US 5,983,334 AGrant
Superscalar Microprocessor for Out-of-Order and Concurrently Executing at Least Two RISC Instructions Translating From in-Order CISC Instructions
Issue Date:1999-11-09
•5 Claims
•17 Drawing Sheets
Abstract
A system and method for extracting complex, variable length computer instructions from a stream of complex instructions each subdivided into a variable number of instructions bytes, and aligning instruction bytes of individual ones of the complex instructions. The system receives a portion of the stream of complex instructions and extracts a first set of instruction bytes starting with the first instruction bytes, using an extract shifter. The set of instruction bytes are then passed to an align latch where they are aligned and output to a next instruction detector. The next instruction detector determines the end of the first instruction based on said set of instruction bytes. An extract shifter is used to extract and provide the next set of instruction bytes to an align shifter which aligns and outputs the next instruction. The process is then repeated for the remaining instruction bytes in the stream of complex instructions. The isolated complex instructions are decoded into nano-instructions which are processed by a RISC processor core.
Metadata
Assignee
- Seiko Epson Corporation
Inventors
- Brett Coon
- Yoshiyuki Miyayama
- Le Trong Nguyen
- Johannes Wang
Application Information
Application Number:US 7843399
Filing Date:1997-01-16
Priority Date:1992-03-31
Art Unit:278
Classifications
IPC:
G06F 930G06F 938
Field of Search:
71223;41;206;215;217
Patent Drawings (17 sheets)
Description
Background of the Invention
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 byte;. 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.TM. Programmer's Reference Guide, several conclusions can be drawn regarding alignment unit present in the i486.TM.. The i486.TM.'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.TM. AU include the presence of prefixed and escaped (2 byte) opcodes. Both of these are common in i486.TM. programs. In addition, complex instructions may also comprise displacement and immediate data. The i486.TM. requires additional time to extract this data.
An example format for a CISC processor instruction is shown in FIG. 1. The example depicts the potential bytes of a variable length i486.TM. 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.
Summary of the Invention
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.times.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 on 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 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.
Brief Description of the Drawings
The invention will be better understood if reference is made to the accompanying drawings in which:
FIG. 1 shows the data structure format for a conventional CISC instruction.
FIG. 2 shows a block diagram of the instruction prefetch buffer of the present invention.
FIG. 3 shows a block diagram of the instruction alignment unit of the present invention.
FIG. 4 shows a representative flow chart of the instruction extraction and alignment method of the IAU of the present invention.
FIG. 5 shows a simplified timing diagram associated with the block diagram of FIG. 3 and the flow chart of FIG. 4.
FIG. 6 is a block diagram of the STACK of the present invention.
FIG. 7A is a block diagram of the Next Instruction Decoder (NID) of the present invention.
FIG. 7B is a block diagram of the Remaining Next Instruction Decoder (RNID) of the present invention.
FIG. 8 is a block diagram of the Immediate Data and Displacement Decoder (IDDD) of the present invention.
FIG. 9 is a block diagram of a Prefix Decoder (PD) of the present invention.
FIG. 10 is a block diagram of the PReFiX Number (PRFXk.sub.-- NO) decoder of the present invention.
FIG. 11 is a block diagram of a nano-instruction bucket of the present invention.
FIG. 12 is a representative block diagram of the instruction decode unit (IDU) of the present invention.
FIGS. 13A, 13B, 13C, 13D, and 13E show instruction bit maps of the present invention.
FIG. 14 shows an example block diagram of the Instruction Decoder section of the IDDD of the present invention.
FIG. 15 depicts a representative block and logic diagram of a set of decoders of the Instruction Decoder shown in FIG. 14.
FIGS. 16A, 16B and 16C show a conceptual block diagram of the decode FIFO of the present invention.
FIG. 17 shows examples of the nano-instruction field formats of the present invention.
Table of Contents
Detailed Description of the Preferred Embodiments
1.0 The Instruction Fetch Unit
2.0 Instruction Alignment Unit Overview
2.1 Instruction Alignment Unit Block Diagrams
3.0 Instruction Decode Unit Overview
3.1 Microcode Dispatch Logic
3.2 Mailboxes
3.3 Nano-Instruction Format
3.4 Special Instructions
3.5 Instruction Decode Unit Block Diagrams
4.0 Decoded Instruction FIFO
Detailed Description of the Preferred Embodiments
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 Cliff, N.J., 1991); John L. Hennessy et al., Computer Architecture--A Quantitative Approach" (Morgan Kaufmann Publishers, Inc., San Mateo, Calif., 1990); and the i486.TM. Microprocessor Programmer's Reference Manual and the i486.TM. 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. FIG. 2 shows a block diagram of three Instruction Prefetch Buffers 200 within the IFU, which comprises: a Main instruction BUFfer (MBUF) 204, an Emulation instruction BUFfer (EBUF) 202, and a Target instruction BUFfer (TBUF) 206. The Prefetch Instruction Buffers can load 128 bits (16 bytes) of an instruction stream from an instruction cache in a single cycle. This data is held in one of the three buffers for use by the IAU.
During normal program execution, the MBUF 202 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 202. 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 204 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 204. (Emulation mode and exception handling will both be discussed below in detail.) Execution continues out of the EBUF 204 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 202. 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 prefix(es) 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.TM.) 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:
detect the presence and the length of prefix bytes;
isolate the Opcode, ModR/M and SIB (scale, index, base) bytes;
detect the length of instructions (which indicates the location of the next instruction); and
send the following information to an Instruction Decode Unit (IDU):
Opcode, eight bits plus 3 optional extension bits. For 2 byte opcodes, the first byte is always OF hex, so the second byte is sent as the opcode.
ModR/M byte, SIB byte, and Displacement and Immediate data; and
Information concerning the number and type of prefixes.
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 Instuction Alignment Unit Block Diagrams
A block diagram of the IAU is shown in FIG. 3. The diagram is divided into two sections: a MAIN DATAPATH 302 (indicated by the dashed line box) and a PREDECODER 304 (indicated by the dashed line box). Instruction shifting and extraction occurs in the MAIN DATAPATH 302, while length determination and datapath control are handled by the PREDECODER 304.
