GB2343032A

GB2343032A - Minimizing exception traps from a top-of-stack cache

Info

Publication number: GB2343032A
Application number: GB9917120A
Authority: GB
Inventors: Peter C Damron
Original assignee: Sun Microsystems Inc
Current assignee: Sun Microsystems Inc
Priority date: 1998-07-30
Filing date: 1999-07-21
Publication date: 2000-04-26
Anticipated expiration: 2019-07-21
Also published as: DE19935378A1; GB2343032B; CA2277636A1; GB9917120D0

Abstract

Reducing the number of overflow and underflow exception traps generated during the execution of a program on a computer that uses a top-of-stack cache, e.g. in a register window file architecture whereby the top of the stack is maintained in registers directly accessible by the CPU and the rest of the stack is maintained in memory. The number of stack elements spilled or filled from/to the top-of-stack cache to/from memory, in response to an overflow or underflow trap respectively, is controlled by a predictor value. The predictor value reflects the history of the trap exceptions, in order to minimize future exceptions. Different predictor values can select different overflow and underflow vectors (Fig. 4), or a hash mechanism (Fig. 6) enables multiple predictors to separately control the spill/fill of the stack file dependant on where in memory the exceptions occur. Further, an exception history can be maintained (Fig. 7) and hashed with the address of the computer instruction that caused the exception, to generate an index into a set of predictors.

Description

2343032 Page: I Reld of the Invention This invention relates to the field

of computer architecture. Specifically, this invention is a method, apparatus and computer program product for reducing overflow and io underflow exception traps resulting from the use of a stack file.

Background

Some computer architectures include a windowed register file (the register window file) that contains a set of registers. A register window is a selection of registers that are available to a program at a particular momem Register windows are used to improve is performance of subroutine calls (arnong other reasons). The SPARC" architecture is an example of a computer architecture that uses a register window file.

A register window file is one example of a stack file. A "stack file" consists of a stack structure that is partially stored in memory and partially stored in a register file for faster access. The "top-of-stack cache" refers to the registers of the stack file. Stack elements are moved from the registers to memory and back again based on stack underflow and overflow exception traps.

Generally, on entering a subroutine, the subroutine invokes a "save" instruction (or similar instruction). The' 'save' instruction provides the subroutine with a new register window. If the register window file does not contain enough registers to hold the new register window, the "save" instruction causes an overflow trap. When the register window file overflows, the registers in at least one older register window in the register window file must be saved (spilled) to memory to release registers for a new register window. In a similar Page: 2 manner the register window file underflows when register window file is empty and the computer executes an instruction (such as a '.'restore" instruction) to restore a previously saved register window.

In both of these circumstances, the computer generates a register window exception trap. This stack exception trap invokes a trap handler that executes instructions to handle the exception. For a register window overflow condition, the contents of one or more existing register windows are spilled (saved) to memory. For a register window underflow condition, the previously saved contents of one or more register windows are restored to the stack file (a fill).

Prior art operating systems spill and fill a fixed number of register windows at each register window exception trap (often the trap only affects a single register window). This is inefficient when there are deeply nested or recursive subroutine calls. Historically, a single fill_ or spill was considered appropriate because most traditional programming methodologies did not generate deep subroutine call chains. Modem pro i g methodologies (in particular object-oriented programs, and programs that use recursion) often generate deep call chains. For these programming methodologies, window exceptions would be reduced if more than one spill/fill occurred on a register window exception trap. However, the program mix on most computer systems includes some programs that use the traditional methodology and other programs that use the modem methodology. In addition, a single program often includes both methodologies. Thus, simply spilling or filling a fixed number of register windows does not improve the overall system efficiency.

The register window file is but one example of the use of a "top-of-stack cache." Another example is illustrated by the Intel@ floating point unit (FPU) architecture. Yet another example is the hardware stack provided by a stack oriented computer architecture (such as a Forth Computer). Still another example is a return address top- of-stack cache (such as those used in some Forth computer architectures). Each of these hardware stacks are a form of a "top-of-stack cache" in that the top of the stack can be maintained in registers directly accessible by the CPU and the rest of the stack can be maintained in memory and that each stack can invoke an overflow trap or an underflow trap.

Page: 3 It would be advantageous to provide a spill/fill. handier for top-of- stack caches that adapt to the characteristics of the currently executing program. This adaptation would improve the efficiency of the computer system by reducing the number of top-of-stack cache exceptions generated by the executing program.

SUMMary of the Invention

The present invention improves the performance of a computer (that includes a top-ofstack cache) by adjusting the contents of the top-ofstack cache dependant on a predictor that contains information about the past usage of the top-of-stack cache. One aspect of the invention is a computer controfled method for selecting a predictor from a set of predictors to minimize exceptions in a computer having a memory and a top-of-stack cache. The method initializes an exception history. The exception history is used to track occurrences of a plurality of exception traps from the top-of-stack cache. Ile method also invokes an exception trap and updates the exception history dependent on the exception trap that was invoked. Ile method also includes the step of selecting the predictor from the set of predictors based on the exception history. Once the predictor is selected, the method processes the exception trap depen6n"I on the predictor.

