Information
-
Patent Application
-
20020188931
-
Publication Number
20020188931
-
Date Filed
February 28, 200123 years ago
-
Date Published
December 12, 200221 years ago
-
Inventors
-
Original Assignees
-
CPC
-
US Classifications
-
International Classifications
Abstract
A method is disclosed for instructing a computing system to unwind a program call stack that lacks explicit Stack Frame Backchain Pointers, including finding the called function's entry point, determining the return point in the calling function, placing a NOP instruction at the return point that contains embedded information about what type of call was used, and backing up in storage by an amount determined by using the address of the called function's entry point to locate the static data item containing the called function's DSA size.
Description
BACKGROUND
[0001] When a function is initially called, the first thing it typically does in its prolog code is to save the contents of the hardware registers. The register contents at function call time represent the state of the calling function, and this state must be restored when the called function returns. Since the called function will most likely need to use the hardware registers for its own purposes, the contents of any registers (general purpose, floating point, etc.) that may be changed must first be saved in what is referred to as a register save area.
[0002] A called function may also need some working storage where it can place its automatic data, which is data with a scope limited to the duration of the called function. When the called function returns, its automatic data is no longer required.
[0003] Both of these storage requirements can be satisfied through the use of a dynamic storage area or Stack. The Stack can be comprised of several Stack Segments that are considered contiguous in storage from the application's perspective, although they may reside in real memory at discontinuous storage locations. The Stack consists of storage that is specific to a single thread of program execution. Typically, every function that has been called but not yet returned during the current program's execution is represented in the Stack by a Stack Frame.
[0004] Typically, the prolog code for a function that has just been called will save the calling function's registers and/or allocate storage for its Stack Frame before it does anything else. For example, function prolog code may save the calling function's registers into a register save area in a Stack Frame that was allocated by the calling function and is pointed to by a register. The Stack Frame size of the called function, which is known at compile time, is used to determine if the called function's Stack Frame will fit within the current Stack Segment. If it does not, a new Stack Segment is allocated. The code also saves the address of the calling function's Stack Frame into the new Stack Frame allocated by the called function. This saved address, known as a Stack Frame Backchain Pointer, can be used to find the calling function's Stack Frame. The process of traversing backwards through the Stack towards older calling functions is known as Stack Unwinding. The called function's epilog code restores the register containing the calling function's Stack Frame address, using the Stack Frame Backchain Pointer, before it returns control to the calling function.
[0005] Unfortunately, saving and restoring the Stack Frame Backchain Pointer in a called function's Stack Frame adds instruction cycles to the prolog and epilog code sections of every function that is called during the execution of the application. For programming languages built around an object-oriented paradigm (e.g., C++ and Java), there may be many function calls, some to functions having an actual “function body” (the executable instructions between the function prolog and epilog) that is small relative to the sizes of the prolog and epilog sections. In these cases, there can be a significant performance benefit by eliminating unnecessary instruction cycles (“call linkage overhead”) from the prolog and epilog code sections. Accordingly, for such function calls it is desirable to provide a method to unwind Stack Frames without the need for a Stack Frame Backchain Pointer in each Stack Frame.
SUMMARY
[0006] This disclosure presents a method for instructing a computing system to unwind a program call stack that lacks explicit Stack Frame Backchain Pointers, including finding the called function's entry point, determining the return point in the calling function, placing a No Operation (“NOP”) instruction at the return point that contains embedded information about what type of call was used, and backing up in storage by an amount determined by using the address of the called function's entry point to locate the static data item containing the called function's DSA size.
BRIEF DESCRIPTION OF THE DRAWINGS
[0007] The present disclosure will be better understood, and its numerous features and advantages will become apparent to those skilled in the pertinent art by referencing the accompanying drawings, in which:
[0008]
FIG. 1 is a schematic diagram of a typical program call stack.
[0009]
FIG. 2 is a sample pseudocode listing of prior art function prolog code.
[0010]
FIG. 3 is a sample pseudocode listing of prior art function epilog code.
[0011]
FIG. 4 is a sample pseudocode listing of new function prolog code.
[0012]
FIG. 5 is a sample pseudocode listing of new function epilog code.
