WO2010123541A2 - Method and structure for solving the evil-twin problem - Google Patents
Method and structure for solving the evil-twin problem Download PDFInfo
- Publication number
- WO2010123541A2 WO2010123541A2 PCT/US2010/001099 US2010001099W WO2010123541A2 WO 2010123541 A2 WO2010123541 A2 WO 2010123541A2 US 2010001099 W US2010001099 W US 2010001099W WO 2010123541 A2 WO2010123541 A2 WO 2010123541A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- register
- registers
- size
- instructions
- processor
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 15
- 238000013507 mapping Methods 0.000 claims abstract description 25
- 101150071111 FADD gene Proteins 0.000 description 5
- 238000004590 computer program Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/30098—Register arrangements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/30098—Register arrangements
- G06F9/30105—Register structure
- G06F9/30112—Register structure comprising data of variable length
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/30098—Register arrangements
- G06F9/3012—Organisation of register space, e.g. banked or distributed register file
- G06F9/30138—Extension of register space, e.g. register cache
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline or look ahead
- G06F9/3836—Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution
- G06F9/3838—Dependency mechanisms, e.g. register scoreboarding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/30—Arrangements for executing machine instructions, e.g. instruction decode
- G06F9/38—Concurrent instruction execution, e.g. pipeline or look ahead
- G06F9/3836—Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution
- G06F9/3838—Dependency mechanisms, e.g. register scoreboarding
- G06F9/384—Register renaming
Definitions
- the present invention relates generally to register files of microprocessors, and more particularly to register files that support multiple register sizes simultaneously.
- processors As processors have become more powerful, the size of a data unit processed by such processors has continually increased. For example, some of the earlier processors included register files where each register entry stored a byte of data. The register size in register files in some processors is sixty-four bits for each register.
- Fig. 1 is an example of a register file 100 in which the size of a register has doubled from one generation of a processor to the next generation of that processor.
- the older processor utilized thirty-two thirty-two bit registers, f0 to f31.
- the new processor utilizes thirty-two sixty-four bit registers d0 to d31.
- register d0 includes register f0 and register fl.
- register fl is referred to as the twin of register f0.
- Each of registers d0 to dl5 includes two thirty-two bit registers and so each of registers d0 to dl5 includes twin registers.
- the thirty-two bit register in the most significant bits of the sixty-four bit register is referred to as the evil twin of the thirty-two bit register in the least significant bits of that sixty- four bit register.
- Fig. IB is a segment of a computer program 150 that includes instructions that are coupled but the coupling is not detected by the comparators and so computer program 150 suffers from the evil-twin problem.
- load instruction Id loads a value in register fl, which is the thirty-two most significant bits of register d0 and is the evil-twin register.
- Floating pointing double precision add instruction fdadd in the next line, adds the value in register dO to the value in register d4 and places the resulting value in register d8.
- Register dO is a sixty-four bit register and so includes any value in register fl, as indicated in Fig. IA.
- the comparators which are used in determining instructions on which floating point double precision add instruction fdadd depends, do not detect the dependence of floating point double precision add instruction fdadd on the result of load instruction Id in evil-twin register fl.
- the floating point double precision add instruction fdadd could be performed before load instruction Id completed.
- the accuracy of the result of the floating point double precision add instruction fdadd is unknown, and is most likely incorrect because the result of instruction Id is not in evil-twin register f1.
- sixty-four bit register d8 includes thirty-two bit register f17 , also as shown in Fig. IA.
- Floating point add instruction fadd adds the value in register fl5 to the value in evil-twin register fl7 and places the resulting value in register f23.
- register fl7 does not include the correct value until after floating point double precision add instruction fdadd completes.
- the dependence of floating point add instruction fadd on the result of floating point double precision add instruction fdadd is not detected by the comparators. Thus, floating point add instruction fadd could be performed before floating point double precision add instruction fdadd completed.
- the problem is the inability of the processor to detect the dependence of an instruction on a value in an evil-twin register. As noted, this problem is referred to as the evil -twin problem.
- each of registers fl, f3 , ... is the referred to as the evil twin of registers f ⁇ , f2, respectively.
- This evil-twin problem has been accepted because the mapping of the different sized registers to the same register file minimizes the area required by the register file while permitting execution of different generations of computer code on a processor.
- some processors for example, prohibit launching a double precision instruction until all single precision instructions have completed, and conversely. The delay- in launching stalls the pipeline until execution is completed for the one set of instructions.
- a processor eliminates the evil twin problem.
- the solution to the evil twin problem (i) minimizes the area on the chip required to support the solution; (ii) can be implemented in a content addressable memory with a single port; and
- the evil twin problem is solved in a processor supporting (i) a first instruction set having first instructions with sources and destinations of a first size, n-bits, where n is an integer, and (ii) a second instruction set having second instructions having sources and destinations of a second size m-bits, where m is an integer.
- the second size, m-bits is an even integer multiple of the first size n-bits.
- a register file in the processor in includes a first plurality of registers of the first size .
- a decoder uses a mapping that divides the register file into a second plurality M of registers having the second size.
- the second size is the largest size register used to store data in the register file.
- each of the registers having the second size, m-bits, is assigned a different name in a continuous name space.
- Each register of the second size includes a third plurality N of registers of the first size, n- bits, where N is an even integer. As indicated above, integer m is an even multiple N of integer n.
- the first size is the smallest size register used to store data in the register file.
- the mapping assigns each register in the third plurality N of registers the same name as the register of the second size that includes that plurality.
- the register file is mapped into (M * N) n-bit registers so that the first plurality is (M * N) registers.
- State information is maintained in the register file for each n-bit register. In one embodiment, the state information is provided by a valid bit.
- intermediate size registers There can be a set of intermediate size registers that are larger than the first size but smaller than the second size, the intermediate size registers also are mapped to the same name as the name of the m-bit register containing the intermediate size registers.