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) IFI0b.sub.-- Bus[127:0] and IFI1b.sub.-- 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; 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 1/2 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.sub.-- 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.sub.-- 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 325, a stack 334 and a further bus 336. The bus 336 also sends the current aligned instruction information to the 1/2 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 1/2 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 oil 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 308 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 oi 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 1/2 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 308 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 FIGS. 7A and 7B.
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 half cycle, which brings its total decode lime 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.
FIG. 4 shows a general flow chart of instruction byte extraction and alignment performed by the IAU of the present invention. When new data enters the lowest line 205 of the IFU's MBUF 204 (called BUCKET.sub.-- #0), the EXTRACT SHIFTER 306 extracts 8 bytes starting with the first instruction, as shown at a step 402. The 8 instruction bytes are passed along to the ALIGN.sub.-- IR latches 318 and 320, while bypassing the ALIGN SHIFTER 310, as shown at a step 404. The IAU then waits for the next clock phase while it holds the aligned instruction irk the ALIGN.sub.-- IR latch, as shown at a step 406.
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 1/2 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 FIG. 4.
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 1/2 cycle delay 316, as shown at a block 416. Next, the IAU uses the output of the MUX 330 as a variable called SHIFT.sub.-- A, which is used to control the ALIGN SHIFTER 310 to align the next instruction. The SHIFT.sub.-- A is also added to the current EXTRACT SHIFTER shift amount (called BUF.sub.-- 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.sub.-- 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 1/2 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 322 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.
FIG. 5 shows a timing diagram associated with the IAU of FIG. 3. Two instruction buckets are shown at the top of FIG. 5. These instruction buckets, labeled BUCKET.sub.-- #0 and BUCKET.sub.-- #1, each comprise, 16 instruction bytes which are provided by the IFU (from an instruction memory not shown) to the IAU in FIG. 3. Instruction alignment is always done from the right out of BUCKET.sub.-- #0, (i.e., the bottom bucket). In this example, BUCKET.sub.-- #0 and BUCKET.sub.-- #1 are the bottom two buckets of the IFU's MBUF 204. Other arrangements are also possible.
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.sub.-- #0. The remaining 3 bytes wrap to the beginning of BUCKET.sub.-- #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 FIG. 5 have been labeled .phi.1, .phi.2 and .phi.3 to aid this discussion.
Traces "2" and "3" in the timing diagram represent instruction data on the input buses IFI1B and IFI0B. A new BUCKET.sub.-- #0 becomes available on bus IFI0B at the beginning of .phi.1, 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.sub.-- #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.sub.-- #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.sub.-- #0 is depleted of instruction bytes, the contents of Bucket.sub.-- #1 are shifted into Bucket.sub.-- #0, and Bucket.sub.-- #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 .phi.1 at the time period shown by a reference numeral 506.
The first 8 bytes of Bucket.sub.-- #0 bypass the ALIGN SHIFTER and are stored in the two ALIGN.sub.-- IR latches 318 or 320 (as shown at traces "6" and "7" in FIG. 3). The ALIGN.sub.-- IR latches receive the instruction bytes in an alternating fashion, based on the timing of clock signals CLK0 and CLK1. ALIGN.sub.-- IR0 318 is a clock signal CLK0 latch, meaning that it is latched while clock signal CLK0 is high. ALIGN.sub.-- IR1 320 is a clock signal CLK1 latch, which latches when clock signal CLK0 is high. The first 8 bytes become valid at the ALIGN.sub.-- IR0 prior to the end of the first clock signal CLK0 phase, as shown by a reference numeral 508 toward the end of .phi.1.
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, .phi.2.
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 .phi.2 of the timing diagram, the EXTRACT SHIFTER shifts out 8 bytes, bytes 15-8 of BUCKET.sub.-- #0. See 510 at FIG. 5. These bytes are sent to the ALIGN SHIFTER, which now has a total of 16 consecutive bytes to work with. The ALIGN SHIFTER looks at the output of the EXTRACT SHIFTER and the valid output of the latches 318 and 320 during .phi.2.
Toward the end of .phi.2, the ALIGN SHIFTER shifts bytes 12-5 of BUCKET.sub.-- #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.sub.-- .sub.-- 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 .phi.2. Therefore, in .phi.3, 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.sub.-- IR latches will output bytes 12-5 during .phi.3; and therefore, bytes 20-5 will be available at the ALIGN SHIFTER.
During .phi.3, the EXTRACT SHIFTER will output bytes 20-13. However, BUCKET.sub.-- #0 only contains bytes 15-0, therefore, bytes 20-16 must be taken from BUCKET.sub.-- #1. As shown at 514 in the timing diagram, BUCKET.sub.-- #1 becomes valid at the beginning of .phi.3. The EXTRACT SHIFTER then shifts bytes 4-0 of BUCKET.sub.-- #1 and bytes 15-13 of BUCKET.sub.-- #0, as shown at 516. If BUCKET.sub.-- #1 was not valid at this time, the IAU would have to wait until it becomes valid.
As noted above, the Shift.sub.-- 5.sub.-- byte signal was generated by the NID during .phi.2. Based on this signal, bytes 12-5 of BUCKET.sub.-- #0 are shifted out by the ALIGN SHIFTER, as shown at 518, and shortly thereafter are latched into ALIGN.sub.-- 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 .phi.3. 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.sub.-- 3.sub.-- bytes signal during the latter half of .phi.3, 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.sub.-- #0 (i.e., BUCKET.sub.-- #0 is completely used), BUCKET.sub.-- #1 will become BUCKET.sub.-- #0 and a new BUCKET.sub.-- #1 will later become valid.
Trace "10" in the timing diagram shows the timing for extraction of bytes from the instruction stream. The Buf.sub.-- Count#0 blocks represent the stored extract shift amount. During each phase the aligned shift amount is added to Buf.sub.-- Count#0, and the result becomes the extract shift amount during the next phase (see the blocks labeled COUNTER.sub.-- SHIFT).