[0013]
FIG. 6 is a sample pseudocode listing for a call using a branch instruction based on the contents of a hardware register.
[0014]
FIG. 7 is a sample pseudocode listing for a call using a relative branch instruction.
[0015]
FIG. 8 is a sample pseudocode listing for the use of a call type descriptor.
DETAILED DESCRIPTION
[0016] An exemplary method is provided for unwinding a program call stack where the function stack frames do not include backchain pointers. The method unwinds one or more stack frames without requiring any backchain pointers in the stack frame memory.
[0017] When a function is initially called, the first thing it normally does is to save the contents of the hardware registers. The register contents at the time the function is called represent the state of the calling function, and this state is to be restored before the called function returns. Since the called function will most likely need to use the hardware registers for its own purposes, the contents of any registers (e.g., general purpose, floating point, etc.) that must be changed must first be saved, typically in a register save area.
[0018] A called function may also require some working storage where it can place its automatic data, which is data with a scope limited to the duration of the called function. When the called function returns, its automatic data is no longer required, so its storage space may be reused. Both of these storage requirements can be satisfied through the use of a Stack Frame (also called a Dynamic Storage Area, or “DSA”).
[0019] Turning to FIG. 1, a program call stack may include several Stack Segments of predefined size (e.g., N words) that are considered contiguous in storage from the application's perspective, although they may reside in real memory at discontinuous storage locations. The Stack consists of storage that is specific to a single thread of program execution. Typically, a separate Stack Frame corresponds to each function that has been called during execution of the current program, but which has not yet returned. However, it is possible that the most recently called function, if it is a “leaf” routine that makes no further calls, will not have a Stack Frame on the Stack.
[0020] Typically, the prolog code for a function that has just been called will save the calling function's registers and/or allocate storage for its Stack Frame before it does anything else. For example, FIG. 2 shows some function prolog code typical for a C function that has been compiled on IBM's OS/390 operating system using a previous method. Line 6 saves the calling function's registers into a register save area in a Stack Frame that was allocated by the calling function and is pointed to by register 13. Lines 7 through 17 use the Stack Frame size of the called function, which is known at compile time, to determine if the called function's Stack Frame will fit within the current Stack Segment. If it will not fit, a new Stack Segment is allocated. Line 15 saves the address of the calling function's Stack Frame, which was in register 13 when the called function got control, into the new Stack Frame allocated by the called function. This provides what is known as a Stack Frame Backchain Pointer that can be used to find the calling function's Stack Frame. This processing of traversing backwards (i.e., towards older calling functions) through the Stack is also known as Stack Unwinding.
[0021]
FIG. 3 shows how the called function's epilog code that corresponds to the prolog code of FIG. 2 restores register 13, which contains the calling function's Stack Frame address, at line 1 using the Stack Frame Backchain Pointer before it returns control back to the calling function.
[0022] Unfortunately, saving and restoring the Stack Frame Backchain Pointer in a called function's Stack Frame adds instruction cycles to every prolog and epilog of every function that is called during the execution of the application. For programming languages built around an object-oriented paradigm (e.g., C++ and Java), there may be many function calls, some to functions whose actual “function body” (the executable instructions between the function prolog and epilog) is small relative to the those in the prolog and epilog. In this case, there can be a significant performance benefit by eliminating as many unnecessary instruction cycles as possible from the prolog and epilog (known as call linkage overhead). With this embodiment, a method is provided to unwind Stack Frames without the need for a Stack Frame Backchain Pointer in each Stack Frame.
[0023]
FIGS. 4 and 5 illustrate function prolog and epilog code, respectively, that could be generated from a C or C++ function compiled with the new XPLINK compiler option on IBM's OS/390 operating system. The new XPLINK linkage takes advantage of a downward growing stack frame (e.g., the Stack Segment shown in FIG. 1 is downward growing) and a write-protected guard page to eliminate the instructions of FIGS. 2 and 3 that were necessary to check for Stack Segment overflow. Also eliminated from this function prolog code are the instructions that would save a Stack Frame Backchain Pointer in each Stack Frame. This elimination further reduces the number of hardware cycles required in each function prolog, but it requires a new technique, such as described in this exemplary embodiment, to unwind the Stack in those situations where unwinding of the stack becomes necessary.
