Patent Grant
7343473
1. Field of the Invention
The field of the invention generally relates to superscalar RISC microprocessors, more specifically, the invention relates to a CISC to RISC microprocessor instruction alignment unit and decode unit for permitting complex instructions to run on RISC-based hardware.
2. Related Art
All complex instruction set computers (CISC computers) which use variable length instructions are faced with the problem of determining the length of each instruction that is encountered in the instruction stream. Instructions are packed into memory as successive bytes of data, so that given the address of an instruction, it is possible to determine the starting address of the next instruction if you know the first instruction's length.
For a conventional processor, this length determination does not have a significant performance impact compared to other stages in the processing of an instruction stream, such as the actual execution of each instruction. As a result, fairly simple circuits are typically used. Superscalar reduced instruction set computers (RISC computers), on the other hand, can process instructions at a much higher rate, requiring instructions to be extracted from memory much more rapidly to keep up with the parallel execution of multiple instructions. This limiting factor imposed by the rate at which instructions can be extracted from memory is referred to as the Flynn Bottleneck.
The task of determining the length of each instruction and extracting that instruction from the instruction stream is performed by a function unit called an Instruction Align Unit (IAU). This block must contain decoder logic to determine the instruction length, and a shifter to align the instruction data with the decoder logic.
For the Intel 80386 microprocessor, the first byte of an instruction can have numerous implications on the overall instruction length, and may require that additional bytes be checked before the final length is known. Furthermore, the additional bytes may specify other additional bytes. It is therefore extremely difficult to quickly determine the length of the X86 instruction because the process is inherently sequential.
Based on the information provided in the i486™ Programmer's Reference Guide, several conclusions can be drawn regarding alignment unit present in the i486™. The i486™'s IAU is designed to look only at the first few bytes of the instruction. In cases where these bytes do not fully specify the length, these initial bytes are extracted and the process is repeated on the remaining bytes. Each iteration of this process requires a full cycle, so it may take several cycles, at worst case, for an instruction to be fully aligned.
Situations that require additional cycles for the i486™ IAU include the presence of prefixed and escaped (2 byte) opcodes. Both of these are common in i486™ programs. In addition, complex instructions may also comprise displacement and immediate data. The i486™ requires additional time to extract this data.
An example format for a CISC processor instruction is shown in
The example depicts the potential bytes of a variable length i486™ CISC instruction. The instructions are stored in memory on byte boundaries. The minimum length of an instruction is 1 byte, and the maximum length of an instruction, including prefixes, is 15 bytes. The total length of the instruction is determined by the Prefixes Opcode, ModR/M and SIB bytes.
The present invention is a subsystem and method of a microprocessor having a superscalar reduced instruction set computer (RISC) processor designed to emulate a complex instruction set computer (CISC), such as an Intel 80×86 microprocessor, or other CISC processors.
The CISC to RISC translation operation of the present invention involves two basic steps. CISC instructions must first be extracted from the instruction stream, and then decoded to generate nano-instructions that can be processed by the RISC processor. These steps are performed by an Instruction Alignment Unit (IAU) and an Instruction Decode Unit (IDU), respectively.
The IAU functions to extract individual CISC instructions from the instruction stream by looking at the oldest 23 bytes of instruction data. The IAU extracts 8 continuous bytes starting with any byte in a bottom line of an Instruction FIFO. During each clock phase, the IAU determines the length of the current instruction and uses this information to control two shifters to shift out the current instruction, leaving the next sequential instruction in the stream. The IAU therefore outputs an aligned instruction during each clock phase, for a peak rate of two instructions per cycle. Exceptions to this best case performance are discussed below in sections 2.0 and 2.1.
After CISC instructions have been extracted from memory, the IDU functions to convert these aligned instructions to equivalent sequences of RISC instructions, called nano-instructions. The IDU looks at each aligned instruction as it is output by the IAU, and decodes it to determine various factors such as the number and type of nano-instruction(s) required, the size of the data operands, and whether or not a memory access is required to complete the aligned instruction. Simple instructions are directly translated by decoder hardware into nano-instructions, while more complex CISC instructions are emulated by subroutines in a special instruction set, called microcode routines, which are then decoded into nano-instructions. This information is collected for two instructions during a complete cycle, and then combined together to form an instruction bucket, containing the nano-instructions corresponding to both source instructions. This bucket is then transferred to an Instructions Execution Unit (IEU) for execution by a RISC processor. The execution of the nano-instruction buckets is outside the scope of the present invention.
The foregoing and other features and advantages of the invention will be apparent from the following more particular description of preferred embodiments of the invention, as illustrated in the accompanying drawings.
The invention will be better understood if reference is made to the accompanying drawings in which:
A more detailed description of some of the basic concepts discussed in this section is found in a number of references, including Mike Johnson, Superscalar Microprocessor Design (Prentice-Hall, Inc., Englewood Cliffs, N.J., 1991); John L. Hennessy et al., Computer Architecture—A Quantitative Approach” (Morgan Kaufmann Publishers, Inc., San Mateo, Calif., 1990); and the i486™ Microprocessor Programmer's Reference Manual and the i486™ Microprocessor Hardware Reference Manual (Order Nos. 240486 and 240552, respectively, Intel Corporation, Santa Clara, Calif., 1990). The disclosures of these publications are incorporated herein by reference.
1.0 The Instruction Fetch Unit
An instruction Fetch Unit (IFU) of the present invention is used to fetch instruction bytes from an instruction stream stored in an instruction memory, instruction cache, or the like, and provide the instruction bytes to a decoder section for execution. Instructions to be aligned by the Instruction Alignment Unit are therefore supplied by the IFU.
During normal program execution, the MBUF 204 is used to supply instruction bytes to the IAU. When conditional control flow (i.e., a conditional branch instruction) is encountered, instructions corresponding to the branch target address are stored in the TBUF 206 while execution continues from the MBUF 204. Once the branch decision is resolved, either the TBUF 206 is discarded if the branch is not taken, or the TBUF 206 is transferred to the MBUF if the branch is taken. In either case, execution continues from the MBUF.
The EBUF 202 operates in a slightly different way. When emulation mode is entered, whether due to an emulation instruction or an exception, both instruction fetching and execution are transferred to the EBUF 202. (Emulation mode and exception handling will both be discussed below in detail.) Execution continues out of the EBUF 202 as long as the processor is in emulation mode. When the emulation routine finishes, execution is continued from the instruction data remaining in the MBUF 204. This eliminates the need to refetch the main instruction data after executing an emulation routine.
2.0 Instruction Alignment Unit Overview
An Instruction Alignment Unit subsystem in combination with the present invention uses the RISC strategy of making the common case fast to deal with by using the superior per-cycle instruction throughput of a superscalar processor.
In the context of the present invention, the term “align” means to position an instruction's bytes so that they can be distinguished from adjacent bytes in the instruction stream for later decoding. The IAU distinguishes the end of the current instruction from the beginning of the next instruction by determining the number of bytes in the current instruction. The IAU then aligns the current instruction so that the least significant byte presented to the IDU is the first byte of the current instruction. Different ordering of the bytes as they are presented to the IDU is also possible.
The IAU subsystem of the present invention is capable of aligning most common instructions at a rate of two per cycle at all clock rates, and provides the capability of aligning most other instructions at this same rate at reduced clock speeds. Instructions including prefixes require an additional half cycle to align. Immediate data and displacement fields are extracted in parallel, and thus, require no extra time.
Additionally, the IAU worst-case alignment time is only 2.0 cycles for an instruction, which is less than the time required to align many common instructions in conventional CISC processors. The worst-case occurs when the instruction has one or more prefixes (half cycle total to align), the instruction is from the set that requires a full cycle to determine the length, and the instruction (not including the prefixes) is greater than eight bytes in length (which requires an extra half cycle, thus totaling 2 full cycles).
This performance is achieved through several architectural features. First, the IAU is designed to perform a complete alignment operation during each phase of the clock by using alternate phase latches and multiplexers in the alignment circuitry. Second, the decode logic divides CISC instructions into two categories based on the number of bits that must be considered to determine each instruction's length: instructions with length specified by a small number of bits are aligned in a single phase (halfcycle), whereas other instructions typically require an additional clock phase. Finally, the IAU extracts up to eight bytes from the instruction stream in a single shift, allowing long instructions (up to 15 bytes for i486™) to be aligned in a small number of shift operations, and most instructions to be aligned with a single shift.
The following tasks are carried out by the IAU in order to quickly and accurately decode a CISC instruction:
The opcode byte or bytes specify the operation performed by the instruction. The Mod R/M byte specifies the address form to be used if the instruction refers to an operand in memory. The Mod R/M byte can also refer to a second addressing byte, the SIB (scale, index, base) byte, which may be required to fully specify the addressing form.
2.1 Instruction Alignment Unit Block Diagrams
A block diagram of the IAU is shown in
The MAIN DATAPATH 302 comprises several shifters, latches and multiplexers. An EXTRACT SHIFTER 306 receives instruction data arranged in bytes from the IFU. Two buses (shown generally at 303) IFI0b13 Bus[127:0] and IFI1b_Bus[55:0] represent instruction data outputs of the IFU. The IFU updates this instruction information in response to requests from the IAU on an ADVance BUFfer REQuest (ADVBUFREQ) line 308. Generation of the ADVBUFREQ signal will be discussed below. Eight bytes of data, corresponding to the current instruction, are output from the EXTRACT SHIFTER and are sent to an ALIGN SHIFTER 310 on a bus 307. The ALIGN SHIFTER holds a total of 16 bytes of instruction data and can shift up to 8 bytes per phase. The ALIGN SHIFTER is used to separate prefixes from their instruction if they are detected by shifting them out. The ALIGN SHIFTER is also used to align the instruction to its lower order bytes and shift-out the entire instruction after it has been aligned.
