US20060282821A1

US20060282821A1 - Efficient subprogram return in microprocessors

Info

Publication number: US20060282821A1
Application number: US11/149,611
Authority: US
Inventors: Erik Renno; Oyvind Strom; Morten Lund
Original assignee: Atmel Corp
Current assignee: Atmel Corp
Priority date: 2005-06-10
Filing date: 2005-06-10
Publication date: 2006-12-14
Also published as: WO2006135549A3; CN101194228B; EP1891519A4; WO2006135549A8; WO2006135549A2; US8555041B2; CN101194228A; EP1891519A2; EP1891519B1; US20100250904A1; JP2009508180A; TW200709043A; KR20080014062A

Abstract

A method and medium for performing subroutine return operations. Test operations are performed in parallel with other operations in a return operation. These test operations and the return operations are performed in response to a single instruction.

Description

FIELD OF THE INVENTION

This invention relates to subprogram return operations in microprocessors.

BACKGROUND

Programs frequently feature subroutines which perform a specific task. After the task is performed, program flow returns from the subroutine to the main program. One common mechanism for performing a subroutine return involves conditionally or unconditionally moving the contents of a return address register into a program counter and then continuing program execution. A return value register may also be updated with a constant literal that may represent a Boolean value. Another approach to subprogram returns is to “pop” a return address from the stack and into the program counter and continue program execution from there. This operation may also pop any spooled-out register file contents from the stack into the register file.
These methods for performing subroutine returns take several cycles to execute. In FIG. 1, when a traditional return (“RET”) instruction is executed in a typical pipelined CPU, five cycles 10 are required to execute the instruction. In FIG. 2, a typical pipelined CPU contains a Program Counter (“PC”) 42 and an instruction memory 44. The CPU has four different pipeline registers 46, 52, 56, and 60 separating the different pipeline stages. The Instruction Decode stage (between registers 46 and 52) contains both a control/decode unit (“CU”) 48 for decoding the current instruction and generating control signals and a register file 50. The Execution Stage (between registers 52 and 56) contains an Arithmetic Logic Unit (“ALU”) 54. The Memory Stage (between registers 56 and 60) contains a data memory 58.) With continued reference to FIG. 1, during cycle 1, in the Instruction Fetch (“IF”) stage 12, the RET instruction is fetched (block 22). In cycle 2, in the Instruction Decode (“ID”) stage 14, correct control signals are generated and the return address register is read from the register file (block 24). In cycle 3, in the Execution (“EX”) stage 16, the return address register content is written through the Arithmetic Logic Unit (“ALU”) with no change (block 26). During cycle 4, in the Memory (“MEM”) stage, the return address register content is written past the data memory. Finally, in cycle 5, in the Writeback (“WB”) 20 stage, the return address register content is written to the Program Counter (“PC”) and the pipeline is flushed (block 30). Once the pipeline is flushed, the pipeline does not contain any instructions until the instruction at the return address is read from program memory. Therefore, several clock cycles are wasted in the pipeline flush process.
A similar issue exists for a return instruction (“RETMEM”) popping the return address register from a stack in memory. As shown in FIG. 3, in cycle 1, in the IF stage, the RETMEM instruction is fetched (block 32). During cycle 2, in the ID stage, the correct control signals are generated. In cycle 3, during the EX stage, the control signals to the data memory are routed past the ALU (block 36). In cycle 4, in the MEM stage, the return address is read from data memory (block 38). Finally, in cycle 5, in the WB stage, the return address read from memory is written to PC and the pipeline is flushed (block 40). As with the return instruction discussed in FIG. 1, several cycles are wasted after the pipeline flush.
It would be advantageous to provide a more efficient subroutine return operation.

SUMMARY

In an exemplary embodiment, an instruction is fetched which requires a return operation and sets status flags based on the contents of a return value register. The status flags are set in parallel with at least one other operation required to process the return instruction. The status flags are set before one of the following occurs: i) contents of a return address register are moved into a program counter; or ii) a return address is popped from a stack and into the program counter. In another embodiment, a processor-readable storage medium causes a processor to perform this subroutine return operation.
In yet another exemplary embodiment of the invention, a subroutine return operation places a return address into a program counter. A test operation is executed on a return value register; the test operation is performed in parallel with at least one other operation required to process the return operation. The program flow is changed to a target address. Each of the above-mentioned steps is performed in response to a single instruction. In one embodiment, a processor-readable storage medium stores an instruction that causes a processor to perform this subroutine return operation.