[0024] Since the stack is downward growing, the new Stack Frames are allocated by subtracting the DSA size (i.e., the size of the new Stack Frame) from the value of the current Stack Pointer. Conversely, the calling function's Stack Frame can be found by adding the DSA size of the called function to the value of the Stack Pointer. While the DSA size of the called function is known by the compiler and can be used in the function epilog code, it is also possible to locate each Stack Frame during Stack unwinding. The process of unwinding a Stack is normally done by an external function that has no direct knowledge of any called function's Stack Frame.
[0025] The technique described here for Stack unwinding without a backchain pointer could also work for upward growing Stack Frames. In this case Stack Frames are allocated by adding the DSA size, and Stack unwinding occurs by subtracting the DSA size. Upward growing stack frames are currently less efficient though, as a write-protected guard page is not used to implicitly check for stack overflow.
[0026] To illustrate this method on a downward growing Stack, assume that a program exception has interrupted the execution of a function, and an exception handler routine has requested that the operating system resume execution at a point earlier in the Stack (e.g., by issuing a UNIX longjmp( ) service). This will require that the operating system be able to unwind the chain of Stack Frames. It then becomes necessary to be able to find the DSA size of each unreturned function. Since each size is known at compilation time, it can be included as static data in the executable function. One place to put this is in a function prolog area that just precedes the function's main entry point. Assuming this exemplary location, then one method to Stack unwinding without an explicit Stack Frame Backchain Pointers is described below.
[0027] 1. Start by finding the called function's entry point. Although it can not be assumed that the called function's entry point is in the register save area because the function may have been called using a relative branching technique, the address of the return point in the calling function is known.
[0028] 2. Next, determine the return point in the calling function. The return point in the calling function is right after the instruction that was used to make the function call. Normally this can be found in the called function's register save area. Although the present example assumes that the interrupted function has completed its prolog code, if the interrupt occurred in the prolog of the called function before the registers were saved then the return address is still contained in a register.
[0029] 3. The compiler can place a NOP instruction at the return point that contains embedded information about what type of call was used. If the embedded NOP information indicates the call was made using an instruction that branched to the called function's entry point based on the contents of a hardware register, then the called function's entry point can be obtained from the called function's register save area. The code sample of FIG. 6 shows an example of this type of call site. If the embedded NOP information indicates the call was made using a relative branch instruction, then the called function's entry point can be calculated from the address of the call site and the size of the relative branch, positive or negative (this size can be obtained by decoding the instruction used to branch to the called function's entry point). The code sample of FIG. 7 shows an example of this type of call site.
[0030] 4. Next, using the address of the called function's entry point, we can back up in storage by a known fixed amount to locate the static data item containing the called function's DSA size.
[0031] 5. For a downward growing Stack, add the DSA size of the called function to the called function's Stack pointer to obtain the address of the calling function's Stack Frame.
[0032] 6. Repeat steps 2 through 5 as necessary.
[0033] This method exemplifies how a Stack can be unwound without the use of a Stack Frame Backchain Pointer. Since Stack unwinding is an operation that is typically performed only in cases of exception, the tradeoff between eliminating cycles in heavily executed prolog and epilog code paths as in the exemplary embodiment versus the potentially unrealized cost of unwinding the entire chain of Stack Frames is a worthwhile one for many function-call intensive programs.
[0034] In operation, the exemplary method allows compilers and others to create prolog code that does not need to set a backchain pointer while providing a means for the runtime environment to continue to be able to unwind the stack, which is necessary for exception and jump processing. Stack unwinding without a backchain pointer uses a stack structure that grows from higher to lower addresses as save areas are pushed on. On function entry, a save area is acquired by simply subtracting the function's save area size from the current stack pointer. Subsequently, on function exit, a save area is deleted by adding the function's save area size to the current stack pointer. This pops the save area off the stack which then points back on the caller's save area.
[0035] There are several new structures created to support unwinding the stack from the perspective of the runtime environment in order to support exception and jump processing. They are the Call Type Descriptor, Call Descriptor, and Entry Point Marker.