The 8-bytes are also sent via a bus 309 to an IMMediate Data SHIFTER (IMM SHIFTER 312), which extracts immediate data from the current instruction, and to a DISPlacement SHIFTER (DISP SHIFTER 314), which extracts displacement data from the current instruction. Data to these two shifters is delayed by a ½ cycle delay element 316 to keep it synchronized with the aligned instruction.
The ALIGN SHIFTER 310 outputs the next aligned instruction on a bus 311 to two ALIGN_IR latches 318 or 320. These latches operate on opposite phases of the system clock, allowing two instructions to be latched per cycle. The ALIGN_IR latches 318 and 320 output aligned instruction bytes on two output buses 321. During the phase in which one of the latches is receiving a new value, the output of the other latch (which is the current aligned instruction) is selected by a multiplexer (MUX 322). The MUX 322 outputs the current aligned instruction on an aligned instruction bus 323. The output 323 is the primary output of the IAU. This output is used by the PREDECODER 304 to determine the length of the current instruction, and it is fed back into the ALIGN SHIFTER 310 as data from which the next instruction is extracted. The current aligned instruction is fed back to the ALIGN SHIFTER 310 via bus 323, a stack 334 and a further bus 336. The bus 336 also sends the current aligned instruction information to the ½ cycle data delay 316.
The IMM and DISP SHIFTERS 312 and 314, respectively, can therefore shift the immediate and displacement data, because they also require 16 total bytes to shift. The ½ cycle data delay 316 outputs instruction bytes to the shifters on a bus. The IMM SHIFTER 312 outputs immediate data corresponding to the current instruction on an IMMEDIATE DATA bus 340. The DISP SHIFTER 314 outputs displacement data corresponding to the current instruction on a DISPLACEMENT DATA bus 342.
The PREDECODER 304 comprises three decoder blocks: a Next Instruction Detector (NID) 324, an Immediate Data and Displacement Detector (IDDD) 326, and a Prefix Detector (PD) 328. The NID and PD control the ALIGN SHIFTER and the EXTRACT SHIFTER, while the IDDD controls the IMM SHIFTER 312 and the DISP SHIFTER 314.
The PD 328 is designed to detect the presence of prefix bytes in an instruction. It determines the number of prefixes present, and provides shift control signals to the ALIGN SHIFTER 310 and the COUNTER SHIFTER 332 via a line 331, a MUX 330 and a line 333, for extraction of the prefixes from the instruction stream in the next half cycle. In addition, the PD 328 decodes the prefixes themselves and provides this prefix information on an output line 329 to the IDU.
The basic architecture of the PD 328 consists of four identical prefix detection units (to detect up to four prefixes), and a second block of logic to decode the prefixes themselves. The CISC format defines the order in which prefixes can occur, but the present invention checks for the presence of all prefixes in each of the first four byte positions. Furthermore, the functions of detecting the presence of prefixes and decoding the prefixes are separated to take advantage of the reduced speed requirements for the decoder. A more detailed description of the architecture of the PD 328 will be addressed below.
The IDDD 326 is designed to extract immediate data and displacement data from each instruction. The IDDD always attempts to extract both fields, whether they are present or not. The IDDD 326 controls the IMM SHIFTER 312 and the DISP SHIFTER 314 on a pair of lines 344 and 346, respectively. The IDU requires a half cycle to process the aligned instruction, but has no use for the immediate and displacement data. The immediate and displacement data is therefore delayed by the ½ cycle data delay 316 to allow more time for the IDDD 326 to compute shift amounts, because the shift occurs during the following phase, unlike the NID 324 which decodes and shifts in the same phase.
The NID 324 is the heart of the PREDECODER. The NID 324 determines the length of each instruction once the prefixes have been removed. The NID 324 controls the ALIGN SHIFTER 310 and a COUNTER SHIFTER 332 via a control line 325, MUX 330 and line 333. The NID comprises two sub-blocks, a Subset Next Instruction Detector (SNID 702) and a Remaining Next Instruction Detector (RNID 704), which will be discussed in conjunction with
The SNID 702, as its name implies, determines the lengths of a subset of the CISC instruction set. Instructions in the subset can be aligned at a rate of two per cycle by the SNID.
The RNID 704 determines the lengths of all remaining instructions, and requires an additional halfcycle, which brings its total decode time to a full cycle. The determination of whether or not an instruction is in the subset is made by the SNID, and this signal is used within the NID to select the outputs of either the SNID or the RNID.
When a new instruction is being aligned, it is initially assumed to be in the subset, and thus the output of the SNID is selected. If the SNID determines (during this same half-cycle) that the instruction must be handled by the RNID, a signal is asserted and the IAU loops the current instruction to hold it for another half-cycle. During this second half-cycle, the RNID output is selected, and the instruction is properly aligned.
This architecture of the NID has several benefits. One, which was mentioned earlier, is that the selection between the SNID and the RNID can be made during a single half cycle if the cycle time is sufficiently long, allowing all instructions to be aligned in a single phase (not including the time to extract prefixes and instructions longer than eight bytes). This provides a per-cycle performance increase at lower cycle rates, without additional hardware.
A second advantage is that the selection signal can be used as an alignment cancel signal, because it causes the IAU to ignore the SNID shift outputs and hold the current instruction for an additional half cycle. The SNID could be designed to predict certain instruction combinations or lengths, and then generate the cancel signal if these predictions were incorrect. This could be used to align multiple instructions in a single half cycle, for example, which would further boost performance.
The IAU also comprises a COUNTER SHIFTER 332. The COUNTER SHIFTER 332 is used to determine the shift amount for the EXTRACT SHIFTER 306 via a line 335, and request additional CISC instruction bytes from the IFU using the ADVBUFREQ line 308. The functionality of the COUNTER SHIFTER 332 will best be understood by reviewing the following flow chart of the IAU operation and a timing diagram example.
During the next clock phase, the IAU outputs the aligned instruction to the IDU, the STACK 334, the IDDD 326, the NID 324, the PD 328 and the ½ cycle data delay 316. The immediate data and displacement information is then output to the IDU on buses 340 and 342, respectively. This data corresponds to the instruction aligned during the previous phase, if there was one. These operations are shown generally at a step 408 of
A conditional statement 409 is then entered by the IAU to determine if a prefix or prefixes are present. This determination is made by the PD (prefix decoder) 328. If one or more prefixes are detected by the PD, as indicated by a “YES” arrow exiting the conditional statement 409, the process proceeds to a step 410 in which the IAU selects the output of the PD with the MUX 330. The decoded prefix information is then latched to be sent to the IDU during the next phase with the corresponding aligned instruction, as shown at a step 412. If no prefix instruction bytes were detected, as indicated by a “NO” arrow exiting the conditional statement 409, the output of the NID 324 is selected with the MUX 330, as shown at a step 414.
Once the steps 412 or 414 are completed, the current output of the COUNTER SHIFTER 332 is used to control the EXTRACT SHIFTER 306 to provide the next 8 bytes of instruction data to the ALIGN SHIFTER 310 and the ½ cycle delay 316, as shown at a block 416. Next, the IAU uses the output of the MUX 330 as a variable called SHIFT_A, which is used to control the ALIGN SHIFTER 310 to align the next instruction. The SHIFT_A is also added to the current EXTRACT SHIFTER shift amount (called BUF_COUNT) to compute the shift amount for use during the next phase. This addition is performed in the COUNTER SHIFTER 308, as shown at a step 418.
The next operational step performed by the IAU is to latch the output of the ALIGN SHIFTER in the ALIGN_IR latch, as shown at a step 420. The position of the immediate data and displacement data in the IDDD 326 is then computed, and this shift amount is delayed by a ½ cycle, as shown at a step 422. Next, the IAU uses the shift amount computed during the previous half cycle to shift the data currently entering the IMM SHIFTER 312 and DISP SHIFTER 314, as shown at a step 424. Finally, the process repeats beginning at step 406 to wait for the next clock phase. The steps 408 through 424 are repeated for the remaining instruction bytes in the instruction stream.
In this example, the first three instructions sent to the IAU are OP0, OP1, and OP2, which have lengths of 5 bytes, 3 bytes and 11 bytes, respectively. Note that only the first 8 bytes of instruction OP2 fit in BUCKET_#0. The remaining 3 bytes wrap to the beginning of BUCKET_#1. To simplify this example, it is assumed that these three instructions have no prefix bytes. An additional phase would be required for the alignment of an instruction if prefixes are detected.
Instructions can start at any position of a bucket. Instructions are extracted up to 8 bytes at a time from the bottom bucket beginning with any instruction in that bucket. The IAU looks at two buckets to accommodate instructions which extend into the second bucket, such as OP2 in the present example.
Trace “1” in the timing diagram is one of two system clocks CLK0. In this example, the system clock has a 6 nano second (ns) half cycle. CLK0, which has opposite phase compared to the other system clock CLK1, rises at T6 and falls at T0, where T0 is the rising edge of CLK1 and T6 is the rising edge of CLK0. The three main clock phases of
Traces “2” and “3” in the timing diagram represent instruction data on the input buses IFI1B and IFI0B. A new BUCKET_#0 becomes available on bus IFI0B at the beginning of φ1, as shown at 502. A short time later, the first 8 bytes starting with OP0 (B#0; 7-0) are extracted by the EXTRACT SHIFTER 306 at 504. BUCKET_#0 bytes 7-0 are shown valid. The EXTRACT SHIFTER timing is shown at a trace “4”.