Trace "11" in the timing diagram shows instruction alignment timing. The blocks labeled IR.sub.-- Latch.sub.-- #0 and IR.sub.-- Latch.sub.-- #1 represent the time during which the instructions in the corresponding ALIGN.sub.-- 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.sub.-- SHIFT represent the time when the ALIGN SHIFTER begins to output the instruction.
S 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 FIG. 6. The STACK comprises 64 1-bit stacks that are arranged in parallel. Each 1 bit stack 600 comprises two latches 602 and 604, and a three input MUX 606. The aligned instructions are input to the latches and the MUX on a bus 607 labeled IN. The loading of the two latches may be done independently on either clock phase. In addition, the MUX 606 has three MUX control lines 608 to select the output of either latch, or bypass the IN data directly to an output 610 labeled OUT.
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 (ISC 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 (ISC instruction is completed.
The 1/2 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 operation,; 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.TM. 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 FIG. 7A. The following discussion of the NID is specific to alignment of i486.TM. instructions. Alignment of other CISC instructions would likely employ a different NID architecture. The techniques discussed below should therefore serve as a guide to those skilled in the art, but should not be considered to limit the scope of the present invention.
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.) FIG. 7A shows a 4 byte (32 bit) bus 701 representing the first 4 bytes of an instruction received from the MUX 322. The first 2 bytes, are sent to the SNID 702 on a bus 703. The SNID determines the length of a first subset of instructions that are, by definition, identifiable based on the first 2 bytes. The SNID can determine the length of this subset of instructions in a half cycle. The length of the subset instructions is output by the SNID on a bus 705. The width of the bus may correspond to the maximum number of instruction bytes detected by the SNID. The SNID also has a 1 bit MOD DETect (MOD.sub.-- DET) output line 707 to indicate whether a ModR/M byte is present in the instruction. In addition, the SNID has a 1 bit NID.sub.-- WAIT line 709 to signal the control logic that the instruction is not in the subset (i.e., use the RNID's output instead). The IAU must therefore wait a half cycle for the RNID to decode the instruction if NID.sub.-- WAIT is true.
The subset of instructions decoded by the SNID are those (ISC instructions that can be decoded in a half cycle using a minimum of 1, 2 and 3 input gates (NANDs, NORs and inventors), with a maximum of 5 gate delays based on an 16.times.16 Karnaugh map of the 256 instructions. Blocks of the map including most 1 byte opeode 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.sub.-- 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.sub.-- 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.sub.-- WAIT signal 709. The output of the SNID is selected during the first clock phase because those results will be complete. If the NID.sub.-- 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 )ET) input signal indicates whether the opcode is 1 byte or 2 bytes in length. For example, in the i486.TM. 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.sub.-- 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 FIG. 7B. The RNID comprises an RNID.sub.-- 1OP decoder 752, which decodes the first opcode byte, an RNID.sub.-- 2OP decoder 754, which decodes the second opcode byte, two identical RNID.sub.-- MOD decoders 756 and 758, which decode the ModR/M bytes in either of the two positions determined by the number of opcode bytes present, and an RNID.sub.-- SUM summer 760. Based on the outputs of the four RNID decoders 752-758, the RNID.sub.-- SUM summer 760 outputs the total length of the instruction on a bus 762. The RNID.sub.-- SUM summer 760 has an additional output line 764 labeled OVER8, to indicate whether the instruction is over 8 bytes in length.
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.sub.-- 1OP 752 on a bus 766. A further input line 768 called DATA.sub.-- SZ to the RNID.sub.-- 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.sub.-- 1OP assumes that the instruction has a 1 byte opcode, and based on that information and the 3 extension bits, RNID.sub.-- 1OP attempts to determine the length of the instruction.
The RNID.sub.-- MOD decoder 754 decodes the ModR/M byte of the instruction input on a bus 770. The RNID.sub.-- MOD decoder has an additional input bus 772 labeled ADD.sub.-- SZ, which indicates whether the address size is 16 or 32 bits. The address size is independent of the data size.
The ESC.sub.-- DET signal 774 is also input to block 760. When the ESC.sub.-- DET signal is logic HIGH, for example, the RNID.sub.-- SUM block knows that the opcode is actually in the second byte.
The RNID.sub.-- 2OP decoder 754 assumes that the opcode is 2 bytes, and therefore decodes the second byte (see bus 776) of the opcode. RNID.sub.-- 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.sub.-- 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.sub.-- MOD decoders are identical, but decode different bytes in the instruction stream.
Again, based on the ESC.sub.-- DET signal 774, the RNID.sub.-- SUM 760 selects the outputs of the appropriate opcode and ModR/IM 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.sub.-- NO[7:0] bus 762 indicates the number of instruction bytes over 8.
The RNID.sub.-- 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.TM. 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.sub.-- 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.TM. instruction set if the opcode is 2 bytes long.
Outputs 784 and 786 of the two RNID.sub.-- MOD decoders 756 and 758 indicate to the RNID.sub.-- SUM 760 the five possible additional lengths that can be specified by the ModR/M byte. Each RNID.sub.-- 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.
FIG. 8 shows a block diagram of the IDDD 326. The IDDD determines the shift amounts for the IMM SHIFTER 312 and the DISP SHIFTER 314. The shift amount is determined by the ModR/M byte of the instruction.
The i486.TM. instruction set includes two special instructions, the enter.sub.-- detect and jump.sub.-- call.sub.-- 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.sub.-- DET and JMP.sub.-- CL.sub.-- DET (820 and 822, respectively) indicate whether one of the corresponding instructions is detected.
MOD.sub.-- DEC decoders 804 and 806 are identical and decode the immediate and displacement data. Based on ADD.sub.-- 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.sub.-- 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.sub.-- DET lines 707 and 714 are also input to the SELECT block 812.