Another aspect of the invention is an apparatus, that has a central processing unit (CPU) and a memory coupled to the CPU, for selecting a predictor from a set of predictors to minimize exceptions in a computer having a memory and a top-of-stack cache. The apparatus includes an initialization mechanism that is configured to initialize an exception history. The exception history is used to track occurrences of a plurality of exception traps from the top-ofstack cache. The apparatus also includes a trap mechanism that is configured to invoke an exception trap. In addition, the apparatus includes a history tracking mechanism that is configured to update the exception history dependent on the exception trap. A predictor selection mechanism is configured to select the predictor from the set of predictors based on the exception history. In addition, a trap handler mechanism, in the computer, is configured to process the exception trap dependent on the predictor.

Yet a further aspect of the invention is a computer program product embodied on a computer usable medium for causing a computer to select a predictor from a set of predictors Page: 4 to minin-Aze exceptions in a computer having a memory and a top-of-stack cache. When executed on a computer, the computer readable code causes the computer to effect an initialization mechanism, a trap mechanism, a history tracking mechanism, a predictor selection mechanism and a trap handler mechanism. Each of these mechanisms having the s same functions as the corresponding mechanisms for the previously described apparatus.

Another aspect of the invention is a computer data signal embodied in a carrier wave having computer readable code embodied therein for causing a computer to select a predictor from a set of predictors to minimize exceptions in a computer having a memory and a top-ofstack cache. When executed on a computer, the computer readable code causes the computer lo to effect an initialization mechanism 4_ trap mechanism, a history tracking mechanism, a predictor selection mechanism and a trap handler mechanism. Each of these mechanisms having the same functions as the corresponding mechanisms for the previously described apparatus.

Yet another aspect of the invention is a computer controlled method for minimizing 15, exceptions in a computer having a memory and a return address top-of- stack cache. The method includes the step of initializing a predictor for tracking exceptions from the return address top-of-stack cache. The method also invokes an exception trap that is processed dependent on the predictor. In addition, the method includes the step of changing the predictor responsive to the exception trap.

Still another aspect of the invention is an apparatus, that has a central processing unit (CPU) and a memory coupled to the CPU, for minimizing exceptions in a computer having a return address top-of-stack cache. The apparatus includes a predictor initialization mechanism that is configured to initialize a predictor. The predictor is used for tracking exceptions from the return address top-of-stack cache. The apparatus also includes a trap mechanism, that is configured to invoke an exception trap in the computer, and a trap handler mechanism, in the computer, that is configured to process the exception trap dependent on the predictor. In addition the apparatus includes a predictor maintenance mechanism that is configured to change the predictor responsive to the exception trap.

Page: 5 Yet a further aspect of the invention is a computer program product embodied on a computer usable medium for causing a co ' mputer to minimize exceptions. The computer includes a return address top-of-stack,ca.che. When executed on a computer, the computer readable code causes the computer to effect a predictor initialization mechanism, a trap mechanism, a trap handler mechanism and a predictor maintenance mechanism. Each of these mechanisms having the same functions as the corresponding mechanisms for the previously described apparatus.

The foregoing and many other aspects of the present invention will become obvious to those of ordinary skill in the art after having read the following detailed description of lo preferred embodiments that am illustrated in the various drawing figures.

Description of the Dmw-Ings

Fig. I illustrates a computer system capable of using the invention mi accordance with, a preferred embodiment; Fig. 2 illustrates an overview of the stack exception handling process in accordance with a preferred embodiment; Fig. 3A illustrates the operation of a stack overflow trap handler process used in Fig. 2; Fig. 3B illustrates the operation of a stack underflow trap handler process used in Fig. 2; Fig. 4 illustrates a predictor dependent vector in accordance with a preferred embodiment; Fig. 5 illustrates an 'adaptive stack use adjustment' process in accordance with a preferred embodiment; Fig. 6A illustrates a hash mechanism used in accordance with a preferred embodiment; Page: 6 Fig. 6B illustrates a predictor selection process used in accordance with a preferred embodiment; Fig. 7A illustrates components of a predictor selection process used in accordance with a preferred embodiment; Fig. 7B illustrates a predictor selection process used in accordance with a preferred embodiment; and Fig. 7C illustrates a exception history maintenance process used in accordance with a preferred embodiment.

Nscription of the Preferred Fmbodiments lo Operating Environment One example of a computer architecture that uses a register window file is the SPARC architecture. This architecture is defined in The SPARC Architecture Manual.

Version 9, Weaver & Germond, Q 1994 SPARC International, Inc., ISBN 0- 13099227-5.

The following sections of this document are related to register window files and are is incorporated by reference in their entirety: 2, 3.2.7, 3.3, 5.1.1-5. 1.4, 5.2.10-5.2.11, 6.3.6. 1 6.3.6.4,6.4,7.5.2.1, A.8, A.21, A.24, A.44 46, H.

Intelo processors use a register stack for floating point operations that can be organized as a top-of-stack cache. This register stack organization is described by Chapter 7 of Intel Architecture Software Developer's Manual: Basic Architecture, order number 243190, @1996, 1997, hereby incorporated by reference in its entirety.