[0036] A Call Type Descriptor consists of a NOP located at the return point of a call site that is encoded to provide information for the runtime to use to locate the entry point of the function. Once the entry point of the function is located, the save area size for the function can be identified and arithmetic used to locate the save area of the function's caller.
[0037] A Call Type Descriptor is a NOP instruction immediately following a function call with 4 bits reserved to indicate the ‘type’ of call and 28 bits to contain an offset to either a ‘Call Descriptor’, which contains parameter mapping information about the call and an offset to the ‘Entry Point Marker’, or if the offset is negative, the offset is to the ‘Entry Point Marker’ for the function making the call.
[0038] An exemplary call type descriptor would be used in the code of a calling program as shown in FIG. 8. The 4 bits of “call type” information are currently defined in this exemplary embodiment as:
[0039] 0000=Call using standard branch and save or BASR
[0040] 0001=Call using branch relative or BRAS
[0041] The ‘Entry Point Marker’ starts 16 bytes behind a function entry point. This consists of an eyecatcher that is designed to never be confused as code and some encoded information, including the size of the function's save area. It is arranged as follows:
1|
|
Entry Point Marker:
|
|
+07 byte unique eyecatcher
+7One byte marker identifier
+8Offset to Program Prolog Area
+12 Save area size
|
[0042] Using a save area, the entry point marker of the callee, or the routine that owns a specific save area, can be found by either walking the code from a known point of interruption back to the entry point marker's eyecatcher, or using the return register in the save area to locate the call type descriptor to determine whether the call was done using a branch and save or branch relative. If the call was via branch and save (“BASR”) the entry point of the callee is available in a copy of the entry register in the callee's save area. If the call was done via branch relative (“BRAS”), the entry point of the callee can be computed as an offset from the call point. Once the entry point marker is located, the save area size of the callee can be retrieved and used arithmetically (added to the current save area) to locate the caller's save area.
[0043] In a stack walking scenario, once the size of the save area for the routine at the top of the call chain has been established, the remainder of the call chain can be traversed using either the same technique for finding the entry point markers or the slightly more efficient method, which uses the call offset information stored in the call type descriptor.
[0044] An advantage of the described exemplary embodiment is that a program compiled using the method of the XPLINK compiler option may execute more quickly at run-time, particularly if the program will encounter multiple function calls during execution.
[0045] While exemplary embodiments have been shown and described, various modifications and substitutions may be made thereto without departing from the spirit and scope of the disclosure. Accordingly, it is to be understood that the present disclosure has been made by way of illustration only, and such illustrations and embodiments as have been disclosed herein are not to be construed as limiting to the claims.
Claims
- 1. A method for unwinding a program call stack, comprising:
predetermining the size of at least one stack frame; receiving a current stack pointer; and offsetting the current stack pointer by the size of the at least one stack frame to obtain a pointer to the previous stack frame.
- 2. A method as defined in claim 1 wherein the program call stack is downward growing.
- 3. A method as defined in claim 1 wherein the size of the at least one stack frame is known to the compiler.
- 4. A method for unwinding a program call stack that lacks explicit Stack Frame Backchain Pointers, the method comprising:
finding the called function's entry point; determining the return point in the calling function; placing a NOP instruction at the return point that contains embedded information about what type of call was used; and backing up in storage by an amount determined by using the address of the called function's entry point to locate the static data item containing the called function's DSA size.
- 5. A method as defined in claim 4 wherein the program call stack is a downward growing Stack.
- 6. A method as defined in claim 5, further comprising:
adding the DSA size of the called function to the called function's Stack pointer to obtain the address of the calling function's Stack Frame.
- 7. A method as defined in claim 4 wherein finding the called function's entry point does not assume that the called function's entry point is in the register save area because the function may have been called using a relative branching technique, but the address of the return point in the calling function is known.
- 8. A method as defined in claim 4 wherein determining the return point in the calling function comprises at least one of accessing the called function's register save area and accessing a register.
- 9. A method as defined in claim 4 wherein placing a NOP instruction at the return point that contains embedded information about what type of call was used comprises at least one of accessing the called function's register save area and accessing the size of the relative branch.
- 10. A method as defined in claim 9 wherein accessing the size of the relative branch comprises decoding the instruction used to branch to the called function's entry point and calculating the offset to the called function's entry point from the address of the call site.