When CISC to RISC decoding of an instruction stream begins, the COUNTER SHIFTER 332 controls the EXTRACT SHIFTER 306 to extract the first 8 bytes from Bucket_#0. The COUNTER SHIFTER signals the EXTRACT SHIFTER to shift and extract further bytes of the buckets as the alignment of instructions progresses. When Bucket_#0 is depleted of instruction bytes, the contents of Bucket_#1 are shifted into Bucket_#0, and Bucket_#1 is refilled from the instruction stream. After the initial extraction of 8 bytes, the EXTRACT SHIFTER extracts and shifts bytes under control of the COUNTER SHIFTER on line 335, based on instruction length, prefix length and previous shift information.
For this example, however, the COUNTER SHIFTER signals the EXTRACT SHIFTER to shift zero to align the first instruction. Thus, the EXTRACT SHIFTER shifts-out the first 8 bytes of the first instruction to the ALIGN SHIFTER 310. The timing of signals at the ALIGN SHIFTER are shown at trace “5” of the timing diagram. These 8 bytes become valid at the ALIGN SHIFTER during φ1 at the time period shown by a reference numeral 506.
The first 8 bytes of Bucket_#0 bypass the ALIGN SHIFTER and are stored in the two ALIGN_IR latches 318 or 320 (as shown at traces “6” and “7” in
The MUX 322 selects the latch that was latching during the previous phase. Thus, in this example, MUX 322 outputs the first eight bytes of OP0 during the second full phase, φ2.
The first 8 bytes of OP0 then flow to the NID 324 and the STACK 334. The NID 324 detects that the first instruction is 5 bytes long and sends this information back to the ALIGN SHIFTER and to the COUNTER SHIFTER via line 325, MUX 330 and line 333. At the same time the first 8 bytes flow through the stack and are fed back to the ALIGN SHIFTER, as discussed above. Thus, the ALIGN SHIFTER receives instruction bytes from the EXTRACT SHIFTER, and itself, indirectly. This is because the ALIGN SHIFTER needs 16 bytes of input in order to shift a maximum of 8 bytes per cycle. When the ALIGN SHIFTER shifts right X number of bytes, it discards the least significant X number of bytes, and passes the next 8 bytes of data to the latches 318 and 320. In this case, the STACK 334 provides bytes 0-7 to the ALIGN SHIFTER 310.
A bypass 336 around the ALIGN SHIFTER is used in the initial case when the EXTRACT SHIFTER extracts the first instruction from the instruction stream. It is not necessary for the ALIGN SHIFTER to shift in the initial case, because, excluding prefix bytes, the first instruction is aligned.
During φ2 of the timing diagram, the EXTRACT SHIFTER shifts out 8 bytes, bytes 15-8 of BUCKET_#0. See 510 at
Toward the end of φ2, the ALIGN SHIFTER shifts bytes 12-5 of BUCKET_#0 to its outputs, based on the signal from the NID, which indicated to the ALIGN SHIFTER to shift 5 bytes to the right, thereby discarding the 5 least significant bytes corresponding to instruction OP0. See the Shift_5_byte signal 512 at trace “8” in the timing diagram. The 8 bytes of remaining instruction data, bytes 12-5, then flow through the ALIGN SHIFTER. Note that byte 5 is the first byte of the next instruction, OP1.
The COUNTER SHIFTER 332 then shifts the EXTRACT SHIFTER 306 8 bytes, because the first 8 bytes are now available from the ALIGN_IR latches, thus the next bytes are needed. Beginning at phase 3, the COUNTER SHIFTER will signal the EXTRACT SHIFTER to increase its shift amount by the number of bytes shifted out by the ALIGN SHIFTER 310 during the previous phase. The COUNTER SHIFTER must therefore comprise logic to store the previous EXTRACT SHIFTER shift amount, and add the ALIGN SHIFTER shift amount to this value.
Each time there is a new value for the ALIGN SHIFTER, the COUNTER SHIFTER adds that amount to its old shift amount. In this example, it shifted 8 bytes during φ2. Therefore, in φ3, it must tell the EXTRACT SHIFTER to shift 8+5, or 13 bytes. The bytes output by the EXTRACT SHIFTER are bytes 20-13. Note that the ALIGN_IR latches will output bytes 12-5 during φ3; and therefore, bytes 20-5 will be available at the ALIGN SHIFTER.
During φ3, the EXTRACT SHIFTER will output bytes 20-13. However, BUCKET_#0 only contains bytes 15-0, therefore, bytes 20-16 must be taken from BUCKET_#1. As shown at 514 in the timing diagram, BUCKET_#1 becomes valid at the beginning of φ3. The EXTRACT SHIFTER then shifts bytes 4-0 of BUCKET_#1 and bytes 15-13 of BUCKET_#0, as shown at 516. If BUCKET_#1 was not valid at this time, the IAU would have to wait until it becomes valid.
As noted above, the Shift—5_byte signal was generated by the NID during φ2. Based on this signal, bytes 12-5 of BUCKET_#0 are shifted out by the ALIGN SHIFTER, as shown at 518, and shortly thereafter are latched into ALIGN_IR1, as shown at 520.
Bytes 12-5 are sent to the STACK 334 and the NID 324 by the MUX 322 at the beginning of φ3. The STACK feeds bytes 12-5 back to the ALIGN SHIFTER as shown at 336, and the NID determines the length of OP1 to be 3 bytes and outputs the Shift—3_bytes signal during the latter half of φ3, as shown in trace “9” at 522. The ALIGN SHIFTER shifts 3 bytes (15-8), and this amount is added to the COUNTER SHIFTER.
The above process then repeats. Once an instruction advances beyond BUCKET #0 (i.e., BUCKET_#0 is completely used), BUCKET_#1 will become BUCKET_#0 and a new BUCKET_#1 will later become valid.
Trace “10” in the timing diagram shows the timing for extraction of bytes from the instruction stream. The Buf_Count#0 blocks represent the stored extract shift amount. During each phase the aligned shift amount is added to Buf_Count#0, and the result becomes the extract shift amount during the next phase (see the blocks labeled COUNTER_SHIFT).
Trace “11” in the timing diagram shows instruction alignment timing. The blocks labeled IR_Latch_#0 and IR_Latch_#1 represent the time during which the instructions in the corresponding ALIGN_IR latch become valid. The small blocks labeled MUX1 represent the time when the MUX 322 begins to select the valid align latch. The small blocks labeled MUX2 represent the time when the MUX 330 begins to select the shift amount determined by the NID 324. Finally, the blocks labeled ALIGN_SHIFT represent the time when the ALIGN SHIFTER begins to output the instruction.
Prefixes are extracted using the same technique by which instructions are aligned, but the output of PD 328 is selected by MUX 330 rather than the output of NID 324.
A block diagram of a section of the STACK 334 is shown in
The IAU may periodically transfer to a different instruction stream. The STACK allows the IAU to store two sets of 8 bytes of instruction data from the MUX 322. This feature is generally used during CISC instruction emulation. When the IAU must branch to process a microcode routine for emulation of a complex CISC instruction, the state of the IAU can be stored and re-initiated once the emulation of the CISC instruction is completed.
The ½ cycle data delay 316 is used to delay the immediate data and displacement information. Placing the delay in the IAU before the shifters pipelines the immediate data and displacement logic in order to do the shift during the following phase, rather than determining the instruction length and the shift in the same half cycle. The operations can be spread across the cycle, thus making the timing requirement easier to meet for that logic. The IDDD block 326 controls the IMM Shifter 312 and the DISP Shifter 314 to extract the immediate data and displacement data from the instructions. For example, if the first 3 bytes of the instruction are opcode, followed by 4 bytes of displacement and 4 bytes of immediate data, the shifters would be enabled to shift out the appropriate bytes.
The shifters 312 and 314 always output 32 bits whether the actual data size is 8, 16 or 32 bits, with the immediate and displacement data appropriately aligned to the low order bits of the 32 bit output. The IDU determines whether the immediate and displacement data is valid, and if so, how much of the data is valid.
The determination of the length of any prefixes, immediate data, displacement data, and the actual length of the instructions is a function of the actual CISC instruction set being aligned and decoded. This information may be obtained by one skilled in the art by studying the CISC instruction set itself, the manufacture's user manuals, or other common reference material. Those skilled in the art will readily recognize how to accomplish this, as well as how to convert the information into random logic to implement the above described IAU subsystem, the IDU subsystem described below, and how to generate the control logic and signals used to control data flow.
Furthermore, once such random logic is generated, commercially available engineering software applications (e.g., Verilog manufactured by Cadence Design Systems, Inc., San Jose, Calif.), may be used to verify the logic, and can aid in defining the timing and generation of the control signals and associated random logic. Other commercially available engineering software applications are available to generate gate and cell layouts to optimize the implementation of the functional blocks and control logic.
The i486™ instruction set supports 11 prefixes that have a defined order when used together in an instruction. The format defines that up to four prefixes can be included in a single instruction. Thus, the PREFIX DETECTOR 328 of the present invention comprises four identical prefix detect circuits. Each circuit looks for any of the 11 prefix codes. The first four bytes passed to the prefix detector are evaluated, and the outputs of the four prefix detect circuits are combined to determine the total number of prefixes present. The result is used as the shift amount that is passed through the MUX 330.
A block diagram of the NID is shown in
Only 4 bytes are required to determine the length of an instruction. (As noted above, the 4 bytes comprise two Opcode bytes, an optional ModR/M byte and a SIB byte.)