BRIEF DESCRIPTION OF HTE DRAWINGS

FIG. 1 is a chart showing how a return instruction is executed in the prior art.
FIG. 2 is a block diagram of a pipelined CPU in the prior art.
FIG. 3 is a chart showing how a return instruction popping the return address register from a stack in memory is executed in the prior art.
FIG. 4 is a block diagram of a pipelined CPU in an exemplary embodiment of the present invention.
FIG. 5 is a chart showing an exemplary execution of a return instruction in an embodiment of the present invention.
FIG. 6 is a chart showing an exemplary execution of a return instruction popping the return address register from a stack in memory is executed in an embodiment of the present invention.

DETAILED DESCRIPTION

A more efficient subroutine return operation is provided in which status flags are updated (in the processor's status register) according to a test of the return value register during the subroutine return operation. (In the prior art, test operations, for instance, a test of the return value register, are performed in response to a separate instruction.) In one embodiment, the status flags are set in parallel with operations to execute single instructions such as conditional return instructions as well as single instructions incorporating a return operation. The instructions are stored in a processor-readable medium, which includes any medium that can store or transfer information, such as an electronic circuit, a semiconductor memory device, a ROM, a flash memory, a floppy diskette, a compact disc, an optical disc, etc.
These instructions can be executed by existing hardware. In FIG. 4, an exemplary CPU for executing these instructions includes a PC 62 and instruction memory 64. The CPU contains four pipeline registers (IF/ID 66, ID/EX 70, EX/MEM 74, and MEM/WB 78) separating the different stages. The ID stage, between registers 66 and 70, contains a control/decode unit 68 for decoding the current instruction and generating the correct control signals. The ID stage also contains a register file 132. The EX stage, between registers 70 and 74, contains an ALU 72 and a flag register 84. The MEM stage, between registers 74 and 78, contains data memory 76. When an address has reached the WB stage (after register 78), the pipeline has been flushed and the fetch address is written to PC 62. A multiplexer 118 determines which address is written into the register file or the program counter (this is discussed in greater detail, below). In other embodiments, the processor may have different features, such as data forwarding; as noted above, the CPU described in FIG. 4 is exemplary and is not the only processor which can execute the more efficient subroutine return operation described herein.
In one embodiment of the invention, test operations are performed in parallel with other operations during execution of instructions with the more efficient subroutine return operation. In one embodiment, shown in FIG. 5, when a return (“return_with_test”) instruction is executed, during cycle 1 in the IF stage, the return_with_test instruction is fetched (block 86). In cycle, 2, the return_with_test has entered the ID stage; the correct control signals are generated and the return address register is read from the register file (block 88). In cycle 3, the return_with_test is kept in the ID stage an additional cycle (in one embodiment, this may be done in the decode stage by splitting the instruction into two “micro-operations”: one micro-operation performs the test operation, the other micro-operation performs the return operation); in this second cycle, the return value register is read from the register file and control signals to instruct the ALU to perform the test operation are generated (block 90). In cycle 3 in the EX stage, the return address content is written through the ALU with no change (block 92). During cycle 4 in the EX stage, the ALU sets the flags corresponding to the test of the value register (block 94). (In this embodiment, the status flags are set according to a comparison of the return value register's contents with zero. Status flags used in this embodiment indicate overflow (“V”), a negative value (“N”), a zero result (“Z”), and a carry after an arithmetic or logic operation (“C”). Different status flags may be used in other embodiments and/or status flags may be set differently in other embodiments.) During cycle 4 in the MEM stage, the return address register content is written past the data memory (block 96). During cycle 5, in the WB stage, the return address register content is written to the PC and the pipeline is flushed (block 98). A test operation has been performed using cycles that would otherwise be unused due to the pipeline flush.
In another embodiment, a test operation may be performed during execution of a return instruction (“pop_with_test”) popping the return address register from a stack in memory. In FIG. 6, in cycle 1 in the IF stage, the pop_with_test instruction is fetched (block 100). During cycle 2, in the ID stage, the correct control signals are generated (block 102). In cycle 3, the pop_with_test is kept in the ID an additional cycle (in one embodiment, the instruction is decoded into two micro-operations (the subroutine return operation and the test operation) in the ID stage); the return value register is read from the register file and control signals to instruct the ALU to perform the test operation are generated (block 104). In cycle 3, in the EX stage, the control signals to data memory are routed past the ALU (block 106). In cycle 4, in the EX stage, the ALU sets the flags corresponding to the test of the return value register (block 108). During the same cycle, in the MEM stage, the return address is read from data memory (block 110). In cycle 5, in the WB stage, the return address read from memory is written to PC and the pipeline is flushed (block 112).
Other embodiments of the invention may vary from the embodiments discussed above. These embodiments may require fewer or additional clock cycles to execute instructions. Other embodiments may require different hardware to execute the instructions. Still other embodiments may be incorporated into different subprogram return operations and instructions.
FIGS. 5 and 6 are exemplary embodiments of “return_with_test” and “pop_with_test” instructions, respectively. In the “return_with_test” instruction, a “return” is performed together with testing the value in the return value register. The “test” tests the specified register and sets the condition code flags accordingly. The “pop_with_test” instruction performs a “pop” (loading a word from the stack into a specified register or a program counter; popping to PC flushes the pipeline and starts fetching instructions from the address loaded from the stack) together with testing the value in the return value register.
Various signals are required from the control/decode unit. Returning to FIG. 4, the following signals required in one embodiment are:

pcmux_sel 114—Selector signal used to choose if the program counter is going to be updated with the sequential program address or the address given by the return instruction.
wbmux_sel 116—Selector signal used by the writeback stage to determine which address is to be written into the register file or into the program counter. If signal is logic “0,” the address comes from the ALU result from the EX/MEM pipeline stage. If the signal is logic “1,” the address comes from the data memory.
as_ctrl 120—Control signal used to choose if the adder in the ALU will perform subtraction or addition on the operands from the register file.
zeromux_sel 122—Signal used to force input operand B to the ALU to integer value zero.
readreg1 124—Register file register number for operand 1.
readreg2 126—Register file register for operand 2.
loadflag 128—Control signal to allow the status register to update the flag settings.
writeaddr130—Register file register number for the register where the result is written back.

The following table lists exemplary outputs from the control/decode unit in the cycles of the RET instruction. The registers identified in the table are:

R12—the Return Value Register. Test operations are performed on this register.
LR—the Link Register. Keeps the address to return to after the subprogram has completed. LR may also be referred to as the Return Address Register (RAR)
PC—the Program Counter. Holds the address of the currently executing instruction.

The following table lists exemplary outputs from the control/decode unit in the cycles of the return_with_test instruction.



		Textual
Cycle	Control signal output	instruction

1	pcmux_sel = 1	Write the contents
	wbmux_sel = 0	of LR into the PC
	as_ctrl = add	register so that
	zeromux_sel = 0	instruction fetch
	readreg1 = X (don't care value)	will restart from
	readreg2 = LR	this address.
	loadflag = 0
	writeadr = X (don't care value)
2	pcmux_sel = 0	Test the contents
	wbmux_sel = X (don't care value)	of the Return
	as_ctrl = sub	Value Register by
	zeromux_sel = 0	comparing it with
	readreg1 = X (don't care value)	the value 0.
	readreg2 = Return Value	Write the
	Register	resulting flags
	loadflag = 1	into the Flag
	writeadr = X (don't care value)	Register.

The following table lists exemplary outputs from the control/decode unit in the cycles of the pop_with_test instruction.



		Textual
Cycle	Control signal output	instruction

1	pcmux_sel = 1	Write the contents
	wbmux_sel = 1	of the return
	as_ctrl = add	address read from
	zeromux_sel = 0	memory into the PC
	readreg1 = X (don't care value)	register so that
	readreg2 = Pointer Register	instruction fetch
	loadflag = 0	will restart from
	writeadr = X (don't care value)	this address. The
		return address
		resides in a
		memory address
		pointed to by the
		pointer register.
2	pcmux_sel = 0	Test the contents
	wbmux_sel = X (don't care value)	of the Return
	as_ctrl = sub	Value Register by
	zeromux_sel = 0	comparing it with
	readreg1 = X (don't care value)	the value 0.
	readreg2 = Return Value	Write the
	Register	resulting flags
	loadflag = 1	into the Flag
	writeadr = X (don't care value)	Register.

