WO2002057909A2 - Value speculation on an assist processor to facilitate prefetching for a primary processor - Google Patents

Value speculation on an assist processor to facilitate prefetching for a primary processor Download PDF

Info

Publication number
WO2002057909A2
WO2002057909A2 PCT/US2002/000643 US0200643W WO02057909A2 WO 2002057909 A2 WO2002057909 A2 WO 2002057909A2 US 0200643 W US0200643 W US 0200643W WO 02057909 A2 WO02057909 A2 WO 02057909A2
Authority
WO
Grant status
Application
Patent type
Prior art keywords
executable code
processor
long latency
latency operation
assist
Prior art date
Application number
PCT/US2002/000643
Other languages
French (fr)
Other versions
WO2002057909A3 (en )
Inventor
Shailender Chaudhry
Marc Tremblay
Original Assignee
Sun Microsystems, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3877Concurrent instruction execution, e.g. pipeline, look ahead using a slave processor, e.g. coprocessor
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3824Operand accessing
    • G06F9/383Operand prefetching
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3824Operand accessing
    • G06F9/383Operand prefetching
    • G06F9/3832Value prediction for operands; operand history buffers
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3836Instruction issuing, e.g. dynamic instruction scheduling, out of order instruction execution
    • G06F9/3842Speculative instruction execution

Abstract

One embodiment of the present invention provides a system that prefetches from memory by using an assist processor that performs data speculation and that executes in advance of a primary processor. The system operates by executing executable code on the primary processor while simultaneously executing a reduced version of the executable code on the assist processor. This allows the assist processor to generate the same pattern of memory references that the primary processor generates in advance of when the primary processor generates the memory references. While executing the reduced version of the executable code, the system predicts a data value returned by a long latency operation within the executable code. The system subsequently uses the predicted data value to continue executing the reduced version of the executable code without having to wait for the long latency operation to complete.

Description

VALUE SPECULATION ON AN ASSIST

PROCESSOR TO FACILITATE PREFETCHING

FOR A PRIMARY PROCESSOR

Inventor(s): Shailender Chaudhry and Marc Tremblay

BACKGROUND

Field of the Invention

The present invention relates to techniques for improving computer system performance. More specifically, the present invention relates to a method and an apparatus for performing value speculation on an assist processor in order to prefetch data and/or instructions from memory for subsequent use by a primary processor.

Related Art

As increasing semiconductor integration densities allow more transistors to be integrated onto a microprocessor chip, computer designers are investigating different methods of using these transistors to increase computer system performance. To this end, computer designers are beginning to incorporate multiple central processing units (CPUs) into a single semiconductor chip. This can result in performance gains for computational tasks that can be parallelized (divided) into separate pieces that can be concurrently executed.

Unfortunately, performance gains from parallelization can be limited for many applications that contain inherently serial portions of code. For these inherently serial portions of code, performance is further limited by memory latency problems.

Memory latency problems are growing progressively worse as processor clock speeds continue to improve at an exponential rate. At today's processor clock speeds, it can take as many as 200 processor clock cycles to pull a cache line in from main memory.

Computer designers presently use a number of techniques to decrease memory latency delays. (1) Out-of-order execution can be used to schedule loads and stores so that memory latency is hidden as much as possible. Unfortunately, out-of-order execution is typically limited to hiding a few clock cycles of memory latency. (2) A non-faulting load instruction can be used to speculatively load a data value, without causing a miss when the address is not valid. (3) A steering load instruction can be used to speculatively load a data value into L2 cache, but not LI cache, so that LI cache is not polluted by unused data values. Unfortunately, using non-faulting loads and steering loads can result in unnecessary loads. This wastes instruction cache space and ties up registers. (4) Some researchers have investigated using hardware prefetch engines, but these hardware prefetch engines are typically ineffective for irregular memory access patterns.

What is needed is a method and an apparatus that reduces memory latency delays in fast processor systems without the limitations and costs involved in using the above-discussed techniques.

SUMMARY

One embodiment of the present invention provides a system that prefetches from memory by using an assist processor that performs data speculation and that executes in advance of a primary processor. The system operates by executing executable code on the primary processor while simultaneously executing a reduced version of the executable code on the assist processor. This allows the assist processor to generate the same pattern of memory references that the primary processor generates in advance of when the primary processor generates the memory references. While executing the reduced version of the executable code, the system predicts a data value returned by a long latency operation within the executable code. The system subsequently uses the predicted data value to continue executing the reduced version of the executable code without having to wait for the long latency operation to complete. The system also stores results of memory references generated by the assist processor into a store that is shared with the primary processor so that the primary processor is able to access the results of the memory references. In one embodiment of the present invention, the system additionally executes the long latency operation while the assist processor continues executing the reduced version of the executable code. Next, the system compares a result of the long latency operation with the predicted data value. If the result of the long latency operation does not match the predicted data value, the system suspends further execution by the assist processor. In a variation on this embodiment, after suspending further execution by the assist processor, the system recommences execution by the assist processor using the result of the long latency operation, wherein execution is recommenced at a point prior to use of the predicted data value.