The subset of instructions decoded by the SNID are those CISC instructions that can be decoded in a half cycle using a minimum of 1, 2 and 3 input gates (NANDs, NORs and invertors), with a maximum of 5 gate delays based on an 16×16 Karnaugh map of the 256 instructions. Blocks of the map including most 1 byte opcode instructions can be implemented in this fashion. The remainder of the instructions are decoded by the RNID using a logic array with a longer gate delay.
The RNID 704 receives the first 4 bytes on the bus 701. The RNID performs length determination decoding for the remaining instructions that requires more that one phase to decode. The RNID has outputs that are similar to the outputs of the SNID.
The RNID detects instruction lengths and outputs the result on a bus 711. A 1 bit OVER8 output 712 indicates that the instruction is over 8 bytes in length. The RNID also has a 1 bit MOD_DET output 714 that indicates whether the instruction includes a ModR/M byte.
The length decoded by either the SNID or the RNID is selected by a MUX 706. A control line 708 for the MUX 706, called SELect_DECoder for current InstRuction (SELDECIR), switches the MUX 706 between the two decoders to get the actual length which is 1 to 11 bytes. An 11 byte-long instruction, for example, would cause the RNID to output the OVER8 signal and a 3 on bus 711. The instruction length (ln) is sent to the MUX 330 on a bus 716, and is used by the ALIGN SHIFTER 310 and the COUNTER SHIFTER 332. The 8 bits output by the top MUX 706 are used as shift controls (enables) for the ALIGN and COUNTER SHIFTERs.
The ModR/M bytes are also selected in a similar fashion. The SELDECIR signal 708 controls a second MUX 710 to choose the appropriate MOD line to indicate whether a ModR/M byte is present. The MOD line output 718 is used by the IDDD.
The SELDECIR signal 708 is generated based on the NID_WAIT signal 709. The output of the SNID is selected during the first clock phase because those results will be complete. If the NID_WAIT signal 709 indicates that the instruction was not decoded, the MUXs 706 and 710 are switched to select the output 711 of the RNID, which will become available at the beginning of the next clock phase.
The RNID 704 essentially comprises two parallel decoders, one decodes the instructions as if there is a 1 byte opcode and the other decodes as if there is a 2 byte opcode. An ESCape DETect (ESC_DET) input signal indicates whether the opcode is 1 byte or 2 bytes in length. For example, in the i486™ instruction set, the first byte in all 2 byte opcodes (called the ESCAPE byte) has the value 0F hex that indicates the instruction has a 2 byte opcode. The RNID outputs a valid instruction length based on an ESC_DET signal. This signal indicates that the first opcode byte is an ESCAPE (0F hex), which indicates a 2 byte opcode, thereby enabling the second byte decoder. Decoding logic for generating the ESC-DET signal should be evident to those skilled in the art.
A block diagram of the RNID is shown in
The first opcode byte of the instruction and 3 bits (bits [5:3] called extension bits) of the ModR/M byte are input to the RNID—1OP 752 on a bus 766. A further input line 768 called DATA_SZ to the RNID—1OP indicates whether the operand size of the instruction is 16 or 32 bits. The data size is determined based on the memory protection scheme used, and whether prefixes are present to override the default data size. RNID—1OP assumes that the instruction has a 1 byte opcode, and based on that information and the 3 extension bits, RNID—1OP attempts to determine the length of the instruction.
The RNID_MOD decoder 754 decodes the ModR/M byte of the instruction input on a bus 770. The RNID_MOD decoder has an additional input bus 772 labeled ADD_SZ, which indicates whether the address size is 16 or 32 bits. The address size is independent of the data size.
The ESC_DET signal 774 is also input to block 760. When the ESC_DET signal is logic HIGH, for example, the RNID_SUM block knows that the opcode is actually in the second byte.
The RNID—2OP decoder 754 assumes that the opcode is 2 bytes, and therefore decodes the second byte (see bus 776) of the opcode. RNID—2OP decoder also has the input 768 identifying the data size.
Since the decoders themselves do not know the length of the opcode, i.e., 1 or 2 bytes, and since the ModR/M byte always follows the opcode, the second RNID_MOD decoder 758 is used to decode the byte (see bus 778) following the 2 byte opcode, again assuming that it is there. The two RNID_MOD decoders are identical, but decode different bytes in the instruction stream.
Again, based on the ESC_DET signal 774, the RNID_SUM 760 selects the outputs of the appropriate opcode and ModR/M byte decoders, and outputs the length of the instruction on bus 762. The output 764 labeled OVER8 indicates whether the instruction is over 8 bytes. If the instruction is over 8-bytes in length, the IR_NO[7:0] bus 762 indicates the number of instruction bytes over 8.
The RNID—1OP decoder 752 has an output bus 780 that is 9 bits wide. One line indicates whether the instruction is 1 byte long. The second line indicates that the instruction is 1 byte long and that a ModR/M byte is present, and thus, information from the ModR/M decoder should be included in the determination of the length of the instruction. Similarly, the remaining output lines of bus 780 indicate the following number of bytes: 2, 2/MOD, 3, 3/MOD, 4, 5, and 5/MOD. If the instruction is 4-bytes long there cannot be a ModR/M byte; this is inherent in the i486™ instruction set. However, the present invention is in no way limited to any specific CISC instruction set. Those skilled in the art will be able to apply the features of the present invention to align and decode any CISC instruction set.
The RNID—2OP decoder 754 has an output bus 782 that is 6 bits wide. One line indicates whether the instruction is 1 byte long. The second line indicates that the instruction is 1 byte long and includes a ModR/M byte, which should be included in the determination of the length of the instruction. Similarly, the remaining output lines of bus 782 indicate that there are 2, 2/MOD, 3, and 5/MOD. There are no other possible instruction lengths supported by the i486™ instruction set if the opcode is 2 bytes long.
Outputs 784 and 786 of the two RNID_MOD decoders 756 and 758 indicate to the RNID_SUM 760 the five possible additional lengths that can be specified by the ModR/M byte. Each RNID_MOD decoder has a 5 bit wide output bus. The five possible additional lengths are: 1, 2, 3, 5 and 6-bytes. The ModR/M byte itself is included in the total length determination. Any remaining bytes comprise immediate or displacement data.
The i486™ instruction set includes two special instructions, the enter_detect and jump_call_detect instructions. The IDDD therefore has a block called the Immediate Special Detector (ISD) 802 to handle decoding of these instructions. An input 803 to the ISD is the first byte of the instruction. Two output lines EN_DET and JMP_CL_DET (820 and 822, respectively) indicate whether one of the corresponding instructions is detected.
MOD_DEC decoders 804 and 806 are identical and decode the immediate and displacement data. Based on ADD_SZ 772, decoder 804 looks at the ModR/M byte assuming a 1 byte opcode and decoder 806 looks at the ModR/M byte assuming a 2 byte. The instruction byte inputs to MOD_DEC 804 and 806 are 805 and 807, respectively. These decoders determine the displacement position and the immediate data position in the instruction stream. Two seven line outputs 824 and 826 indicate at what position the displacement and immediate data starts: the displacement can start at position two or position three; and immediate data can start at position two, three, four, six or seven.
The MOD_DET lines 707 and 714 are also input to the SELECT block 812.
The SELECT block 812 combines the EN_DET and JMP_CL_DET signals, the MOD_DET and MOD_DEC results, and the ADD_SZ and outputs its results on four buses 832-838. A DISPlacement 1 (DISP_1) bus 832 outputs the displacement shift results assuming a 1 byte opcode. A DISPlacement 2 (DISP_2) bus 834 outputs the displacement shift results assuming a 2 byte opcode. IMMediate 1 and 2 (IMM_1 and IMM_2) buses 836 and 838 output the immediate data shift information assuming a 1 byte and a 2 byte opcode, respectively.
A last block 814 labeled MOD_SEL/DLY actually selects the appropriate shift amounts and delays these results a halfcycle. The halfcycle delay performed by MOD_SEL/DLY 816 represents the delay 316 shown in
A block diagram of the PREFIX DETECTOR 328 is shown in
The i486™ instruction set, for example, includes 11 possible prefixes. Four total prefixes can be included per instruction, because there are several invalid prefix combinations. The ordering of the four prefixes is also defined by the instruction set. However, rather than detect only the legitimate prefix permutations, the PREFIX DETECTOR uses the four prefix detectors 904-910 to decode each of the first 4 bytes of the instruction. The first 4 bytes of the instruction are input to the PREFIX DETECTOR on a bus 901. Each detector 904-910 has an output bus (905, 907, 909 and 911, respectively) that is 12 bits wide. The 12 outputs indicate which prefix(es) are present, if any are actually decoded at all. The twelfth prefix is called UNLOCK, which is the functional complement of the i486™ LOCK prefix, and is only available to microcode routines during emulation mode.
An ALIGN_RUN control signal 920 may be included to enable/disable the prefix decoder, and can be used to mask-out all of the prefixes. A HOLD_PRFX control signal 922 is used to latch and hold the prefix information. Generally, for alignment of an instruction if the PREFIX DETECTOR 328 indicates that there are prefixes present, the control logic must latch the prefix information. The prefix information is then used by the ALIGN SHIFTER 310 to shift-out the prefixes. In the following cycle, the IAU determines the length of the instruction, aligns it, and passes it to the IDU.