The SELECT block 812 combines the EN.sub.-- DET and JMP.sub.-- CL.sub.-- DET signals, the MOD.sub.-- DET and MOD.sub.-- DEC results, and the ADD.sub.-- SZ and outputs its results on four buses 832-838. A DISPlacement 1 (DISP.sub.-- 1) bus 832 outputs the displacement shift results assuming a 1 byte opcode. A DISPlacement 2 (DISP.sub.-- 2) bus 834 outputs the displacement shift results assuming a 2 byte opcode. IMMediate 1 and 2 (IMM.sub.-- 1 and IMM.sub.-- 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.sub.-- SEL/DLY actually selects the appropriate shift amounts and delays these results a half cycle. The half cycle delay performed by MOD.sub.-- SEL/DLY 816 represents the delay 316 shown in FIG. 3. The ESC.sub.-- DET signal 774 described above is used by the MOD.sub.-- SEL/DLY block to perform the shift selection. The results are clocked out of the MOD.sub.-- SEL/DLY 814 by the clock signals CLK0 and CLK1 after a half cycle delay. The immediate data shift control signal and the displacement shift control signal are sent to the DISP SHIFTER and the IMM SHIFTER via a SHIFT.sub.-- D[3:0] bus 840 and a SHIFT[7:0] bus 842, respectively. The number of possible positions within the CISC instruction of the immediate and displacement data define the number of bits required to specify the amount of shift
A block diagram of the PREFIX DETECTOR 328 is shown in FIG. 9. The PREFIX DETECTOR 328 comprises a Prefix.sub.-- Number decoder (PRFX.sub.-- NO) 902, four Prefix.sub.-- Detector decoders (PRFX.sub.-- DECs 904-910), and a Prefix-Decoder (PRFX.sub.-- SEL) 912.
The i486.TM. 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.TM. LOCK prefix, and is only available to microcode routines during emulation mode.
An ALIGN.sub.-- 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.sub.-- 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.sub.-- 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.sub.-- NO decoder 902 is shown in FIG. 10. The PRFX.sub.-- NO decoder comprises four identical decoders 1002-1008 and a set of logic gates 1010. The four decoders 1002-1008 each look at one of the first four bytes (1010-1013) and determine if a prefix is present. Since it is possible for a prefix byte to follow an opcode byte, the logic gates 1010 are used to output a result representing the total number of prefixes before the first opeode byte, because prefixes following an opcode apply only to the next instruction's opcode.
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.sub.-- NO decoder 902 also includes a Prefix.sub.-- Present (PRFX.sub.-- P) output bus 1020 (which is also 4 bits wide). Four PRFX.sub.-- P output lines 1020-1023 indicate whether or not there is a prefix ire the given position, regardless of what the other positions output. The PRFX.sub.-- P outputs are tapped directly off the four decoder (1002-1008) outputs.
The PRFX.sub.-- NO decoder results (to be discussed in connection with FIG. 10) and the information from the PRFX.sub.-- DEC detectors 904-910 are combined by the PRFX.sub.-- SEL decoder 912. The prefix information is combined to form one 13 bit output bus 924 that indicates whether or not there are prefix signals and which prefixes are present.
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.times.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, Attorney Docket No. SP024; 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 ill 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.sub.-- 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.sub.-- 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.sub.-- 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-Ins ction 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 FIG. 11. Each bucket consists of two packets, plus general information pertaining to the entire bucket Packet #0 always contains three nano-instructions which are executed in-order: a LOAD instruction 1102, an ALU-type instruction 1104, and a STORE instruction 1106. Packet #1 consists of a single ALU-type instruction 1108.
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:
only one of the two basic instructions can reference a memory operand (there is only one load/store operation per bucket), and
both instructions must consist of a single ALU-type operation (as opposed to one instruction requiring two ALU-type operations).
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, 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, 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.sub.-- desc.sub.-- 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.TM. 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:
Decode one CISC instruction per half cycle;
Decode the 1st CISC instruction in a first phase;
Hold as valid the decoded results of the 1st (CISC instruction through the second phase;
Decode the 2nd CISC instruction in the second phase;
Combine the outputs of two instructions, if possible in the third phase; and
Output one bucket comprising four nano-instructions per cycle.
3.5 Instruction Decode Unit Block Diagrams
A block diagram of the IDU is shown in FIG. 12. Aligned instructions from the IAU arrive at the IDU on a bus 1201 which is 32 bits wide ([31:0] or 4 bytes). The aligned instructions are received by an Instruction Decoder 1202. The IDU 1202 only looks at the first four bytes of an aligned instruction in order to perform the CISC to RISC transformation.
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:
both only use A0, or
one only uses A0 and the other uses only A0, LD and ST.
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.sub.-- 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 FIGS. 13A-13E. The numbers along the left hand side and the top of the Karnaugh maps represent the opcode bits. For example, a one-byte opcode coded as hex 0F corresponds to the first row and 11th column, which is the "2 byte escape" instruction.
The instruction boxes that are shaded gray in the Karnaugh map of FIGS. 13A-13E represent basic instructions and the white boxes are those instructions which must be emulated.
A block diagram of the IDU's Instruction Decoder 1202 is shown in FIG. 14. The Instruction Decoder 1202 includes a plurality of decoders that are used to decode the CISC instructions and microcode routines.
A TYPE GENerator (TYPE.sub.-- GEN) decoder 1402 receives the first full aligned instructions on the ALIGN.sub.-- 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.sub.-- 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.sub.-- 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.sub.-- 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 a 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 FIG. 14 that include sections labeled 1, 2 and F decode assuming a 1 byte opcode, a 2 byte opcode and a floating point instruction, respectively. The invalid results are merely not selected. A multiplexer selects the output of the correct decoder.
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.sub.-- 0UNIT) decoder 1410, which comprises decoders F.sub.-- 0UNIT1, F.sub.-- 0UNIT2 and F.sub.-- 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.sub.-- 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.sub.-- GEN) decoder 1414 is therefore included to generate register indices based on the opcode of the CISC instruction. The CST.sub.-- 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 FIG. 15.
An additional two bit control signal, TempCount (TC), is input to the CST.sub.-- 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.sub.-- 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.sub.-- GEN) decoder, shown generally at block 1416, will now be discussed. The SEL.sub.-- GEN decoder includes a FlaG Need Modify (FG.sub.-- NM) decoder 1418. The FG.sub.-- NM decoder decodes for a one-byte opcode, a 2 byte opcode and a floating point instruction. In the i486.TM. 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.sub.-- 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.sub.-- 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 specie which register(s) are required by the operation.