A general register top-of-stack cache is also used in a microprocessor designed to directly execute Forth programs. This device is described by An Architecturefor the Direct Execution of the Forth Programming Language, by Hayes, Fraeman, Williams and Zaremba, 0 1987 ACM 0-89791-2381/87/1000-0042, Proceedings Second International Conference on Architectural Support for Programming Languages and Operating Systems, October 1987, hereby incorporated by reference in its entirety.

Pap: 7 Branch prediction technology improves performance in pipelined computers. This technology can be applied to minimizing exception traps resulting from overflow and underflow conditions of a top-of-stack cache. An overview of branch prediction approaches is provided by A study of Branch Prediction Strategies, by James E. Smith, 0 1981 by The Institute of Electrical and Electronics Engineers, Inc., New York, N.Y. IEEE order number: 8 1 CH 1593-3 hereby incorporated by reference in its entirety.

Fig. 1 illustrates a computer, indicated by general reference character 100, that incorporates the invention. The computer 100 includes a processor 101 that incorporates a central processor unit (CPU) 103, a memory section 105 and an input/output (1/0) section io 107. The 1(0 section 107 is connected to a keyboard 109, a display unit 111, a disk storage unit 113 and a CD-ROM drive unit 115. The CD-ROM drive unit 115 can read a CD-ROM medium 117 that typically contains a program and data 119. The CD-ROM drive unit 115 (along with the CD-ROM medium 117 and the disk storage unit 113 comprise a filestorage mechanism. Some embodiments of the invention include a network interface 121 that is connects the computer 100 to a network 123. The processor 101 includes the capability of generating exception traps that are handled by a trap handler 125. The trap handler 125 is generally (although not always) part of an operating system 127 within the memory section 105. An application program 129 generally makes system calls to the operating system 127 to perform services and provide resources. The CPU 103 includes a 'top-of-stack cache' 131.

In the case of computers conforming to the SPARC architecture the 'top-ofstack cache' 131 is a register file that is organized into register windows. One skilled in the art will understand that not all of the displayed features of the computer 100 need to be present to practice the invention.

Much of the invention is described with reference to the register window file of the SPARC architecture. One skilled in tlfe art will understand how to apply the invention as described with respect to the register file to other top- of-stack cache architectures. One example architecture uses a general purpose register file that invokes a general purpose register file exception trap when the general purpose top-of-stack cache overflows or underflows. Another example architecture includes a floating point top-of- stack cache capable of generating a floating point register file exception trap when the floating point top- Page: 8 of-stack cache overflows or underflows. Yet another example architecture includes a return address top-of-stack cache (similar to those used in some Forth computer architectures).

Fig. 2 illustrates a 'stack exception handling' process, indicated by general reference character 200, for minimizing the number of overflow and underflow exceptions. The 'stack exception handling' process 200 initiates at a 'start' terminal 201 and continues to an initialization' procedure 203. A procedure is any sequence of instructions that can be executed by a computer. The 'initialization' procedure 203 initializes a predictor, initializes stack element management values 'associated with the predictor, and initializes the top-of stack cache overflow and underflow trap vectors - After some period of operation the top-of-stack cache overflows (or, at a later time, the top-of-stack cache may underflow) invoking a stack exception an through one of the trap vectors initialized by the 'initialization' procedure 203. A 'receive stack trap' event 205 receives the trap, saves the computer's state as trap information and causes the CPU 103 to execute computer instructions residing in the trap handler W. The trap handler M includes a 'process stack trap' procedure 207. The 'process stack trap' procedure 207 processes the stack exception trap to spill or fill one or more stack elements (such as register windows) depending on the predictor. Next, the 4 process stack trap' procedure 207 changes the predictor depending on whether the stack exception trap was caused by a stack overflow or underflow. When the next stack exception trap occurs, the 'stack exception handling' process 2D 200 repeats starting at the 'receive stack UV' event 205. One skilled in the art will understand that the saved trap information is used to restore the computer's state (generally to re-execute the trapped instruction) when the trap returns.

One of the advantages of the invention is that it determines the number of stack elements that are to be spilled/filled dependent on the spill/fill history reflected by the value of the predictor. Thus, if more overflow traps than underflow traps have occurred, more stack elements can be spilled to memory on each overflow trap, For register window stacks the stack elements are register windows. This aspect of the invention becomes very important with modem programming methodologies that generate deep call sequences (for example, object-oriented and/or recursive programming).

1.

Page: 9 The stack elements in other t6pof-stack cache architectures need not be register windows. For example, the Intel FPU contains a register stack having stack elements that are floating point registers while the Forth architecture contains general purpose registers. The invention operates on these stack files to spill or fill a specified number of registers (stack s elements) to or from memory at every overflow or underflow exception trap dependent on a predictor.