The PRFX_NO decoder 902 indicates where and how many prefixes are present by decoding the first 4 bytes of the opcode. A logic diagram of the PRFX_NO decoder 902 is shown in
The total number of prefixes is one if the first byte (position) is a prefix and there is no prefix in the second position. As a further example, a prefix in the fourth position does not matter, unless there are prefixes in the first three positions. A logic HIGH (1) output from the bottom NAND 1014 gate indicates that there are four prefixes; a HIGH output from the second last NAND gate 1015 indicates that there are three prefixes, and so on. The four NAND gate outputs are combined to form a PREFIX_NO bus 1018 to indicate the total number of valid prefixes that precede the first opcode byte, i.e, the shift amount output of the PREFIX DETECTOR 328.
The PRFX_NO decoder 902 also includes a Prefix_Present (PRFX_P) output bus 1020 (which is also 4 bits wide). Four PRFX_P output lines 1020-1023 indicate whether or not there is a prefix in the given position, regardless of what the other positions output. The PRFX_P outputs are tapped directly off the four decoder (1002-1008) outputs.
The PRFX_NO decoder results (to be discussed in connection with
3.0 Instruction Decode Unit Overview
All instructions are passed from the IAU to an Instruction Decode Unit (IDU), and are directly translated into RISC instructions. All instructions to be executed by the IEU are first processed by the IDU. The IDU determines whether each instruction is an emulated or a basic instruction. If it is emulated, the microcode emulation routine consisting entirely of basic instructions is processed. If the instruction is basic, it is directly translated by hardware into one to four nano-instructions and sent to the IEU. It is these nano-instructions, rather than the original CISC or microcode instructions, that the IEU actually executes.
The partitioning of instructions has two key benefits: the hardware is kept small because it only needs to support simple operations, and bugs are less troublesome because they are more likely to occur in the complex microcode routines, which can easily be changed.
The IDU's microcode routine support hardware in conjunction with the present invention has several features which make it unique. Typically, microcode instructions consist of control bits for the various datapaths present in a processor, with little or no encoding. The microcode of the present invention, in contrast, is a comparatively high-level machine language designed to emulate a specific complex instruction set. Whereas typical microcode is routed directly to a processor's function units, the microcode of the present invention is processed by the same decoder logic that is used for the target CISC (e.g., 80×86) instructions. This gives the microcode of the present invention much better code-density than is achieved by typical microcode, and makes the microcode easier to develop due to its similarity with the target CISC instruction set. Furthermore, the present invention provides hardware support for microcode revisions: part or all of the on-chip ROM-based microcode can be replaced with external RAM-based microcode under software control. (See commonly owned, co-pending application titled, “A ROM With RAM Cell and Cyclic Redundancy Check Circuit”, Ser. No. 07/802,816, filed Dec. 6, 1991; the disclosure of which is incorporated herein by reference.)
The microcode routine language is designed to be a set of instructions that can be executed by the RISC core to perform the functions required by all of the complex emulated instructions, plus the various control and maintenance functions associated with exception handling. Although emulated instructions are typically less performance sensitive than non-emulated (basic) instructions, and exceptions, (which are handled by microcode routines) occur infrequently, it is still critical to the overall system throughput that both be handled efficiently. This goal is achieved through the use of various forms of hardware support for the microcode routines. The present invention comprises four areas of hardware support for microcode: dispatch logic, mailboxes, a nano-instruction format, and special instructions.
The microcode dispatch logic controls the efficient transfer of program control from the target CISC instruction stream to a microcode routine and back to the target instruction stream. It is handled with a small amount of hardware, and in a manner that is transparent to the RISC core's Instruction Execution Unit (IEU). (The IEU executes the RISC instructions. The “RISC core” mentioned above is synonymous with the IEU. The details of the IEU are not necessary for one skilled in the art to practice the present invention. The features of the present invention are applicable to RISC processors in general.)
The mailboxes comprise a system of registers used to transfer information from the instruction decode hardware to microcode routines in a systematic way. This allows the hardware to pass instruction operands and similar data to the microcode routines, saving them the task of extracting this data from the instruction.
The nano-instruction format describes the information that passes from the IDU to the IEU. This format was chosen to allow it to be efficiently extracted from the source CISC instructions, but still provide adequate information to the IEU for dependency checking and function unit control.
Finally, the special instructions are a set of additional instructions provided to allow complete control of the RISC hardware and support certain unique emulation tasks in hardware, and are CISC instruction set specific.
3.1 Microcode Dispatch Logic
The first step in dispatching to microcode is to determine the address of the microcode routine. This step has two important requirements: each microcode routine must have a unique starting address, and these addresses must be generated quickly. This is fairly easy to achieve for exception handling routines, since the small number of cases that must be handled allows the hardware to store the addresses as constants and merely select between them. Determining the addresses for emulated instructions is more difficult, however, because there are too many to make storing all the addresses feasible.
The microcode dispatch logic meets the requirements by basing each instruction's dispatch address directly on its opcode. For example, one-byte opcodes are mapped into the address space from 0H to 1FFFH, requiring that the upper three bits of the 16 bit dispatch address be zeroes. These microcode entry points are spaced 64 bytes apart, which requires the six least-significant bits of each entry point address to be zero. This leaves 7 bits undetermined, and they can be taken directly from seven of the opcode bits. Generating the address in this way requires very little logic, as will become evident to those skilled in the art. For example, a multiplexer alone can be used to select the proper bits from the opcode.
Once the dispatch address for a microcode routine has been determined, the microcode must be fetched from memory. Typically, microcode resides in on-chip ROM, but this is not necessarily the case. As detailed in the above referenced application Ser. No. 07/802,816, each entry point is associated with a ROM-invalid bit which indicates whether or not the ROM routine is correct. This bit is fetched in parallel with the ROM access, and functions similarly to a conventional cache-hit indicator. If this bit indicates that the ROM entry is valid, the microcode routine will continue to be fetched from ROM and executed normally. If the bit indicates that the ROM is invalid, however, the microcode is fetched from external memory, such as RAM or the like.
On chip microcode routine addressing is handled by the IDU itself. The IDU generates 16 bit addresses for accesses to the microcode ROM. If the ROM-invalid bit corresponding to the ROM entry being addressed indicates that the microcode is invalid, the address of external microcode residing off-chip in main memory is calculated. A U_Base register holds the upper 16 address bits (called the starting address) of the external microcode residing in main memory. The 16 bit address decoded by the IDU is concatenated with the upper 16 bits in the U_Base register to access the external microcode residing in main memory. If the location of the external microcode residing in main memory is changed, the contents of the U_Base register can be modified to reflect the new main memory location.
This feature allows microcode updates to be performed by replacing certain routines with alternates in external memory, without forcing all microcode to suffer the reduced performance of external memory accesses. It also makes it possible to remove all ROM from the RISC chip and place the entire microcode in external RAM, to reduce the RISC chip's area requirements or to aid in microcode development.
The dispatch logic is also responsible for providing a means for the microcode routine to return to the main instruction stream when its task is finished. To handle this, separate Program Counters (PC's) and instruction buffers are maintained. During normal operation, the main PC determines the address of each CISC instruction in external memory. A section of memory containing these instructions is fetched by the IFU and stored in the MBUF.
When an emulated instruction or exception is detected, the PC value and length of the current instruction are stored in temporary buffers, while the microcode dispatch address is calculated as described above and instructions are fetched from this address into the EBUF. Microcode is executed from the EBUF until a microcode “return” instruction is detected, at which time the preserved PC value is reloaded, and execution continues from the MBUF. Since the MBUF and all other related registers are preserved during the transfer of control to the microcode routine, the transfer back to the CISC program happens very quickly.
There are two return instructions used by microcode routines to support the differences between instruction emulation routines and exception handling routines. When the microcode routine is entered for the purpose of handling an exception, it is important that after the routine is finished, the processor should return to the exact state in which it was interrupted. When the microcode routine is entered for the purpose of emulating an instruction, however, the routine wants to return to the instruction following the emulated instruction. Otherwise, the emulation routine will be executed a second time. These two functions are handled by the use of two return instructions: aret and eret. The aret instruction returns the processor to its state when microcode was entered, while the eret instruction causes the main PC to be updated and control to return to the next instruction in the target stream.
3.2 Mailboxes
For emulation routines to successfully perform the functions of a complex CISC instruction, it is necessary that the microcode have convenient access to the operands referenced by the emulated instruction. In the present invention, this is performed through the use of four mailbox registers. These registers are unique in their use only; they are defined to be the first four of a set of sixteen temporary registers in the integer register file that are available to microcode. Each emulation routine that requires operands or other information from the original instruction can expect to find these values stored in one or more of the mailbox registers upon entry into the routine. When the IDU detects an emulated instruction, it generates instructions which are used by the IEU to load the registers with the values that microcode expects, before execution of the microcode routine itself begins.
For example, consider the emulation of the Load Machine Status Word instruction (lmsw), which specifies any one of the general registers as an operand. Assume the specific instruction to be emulated is lmsw ax, which loads a 16 bit status word from the “ax” register. The same microcode routine is used regardless of the register actually specified in the instruction, so for this instruction mailbox#0 is loaded with the status word before microcode entry. When the IDU detects this instruction, it will generate a mov u0,ax instruction for the IEU to move the status word from the “ax” register to the “u0” register, which is defined to be mailbox #0. After this mov instruction is sent to the IEU, the microcode routine will be fetched and sent. Thus, the microcode can be written as if the emulated instruction were lmsw u0, and it will correctly handle all of the possible operands that may be specified in the original CISC instruction.