- a larger intermediate sized register also is an even integer multiple of the smallest sized register.
- a processor includes a register file comprising a plurality of registers of a first size and a decode unit.
- the decode unit decodes instructions from a first instruction set having sources and destinations of the first size, and instructions from a second instruction set having sources and destinations of a second size.
- the second size is a multiple n of the first size where n is an even integer.
- the decode unit also includes a mapping unit for mapping register names for instructions in the first instruction set and register names for instructions in the second instruction set into a common name space.
- a register name in the common name space is a same name (i) for each register in a set of n registers in the plurality of registers and (ii) for a register of the second size including the set of n registers.
- the common name space is a continuous name space for the register file.
- Each register of the first size has a corresponding valid bit to provide state information for that register.
- the processor also includes a Scoreboard having an entry for each register of the first size, and an execution pipe coupled to the register file.
- a valid bit is asserted only for the one register.
- a valid bit for each of the n registers is asserted.
- the register file is included in a content addressable memory with only a single port.
- the register file can be, for example, a working register file, an architectural register file, or both.
- a method for solving an evil twin problem in a processor, supporting a first instruction set having first instructions with sources and destinations of a first size, and a second instruction set having second instructions having sources and destinations of a second size includes decoding, by a decoding pipe of the processor, the first instructions and the second instructions.
- the second size is a multiple n of the first size where n is an even integer.
- the decoding further includes mapping register names for the first instructions and register names for the second instructions into a common name space for a register file comprising a plurality of registers of the first size.
- a register name in the common name space is a same name (i) for each register in a set of n registers in the plurality of registers and (ii) for a register of the second size including the set of n registers.
- the common name space is a continuous name space for the register file.
- the method also includes maintaining, in the register file, separate state information for each register of the first size in the register file.
- Fig. 1 is a block diagram of a mapping of a prior art register file.
- FIG. 2 is a high level block diagram of a processor including a plurality of cores in which at least one core includes a continuous name space mapped register file according to one embodiment of this invention.
- FIG. 3 is a more detailed block diagram of a representative core of the processor of Fig. 2 according to one embodiment of the invention.
- Fig. 4 is an illustration of the novel mapping of a register file according to one embodiment of the invention.
- FIG. 5 is block diagram of a system that includes a processor that includes and utilizes a continuous name space mapped register file according to an embodiment of this invention.
- a processor 200 includes a plurality of cores 210-i, where i ranges from 1 to S, where S is an integer.
- the plurality of cores 210-i may have a common registers file or each core 210-i may have its own register files, or some combination of global and local register files. Irrespective of the particular configuration of the register files in processor 200, a novel mapping of registers in a register file into a continuous name space allows the evil twin problem to be recognized and to be solved.
- pipeline 310 also includes a fetch pipe 311 having one or more stages.
- Fetch pipe 311 is coupled to an instruction store 318, e.g., a level two (L2) cache memory.
- An output of fetch pipe 311 is provided to a fetch buffer 312.
- fetch buffer 312 is configured to service multiple threads.
- Decode pipe 313 includes one or more stages that function to decode instructions.
- the instructions can include at least instructions from two different instructions sets. Instructions in a first instruction set utilize registers, as sources and destinations, of a first size. Instructions in a second instruction set utilize registers, as sources and destinations, of a second size.
- the registers used for storing inputs to an instruction, i.e., operands, are referred to as sources, while the registers used for storing the results of the executed instructions are referred to as destinations.
- the second size is different from the first size.
- the second size is a multiple n of the first size where n is an even integer.
- one stage of decode pipe 313 includes a map unit 340, which maps the names of registers in register file 320 into a continuous name space to eliminate the evil twin problem associated with the at least two different register sizes.
- Architectural register file 330 is mapped in a similar way.
- Decode pipe 313 is coupled to an instruction queue 314, which serves to decouple decode pipe 313 from later stages of pipeline 310.
- working register file (WRF) 320 and architectural register file (ARF) 330 are coupled to decode pipe 313, an execution pipe 316, and a trap and commit unit 317.
- Instructions stored in instruction queue 314 are grouped, by grouping unit 315, for execution by execution pipe 316.
- Execution pipe 316 is coupled to trap and commit unit 317, which commits executed instructions to architectural state in architectural register file 330.
- a data cache 319 is coupled to execution pipe 316. Data cache 319 provides data to execution pipe 316.
- register files 320 and 330 support at least two register sizes where one size is an even multiple of the other size.
- Decode pipe 313 decodes at least computer programs including instructions from the first instruction set and instructions from the second instruction set.
- the first size is taken as thirty-two bits and even integer n is taken as two.
- the second size is sixty-four bits. The use of these sizes is illustrative only and is not intended to limit the invention to this specific example.
- Fig. 4 is an example illustrating the mapping of working register file 320 by map unit 340 into a continuous name space.
- Working register file 320 includes thirty-two sixty-four bit registers, in this example.
- Register file 320 is mapped into sixty-four thirty-two bit registers, i.e., a plurality of registers of the first size.
- a Scoreboard 350 associated with execution pipe 316, maintains status information for each of the sixty-four thirty-two bit registers .
- each pair of thirty-two registers and the sixty-four bit register including that pair share a common name.
- One register in the pair is the low register for the name and the other register in the pair is the high register for that name.
- the names range from DO to D31 and the lower register is represented by "Low, " while the upper register is represented by "High.”
- the names DO to D31, High, Low, and Valid in Fig. 4 are reference numerals; are used for ease of understanding; and should not be interpreted as requiring the names to be stored in working register file 320.
- a reference numeral of Di_H is used for the high portion of register Di;.
- a reference numeral Di_L is used for the low portion of register Di; and reference numeral Di includes both registers Di_L and Di_H.