Fig. 3A illustrates a stack overflow trap handler process, indicated by general reference character 300, configured to process a stack overflow trap. One skilled in the art will understand that some computer architectures have only one trap vector associated with lo both register window overflow and underflow conditions. Other computer architectures have separate vectors associated with the regisier window underflow and overflow conditions. Still other computer architectures incorporate multiple vectors for both overflow traps and underflow traps. Regardless of the actual trap details, a stack exception trap handler similar to the stack overflow trap handler process 300 is invoked on an overflow trap. For the purposes of the following discussion, the term "stack exception trap" encompasses both a stack underflow trap and a stack overflow trap.

The stack overflow trap handler process 300 initiates at a 'start' terminal 301 after the 'initialization' procedure 203 executes. Once a stack overflow trap occurs, the stack overflow trap handler process 300 continues to a 'receive stack overflow' event 303 that is invoked as a direct or indirect result of the occurrence of a stack overflow trap. Next, the stack overflow trap handler process 300 determines how many stack elements to spill to memory based on the predictor. Because the trap is a stack overflow trap, the top-of-stack cache is full. Thus, although the stack overflow trap handler process 300 must spill at least one stack element, it can spill more than one stack element to make additional room in the top-of-stack cache.

Thus, for a register window file the process can make room for anticipated execution of 66 save" instructions. A 'determine amount of stack to spill' procedure 305 determines how many stack elements to spill based on the predictor.

In one preferred embodimem the predictor is a two-bit variable that is used to select a spill value from a table of stack element management values. For example, Table I illustrates one example of stack element management values indexed by a two-bit predictor.

Page: 10 Table I

Thus, assuming that the predictor is initially set to zero, the first stack overflow trap spills only one stack element. A second or third stack overflow trap without an intervening stack underflow trap will spill two stack elements. A fourth trap (and all subsequent stack overflow traps that do not have an intervening stack underflow trap) will spill three stack elements. However, as is subsequently described, each stack underflow trap will decrement the predictor and potentially change the number of stack elements that are spilled to memory.

One skilled in the art will understand that for register window top-ofstack caches, the number of stack elements is the number of register windows to fill/spill. For other top-of stack caches the stack element management values can determine the number of stack elements to spill from or fill to the top-of-stack cache.

Once the spill value is selected by the 'determine amount of stack to spill' procedure 305, the stack overflow trap handler process 300 continues to a 'spill stack amount' procedure 307 that spills the number of register windows to memory specified by the spill value.

Next the stack overflow trap handler process 300 continues to a 'predictor less than is maximum? decision procedure 309 that determines whether the predictor is less than its maximum. If so, the stack overflow trap handler process 300 continues to an 'increment predictor' procedure 311 that changes the predictor so that the next stack exception trap will use the larger predictor. Once the predictor is incremented, or if the predictor is already at its maximum, the stack overflow trap handler process 300 continues to a 'return from trap' procedure 313. The 'return from trap' procedure 313 returns to the program that attempted to execute an instruction that caused the stack overflow. In the case of a register file, this instruction is a "save" instruction. Thus, the 'return from trap' procedure 313 re-executes the Predictor Spill Fill 00 1 3 01 2 2 2 2 I 1 3 1 Page: 11 66save" instruction. Because the top-of-stack cache now has room for a register window, the 66save" instruction succeeds and the program continues execution. 1"he stack overflow trap handler process 300 then completes through an 'end' terminal 315.

One skilled in the art will understand that the predictor can be of any size, from a single bit to many bits depending on the amount of information needed to anticipate the next stack exception trap. Further, one skilled in the art wW understand that the invention need not decrement or incremeftt the predictor. Instead, one preferred embodiment stores a state value in the predictor and changes the state value dependent on the existing state and whether an overflow or underflow trap occurs.

Fig. 3B i1lustrates a stack un&rflow trap handler process, indicated by general reference character 350 configured toprocess a stack underflow trap. As previously described with respect to Fig. 3A, there are many ways to invoke the stack underflow trap handler process 350. The stack underflow trap handler process 350 initiates at a 'start' terminal 351 after the 'initialization' procedure 203 executes. Once a stack underflow trap occurs, the stack underflow trap handler process 350 continues to a 'receive stack underflow' event 353 that is invoked as a direct or indirect result of the occurrence of a stack underflow trap. Next, the stack underflow trap handler process 350 continues to a 'determine amount of stack to fill' procedure 355 that determines how many register windows to fill from memory based on the predictor.

As was previously described, m- one preferred embodiment, the predictor is a two-bit variable that is used to select a fill value from a table of stack element management values similar to that shown in Table 1. Because the trap is a stack underflow trap, the top-of-stack cache is empty. Thus, although the stack underflow trap handler process 350 must fill at least one stack element it can fill more than one stack element to preload the top-of-stack cache.

In the case of a register window, the register file can be filled in anticipation of the execution of more "restore" instructions than "save" instructions.

Once the fill value is selected by the 'determine amount of stack to fill' procedure 355, the stack underflow trap handler process 350 continues to a 'fill stack amount' procedure 357 that fills the number of stack elements from memory specified by the fill value.