3.3 Nano-Instruction Format
As mentioned above, CISC instructions are decoded by the IDU into nano-instructions, which are processed by the RISC processor core, referred to as the IEU. Nano-instructions are passed from the IDU to the IEU in groups of four, called “buckets”. A single bucket is shown
The IEU can accept buckets from the IDU at a peak rate of one per cycle. The IDU processes basic instructions at a peak rate of two per cycle. Since most basic instructions are translated into a single packet, two basic instructions can usually be placed in one bucket and passed to the IEU together. The primary restriction on this rate is that the basic instructions must match the requirements of a bucket:
If one or both of these restrictions is violated, the bucket is sent to the IEU with nano-instructions corresponding to only one of the basic instructions, and the remaining instruction is sent in a later bucket. These requirements closely mirror the capabilities of the IEU, i.e., an IEU having two ALUs and one Load/Store unit, so in reality they do not present a limitation on performance. An example of this type of IEU is disclosed in commonly owned, co-pending applications titled, “High-Performance, Superscalar-Based Computer System with Out-of-Order Instruction Execution,” application Ser. No. 07/817,810, filed Jan. 8, 1992, now U.S. Pat. No. 5,539,911, which is a continuation of application Ser. No. 07/727,006, filed Jul. 8, 1991, now abandoned, and “High Performance, Superscalar-Based Computer System with Out-of-Order Instruction Execution and Concurrent Results Distribution,” application Ser. No. 08/397,016 filed Mar. 1, 1995, now U.S. Pat. No. 5,560,032, which is a file wrapper continuation of application Ser. No. 07/817,809, filed Jan. 8, 1992, now abandoned, which is a continuation of application Ser. No. 07/727,058 filed Jul. 8, 1991, now abandoned, the disclosures of which are incorporated herein by reference.
3.4 Special Instructions
There are many functions that must be performed by microcode routines which are difficult or inefficient to perform using general-purpose instructions. Furthermore, due to the expanded architecture of the present RISC processor compared to conventional CISC processors, certain functions are useful, whereas such functions would be meaningless for an CISC processor, and thus cannot be performed using any combination of CISC instructions. Together, these situations led to the creation of “special instructions”.
An example of the first category of special instructions is the extract_desc_base instruction. This instruction extracts various bit-fields from two of the microcode general-purpose registers, concatenates them together and places the result in a third general register for use by microcode. To perform the same operation without the benefit of this instruction, microcode would have to perform several masking and shift operations, plus require the use of additional registers to hold temporary values. The special instruction allows the same functionality to be performed by one instruction during a single cycle, and without the use of any scratch registers.
Two examples of the second category of special instructions were already presented: the two return instructions, aret and eret, used to end microcode routines. These instructions are only meaningful in the microcode environment, and thus have no equivalent instructions or instruction sequences in the CISC architecture. In this case, special instructions were required for correct functionality, not just for performance reasons.
Since the special instructions are only available to microcode routines, and emulated instructions can only be encountered in the target CISC instruction stream, the opcodes of emulated instructions are re-used in microcode mode for the special instructions. Thus, when one of these opcodes is encountered in the target CISC instruction stream, it merely indicates that the microcode emulation routine for that instruction should be executed. When the same opcode is encountered in the microcode instruction stream, however, it has a completely different function as one of the special instructions. To support this opcode re-use, the IDU keeps track of the current processor state and decodes the instructions appropriately. This re-use of the opcodes is transparent to the IEU.
The IDU decodes-each CISC instruction (of the i486™ instruction set, for example) and translates each instruction into several RISC processor nano-instructions. As described above, each CISC instruction is translated into 0 to 4 nano-instruction(s), depending on its complexity and functionality. The IDU decodes and translates two CISC instructions per cycle at best case. The basic functions of the IDU can be summarized as follows, it functions to:
A block diagram of the IDU is shown in
The Instruction Decoder 1202 operates in one clock phase (a half cycle). The aligned instruction goes through the decoder and the decoded information that exits is MUXed and fed into a half cycle delay latch 1204 via a bus 1203. The decoded information therefore experiences the equivalent to a one phase pipeline delay.
After the half cycle delay, the decoded information is sent via a bus 1205 to a MUX 1206 to determine the actual register codes used. At this stage of decoding, the decoded information is arranged in the nano-instruction format. The nano-instruction is then latched. Two complete nano-instruction buckets are latched per cycle. The latching of two nano-instruction buckets is shown diagrammatically by 1st IR and 2nd IR buckets 1208 and 1210, respectively.
The IDU attempts to assemble buckets 1208 and 1210 into a single bucket 1212. This assembly is performed by a set of control gates 1214. The IDU first looks at the TYPE of each nano-instruction, and determines if the TYPEs are such that they can be combined. Note that either LoaD (LD) operation of the two latched instructions can be placed in a LD location 1216 of the single bucket 1212; either STore (ST) operation of the latched instructions can be placed in a ST location 1218 of the single bucket; either A0 operation can be placed in an A0 location 1220; and any A0 or A1 operation can be placed in an A1 location 1222.
The IDU treats the instructions as a whole. If the IDU cannot pack the two instructions into one bucket, it will leave one complete instruction behind. For example, if the 1st IR latch has only an A0 operation, and the 2nd IR latch includes all four operations, the IFU will not take the A1 from the 2nd IR latch and merge it with the A0 operation. The A0 operation will be sent by itself and the 2nd IR latch's set of operations will be transferred to the 1st IR latch and sent on the next phase, during which time the 2nd IR latch is reloaded. In other words, the operations stored in the 1st IR latch will always be sent, and the operations stored in the 2nd IR latch will be combined with the 1st IR latch operations if possible. The previous pipeline stages of the IDU and IAU must wait in the event that the 1st and 2nd IRs cannot be combined. The following situations permit the IDU to combine the 1st and 2nd IR latch operations:
Combination logic can readily be designed by those skilled in the art to generate the necessary control signals for the control gates to merge the content of the 1st and 2nd IR latches, based on the functionality discussed above and basic logic design practice.
Emulation mode is entered when the IDU identifies an instruction belonging to the subset of instructions requiring emulation. An EMULation MODE control signal (EMUL_MODE) is sent to the decoders of the IDU once emulation mode is entered. Direct decoding of the CISC instruction stops, and the microcode routine corresponding to the identified instruction is sent to the IDU for decoding. The IDU decoders return to basic mode for decoding further CISC instructions when the microcode routine is finished emulation of the subset instruction. Fundamentally, basic CISC instructions and microcode instructions are handled in the same way by the IDU. Only the interpretation of the opcode changes.
Karnaugh maps of the default (basic) mode for both 1 and 2 byte opcode instructions are shown at
The instruction boxes that are shaded gray in the Karnaugh map of
A block diagram of the IDU's Instruction Decoder 1202 is shown in
A TYPE GENerator (TYPE_GEN) decoder 1402 receives the first full aligned instructions on the ALIGN_IR bus, and decodes instructions one at a time to identify the TYPE field of the instruction.
The identified TYPE field corresponds to the nano-instruction operations discussed above in connection with the IDU. The TYPE is signified by a 4 bit field representing each operation in a bucket (Load, ALU0, Store and ALU1). The TYPE_GEN decoder 1402 specifies which of those four operations are needed to execute the instruction. Depending on the instruction received, any number from 1-4 of the operations may be required to satisfy the CISC instruction.
For example, an add operation, which sums the contents in one register with the contents in another register, requires only one ALU nano-instruction operation. Alternatively, an instruction which requires the addition of the contents of a register with a memory location would require a Load, an ALU operation and then a Store operation, thus totalling three nano-instruction operations. (The data must be read from memory, added to the register, and then stored back in memory). More complicated CISC instructions may require all four nano-instructions.
The TYPE_GEN decoder 1402 comprises three TYPE decoders. A first decoder TYPE1 assumes that the instruction has a one-byte opcode followed by the ModR/M byte, and computes the TYPE based on that assumption. A second decoder TYPE2 assumes that the instruction has a two-byte opcode. The first byte being the ESCAPE byte, followed by the second byte which is the opcode and the third byte which is the ModR/M byte. A third decoder TYPEF assumes that the instruction is a floating point instruction, and decodes the instruction based on that assumption.
The TYPE_GEN decoder has three 4 bit wide TYPE instruction output buses (TYPE1, TYPE2, and TYPEF). Each bit corresponds to one of the 4 nano-instruction operations in a bucket. The specific TYPE field specifies which nano-instruction operations are necessary to carry out the CISC instruction. For example, if all 4 bits are logic HIGH, the CISC instruction requires a Load, a Store and two ALU operations.
The remaining decoders in
The two ALU operations (ALU0 and ALU1) each have an opcode field which is 11 bits long. The 11 bits comprise the 8 bits of the opcode and three opcode extension bits from the adjacent ModR/M byte. For most CISC instructions processed by the IDU, the opcode bits are directly copied to the nano-instruction operations. Some CISC instructions, however, may require opcode substitution; here the IDU unit does not merely filter the CISC opcode to the instruction execution unit (IEU). This will become evident to those skilled in the art, because the type and number of functional units in the IEU will dictate whether or not opcode replacement is required within the IDU for specific CISC instructions.
In order for the IEU to process ALU operations, it must receive information concerning which functional unit is needed to process the specified ALU operation. The IDU therefore includes a Functional zero UNIT (F—0UNIT) decoder 1410, which comprises decoders F—0UNIT1, F—0UNIT2 and F—0UNITF. The outputs of the decoders are multi-byte fields that indicate which functional unit is necessary for processing the A0 ALU operation. The functional unit decoding for the A1 ALU operation is identical, but is handled by a separate decoder F—1UNIT 1412.