The SEL.sub.-- GEN decoder 1416 further includes a FLD.sub.-- CNT decoder 1424 that indicates which of the register fields is required for the CISC instruction. The FLD.sub.-- CNT decoder specifies which of the 2 fields is the source register and which is the destination register.
A Nano-InstRuction GENerator (NIR.sub.-- GEN) decoder is shown generally as block 1426. The data size (DATA.sub.-- SZ) and address size (ADDR.sub.-- 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.sub.-- MODE control signal is also input to the NIR.sub.-- GEN decoder, but it is also used by the other decoders.
The ESCape DETect (ESC.sub.-- DET) input control signal is fed to the NIR.sub.-- GEN decoder to indicate whether the instruction has a 2 byte opcode. In addition, a SELect OPcode EXTension (SEL.sub.-- OP.sub.-- EXT) input control signal is used to generate loading of the mailbox registers when an emulation instruction is detected.
A Floating Point REGister (FP.sub.-- REG) input control signal passes the translated floating point register index to the IDU. The floating point format of the i486.TM., 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, registers being the next top register, etc. This register stack is emulated by using eight linear registers with fixed indices. When the input instruction specifies register0, 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.sub.-- DSIZE) and Address SIZE (EM.sub.-- ASIZE) of the instruction, as well as the Register index of the DESTination (EM.sub.-- RDEST), the source (EM.sub.-- RDEST2) and the Base InDeX information (EM.sub.-- 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.sub.-- ASIZE to determine the address size of the add instruction so that it knows what address size to emulate.
The NIR.sub.-- 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.sub.-- GEN decoder is called a LoaD INForm ation (LD.sub.-- INF) decoder 1430. The LD.sub.-- INF decoder decodes information corresponding to the Load and Store operations. The Load information is used for effective address calculations. The Load information fields (LDk.sub.-- INF1, LD.sub.-- INF2 and LD.sub.-- 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.TM. 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.sub.-- INF fields.
A Nano-InstRuction OPCode (NIR.sub.-- OPC) decoder 1432 transfers opcode for the A1 operation (packet1). The decoded fields (NIR.sub.-- OPC1, NIR.sub.-- OPC2 and NIR.sub.-- OPCF) comprise the first instruction byte (8 bits), plus three extension bits from the second byte.
A MIScellaneous OPCode (MISC.sub.-- 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.sub.-- CODE decoder 1436. The A0 opcode is usually copied directly from the i486.TM. 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.sub.-- 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.sub.-- CODE decoder 1440 extracts the 3 bit opcode extension from the ModR/M byte.
A IN.sub.-- 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.sub.-- JV.sub.-- 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.sub.-- 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.TM. 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:
RC1: ST(0)=Top of stack;
RC2: ST(1)=Second item on stack=next to the top of the stack; and
RC3: ST(i)=The i.sup.th item from the stack, where i is specified in the opcode.
In emulation mode, RC1, RC2 and RC3 are assigned as follows:
RC1: bits [4:0] of byte 3;
RC2: bits [1:0] of byte 2 and bits [7:5] of byte 3; and
RC3: bits [6:1] of byte 2.
FIG. 15 shows a representative block and logic gate diagram for the CST.sub.-- GEN, NIR.sub.-- GEN and SEL.sub.-- GEN decoders (1414, 1438 and 1424 respectively). It should be understood that this FIG. 15 is an example of how the 1 byte opcode, 2 byte opcode and floating point decoded results are selected, delayed, and combined to generate source and destination register indices for nano-instruction operations A0 and A1, and the destination register index for the Load instruction. The methodology of the selection, delay and multiplexing applies to all of the signals generated by the INSTRUCTION DECODER 1202, with the exception of those signals which do not generate separate 1 byte opcode, 2 byte opcode and floating point results. Furthermore, the results generated by this example are application specific, in other words, they apply to decoding of i486.TM. instructions into the nano-instruction format of the present invention. The principles discussed throughout these examples, however, are generally applicable to any CISC to RISC instruction alignment and decoding.
As discussed above, the CST.sub.-- 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.sub.-- 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, CSI2 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 1/2 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 1/2 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 1/2 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.sub.-- rd and a0.sub.-- rs for operation A0, which are shown generally at 1516, as well as the source and destination register indexes a1.sub.-- rd and a1.sub.-- rs for operation A1, which are shown generally at 1518. An index ld.sub.-- 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 FIG. 16A. The DFIFO holds four complete buckets, each of which contains four nano-instructions, two immediate data fields, and one displacement field. Each bucket corresponds to one level of pipeline register in the DFIFO. These buckets are generated in the IDU and pushed to the DFIFO during each cycle that the IEU requests a new bucket. The nano-instructions in a bucket are divided into two groups, called packet0 and packetk1. Packet0 can consist of a Load, ALU, and/or Store operation, which corresponds to one, two, or three nano-instruction. Packetk1 can only be an ALU operation, corresponding to one nano-instruction. As a result of this division, a bucket can only contain two ALU operations, and only one of them can reference memory. If subsequent instructions both require memory operands, they must be placed in separate buckets.
As can be seen from FIG. 16B, there is only a fair amount of general information associated with each packet and with the bucket as a whole. This information is stored in a general information FIFO. By default, the four nano-instructions in a bucket are executed in order, from NIR0 to NIR3. One of the bucket general information bits can be set to indicate that NIR3 should be executed before NIR0-NIR2. This feature makes it much easier to combine subsequent instructions into a single bucket, because their order no longer affects their ability to fit the bucket requirements. FIG. 16C shows an immediate data and displacement FIFO for buckets0-4. IMM0 represents the immediate data corresponding to packet0, and IMM1 represents the immediate data corresponding to packetk1. DISP represents the displacement corresponding to packet0. Packetk1 does not use DISP information because the DISP fields are only used as a part of address calculation.
A specific example of the three types of nano-instruction described above is shown in FIG. 17. The field descriptions and definitions are also described in Appendix A, pages 1-10. These tables provide detailed information about the contents of each bucket.
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.