- register Di includes a set of twins, registers Di_H and Di_L, and all three registers are represented by a common register name Di.
- a separate signal is used to indicate the high and low portions respectively of register Di.
- each portion of a register Di has a valid bit, which is used as described more completely below.
- register fl is mapped to the high portion of register DO and is represented by reference numeral D0_H.
- Registers d ⁇ , d4 , and d8 are mapped to registers DO, D4 , and D8 respectively.
- Register fl5 is mapped to the high portion of register D7 and is represented by reference numeral D7_H;
- register fl7 is mapped to the high portion of register D8 and is represented by reference numeral D8_H;
- register f23 is mapped to the high portion of register D12 and is represented by reference numeral D12_H.
- the instructions after the mapping are shown in Fig. 4B as code segment 450.
- the mapping in the decoder divides a register file into a second plurality of registers M having a second size, where the second size is a second number m of bits.
- the second size is the largest size register used to store data in the register file.
- Each of the registers having the second size, m-bits, is assigned a different name in a continuous name space.
- the register file includes a first plurality of registers of a first size, as defined below.
- Each register of the second size includes a third plurality N of registers of the first size, where the first size is a first number n of bits and integer m is an even multiple N of integer n.
- the first size is the smallest size register used to store data in the register file.
- Each register in the third plurality of registers has the same name as the register of the second size that includes that plurality.
- the register file is mapped into (M * N) n-bit registers. State information is maintained for each n-bit register. There can be intermediate size registers that are larger than the first size but smaller than the second size. A larger intermediate sized register also is an even integer multiple of the smallest sized register. When a particular register is written to, all the valid bits associated with that register size are asserted.
- FIG. 5 depicts an example of a computer system 500 constructed according to various aspects of the present invention.
- Computer system 500 includes a processor 502 with at least one continuous name space mapped register file.
- Processor 502 may include multiple cores, be a single threaded processor, be a multi-threaded processor, etc.
- computer system 500 is a stand alone system, while in another embodiment computer system 500 is a server computer that is networked and is part of a client- server system.
- Computer system 500 also includes system memory 501, e.g., one or more cache levels, SRAM(Static Random Access Memory) , DRAM (Dynamic Random Access Memory) , RDRAM (Rambus DRAM) , EDO RAM (Extended Data Output Random Access Memory) , DDR RAM (Double Data Rate Random Access Memory) , and/or EEPROM (Electrically- Erasable Programmable Read Only Memory), etc.), a system bus 505 (e.g., LDT (Lightning Data Transport), PCI (Peripheral Component Interconnect) , ISA (Industry Standard Architecture), etc.), a network interface 503 (e.g., an ATM (Asynchronous Transfer Mode) interface, an Ethernet interface, a Frame Relay interface, etc.), and storage 504, e.g., optical storage, magnetic storage, etc.
- system memory 501 e.g., one or more cache levels, SRAM(Static Random Access Memory) , DRAM (Dyna
- Computer system realizations of the invention may include fewer or additional components not illustrated in FIG. 5.
- computer system 500 may also include a video card, an audio card " , additional network interfaces, peripheral devices, etc.
- Processor 502, storage 504, network interface 503 and system memory 501 are coupled to system bus 505, which includes a plurality of address, data and control lines.
- Processor 502 includes instruction store 318 and a memory interface that couples the instruction store 318 to a memory hierarchy, i.e., system memory 501.
- the computing apparatus described herein may be embodied as a design file representation including machine readable instructions encoded on one or more computer readable media .
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Executing Machine-Instructions (AREA)
- Advance Control (AREA)
Abstract
A register file, in a processor, includes a first plurality of registers of a first size, n-bits. A decoder uses a mapping that divides the register file into a second plurality M of registers having a second size. Each of the registers having the second size is assigned a different name in a continuous name space. Each register of the second size includes a plurality N of registers of the first size, n-bits. Each register in the plurality N of registers is assigned the same name as the register of the second size that includes that plurality. State information is maintained in the register file for each n-bit register. The dependence of an instruction on other instructions is detected through the continuous name space. The state information allows the processor to determine when the information in any portion, or all, of a register is valid.
Description
METHOD AND STRUCTURE FOR SOLVING THE EVIL-TWIN PROBLEM
Shailender Chaudhry Marc Tremblay
BACKGROUND OF THE INVENTION
Field of the Invention
[0001] The present invention relates generally to register files of microprocessors, and more particularly to register files that support multiple register sizes simultaneously.
Description of Related Art
[0002] As processors have become more powerful, the size of a data unit processed by such processors has continually increased. For example, some of the earlier processors included register files where each register entry stored a byte of data. The register size in register files in some processors is sixty-four bits for each register.
[0003] Typically, when a new processor is introduced, the processor is configured to process computer code that was written for earlier processors with smaller register sizes as well as computer code written for the largest register size available in the new processer, i.e., backward compatibility for computer programs is maintained in the new processor. Fig. 1 is an example of a register file 100 in which the size of a register has doubled from one generation of a processor to the next generation of that processor. For example, the older processor utilized thirty-two thirty-two bit registers, f0 to f31. The new processor utilizes thirty-two sixty-four bit registers d0 to d31.
[0004] As shown in Fig. 1, the least significant thirty-two bits of register dO is register fO and the most significant thirty-two bits of register d0 is register f1. Thus, register d0 includes register f0 and register fl. Herein, register fl is referred to as the twin of register f0. Each of registers d0 to dl5 includes two thirty-two bit registers and so each of registers d0 to dl5 includes twin registers. [0005] The thirty-two bit register in the most significant bits of the sixty-four bit register is referred to as the evil twin of the thirty-two bit register in the least significant bits of that sixty- four bit register. Use of register file 100 for executing instructions that use both thirty-two bit and sixty-four bit registers can lead to a problem, which is referred to as the "evil-twin" problem. [0006] If two instructions are coupled through a thirty-two bit register in the most significant bits for a sixty-four bit register, i.e., coupled by the evil-twin register, the processor normally does not detect the coupling. Thus, incorrect results may be obtained as a result of the evil-twin problem. [0007] As an example, the comparators for instructions using sixty-four bit operands do not detect an instruction using the thirty-two-most significant bits of the sixty-four bit register as a destination. Fig. IB is a segment of a computer program 150 that includes instructions that are coupled but the coupling is not detected by the comparators and so computer program 150 suffers from the evil-twin problem.