Page: 12 Next, the stack underflow trap handler process 350 continues to a 'predictor greater than minimum' decision procedure 359 that determines whether the predictor is greater than its minimum value. If so, the stack underflow trap handler process 350 continues to a 'decrement predictor' procedure 361 that decrements the predictor so that the next stack exception trap will use the new predictor. Once the predictor is decremented, or if the predictor is already at its minimum, the stack underflow trap handler process 350 continues to a 'return from trap' procedure 363. In the case of a register window underflow trap, the trap returns to the program that attempted to execute a "restore" instruction and re-executes the "restore" instruction. Because the register file now contains a register window, the "restore" io instruction succeeds and the program continues execution. The stack underflow trap handler process 350 then completes through an 'end' terminal 365.

In one embodiment the stack overflow trap handler process 300 and the stack underflow trap handler process 350 reside within the operating system and execute in a privileged enviroranent. In another embodiment, the stack overflow trap handler process 300 and the stack underflow trap handler process 350 reside in an application and execute within a protected environment. In this case, the trap is generally first vectored to program instructions in the operating system that re-directs the trap to execute the register window handlers in the application.

One skilled in the art will understand that the stack element management values indicated in Table I are only one possible set of values. The optimum set of values will depend on the number of stack elements in the top-of-stack cache and the characteristics of the types of programs that are executed by the computer. Such a one will also understand that the predictor represents a state. Thus, the invention contemplates storing particular values in the predictor instead of incrementing or decremienting the predictor.

Fig. 4 illustrates a stack exception trap vector architecture, indicated by general reference character 400 that can be used to effectuate a preferred embodiment. A predictor register 401 contains the value of the predictor. The value contained in the predictor register 401 is used to select which overflow vector, in a stack overflow vector array 403, and which underflow vector, in a stack underflow vector array 405, to use when a stack exception trap occurs. For example, the current value in the predictor register 401 selects a selected Page: 13 overflow vector 407 on an overflow exception. The selected overflow vector 407 causes the computer to execute instructions in a 'spill I stack element' handler 409. The 'spill I stack element' handler 409 contains instructions that spill one stack element (for example, a single register or one register window) and that increment the value in the predictor register 401 (up to a maximum). However, the current value in the predictor register 401 also specifies a selected underflow vector 411 that causes the computer to execute instructions in a 'fill 3 stack elements' handler 413. The 'fiIl 3 stack elements' handler 413 contains instructions that will fill three stack elements and that decrements the value in the predictor register 401 (down to a minimum). Thus, as the value in4he predictor register 401 changes (due to stack io exception traps) different spill/fill handlers are selected by specifying which trap vectors in the vector arrays are selected.

Fig. 5 illustrates an 'adaptive stack use adjustment' process, indicated by general reference character 500, for adaptively changing the stack element management values responsive to information gathered about the performance characteristics of a particular executing program.

The 'adaptive stack use adjustment' process 500 initiates at a 'start' terminal 501 and continues to an 'initialize stack element managementvalue' procedure 503. The 'initialize stack element management value' procedure 503 initializes the stack element management values associated with the predictor. This initialization can be accomplished (without limitation) once when the operating system boots, when a new application program process is initiated, or whenever the application program intends to reset the stack element management values. Once the stack element management values are initialized, the 'adaptive stack use adjustment' process 500 continues to a 'processing' procedure 505 that performs the steps involved to effectuate the program's purpose. When the processing is complete, the 'adaptive stack use adjustment' process 500 completes through an 'end' terminal 507. However, after the 'initialize stack element management value' procedure 503 and during the execution of the 'processing' procedure 505 a 'gather stack use information' procedure 509 also executes.

The 'gather stack use information' procedure 509 gathers stack use information resulting from the execution of the 'processing' procedure 505. The stack use information is used by an 'adjust stack element management value' procedure 511 to adjust the stack element Page: 14 management values to optimize the stack file fill/spill characteristics during the execution of the 'processing' procedure 505. The stack element management values can be adjusted through an operating system service invocation or other technique suited to changing these values dependent on the specific characteristics of an embodiment.

Another aspect of the invention provides multiple predictors as a function of the address of the instructions that invoke a stack exception trap. Fig. 6A illustrates a hash mechanism, as indicated by general reference character 600, that selects a predictor as a function of the address of the instruction that caused a stack exception trap. The hash mechanism 600 includes a 'hash index generation' procedure 601, which is subsequently lo described with respect to Fig. 6B, and a hash table 603. The 'hash index generation' procedure 601 takes as input the address in memory of the instruction that caused the stack exception trap (for example, in the case of a register window file, the "save" or "restore" instruction that caused the stack exception trap). Using well known methods, the address is hashed to generate an index to an entry in the hash table 603. The hash table 603 contains a first predictor entry 605 and an e predictor entry 607. The first predictor entry 605 through the e predictor entry 607 comprise a set of predictor entries. Each predictor entry in the hash table 603 can contain the predictor value itself, a pointer to the appropriate predictor value, or other value used to specify a predictor. One skilled in the art will understand that the use of the hash mechanism 600 allows multiple predictors to separately control the spill/fill of the stack file dependant on where in memory the overflow and underflow exceptions occur. Such a one wW also understand that the hash mechanism is but one possible way of using information saved by the trap to direct the selection of a predictor.