Claims
What is claimed is:
1. A superscalar microprocessor for executing instructions obtained from an instruction store, said microprocessor comprising: a fetch circuit to fetch a plurality of CISC instructions from said instruction store, the plurality of CISC instructions being in program order; a decoder to decode said CISC instructions into RISC instructions having a predetermined sequence; and a dispatch circuit to concurrently dispatch more than one of said plurality of RISC instructions decoded by said decoder; and an execution unit comprising a plurality of functional units, each of said plurality of functional units executing one of said plurality of RISC instructions dispatched by said dispatch circuit out of the predetermined sequence, and a register file for storing data from said plurality of functional units in a plurality of registers, and wherein said register file communicates with said plurality of functional units via a plurality of data routing paths for concurrently providing data to more than one of said functional units and thereby enabling concurrent execution of more than one of said plurality instructions by said plurality of functional units, wherein first and second CISC instructions are decoded by said decoder into one or more first RISC instructions and one or more second RISC instructions, respectively, per clock cycle, wherein said execution unit further comprises first and second registers each comprising RISC instruction storage locations, wherein said first RISC instructions and said second RISC instructions are stored in said first register.
2. A superscalar microprocessor according to claim 1, wherein each one of said first and second registers comprises four RISC instruction storage locations.
3. A superscalar microprocessor, comprising: means for storing data in a plurality of registers identifiable by register references, said plurality of registers including a predetermined register and a temporary register; means for fetching CISC instructions to be executed, wherein at least one said CISC instructions includes a register reference; means for decoding said CISC instructions into RISC instructions having a predetermined sequence; and executing means for executing at least two of said RISC instructions concurrently and out of said predetermined sequence said executing means including means for selecting said temporary register where the execution of said instruction provides said register reference to select said predetermined register for the storage of data, wherein first and second CISC instructions are decoded by said decoding means into one or more first RISC instructions and one or more second RISC instructions, respectively, per clock cycle wherein said executing means comprises first and second registers each comprising RISC instruction storage locations, wherein said first RISC instructions and said second RISC instructions are stored in said first register.
4. A superscalar microprocessor according to claim 3, wherein each one of said first and second registers comprises four RISC instruction storage locations.
5. A superscalar microprocessor, comprising: memory to store data in a plurality of registers identifiable by register references, said plurality of registers including a predetermined register and a temporary register; fetch unit to fetch CISC instructions to be executed, wherein at least one of said CISC instructions includes a register reference; decoder to decode said CISC instructions into RISC instructions having a predetermined sequence; and execution unit to execute at least two of said RISC instructions concurrently and out of said predetermined sequence, said execution unit includes a selector to select said temporary register where the execution of said instruction provides said register reference to select said predetermined register for the storage of data, wherein first and second CISC instructions are decoded by said decoder into one or more first RISC instructions and one or more second RISC instructions, respectively, per clock cycle, wherein said execution unit comprises first and second registers each comprising RISC instruction storage locations, wherein said first RISC instructions and said second RISC instructions are stored in said first register.
Patent Citations (93)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US3346851 | 1967-10-01 | Thornton et al. | |
| US3771138 | 1973-11-01 | Celtruda et al. | |
| US3789365 | 1974-01-01 | Jen et al. | |
| US3916388 | 1975-10-01 | Shimp et al. | |
| US4034349 | 1977-07-01 | Monaco et al. | |
| US4084235 | 1978-04-01 | Hirtle | |
| US4189768 | 1980-02-01 | Liptay et al. | |
| US4189772 | 1980-02-01 | Liptay et al. | |
| US4200927 | 1980-04-01 | Highes et al. | |
| US4228495 | 1980-10-01 | Bernhard et al. | |
| US4236206 | 1980-11-01 | Strecker et al. | |
| US4296470 | 1981-10-01 | Fairchild et al. | |
| US4315314 | 1982-02-01 | Russo | |
| US4317170 | 1982-02-01 | Wada et al. | |
| US4410939 | 1983-10-01 | Kawakami | |
| US4434461 | 1984-02-01 | Puhl | |
| US4454578 | 1984-06-01 | Matsumoto et al. | |
| US4456955 | 1984-06-01 | Yanagita et al. | |
| US4459657 | 1984-07-01 | Murao | |
| US4476525 | 1984-10-01 | Ishii | |
| US4514803 | 1985-04-01 | Agnew et al. | |
| US4569016 | 1986-02-01 | Hao et al. | |
| US4587612 | 1986-05-01 | Fisk et al. | |
| US4629989 | 1986-12-01 | Toril | |
| US4654781 | 1987-03-01 | Schwartz et al. | |
| US4675806 | 1987-06-01 | Uchida | |