In one embodiment of the present invention, executing the reduced version of the executable code involves selecting loads to predict based upon: how far the assist processor is ahead of the primary processor; predictability of the data value returned by the long latency operation; or a likelihood of a cache miss while performing the long latency operation.

In one embodiment of the present invention, the long latency operation is either a load operation from memory or a computational operation requiring multiple clock cycles. In one embodiment of the present invention, predicting the data value returned by the long latency operation involves: predicting a value previously returned by the long latency operation; predicting a function of the value previously returned by the long latency operation; or predicting a default value if the long latency operation is being performed for a first time. In one embodiment of the present invention, prior to executing the executable code, the system compiles source code into the executable code for the primary processor. The system also produces the reduced version of the executable code for the assist processor from the executable code for the primary processor by eliminating instructions from the executable code that have no effect on a pattern of memory references generated by the executable code.

In one embodiment of the present invention, producing the reduced version of the executable code involves: converting store instructions into corresponding load instructions; eliminating redundant load instructions directed to previously loaded cache lines; and eliminating code that is used to calculate store values that are not subsequently used in determining address reference patterns for the executable code for the primary processor. In one embodiment of the present invention, the reduced version of the executable code is modified to speculatively execute code down a branch path that is more frequently taken if the reduced version of the executable code is determined to not be faster than the executable code.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 illustrates a computer system in accordance with an embodiment of the present invention.

FIG. 2 illustrates the compilation process in accordance with an embodiment of the present invention.

FIG. 3 is a flow chart illustrating the process of generating the reduced version of the executable code in accordance with an embodiment of the present invention.

FIG. 4 is a flow chart illustrating the process of concurrently executing the executable code and the reduced executable code in accordance with an embodiment of the present invention.

FIG. 5 is a flow chart illustrating operation of the assist processor in accordance with an embodiment of the present invention.

FIG. 6 is a flow chart illustrating the creation of reduced executable code that performs value prediction in accordance with an embodiment of the present invention. FIG. 7 is a flow chart illustrating the execution of reduced executable code that performs value prediction in accordance with an embodiment of the present invention.

FIG. 8 illustrates the structure of reduced executable code that performs value prediction in accordance with an embodiment of the present invention.

DETAILED DESCRIPTION

The following description is presented to enable any person skilled in the art to make and use the invention, and is provided in the context of a particular application and its requirements. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

The data structures and code described in this detailed description are typically stored on a computer readable storage medium, which may be any device or medium that can store code and/or data for use by a computer system. This includes, but is not limited to, magnetic and optical storage devices such as disk drives, magnetic tape,

CDs (compact discs) and DNDs (digital video discs), and computer instruction signals embodied in a transmission medium (with or without a carrier wave upon which the signals are modulated). For example, the transmission medium may include a communications network, such as the Internet.

Computer System

FIG. 1 illustrates a computer system in accordance with an embodiment of the present invention. This computer system includes a primary processor 102 and an assist processor 104, which share a common data cache 106. All of these structures reside on silicon die 100 (although in other embodiments of the present invention they can reside on multiple silicon dies). Note that processors 102 and 104 may generally include any type of computational devices.

Processors 102 and 104 include instruction caches 112 and 120, respectively, which contain instructions to be executed by processors 102 and 104.

Processors 102 and 104 additionally include load buffers 114 and 122 as well as store buffers 116 and 124 for buffering communications with data cache 106. More specifically, primary processor 102 includes load buffer 114 for buffering loads received from data cache 106, and store buffer 116 for buffering stores to data cache 106. Similarly, assist processor 104 includes load buffer 122 for buffering loads received from data cache 106, and store buffer 124 for buffering stores to data cache 106.

Processors 102 and 104 are additionally coupled together by one-way communication channels 128-129, which facilitates rapid communication between primary processor 102 and assist processor 104. Note that communication channel 128 allows primary processor 102 to write into register 126 within assist processor 104, and that communication channel 129 allows assist processor 104 to write into register 127 within primary processor 102.

