the branch instruction corresponds to whatever happened last time. This is also known as 1-bit dynamic prediction. Using 1-bit prediction, if a conditional branch is taken, it is predicted to be taken the next time. Otherwise, if the conditional branch is not taken, it is predicted to not be taken the next time.

A BHT is also used to perform 2-bit dynamic prediction. In 2-bit dynamic prediction, if a given conditional branch is taken twice in succession, it is predicted to be taken next time. Likewise, if the branch is not taken twice in succession, it is predicted to not be taken the next time. If the branch is both taken once and not taken once in the prior two instances, then the prediction for the next instance is the same as the last time. Generally, if the branch is used for loop or exception handling, 2-bit dynamic prediction using a BHT is better than 1-bit because the branch is taken only once per loop or execution. Two-bit prediction tends to be more accurate and have a greater hardware cost. Therefore, if the state does not change frequently, 1-bit prediction may be sufficient for many purposes.

A BHT uses a significant amount of processor hardware resources, and may also result in significant branch penalties.

In a software approach, an instruction is provided that causes a pre-fetch of instructions starting at a specified address. In one implementation, this software takes the form of a HINT(ADDRESS) instruction, wherein the processor automatically pre-fetches the instructions beginning at ADDRESS as soon as the HINT is encountered. By placing the HINT several instructions before the actual branching instruction, the programmer can reduce the number of clock cycles during which the processor is awaiting the fetch of the next instruction.

One drawback to the use of the HINT(ADDRESS) instruction is that it unconditionally causes the pre-fetching of instructions. Thus, whether or not the branch is taken, the instructions at the branch address will be pre-fetched. The programmer must therefore decide where to place the HINT(ADDRESS) instruction, and the programmer may not always make the best predictions as to when the branch will be taken. If the programmer is incorrect, i.e., if the HINT(ADDRESS) instruction is given and the branch is not taken, the pre-fetch is a wasted effort. A significant time penalty is thus incurred from having to squash the erroneous instruction, flush the pipeline and re-load the correct instruction sequence. Depending on the size of the pipeline, this penalty can be quite large.

The hardware and software approaches to minimizing or eliminating branch penalties also do not account for the fact that the probability of conditions for branching varies throughout the execution of a program.

The present invention addresses these drawbacks.

In one aspect, the present invention provides a method that includes specifying a value in a first portion of a conditional pre-fetch instruction associated with a branch instruction used for effectuating a branch operation in a processor. Next a target instruction address is specified in a second portion in the conditional pre-fetch instruction. The value is then evaluated to determine whether a condition is met. If the condition is met, one or more instructions are pre-fetched starting at the target instruction address into an instruction buffer of the processor.

In another aspect, the target instructions are pre-fetched as described above when the value is non-zero.

A method of operating a processor having an instruction cache and using a branch control instruction associated with a conditional pre-fetch instruction. The conditional pre-fetch instruction includes a test condition portion and an instruction address portion. The method includes the steps of: (a) determining if the test condition portion of the conditional pre-fetch instruction evaluates to a TRUE value; and (b) when the test condition portion of the conditional pre-fetch instruction evaluates to a TRUE value, preloading one or more instructions beginning at an address indicated by the instruction address portion of the conditional pre-fetch instruction into an instruction buffer of the processor.

A non-transitory computer readable storage medium is also provided in an aspect of the present invention. The non-transitory computer readable storage medium contains a program including a conditional pre-fetch instruction operable to cause a processor to perform several steps. a value is specified in a first portion of the conditional pre-fetch instruction. The conditional pre-fetch instruction is associated with a branch instruction used for effectuating a branch operation in the processor. A target instruction address is specified in a second portion in the conditional pre-fetch instruction. The value is used to determine whether a condition is met. If the condition is met, one or more instructions starting at the target instruction address are pre-fetched into an instruction buffer of the processor.

A processor under control of a program including a conditional pre-fetch instruction in conjunction with a branch instruction is also provided. The program causes the processor to decode a first portion of the conditional pre-fetch instruction, which specifies a value for evaluation. the value evaluates to a true or false. Next, if the value evaluates to true, a second portion of the conditional pre-fetch instruction is decoded. This second portion identifies an address of a particular target instruction, which is then pre-fetched. The pre-fetching operation is associated with an operation to move the particular target instruction from a cache to an instruction buffer of the processor.

FIG. 1 is a schematic block diagram of a system in accordance with an embodiment of the present invention;

FIG. 2 is an exemplary flowchart of a method in accordance with an embodiment of the present invention; and

FIG. 3 is a flow chart of a method in accordance with an embodiment of the present invention.

As shown in FIG. 1 , a computer system 100 in accordance with one embodiment of the invention comprises a central processing unit (“CPU”) 102 , an instruction cache 104 , a pipeline 106 connecting the CPU 102 to the instruction cache 104 , and a data bus 108 . Although only these components are depicted in FIG. 1 , it should be appreciated that a typical system 100 can include a large number of components, peripherals, and communications buses. In a preferred embodiment, the computer system 100 is a general purpose computer having all the additional internal components normally found in a personal computer such as, for example, a display, a CD-ROM, a hard-drive, a mouse, a keyboard, speakers, a microphone, a modem and all of the components used for connecting these elements to one another. These components have not been depicted in FIG. 1 for clarity. Additionally, the computer system 100 may comprise any workstation, PDA, or other processor-controlled device or system capable of processing instructions.

The CPU 102 may be a processor of any type. The CPU 102 might also be a multiprocessor system. In a multiprocessor system, a more complex instruction cache 104 and pipeline 106 may be used than those depicted here, although the present invention is similarly applicable to such systems.

The instruction cache 104 may also be of any sort. For example, the instruction cache 104 can be a 32 kiloword (KW) instruction cache using four word blocks. Alternatively, any size instruction cache 104 using any size blocks may also be used. The terms “instruction cache” and “instruction buffer” are used herein interchangeably, with no distinction intended between the use of either term.

As depicted in FIG. 1 , the instruction cache 104 and CPU 102 are interconnected by an instruction pipeline 106 . In an embodiment of the invention, this pipeline represents the hardware and/or firmware used by the system 100 to manage and move instructions between the instruction cache 104 and the CPU 102 . For example, many CPUs have specialized circuitry to manage their associated instruction pipelines and buffers.

In a preferred embodiment, the instruction cache 104 and pipeline 106 are packaged with the CPU 102 in a single integrated circuit (not depicted). Such packaging advantageously provides close proximity of the instruction cache 104 , pipeline 106 and CPU 102 , which minimizes power consumption and instruction transfer time.

In an embodiment of the invention, the instruction cache 104 and CPU 102 may each be in communication with a data bus 108 , thereby allowing the transfer of instructions and data from other devices and memory (not depicted).

FIG. 2 is an exemplary flowchart 200 of a method in accordance with an embodiment of the present invention. Preferably, the CPU executes instructions sequentially. Thus, in FIG. 2 , “Inst A 0 ” 202 is first executed. Then the CPU executes the “CPIF(vall, L 1 )” 204 instruction. The CPIF instruction of this example is the conditional instruction pre-fetch. In the present example, a programmer has placed the CPIF instruction several instructions before the “Branch on COND 1 ” instruction 210 . This placement allows time for the instructions at the branch-to address to be pre-fetched, if necessary, prior to the CPU executing the “Branch on COND 1 ” instruction 210 .

In an embodiment, the “CPIF(vall, L 1 )” instruction 204 causes the CPU or hardware instruction pipeline circuitry to evaluate the “vall” component of the “CPIF(vall, L 1 )” instruction 204 . In a preferred embodiment, if the “vall” component evaluates as a non-zero value, i.e., vall is ‘TRUE’ at action 216 , then the CPU or hardware instruction pipeline circuitry begins the pre-fetch of instructions located at the address given by the “L 1 ” component 218 .

The CPU, meanwhile, may continue to execute the “Inst A 1 ” instruction 206 , as well as any others in sequence. At some point, as depicted here by “Inst AN” 208 , the instruction immediately preceding the “Branch on COND 1 ” instruction 210 is executed. In a preferred embodiment, the pre-fetch of instructions located at the address given by the “L 1 ” component 218 of the “CPIF(vall, L 1 )” instruction is completed prior to the CPU execution of the “Branch on COND 1 ” instruction 210 .

When the CPU executes the “Branch on COND 1 ” instruction 210 in an embodiment of the present invention, either the condition given by “COND 1 ” is ‘TRUE’ or it is ‘FALSE’. Preferably, if COND 1 is ‘FALSE’, the branch is not taken, and the CPU continues program execution with the next sequential instructions, “Inst B 0 ” 212 , then “Inst B 1 ” 214 , and so on. Otherwise, if COND 1 is ‘TRUE’, the branch is taken, and the CPU preferably continues execution with “Inst C 0 ” 222 , then “Inst C 1 ” 224 , and so forth. If the pre-fetch of instructions located at the address given by the “L 1 ” component 218 , namely the “Inst C 0 ” 222 , “Inst C 1 ” 224 , etc. instructions is not completed when the CPU executes the “Branch on COND 1 ” instruction 210 , then a branch penalty 220 occurs while the pre-fetch completes.

A greater branch penalty is encountered if the “CPIF(vall, L 1 )” instruction 204 causes the incorrect prediction that the branch will or will not be taken. For example, if the “CPIF(vall, L 1 )” instruction 204 causes the prediction that the branch will be taken and it is not taken, then the instruction cache may not have the “Inst B 0 ” 212 and “Inst B 1 ” 214 instructions when they are needed, likely resulting in the need to flush and re-fill the instruction cache. This is known as an instruction cache “miss”. Similarly, if the “CPIF(vall, L 1 )” instruction 204 causes the prediction that the branch will not be taken and it is taken, then the instruction cache may not have the “Inst C 0 ” 222 and “Inst C 1 ” 224 instructions when they are needed, also likely resulting in the need to flush and re-fill the instruction cache.

The present invention advantageously minimizes the number of instruction cache misses by allowing dynamic prediction of whether a given branch will be taken. In an embodiment of the present invention, this dynamic prediction is enabled by providing for testing a logical condition embedded within the CPIF instruction. By proper selection of the condition to test, a programmer can greatly increase the accuracy of the branching predictions.

In the example shown in FIG. 2 , a relatively simple parameter of the CPIF instruction, ‘vall’, is evaluated to determine if it is non-zero (TRUE) or a zero (FALSE), corresponding to ‘pre-fetch’ and ‘do not pre-fetch’. Likewise, the preferably more complex condition ‘COND 1 ’ of the “Branch on COND 1 ” instruction 210 is evaluated to determine if the branch is taken. By selecting a ‘vall’ whose TRUE or FALSE status closely corresponds with the ‘COND 1 ’ of the “Branch on COND 1 ” instruction, the programmer may improve the branching predictions. Indeed, in an embodiment of the present invention, ‘vall’ may be identical to ‘COND 1 ’.

FIG. 3 presents a more detailed flowchart 300 of a simplified processor executing a conditional pre-fetch instruction in accordance with an embodiment of the present invention. In a general operation, a processor decodes the instruction addressed by its instruction pointer 302 . The processor determines if the decoded instruction is a conditional pre-fetch instruction 304 . If it is not a conditional pre-fetch, the processor proceeds with instruction processing 306 . When the instruction processing is completed, the processor's instruction pointer is incremented 308 to point to the next sequential instruction, which is then decoded by the processor 302 .

If the instruction is a conditional pre-fetch, the processor evaluates the ‘value’ component of the conditional pre-fetch instruction 310 . In a preferred embodiment, the conditional pre-fetch instruction has the form: CPIF(value, address), wherein the ‘CPIF’ is the instruction pneumonic, ‘value’ is the expression to be evaluated, and ‘address’ is the beginning address of the instructions to be pre-fetched if value is TRUE. Although the pneumonic ‘CPIF’ is used herein to represent the conditional pre-fetch instruction, any pneumonic may be employed in an embodiment.

The processor then evaluates the ‘value’ component of the conditional pre-fetch instruction 312 . If it is a non-zero value, which is also herein referred to as ‘TRUE’, the processor then pre-fetches the instructions at the location indicated by the ‘address’ component of the conditional pre-fetch instruction 314 . Any number of instructions may be pre-fetched in a given embodiment of the invention, although the number of instructions pre-fetched is preferably related to the size of the instruction cache and the architecture of the processor's pipeline.

If the ‘value’ component of the conditional pre-fetch instruction is zero, which is referred to herein as ‘FALSE’, the processor does not perform a pre-fetch operation. Instead, it preferably proceeds to increment its instruction pointer 308 and decode the next sequential instruction 302 .

The flowchart of FIG. 3 is simplified for exemplary purposes. For example, as is well known in the art, many processors are of the multiprocessing variety, wherein several instructions are in various stages of execution by the processor at any given time. An embodiment of the present invention also envisions the use of such multiprocessing processors. These embodiments generally use more complex instruction pipeline architectures that allow for several instruction to be in various stages of execution at each processor clock cycle.

Additionally, in one embodiment the processor itself manages the instruction pipeline and the instruction cache.

In another embodiment of the present invention, the instruction pipeline and instruction cache may be managed by hardware associated with the processor but not actually considered part of the processor. As mentioned above, it is possible that one processor may manage these components for another processor.

As an illustrative example of an embodiment of the invention, consider the following set of processor instructions

In this exemplary set of processor instructions, program execution will jump from the “Branch COND, L 1 ” instruction to the “Inst C 0 ” instruction at label ‘L 1 ’ if ‘COND’ is ‘TRUE’, or non-zero. Otherwise, program execution will proceed to the “Inst B 0 ” instruction first. If the instructions at the ‘L 1 ’ label are not in the processor's instruction buffer at the time the “Branch COND, L 1 ” instruction is executed, the processor may incur a branch penalty.

To further elaborate on this exemplary use of the conditional pre-fetch instruction, consider the above sample set of processor instructions at execution time, when the branch is taken and the instructions at ‘L 1 ’ are not present in the processor's instruction buffer. Exemplary clock cycles have been provided for further illustration:

As indicated, the condition pre-fetch operation may advantageously save a significant number of processor clock cycles when compared with a similar scenario that does not use a conditional pre-fetch, as described above.

In another example of using a conditional pre-fetch instruction in accordance with an embodiment of the present invention, use of the conditional pre-fetch by a “C” language programmer is envisioned. Although this example uses the “C” programming language, any programming language, including but not limited to any assembly language, any compiled language, such as “C”, “C++”, Cobol, Fortran, Ada, Pascal, etc., or any interpretive language, such as BASIC, JAVA, XML, or any other language may be used.

Using the “C” language loop as an example (in pseudocode):

A conditional branch is implicit at the closing bracket “}”, where the “/* conditional branch */” comment has been placed. Thus, each time ‘instruction . . . ’ is executed, the variable i is incremented and compared with n. If i is less than n, then the loop repeats. This is often implemented in compiled machine language as a conditional branch operation.

Using an embodiment of the present invention, the programmer may rewrite this “C” language loop thus:

This version of the “C” language adds an address label ‘L 1 ’ at the top of the ‘for’ loop, and a conditional pre-fetch instruction ‘CPIF’ at the outset of the loop. Note that the value to be evaluated for the conditional pre-fetch instruction is (i !=n−1). This expression evaluates to a non-zero value until the last iteration of the loop. Therefore, the instructions at address L 1 will be pre-fetched for each loop iteration except for the last loop iteration. In this manner, the loop may be advantageously processed and executed without the processor incurring any branch penalty.

In an embodiment of the invention, a 1-bit or 2-bit branch history is used together with a CPIF instruction to minimize branch penalties. For example, the branch history may be stored in one or two processor registers. Prior to the next conditional branch, a CPIF instruction may use the stored register values in its expression to be evaluated. A conventional branch history table only uses past history to determine when to pre-fetch instructions. One aspect of the invention, however, uses information contained in the branch history table as only one of the parameters to make a determination of a higher level. In this manner, the CPIF can incorporate the advantages of using 1-bit or 2-bit branch histories.

Most of the foregoing alternative embodiments are not mutually exclusive, but may be implemented in various combinations to achieve unique advantages. As these and other variations and combinations of the features discussed above can be utilized without departing from the invention as defined by the claims, the foregoing description of the embodiments should be taken by way of illustration rather than by way of limitation of the invention as defined by the claims.