The “return_with_test” and “pop_with_test” instructions can be executed as part of other instructions. For instance, the “return_with_test” instruction can be executed as part of a conditional return instruction, in which there is a return from the subroutine if a specified condition is true. Values are moved into the return register, the return value is tested, and flags are set. A specific example of this instruction is the “ret{cond4}” instruction in the ATMEL AVR32 instruction set. The following pseudocode describes the ret{cond4} instruction (SP is the stack pointer register):

Operation:



	I. If (cond4)

If (Rs != {LR, SP, PC})

R12

Rs;

else if (Rs == LR)

R12

−1;

else if (Rs == SP)

R12

0;

else

R12 4

1;

	Test R12 and set flags;
	PC 4 LR;

	Syntax:
	I. ret{cond4} Rs
	Operands:
	I. cond4 ε {eq, ne, cc/hs, cs/lo, ge, lt, mi, pl, ls, gt,
	le, hi, vs, vc, qs, ab}
	s ε{0, 1, . . . , 15}
	Status Flags:
	Flags are set as result of the operation CP R12, 0.
	V: V 0
	N: N 4 RES[31]
	Z: Z 4 (RES[31:0] == 0)
	C: C 4 0

The following table explains some of the mnemonics used above and the pseudocode for the “Load Multiple Registers” instruction, below:



Mnemonic	Meaning

eq	Equal
ne	Not equal
cc/hs	Higher or same
cs/lo	Lower
ge	Greater than
	or equal
lt	Less than
mi	Minus/negative
pl	Plus/positive
ls	Lower or same
gt	Greater than
le	Less than or
	equal
hi	Higher
vs	Overflow
vc	No overflow
qs	Saturation
al	Always

The operation CP R12, 0 is a comparison or subtraction operation without operation. In this particular case, the result of the operation=R12−0.

Another instruction in which the “return_with_test” operation may be employed is the “Load Multiple Registers” instruction from the AVR 32 instruction set. This instruction loads consecutive words pointed to by the register pointer into the register specified in the instruction. The PC can be loaded, resulting in a jump to the loaded target address. If the PC is loaded, the return value in R12 is tested and the flags are updated. The return value optionally may be set to −1, 0, or 1. The following pseudocode describes this instruction (SP is a stack pointer):



	I.	Loadaddress Rp;

if Reglist16[PC] == 1 then

if Rp == PC then

Loadaddress

SP;

	PC * (Loadaddress++);
	if Rp == PC then

if Reglist16[LR,R12] == B′00

R12

0;

else if Reglist16[LR,R12] == B′01

R12

1;

else

R12

−1;

Test R12 and update flags;

else

	Test R12 and update flags;
	if Reglist16[LR] == 1

LR

*(Loadaddress++);

if Reglist16[SP] == 1

SP

*(Loadaddress++);

if Reglist16[R12] == 1

R12

*(Loadaddress++);

else

if Reglist16[LR] == 1

LR

*(Loadaddress++);

if Reglist16[SP] == 1

SP

*(Loadaddress++);

if Reglist16[R12] == 1

R12

*(Loadaddress++);

for (i = 11 to 0)

if Reglist16[i] == 1 then

Ri

*(Loadaddress++);

if Opcode[++] == 1 then

	if Rp == PC then
	SP Loadaddress;

else

Rp

Loadaddress;

	Syntax:
	I. ldm Rp{++}, Reglist16
	Operands:
	I. Reglist16 ε {R0, R1, R2, . . . , R12, LR, SP, PC}
	p ε {0, 1, . . . , 15}
	Status Flags:
	Flags are only updated if Reglist16[PC] == 1.
	They are set as the result of the operation CP R12, 0.
	V: V 0
	N: N RES [31]
	Z: Z (RES[31:0] == 0)
	C: C 0

Similar instructions employing the “pop_with_test” operation may be employed in which words pointed to by SP are loaded into registers specified in the instruction.
An instruction in which the “pop_with_test” operation may be employed is the Pop Multiple Registers from Stack (“POPM”) instruction from the AVR 32 instruction set. This instruction loads the consecutive words pointed to by SP into the registers specified in the instruction.
While specific examples have been cited above showing how the subroutine return operation may be employed in different instructions, other embodiments may incorporate the subroutine operation into different instructions.
One advantage of the more efficient subroutine return operations is the reduction in code size, since an explicit “test return register” instruction can be eliminated since the test operation may be performed implicitly by the return operation. Another advantage is that execution time is reduced since the return register test is performed in parallel with the fetching of the instruction to which the program will return.
The instructions and operations described above may be employed in both RISC and CISC machines.
Although the present invention has been described in terms of specific exemplary embodiments, one skilled in the art will recognize variations and additions to the embodiments may be made without departing from the principles of the present invention. For instance, return operations may require more or fewer cycles to be executed, or the return operations may be part of different instructions, or the processors executing the return operations may have different architectures. In another embodiment, more hardware may be added so the return operations could be completed in one cycle (i.e., the two micro-operations performed in response to a single instruction are completed in one cycle).