Unlike using shared memory communication mechanisms, writing into register 126 (or 127) will not cause a cache miss. Furthermore, primary processor 102 does not have to wait until assist processor 104 receives the communication. Note that if primary processor 102 were to communicate with assist processor 104 through memory, the system would have to wait for store buffer 116 to be cleared in order to communicate.

In the embodiment illustrated in FIG. 1, data cache 106 is fully dual-ported, thereby allowing concurrent read and/or write accesses from processors 102 and 104. This dual porting eliminates cache coherence delays associated with conventional shared memory architectures that rely on coherent caches.

In one embodiment of the present invention, data cache 106 is a 16K-byte 4- way set-associative data cache with 32 byte cache lines. Data cache 106, instruction cache 112 and instruction cache 120 are coupled through switch 110 to memory controller 111. Memory controller 111 is coupled to dynamic random access memory (DRAM) 108, which is located off chip. Switch 110 may include any type of circuitry for switching signal lines. In one embodiment of the present invention, switch 110 is a cross bar switch. DRAM 108 contains executable code 130 for primary processor 102. During system operation executable code 130 is loaded through memory controller 111 and switch 110 into instruction cache 112 of primary processor 102. DRAM 108 also contains reduced executable code 132 for assist processor 104. Reduced executable code 132 is a reduced version of executable code 130 that generates the same pattern of memory references as executable code 130. During system operation, reduced executable code 132 is loaded through memory controller 111 and switch 110 into instruction cache 120 of assist processor 104. DRAM 108 additionally contains data 134 that is moved to and from data cache 106 through memory controller 111 and switch 110. Note that the present invention can be used in any computer system that includes multiple processors and is not limited to the illustrated computing system structure. Also note that in one embodiment of the present invention, assist processor

104 is used to make data accesses in advance of when the data is used by primary processor 102. In this embodiment, assist processor 104 does not have to actually perform the computations specified in a program; assist processor 104 merely has to perform sufficient computations to allow assist processor 104 to determine the memory access pattern of the program. Hence, assist processor 104 can be a much simpler version of primary processor 104, without circuitry that is not required to perform address computations, such as arithmetic circuitry.

Compilation Process

FIG. 2 illustrates the compilation process in accordance with an embodiment of the present invention. During the compilation process, source code 202 feeds through compiler 204 to produce executable code 130 for primary processor 102. ■ Executable code 130 then feeds through reduction module 208 to produce reduced executable code 132 for assist processor 104. Note that reduction module 208 may be part of compiler 204, or alternatively, may be separate from compiler 204.

In one embodiment of the present invention, the compilation and reduction processes take place before run-time. In another embodiment, the compilation and reduction processes take place during run-time, while the program is executing.

Generating Reduced Executable Code

FIG. 3 is a flow chart illustrating the process of generating reduced executable code 132 within reduction module 208 in accordance with an embodiment of the present invention. The system starts by profiling (simulating execution of) executable code 130 to identify hot spots in which load latency is causing delays (step 302).

Next, the system builds instruction traces of the hot spots in executable code 130 (step 304). In one embodiment of the present invention, the system constructs about 100 dispersed traces of about 2000 instructions each. The system then correlates this trace back to the source code (step 305). Next, the system categorizes traces into three types (step 306). In a type one trace, the address stream is not dependent on loaded data. For example, the pattern of data references in the code fragment below is not dependent on the value of the data loaded. for(i=0; i<100; i++) { rnd = generateRandomint(); index 1 = function 1 (rnd) ; index2 = function2(rnd); B [index l] Λ= A[index2];

} In a type two trace, the value of loaded data affects control flow only, but not the next address. In other words, if a load happens it will be to the next address.

However, due to control flow variations, the load may not happen. For example, in the code fragment below, if the hash table entry is found, the variable "found" will be set to I. Otherwise, the variable "found" is not modified.

while (hashTable[I] > 0) { if (hashTable[I] = entry) { found = I; break;

} I -= disp; if(I < 0) I+= hashTableSize; }

In a type three trace, the value loaded affects both the next address and control flow. For example, a piece of code that searches down a linked list is a type three trace because a pointer value loaded from a linked list node determines the address of the next node in the linked list.

Next the system modifies the traces to reduce the number of instructions in the traces while preserving the pattern of memory references generated by the traces. In doing so, the system presumes live variables are stored in registers. The first step is to change stores to loads (step 308). Note that the immediate effect of a store instruction is to pull a target cache line into data cache 106, which is the same as the effect of a load instruction. Hence, substituting a load instruction for a store instruction does not affect the memory reference pattern of the program.