Fig. 6B illustrates a predictor selection process, as indicated by general reference character 650, for accessing a predictor from the hash table 603 of Fig. 6A. The predictor selection process 650 is invoked by the 'hash index generation' procedure 601 and initiates at a 6start' terminal 651. Next, a 'get address of trapping instruction' procedure 653 accesses the information saved as a result of the stack,exception trap to obtain the address of the instruction that caused the trap. Next a 'hash address' procedure 655 uses well understood techniques to generate an appropriate index into the hash table 603 based on the address obtained from the 'get address of trapping instruction' procedure 653. Next, an 'access ftge: Is predictor' procedure 657 uses the contents of the hash table 603 as (or to obtain) the predictor. Once the predictor (or a pointer to the predictor) is selected, the previously described processes can operate on the predictor and appropriately spil.1/fiR the stack file. The predictor selection process 650 completes through an 'end' terminal 659.

Another approach to minimizing exception traps from a top-of-stack cache is to maintain a history of the recent exception traps and use this pattern of overflow traps and underflow traps to better select a predictor.

Fig. 7A illustrates a predictor selection overview, indicated by general reference character 700, that includes a hash table 701 containing predictors or pointers to predictors. A io 'two element hash index generation' procedure 703 uses a trap address 705 and an exception history 707 to generate a hash index into the hash table 701. The trap address 705 is determined as was described with respect to Fig. 6A. The exception history 707 is a variable that contains a number of "places". These places are usually single-bit fields. However, depending on the.number of types of exceptions being tracked each place may contain is multiple bits. Single-bit fields are sufficient if only overflow and underflow traps are tracked.

In response to a tracked exception trap, the contents of the exception history 707 is shifted one place (one bit) and the place freed by & shift is set to a value that identifies the exception trap. Thus, the exception history 707 repre I sents a usage pattern for the top-of-stack cache.

This usage pattern can be used in conjunction with the address of the trapping instruction to select the appropriate predictor for the instant exception trap.

Fig. 7B Wustrates a predictor selection process, indicated by general reference character 720, that selects a predictor from the set of predictors in the hash table 701. The predictor selection process 720 initiates at a 'start' terminal 721 and continues to a 6get trap address' procedure 723 that accesses the information saved by the operation of the trap to determine the address of the instruction that caused the trap. Next, a 'hash exception history with trap address' procedure 725 hashes aU or a portion of the trap address with the exception history 707 to generate a hash index into the hashitable 701. Once the hash index is generated, an 'access predictor' procedure 727 uses the contents of the hash table 701 as (or to obtain) the predictor. Once the predictor (or a pointer to the predictor) is selected, the Page: 16 previously described processes can operate on the predictor and appropriately spill/fill the stack file. The predictor selection process 720 completes through an 'end' terminal 729.

Fig. 7C illustrates an exception history maintenance process, indicated by general reference character 740, that maintains an ordered sequence bits that represent the history of overflow exceptions and underflow exceptions from said top-of-stack cache. The exception history maintenance process 740 is invoked by the 'process stack trap' procedure 207 of Fig. 2 and initiates at a 'start' terminal 741. The exception history maintenance process 740 then continues to a 'shift history' procedure 743 that shifts the contents of the exception history by one place leaving a free place in the exception history. If the exception history is only lo tracking two types of exceptions (overflow and underflow) the shift is one bit. Next, the dexception type' decision procedure 745 determines which exception occurred. If the exception was a stack overflow exception, the exception history maintenance process 740 continues to a 'set overflow indication' procedure 747 that sets the free place in the exception history to indicate that the current exception was an overflow. However, if the exception was an underflow, the exception history maintenance process 740 continues to a 'set underflow indication' procedure 749 that sets the free place in the exception history to indicate that the current exception was an underflow. After the free place in the exception history is set, the exception history maintenance process 740 completes through an 'end' terminal 751. One skilled in the art will understand that if the procedure used to shift the exception history leaves the free place in a known state (that is one or zero) one of the 'set overflow indication' procedure 747 and the 'set underflow indication' procedure 749 need not perform any operation- One skilled in the art will understand that the invention improves the performance of computer systems having a stack Me (or other top-of-stack cache) by reducing the number of 2s stack exception traps resulting from the execution of a program.

Although the present invention has been described in terms of presently preferred embodiments, one skilled in the art will understand that various modifications and alterations may be made without departing from the scope of the invention. Accordingly, the scope of Pap: 17 the invention is not to be limited. to the particular invention embodiments discussed herein, but should be defted only by the appended claims and equivalents thereof Page: 18

Claims

Claims

What is claimed is:

1 1. A computer controlled method for selecting a predictor from a set of predictors to 2 minimize exceptions in a computer having a memory and a top-of-stack cache, said 3 method comprising the steps of 4 (a) initializing an exception history used to track occurrences of a plurality of exception traps from said top-of-stack cache; a (b) invoking an exception trap; 7 (c) updating said exception history dependent on said exception trap; 8 (d) selecting said predictor from said set of predictors based on said exception history; 9 and (e) processing s aid exception trap dependent on said predictor.