Many CISC instructions carry out operations using registers that are implied by the opcode. For example, many instructions imply that the AX register is to be used as an accumulator. A ConSTant GENerator (CST_GEN) decoder 1414 is therefore included to generate register indices based on the opcode of the CISC instruction. The CST_GEN decoder specifies which register(s) are implied based on the specific opcode. Multiplexing for generating the correct source and destination register indices for the nano-instructions will be discussed below in conjunction with
An additional two bit control signal, TempCount (TC), is input to the CST_GEN decoder. The TC control signal is a two bit counter representing 4 temporary registers which may be cycled through for use as dummy registers by the IEU. The temporary (or dummy) registers represent another value of register that can be passed on by the CST_GEN decoder, in addition to the implied registers. The constant generator decoder passes on 4 constant fields because there are 2 ALU operations having 2 registers per operation. Each constant register bus is 20 bits wide, with each constant being a total of 5 bits, thereby permitting selection of one of the 32 registers in the IEU.
A SELect GENerator (SEL_GEN) decoder, shown generally at block 1416, will now be discussed. The SEL_GEN decoder includes a FlaG Need Modify (FG_NM) decoder 1418. The FG_NM decoder decodes for a one-byte opcode, a 2 byte opcode and a floating point instruction. In the i486™ instruction set, for example, there are a total of 6 flags. These flags have to be valid before execution of some instructions begin, while the flags may be modified by some instructions. The FG_NM decoder outputs two signals per flag, one bit indicates whether the flag is needed for execution of this instruction and the other indicates whether or not this instruction actually modifies the flag.
Register INValiDation information concerning the ALU0 and ALU1 operations are decoded by an INVD1 and an INVD2 decoder, shown at 1420 and 1422 respectively. The INVD1 and INVD2 decoders are also part of the SEL_GEN decoder 1416. INVD1 and INVD2 generate control signals for the IEU. These signals indicate whether the ALU registers should be used or not. Three possible register indices can be specified by each ALU operation. One can be used as a source and/or destination register, and the remaining two are limited to specifying source registers. A 4 bit field is uses to specify which register(s) are required by the operation.
The SEL_GEN decoder 1416 further includes a FLD_CNT decoder 1424 that indicates which of the register fields is required for the CISC instruction. The FLD_CNT decoder specifies which of the 2 fields is the source register and which is the destination register.
A Nano-InstRuction GENerator (NIR_GEN) decoder is shown generally as block 1426. The data size (DATA_SZ) and address size (ADDR_SZ) input control signals correspond to the default that the system is operating in. In order to decode the final address and operand size, the default mode must be known and the presence of any prefixes (discussed above in conjunction with the IAU) must be known. The EMUL_MODE control signal is also input to the NIR_GEN decoder, but it is also used by the other decoders.
The ESCape DETect (ESC_DET) input control signal is fed to the NIR_GEN decoder to indicate whether the instruction has a 2 byte opcode. In addition, a SELect OPcode EXTension (SEL_OP_EXT) input control signal is used to generate loading of the mailbox registers when an emulation instruction is detected.
A Floating Point REGister (FP_REG) input control signal passes the translated floating point register index to the IDU. The floating point format of the i486™, for example, has eight registers for floating point numbers, but the registers are accessed like a stack. Accessing these registers is accomplished by using a stack accessing scheme: register0 being the top of the stack, register1 being the next top register, etc. This register stack is emulated by using eight linear registers with fixed indices. When the input instruction specifies registers, a translation block (not shown), translates the stack relative register index into the register index for the linear registers in a known manner. This permits the IDU to keep track of which register is on the top of the stack.
When the system branches to emulation mode, the IDU saves information about the instruction being emulated. The IDU saves the Data SIZE (EM_DSIZE) and Address SIZE (EM_ASIZE) of the instruction, as well as the Register index of the DESTination (EM_RDEST), the source (EM_RDEST2) and the Base InDeX information (EM_BSIDX). This saved information is used by the microcode routine to properly emulate the instruction. Take for example the emulation of an add instruction. The microcode routine may check EM_ASIZE to determine the address size of the add instruction so that it knows what address size to emulate.
The NIR_GEN decoder 1426 includes a SIZE decoder 1428. The fields generated by the SIZE decoder (i.e., SIZE1, SIZE2 and SIZEF) indicate the address size, operand size and immediate data size of the instruction. An address size of 16 or 32 bits, an operand size of 8, 16 or 32 bits and an immediate data field size of 8, 16 or 32 bits are extracted for each instruction.
Another NIR_GEN decoder is called a LoaD INFormation (LD_INF) decoder 1430. The LD_INF decoder decodes information corresponding to the Load and Store operations. The Load information is used for effective address calculations. The Load information fields (LD_INF1, LD_INF2 and LD_INFF) can be used to specify which addressing mode is being used by the CISC instruction, since CISC instruction sets usually support many different addressing modes.
The i486™ basic addressing mode includes a segment field and an offset which are added together to determine the address. An index register can be specified, as well as a scale for the index register (e.g., if the index registers are elements in an array), the elements can be specified as 1, 2, 4 or 8-bytes in length, thus the index register can be scaled by 1, 2, 4 or 8 before it is added to determine the address. The base and index are also specified by the LD_INF fields.
A Nano-InstRuction OPCode (NIR_OPC) decoder 1432 transfers opcode for the A1 operation (packet1). The decoded fields (NIR_OPC1, NIR_OPC2 and NIR_OPCF) comprise the first instruction byte (8 bits), plus three extension bits from the second byte.
A MIScellaneous OPCode (MISC_OPC) decoder 1434 indicates whether the instruction is a floating point instruction and whether a load instruction is actually present. The field generated by the MISC_OPC decoder will indicate whether conversion of the floating data is necessary. Multiplexing is not necessary for this decoder, because this information is easily extracted, regardless of the format of the instruction.
The opcode for the A0 operation of packet0 is specified by an OP_CODE decoder 1436. The A0 opcode is usually copied directly from the i486™ input opcode, but for some instructions the opcode is replaced by an alternate opcode. (As noted above, the functionality of the signals generated by the NIR_GEN decoder are specific to the CISC instruction set being decoded, and thus should become evident to those skilled in the art upon review of the CISC instruction set and the nano-instruction format of the present invention.)
An EXT_CODE decoder 1440 extracts the 3 bit opcode extension from the ModR/M byte.
A IN_ORDER decoder 1442 decodes the instruction to determine whether the instruction must be executed “in order”. This instructs the IEU not to do anything with this instruction until all the previous instructions have been executed. Once the execution of the instruction is completed, execution of subsequent instructions is started.
A Control Flow Jump Size decoder 1444 indicates the displacement size for jumps that specify an address. This field, labeled CF_JV_SIZE, specifies the size of the address for the jump. This is specific to the type of addressing scheme employed by the CISC instruction set.
A 1 bit decoder labeled DEC_MDEST 1446 indicates whether or not the destination of the instruction is a memory address.
Finally, the Instruction Decoder includes three Register Code decoders 1438 to select the register codes (indices). The i486™ instruction format encodes the index of the register fields in various places within the instruction. The indices of these fields are extracted by the RC decoder. The ModR/M byte also has two register indices, which are used as the destination/source as specified by the opcode itself. The Register Code decoder 1438 generates three RC fields RC1, RC2 and RC3. RC1 and RC2 are extracted from the ModR/M byte as follows, if the processor is not in emulation mode, and that instruction is not a floating point instruction: RC1=bits [2:0] of the ModR/M byte; RC2=bits [5:3] of the ModR/M byte; and RC3=bits [2:0] of the opcode. For floating point instructions in basic (not emulation) mode, RC1, RC2 and RC3 are assigned as follows:
In emulation mode, RC1, RC2 and RC3 are assigned as follows:
As discussed above, the CST_GEN decoder 1414 generates three outputs, CST1, CST2 and CSTF, each of which comprise four constant 5 bit register fields (20 bits total). The SEL_GEN generates register field control signals (FLD1, FLD2, and FLD3 for the selection of the multiplexers in a further section MUX 1512. The selection of the CST1, CST2 or CSTF results and the FLD1, FLD2, and FLDF results is shown generally at the multiplexer block 1502. A 3 bit MUX select line 1504 is used to select the results depending on whether the instruction has a 1 byte opcode, 2 byte opcode, or is a floating point instruction.
A ½ cycle pipeline delay latch 1504 is used to delay the results selected by the multiplexer 1502, and the three register control fields RC1, RC2, and RC3. Each input to the ½ pipeline delay 1504 is sent to a pair of oppositely clocked latches 1508. The contents of the latches are selected by a multiplexers 1510. This arrangement is similar to the ½ cycle delay 316 discussed above in connection to the IAU.
A further multiplexing stage is shown in block 1512. The constant register fields selected by the multiplexer 1502 are input to the multiplexer 1512 as four separate fields labeled regc1 through regc4, respectively, as shown generally at 1514. Also shown as inputs to the block 1512 are the EXTRACT REGISTER fields RC1, RC2, and RC3 from the opcode and ModR/M bytes. The regc fields and RC fields are combined by logic in the block 1512 under control of an FLD control signal 1520 to generate the source and destination register indexes a0_rd and a0_rs for operation A0, which are shown generally at 1516, as well as the source and destination register indexes a1_rd and a1_rs for operation A1, which are shown generally at 1518. An index Id_rd, which is the destination register index for the Load instruction, is also selected in the block 1512.
4.0 Decoded Instruction FIFO
A block diagram of a Decode FIFO (DFIFO) in conjunction with the present invention is shown in
As can be seen from
A specific example of the three types of nano-instruction described above is shown in
While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example, and not limitation. Thus the breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.