Claims

1. A method for performing a subroutine return operation, the method comprising:

a) fetching an instruction which requires the return operation; and

b) setting status flags based on contents of a return value register in parallel with at least one other operation required to process the return operation, the status flags set before one of the following occurs:

i) contents of a return address register are moved into a program counter; or

ii) a return address is popped from a stack and into the program counter.

2. The method of claim 1 wherein either the contents of the return address register are moved into the program counter or a return address is popped from the stack and into the program counter only when a test operation determines return conditions are satisfied.

3. The method of claim 1 wherein the return operation is split into two micro-operations, one micro-operation for performing the return operation, the other micro-operation for performing a test operation.

4. The method of claim 1 wherein the status flags are set in parallel with writing contents of a return address register through an Arithmetic Logic Unit of a microprocessor.

5. The method of claim 1 wherein the status flags are set in parallel with routing control signals to data memory of a microprocessor past an Arithmetic Logic Unit of the microprocessor.

6. The method of claim 1 wherein steps a) and b) are performed in response to a single instruction.

7. A processor-readable storage medium storing an instruction that, when executed by a processor, causes the processor to perform a method for performing a subroutine return operation, the method comprising:

a) fetching an instruction which requires the return operation; and

b) setting status flags based on contents of a return value register in parallel with at least one other operation required to process the return instruction, the status flags set before one of the following occurs:

i) contents of a return address register are moved into a program counter; or

ii) a return address is popped from a stack and into the program counter.

8. The processor-readable storage medium of claim 7 wherein either the contents of the return address register are moved into the program counter or a return address is popped from the stack and into the program counter only when a test operation determines return conditions are satisfied.

9. The processor-readable storage medium of claim 7 wherein the return operation is split into two micro-operations, one micro-operation for performing the return operation, the other micro-operation for performing a test operation.

10. The processor-readable storage medium of claim 7 wherein the status flags are set in parallel with writing contents of a return address register through an Arithmetic Logic Unit of a microprocessor.

11. The processor-readable storage medium of claim 7 wherein the status flags are set in parallel with routing control signals to data memory of the microprocessor past the Arithmetic Logic Unit of the microprocessor.

12. The processor-readable storage medium of claim 7, the method further comprising flushing a pipeline before either the contents of the return address register or the return address is written into the program counter.

13. The processor-readable storage medium of claim 7 wherein steps a) and b) are performed in response to a single instruction.

14. A method for performing a subroutine return operation comprising:

a) placing a return address into a program counter;

b) executing a test operation on a return value register, the test operation performed in parallel with at least one other operation required to process a return operation;

c) changing a flow of a program to a target address; and

d) performing steps a), b), and c) in response to a single instruction.

15. The method of claim 14 wherein the step of changing the flow of the program is performed only if the test operation determines return conditions are satisfied.

16. The method of claim 14 wherein the subroutine return operation is kept in an instruction decode stage for an additional cycle.

17. The method of claim 14 wherein status flags for the test operation are set in parallel with writing contents of a return address register through an Arithmetic Logic Unit of a microprocessor.

18. The method of claim 14 wherein status flags for the test operation are set in parallel with routing control signals to data memory of a microprocessor located downstream of an Arithmetic Logic Unit of the microprocessor.

19. The method of claim 14 further comprising loading multiple registers.

20. The method of claim 14 further comprising setting a return value in a return value register.

21. A processor-readable storage medium storing an instruction that, when executed by a processor, causes the processor to perform a method for performing a subroutine return operation, the method comprising:

a) placing a return address into a program counter;

c) changing a flow of a program to a target address; and

d) performing steps a), b), and c) in response to a single instruction.

22. The processor-readable storage medium of claim 21 wherein the step of changing the flow of the program is performed only if the test operation determines return conditions are satisfied.

23. The processor-readable storage medium of claim 21 wherein the return operation is kept in an instruction decode stage for an additional cycle.

24. The processor-readable storage medium of claim 21 wherein status flags for the test operation are set in parallel with writing contents of a return address register through an Arithmetic Logic Unit of a microprocessor.

25. The processor-readable storage medium of claim 21 wherein status flags for the test operation are set in parallel with routing control signals to data memory of the microprocessor located downstream of Arithmetic Logic Unit of the microprocessor.

26. The processor-readable storage medium of claim 23, the method further comprising loading multiple registers.

27. The processor-readable storage medium of claim 23, the method further comprising setting a return value in a return value register.