[0008] In Fig. IB, load instruction Id loads a value in register fl, which is the thirty-two most significant bits of register d0 and is the evil-twin register. Floating pointing double precision add instruction fdadd, in the next line, adds the value in
register dO to the value in register d4 and places the resulting value in register d8. Register dO is a sixty-four bit register and so includes any value in register fl, as indicated in Fig. IA.
[0009] However, the comparators, which are used in determining instructions on which floating point double precision add instruction fdadd depends, do not detect the dependence of floating point double precision add instruction fdadd on the result of load instruction Id in evil-twin register fl. Thus, the floating point double precision add instruction fdadd could be performed before load instruction Id completed. In this case, the accuracy of the result of the floating point double precision add instruction fdadd is unknown, and is most likely incorrect because the result of instruction Id is not in evil-twin register f1.
[0010] Similarly, sixty-four bit register d8 includes thirty-two bit register f17 , also as shown in Fig. IA. Floating point add instruction fadd adds the value in register fl5 to the value in evil-twin register fl7 and places the resulting value in register f23. However, register fl7 does not include the correct value until after floating point double precision add instruction fdadd completes. [0011] The dependence of floating point add instruction fadd on the result of floating point double precision add instruction fdadd is not detected by the comparators. Thus, floating point add instruction fadd could be performed before floating point double precision add instruction fdadd completed. In this case, the accuracy of the result of floating point add instruction fadd is unknown, and is most likely incorrect because the result of instruction fdadd is not in evil-twin register f17.
[0012] In each of these instances, the problem is the inability of the processor to detect the dependence of an instruction on a value in an evil-twin register. As noted, this problem is referred to as the evil -twin problem.
[0013] In this example, each of registers fl, f3 , ... is the referred to as the evil twin of registers fθ, f2, respectively. This evil-twin problem has been accepted because the mapping of the different sized registers to the same register file minimizes the area required by the register file while permitting execution of different generations of computer code on a processor. Moreover, to avoid the evil -twin problem, some processors, for example, prohibit launching a double precision instruction until all single precision instructions have completed, and conversely. The delay- in launching stalls the pipeline until execution is completed for the one set of instructions.
SUMMARY OF THE INVENTION
[0014] In one embodiment, a processor eliminates the evil twin problem. The solution to the evil twin problem (i) minimizes the area on the chip required to support the solution; (ii) can be implemented in a content addressable memory with a single port; and
(iii) does not require stalling the pipeline. Thus, in contrast to the prior art, instructions in a computer program having different size operands and different size results can be intermingled. The intermingled instructions are executed properly without the prior art special precautions to avoid the evil-twin problem.
[0015] In one embodiment, the evil twin problem is solved in a processor supporting (i) a first instruction set having first instructions with sources and destinations of a first size, n-bits, where n is an
integer, and (ii) a second instruction set having second instructions having sources and destinations of a second size m-bits, where m is an integer. The second size, m-bits, is an even integer multiple of the first size n-bits. A register file in the processor in includes a first plurality of registers of the first size .
[0016] A decoder uses a mapping that divides the register file into a second plurality M of registers having the second size. The second size is the largest size register used to store data in the register file. In the mapping, each of the registers having the second size, m-bits, is assigned a different name in a continuous name space.
[0017] Each register of the second size includes a third plurality N of registers of the first size, n- bits, where N is an even integer. As indicated above, integer m is an even multiple N of integer n. The first size is the smallest size register used to store data in the register file. The mapping assigns each register in the third plurality N of registers the same name as the register of the second size that includes that plurality.
[0018] Thus, the register file is mapped into (M * N) n-bit registers so that the first plurality is (M * N) registers. State information is maintained in the register file for each n-bit register. In one embodiment, the state information is provided by a valid bit.
[0019] There can be a set of intermediate size registers that are larger than the first size but smaller than the second size, the intermediate size registers also are mapped to the same name as the name of the m-bit register containing the intermediate size registers. A larger intermediate sized register also is an even integer multiple of the smallest sized
register. When a particular register is written to, all the valid bits associated with that register size are asserted.
[0020] Since a register of the second size and all registers of the first size included in that register have a common name, register dependencies between instructions that utilize different sized registers is easily ascertained. Also, since state information is maintained in the register file for each register of the first size, the processor can determine when an instruction, which utilizes all or part of a register of the second size, is ready and so can issue the instruction for execution at an appropriate time without stalling the pipeline. The unique mapping in combination with the state information eliminates the evil-twin problem for the processor. [0021] In one embodiment, a processor includes a register file comprising a plurality of registers of a first size and a decode unit. The decode unit decodes instructions from a first instruction set having sources and destinations of the first size, and instructions from a second instruction set having sources and destinations of a second size. The second size is a multiple n of the first size where n is an even integer.
[0022] The decode unit also includes a mapping unit for mapping register names for instructions in the first instruction set and register names for instructions in the second instruction set into a common name space. A register name in the common name space is a same name (i) for each register in a set of n registers in the plurality of registers and (ii) for a register of the second size including the set of n registers. The common name space is a continuous name space for the register file. Each register of the
first size has a corresponding valid bit to provide state information for that register. [0023] The processor also includes a Scoreboard having an entry for each register of the first size, and an execution pipe coupled to the register file. Upon the execution pipe writing a result to only one register in a set of n registers, a valid bit is asserted only for the one register. Upon the execution pipe writing a result to all registers in a set of n registers, a valid bit for each of the n registers is asserted.