1
2. The computer controlled method of claim I wherein step (d) comprises:

2 (dl) accessing trap information saved by said exception trap; and 3 (d2) selecting said predictor from said set of predictors, said selection based on said trap 4 information and said exception history.

1
3. The computer controlled method of claim 1 wherein said exception history 2 represents an ordered sequence of overflow exceptions and underflow exceptions 3 from said top-of-stack cache.

1
4. The computer controlled method of claim I further comprising changing said 2 predictor responsive to said exception trap.

Page: 19 1
5. An apparatus having a central processing unit (CPU) and a memory coupled to said 2 CPU for selecting a predictor from a set of predictors to minimize exceptions in a 3 computer having a memory and a top-of-stack cache, said apparatus comprising:

4 an initialization mechanism configured to initialize an exception history used to track occurrences of a plurality of exception traps from said top-of- stack cache; 6 a trap mechanism configured to invoke an exception trap; 7 a history tracking mechanism configured to update said exception history 8 dependent on said exception trap; 9 a predictor selection mechanism configured to select said predictor from said set of predictors based on said exception history; and I I a trap handler mechanism in said computer, configured to process said 12 exception trap dependent on said predictor.

1
6. The apparatus of claim 5 w6rein the predictor selection mechanism comprises:

2 a trap information access mechanism configured to access trap information 3 saved by said exception trap; and 4 a directed selection mechanism configured to select said predictor from said set of predictors, said selection based on said trap information and said exception 6 history.

1
7. The apparatus of claim 5 wherein said exception history represents an ordered 2 sequence of overflow exceptions and underflow exceptions from said top- of-stack 3 cache.

1
8. The apparatus of claim 5 further comprising a predictor maintenance mechanism 2 configured to change said predictor responsive to said exception trap.

Pap: 20 1
9. A computer program product comprising:

2 a computer usable storage medium having computer readable code embodied 3 therein for causing a computer to select a predictor from a set of predictors to 4 minimize exceptions in a computer having a memory and a top-of-stack cache, said computer readable code comprising:

6 computer readable program code configured to cause said computer to effect 7 an initialization mechanism configured to initialize an exception history used to 8 track occurrences of a plurality of exception traps from said top-of- stack cache; 9. computer readable program code configured to cause said computer to effect a trap mechanism configured to invoke an exception trap; 11 computer readable program code configured to cause said computer to effect a 12 history tracking mechanism configured to update said exception history dependent 13 on said exception trap; 14 computer readable program code configured to cause said computer to effect a is predictor selection mechanism configured to select said predictor from said set of 16 predictors based on said exception history; and 17 computer readable program code configured to cause said computer to effect a 18 trap handler mechanism, in said computer, configured to process said exception 19 trap dependent on said predictor.

Page: 21 1 1
10. The computer program product of claim 9 wherein the predictor selection 2 mechanism comprises:

3 computer readable program code configured to cause said computer to effect a 4 trap information access mechanism configured to access trap information saved by said exception trap; and 6 computer readable program code configured to cause said computer to effect a 7 directed selection mechanism configured to select said predictor from said set of 8 predictors, said selection based on said trap information and said exception history.

1
11. The computer program product of claim 9 wherein said exception history represents 2 an ordered sequence of overflow exceptions and underflow exceptions from said 3 top-of-stack cache.

1
12. The computer program product of claim 9 further comprising computer readable 2 program code configured to cause said computer to effect a predictor maintenance 3 mechanism configured to change said predictor responsive to said exception trap.

Page: 22 1
13. A computer program product comprising:

2 a computer data signal embodied in a carrier wave having computer readable 3 code embodied therein for causing a computer to select a predictor from a set of 4 predictors to minimize exceptions in a computer having a memory and a top-of stack cache, said computer readable code comprising:

6 computer readable program code configured to cause said computer to effect 7 an initialization mechanism configured to initialize an exception history used to 8 track occurrences of a plurality of exception traps from said top-of- stack cache; 9 computer readable program code configured to cause said computer to effect a trap mechanism configured to invoke an exception trap; 11 computer readable program code configured to cause said computer to effect a 12 history tracking mechanism configured to update said exception history dependent 13 on said exception trap; 14 computer readable program code configured to cause said computer to effect a is predictor selection mechanism configured to select said predictor from said set of 16 predictors based on said exception history; and 17 computer readable program code configured to cause said computer to effect a trap handler mechanism, in said computer, configured to process said exception 19 trap dependent on said predictor.

Page: 23 1
14. A computer controlled method for minimizing exceptions in a computer having a 2 memory and a return address top-of-stack cache, said method comprising the steps 3 of 4 (a) initializing a predictor for tracking exceptions from said return address top-of-stack cache; 6 (b) invoking an exception trap; 7 (c) processing said exception trap dependent on said predictor; and 8 (d) changing said predictor responsive to said exception trap.