| US4714994 | 1987-12-01 | Oklobzija et al. | |
| US4722049 | 1988-01-01 | Lahti | |
| US4739471 | 1988-04-01 | Baum et al. | |
| US4752873 | 1988-06-01 | Shonai et al. | |
| US4758948 | 1988-07-01 | May et al. | |
| US4766566 | 1988-08-01 | Chuang | |
| US4771376 | 1988-09-01 | Kamiya | |
| US4807115 | 1989-02-01 | Torng | |
| US4814976 | 1989-03-01 | Hansen et al. | |
| US4858105 | 1989-08-01 | Kuriyama et al. | |
| US4897810 | 1990-01-01 | Nix | |
| US4901228 | 1990-02-01 | Kodama | |
| US4903196 | 1990-02-01 | Pomerene et al. | |
| US4924376 | 1990-05-01 | Ooi | |
| US4926323 | 1990-05-01 | Baror et al. | |
| US4942525 | 1990-07-01 | Shintani et al. | |
| US4985825 | 1991-01-01 | Webb, Jr. et al. | |
| US4992931 | 1991-02-01 | Hirasawa | |
| US4992934 | 1991-02-01 | Portanova et al. | |
| US4992938 | 1991-02-01 | Cocke et al. | |
| US5003462 | 1991-03-01 | Blaner et al. | |
| US5101341 | 1992-03-01 | Circello et al. | |
| US5127091 | 1992-06-01 | Boufarah et al. | |
| US5133072 | 1992-07-01 | Buzbee | |
| US5148528 | 1992-09-01 | Fite et al. | |
| US5168571 | 1992-12-01 | Hoover et al. | |
| US5193206 | 1993-03-01 | Mills | |
| US5226126 | 1993-07-01 | McFarland et al. | |
| US5226170 | 1993-07-01 | Rubinfeld | |
| US5230068 | 1993-07-01 | Van Dyke et al. | |
| US5335460 | 1994-08-01 | Eickemeyer et al. | |
| US5390355 | 1995-02-01 | Horst | |
| US5442757 | 1995-08-01 | McFarland et al. | |
| US5487156 | 1996-01-01 | Popescu et al. | |
| US5539911 | 1996-07-01 | Nguyen et al. | |
| US5546552 | 1996-08-01 | Coon et al. | |
| US5561776 | 1996-10-01 | Popescu et al. | |
| US5574927 | 1996-11-01 | Scantlin | |
| US5592636 | 1997-01-01 | Popescu et al. | |
| US5625837 | 1997-04-01 | Popescu et al. | |
| US5627983 | 1997-05-01 | Popescu et al. | |
| US5651125 | 1997-07-01 | Witt et al. | |
| US5689720 | 1997-11-01 | Nguyen et al. | |
| US5708841 | 1998-01-01 | Popescu et al. | |
| US5768575 | 1998-06-01 | McFarland et al. | |
| US5778210 | 1998-07-01 | Henstrom et al. | |
| US5797025 | 1998-08-01 | Popescu et al. | |
| US5832205 | 1998-11-01 | Kelly et al. | |
| US5832293 | 1998-11-01 | Popescu et al. | |
| EPX136179A2 | 1985-04-01 | ||
| EPX170398 | 1986-02-01 | ||
| EPX171595A2 | 1986-02-01 | ||
| EPX272198 | 1988-06-01 | ||
| EPX354585A2 | 1990-02-01 | ||
| EPX368332A2 | 1990-05-01 | ||
| EPX372751A2 | 1990-06-01 | ||
| EPX377991A3 | 1990-07-01 | ||
| EPX380854 | 1990-08-01 | ||
| EPX402856A2 | 1990-12-01 | ||
| EPX419105A2 | 1991-03-01 | ||
| EPX426393A2 | 1991-05-01 | ||
| EPX461257 | 1991-12-01 | ||
| EPX473420 | 1992-03-01 | ||
| EPX479390A2 | 1992-04-01 | ||
| FRX2575564A1 | 1986-04-01 | ||
| GBX2011682 | 1979-07-01 | ||
| GBX2230116 | 1990-10-01 |
Non-Patent Literature (57)
- Faix et al., "Combined Macro/Micro Program Machine", IBM Technical Disclosure Bulletin, vol. 14, No. 1, p. 298, Jun. 1971.
- Patent Abstracts of Japan, vol. 009, No. 006 (p-326), Jan. 11, 1985.
- S. Ohr, "Superlong Instructions Help Supermini Refine Number Crunching", Electronic Design, vol. 33, No. 26, pp. 41-43, Nov. 1985.
- McNeley et al., "Emulating a Complex Instruction Set Computer with a Reduced Instruction Set Computer", IEEE Micro, vol. 7, No. 1, pp. 60-71, Feb. 1987.
- J. Wharton, "Parallel 486 Pipelines Produce Peak Processor Performance," Microprocessor Report, v3 n6 p. 13(5), Jun. 1989.
- S. Garth, "Combining RISC and CISC in PC Systems," IEEE, pp. 10/1-10/5.
- i486.TM. Microprocessor Programmer's Reference Manual, Order No. 240486, Intel Corporation, Santa Clara, California, 1990.
- J. Hennessy et al., Computer Architecture --A Quantitative Approach, Morgan Kaufman Publishers, Inc. San Mateo, California, 1990.
- i486.TM. Microprocessor Hardware Reference ManualOrder No. 240552, Intel Corporation, Santa Clara, California, 1990.
- J. Crawford, "The Execution Pipelines of the Intel i486.TM. CPU," IEEE, pp. 254-258, Feb. 26 --Mar. 2, 1990.
- M. Johnson, Superscalar Microprocessor Design, Prentice-Hall, Inc., Englewood Cliffs, New Jersey, 1991.
- S. Heath, "Architecture Considerations for a M88000 Superscalar RISC Processor," IEEE, London, England, pp. 1-6, Nov. 4, 1991.
- Patent Abstracts of Japan, vol. 017, No. 082 (P-1489), Feb. 18, 1993.
- Patent Abstracts of Japan vol. 017, No. 084 (P-1490), Feb. 19, 1993.
- "High Performance Dual Architecture Procesor," IBM Technical Disclosure Bulletin, vol. 36, No. 2, Feb. 1993, pp. 231-234.
- Bakoglu, H.B. et al.,. "IBM Second-Generation RISC Machine Organization," 1990 IEEE, pp. 173-178.
- McGeady, S., "The i960CA SuperScalar Implementation of the 80960 Architecture," 1990 IEEE, pp. 232-240.
- Nakano, H., et al., "A 80 Mflops 64-Bit Microprocessor For Parallel Computer," 1991 IEEE pp. 15.2.1--15.2.4.
- Andrews, W., "Distinctions Blur Between DSP Solutions," Computer Design, vol. 28, No. 9, May 1, 1989, 11 pages.
- Coldwell, R.P. et al., "A VLIW Architecture for a Trace Scheduling Compiler," Proceedings 2.sup.nd Interantional Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS II), Oct. 5-8, 1987, California, pp. 180-192.
- Lee, Ruby B., "Precision Architecture," Computer, Jan. 1989, pp. 78-91.
- IBM Technical Disclosure Bulletin, "Overlap of Store Multiple Operation With Succeeding Operations Through Second Set of General Purpose Registers", vol. 28, No. 6, Nov. 1985, New York, NY, pp. 2727-2728.
- Jelemensky, J. et al., "New Microcomputer Features Advanced Hardware For Real-Time Control Applications", Electro Conference Record, vol. 14, Apr. 11, 1989, Los Angeles, CA, pp. 1,3,5,7 and 9.
- Odnert, D. et al., "Architecture and Compiler Enhancements for PA-RISC Workstations," Proceedings from IEEE Compcon, Feb. 25 --Mar. 1, 1991, San Francisco, CA, pp. 214-218.
- Melear, C., Motorola, Inc., "The Design of the 88000 RISC Family," IEEE Micro, vol. 9, No. 2, Apr. 1989, Los Alamitos, CA, pp. 26-38.
- Abstract for Publication No. JP60225943, Published Nov. 11, 1985, from Patents Abstracts of Japan, Abstract Vol. No. 010089, p. 1.
- Popescu, V. et al., "The Metaflow Architecture," IEEE Micro, Jun. 1991, pp. 10-13 and 63-73.
- Smith, J.E. et al., "Implementing Precise Interrupts in Pipelined Processors," IEEE Transactions On Computers, vol. 37, No. 5, May 1988, New York, NY, pp. 562-573.
- Weiss, R., "Third-generation of R.I.S. C. processors", EDN, Mar. 30, 1992, pp. 96-108.
- 8-Bit Embedded Controller Handbook, Intel, 1989, pp. 1-1 --1-19 and 3-1 --3-37.
- Smith et al., "Implementation of Precise Interrupts in Pipelined Processors," Proceedings of the 12th Annual International Symposium on Computer Architecture, Jun. 1985, pp. 36-44.
- Wedig, R.G., Detection of Concurrency In Directly Executed Language Instructions Streams, (Dissertation), Jun. 1982, pp. 1-179.
- Agerwala et al., "High-Performance Reduced Instruction Set Processors," IBM Research Division, Mar. 31, 1987, pp. 1-61.
- Gross et al., "Optimizing Delayed Branches,"Proceedings of the 5th Annual Workshop on Microprogramming, Oct. 5-7, 1982, pp. 114-120.
- Tjaden et al., "Representation of Concurrency with Ordering Matrices," IEEE Trans. On Computers, vol. C-22, No. 8, Aug. 1973, pp. 752-761.
- Tjaden, Representation and Detection of Concurrency Using Ordering Matrices, (Dissertation), 1972, pp. 1-199.
- Foster et al., "Percolation of Code to Enhance Parallel Dispatching and Execution," IEEE Trans. On Computers, Dec. 1971, pp. 1412-1415.
- Thornton, J.E., Design of a Computer: The Control Data 6600, Control Data Corporation, 1970, pp. 58-140.
- Weiss et al., "Instruction Issue Logic in Pipelined Supercomputers," Reprinted from IEEE Trans. on Computers, vol. C-33, No. 11, Nov. 1984, pp. 1013-1022.
- Tomasulo, R.M., "An Efficient Algorium for Exploiting Multiple Arithmetic Units," IBM Journal, vol. 11, Jan. 1967, pp. 25-33.
- Tjaden, et al. "Detection and Parallel Execution of Independent Instructions," IEEE Trans. On Computers, vol. C-19, No. 10, Oct. 1970, pp. 889-895.
- Smith et al., "Limits on Multiple Instruction Issue," Proceedings of the 3rd International Conference on Architectural Support for Programming Languages and Operating Systems, Apr. 1989, pp. 290-302.
- Pleszkun et al., "The Performance Potential of Multiple Functional Unit Processors," Proceedings of the 15th Annual Symposium on Computer Architecture, Jun. 1988, pp. 37-44.
- Pleszkun et al., "WISQ: A Restartable Architecture Using Queues," Proceedings of the 14th International Symposium on Computer Architecture, Jun. 1987, pp. 290-299.
- Patt et al., "Critical Issues Regarding HPS, A High Performance Microarchitecture," Proceedings of the 18th Annual Workshop on Microprogramming, Dec. 1985, pp. 109-116.
- Hwu et al., "Checkpoint Repair for High-Performance Out-of-Order Execution Machines," IEEE Trans. On Computers, vol. C-36, No. 12, Dec. 1987, pp. 1496-1514.
- Patt et al., "HPS, A New Micoarchitecture: Rationale and Introduction," Proceedings of the 18th Annual Workshop on Microprogramming, Dec. 1985, pp. 103-108.
- Keller, R.M., "Look-Ahead Processors,"Computing Surveys, vol. 7, No. 4, Dec. 1975, pp. 177-195.
- Jouppi et al., "Available Instruction-Level Parallism for Superscalar and Superpipelined Machines," Proceedings of the 3rd International Conference on Architectual Support for Programming Languages and Operating Sytems, Apr. 1989, pp. 272-282.
- Hwu et al., "HPSm, a High Performance Restricted Data Flow Architecture Having Minimal Functionality," Proceedings from ISCA-13, Tokyo, Japan, Jun. 2-5, 1986, pp. 297-306.
- Hwu et al., "Exploiting Parallel Microprocessor Microarchitectures with a Compiler Code Generator," Proceedings of the 15th Annual Symposium on Computer Architecture, Jun. 1988, pp. 45-53.
- Coldwell, et al., "A VLIW Architecture for a Trace Scheduling Compiler," Proceedings of the 2nd International Conference on Architectural Support for Programming Languages and Operating Systems, Oct. 1987, pp. 180-192.
- Uht, A.K., "An Efficient Hardware Algorithm to Extract Concurrency From General-Purpose Code," Proceedings of the 19th Annual Hawaii International Conference on System Sciences, 1986, pp. 41-50.
- Charlesworth, A.E., "An Approach to Scientific Array Procesing: The Architectual Design of the AP-120B/FPS-164 Family," Computer, vol. 14, Sep. 1981, pp. 18-27.
- Acosta, Ramon D. et al., "An Instruction Issuing Approach to Enhancing Performance in Multiple Functional Unit Processors," IEEE Transactions On Computers, vol. C-35, No. 9, Sep. 1986, pp. 815-825.
- Johnson, William M., Super-Scalar Processor Design, (Dissertation), Copyright 1989, 134 pages.
- Sohi, Gurindar S. and Sriram Vajapeyam, "Instruction Issue Logic For High-Performance, Interruptable Pipelined Processors," Conference Proceedings of the 14th Annual International Symposium on Computer Architecture, Jun. 2-5, 1987, pp. 27-34.