[0024] The register file is included in a content addressable memory with only a single port. The register file can be, for example, a working register file, an architectural register file, or both. [0025] A method for solving an evil twin problem in a processor, supporting a first instruction set having first instructions with sources and destinations of a first size, and a second instruction set having second instructions having sources and destinations of a second size, includes decoding, by a decoding pipe of the processor, the first instructions and the second instructions. The second size is a multiple n of the first size where n is an even integer. [0026] The decoding further includes mapping register names for the first instructions and register names for the second instructions into a common name space for a register file comprising a plurality of registers of the first size. A register name in the common name space is a same name (i) for each register in a set of n registers in the plurality of registers and (ii) for a register of the second size including the set of n registers. The common name space is a continuous name space for the register file. The method also includes maintaining, in the register file,
separate state information for each register of the first size in the register file.
BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Fig. 1 is a block diagram of a mapping of a prior art register file.
[0028] Fig. 2 is a high level block diagram of a processor including a plurality of cores in which at least one core includes a continuous name space mapped register file according to one embodiment of this invention.
[0029] Fig. 3 is a more detailed block diagram of a representative core of the processor of Fig. 2 according to one embodiment of the invention. [0030] Fig. 4 is an illustration of the novel mapping of a register file according to one embodiment of the invention.
[0031] Fig. 5 is block diagram of a system that includes a processor that includes and utilizes a continuous name space mapped register file according to an embodiment of this invention.
[0032] In the drawings, elements with the same reference numeral are the same or equivalent elements. Also, the first digit of a reference numeral is the figure number of the figure in which that element first appears .
DETAILED DESCRIPTION
[0033] According to one embodiment of this invention, a processor 200 includes a plurality of cores 210-i, where i ranges from 1 to S, where S is an integer. The plurality of cores 210-i may have a common registers file or each core 210-i may have its own register files, or some combination of global and local register files. Irrespective of the particular
configuration of the register files in processor 200, a novel mapping of registers in a register file into a continuous name space allows the evil twin problem to be recognized and to be solved.
[0034] The solution of utilizing a continuous name space for the register file allows sharing of a CAM (Content -Addressable Memory) port when the register file is implemented in a CAM and so does not require an additional CAM port. Thus, the amount of space on the die required to implement the solution to the evil twin problem is limited to the space required to implement the mapping in the decoder and a small amount of additional space in the CAM for additional valid bits, as described more completely below. Also, unlike the prior art, the solution does not requiring stalling the pipeline until an instruction or instructions using a particular size of register have completed execution. [0035] Before considering the solution to the evil twin problem in further detail, normal operation of a pipeline 310 (Fig. 3) in a core 210-i is considered. [0036] Pipeline 310 executes instructions and writes the results to a working register file 320. When an instruction is retired, the result for that instruction in working register file 320 is written to an architectural register file 330.
[0037] As illustrated in Fig. 3A, pipeline 310 also includes a fetch pipe 311 having one or more stages. Fetch pipe 311 is coupled to an instruction store 318, e.g., a level two (L2) cache memory. An output of fetch pipe 311 is provided to a fetch buffer 312. In at least one embodiment, fetch buffer 312 is configured to service multiple threads.
[0038] Decode pipe 313 includes one or more stages that function to decode instructions. The instructions can include at least instructions from two different instructions sets. Instructions in a first instruction
set utilize registers, as sources and destinations, of a first size. Instructions in a second instruction set utilize registers, as sources and destinations, of a second size. The registers used for storing inputs to an instruction, i.e., operands, are referred to as sources, while the registers used for storing the results of the executed instructions are referred to as destinations. The second size is different from the first size. The second size is a multiple n of the first size where n is an even integer.
[0039] As explained more completely below, one stage of decode pipe 313 includes a map unit 340, which maps the names of registers in register file 320 into a continuous name space to eliminate the evil twin problem associated with the at least two different register sizes. Architectural register file 330 is mapped in a similar way. Decode pipe 313 is coupled to an instruction queue 314, which serves to decouple decode pipe 313 from later stages of pipeline 310. [0040] In this example, working register file (WRF) 320 and architectural register file (ARF) 330 are coupled to decode pipe 313, an execution pipe 316, and a trap and commit unit 317. [0041] Instructions stored in instruction queue 314 are grouped, by grouping unit 315, for execution by execution pipe 316. Execution pipe 316 is coupled to trap and commit unit 317, which commits executed instructions to architectural state in architectural register file 330. A data cache 319 is coupled to execution pipe 316. Data cache 319 provides data to execution pipe 316.
[0042] Hence, in this example, register files 320 and 330 support at least two register sizes where one size is an even multiple of the other size. Decode pipe 313 decodes at least computer programs including
instructions from the first instruction set and instructions from the second instruction set. [0043] In the following example, the first size is taken as thirty-two bits and even integer n is taken as two. Thus, the second size is sixty-four bits. The use of these sizes is illustrative only and is not intended to limit the invention to this specific example. In view of the disclosure, those knowledgeable in the field can extend the application, for example, to four sets of instructions with sources and destinations that are even integer multiples of each other, e.g., instructions with eight bit sources and destinations; instructions with sixteen bit sources and destinations; instructions with thirty-two bit sources and destinations; and instructions with sixty- four bit sources and destinations. Alternatively, instructions with thirty-two bit sources and destinations; instructions with sixty- four bit sources and destinations; and instructions with one-hundred- twenty-eight bit sources and destinations could be utilized.
[0044] Fig. 4 is an example illustrating the mapping of working register file 320 by map unit 340 into a continuous name space. Working register file 320 includes thirty-two sixty-four bit registers, in this example.