1
15. The computer controlled method of claim 14 wherein said predictor is associated 2 with at least one stack element management value that includes a fill value, the 3 stack exception trap is a stack underflow trap, said return address top- of-stack 4 cache is organized into a plurality of stack elements and step (c) further comprises:

(cl) determining said fill value dependant on said predictor, said fill value used to 6 specify how many of said plurality of stack elements are to be filled; and 7 (c2) fiffing at least one of said plurality of stack elements dependant on said fill value.

1
16. The computer controlled method of claim 14 wherein said predictor is associated 2 with at least one stack element management value that includes a spill value, the 3 stack exception trap is a stack overflow trap, said return address top- of-stack cache 4 is organized into a plurality of stack elements and step (c) further comprises:

(c 1) determining said spill value dependant on said predictor, said spill value used to 6 specify how many of said plurality of stack elements are to be spilled; and 7 (c2) spilling to said memory at least one of said plurality of stack elements dependan't on 8 said spill value.

Page: 24 1
17. The computer controlled method of claim 14 wherein said predictor is associated 2 with at least one stack element management value and said method further 3 comprises:

4 (e) adjusting said at least one stack element management value.

1
18. An apparatus having a central processing unit (CPU) and a memory coupled to said 2 CPU for minimizing exceptions in a computer having a return address top- of-stack 3 cache, said apparatus comprises:

4 a predictor initialization mechanism configured to initialize a predictor for tracking exceptions from said return address top-of-stack cache; 6 a trap mechanism configured to invoke an exception trap in said computer; 7 a trap handler mechanism in said computer, configured to process said a exception trap dependent on said predictor, and 9 a predictor maintenance mechanism configured to change said predictor responsive to said exception trap.

1
19. The apparatus of claim 18 wherein said predictor is associated with at least one 2 stack element management value that includes a flu value, the stack exception trap 3 is a stack underflow trap, said return address top-of-stack cache is organized into a 4 plurality of stack elements and the trap handler mechanism further comprises:

a fill determination mechanism configured to determine said fill value 6 dependant on said predictor, said fill value used to specify how many of said 7 plurality of stack elements are to be filled; and a a fill mechanism configured to fill at least one of said plurality of stack 9 elements dependant on said fill value.

I-.

Page: 25 1
20. The apparatus of claim 18 wherein said predictor is associated with at least one 2 stack element management value that includes a spill value, the stack exception trap 3 is a stack overflow trap, said return address top-of-stack cache is organized into a 4 plurality of stack elements and the trap handler mechanism further comprises:

a spill determination mechanism configured to determine said spill value 6 dependant on said predictor, said spill value used to specify how many of said 7 plurality of stack elements are to be spilled; and 8 a spill mechanism configured to spill to said memory at least one of said 9 plurality of stack elements dependant on said spill value.

1
21. The apparatus of claim 18 wherein said predictor is associated with at least one 2 stack element management value and said apparatus further comprises:

3 an adjustment mechanism configured to adjust said at least one stack element 4 management value.

Page: 26 1
22. A computer program product comprising:

2 a computer usable storage medium having computer readable code embodied 3 therein for causing a computer to minimize exceptions, said computer having a 4 return address top-of-stack cache, said computer readable code comprising:

computer readable program code configured to cause said computer to effect a 6 predictor initialization mechanism configured to initialize a predictor for tracking 7 exceptions from said return address top-of-stack cache; a computer readable program code configured to cause said computer to effect a 9 trap mechanism configured to invoke an exception trap in said computer, computer readable program code configured to cause said computer to effect a 11 trap handler mechanism, in said computer, configured to process said exception 12 trap dependent on said predictor; and 13 computer readable program code configured to cause said computer to effect a 14 predictor maintenance mechanism configured to change said predictor responsive is to said exception trap.

1
23. The computer program product of claim 22 wherein said predictor is associated 2 with at least one stack element management value that includes a fin value, the 3 stack exception trap is a stack underflow trap, said return address top- of-stack 4 cache is organized into a plurality of stack elements and the trap handler mechanism further comprises:

6 computer readable program code configured to cause said computer to effect a 7 fill determination mechanism configured to determine said fill value dependant on 8 said predictor, said fill value used to specify how many of said plurality of stack 9 elements are to be filled; and computer readable program code configured to cause said computer to effect a 11 fill mechanism configured to fill at least one of said plurality of stack elements 12 dependant on said fill value.

Page: 27 1
24. The computer program product of claim 22 wherein said predictor is associated 2 with at least one stack element management value that includes a spill value, the 3 stack exception trap is a stack overflow trap, said return address top- of-stack cache 4 is organized into a plurality of stack elements and the trap handler mechanism further comprises:

6 computer readable program code configured to cause said computer to effect a 7 spill determination mechanism configured to determine said spill value dependant 8 on said predictor, said spill value used to specify how many of said plurality of 9 stack elements are to be spilled; and computer readable program code configured to cause said computer to effect a 11 spill mechanism configured to spill to said memory at least one of said plurality of 12 stack elements dependant on said spill value.

1
25. The computer program product of claim 22 wherein said predictor is associated 2, with at least one stack element management value and said Product further 3 comprises: ' 4 computer readable program code configured to cause said computer to effect an adjustment mechanism configured to adjust said at least one stack element 6 management value.