This application is a continuation of application Ser. No. 10/061,295, filed Feb. 4, 2002, now allowed, which is a continuation of application Ser. No. 09/852,295, filed May 10, 2001, now abandoned, which is a continuation of application Ser. No. 09/401,860, filed Sep. 22, 1999, now U.S. Pat. No. 6,263,423, which is a continuation of application Ser. No. 08/784,339, filed Jan. 16, 1997, now U.S. Pat. No. 5,983,334, which is a continuation of application Ser. No. 08/460,272, filed Jun. 2, 1995, now U.S. Pat. No. 5,619,666, which is a continuation of application Ser. No. 07/857,599, filed Mar. 31, 1992, now U.S. Pat. No. 5,438,668. The disclosures of each of the above-referenced applications are incorporated herein by reference. The following are related applications: “A ROM With RAM Cell and Cyclic Redundancy Check Circuit,” application Ser. No. 07/802,816, filed Dec. 6, 1991, now abandoned; “High-Performance, Superscalar-Based Computer System with Out-of-Order Instruction Execution,” application Ser. No. 07/817,810, filed Jan. 8, 1992, now U.S. Pat. No. 5,539,911, which is a continuation of application Ser. No. 07/727,006, filed Jul. 8, 1991, now abandoned; “High-Performance, Superscalar-Based Computer System with Out-of-Order Instruction Execution and Concurrent Results Distribution,” application Ser. No. 08/397,016, filed Mar. 1, 1995, now U.S. Pat. No. 5,560,032, which is a file wrapper continuation of application Ser. No. 07/817,809, filed Jan. 8, 1992, now abandoned, which is a continuation of application Ser. No. 07/727,058 filed Jul. 8, 1991, now abandoned. The disclosures of each of the above-referenced applications are also incorporated herein by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 3346851 | Thornton et al. | Oct 1967 | A |
| 3771138 | Celtruda et al. | Nov 1973 | A |
| 3789365 | Jen et al. | Jan 1974 | A |
| 3916388 | Shimp et al. | Oct 1975 | A |
| 4034349 | Monaco et al. | Jul 1977 | A |
| 4084235 | Hirtle et al. | Apr 1978 | A |
| 4189768 | Liptay et al. | Feb 1980 | A |
| 4189772 | Liptay | Feb 1980 | A |
| 4200927 | Hughes et al. | Apr 1980 | A |
| 4228495 | Bernhard et al. | Oct 1980 | A |
| 4236206 | Strecker et al. | Nov 1980 | A |
| 4296470 | Fairchild et al. | Oct 1981 | A |
| 4315314 | Russo | Feb 1982 | A |
| 4317170 | Wada et al. | Feb 1982 | A |
| 4410939 | Kawakami | Oct 1983 | A |
| 4434461 | Puhl | Feb 1984 | A |
| 4454578 | Matsumoto et al. | Jun 1984 | A |
| 4456955 | Yanagita et al. | Jun 1984 | A |
| 4459657 | Murao | Jul 1984 | A |
| 4476525 | Ishii | Oct 1984 | A |
| 4514803 | Agnew et al. | Apr 1985 | A |
| 4569016 | Hao et al. | Feb 1986 | A |
| 4587612 | Fisk et al. | May 1986 | A |
| 4626989 | Torii | Dec 1986 | A |
| 4654781 | Schwartz et al. | Mar 1987 | A |
| 4675806 | Uchida | Jun 1987 | A |
| 4714994 | Oklobdzija et al. | Dec 1987 | A |
| 4722049 | Lahti | Jan 1988 | A |
| 4739471 | Baum et al. | Apr 1988 | A |
| 4752873 | Shonai et al. | Jun 1988 | A |
| 4758948 | May et al. | Jul 1988 | A |
| 4766566 | Chuang | Aug 1988 | A |
| 4771376 | Kamiya | Sep 1988 | A |
| 4807115 | Torng | Feb 1989 | A |
| 4814976 | Hansen et al. | Mar 1989 | A |
| 4841476 | Mitchell et al. | Jun 1989 | A |
| 4858105 | Kuriyama et al. | Aug 1989 | A |
| 4897810 | Nix | Jan 1990 | A |
| 4901228 | Kodama | Feb 1990 | A |
| 4903196 | Pomerene et al. | Feb 1990 | A |
| 4924376 | Ooi | May 1990 | A |
| 4926323 | Baror et al. | May 1990 | A |
| 4942525 | Shintani et al. | Jul 1990 | A |
| 4985825 | Webb, Jr. et al. | Jan 1991 | A |
| 4992931 | Hirasawa | Feb 1991 | A |
| 4992934 | Portanova et al. | Feb 1991 | A |
| 4992938 | Cocke et al. | Feb 1991 | A |
| 5003462 | Blaner et al. | Mar 1991 | A |
| 5019967 | Wheeler et al. | May 1991 | A |
| 5101341 | Circello et al. | Mar 1992 | A |
| 5127091 | Boufarah et al. | Jun 1992 | A |
| 5133072 | Buzbee | Jul 1992 | A |
| 5148528 | Fite et al. | Sep 1992 | A |
| 5168571 | Hoover et al. | Dec 1992 | A |
| 5193206 | Mills | Mar 1993 | A |
| 5226126 | McFarland et al. | Jul 1993 | A |
| 5226170 | Rubinfeld | Jul 1993 | A |
| 5230068 | Van Dyke et al. | Jul 1993 | A |
| 5307492 | Benson | Apr 1994 | A |
| 5355460 | Eickemeyer et al. | Oct 1994 | A |
| 5390355 | Horst | Feb 1995 | A |
| 5430862 | Smith et al. | Jul 1995 | A |
| 5438668 | Coon et al. | Aug 1995 | A |
| 5442757 | McFarland et al. | Aug 1995 | A |
| 5450575 | Sites | Sep 1995 | A |
| 5487156 | Popescu et al. | Jan 1996 | A |
| 5507030 | Sites | Apr 1996 | A |
| 5539911 | Nguyen et al. | Jul 1996 | A |
| 5546552 | Coon et al. | Aug 1996 | A |
| 5561776 | Popescu et al. | Oct 1996 | A |
| 5574927 | Scantlin | Nov 1996 | A |
| 5592636 | Popescu et al. | Jan 1997 | A |
| 5619666 | Coon et al. | Apr 1997 | A |
| 5625837 | Popescu et al. | Apr 1997 | A |
| 5627983 | Popescu et al. | May 1997 | A |
| 5651125 | Witt et al. | Jul 1997 | A |
| 5655097 | Witt et al. | Aug 1997 | A |
| 5689720 | Nguyen et al. | Nov 1997 | A |
| 5708841 | Popescu et al. | Jan 1998 | A |
| 5751981 | Witt et al. | May 1998 | A |
| 5768575 | McFarland et al. | Jun 1998 | A |
| 5778210 | Henstrom et al. | Jul 1998 | A |
| 5797025 | Popescu et al. | Aug 1998 | A |
| 5832205 | Kelly et al. | Nov 1998 | A |
| 5832293 | Popescu et al. | Nov 1998 | A |
| 5983334 | Coon et al. | Nov 1999 | A |
| 6141742 | Favor | Oct 2000 | A |
| 6253309 | Mahalingaiah | Jun 2001 | B1 |
| 6263423 | Coon et al. | Jul 2001 | B1 |
| 20030084270 | Coon et al. | May 2003 | A1 |
| Number | Date | Country |
|---|---|---|
| 0 136 179 | Apr 1985 | EP |
| 0 170 398 | Feb 1986 | EP |
| 0 171 595 | Feb 1986 | EP |
| 0 272 198 | Jun 1988 | EP |
| 0 272 198 | Jun 1988 | EP |
| 0 354 585 | Feb 1990 | EP |
| 0 354 585 | Feb 1990 | EP |
| 0 368 332 | May 1990 | EP |
| 0 368 332 | May 1990 | EP |
| 0 372 751 | Jun 1990 | EP |
| 0 372 751 | Jun 1990 | EP |
| 0 377 991 | Jul 1990 | EP |
| 0 377 991 | Jul 1990 | EP |
| 0 380 854 | Aug 1990 | EP |
| 0 380 854 | Aug 1990 | EP |
| 0 402 856 | Dec 1990 | EP |
| 0 402 856 | Dec 1990 | EP |
| 0 419 105 | Mar 1991 | EP |
| 0 419 105 | Mar 1991 | EP |
| 0 426 393 | May 1991 | EP |
| 0 426 393 | May 1991 | EP |
| 0 459 232 | Dec 1991 | EP |
| 0 461 257 | Dec 1991 | EP |
| 0 473 420 | Mar 1992 | EP |
| 0 473 420 | Mar 1992 | EP |
| 0 479 390 | Apr 1992 | EP |
| 0 479 390 | Apr 1992 | EP |
| 2 575 564 | Jul 1986 | FR |
| 2 011 682 | Jul 1979 | GB |
| 2 230 116 | Oct 1990 | GB |
| 62-219128 | Sep 1987 | JP |
| 63-163634 | Jul 1988 | JP |
| 63-192135 | Aug 1988 | JP |
| 3-201030 | Sep 1991 | JP |
| 5-197544 | Aug 1993 | JP |
| Number | Date | Country | |
|---|---|---|---|
| 20050251653 A1 | Nov 2005 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10061295 | Feb 2002 | US |
| Child | 11167289 | US | |
| Parent | 09852295 | May 2001 | US |
| Child | 10061295 | US | |
| Parent | 09401860 | Sep 1999 | US |
| Child | 09852295 | US | |
| Parent | 08784339 | Jan 1997 | US |
| Child | 09401860 | US | |
| Parent | 08460272 | Jun 1995 | US |
| Child | 08784339 | US | |
| Parent | 07857599 | Mar 1992 | US |
| Child | 08460272 | US |