[0045] Register file 320 is mapped into sixty-four thirty-two bit registers, i.e., a plurality of registers of the first size. A Scoreboard 350, associated with execution pipe 316, maintains status information for each of the sixty-four thirty-two bit registers .
[0046] To allow the use of a single CAM port for working register file 320, each pair of thirty-two registers and the sixty-four bit register including that pair share a common name. One register in the
pair is the low register for the name and the other register in the pair is the high register for that name. In the example of Fig. 4, the names range from DO to D31 and the lower register is represented by "Low, " while the upper register is represented by "High." The names DO to D31, High, Low, and Valid in Fig. 4 are reference numerals; are used for ease of understanding; and should not be interpreted as requiring the names to be stored in working register file 320.
[0047] Also, herein to differentiate between the high and low portions of a register Di, where i ranges from 0 to 31, a reference numeral of Di_H is used for the high portion of register Di;. a reference numeral Di_L is used for the low portion of register Di; and reference numeral Di includes both registers Di_L and Di_H. In this implementation, register Di includes a set of twins, registers Di_H and Di_L, and all three registers are represented by a common register name Di. A separate signal is used to indicate the high and low portions respectively of register Di. Also, each portion of a register Di has a valid bit, which is used as described more completely below.
[0048] Consider code segment 150 in Fig. IB and the results obtained by map unit 340 for mapping the registers in code segment 150 to working register file 320 in Fig. 4. Register fl is mapped to the high portion of register DO and is represented by reference numeral D0_H. Registers dθ, d4 , and d8 are mapped to registers DO, D4 , and D8 respectively. Register fl5 is mapped to the high portion of register D7 and is represented by reference numeral D7_H; register fl7 is mapped to the high portion of register D8 and is represented by reference numeral D8_H; and register f23 is mapped to the high portion of register D12 and is
represented by reference numeral D12_H. The instructions after the mapping are shown in Fig. 4B as code segment 450.
[0049] Since the comparators compare the register names, i.e., DO to D31, the comparators detect the dependency of the various instructions. Thus, instruction fdadd is not executed until instruction Id completes, and instruction fadd is not executed until instruction fdadd completes. The mapping to a continuous name space permits detection of the evil twin problem and a solution by preventing execution until the proper data becomes available. [0050] Specifically, when a value is loaded to register D0_H, the valid bit for register D0_H is asserted and value of the valid bit for register D0_L is unchanged. Similarly, when a value is loaded or written to register D0_L, the valid bit for that register is asserted and the valid bit for register D0_H is unchanged. If a value is loaded or written to a register Di, where i ranges from 0 to 31, both the upper and lower valid bits are asserted. A valid bit is de-asserted when the destination register becomes available for bypass, either from an intermediate register or from the register file itself. [0051] Thus, not only can the floating point double precision add instruction fdadd detect the dependence of the instruction on the value in register D0_H based on the common name, but also when the valid bit for register D0_H is read, the instruction can determine whether the read value is valid. Similarly, if an instruction needs a value in register Di, the value in register Di is not valid until both valid bits are asserted. Thus, the dependence of an instruction is not only detected through the continuous name space, but also the valid bits allow the processor to
determine when the information in any portion, or all, of register Di is valid.
[0052] Since a register of the second size and all registers of the first size included in that register have a common name, dependencies between instructions that utilize different sized registers is easily ascertained as illustrated above. Also, since state information is maintained in the register file for each register of the first size, the processor can determine when an instruction that utilizes all or part of a register of the second size is ready and so can issue the instruction for execution at an appropriate time without stalling the pipeline. The unique mapping in combination with the state information eliminates the evil-twin problem for the processor. [0053] The mapping to a continuous name space in conjunction with the use of the valid bits is different from register renaming in an out-of-order processor. In register renaming, the registers are assigned names based on program order and typically are not continuous and so register renaming requires double the number of CAM ports.
[0054] In more general terms, the mapping in the decoder divides a register file into a second plurality of registers M having a second size, where the second size is a second number m of bits. The second size is the largest size register used to store data in the register file. Each of the registers having the second size, m-bits, is assigned a different name in a continuous name space. The register file includes a first plurality of registers of a first size, as defined below.
[0055] Each register of the second size includes a third plurality N of registers of the first size, where the first size is a first number n of bits and integer m is an even multiple N of integer n. The
first size is the smallest size register used to store data in the register file. Each register in the third plurality of registers has the same name as the register of the second size that includes that plurality.
[0056] Thus, the register file is mapped into (M * N) n-bit registers. State information is maintained for each n-bit register. There can be intermediate size registers that are larger than the first size but smaller than the second size. A larger intermediate sized register also is an even integer multiple of the smallest sized register. When a particular register is written to, all the valid bits associated with that register size are asserted.
[0057] Fig. 5 depicts an example of a computer system 500 constructed according to various aspects of the present invention. Computer system 500 includes a processor 502 with at least one continuous name space mapped register file. Processor 502 may include multiple cores, be a single threaded processor, be a multi-threaded processor, etc. In one embodiment computer system 500 is a stand alone system, while in another embodiment computer system 500 is a server computer that is networked and is part of a client- server system.
[0058] Computer system 500 also includes system memory 501, e.g., one or more cache levels, SRAM(Static Random Access Memory) , DRAM (Dynamic Random Access Memory) , RDRAM (Rambus DRAM) , EDO RAM (Extended Data Output Random Access Memory) , DDR RAM (Double Data Rate Random Access Memory) , and/or EEPROM (Electrically- Erasable Programmable Read Only Memory), etc.), a system bus 505 (e.g., LDT (Lightning Data Transport), PCI (Peripheral Component Interconnect) , ISA (Industry Standard Architecture), etc.), a network interface 503 (e.g., an ATM (Asynchronous Transfer Mode) interface,
an Ethernet interface, a Frame Relay interface, etc.), and storage 504, e.g., optical storage, magnetic storage, etc.