This involves determining which load instructions are directed to the same cache line and getting rid of any load instructions that are directed to a cache line that has been previously loaded by a preceding load instruction. The third step is to eliminate dead code (step 312). This dead code includes code for generating values for store instructions that were changed to load instructions. For example, code on different sides of a branch often generate the same address streams but perform different calculations. In this case, when a store is changed to a load, the branch is not needed (nor is the compare to set the branch condition) and only code on one side of the branch is included in reduced executable code 132.

Also note that for type three traces, store instructions which store values that influence subsequent address calculations are not changed. These values are calculated and stored in registers to enable the system to follow subsequent address calculations.

Next, the system performs different actions for different types of traces. For a type one or type two trace, the system estimates if the reduced executable code 132 is faster than the original executable code 130 assuming that reduced executable code 132 will encounter only cache misses during load instructions, and assuming executable code 130 will encounter only cache hits (step 316). If reduced executable code 132 is faster, the process is complete. Otherwise, the system removes control flow instructions (e.g., branches) with a bias toward the most frequently taken branch paths as is determined during the profiling process (step 318). In doing so, reduced executable code 132 effectively speculates on the outcome of conditional branch instructions in order to gain more speed.

For type three traces, the system adds code to predict results of long latency operations, and modifies code to remove dependencies on long latency operations from reduced executable code 132 (step 319). This process is described in more detail below with reference to FIGs. 6-8. The system also adds code to both executable code 130 and reduced executable code 132 that allows assist processor 104 to detect if assist processor 104 is less than a minimum number of instructions ahead of primary processor 102, and if so to bail out of executing the trace (step 320). Note that communications from primary processor 102 to assist processor 104 take place through the one-way communication channel 128. Concurrent Execution

FIG. 4 is a flow chart illustrating the process of concurrently executing executable code 130 and reduced executable code 132 in accordance with an embodiment of the present invention. The system starts by executing executable code 130 on primary processor 102 (step 402). The system also concurrently executes reduced executable code 132 on assist processor 104 (step 404). When results of memory references generated by assist processor 104 return from memory, they are stored in data cache 106, which is shared with primary processor 102 (step 406). This allows primary processor 102 to reference the pre- fetched data values from data cache 106 without having to wait for the data values to be fetched from DRAM 108.

Operation of Assist Processor

FIG. 5 is a flow chart illustrating in more detail the operation of assist processor 104 in accordance with an embodiment of the present invention. Assist processor 104 first receives an interrupt from primary processor 102 (step 502). This causes assist processor 104 to examine a shared register value to determine which trace to execute (step 504). This implicitly presumes primary processor writes a trace identifier to the shared register, and then sends an interrupt to assist processor 104. Primary processor 102 attempts to send the interrupt to assist processor 104 as soon as all the live values of executable code 130 are known and are loaded into registers in primary processor 102. In one embodiment of the present invention, the a compiler within the system moves code around to calculate the live values as soon as possible.

After receiving the interrupt, assist processor 104 executes the designated trace (step 506). During trace execution, assist processor 104 keeps track of how far ahead of primary processor 102 it is (step 508). This is accomplished by primary processor 102 periodically sending progress indicators to assist processor 104 through communication channel 128.

If assist processor 104 is not a minimum number of instructions ahead of primary processor 102 during a type three trace, assist processor 104 bails out of the trace and waits for primary processor 102 to specify another trace (step 510). In some cases, this allows assist processor 104 to begin executing the next trace prior to primary processor 102 completing the previous trace.

Note that although the present invention is described in terms of a shared data cache 106, the present invention can generally be applied to any type of system in which processors share information retrieved from memory. In one embodiment of the present invention, processors 102 and 104 share an instruction cache. In another embodiment, processors 102 and 103 share a branch history table.

Also note that the term "shared cache" as used in this patent application can refer to systems in which processors 102 and 104 contain separate local caches containing mirrored copies of the same data. In this case, a load from memory into the local cache of assist processor 104 causes a corresponding load to the local cache of primary processor 102.

Value Prediction

FIG. 6 is a flow chart illustrating the creation of reduced executable code 132 that performs value prediction in accordance with an embodiment of the present invention. FIG. 6 illustrates in more detail the process carried out in step 319 of FIG. 3. In producing code to perform value prediction, the system starts by adding code to determine which long latency operations to predict (step 602). This determination can be based upon a number of factors, including how far assist processor 104 is ahead of the primary processor 102. Note that if assist processor 104 is many memory latencies ahead of primary processor 102, it is less advantageous to perform value prediction. This determination can also be based upon the predictability of the result of the long latency operation. Note that some long latency operations generate predictable results, such as loading a counter variable that is successively incremented in a predictable manner. Also note that it is more advantageous to attempt to predict results that are predictable. The determination can also be based on the likelihood of a cache miss being generated by the long latency operation. For example, a load from a successive memory location is unlikely to generate a cache miss, so performing value prediction for such a load operation tends to be less advantageous because the load is likely to return quickly. (Note that a long latency operation can include any operation that takes multiple clock cycles to accomplish and that returns a result. This can include a memory load operation, as well as a complex arithmetic operation, such as a square root operation.)

Next, the system adds code to predict results of the selected long latency operations (step 604). This code can use any known technique for performing value prediction, including simply predicting the last value returned by the long latency operation, or predicting a function of the last value returned by the long latency operation, such as the last value + 1.

The system also adds code to all execution paths within reduced executable code 132 that use predicted values in order to suspend execution if the predicted values do not match corresponding actual values returned by the long latency operations (step 606). This code can be added at multiple points in each relevant execution path to periodically check the actual values returned by the long latency operations. FIG. 7 is a flow chart illustrating the execution of reduced executable code

132 that performs value prediction in accordance with an embodiment of the present invention. The system starts by determining which long latency operation (or operations) to predict (step 702). Next, the system predicts the result of the long latency operation (step 704), and then uses the result to continue executing reduced executable code 132 without having to wait for the long latency operation to complete (step 706). At some time in the future, the system compares the actual result of the long latency operation with the predicted result (step 708). Assist processor 104 can receive the actual result from primary processor 102 through one-way communication channel 129. Alternatively, if the long latency operation is a load from memory, assist processor 104 can receive the actual result from a load operation to a temporary register previously executed by assist processor 104 itself (see FIG. 8).

If the actual result does not match the predicted result, assist processor 104 has been assuming an erroneous predicted result. At this point, the system can system can do a number of things (step 710). (1) The system can suspend continuing execution by assist processor 104. (2) The system can recommence execution by assist processor 104 using the result of the long latency operation, wherein execution is recommenced at a point prior to use of the predicted result. (3) Alternatively, the system can simply continue executing reduced executable code 132 without taking any corrective action.

Note that because assist processor 104 does not actually change the program state for primary processor 102, no corrective action is required. In the worst case, assist processor 104 merely prefetches a small number of cache entries incorrectly.

Operations performed by assist processor 104 do not affect the correctness of the execution of executable code 130 on primary processor 102.

FIG. 8 illustrates the structure of reduced executable code 132 that performs value prediction in accordance with an embodiment of the present invention. In this example, primary processor 102 executes a portion of executable code 130 which loads a value into R5. If this value is greater than or equal to zero, a first branch path is executed. Otherwise, a second branch path is executed.

Assist processor 104 executes a corresponding portion of reduced executable code 132 in advance of when primary processor executes the portion of executable code 130. This portion of reduced executable code 132 predicts a constant value to be returned by the load operation. This predicted constant is stored in R5. This portion of reduced executable code 132 also performs the load operation to a temporary register. Note that assist processor 104 does not have to wait for this load operation to return until the temporary register is subsequently accessed. Assist processor 104 then speculatively executes code for the first branch path. At some time in the future, assist processor 104 compares the predicted constant with the value returned by the load operation into the temporary register. If these values are not equal, the system can suspend execution, recommence execution, or do nothing, as is discussed above with reference to step 710 in FIG. 7. The foregoing descriptions of embodiments of the invention have been presented for purposes of illustration and description only. They are not intended to be exhaustive or to limit the invention to the forms disclosed. Accordingly, many modifications and variations will be apparent to practitioners skilled in the art. Additionally, the above disclosure is not intended to limit the invention. The scope of the invention is defined by the appended claims.

Claims

What Is Claimed Is:
1. A method for prefetching from memory by using an assist processor that performs data speculation and executes in advance of a primary processor, comprising: executing executable code on the primary processor; simultaneously executing a reduced version of the executable code on the assist processor, so that the assist processor generates the same pattern of memory references that the primary processor generates in advance of when the primary processor generates the memory references; wherein executing the reduced version of the executable code involves, predicting a data value returned by a long latency operation within the executable code; using the predicted data value to continue executing the reduced version of the executable code without having to wait for the long latency operation to complete; and storing results of memory references generated by the assist processor into a store that is shared with the primary processor so that the primary processor is able to access the results of the memory references.
2. The method of claim 1, further comprising: executing the long latency operation while the assist processor continues executing the reduced version of the executable code; comparing a result of the long latency operation with the predicted data value; and if the result of the long latency operation does not match the predicted data value, suspending further execution by the assist processor.
3. The method of claim 2, further comprising, after suspending further execution by the assist processor, recommencing execution by the assist processor using the result of the long latency operation, wherein execution is recommenced at a point prior to use of the predicted data value.
4. The method of claim 1, wherein executing the reduced version of the executable code involves selecting loads to predict based upon: how far the assist processor is ahead of the primary processor; predictability of the data value returned by the long latency operation; and a likelihood of a cache miss while performing the long latency operation.
5. The method of claim 1, wherein the long latency operation is one of: a load operation from memory; and a computational operation requiring multiple clock cycles.
6. The method of claim 1 , wherein predicting the data value returned by the long latency operation involves at least one of: predicting a value previously returned by the long latency operation; predicting a function of the value previously returned by the long latency operation; and predicting a default value if the long latency operation is being performed for a first time.
7. The method of claim 1, further comprising, prior to executing the executable code: compiling source code into the executable code for the primary processor; and producing the reduced version of the executable code for the assist processor from the executable code for the primary processor by eliminating instructions from the executable code that have no effect on a pattern of memory references generated by the executable code.
8. The method of claim 7, wherein producing the reduced version of the executable code involves: converting store instructions into corresponding load instructions; eliminating redundant load instructions directed to previously loaded cache lines; and eliminating code that is used to calculate store values that are not subsequently used in determining address reference patterns for the executable code for the primary processor.
9. The method of claim 1, wherein the reduced version of the executable code is modified to speculatively execute code down a branch path that is more frequently taken if the reduced version of the executable code is determined to not be faster than the executable code.
10. A computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method for prefetching from memory by using an assist processor that performs data speculation and executes in advance of a primary processor, the method comprising: executing executable code on the primary processor; simultaneously executing a reduced version of the executable code on the assist processor, so that the assist processor generates the same pattern of memory references that the primary processor generates in advance of when the primary processor generates the memory references; wherein executing the reduced version of the executable code involves, predicting a data value returned by a long latency operation within the executable code; using the predicted data value to continue executing the reduced version of the executable code without having to wait for the long latency operation to complete; and storing results of memory references generated by the assist processor into a store that is shared with the primary processor so that the primary processor is able to access the results of the memory references.
11. The computer-readable storage medium of claim 10, wherein the method further comprises: executing the long latency operation while the assist processor continues executing the reduced version of the executable code; comparing a result of the long latency operation with the predicted data value; and if the result of the long latency operation does not match the predicted data value, suspending further execution by the assist processor.
12. The computer-readable storage medium of claim 11, wherein after suspending further execution by the assist processor, the method further comprises, recommencing execution by the assist processor using the result of the long latency operation, wherein execution is recommenced at a point prior to use of the predicted data value.
13. The computer-readable storage medium of claim 10, wherein executing the reduced version of the executable code involves selecting loads to predict based upon: how far the assist processor is ahead of the primary processor; predictability of the data value returned by the long latency operation; and a likelihood of a cache miss while performing the long latency operation.
14. The computer-readable storage medium of claim 10, wherein the long latency operation is one of: a load operation from memory; and a computational operation requiring multiple clock cycles.
15. The computer-readable storage medium of claim 10, wherein predicting the data value returned by the long latency operation involves at least one of: predicting a value previously returned by the long latency operation; predicting a function of the value previously returned by the long latency operation; and predicting a default value if the long latency operation is being performed for a first time.
16. The computer-readable storage medium of claim 10, wherein prior to executing the executable code, the method further comprises: compiling source code into the executable code for the primary processor; and producing the reduced version of the executable code for the assist processor from the executable code for the primary processor by eliminating instructions from the executable code that have no effect on a pattern of memory references generated by the executable code.
17. The computer-readable storage medium of claim 16, wherein producing the reduced version of the executable code involves: converting store instructions into corresponding load instructions; eliminating redundant load instructions directed to previously loaded cache lines; and eliminating code that is used to calculate store values that are not subsequently used in determining address reference patterns for the executable code for the primary processor.
18. The computer-readable storage medium of claim 10, wherein the reduced version of the executable code is modified to speculatively execute code down a branch path that is more frequently taken if the reduced version of the executable code is determined to not be faster than the executable code.
19. An apparatus that facilitates prefetching from memory by using an assist processor that performs data speculation and executes in advance of a primary processor, comprising: the primary processor that is configured to execute executable code; the assist processor that is configured to simultaneously execute a reduced version of the executable code, so that the assist processor generates the same pattern of memory references that the primary processor generates in advance of when the primary processor generates the memory references; wherein the reduced version of the executable code includes, a prediction mechanism that predicts a data value returned by a long latency operation within the executable code, and an execution mechanism that uses the predicted data value to continue executing the reduced version of the executable code without having to wait for the long latency operation to complete; and a store that is configured to store results of memory references generated by the assist processor; wherein the store is configured to be shared between the assist processor and the primary processor, so that the primary processor is able to access the results of the memory references generated by the assist processor.
20. The apparatus of claim 19, wherein the apparatus is configured to: execute the long latency operation while the assist processor continues executing the reduced version of the executable code; compare a result of the long latency operation with the predicted data value; and to suspend further execution by the assist processor, if the result of the long latency operation does not match the predicted data value.
21. The apparatus of claim 20, wherein the execution mechanism is further configured to recommence execution by the assist processor using the result of the long latency operation after further execution by the assist processor is suspended, wherein execution is recommenced at a point prior to use of the predicted data value.
22. The apparatus of claim 19, wherein the execution mechanism is configured to select loads to predict based upon: how far the assist processor is ahead of the primary processor; predictability of the data value returned by the long latency operation; and a likelihood of a cache miss while performing the long latency operation.
23. The apparatus of claim 19, wherein the long latency operation is one of: a load operation from memory; and a computational operation requiring multiple clock cycles.
24. The apparatus of claim 19, wherein the prediction mechanism is configured to predict at least one of: a value previously returned by the long latency operation; a function of the value previously returned by the long latency operation; and a default value if the long latency operation is being performed for a first time.
25. The apparatus of claim 19, further comprising a compilation mechanism that is configured to: compile source code into the executable code for the primary processor; and to produce the reduced version of the executable code for the assist processor from the executable code for the primary processor by eliminating instructions from the executable code that have no effect on a pattern of memory references generated by the executable code.
26. The apparatus of claim 25, wherein producing the reduced version of the executable code involves: converting store instructions into corresponding load instructions; eliminating redundant load instructions directed to previously loaded cache lines; and eliminating code that is used to calculate store values that are not subsequently used in determining address reference patterns for the executable code.
27. The apparatus of claim 19, wherein the reduced version of the executable code is modified to speculatively execute code down a branch path that is more frequently taken if the reduced version of the executable code is determined to be not faster than the executable code.
28. The apparatus of claim 19, wherein the assist processor is a simplified version of the primary processor, excluding circuitry that is not required to perform address computations.
29. The apparatus of claim 19, wherein the primary processor and the assist processor reside on the same semiconductor chip.
30. The apparatus of claim 19, wherein the primary processor and the assist processor reside on distinct semiconductor chips.
PCT/US2002/000643 2000-05-04 2002-01-09 Value speculation on an assist processor to facilitate prefetching for a primary processor WO2002057909A3 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US09/761,360 2001-01-16
US09761360 US6772321B2 (en) 2000-05-04 2001-01-16 Method and apparatus for using an assist processor and value speculation to facilitate prefetching for a primary processor

Publications (2)

Publication Number Publication Date
WO2002057909A2 true true WO2002057909A2 (en) 2002-07-25
WO2002057909A3 true WO2002057909A3 (en) 2003-08-07

Family

ID=25061970

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2002/000643 WO2002057909A3 (en) 2000-05-04 2002-01-09 Value speculation on an assist processor to facilitate prefetching for a primary processor

Country Status (1)

Country Link
WO (1) WO2002057909A3 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140053056A1 (en) * 2012-08-16 2014-02-20 Qualcomm Incorporated Pre-processing of scripts in web browsers
US9898446B2 (en) 2012-08-16 2018-02-20 Qualcomm Incorporated Processing a webpage by predicting the usage of document resources
US9898445B2 (en) 2012-08-16 2018-02-20 Qualcomm Incorporated Resource prefetching via sandboxed execution

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1996020440A1 (en) * 1994-12-23 1996-07-04 The Victoria University Of Manchester Dual processor decoupling
US5996060A (en) * 1997-09-25 1999-11-30 Technion Research And Development Foundation Ltd. System and method for concurrent processing
WO2001052061A2 (en) * 2000-01-14 2001-07-19 Sun Microsystems, Inc. Method and apparatus for using an assist processor to pre-fetch data values for a primary processor
WO2001093029A2 (en) * 2000-05-31 2001-12-06 Sun Microsystems, Inc. Speculative program execution with value prediction

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1996020440A1 (en) * 1994-12-23 1996-07-04 The Victoria University Of Manchester Dual processor decoupling
US5996060A (en) * 1997-09-25 1999-11-30 Technion Research And Development Foundation Ltd. System and method for concurrent processing
WO2001052061A2 (en) * 2000-01-14 2001-07-19 Sun Microsystems, Inc. Method and apparatus for using an assist processor to pre-fetch data values for a primary processor
WO2001093029A2 (en) * 2000-05-31 2001-12-06 Sun Microsystems, Inc. Speculative program execution with value prediction

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
SAKALAY F E: "STORAGE HIERARCHY CONTROL SYSTEM" IBM TECHNICAL DISCLOSURE BULLETIN, IBM CORP. NEW YORK, US, vol. 15, no. 4, 1 September 1972 (1972-09-01), pages 1100-1101, XP002002415 ISSN: 0018-8689 *
SATHE R ET AL: "AVAILABLE PARALLELISM WITH DATA VALUE PREDICTION" PROCEEDINGS. INTERNATIONAL CONFERENCE ON HIGH PERFORMANCE COMPUTING, XX, XX, 17 December 1998 (1998-12-17), pages 194-201, XP001001139 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140053056A1 (en) * 2012-08-16 2014-02-20 Qualcomm Incorporated Pre-processing of scripts in web browsers
US9898446B2 (en) 2012-08-16 2018-02-20 Qualcomm Incorporated Processing a webpage by predicting the usage of document resources
US9898445B2 (en) 2012-08-16 2018-02-20 Qualcomm Incorporated Resource prefetching via sandboxed execution

Also Published As

Publication number Publication date Type
WO2002057909A3 (en) 2003-08-07 application

Similar Documents

Publication Publication Date Title
US6718440B2 (en) Memory access latency hiding with hint buffer
US5446850A (en) Cross-cache-line compounding algorithm for scism processors
US5349651A (en) System for translation of virtual to physical addresses by operating memory management processor for calculating location of physical address in memory concurrently with cache comparing virtual addresses for translation
US5761468A (en) Hardware mechanism for optimizing instruction and data prefetching by forming augmented prefetch instructions
US6785772B2 (en) Data prefetching apparatus in a data processing system and method therefor
US5765208A (en) Method of speculatively executing store instructions prior to performing snoop operations
US6684398B2 (en) Monitor entry and exit for a speculative thread during space and time dimensional execution
US6574712B1 (en) Software prefetch system and method for predetermining amount of streamed data
US7493452B2 (en) Method to efficiently prefetch and batch compiler-assisted software cache accesses
US7681188B1 (en) Locked prefetch scheduling in general cyclic regions
US6230260B1 (en) Circuit arrangement and method of speculative instruction execution utilizing instruction history caching
US20040268051A1 (en) Program-directed cache prefetching for media processors
US6308261B1 (en) Computer system having an instruction for probing memory latency
US6353881B1 (en) Supporting space-time dimensional program execution by selectively versioning memory updates
US5953512A (en) Microprocessor circuits, systems, and methods implementing a loop and/or stride predicting load target buffer
US6055622A (en) Global stride prefetching apparatus and method for a high-performance processor
US5694568A (en) Prefetch system applicable to complex memory access schemes
US20050102294A1 (en) Method for prefetching recursive data structure traversals
US20090132796A1 (en) Polling using reservation mechanism
US6928645B2 (en) Software-based speculative pre-computation and multithreading
US4755966A (en) Bidirectional branch prediction and optimization
US6684319B1 (en) System for efficient operation of a very long instruction word digital signal processor
US6918010B1 (en) Method and system for prefetching data
US6128703A (en) Method and apparatus for memory prefetch operation of volatile non-coherent data
US5615386A (en) Computer architecture for reducing delays due to branch instructions

Legal Events

Date Code Title Description
AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG

AK Designated states

Kind code of ref document: A2

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ OM PH PL PT RO RU SD SE SG SI SK SL TJ TM TN TR TT TZ UA UG UZ VN YU ZA ZM ZW

121 Ep: the epo has been informed by wipo that ep was designated in this application
REG Reference to national code

Ref country code: DE

Ref legal event code: 8642

122 Ep: pct application non-entry in european phase
WWW Wipo information: withdrawn in national office

Country of ref document: JP

NENP Non-entry into the national phase in:

Ref country code: JP