[0059] Computer system realizations of the invention may include fewer or additional components not illustrated in FIG. 5. For example, computer system 500 may also include a video card, an audio card", additional network interfaces, peripheral devices, etc. Processor 502, storage 504, network interface 503 and system memory 501 are coupled to system bus 505, which includes a plurality of address, data and control lines. Processor 502 includes instruction store 318 and a memory interface that couples the instruction store 318 to a memory hierarchy, i.e., system memory 501. It should be appreciated that the computing apparatus described herein may be embodied as a design file representation including machine readable instructions encoded on one or more computer readable media .
Claims
1. A processor comprising: a register file comprising a plurality of registers of a first size,- and a decode unit for decoding first instructions having sources and destinations of the first size, and second instructions having sources and destinations of a second size, where the second size is an even integer multiple n of the first size, the decode unit further comprising: a mapping unit for mapping register names for the first instructions and register names for the second instructions into a common name space, wherein a register name in the common name space is a same register name (i) for each register in a set of n registers in the plurality of registers and (ii)for a register of the second size including the set of n registers; and said common name space is a continuous name space for the register file.
2. The processor of Claim 1 wherein each register of the first size has a corresponding valid bit.
3. The processor of Claim 1 further comprising: a Scoreboard having an entry for each register of the first size.
4. The processor of Claim 1 further comprising: an execution pipe coupled to the register file, wherein upon the execution pipe writing a result to only one register in a set of n registers, a valid bit is asserted only for the one register.
5. The processor of Claim 1 further comprising: an execution pipe coupled to the register file, wherein upon the execution pipe writing a result to all registers in a set of n registers, a valid bit for each of the n registers is asserted.
6. The processor of Claim 4 further wherein upon the execution pipe writing a result to all registers in the set of n registers, a valid bit for each of the n registers is asserted.
7. The processor of Claim 1 wherein n is two and the first size is thirty-two bits.
8. The processor of Claim 1 further comprising: a content addressable memory with only a shared port, wherein said register file is included in said content addressable memory.
9. The processor of Claim 1 wherein the register file is a working register file.
10. A method for solving an evil twin problem in a processor supporting first instructions with sources and destinations of a first size, and second instructions having sources and destinations of a second size, where the second size is an even integer multiple n of the first size, the method comprising decoding, by a decoding pipe of the processor, the first instructions with sources and destinations of the first size and the second instructions having sources and destinations of the second size, wherein the decoding further comprises : mapping register names for the first instructions and register names for the second instructions into a common name space for a register file comprising a plurality of registers of a first size, wherein a register name in the common name space is a same register name (i) for each register in a set of n registers in the plurality of registers and (ii)for a register of the second size including the set of n registers; and said common name space is a continuous name space for the register file.
11. The method of Claim 10 further comprising: maintaining separate state information for each register of the first size in the register file.
12. The method of Claim 10 further comprising: maintaining an entry for each of the registers of the first size in a Scoreboard.
13. The method of Claim 10 further comprising: asserting only a valid bit for one register in a set of n registers upon an execution pipe in the processor writing a result to only the one register in the set of n registers.
14. The method of Claim 10 further comprising: asserting a valid bit for each register in a set of n registers upon an execution pipe writing a result to all registers in the set of n registers .
15. The method of C] aim 13 further comprising: asserting a valid bit for each register in the set of n registers upon the execution pipe writing a result to all registers in the set of n registers.
16. The method of Claim 10 wherein n is two and the first size is thirty-two bits.
17. The method of Claim 10 wherein the register file is included in a content addressable memory with only a shared port .
18. The method of Claim 10 wherein the register file is a working register file.
19. A system comprising: a memory; and a processor, coupled to the memory, comprising: a register file comprising a plurality of registers of a first size; a decode unit for decoding first instructions having sources and destinations of the first size, and second instructions having sources and destinations of a second size, where the second size is an even integer multiple n of the first size, the decode unit further comprising: a mapping unit for mapping register names for the first instructions and register names for the second instructions into a common name space, wherein a register name in the common name space is a same register name (i) for each register in a set of n registers in the plurality of registers and (ii)for a register of the second size including the set of n registers; and said common name space is a continuous name space for register file.
20. The system of Claim 18 further wherein the processor further comprises: an execution pipe coupled to the register file, wherein upon the execution pipe writing a result to only one register in a set of n registers, a valid bit is asserted for only the one register; and upon the execution pipe writing a result to all registers in the set of n registers, a valid bit for each of the n registers is asserted.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP10721863A EP2422270A2 (en) | 2009-04-20 | 2010-04-13 | Method and structure for solving the evil-twin problem |
CN201080010412.3A CN102356377B (en) | 2009-04-20 | 2010-04-13 | For solving the method and structure of harmful twin problem |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US12/426,550 US8898436B2 (en) | 2009-04-20 | 2009-04-20 | Method and structure for solving the evil-twin problem |
US12/426,550 | 2009-04-20 |
Publications (2)
Publication Number | Publication Date |
---|---|
WO2010123541A2 true WO2010123541A2 (en) | 2010-10-28 |
WO2010123541A3 WO2010123541A3 (en) | 2010-12-16 |
Family
ID=42790883
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2010/001099 WO2010123541A2 (en) | 2009-04-20 | 2010-04-13 | Method and structure for solving the evil-twin problem |
Country Status (4)
Country | Link |
---|---|
US (1) | US8898436B2 (en) |
EP (1) | EP2422270A2 (en) |
CN (1) | CN102356377B (en) |
WO (1) | WO2010123541A2 (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9727336B2 (en) * | 2011-09-16 | 2017-08-08 | International Business Machines Corporation | Fine-grained instruction enablement at sub-function granularity based on an indicated subrange of registers |
US8914615B2 (en) * | 2011-12-02 | 2014-12-16 | Arm Limited | Mapping same logical register specifier for different instruction sets with divergent association to architectural register file using common address format |
US9304749B2 (en) * | 2013-09-12 | 2016-04-05 | Marvell World Trade Ltd. | Method and system for instruction scheduling |
US10318299B2 (en) * | 2013-10-31 | 2019-06-11 | International Business Machines Corporation | Reading a register pair by writing a wide register |
US11455171B2 (en) * | 2019-05-29 | 2022-09-27 | Gray Research LLC | Multiported parity scoreboard circuit |
US11432152B2 (en) | 2020-05-04 | 2022-08-30 | Watchguard Technologies, Inc. | Method and apparatus for detecting and handling evil twin access points |
CN113281754B (en) * | 2021-07-26 | 2021-10-01 | 中国水利水电科学研究院 | WRF-Hydro key parameter calibration method for quantitatively estimating rainfall by integrating rainfall station with radar |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0640912A1 (en) | 1993-08-31 | 1995-03-01 | Sun Microsystems, Inc. | Memory addressing scheme for increasing the number of memory locations available in a computer for storing higher precision numbers |
US20010044891A1 (en) | 2000-01-14 | 2001-11-22 | Mcgrath Kevin J | Establishing an operating mode in a processor |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1005657B (en) * | 1987-07-20 | 1989-11-01 | 株式日立制作所 | Information processing apparatus |
US5781753A (en) * | 1989-02-24 | 1998-07-14 | Advanced Micro Devices, Inc. | Semi-autonomous RISC pipelines for overlapped execution of RISC-like instructions within the multiple superscalar execution units of a processor having distributed pipeline control for speculative and out-of-order execution of complex instructions |
US5897665A (en) * | 1995-12-15 | 1999-04-27 | Intel Corporation | Register addressing for register-register architectures used for microprocessors and microcontrollers |
US5790827A (en) * | 1997-06-20 | 1998-08-04 | Sun Microsystems, Inc. | Method for dependency checking using a scoreboard for a pair of register sets having different precisions |
US7340590B1 (en) * | 2003-12-11 | 2008-03-04 | Sun Microsystems, Inc. | Handling register dependencies between instructions specifying different width registers |
US8417922B2 (en) | 2006-08-02 | 2013-04-09 | Qualcomm Incorporated | Method and system to combine multiple register units within a microprocessor |
-
2009
- 2009-04-20 US US12/426,550 patent/US8898436B2/en active Active
-
2010
- 2010-04-13 WO PCT/US2010/001099 patent/WO2010123541A2/en active Application Filing
- 2010-04-13 EP EP10721863A patent/EP2422270A2/en not_active Withdrawn
- 2010-04-13 CN CN201080010412.3A patent/CN102356377B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0640912A1 (en) | 1993-08-31 | 1995-03-01 | Sun Microsystems, Inc. | Memory addressing scheme for increasing the number of memory locations available in a computer for storing higher precision numbers |
US20010044891A1 (en) | 2000-01-14 | 2001-11-22 | Mcgrath Kevin J | Establishing an operating mode in a processor |
Also Published As
Publication number | Publication date |
---|---|
US8898436B2 (en) | 2014-11-25 |
EP2422270A2 (en) | 2012-02-29 |
CN102356377B (en) | 2016-01-20 |
WO2010123541A3 (en) | 2010-12-16 |
CN102356377A (en) | 2012-02-15 |
US20100268919A1 (en) | 2010-10-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7437532B1 (en) | Memory mapped register file | |
EP0851343B1 (en) | System for processing floating point operations | |
US8225076B1 (en) | Scoreboard having size indicators for tracking sequential destination register usage in a multi-threaded processor | |
US20200004549A1 (en) | Multi-threaded instruction buffer design | |
US8898436B2 (en) | Method and structure for solving the evil-twin problem | |
US20120005459A1 (en) | Processor having increased performance and energy saving via move elimination | |
US6968444B1 (en) | Microprocessor employing a fixed position dispatch unit | |
US20100274961A1 (en) | Physically-indexed logical map table | |
US20100274993A1 (en) | Logical map table for detecting dependency conditions | |
US9454371B2 (en) | Micro-architecture for eliminating MOV operations | |
US11068271B2 (en) | Zero cycle move using free list counts | |
US7096345B1 (en) | Data processing system with bypass reorder buffer having non-bypassable locations and combined load/store arithmetic logic unit and processing method thereof | |
WO2005013118A2 (en) | Renaming for register with multiple bit fields | |
US9317285B2 (en) | Instruction set architecture mode dependent sub-size access of register with associated status indication | |
US10877755B2 (en) | Processor load using a bit vector to calculate effective address | |
CN115640047B (en) | Instruction operation method and device, electronic device and storage medium | |
US6266763B1 (en) | Physical rename register for efficiently storing floating point, integer, condition code, and multimedia values | |
US9170638B2 (en) | Method and apparatus for providing early bypass detection to reduce power consumption while reading register files of a processor | |
US7406587B1 (en) | Method and system for renaming registers in a microprocessor | |
US8909904B2 (en) | Combined byte-permute and bit shift unit | |
EP1324192B1 (en) | Computer system for execution of operations out of order | |
US20120144175A1 (en) | Method and apparatus for an enhanced speed unified scheduler utilizing optypes for compact logic | |
US11451241B2 (en) | Setting values of portions of registers based on bit values | |
US20050102494A1 (en) | Method and apparatus for register stack implementation using micro-operations | |
US20120179895A1 (en) | Method and apparatus for fast decoding and enhancing execution speed of an instruction |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 201080010412.3 Country of ref document: CN |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 10721863 Country of ref document: EP Kind code of ref document: A2 |
|
REEP | Request for entry into the european phase |
Ref document number: 2010721863 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2010721863 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |