US 5,453,927 AGrant
Data Processor for Processing Branch Instructions
Issue Date:1995-09-26
•2 Claims
•61 Drawing Sheets
Abstract
A data processor capable of efficiently processing jump instructions including an instruction fetch unit for fetching instructions from a memory, an instruction decoding unit for decoding instructions fetched by the instruction fetch unit, an instruction executing unit for executing instructions according to the decoding result of the instructions, a PC calculation unit for holding a head address of an instruction being decoded at the instruction decoding unit, adders for selectively inputting respective values of branch displacement fields transferred from the instruction fetch unit responsive to a part of a value of an instruction code field transferred from the instruction fetch unit so as to add it to the head address of the instruction transferred from the PC calculation unit, and a JA bus for transferring the addition results to the instruction fetch unit.
Metadata
Assignee
- Mitsubishi Denki Kabushiki Kaisha
Inventor
- Masahito Matsuo
Application Information
Application Number:US 2239094
Filing Date:1994-04-06
Priority Date:1993-04-13
Art Unit:235
Classifications
IPC:
G06T 942
Field of Search:
395375;800
Patent Drawings (61 sheets)
Description
Background of the Invention
1. Field of the Invention
The present, invention relates to a data processor, specifically, to a data processor including a pipeline processing mechanism which processes a jump instruction rapidly, and more particularly, to a data processor capable of reducing overheads of pipeline processing of the case where the jump instruct ion is executed by performing jump processing in the initial pipeline stage.
2. Description of the Related Art
In a conventional data processor, by dividing the processing into a plural number of steps with a flow of data processing, and processing the steps of different instructions simultaneously in respective corresponding stages of the pipeline, a mean processing time necessary for one instruction is shortened and a processing performance is improved as a whole.
However, in case of executing an instruction which disturbs an instruction processing sequence such as a jump instruction, since an instruction processing sequence is switched at executing stages of the instruction, an overhead of the pipeline processing increases and a pipeline processing can not be performed efficiently. Besides, a frequency of appearance of the jump instruction in executing practical programs is very high, thus an increase in processing speed of the jump instruction is one of the most important items to improve the performance of the data processor.
In order to improve performance of a data processor, various kinds of devices have been made about reducing overhead in performing instructions such as unconditional branch instructions, conditional branch instructions and the like.
For example, FIG. 1 is a block diagram showing a configuration for performing jump processing in the stage of instruction decoding of the invention of Japanese Patent Application No. 4-22695 as an example of a conventional data processor.
In FIG. 1, reference numeral 401 designates a branch target address calculation unit, 402 a PC calculation unit, 403 an instruction decoding unit, 404 an instruction fetch unit, respectively.
These blocks are connect with each other by an II bus 433 for transferring instruction codes, a DISP bus 434 for transferring a branch displacement or an immediate cut out according to the instruction decoding result, an ILEN bus 435 for transferring a length of an instruction having been decoded, a PI bus 431 for transferring a PC value of an instruction being decoded, a JA bus 436 for transferring a jump target address, and the like, and data are given or received between them.
By adopting a configuration shown in FIG. 1, when a jump target address is designated by PC relative address or an absolute address, the jump target address is generated in parallel with the instruction decoding processing so that the jump target instruction can be fetched immediately after the instruction decoding.
The instruction fetch unit 404 is provided with an address translation mechanism of an instruction address, a built-in instruction cache, an instruction MMU (Memory Management Unit)/cache unit including an instruction TLB (Translation Lookaside Buffer) and its control unit, and with an instruction queue unit. The instruction fetch unit 404 fetches an instruction code from the built-in instruction cache or an external memory by address conversion of an instruction fetch address, and sends the instruction code to the instruction decoding unit 403 and the branch target address calculation unit 401 through the II bus 433.
The instruction decoding unit 403 decodes instruction codes taken from the II bus 433 in 16-bit (half word) unit. The instruction decoding unit 403 decodes instruction codes outputted from the instruction fetch unit 404 by zero to eight bytes per one clock.
The information as to a PC calculation is outputted to the PC calculation unit 402, a signal for controlling a branching when a branch instruction is decoded to the branch target address calculation unit 401 or to the instruction fetch unit 404, and an output pointer update information of the instruction queue to the instruction fetch unit 404, respectively.
The PC calculation unit 402 is controlled by hardwired and calculates the PC value of an instruction by the information as to the PC calculation outputted from the instruction decoding unit 403. The instruction to be processed is the variable length instruction, and the instruction length does not become clear until the instruction is decoded. The PC calculation unit 402 calculates the PC value of the next instruction by adding the instruction length outputted from the instruction decoding unit 403 to the PC value of the instruction being decoded.
At the PC calculation unit 402, a DPC 79 holds a head address of an instruction being decoded. And a TPC 77 holds a head address of a code to be decoded including also a case where an instruction code of one instruction is divided into plural processing units and processed. That is, a head address of the code taken into the instruction decoding unit 403 from the II bus 433 is hold.
A PC adder 73 adds an instruction length taken from the instruction decoding unit 403 through the ILEN bus 435 and a value of the TPC 77 to each other, and the addition result is rewritten in the TPC 77 and an ATPC 78. When decoding of one instruction is finished, since the addition result shows a head address of the next instruction, it is also rewritten in the DPC 79. In addition, the ATPC 78 is referred to as a PC value of the next instruction at the time of calculating an address.
When a jump is generated, the jump target address is taken in from the JA bus 436, and the TPC 77 and the DPC 79 ape initialized, but the ATPC 78 is not initialized even when a jump is generated.
By such a control, a PC value of an instruction being decoded is always held in the DPC 79.
The branch target address calculation unit 401 calculates a branch target address and cuts out an absolute address field in an instruction decoding stage. Data of a position of a branch displacement is cutout from an instruction code on the II bus 433 which is sent to the instruction decoding unit 403 from the instruction fetch unit 404, and it is added to a value of the DPC 79 being a PC value of an instruction to be decoded, thereby a branch target address is calculated.
For high frequency used branch instructions, four adders, that is, first to fourth adders A1, A2, A3, A4 are provided so that the jump target address can be calculated at the same time of an instruction decoding.
In addition, reference numerals 409 to 416 designate input latches of the respective adders A1, A2, A3, A4, and numerals 417 to 420 designate output latches of the respective adders A1, A2, A3, A4.
The branch target address calculation unit 401 operates in synchronism with the decoding process of an instruction code. In synchronism with the decoding cycles of respective instructions, a value of the DPC 79 is taken in by the input latches 413, 414, 415, 416 through the PI bus 431. Data on the II bus 433 is sign-extended by the four input latches 409, 410, 411, 412 corresponding to the four positions of branch displacement fields and taken therein.
The branch displacement fields taken in the input latches 411, 412 correspond to absolute address fields in the case where jump target addresses of jump instructions are designated by absolute addresses. In the latch 421, the sign-extended data taken in the input latch 411 is held. And in the latch 422, the data taken in the input latch 412 is held intact.
In the case where a branch instruction by which a branch target address can be calculated or a jump instruction by which a jump target address is designated by an absolute address, is decoded in parallel with decoding an instruction at the first to fourth adders A1, A2, A3, A4, a jump processing or a prefetch of a branch target instruction is performed in that cycle.
When branch target addresses cannot be calculated at the first to fourth adders A1, A2, A3, A4 during a branch instruction decoding cycle, a branch target address is calculated in the next cycle, and a branch processing or prefetching of a branch target instruction is performed. In this case, after decoding a branch instruction, the input latch 412 takes in a branch displacement cut out according to the instruction decoding result through the DISP bus 43 from the instruction decoding unit 403. The input latch 416 takes in a value of the DPC 79 through the PI bus 431. The fourth adder A4 adds the contents of the input latches 412 and 416 to each other and outputs the result to the output latch 420.
When a jump processing is performed in an instruction decoding stage, one of the output latches 417 to 422 is selected by the control signal inputted from the instruction decoding unit 403 and its content is outputted to the JA bus 436, and a jump target address is transferred to the instruction fetch unit 404. The instruction fetch unit 404 starts to fetch a jump target instruction according to the address taken in from the JA bus 436. And a value on the JA bus 436 is taken also in the TPC 77 and the DPC 79 and the PC calculation unit 402 is initialized.
In such a way, the conventional data processor is capable of fetching a branch target instruction immediately after the instruction decoding, by calculating a branch target address of a branch instruction by which a branch target address is designated by PC relative or by cutting out an absolute address field of a jump instruction by which a jump target address is designated by an absolute address in parallel with decoding an instruction. Accordingly, by providing an aforementioned jump processing mechanism, a data processor of high performance can be obtained since a wasteful time until decoding a branch target instruction after decoding a branch instruction can be reduced to a time required only for fetching a branch target instruction.
By the way, in the aforementioned conventional data processor, the same number of adders as the number of branch displacement fields of branch instructions by which branch target addresses are calculated in parallel with decoding instructions are required, and the problem was that a quantity of hardware is increased in proportion to the increase of the number of instructions being an object to be processed at high speed.
And there was also a problem as to an instruction by which a jump target address is designated by an absolute address, that a quantity of hardware increases since latches are necessary by the number of fields which are cut out.
Summary of the Invention
The present invention has been devised in view of such circumstances, and a main object thereof is to provide a data processor of high performance capable of pipeline processing a jump instruction efficiently by smaller quantity of hardware.
Another object of the invention is to provide a data processor capable of generating a jump target address in parallel with decoding an instruction by adding smaller quantity of hardware compared with that to the conventional data processor.
A data processor of the first invention of the present invention which processes a first branch instruction including an operation code field and a first branch displacement field and branching to an address obtained by adding a value of the first branch displacement field to a head address thereof, and a second branch instruction including an operation code field and a second branch displacement field and branching to an address obtained by adding a value of the second branch displacement field to a head address thereof, comprising: a memory which stores instructions including the first and second branch instructions; an instruction fetch unit which fetches instructions from the memory; an instruction decoding unit which decodes instructions fetched by the instruction fetch unit; an instruction executing unit which executes instructions on the basis of the decoding result of instructions by the instruction decoding unit; a program counter unit which holds a head address of an instruction being decoded at the instruction decoding unit; and a jump target address generating unit, including: adding means for calculating an address by selectively inputting either a value of the first branch displacement field or a value of the second branch displacement field from the instruction fetch unit according to a value of a part of the field of the operation code field transferred from the instruction fetch unit, by inputting a head address of the instruction from the program counter unit, and executing the addition of the two in parallel with decoding processing of the instruction by the instruction decoding unit, regarding that a new instruction is the first branch instruction or the second branch instruction every time a new instruction is decoded; and transferring means for transferring the address calculated by the adding means to the instruction fetch unit.
A data processor of the second invention of the present invention which processes a first branch instruction including an operation code field and a first branch displacement field and branching to an address obtained by adding a value of the first branch displacement field to a head address thereof, a second branch instruction including an operation code field and a second branch displacement field and branching to an address obtained by adding a value of the second branch displacement field to a head address thereof, and a third branch instruction including an operation code field and a third branch displacement field and branching to an address obtained by adding a value of the third branch displacement field to a head address thereof, comprising: a memory which stores instructions including the first, second and third branch instructions; an instruction fetch unit which fetches instructions from the memory; an instruction decoding unit which decodes instructions fetched by the instruction fetch unit; an instruction executing unit which executes instructions on the basis of the decoding result of instructions by the instruction decoding unit; a program counter unit which holds a head address of an instruction being decoded at the instruction decoding unit; and a jump target address generating unit, including: first adding means for calculating a first address by selectively inputting either a value of the first branch displacement field or a value of the second branch displacement field from the instruction fetch unit according to a value of a part of the field of the operation code field transferred from the instruction fetch unit, by inputting a head address of the instruction from the program counter unit, and executing the addition of the two in parallel with decoding processing of the instruction by the instruction decoding unit, regarding that a new instruction is the first branch instruction or the second branch instruction every time a new instruction is decoded; second adding means for calculating a second address by inputting a value of the third branch displacement field from the instruction fetch unit, by inputting a head address of the instruction from the program counter unit, and executing the addition of the two in parallel with decoding processing of the instruction by the instruction decoding unit, regarding that a new instruction is the third branch instruction every time a new instruction is decoded; and transferring means for selectively transferring one of address calculated by-the first adding means and the second adding means to the instruction fetch unit.
A data processor of the third invention of the present invention which processes a first jump instruction designating a jump target by a first absolute address field and a second jump instruction designating a jump target by a second absolute address field, comprising: a memory which stores instructions including the first and second jump instructions; an instruction fetch unit for fetching instructions from the memory; an instruction decoding unit which decodes instructions fetched by the instruction fetch unit; an instruction executing unit which executes instructions on the basis of the decoding result of instructions by the instruction decoding unit; arid a jump target address generating unit, including: field cutting out means for generating an address by executing a processing of selectively cutting out either a value of the first absolute address field or a value of the second absolute address field from an instruction data transferred from the instruction Fetch unit according to a value of a part of the field of the operation code field transferred From the instruction fetch unit in parallel with decoding the instruction by the instruction decoding unit, regarding that a new instruction is the first jump instruction or the second jump instruction every time a new instruction is decoded; and transferring means for transferring the address cut out by the Field cutting out means to the instruction fetch unit.
In the data processor related to the first invention of the present invention, when an instruction being decoded at the instruction decoding unit is judged to be the first branch instruction or the second branch instruction, since the jump target address generating unit transfers the address to the instruction fetch unit through the transferring means and the instruction fetch unit fetches an instruction of the address transferred from the jump target address generating unit, at the time when a branch instruction is processed, a branch target instruction is fetched from the memory immediately after the branch instruction is decoded.
In the data processor related to the second invention of the present invention, when an instruction being decode at the instruction decoding unit is judged to be the first branch instruction or the second branch instruction, since the jump target address generating unit transfers a first address to the instruction fetch unit through the transferring means and the instruction fetch unit fetches the instruction of the first address transferred from the jump target address generating unit, and when an instruction being decoded at the instruction decoding unit is judged to be the third branch instruction, the jump target address generating unit transfers the second address to the instruction fetch unit through the transferring means and the instruction fetch unit fetches the second address transferred form the jump target address generating unit, at the time when a branch instruction is processed, a branch target instruction is fetched from the memory immediately after the branch instruction is decoded.
In the data processor related to the third invention of the present invention, when an instruction being decoded at the instruction decoding unit is judged to be the first jump instruction or the second jump instruction, since the jump target address generating means transfers an address to the instruction fetch unit through the transferring means and the instruction fetch unit fetches the instruction of the address transferred from the jump target address generation unit, at the time when a jump instruction is processed, a jump target instruction is fetched from the memory immediately after the jump instruction is decoded.
The above and further objects and features of the invention will more fully be apparent from the following detailed description with accompanying drawings.
Brief Description of the Drawings
FIG. 1 is a block diagram showing a configuration of a jump instruction processing mechanism of a conventional data processor which calculates a branch target address at the instruction decoding stage so as to perform jump processing,
FIG. 2 is a block diagram showing a configuration example of a system using an embodiment example of a data processor of the present invention,
FIG. 3 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 4 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 5 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 6 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 7 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 8 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 9 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 10 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 11 is a schematic diagram of an instruction format according to an example of one embodiment of a data processor of the present invention,
FIG. 12 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 13 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 14 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 15 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 16 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 17 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 18 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present, invention,
FIG. 19 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 20 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 21 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 22 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 23 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 24 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 25 is a schematic diagram showing a format of an addressing mode according to an example of one embodiment of a data processor of the present invention,
FIG. 26 is a schematic diagram showing bit pattern of BRA instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 27 is a schematic diagram showing bit pattern of BRA instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 28 is a schematic diagram showing bit pattern of Bcc instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 29 is a schematic diagram showing bit pattern of Bcc instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 30 is a schematic diagram showing bit pattern of BSR instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 31 is a schematic diagram showing bit pattern of BSR instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 32 is a schematic diagram showing bit pattern of JMP instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 33 is a schematic diagram showing bit pattern of JSR instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 34 is a schematic: diagram showing bit pattern of ACB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 35 is a schematic diagram showing bit pattern of ACB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 36 is a schematic diagram showing bit pattern of ACB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 37 is a schematic diagram showing bit pattern of ACB instruction according to an example of one embodiment f a data processor of the present invention, FIG. 38 is a schematic diagram showing bit pattern of SCB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 39 is a schematic diagram showing bit pattern of SCB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 40 is a schematic diagram showing bit pattern of SCB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 41 is a schematic diagram showing bit pattern of SCB instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 42 is a schematic diagram showing bit pattern of FBcc instruction according to an example of one embodiment of a data processor of the present invention,
FIG. 43 is a block diagram showing a configuration of an example of one embodiment of a data processor of the present invention,
FIG. 44 is a block diagram showing a configuration of an instruction fetch unit based upon an example of one embodiment of a data processor of the present invention,
FIG. 45 is a block diagram showing a configuration of an operand address calculation unit based upon an example of one embodiment of a data processor of the present invention,
FIG. 46 is a block diagram showing a configuration of a PC processing unit based upon an example of one embodiment of a data processor of the present invention,
FIG. 47 is a block diagram showing a configuration of a PC processing unit based upon an example of one embodiment of a data processor of the present invention,
FIG. 48 is a schematic diagram showing a bit position on an II bus which is inputted as a branch displacement to each adder of a jump target address generating unit based upon an example of one embodiment of a data processor of the present invention,
FIG. 49 is a schematic diagram showing a pipeline configuration based upon an example of one embodiment of a data processor of the present invention,
FIG. 50 is a block diagram showing a relation of delivering and receiving of control signals relating to jump processing and jump target address between each block based upon an example of one embodiment of a data processor of the present invention,
FIG. 51 is a schematic diagram showing branch displacement positions on an II bus at processing of BRA, BSR and Bcc instruction based upon an example of one embodiment of a data processor of the present invention,
FIG. 52 is a timing chart at processing of BRA instruction based upon an example of one embodiment of a data processor of the present invention,
FIG. 53 is a timing chart at processing of BSR instruction based upon an example of one embodiment of a data processor of the present invention,
FIG. 54A is a timing chart of processing of Bcc instruction when it is predicted not to branch based upon an example of one embodiment of a data processor of the present invention,
FIG. 54B is a timing chart of processing of Bcc instruction when it is predicted to branch based upon an example of one embodiment of a data processor of the present invention,
FIG. 55 is a schematic diagram showing branch displacement positions on an II bus at processing of JMP and JSR instructions based upon an example of one embodiment of a data processor of the present invention,
FIG. 56 is a timing chart at processing of JMP instruct invention, based upon an example of one embodiment of a data processor of the present invention,
FIG. 57 is a timing chart at processing of JSR instruction based upon an example of one embodiment of a data processor of the present invention,
FIG. 58 is a schematic diagram showing branch displacement positions on an II bus at processing of ACB and SCB instructions based upon an example of one embodiment of data processor of the present invention,
FIG. 59A is a timing chart of processing of ACB instruct invention, when displacement is 8 bits based upon an example of one embodiment of a data processor of the present invention,
FIG. 59B is a timing chart of processing of ACB instruction when displacement is 16 or 32 bits based upon an example of one embodiment of a data processor of the present invention,
FIG. 60 is a schematic diagram showing branch displacement positions on an II bus at processing of FBcc instruction based upon an example of one embodiment of a data processor of the present invention, and
FIG. 61 is a timing chart at processing of FBcc instruction based upon an example of one embodiment of a data processor of the present invention.
Description of the Preferred Embodiments
In the following, the present invention will be particularly described with reference to the drawings showing its embodiments.
(1) "Configuration of a System using a Data Processor of the Present Invention"
FIG. 2 is a block diagram showing an example of a configuration of a system using a data processor 100 of the present invention.
In this embodiment, a data processor 100 of the present invention, an instruction cache 106, data caches 107 and 108, and a main memory 109 are connected by an address bus 101, a data bus 102, an instruction bus 103, a memory address bus 104, and a memory data bus 105.
The address bus 101 inputs the address outputted from the data processor 100 of the present invention into the instruction cache 106 and the data caches 107, 108. The instruction bus 103 transfers the instruction code outputted from the instruction cache 106 to the data processor 100 of the present invention. The data bus 102 transfers data outputted from the data processor 100 of the present invention to the data caches 107, 108, or data outputted from the data caches 107, 108 to the data processor 100 of the present invention. The memory address bus 104 transfers the address outputted from the instruction cache 106 or the data caches 107, 108 to the main memory 109. The memory data bus 105 transfers the instruction or data between the main memory 109 and the instruction cache 106 or the data caches 107, 108.
When the instruction cache 106 or the data caches 107, 108 misses, the caches 106, 107 or 108 arbitrate the memory address bus 104 and the memory data bus 105 so as to access the main memory 109.
In the data caches 107, 108, on the side of the data processor 100 of the present invention, two data caches 107, 108 are operated in cooperation with each other for they associate with a 64-bit bus. For each of the 64-bit data, the data cache 107 treats the high-order 32-bit data and the data cache 108 treats the low-order 32-bit data.
In the following, first, instruction system and a processing mechanism of the data processor 100 of the present invention will be explained, and next, processing method of jump instruction will be explained.
(2) "Allocation of Instruction Bits of a Data Processor of the Present Invention"
(2.1) "Instruction Format"
An instruction of a data processor of the present invention is formed on a 16-bit unit basis, being variable in length. No instruction with an odd number of bytes is used herein.
The data processor of the present invention has an instruction format specifically devised for the purpose of highly frequent instructions in a short format. For example, as to a two-operand instruction, two formats are provided; a general format which has basically a configuration of "four bytes+extension part(s)" and allows the utilization of all addressing modes and a short format allowing only the use of a highly frequent instruction and an addressing mode.
The meanings of symbols appearing in the instruction formal of the data processor of the present invention shown in FIG. 3 to FIG. 11 are as follows:
-: Field for operation code
#: Field in which literal or immediate enters
Ea: Field for specifying an operand in a general type 8-bit addressing mode
Sh: Field for specifying an operand in a short type 6-bit addressing mode
Rn: Field for specifying an operand on a register file by the register number
*: Field in which anything may enter
In the format, as shown in FIG. 3, the right side 602 is the LSB side and is higher address. The instruction format can be discriminated only after an address N 604 and an address N+1 606 are checked. This is because of a premise that the instruction is fetched and decoded on a 16-bit (half word) unit basis without fail.
In the data processor of the present invention, in the case with any format, the extension part of Ea or Sh of each operand is sure to be positioned immediately after 16 bits (half word) comprising the basic part of that Ea or Sh. This takes precedence over the immediate data specified implicitly by the instruction or the extension part of the instruction. Accordingly, as to an instruction of four or more bytes, there are cases where the operation code of the instruction is divided by the extension part of Ea.
Also, as described later, in the case where another extension part is further appended to the extension part, of Ea at a chained addressing mode, this part, takes precedence over the next operation field. For example, consideration is made on the case of 6-byte instruction which comprises Ea1 in a first half word, Ea2 in a second half word, and ends with a third half word. In this case, the chained addressing mode is used for Ea1, the extension part of the chained addressing mode is also appended besides the general extension part. In this case, the actual instruction bit pattern is composed in the sequence of the first half word of instruction (including the basic part of Ea1), the extension part of the chained addressing mode of Ea1, the second half word of instruction (including the basic part of Ea2), the extension part of Ea2 and the third half word of instruction.
(2.1.1) "Short-Format Two-Operand Instruction"
FIG. 4 to FIG. 7 are schematic diagrams showing short format of the two-operand instruction.
FIG. 4 is a schematic diagram showing a format 608 of instruction of operation between memory and register. This format includes an L-format wherein the source operand side is memory and an S-format wherein the destination operand side is memory.
In the S-format, symbol Sh 610 represents the specifying field of the source operand, symbol Rn 612 represents the specifying field of the register of the destination operand and RR 614 represents specifying of the operand size of Sh 610, respectively. The size of the destination operand located in the register is fixed to 32 bits. In the case where the size of the register side differs from that of the memory side and the size of the source side is smaller, sign extension is performed. And, in some instructions (addition instruction, subtraction instruction), the operand size of the source is fixed to a word. In this case, the RR field is the operation code.
In the S-format, symbol Sh 610 represents the specifying field of the destination operand, symbol Rn 612 represents the register specifying field of the source operand and RR 614 represents specifying field of the operand size of Sh 610, respectively. The size of the source operand located in the register is fixed to 32 bits. In the case where the size of the register side differs from that of the memory side and the size of the source side is larger, truncating of the overflow portion and overflow check are performed.
FIG. 5 is a schematic diagram showing a format (R-for-mat) 616 of instruction of operation between register and register. In this format, symbol Rn 618 represents the specifying field of the destination register and symbol Rm 620 represents the specifying register of the source register. The operand size is limited to only 32 bits.
FIG. 6 is a schematic diagram showing a format (Q-format) 622 of instruction of operation between literal and memory. In this format, symbol MM 624 represents the specifying field of the destination operand size, symbol ### 626 represents the specifying field of the source operand and symbol Sh 628 represents the specifying register of the destination operand.
FIG. 7 is a schematic diagram showing a format (I-format.) 630 of instruction of operation between immediate and memory. In this format, symbol MM 632 represents the specifying field of the operand size (common in source and destination), and symbol Sh 634 represents the specifying field of the destination operand. The I-format immediate size is common with the operand size on the destination side and is 8, 16 and 32 bits, and zero extension and sign extension are not performed.
(2.1.2) "General-Format One-Operand Instruction"
FIG. 8 is a schematic diagram showing a general format 636 of one-operand instruction (G1-format). In this format, symbol MM 638 represents the specifying field of the operand size. Some of the G1-format instructions comprise an extension part besides the extension part of Ea 640. There are also instructions using no MM 638.
(2.1.3) "General-Format Two-Operand Instruction"
FIG. 9 to FIG. 11 are schematic diagrams showing general format of two-operand instruction.
This format comprises an instruction having a maximum of two operands of the general-type addressing mode which are specified by eight bits. There are cases where the total number of operands becomes three or more.
FIG. 9 is a schematic diagram showing a format (G-format) 642 of instruction in which the first operand needs trading of memory. In this format, symbol EaM 644 represents the specifying field of the destination operand, symbol MM 646 represents the specifying field of the destination operand size, EaR 648 represents the specifying field of the source operand, and RR 649 represents the specifying field of the source operand size. In addition, some of the G-format instructions comprise &n extension part besides the extension part of EaM 644 and EaR 648.
FIG. 10 is a schematic diagram showing a format (E-format) 650 of instruction whose first operand is 8-bit immediate. In this format, symbol EaM 652 represents the specifying field of the destination operand, symbol MM 656 represents the specifying field of the destination operand size, and the symbol #### . . . 656 represents the source operand.
FIG. 11 is a schematic diagram showing a format (GA-format) 658 of instruction whose first operand is only address calculation. In this format, symbol EaW 660 represents the specifying field of the destination operand, symbol WW 662 represents the specifying field of the destination operand size, and symbol EaA 659 represents the specifying field of the source operand. In addition, as the source operand, the calculation result itself of the effective address is used.
(2.1.4) "The Other Instruction"
Besides the instruction aforementioned, there is such instruction as specifies irregular operand, or instruction with only operation code and with no operand specification.
(2.2) "Addressing Mode"
The methods of specifying the addressing mode of the data processor 100 of the present invention include the short type specified by six bits including the register and the general type specified by eight bits.
In the case where an undefined addressing mode has been specified, or where a combination of addressing modes obviously unsuitable has been specified, a reserved instruction exception is generated, likewise, the case where the undefined instruction has been executed, and exception processing is started. Examples of these cases are the case where the destination is the immediate mode and the case where the immediate mode is used in the specifying field of the addressing mode to be accompanied by address calculation.
Meanings of symbols used in the format diagram shown in FIG. 12 to FIG. 22 is as follows:
Rn: Register specifying field
(Sh): Specifying method by the short-type addressing mode of six bits
(Ea): Specifying method by the general-type addressing mode of eight bits
In the format diagram, the portion surrounded by a broken line shows the extension part.
(2.2.1) "Basic Addressing Modes"
The data processor of the present invention supports various addressing modes. Among them, the basic addressing modes supported by the data processor 100 of the present invention include a register direct mode, a register indirect mode, a register relative indirect mode, an immediate mode, an absolute mode, a Program Counter (PC) relative indirect mode, a stack pop mode and a stack push mode.
The register direct mode is an addressing mode which takes the content of a register intact as an operand. FIG. 12 is a schema-tic diagram of the format thereof. Each symbol Rn 672a, 672b shows the number of the general-purpose register or the FPU register.
The register indirect mode is an addressing mode which takes as an operand the content of the memory whose address is the content of a register. FIG. 13 is a schematic diagram of the format thereof. Symbol Rn 674a, 674b shows the number of the general-purpose register.
The register relative indirect mode includes two kinds due to 16 bits or 32 bits of the displacement value. Each of them is an addressing mode which makes s an operand the content of the memory whose address is the value obtained by adding the value of displacement of 16 bits or 32 bits to content of a general-purpose register. FIG. 14 is a schematic diagram of the format thereof. In the figure, symbol Rn 676a, 676b, 676c shows the number of the general-purpose register. Each symbol disp: 16 678a, 678b and disp: 32 680 show the displacement value of 16 bits and the displacement value of 32 bits, respectively. The displacement value is handled as a signed number.
The immediate mode is an addressing mode which makes as an operand the bit pattern specified in the instruction code assuming it to be a binary number. FIG. 15 is a schematic diagram of the format thereof. Symbol imm.sub.-- data 682a, 682b shows the immediate. The size of imm.sub.-- data is specified in the instruction as the operand size.
The absolute mode includes two kinds of field due to 16 bits and 32 bits of the address value. Each of them is an addressing mode which makes as an operand the content of the memory whose address is the bit pattern of 16 bits or 32 bits specified in the instruction code. FIG. 16 is a schematic diagram showing the format thereof. Each symbol abs: 16 684a, 684b and abs: 32 686a shows the address value of 16 bits and 32 bits, respectively. When the address is shown by abs: 16, the specified address value is sign-extended to 32 bits.
The PC relative indirect mode includes two kinds field due to 16 bits or 32 bits of the displacement value. Each of them is an addressing mode which makes as an operand the content of the memory whose address is the value of the content of a program counter whereto the displacement value of 16 bits or 32 bits is added. FIG. 17 is a schematic diagram showing the format thereof. Each symbol disp: 16 687a, 687b and disp: 32 688a, 688b shows the displacement value of 16 bits and the displacement value of 32 bits, respectively. The displacement value is handled with a signed number. In the PC relative indirect mode, the value of the program counter to be referred is the start address of the instruction comprising the operand. In the case where the value of the program counter is referred in the chained addressing mode, the start address of the instruction is also used as a reference value of PC relative mode.
The stack pop mode is an addressing mode which makes as an operand the content of the memory whose address is the content of a stack pointer (SP). After operand access, the SP is incremented by the operand size. For example, in the case where 32-bit data is handled, the SP is renewed (incremented) by +4 after the operand access. Designating of the stack pop movie for operands of sizes of 8 bits, 16 bits and 64 bits can also be performed, and the SP is renewed by +1, +2 and +8, respectively. FIG. 18 is a schematic diagram of the format thereof. As to the case where the stack pop mode is invalid for the operand, a reserved instruction exception is generated. Specifically, the stack pop mode specifying for a write operand and a read-modify-write operand becomes the reserved instruction exception.
The stack push mode is an addressing mode which makes as an operand the content of the memory whose address is the content of the SP decremented by the operand size. In the stack push mode, the SP is decremented before operand access. For example, when handling 32-bit data, the SP is renewed by -4 before operand access. It is also possible to specify the stack push mode for operands of sizes of 8 bits, 16 bits and 64 bits, and the SP is renewed by -1, -2 and -8, respectively. FIG. 19 is a schematic diagram showing the format thereof. As to the case where the stack push mode is invalid for the operand, a reserved instruction exception is generated. Specifically, the stack push mode specifying a read operand anti a read-modify-write operand becomes the reserved instruction exceptions.
(2.2.2) "Chained Addressing Mode"
Addressing, whatever complicated, can basically be decomposed into combinations of addition and indirect reference. Accordingly, operations of addition and indirect reference are given in advance as primitives of addressing, and in the case where they can be combined arbitrarily, any complicated addressing mode can be realized. The chained addressing mode of the data processor 100 of the present invention is an addressing mode based on such a conception. A complicated addressing mode is particularly useful for data reference between modules or the processing system of AI (Artificial Intelligence) language.
When specifying the chained addressing mode, in the basic addressing mode specifying field, one is specified from among three kinds of specifying methods viz. a register base chained addressing mode, a PC base chained addressing mode and an absolute chained addressing mode.
The register base chained addressing mode is an addressing mode making the value of general-purpose register as the base value of the chained addressing to be extended. FIG. 20 is a schematic diagram showing the format 694 thereof. Symbol Rn 696 shows the number of the general-purpose register.
The PC base chained addressing mode is an addressing mode making the value of the program counter (PC) as the base value of the chained addressing to be extended. FIG. 21 is a schematic diagram showing the format 698 thereof.
The absolute base chained addressing mode is an addressing mode making zero as the base value of the chained addressing to be extended. FIG. 22 is a schematic diagram of the format 700 thereof.
The chained addressing mode specifying field to be extended takes 16 bits as a unit, and this is repeated an arbitrary number of times. In each stage of the chained addressing mode, addition of displacement, scaling (.times.1, .times.2, .times.4, .times.8) and addition of an index register, and indirect reference of memory are performed. FIG. 22 is a schematic diagram showing the format 702 of the chained addressing mode. Each field has meanings as shown below.
E=0: Chained addressing mode is continued.
E=1: Address calculation ends.
tmp==> address of operand
I=0: No memory indirect reference is performed.
tmp+disp+Rx*Scale==> tmp
I=1: Memory indirect reference is performed.
mem [tmp+disp+Rx*Scale]==> tmp
M=0: <Rx> is used as an index.
M=1: Special index
<Rx>=0 Index value is not added (Rx=0).
<Rx >=1 Program counter is used as an index value (Rx=PC).
<Rx>=2 or more Reserved.
D=0: The value of field d4 of four bites in the chained addressing mode is quadrupled to be taken as a displacement value, and this value is added. The d4 is handled with a signed number, and is used by surely quadrupling it irrespective of the size of the operand.
D=1: The dispx (16/32 bits) specified in the extension part of the chained addressing mode is taken as a displacement value and this value is added. The size of the extension part is added is specified in the d4 field.
XX: Index scale (scale=1/2/4/8) In the case where scaling of .times.2, .times.4, .times.8 has been performed for the program counter, an indefinite value is entered as an intermediate value (tmp) after completion of processing of that stage. The effective address obtained by this chained addressing mode becomes an unpredictable value, but no exception is generated. Designating of scaling for the program counter should not be performed.
FIG. 24 and FIG. 25 show variations on the instruction format formed by the chained addressing mode.
FIG. 24 shows variations of continuation 704 and completion of the chained addressing mode.
FIG. 25 shows variations 708a, 708b, 708c on the size of displacement.
In the case where the chained addressing mode of an arbitrary number of stages can be utilized, case-sorting in the compiler, on the basis of the number of stages, can be dispensed with, and therefore an advantage of alleviating the burden on the compiler is obtainable. This is because, even if frequency of chained reference is very low, the compiler has to be able to generate a correct code without fail. For this reason, up to an arbitrary number of stages can be applied in the format.
(2.3) "Jump Instruction"
Schematic views of FIG. 26 to FIG. 42 show instruction formats of the jump instruction of a data processor 100 of the present invention. Hereupon, the instructions are limited to those which are subject to processings associated with pre-jump described hereinbelow. Mark "*" " shown in FIG. 26 to FIG. 42 designates a portion wherein anything can be filled.
FIG. 26 is a schematic view showing an instruction format of a contracted type (D format) unconditional branch instruction (BRA instruction). The BRA instruction is the jump instruction where a branch target address is designated by PC relative. In the figure, symbol pcdisp is a branch displacement and designated by an 8-bit immediate value. The branch displacement is treated as signed numbers. In the D format instruction, a doubled branch displacement is used.
FIG. 27 is schematic view showing an instruction format of a general type (G format) unconditional branch instruction (BRA instruction). In the figure, pcdisp is a branch displacement, designated by an 8-, 16- or 32-bit immediate value and treated as signed numbers. A branch displacement size is designed by sx. Bit allocations of sx are shown hereinbelow.
sx=00: 8 bits
sx=01: 16 bits
sx=10: 32 bits
In the data processor 100 of the present invention, since the instruction of the byte length of odd number are forbidden, even when the 8-bit displacement is designated, it is designated by 16 bits as a field. In this case, however, only the lower 8 bits of the 16-bit field are valid and the upper 8 bits are neglected whatever values they may be.
FIG. 28 is a schematic view showing an instruction format of a contracted type (D format) conditional branch instruction (Bcc instruction), and FIG. 29 is a schematic view showing an instruction format of a general type (G format,) conditional branch instruction (Bcc instruction). As the instruction format, they are same as the BRA instruction. Branch conditions are designated by cccc fields in the figures. Branch processing is implemented only when the designated condition is realized. When the condition is not realized, the next instruction is executed.
FIG. 30 is a schematic view showing an instruction format of a contracted (D format) subroutine branch instruction (BSR instruction.), and FIG. 31 is a schematic view showing an instruction format of a general type (G format) subroutine branch instruction (BSR instruction). As the instruction format, they are same as the BRA instruction. Besides the branch processing, a return target address is pushed to a stack top.
FIG. 32 is a schematic view showing a format of the jump instruction (JMP instruction), and FIG. 33 is a schematic view showing a format of the subroutine jump instruction (JSR instruction). They are both the instructions of the above-mentioned general type 1 operand instruction (G1 format). In the figures, symbol pcadr shows a branch target address and is designated by the aforementioned general type addressing mode. They jump to the operand address calculated according to designation of the addressing mode. It is also possible to designate a chained mode as the addressing mode. In the JSR instruction, the return target address is pushed to the stack top.
FIG. 34 through FIG. 37 are schematic views showing an instruction format of the ACB instruction which adds, compares and performs conditional branch. This instruction adds a step value to an xreg value and writes the addition result back to xreg, and when the addition result is less than a limit, branches to address obtained by adding the branch displacement designated by pcdisp to a PC value, and is used as a loop control primitive. The branch displacement is designated by sz. Bit allocations of sz are shown hereinbelow.
sz="00": 8 bits
sz="01": 16 bits
sz="10": 32 bits
When the displacement size is 16 bits or 32 bits, a 16-bit or 32-bit expanded portion is added. When the displacement size is 8 bits, displacement is designated by the field off pcdisp and the expanded portion is not added. The branch displacement is treated as signed number.
The ACB instruction has four instruction formats. xreg is designated by a register for all formats. FIG. 34 is a schematic view showing a contracted type format (Q format), whereby a step is fixed at "1" and a limit value is designated by a 6-bit literal (a small immediate value encoded in an operation code). FIG. 35 is a schematic view showing a contracted type format (R format), whereby step is fixed at "1" and the limit value is designated by the register.
FIG. 36 is a schematic view showing an instruction format of the aforementioned G format. The step and limit are designated by a general type addressing mode. FIG. 37 is a schematic view showing an instruction format of the aforementioned E format. The step value is designated by an 8-bit, literal value, and the limit value is designated by the general type addressing mode.
FIG. 38 through FIG. 41 are schematic views showing instruction formats of the SCB instruction which subtracts, compares and performs conditional branch. The SCB instruction subtracts the step value from the xreg value and writes the subtraction result back to the xreg, and when the subtraction result is above the limit, branches to address obtained by adding the branch displacement designated by pcdisp to the PC value. It is used as a loop control primitive in the same way as the ACB instruction.
The SCB instruction has four formats as same as the ACB instruction. FIG. 38 is a schematic view showing an instruction format of the Q format, FIG. 39 is a schematic view showing an instruction format of the R format, FIG. 40 is a schematic view showing an instruction format of the G format, and FIG. 41 is a schematic view showing an instruction format of the E format.
FIG. 42 is a schematic view showing an instruction format of the FBcc instruction which executes the conditional branch by the operation result of floating-point numbers. The branch condition is designated by ccccc. When the condition is realized, it branches to address obtained by adding the branch displacement indicated by pcdisp to the PC value. The branch displacement size is designated by sy. The branch displacement is treated as the signed numbers. Though sx is a field from relation with the instruction format of the other instruction, it has no particular meaning in the FBcc instruction. A designating process of the branch displacement is similar to the G format BRA instruction.
sy="00": 8 bits
sy="01": 16 bits
sy="10": 32 bits
In the data processor 100 of the present invention, the instruction is in 16-bit unit, and it is forbidden to jump to odd addresses. Thus, when jumping to the odd address occurs, an odd-address jump trap (OAJT) starts to perform exceptional processing.
(3) "Configuration of Function Blocks of the Data Processor of the Present Invention"
FIG. 43 is a block diagram showing a configuration of the data processor 100 of the present invention.
When roughly dividing the inner portion of the data processor of the present invention functionally, it can be divided into an instruction input unit 110, an instruction fetch unit, 111, an instruction decoding unit 112, a first micro-ROM 113, a second micro-ROM 114, an operand address calculation unit 115, a PC processing unit 116, an integer operating unit 117, a floating-point operating unit 118, an address input/output unit 119, an operand accessing unit 120 and a data input/output unit 121.
Each of the blocks transfers addresses and data via an S1 bus 201, an S2 bus 202, a DO bus 203, an II bus 204, a DISP bus 205, an IX bus 206, an AG bus 207, a JA bus 208, an AA bus 209, a DD bus 210, an EI bus 211, an EA bus 212, an ED bus 213, an ADJST bus 214, an ILEN bus 215 and an AO bus 216. A system configuration shown in FIG. 2 can be accomplished by, connecting the address input/output unit 119 to an address bus 101, the data input/output unit 121 to a data bus 102 and the instruction input unit 110 to an instruction bus 103.
(3.1) "Instruction Input Unit"
The instruction input unit 110 inputs the instruction code from the external instruction bus 103 to the data processor 100 of the present invention in 32-bit unit. For accessing the instruction cache 106, there are a standard accessing mode which accesses 32-bit instruction code for one address and a good accessing mode which accesses 32-bit instruction code four times sequentially for one address, and in either case, the instruction input unit 110 outputs the inputted instruction code to the instruction fetch unit 111 via the EI bus 211.
(3.2) "Instruction Fetch Unit"
FIG. 44 is a block diagram showing a configuration of the instruction fetch unit 111. The instruction fetch unit 111 includes an instruction MMU (Memory Management Unit)/cache unit 81 containing an address conversion mechanism of instruction address, internal instruction cache and instruction TLB and its control unit, and an instruction queue unit 82.
The instruction MMU/cache unit 81 converts address of the instruction address and fetches the instruction. Usually, a PC value of the next instruction to be fetched is transferred to the instruction MMU/cache unit 81 from a QINPCA 86 or a QINPCB 87 via a PFA bus 91. However, when a jump occurs and the jump target address is taken in immediately, a jump target address value outputted from the operand address calculation unit 115, PC processing unit 116 or integer operating unit 117 is taken in directly by the MMU/cache unit 81 via the JA bus 208. In the instruction MMU/cache unit 81, the PC value is converted into the physical address, fetching the instruction code from the internal instruction cache and sending instruction code to an instruction queue A 83 or an instruction queue B 84 via an ICO bus 93.
When the internal instruction cache is missed, physical address is outputted to the address input/output unit 119, the instruction access is required to the outside, and the instruction code inputted through the instruction input unit 110 is registered in the internal instruction cache and sent also to the instruction queue A 83 or the instruction queue B 84. Address conversion and instruction TLB updating by the paging when the instruction TLB is missed are also performed by the control circuit included in the instruction MMU/cache unit 81.
The instruction queue unit 82 performs instruction fetch address control and buffering of the fetched instructions. The instruction queue unit 82 includes two instruction queue A 83 and instruction queue B 84 and two address registers QINPCA 86 and QINPCB 87 holding an instruction queue fetch address, so as to enable queuing of the two-way branch/no-branch instruction of the conditional branch instruction. A QINPCC 88 which is a latch for saving the branch target address of the second conditional branch instruction is also included.
Numeral 89 designates a counter for incrementing the instruction pre-fetch address, and numeral 90 designates an input latch of the counter 89.
The instruction queue A 83 and instruction queue B 84 have a 32-byte capacity, and input/output control is effected by an input/output pointer unit 85. An input pointer of the instruction queue is controlled by a word (4 bytes) unit, and at cache hit, the 128-bit instruction code is taken in from the ICO bus 93 and at cache miss, the 32-bit instruction is fetched from the EI bus 211 into the instruction queue A 83 or instruction queue B 84. An input pointer is updated responsive to the hit/miss of the instruction cache. An output pointer is controlled by a half word (2 bytes) unit, and the instruction code fetched into the instruction queues A 83 and B 84 are arranged at a 16-bit unit, outputted to the II bus 204 and transferred to the instruction decoding unit, 112 and PC processing unit 116. The output pointer is updated responsive to output pointer update information sent from the instruction decoding unit 112.
The instruction queues A 83 and B 84 are also able to bypass data fetched from the ICO bus 93 or EI bus 211 and output to the II bus 204, when the instruction queue is vacant as immediately after a jump.
The input/output pointer unit 85 also includes two sets of input/output pointers which control the instruction queue A 83 and instruction queue B 84, and a latch saving a pointer value corresponding to the QINPC 88. When a jump occurs, the pointer is initialized by the value fetched from the JA bus 208.
The instruction queue unit 82 processes the branch instruction efficiently by using two instruction queue A 83 and instruction queue B 84 and the QINPCC 88. Though the instruction queue unit 82 fetches two-way instruction at processing of the conditional branch instruction, it preferentially fetches the sequence instruction on the predicted side of the next execution, and when the sequence instruction queue on the predicted side is full, fetches the other sequence instruction. Furthermore, the instruction queue unit 82 saves the branch target address of the second conditional branch instruction in the QINPCC 88 so as to enable pipeline processing of the two conditional branch instructions, and when one of the two instruction queues becomes vacant, fetches the instruction of address saved in the QINPCC 88 immediately.
Usually, when the instruction queue unit 82 fetches the instruction continuously, the value of QINPCA 86 or QINPCB 87 is outputted to the PFA bus 91 and sent to the instruction MMU/cache unit 81, and at the same time, fetched into the counter 89 via the latch 90 so as to be incremented, and written back to the QINPCA 86 or QINPCB 87. When the jump target address is transferred via the JA bus 208, a value on which is fetched into either of the QINPCA 86, QINPCB 87 or QINPCC 88. That to which register is written are dependent on the kind of requests received by the instruction fetch unit 111 and the internal state of the instruction queue unit 82. When the instruction of the jump target address on the JA bus 208 is fetched in the next cycle, the jump target address on the JA bus 208 is fetched directly into the counter 89 via the latch 90 so as to be incremented, and written into the QINPCA 86 or QINPCB 87.
When a valid address is taken into the QINPCA 86, QINPCB 87 and QINPCC 88 during processing two conditional branch instructions in the instruction queue unit 82, and a branch prediction of the first conditional branch instruct, is right, either the instruction queue A 83 or instruction queue B 84 becomes void and the value of QINPCC 88 is written into the QINPCA 86 or QINPCB 87 on the voided side. When starting to fetch the instruction of the address saved in the QINPCC 88 immediately, the value of the QINPCC 88 is outputted to the PFA bus 91 and sent to the instruction MMU/cache unit 81, and at the same time, taken into the counter 89 via the latch 90 so as to be incremented, and written back to the QINPCA 86 or QINPCB 87.
When the data processor 100 of the present invention is in a bus-watch mode, entry of the internal instruction cache hit by physical address inputted through the address input/output unit 119 is voided.
(3.3) "Instruction Decoding Unit"
In the instruction decoding unit 112, the instruction code fetched from the II bus 204 is decoded in a 16-bit half word) unit. The instruction decoding unit 112 includes, an OPDEC1 which decodes an operation code included in the first half word, an OPDEC" which decodes an operation code included in the second and third half words an addressing mode decoder which decodes the addressing mode and a SUBDEC1 and SUBDEC2 which decode only the 2-byte instruction limited for decoding the two instructions in parallel.
Furthermore, a second decoder which further decodes outputs of the OPDEC1 and OPDEC2 to calculate entry address of the micro-ROM, a branch predicting unit which performs branch prediction of the conditional branch instruction, and an address calculation conflict check mechanism which checks a pipeline conflict at operand address calculation are also included. In the branch predicting unit, branch prediction of the conditional branch instruction is performed by the past one branch hysteresis. This branch prediction information is referred to at decoding of the conditional branch instruction. The branch histories stores as a tag a part of instruction address executed immediately before the conditional branch instruction, and PC value which registers and refers to is sent from the PC processing unit 116.
The instruction decoding unit 112 decodes the instruction code outputted from the instruction fetch unit 111 at 0 to 8 bytes for one clock. Information related to operation in the integer operating unit 117 among the decoded results is outputted to the first micro-ROM 113, information related to operation in the floating-point operating unit 118 to the second micro-unit 114, information related to the operand address calculation to the operand address calculation unit 115, information related to the PC calculation to the PC processing unit 116 and output pointer update information of the instruction queue to the instruction fetch unit 111. The instruction decoding unit 112 is able to decode the two instructions in parallel, only when the rear side instruction is the 2-byte instruction having no memory operand, the decade result (called a sub-code) related to the rear side instruction is sent to the first micro-ROM 113. Information related to the jump processing performed at an instruction decoding stage is outputted to the instruction fetch unit 111, PC processing unit 116 or operand address calculation unit 115.
Though the branch prediction unit performs dynamic branch prediction of the conditional branch instruction, it can be set, by the control register in the integer operating unit 117, to predict statically not to branch.
(3.4) "First Micro Unit"
The first micro ROM unit 113 comprises a micro ROM storing microprogram which mainly controls the integer operation unit 117, a microsequencer, a microinstruction decoder, sub-code decoder and the like. A microinstruction is read out from the micro ROM once at every one clock. The microsequencer accepts processings of exception, interruption, and trap (these three are called EIT), besides the sequential processing for execution of the microprogram relating to instruction execution.
To the first micro ROM unit 113, there are inputted interruption independent, of the instruction code and branch condition of the microprogram obtained by the result of integer operation execution.
Output of the microinstruction decoder is mainly given to the integer operation unit 117, but some information is outputted also to other blocks at execution of jump instruction and acceptance of execution.
The sub-decoder decodes sub-code. When the sub-code is effective, operation information relating to the sub-code becomes effective in synchronism with the last microstep of one instruction. The output of the sub-decoder is output to the integer operation unit 117.
(3.5) "Second Micro ROM Unit"
In the second micro ROM unit 114, a micro ROM in which various microprogram routines for controlling a floating-point operation unit 118 are stored, a microsequencer, a microinstruction decoder and so on are included. One microinstruction is read from the micro ROM in one clock cycle. The microsequencer also processes the exception relating to the floating-point operation besides the sequence processing specified by the microprogram, and when the unmasked floating point exception is detected, requests the exception processing to the first micro ROM unit 11.3. The microsequencer of the second micro ROM unit 114 operates in parallel to the microsequencer of the first micro ROM unit 113, and controls the floating-point operation unit 118 in parallel to the integer operation unit 117.
To the second micro ROM unit 113, flag information obtained from the result, of the floating-point operation execution is also inputted.
Though out, put of the microinstruction decoder is mainly outputted to the floating-point operation unit 118, a part of information such as detection of the exception relating to the floating-point operation is also outputted to the other functional blocks.
(3.6) "Operand Address Calculation Unit"
The operand address calculation unit 115 is controlled in hardwired logic by information on operand address calculation outputted from the addressing mode decoder of the instruction decoding unit 112. In this operand address calculation unit 115, operand address calculation other than the memory access for the memory indirect addressing, and jump target address calculation of the jump instruction are executed.
FIG. 45 is a block diagram showing a detailed configuration of the operand address calculation unit 115. In the figure, reference numeral 61 designates an address adder performing address calculation, 62, 63 and 64 input latches, 65 an output latch (AO latch), 66 an ADJST latch holding compensation values such as a pre-decrement value at push mode, and numeral 57 designates a DISP latch holding displacement and immediate for address calculation, respectively. In addition, in order to process in one cycle address calculation of one stage in chained mode, a ternary adder is used in the address adder 61.
The value of general-purpose register and the PC value referred to at address calculation are taken in from the latch 64 or the latch 63 from the integer operation unit 117 and the PC processing unit 116 through IX bus 206. Into the latch 62, the data held in the ADJST latch 66 or the DISP latch 67 is sign-extended to 32 bits and taken. When indirect reference is performed in chained mode, fetched address is transferred to the operand access unit 120 through the AO latch 65 and the AA bus 209 so as to perform data fetch. The fetched data is taken into the input latch 64 through the AG bus 207. When indirect reference is not performed in chain mode, the address calculation result at previous step is taken intact to the input latch 64. The address calculation result is sent to the integer operation unit 117 through the AO bus 216. And when pre-fetch of operand is performed, the operand address is transferred to the operand access unit 120 through the AO latch 65 and the AA bus 209.
When the operand is an immediate, the data taken in the DISP latch 67 is sent to the integer operation unit 117 through the address adder 61, the AO latch 65 and the AO bus 216. And when an immediate is a double precision floating point number (64 bits), it is divided into two parts, that is, upper 32 bits and lower 32 bits, and taken into the DISP latch 67 at two times. In this case, the upper 32 bits are saved in the AO latch 65 in the first cycle and the upper 32 bits are transferred to the integer operation unit 117 through the AG bus 207 and the lower 32 bits are transferred to the integer operation unit 117 through the AO bus 216 in the second cycle.
When jump is performed at the stage where the operand address calculation is finished, the calculation result of the jump target address is outputted to the instruction fetch unit 111 and the PC processing unit 116 through the AO latch 65 and the JA bus 208.
(3.7) "PC Processing Unit"
FIG. 46 and FIG. 47 are block diagrams showing configuration of the PC processing unit 116. The PC processing unit 116 roughly comprises a jump target address generating unit 1, a PC calculation unit 2, a step PC transferring unit 3 and a PC stack 4.
The PC calculation unit 2 is controlled in hardwired logic using information on PC calculation outputted from the instruction decoding unit 112 so as to calculate the PC value of an instruction. The instruction of the data processor 100 of the present invention is a variable-length instruction, and the length of that instruction can be found only after the instruction is decoded. The PC calculation unit 2 calculates the PC value of the next instruction by adding the length of the instruction outputted from the instruction decoding unit 112 to the PC value of the instruction being decoding.
In the PC calculation unit 2, the DPC 29 holds a head address of the instruction being decoded. The TPC 27 holds a head address of the code to be decoded, including the case where an instruction code of one instruction is divided into a plural number of processing units so as to be processed. That is, the head address of the code taken in from II bus 204 into the instruction decoding unit 112 is held.
The PC adder 23 adds an instruction length taken in from the instruction decoding unit 112 through ILEN bus 215 and a value of the TPC 2? , and the addition result is written back to the TPC 27 and the ATPC 28. And, at decode completion of one instruction, since the addition result indicates a head address of the next instruction, it is also written back to the DPC 29. The ATPC 28 is referred to as the PC value of the next instruction at address calculation. When jump occurs, a jump target address is taken in from the JA bus 208, and the TPC 27 and the DPC 29 are initialized. But the ATPC 28 is not initialized even when jump occurs. A part, of the DPC 29 is also transferred to the instruction decoding unit 112 so as to be used for branch prediction.
The jump target address generating unit 1 is a block in which branch target address is calculated at instruction decoding stage. The jump target address generating unit 1 calculates a branch target address by cutting out data of a branch displacement position from an instruction code on the II bus 204 to be sent from the instruction fetch unit 111 to the instruction decoding unit 112 and by adding it to the DPC 29. The data processor 100 of the present invention comprises two adders of the first and second adders 5 and 6 so as that jump target address can be calculated at the same time with instruction decoding as to all instruction of BRA, BSR and Bcc instructions.
In addition, reference numerals 7 and 8 designate input latches of the adders 5 and 6 respectively, and numerals 10 and 11 designate output latches of the adders 5 and 6 respectively.
An input latch 9 and an output latch 12 both for holding absolute addresses are also provided.
The jump target address generating unit 1 operates in synchronism with decoding of an instruction code. The jump target address generating unit I is provided with two adders one of which is the first adder 5 for mainly calculating a branch target address of the BRA instruction, BSR instruction, Bcc instruction of the D format and the other is the second adder 6 for mainly calculating a branch target address of the BRA instruction, BSR instruction, Bcc instruction of the G format.
In synchronism with the respective instruction decoding cycles, a value of the DPC 29 is taken into the first adder 5 and the second adder 6 through the PI bus 40. And data on the II bus 204 is taken into the input latches 7, 8 and 9.
In FIG. 48, bit positions at the II bus 204 taken in the respective input latches 7, 8 and 9 as branch displacements and absolute addresses are shown.
The input latch 7 doubles (to shift to the left by one bit) bits (8:15) (eight bits from bit 8 to bit 15, hereinafter to be called II(8:15)) of the II bus 204 which are the branch displacement field of the BRA instruction, BSR instruction, and Bcc instruction of the D format, and takes in data which have been sign-extended to 32 bits. And the input latch 8 selectively takes in data on the II bus 204 responsive to values of bits (6:7) (II(6:7)) of the II bus 204 which correspond to the branch displacement field of the BRA instruction, BSR instruction and Bcc instruction. When II(6:7)="00" or II(6:7)="11", the input latch 8 takes in data obtained by making bits (24:31) (II(24:31)) of the II bus 204 sign-extended to 32 bits. And when II(6:7)="01", the input latch 8 takes in data obtained by making II(16:31) sign-extended to 32 bits. Further when II(6:7)="10", the input leach 8 takes in II(16:47) intact.
The first adder 5 adds the content of the input latch 7 and a value of the DPC 29, and the second adder 6 adds the content of the input latch 8 and a value of the DPC 29 respectively, and they transfer the respective addition result to the output latches 10 and 11 respectively.
The input latch 9 selectively takes in, as shown in FIG. 48, data on the II bus 204 responsive to a value of bit 15 (II(15)) of the II bus 204 capable of distinguishing an address data size when an addressing mode is the absolute mode. When the II(15)="1", the input latch 9 takes in data obtained by making II(16:31) sign-extended to 32 bits. When the II(15)="0", the input latch 9 takes in II(16:47) intact. To the output latch 12, data taken in the input latch 9 is transferred intact.
During the branch instruction decoding cycle, when a branch target addresses cannot be calculated at the first and second adders 5 and 6, the input latch 7 takes in a branch displacement from the instruction decoding unit 112 through the DISP bus 205 in the next cycle. The first adder 5 adds the content of the input latch 7 and a value of the DPC 29 taken in through the PI bus 40 to each other, and transfers the addition result to the output latch 10.
When a jump processing is performed in the stage of instruction decoding, one of the output latches 10, 11, 12 is selected and the content thereof is outputted to the JA bus 208, thereby a jump target address is transferred to the instruction fetch unit 111. And when a processing sequence of an instruction is switched, a value on the JA bus 208 is taken in the TPC 27 and DPC 29.
The PC stack 4 is a buffer for holding a copy of a return target PC value from a subroutine pushed to the stack. For a return instruction from the subroutine, a return target PC value is read out from the PC stack and transferred to the instruct ion fetch unit 111 through the JA bus 208. And a value on the JA bus 208 is taken in the TPC 27 and DPC 29.
The step PC transferring unit 3 is provided with plural number of latches for holding PC values corresponding to the respective pipeline stages, and values of the DPC 29 calculated in the stage of instruction decoding are transferred as the PC values of the respective instructions with the pipeline processing proceeds. An OLDPC 36 holds a PC value of an instruction executes immediately before by the execution stage and outputs it to the instruction decoding unit 112. By using this value, a branch hysteresis table in the instruction decoding unit 112 is updated.
(3.8) "Integer Operation Unit"
The integer operation unit 117 is controlled by the microprogram stored in a micro ROM of the first micro ROM unit 113, and executes operations necessary for realizing the function of each integer operation instruction, using the register file and arithmetic unit in the integer operation unit 117. In the register file, a general register and a working register are included. The integer operation unit 117 is provided with as operation devices, a main operation device comprising an ALU, a shifter, a priority encoder and the like controlled by microprogram, and with a sub-operation device comprising an ALU, a shifter and the like for executing sub-code.
And the integer operation unit 117 is so configurated that two instructions can be executed i parallel with each other by using the main operation device and the sub-operation device. The sub-operation device can also be controlled by a microprogram. In the integer operation unit 117, a processor status word (PSW) including a flag which is varied by the result of integer operation and the bits which decide the external interruption mask level, and the buffer memory control register are included.
When an operand to be calculated by an instruction is an address or an immediate, the immediate or calculated address is inputted from the operand address calculation unit 115. Also, when an operand to be calculated by an instruction is data on the memory, its address calculated in the address calculation unit 115 is outputted to the operand access unit 120, and the operand fetched from a built-in data cache or the outside is inputted to the integer operation unit 117 through the DD bus 210.
In operation, in the case where the built-in data cache, external data caches 107, 108 or the main memory 109 must be read, the address is outputted to the operand access unit 120 through the AA bus 209 and target data is taken in from the DD bus 210, under the control of the microprogram. When it is necessary to store the operation result in the built-in data cache, external data caches 107, 108 or the main memory 109, the address is inputted through the AA bus 209 to the operand access unit 120, and the data through the bus 210 to the same, under the control of the microprogram.
When external interruption or exception is processed and the integer operation unit 117 and receives a new instruction address, the integer operation unit 117 outputs the new address to the instruction fetch unit, 111 and the PC processing unit 116.
(3.9) "Floating Point Operation Unit"
The floating-point operation unit 118 is controlled by the microprogram stored in a micro ROM of the second micro ROM unit 114, and executes operations necessary for realizing the function of each floating-point operation instruction, using the register file and operation device in the floating-point operation unit 118. Also, the floating-point operation unit is provided with a floating-point operation mode control register (FMC) which sets a mode of the rounding method of flowing-point operation and detection allowance of floating-point operation exception and with a floating-point operation status word (FSW) comprising status bits indicating generation states of flag for the floating-point operation result and of the floating-point exception.
When an operand to be calculated by an instruction is an immediate or data on the memory, the operand value is inputted to the floating-point operation unit 118 from the integer operation unit 117 through the bus 201 or the S2 bus 202.
In storing-operation, the floating-point operation unit 118 and the integer operation unit 111 operate in cooperation with each other. When it is necessary to store the operand in the built-in data cache, external data caches 107, 108 or the main memory 109, the floating-point operation unit 118 outputs data to the integer operation unit 117 through the DO bus 203 under the control of the microprogram. The integer operation unit 117 outputs an operand address through the AA bus 209 and operand data through the DD bus 210 to the operand access unit 120, respectively.
(3.10) "Operand Access Unit"
An operand access unit 120 includes the operand address translation mechanism, the built-in data cache, a data TLB, a pre-fetch data buffer, a store buffer and an operand breakpoint register and their control units. In loading-operation of data, a logical address of data to be loaded outputted from the operand address calculation unit 115 or the integer operation unit 117 to the AA bus 209 is converted into a physical address, thereby data is fetched from the built-in data cache and is outputted to the integer operation unit 117.
When cache miss occurs at the built-in data cache, the physical address is outputted to the address input/output unit, 119, and the data access to the outside is required and data input ted through the data input/output unit 121 is registered in the built-in data cache. When storing the data, the logical address of data to be stored outputted from the integer operation unit 117 is converted into the physical address, and data outputted from the integer operation unit 117 or the floating-point operation unit 118 is stored in the built-in data cache, and the physical address is outputted to the address input/output unit 119 through the store buffer and the data is outputted to the outside through the data i put/output unit 121. When a miss occurs in the storing operation, the data is not updated.
In the store buffer, data to be stored and its address, and more over, the PC value of the instruction having executed the storing operation are managed in a set. The storing operation in the store buffer is managed in a first-in first-out (FIFO) manner.
Address translation by paging fox the data TLB miss and the update of the data TLB are also performed by the control circuit in the operand access unit 120. It is also checked whether or not the memory access address is in an I/0 area mapped in the memory.
When the data buffer is made to be operated as the built-in data cache, in the case where the data processor 100 of the present invention is under a bus watching mode, the entry of the built-in data cache hit by the physical address inputted through the address input/output unit 119 invalidated.
(3.11) "Address Input/Output Unit"
The address input/output unit 119 outputs the address outputted from the instruction fetch unit 111 and the operand access unit 120 to the outside of the data processor 100 of the present invention. The address is outputted in accordance with a bus protocol defined in the data processor 100 of the present invention. The bus protocol is controlled by an external bus control circuit in the address input/output unit 119. In the external bus control circuit, page fault exception, bus access exception and external interruption are also received.
When external device other than the data processor 100 of the present invention is a bus master and the data processor 100 of the present invention is under the bus watching mode, the address outputted on the address bus 101 is taken in when the external device executes the data write cycle, and transfers it to the instruction fetch unit 111 and the operand access unit 120.
(3.12) "Data Input/Output Unit"
The data input/output unit 121, at operand loading operation, fetches data from the data bus 102 and transfers it to the operand access unit 120, and at operand storing operation, outputs the operand outputted from the operand access unit 120 to the data bus 102.
As the accessing methods of the data caches 107, 108, there are a single access mode in which a 64-bit data is accessed for one address, and a quad access mode in which four 64-bit data are accessed continuously for one address, and in either case, the data input/output unit 121 controls the input and output of the data between the operand access unit 120 and the external memory.
(4) "Pipeline Processing"
The data processor 100 of the present invention performs the pipeline processing of the instruction and operates very efficiently, by effective accessing of the memory using various buffer memories, the instruction bus 103 and the data bus 102. Here, a pipeline processing method of the data processor 100 of the present invention will be described.
(4.1) "Pipeline Processing Mechanism"
A pipeline processing mechanism of the data processor 100 of the present invention is constituted as shown schematically in FIG. 25.
The pipeline processing is executed in six-stage configuration of an instruction fetch stage (IF stage) 131 which prefetches instructions, a decoding stage (D stage) 132 for decoding the instruction, an operand address calculation stage (A stage) 133 which performs address calculation of the operand, an operand fetch stage (F stage) 134 which performs the micro ROM access (particularly referred to as an R stage 137) and the operand prefetch (particularly referred to as an OF stage 138), an execution stage (E stage) 135 for executing the instruction and a store stage (S stage) 136 which stores the memory operand. In addition, in the S stage 136, there is a three step store buffer.
Each stage operates independently of the other stages and, theoretically, the six stages operate completely independently.
Each stage other than the S stage 136 performs one processing in one clock cycle at a minimum. The S stage 136 performs one operand storing in two clock cycle at a minimum. But, until the store buffer becomes full, the S stage 136 can receive one storing in every clock cycle. Accordingly, theoretically, the pipeline processing is proceeded in every clock cycle.
In the data processor 100 of the present invention, though there are instructions which can not be processed by only one basic pipeline processing such as the memory-memory operation or the memory indirect addressing, it is so constituted that the balanced pipeline processing can also be performed for these instruction if possible. For the instruction having a plural number of memory operands, the data processor 100 of the present invention performs pipeline processing by decomposing it into plural pipeline processing unit (step code) at the decoding stage in response to the number of memory operands.
Information given to the D stage 132 from the IF stage 131 are instruction code. Informations given from the D stage 132 to the A stage 133 are the one (referred to as a D code 141) relating to the operation specified by the instruction, the one (referred to as an A code 142) relating the operand address calculation and a program counter value (PC) of the instruction being processed. In the A code 142, information relating to pre-jump processing performed at the A stage 133 is also included.
Informations given to the F stage 134 from the A stage 133 are, an R code 143 including the entry address of the microprogram routine and parameters to the microprogram, an F code 144 including the operand address and accessing indicating information, and the program counter (PC) value of the instruction being processed. Informations given to the E stage 135 from the F stage 134 are, an E code 145 including operation control information and literal, S code 146 including the operand and its address and the program counter (PC) value of the instruction being processed. The S codes 146 comprises an address and data.
Information given to the S stage 136 from the E stage 135 are W code 147 which is the operation result to be stored and the PC value of the instruction being outputted the operation result. The W codes 147 comprises an address and data.
An EIT (Exception, Interruption and Trap) detected at the stages before the E stage 135 does not start EIT processing until the code thereof reaches the E stage 135. It is because that, only the instruction processed at the E stage 135 is the instruction at the execution step, and the instruction processed at the IF stage 131 through F stage 134 is still not in the execution step. Accordingly, detection of the EIT before the E stage 135 is registered in the step code and just transmitted to the next stage. The EIT detected at the S stage 136 is received at the time point when execution of the instruction being processed in the E stage 135 is completed or at cancellation of the processing of the instruction, and restored to the E stage 135 for processing.
(4.2) "Processings in Each Pipeline Stage"
The input/output step codes to respective pipeline stages are named, as shown in FIG. 49, fop the sake of convenience. In the step codes, there are two kinds, the one which becomes the entry address of micro-ROM or its parameters for the E stage 135, and the other which becomes the operand to be processed at the E stage 135. Between the D stage 132 and the S stage 136, the PC value of the instruction being processed is received and given.
(4.2.1) "Instruction Fetch Stage"
At the instruction fetch stage (IF stage) 131, the instruction fetch unit 111 is operated. The IF stage 31 fetches the instruction from the built-in instruction cache or the outside and inputs it to an instruction queue, and outputs the instruction code to the D stage 132 in a unit of 0 to 8 bytes. Input to the instruction queue is performed in a unit of aligned 4 bytes.
When the instruction from the outside under a standard access mode, it requires at least 2 clock cycles for the aligned 4 bytes. Under a quad access mode, at least 5 clock cycles are necessary for 16 bytes. When the built-in instruction cache is hit, fetch is possible in one clock cycle for the aligned 16 bytes. Output unit of the instruction queue is variable by every 2 bytes, and can be outputted to 8 bytes in 1 clock at a maximum. Right after the jump, the instruction queue may be bypassed and the 2-byte or 4-byte instruction code can be transferred directly to the instruction decoder.
Translation of the logical address of the instruction into the physical address, control of the built-in instruction on cache and instruction TLB, management of the prefetch target instruction address, and control of the instruction queue are performed at the IF stage 131.
In addition, the EIT detected at the IF stage 131 includes an address translation exception relating to the instruction fetch page fault exception and bus access exception.
(4.2.2) "Instruction Decoding Stage"
The instruction decoding stage (D stage) 132 decodes an instruction code inputted from the IF stage 131. The instruction code is decoded once-for one clock using the OPDEC1, OPDEC2, addressing mode decoder, SUBDEC1 and SUBDEC2 of the instruction decoding unit 112, and the instruction codes of 0 to 8 bytes are consumed at one decoding processing. The instruction succeeding the instruction decoded in the OPDEC1 or OPDEC2 is decoded in the SUBDEC1 or SUBDEC2, and when an interference does not occur related to the register operand, two instructions are processed at one decoding processing (the decade result by the OPDEC1 or OPDEC2 is called a main code, the decade result by the SUBDEC1 or SUBDEC2 is called a sub-code).
In some instructions, there is the case that a step code is generated without consuming the instruction code after a second decoding cycle of one instruction. An A code 142 which is address calculating information and a D code 141 which is an intermediate decade result of the operation code are outputted to the A stage 133 at one decoding.
At the D stage 132, the PC processing unit 116 is controlled and output control of the instruction code from the instruction queue is also performed. At the D stage 132, pre-jump processing for the unconditional branch instruction and the return instruction from the subroutine is also performed. For the unconditional branch instruction to which pre-jump processing is performed, neither D code 141 nor A code 142 is outputted and the instruction processing is finished at the D stage 132.
However, when the jump target address is odd numbers or the debug event detecting mode is effective, the D code 141 and A code 142 are outputted for the unconditional branch instruction.
In exception detected at the D stage 132, there is an odd address jump trap related to the prediction instruction exception and pre-jump.
(4.2.3) "Operand Address Calculation Stage"
Processings of an operand address calculation stage (A stage) 133 are roughly divided into two. First is the post-stage decoding processing of the operation code using the second decoder of the instruction decoding unit 112, and second one is the calculation processing of the operand address in the operand address calculation unit 54.
In the post-stage decoding processing of the operation code, the D code 141 is inputted and the R code 143 including the address of the write reservation of the register and memory, the entry address of microprogram routine, and parameters for the microprogram is outputted. Incidentally, the write reservation of the register and memory is for preventing, when the content of the register or the memory referred all address calculation is rewritten by the preceding instruction in the pipeline, the wrong address calculation.
In the operand address calculation processing, the A code 142 is inputted and in accordance with the A code 142, the operand address is calculated in the operand address calculation unit 54 to output the calculation result as the F code 144. At this time, the write reservation is checked when the register is read in connection with the address calculation, and when the preceding instruction has been indicated that there is the reservation because the writing processing for the register or memory is not completed, the following instruction is in the waiting states until the writing processing of the preceding instruction is completed at the E stage 135.
At the A stage 133, for the jump instruction which has not pre-jumped at the D stage 132, the pre-jump processing is performed. For a jump of the register indirect address or the like, pre-jumping is performed at the A stage 133. To an unconditional jump instruction having performed pre-jumping, the R code 143 and the F code 144 are not outputted, and the processing of the instruction is finished at the A stage 133.
The EIT detected at the A stage 133 includes a reservation instruction exception, a privileged instruction exception, an odd number address jump trap relating to the pre-jump, an address translation exception relating to indirect reference, a page fault exception, a bus access exception. Also, a comparison checking of the indirect reference address and the operand breakpoint is performed at the A stage 133.
(4.2.4) "Micro ROM Access Stage"
Processings of an operand fetch stage (F stage) 134 are also divided roughly into two. First is the access processing of the micro ROM, particularly referred to as an R stage 137, and second one is the operand prefetch processing, particularly referred to as an OF stage 138. The R stage 137 and the OF stage 138 do not necessarily operate simultaneously, the operation timing being different from each other depending upon miss and hit of the data cache, and miss and hit of data TLB.
The micro ROM access processing which is the processing at the R stage 137 is the micro ROM access and the microinstruction decode processing for generating the E code 145, which is the execution control code used in execution at the next E stage 135 for the R code 143.
In the case where one processing for one R code is decomposed into two or more microprogram steps, there may be a case where the first micro ROM unit 113 and the second micro ROM unit 114 are used at the E stage 138 and the next R code 143 is in the waiting state of micro ROM access. Micro ROM access for the R code 143 is performed when it is not performed at the E stage 135. In the data processor 100 of the present invention, since a number of integer operation instructions are executed in one microprogram step and many floating-point operation instructions are executed in two microprogram steps, in practice, there is a high frequency of performing of the micro ROM access sequentially for the R code 143.
When two instruction are decoded in parallel with each other and the sub-code is effective, the decode result at the sub-decoder is effective. When a processing for one R code is divided into two or more microprogram steps, the E code 145 relating to the sub-code at the last microprogram step is effective.
At the R stage 137, a breakpoint check relating to the execution PC is performed.
(4.2.5) "Operand Fetch Stage"
The operand fetch stage (OF stage) 138 executes the operand prefetch processing of the aforesaid two processings performed at the F stage 134.
At the OF stage 138, the logical address of the F code 144 is translated into the physical address by data TLB, and by the physical address, the built-in data cache or the external memory is accessed to fetch the operand, which is combined with the logical address transferred as the F code 144 and outputted as the S codes 146.
At one F code 144, the operand may cross an 8-byte boundary, but the operand fetch equal to or less than 8 bytes is fetched. At the F code 144, selection whether or not to access the operand is involved, and when the operand address itself or the immediate calculated at the A stage 133 is transferred to the E stage 135, the operand prefetch is not performed and the content of F code 144 is transferred as the S code 146.
The EIT detected at the OF stage 138 includes an address translation exception relating to operand prefetch, page fault exception and bus access exception.
(4.2.6) "Execution Stage"
The execution stage (E stage) 135 operates with the E code 145 and the S code 146 as inputs. The E stage 135 is the instruction executing stage, thus the processings performed at stages before and at the F stage 134 are all processings for the E stage 135. When a jump is performed or the EIT processing is started at the E stage 135, the processings from the IF stage 131 to the F stage 134 are all repealed. The E stage 135 is controlled by the microprogram and executes the instruction by executing a series of microinstructions from the entry address of the microprogram routine specified in the E code 145. P In the E code 145, there are a code (particularly referred to as an EI code) which controls the integer operation unit 117, and a code (particularly referred to as an EF code) which controls the floating-point operation unit 118. The EI code and EF code are able to outputted independently, and at this time, at the E stage 135, the integer operation unit 117 and the floating-point operation unit 118 operate in parallel. For example, when executing the floating-point instruction having no memory operand in the floating-point operation unit 118, this operation is executed independently from the operation of the integer operation unit 117.
In both the integer operation and the floating-point operation, read-out of the micro ROM and execution of the microinstruction are executed in the pipeline processing. Accordingly, when the branch occurs in the microprogram, there is a space of one microstep, however, it is so configurated that an effective processing can be performed in this spaced step. In the E stage 135, the write reservation for the register or memory performed at the A stage 133 is released after writing the operand.
When two instructions are decoded in parallel with each other, and sub-code is effective, the two instructions are executed in parallel by the main operation device and the sub-operation device. When a processing for one R code is divided into two or more microprogram steps, operation relating the sub-decode is executed in the last microprogram step.
Each kind of the interruption is directly received by the E stage 135 at the timing of the instruction end, and a necessary processing is executed by the microprogram. The other processings of EITs are performed by the microprogram at the E stage 135.
When the operation result must be stored in the memory, the E stage 135 outputs the W code 147 and the program counter value of the instruction performing the storing processing Lo the S stage 136.
(4.2.7) "Operand Store Stage"
The operand store stage (S stage) 136 translates the logical address of the W code 147 into the physical address by data TLB, and stores the data of the W code 147 in the built-in data cache by address obtained in the above. Simultaneously, the W code 147 and the value of the program counter are inputted to the store buffer, and processes to store the data of the W code 147 to the external memory using the physical address outputted from the data TLB.
The operation of the S stage 136 is performed in the operand access unit 120, and also the address translation processing and the permuting processing of the built-in data buffer, in the case where a miss of the data TLB or the built-in data cache occurred.
When an EIT is detected at the store processing of the operand, while holding the W code and the program counter value in the store buffer, the EIT detection is noticed to the E stage 135.
(4.3) "State control of Each Pipeline Stage"
Each pipeline stage includes an input latch and an output latch, and basically, operates independently of the other stages. Each of the stages, when the processing performed one before is completed, transfers the processing result to the input latch of the next stage from the own output latch, and starts the next processing when all of the input signals necessary for the next processing are ready in the input latch of the own stage.
That is, each stage starts the next processing when all of the input signals for the next processing outputted from the preceding stage become effective, and the processing resulting at that time point is transferred to the input latch of the later stage, and the output latch becomes empty.
In other words, all input signals must be ready at the timing immediately before each of the stages starts the operation. In the case where the input signals are not ready, this stage becomes the waiting state (input waiting). When transferring from the output latch to the input latch of the next stage, the input latch of the next stage must be empty. Even in the case where the next stage input latch is not empty, the pipeline stage is in the waiting state (output waiting). When the cache or the TLB misses or data interference occurs between the instructions being processed the pipeline, a plural number of clocks are necessitated for the processing of one stage, results in delay of the pipeline processing.
And in executing the BRA instruction or JMP instruction, when an odd number address jump trap (OAJT) is not detected, a step code after the finish of a preceding jump processing is not outputted, and an instruction processing is finished in that stage. And a NOP (no operation) instruction is finished usually in the D stage 132.
(5) "Detailed Operation of Pre-jump Processing Mechanism"
In the data processor 100 of the present invention, jump processings performed before the E stage 135 are generally called the pre-jump processings. In the following description, the instruction generating the jump is generally PG,85 called the jump instruction, and the jump instruction whose jump target address is limited to the PC relative mode is defined as the branch instruction.
As shown in Table 1, the instructions which are subject to the pre-jump processing are roughly classified into three. First one is the unconditional jump instruction which jumps surely to the designated jump target. As the competent instructions, there are BRA, BSR, JMP and JSR instructions. Second one is the conditional branch instruction which is dependent on a flag state at execution whether to branch to the branch target address or not. As the competent instructions, there are the Bcc, ACB, SCB and FBcc instructions. Third one is the return instruction from a subroutine. As the competent instructions, there are RTS and EXITD instructions.
In the data processor 100 of the present invention, all instructions shown in Table 1 re subject to pre-jump processing.
As to the unconditional jump instruction, an unconditional pre-jump is performed. Furthermore, there are the case where the unconditional pre-jump is performed at, the D stage 132 and the case where the unconditional pre-jump is performed at the A stage 133. As to the conditional jump instruction, a branch prediction is performed, and when it is predicted to branch, a conditional pre-branch is performed and when it is predicted not to branch, a branch target instruction pre-fetch is performed.
A branch prediction method of the conditional branch instruction of the data processor 100 of the present invention can be classified into three cases; predicting statically to branch, predicting statically not to branch and predicting dynamically according to the past branch hysteresis as shown in Table 2.
However, it is possible to set to predict statically that the conditional branch instruction does not branch by the control register in the integer operating unit 117.
A conditional pre-branch is performed at the D stage 132. The branch target instruction pre-fetch is performed in two cases, one at the D stage 132 and the other at the A stage 133. As to the subroutine return instruction, pre-return processing is performed in the D stage 132 using a PC stack 4 in which return target address is registered at subroutine call.
The jump target address at the time of a preceding jump processing other than a pre-returning performed in the D stage 132, is generated in the jump target address generating unit 1. To the jump target address generating unit 1, two adders for processing the BRA instruction, BSR instruction, Bcc instruction which are frequently appeared, that is, the aforementioned first adder 5 and the second adder 6 are implemented.
The second adder 6 is mainly used for calculating the branch target address of the BRA instruction, BSR instruction, Bcc instruction of the G format. In these instructions, bits (6:7) of the instruction code is a field which designates the branch displacement size. Therefore, the input latch 8 of the second adder 6 is so constructed as to select the branch displacement field according to the values of the two bits and takes in it after extending it to 32 bits.
The first adder 5 is used for calculating the branch target address of the BRA instruction, BSR instruction, Bcc instruction of the D format. In the branch instructions of the D format, since bits (6:7) of the instruction code is allocated to a field for distinguishing an instruction, the second adder 6 cannot cope with the D format. Therefore, an adder exclusive for the D format, the first adder 5 is provided.
In this way, the hardware configuration is so designed as to be able to process a jump instruction which is frequently appeared in high speed.
In addition, for an instruction in which the branch target address is designated by PC relative other than the BRA instruction, BSR instruction and Bcc instruction, in the case where an branch target address can be calculated in parallel with instruction decoding at the aforementioned first adder 5 and the second adder 6, the branch target address is calculated in that cycle, however, not in that case, the branch target address is calculated after a branch instruction is decoded and a branch displacement is cut out.
Classifications by the processing method of the jump instructions which is subject to the pre-jump processing are shown in Table 3. The method of each pre-jump processing is described in the following. However, as to the pre-return processing, since it is not so important in the present invention, it is not described particularly.
FIG. 50 is a block diagram showing, pre-jump processings between the instruction fetch unit 111, instruction decoding unit 112, PC processing unit 116, operand address calculation unit 115 and integer operating unit 117, or control signals associated with the instruction processing sequence control of the execution stage jump and the like, and relationships of giving and receiving jump target addresses.
(5.1) "D Stage Unconditional Pre-jump"
In the BRA instruction and BSR instruction, and in the JMP instruction and JSR instruction whose jump target addresses are shown by the absolute mode or PC relative indirect mode, unconditional pre-jump processing is performed at the D stage 132. In processing these instructions, the valid jump target address is generated at the jump target address generating unit 1 in an instruction decoding cycle.
The instruction decoding unit 112, when decoding these instructions, transmits a D stage unconditional pre-jump quest to the instruction fetch unit 111 by making a DJMPAREQ signal to be "1". The instruction decoding unit 112 outputs JASEL signal 313 for selecting a storing position of jump target address to be outputted to the JA bus 208 to the PC processing unit 116. The instruction fetch unit 111 arbitrates jump requests from respective stages, and when the unconditional pre-jump request at the D stage can be accepted, transmits to the instruction decoding unit 112 and the PC processing unit 116 that the D stage unconditional pre-jump request is accepted, by making a DPJMPACK signal 305 to be "1". The instruction decoding unit 112 maintains the DJMPAREQ signal 301 at "1" till the DPJMPACK signal 305 becomes "1".
When the DPJMPACK signal 305 becomes "1", in the PC processing unit 116, an output latch value in the jump target address generating unit 1 selected by the JASEL signal 313 is outputted to the JA bus 208 and transferred to the instruction fetch unit 111.
The instruction fetch unit 111 starts instruction fetching processing according to the jump target address taken in from the JA bus 208, and takes the fetched instruction into an instruction queue on the side being outputted to the instruction decoding unit 112. The instruction fetch unit 111 initializes an input/output pointer of the instruction queue and an instruction queue fetch address register, according to the jump target address taken in from the JA bus 208.
The instruction decoding unit 112 makes a PCINIT signal to be "1" by the fact that the DPJMPACK signal 305 becomes "1" so as to instruct the PC processing unit 116 to initialize the PC value. When the PCINIT signal 314 becomes "1", the PC processing unit 116 takes the value on the JA bus 208 into the TPC 27 and DPC 29 for initialization. After these processings, the instruction decoding unit 112 is in instruction waiting state and decodes the jump target instruction as soon as the valid instruction is fetched and transferred from the instruction fetch unit 111.
The instruction decoding unit 112 judges whether or not the jump address is the odd number, and transfers information showing whether a jump occurs to the odd address or not to the D code 141.
In the BRA instruction or in the JMP instruction in which the jump target address are shown in an absolute mode or PC relative indirect mode (32 bits displacement), except the case where the OAJT is detected, the D code 141 or the A code 142 is not outputted for the BRA instruction and JMP instruction and an instruction processing is finished in the D stage 132.
(5.2) "A stage Unconditional Pre-jump"
In processing the JMA and JSR instructions in which the jump target address is shown in addressing modes other than the absolute mode and PC relative mode (32 bits displacement), the unconditional pre-jump is processed at the A stage 133.
When the instruction decoding unit 112 decodes the instructions, information showing that the unconditional pre-jump is to be made at the A stage 133 is sent to the operand address calculation unit 115 as the A code 142. In this case, since the pipeline processing sequence is changed, the instruction decoding unit 112 stops the decoding processing temporarily. In processing the instructions, the jump target address is calculated in the operand address calculation unit 115. It is also possible to designate the jump target address using the chained mode.
The operand address calculation unit 115 transmits an A stage unconditional pre-jump request to the instruction fetch unit 111 by making the AJMPAREQ signal 308 to be "1" in a cycle where a valid jump target address is calculated. The instruction fetch unit 111 arbitrates jump requests from respective stages, and when the A stage unconditional pre-jump request can be accepted, transmits to the instruction decoding unit 112 and PC processing unit 116 that the A stage unconditional pre-jump request has been accepted, by making the APJMPACK signal 307 to be "1".
The operand address calculation unit, 115 maintains the AJMPAREQ signal 308 at "1" till the APJFPACK signal becomes "1". In the operand address calculation unit 115, when the APJMPACK signal 307 becomes "1", a value of the A0 latch 65 showing the jump target address is outputted to the JA bus 208 and transferred to the instruction fetch unit 111. The instruction fetch unit 111 starts instruction fetching according to the jump target address taken in from the JA bus 208, and takes the fetched instruction into an instruction queue on the outputting side to the instruction decoding unit 112. The instruction fetch unit 111 also initializes the input/output pointer of the instruction queue and the instruction queue fetch address register, according to the jump target address taken in from the JA bus 208.
The instruction decoding unit 112 restarts instruction decoding which has been interrupted, by the fact that the APJMPACK signal 307 becomes "1", and makes the PCNIT signal 314 to be "1" to instruct initialization of the PC value to the PC processing unit 116. When the PCINIT signal 314 becomes "1", the PC processing unit 116 takes a value on the JA bus 208 into the TPC 27 and DPC 29 for initialization. After these processings, the instruction decoding unit 112 is in instruction wait state, and decodes the jump target instruction as soon as the valid instruction code is fetched and transferred from the instruction fetch unit 111.
The operand address calculation unit. 115 judges whether or not jump address is the odd address, and transfers information showing whether a jump occurs to the odd address or not as the R code 143.
In processing the JMP instruction in which the jump target address is shown in the mode other than the absolute mode or the PC relative indirect mode (32 bits displacement), neither the R code 143 nor the F code 144 is outputted for the JMP instruction, and instruction processing is finished at the A stage 133.
(5.3) "D stage Conditional Pre-jump"
When it is predicted to branch in processing the Bcc instruction and when the ACB instruction and SCB instruction are processed, conditional pre-jump processing is performed at the D stage 132. In the conditional pre-jump, following processings vary according to the fact that whether or not the branch target address is calculated in which of the two adders 5 and 6 of the jump target address generating unit 1, the aforesaid calculation is performed in a cycle wherein the instruction decoding unit 112 takes in branch displacement.
When the branch target address is calculated in the cycle wherein the branch displacement is taken in, the instruction decoding unit 112, after decoding the instructions, transmits the D stage conditional pre-jump request to the instruction fetch unit 111 by making the DJMPCREQ signal 302 to be "1". The instruction decoding unit 112 outputs the JASEL signal 313 for selecting a storing position of the jump target address to be output ted to the JA bus 208 to the PC processing unit 116.
Meanwhile, when the branch target address can not be calculated in the cycle wherein the branch displacement is taken in, only cut-out processing of the branch displacement is performed in that cycle. The cut-out branch displacement is sent to the jump target address generating unit 1 via the DISP bus 205. The LDDISP signal 315 is maintained at "1". In this cycle, the addition result by the PC adder 23 is never written back to the DPC 29, which holds the PC value of the jump instruction.
The jump target address generating unit 1 calculates the branch target, address in the next cycle. By the fact that the LDDISP signal 315 is "1", the branch displacement is taken into the input latch 7 of the first adder 5 from the DISP bus 205 in place of the II bus 204. The branch displacement and the value of DPC 29 are added in the first adder 5, and the calculated branch target address value is taken into the output latch 10. The instruction decoding unit 112 transmits the D stage conditional pre-jump request to the instruction fetch unit 111, by making the DJMPCREQ signal 302 to be "1" in this cycle. A value selecting the output latch 10 is outputted to the PC processing unit 116 as the JASEL signal 313.
Processings hereinafter are same in all cases.
The instruction fetch unit 111 arbitrates jump requests from respective stages, and when the D stage conditional pre-jump request can be accepted, by making the DPJMPACK signal 305 to be "1", transmits to the instruction decoding unit 112 and PC processing unit 116 that the D stage conditional pre-jump request is accepted.
The instruction fetch unit 111 includes two instruction queues and one address saving mechanism (QINPCC) 88, and is constituted such that three sequence information related to two conditional branch instructions can be held. When three sequence information is taken in already, the D stage conditional pre-jump request is not accepted till branch predicted on hit-miss of the preceding conditional branch instruction is transmitted from the integer operating unit 117. The instruction decoding unit 112 maintains the DJMPCREQ signal 302 at "1" till the DPJMPACK signal 305 becomes "1".
When the DPJMPACK signal 305 becomes "1", in the PC processing unit 116, a value of output latch 10, 11 or 12 in the jump target, address generating unit 1 selected by the JASEL signal 313 is outputted to the JA bus 208 and transferred to the instruct ion fetch unit 111.
When the instruction fetch unit 111 is holding only one sequence information, the instruction fetch unit 111 switches the instruction queue. With maintaining the contents of the instruction queue A 83 or instruction queue B 84 used at present, values of the two registers QINPCA 86 or QINPCB 87 holding the instruction queue fetch address and an input/output pointer value of the instruction queue, the instruction fetch unit 111 starts instruction fetch according to the jump target address taken in from the JA bus 208, and takes the fetched instruction into the instruction queue on the unused side so as to output to the instruction decoding unit 112. Also, according to the jump target address taken in from the JA bus 208, it initializes the input/output pointer of the instruction queue on the unused side and the instruction queue fetch address register. The instruction fetch unit 111 fetches the sequence instruction on the branch target instruction side preferentially, and by utilizing a period during which the instruction queue is full, performs instruction fetch of the sequence on the side succeeding the conditional branch instruction.
When the instruction fetch unit 111 holds two sequence information already, the instruction fetch unit 111 saves the jump target address taken in from the JA bus 208 in the QINPCC 88 and input/output pointer unit 85, and brings switching processing of the instruction queue in the pending state. When the branch condition of the preceding conditional branch instruction is fixed and the branch prediction is true, the pending processing is restarted. When the prediction is missed, the address value having been saved in the QINPCC 88 is also voided.
The instruction decoding unit 112, by the fact that the DPJMPACK signal 305 becomes "1", makes the PCINIT signal 314 to be "1" and instructs the PC processing unit 116 to initialize the PC value. When the PCINIT signal 314 becomes "1", the PC processing unit 116 takes the value on the JA bus 208 into the TPC 27 and DPC 29 for initialization. The instruction decoding unit 112, after performing these processings, becomes the instruction code waiting state, and decodes the jump target instruction as soon as the valid instruction code is fetched and transferred from the instruction fetch unit 111.
The instruction decoding unit 112 judges whether the jump address is the odd address or not, and transfers information showing whether or not a jump occurs to the odd address as the D code 141.
In the instruction fetch unit 111, though the sequence on the predicted side is fetched preferentially till the conditional branch instruction which has performed conditional pre-jump processing is processed at the E stage 135 and the branch condition is fixed, during the period in which the instruction queue is full, the original sequence is also fetched continuously.
At processing the instruction which has performed the conditional pre-jump processing, since it is necessary to initialize the PC calculation unit 2 when the branch prediction is missed, next instruction address of the conditional branch instruction is calculated at the A stage 133. This processing is implemented by transferring the value of ATPC 28 to the integer operating unit 117 via the IX bus 206, address adder 61 and A0 bus 216.
Judging the branch condition of the conditional branch instruction which has performed the conditional pre-jump processing is performed at the E stage 135. When the branch prediction is right, the integer operating unit 117, by making the QCANCEL signal to be "1", instructs the instruction fetch unit 111 to void sequence information succeeding the conditional branch instruction. When the QCANCEL, signal 312 becomes "1", the instruction fetch unit 111 releases the instruction queue on the sequence side succeeding the conditional branch instruction. When the address has been saved in the QINPCC 88, information related to the instruction queue on the side voided by the information is initialized. Instruction being processed in the pipeline is processed continuously as it is.
When the branch prediction is missed, switching processing of the instruction queue is performed.
The integer operating unit 117, by making the QEXCHG signal 311 to be "1", informs the instruction fetch unit 111 that the branch prediction is missed. The integer operating unit 117 makes the INIT signal 300 to be "1" so as to void pipeline processings after the D stage 132, and outputs a PC value of the next instruction of the conditional branch instruction to the JA bus 208 so as to initialize the PC calculation unit 2. In respective stages after the I) stage 132 of the pipeline mechanism, initialization takes place by the fact that the INIT signal 300 becomes "1".
When the QEXCHG signal 311 becomes "1", the instruction fetch unit 111 releases the instruction queue on the branch target sequence side, fetches on the sequence side succeeding the branch target instruction and sends the instruction code to the instruction decoding unit 112. When the address has been saved in the QINPCC 88, which is also voided. In many cases the instruction code is already taken into the instruction queue on the sequence side succeeding the conditional branch instruction.
The instruction decoding unit 112 performs initialization by the fact that the INIT signal 300 becomes "1", and instructs initialization of the PC calculation unit 2 by making the PCINIT signal 314 to be "1". When the PCINIT signal 314 becomes "1", the PC processing unit 116 takes the value on the JA bus 208 into the TPC 27 and DPC 29 for initialization.
(5.4) "D Stage Branch Destination Instruction Pre-fetch"
When it is predicted that a branch does not occur by the Bcc instruction, processing of the branch target instruction pre-fetch is processed at the D stage 132. In this case, the branch target address is calculated by either of two adders 5 and 6 of the jump target address generating unit 1 in an instruction decoding cycle.
The instruction decoding unit 112 decodes the Bcc instruction, and when it has been predicted that the branch does not occur according to the past branch hysteresis, by making the DPFETREQ signal 303 to be "1", transmits the D stage branch target instruction pre-fetch request to the instruction fetch unit 111. The instruction decoding unit 112 outputs the JASEL signal 313 for selecting a storing position of the jump target address outputted to the JA bus 208 to the PC processing unit 116.
The instruction fetch unit 111 arbitrates jump requests from respective stages, and when the D stage branch target instruction pre-fetch request can be accepted, by making the DPJMPACK signal 305 to be "1", transfer to the instruction decoding unit 112 and PC processing unit 116 that the D stage branch target instruction pre-fetch request has been accepted.
The instruction Fetch unit 111 is designed to hold three sequence information related to two conditional branch instructions as aforementioned. When three sequence information is already taken in, the instruction fetch unit 111 does not accept the D stage conditional pre-jump request, till branch prediction hit-miss of the preceding conditional branch instruction is transmitted from the integer operating unit 117. The instruction decoding unit 112 maintains the DPFETREQ signal 303 at "1" till the DPJMPACK signal 305 becomes "1".
When the DPJMPACK signal 305 becomes "1", the PC processing unit 116 outputs a value of a output latch 10, 11 or 12 in the jump target address generating unit 1 selected by the JASEL signal 313, and transfers it to the instruction fetch unit 111.
When the instruction fetch unit 111 holds only one sequence information, the instruction fetch unit 111 initializes the Fetch address registers QINPCA 86 or QINPCB 87 of the instruction queue on the unused side and the input/output pointer value of the instruction queue, according to the jump target address taken in from the JA bus 208. The instruction queue A 83 or instruction queue B 84 used at present fetch the succeeding instruction and output the instruction which has been fetched in the instruction decoding unit 112. The instruction fetch unit 111 fetches the sequence instruction on the side succeeding the conditional branch instruction preferentially, and by utilizing the period during which the instruction queue is full, fetches the branch target instruction.
When the instruction fetch unit 111 already holds two sequence information, the instruction fetch unit 111 saves the jump target address taken in from the JA bus 208 in the QINPCC 88 and input/output pointer unit 85, and leaves pre-fetch processing of the instruction queue in a pending state. When the branch condition of the preceding conditional branch instruction is fixed and the branch prediction is right, the pending processing is started again. When the prediction is missed, an address value having been saved in the QINPCC 88 is also voided.
In the branch target instruction pre-fetch, since the sequence itself which performs pipeline processing does not change, the PC processing unit 116 does not initialize the PC value. The instruction decoding unit 112 processes the succeeding instruction. In this case, since the processing sequence is not-switched, it is not judged in the instruction decoding unit, 112 whether the jump address is an odd address or not.
Since it is necessary to initialize the PC calculation unit 2 when the branch prediction is missed, in the instruction having performed the branch target pre-fetch processing, the branch target address is recalculated at the A stage 133. This processing is implemented by adding a value of APC 30 inputted via the IX bus 206, and a value of branch displacement taken in from the instruction decoding unit 112 via the DISP bus 205 and DISP latch 67.
The branch condition of the conditional branch instruction which has performed branch target, instruction pre-fetch processing is judged at the E stage 135. When the branch prediction is right, the integer operating unit, 117, by making the QCANCEL signal 312 to be "1", instructs the instruction fetch unit 111 to void sequence information succeeding the branch target instruction. When the QCANCEL signal 312 becomes "1", the instruction Fetch unit 111 releases the instruction queue on the sequence side succeeding the branch target instruction. When address is saved in the QINPCC 88, information related to the instruction queue on the side voided by the information is initialized. Instruction processing being processed in the pipeline is continued as it is.
When the branch prediction is missed, the instruction queue is switched.
The integer operating unit 117, by making the QEXCHG signal 311 to be "1", informs to the instruction fetch unit 111 that the branch prediction has been missed. The integer operating unit 117 makes the INIT signal 300 to be "1" so as to void the pipeline processing after the D stage 132, and outputs the branch target address to the JA bus 208 so as to initialize the PC calculation unit 2. Respective stages performing the pipeline processing after the D stage 132 of the pipeline processing mechanism are initialized by the fact that the INIT signal 300 becomes "1".
When the QEXCHG signal 311 becomes "1", the instruction fetch unit 111 releases the instruct ion queue on the sequence side succeeding the conditional branch instruction, fetches the sequence side succeeding to branch target instruction, and sends the instruction code to the instruction decoding unit 112. When address is saved in the QINPCC 88, which is voided. In most cases the instruction code is already fetched in the instruction queue on the sequence side succeeding the conditional branch instruction.
The instruction decoding unit 112 performs initialization by the fact that the INIT signal 300 becomes "1", and by making the PCINIT signal 314 to be "1", instructs initialization of the PC calculation unit 2. When the PCINIT signal 314 becomes "1", the PC processing unit, 116 takes the value on the JA bus 208 into the TPC 27 and DPC 29 for initialization.
(5.5) "A Stage Branch Destination Instruction Pre-fetch"
In processing the FBcc instruction, branch target instruction pre-fetch processing is performed at the A stage 133. A reason why the branch target instruction is prefetched at the A stage 133 and not at the D stage 132 is that, the branch target address can not be calculated by either of the two adders 5 or 6 of the jump target address generating unit 1 in an instruction decoding cycle by the FBcc instruction.
When the instruction decoding unit 112 decodes the FBcc instruction, information showing that the branch target instruction pre-fetch is performed at the A stage 133 is sent to the operand address calculation unit. 115 as the A code 142. In this case, the instruction decoding unit, 112 inhibits to output the pre-jump processing request at the I) stage 132 to the instruction fetch unit 111, till the pre-fetch request of the instruction is accepted, so that the pre-jump processing request output Led to the instruction fetch unit 111 is in processing order of the instruction.
By the instruction performing the branch target instruction pre-fetch processing at, the A stage 133, the branch target address is calculated at the A stage 133. This processing is implemented by adding the value of APC 30 inputted via the IX bus 206 and the branch displacement value taken in from the instruction decoding unit 112 via the DISP bus 205 and the DISP latch 67.
The operand address calculation unit 115, by making the APFETREQ signal 309 to be "1", in the cycle wherein the branch larger address is calculated, transmits the A stage branch target instruction pre-fetch request to the instruction fetch unit 111.
The instruction fetch unit 111 arbitrates jump requests from respective stages, and when the A stage branch target pre-fetch request can be accepted, by making the APJMPACK signal 307 to be "1", transmits that the A stage branch target instruction pre-fetch request has been accepted to the instruction decoding unit 112 and the operand address calculation unit 115. The operand address calculation unit. 115 maintains the APFETREQ signal 309 at "1" till the APJMPACK signal 307 becomes "1".
When the APJMPACK signal 307 becomes "1", in the operand address calculation unit 115, a value of the AO latch 65 is outputted to the JA bus 208 and transferred to the instruction fetch unit 111. Processings in the instruction fetch unit 111 are same as the case of D stage branch target instruction pre-fetch. Since the PC calculation unit 2 must be initialized when the branch prediction is missed, the branch target address is also sent to the integer operating unit 117 via the A0 bus 216.
The instruction decoding unit 112, by the fact that the DPJMPACK signal 305 becomes "1", stops to inhibit the pre-jump processing at the D stage 132. At the branch target pre-fetch, since the sequence itself which performs pipeline processing does not change, the PC processing unit 116 does not initialize the PC value. In this case, since the processing sequence is not switched, in the operand address calculation unit 115 it is not judged whether the jump address is an odd address or not.
The branch condition of the conditional branch instruction having performed the branch target instruction pre-fetch processing is judged at the E stage 135. Processings at the E stage 135 are entirely same as the case of processing the D stage branch target instruction pre-fetch.
(6) "Processing Methods of Jump Instruction"
Respective processing methods for the jump instructions are particularly described in the following.
(6.1) "BRA Instruction"
Though instruction formats of the BRA instruction are shown in FIG. 26 and FIG. 27, this instruction is decoded in one cycle. Branch displacement positions on the II bus 204 at processing the BRA instruction is shown in a schematic view of FIG. 51. In FIG. 51, reference 51a shows the case of D format, and references 51b through 51d respectively show the case where the branch displacements of the G format are respectively 8 bits, 16 bits and 32 bits. As it is apparent from comparing these with FIG. 48, the branch target address can be calculated about at instruction decoding cases.
Timings at processing the BRA instruction are shown in a timing chart in FIG. 52. As shown in FIG. 52 by reference 52b, at the D stage 132, the instruction decoding unit 112 decodes the ERA instruction, and the jump target address generating unit 1 calculates the branch target address in the C1 cycle. In case of the D format, the branch target address is calculated by the first adder 5. In case of the format, the branch target address is calculated by the second adder 6.
In the G format, II(6:7) is allocated as a branch displacement size field. As aforementioned, the branch displacement size is designated 8 bit when the II(6:7) is "00", 16 bits when the II(6:7) is "01", and 32 bits when the II(6:7) is "10".
The input latch 8 takes in data obtained by making the II(24:31) sign-extended to 32 bits when II(5:7)="00", data obtained by making II(16:31) sign-extended to 32 bits when I(6:7)="01", and II(16:47) intact when 1I(6:7)="10", respectively. Accordingly, the respective proper branch displacements for all cases are sign-extended to 32 bits and taken in the input latch 8 and proper branch target addresses are calculated at the second adder 6.
Since the decoded instruction is the BRA instruction, the D stage unconditional jump processing is performed in the C1 cycle. Processing of the BRA instruction is finished at the D stage 132, and neither the D code 141 nor the A code 142 is output ted. As shown in FIG. 52 by reference 52a, a branch target instruction fetch is performed at the IF stage 131 in the C2 cycle, and at cache hit, as shown in FIG. 52 by reference 52b, decoding of the branch target instruction is started in the C3 cycle. That is, processing of the BRA instruction is finished by two cycles at cache hit.
(6.2) "BSR instruction"
Though instruction formats of the BSR instruction are shown in FIG. 30 and FIG. 31, this instruction is decoded in one cycle. Branch displacement positions on the II bus 204 at BSR instruction processing are same as the BRA instruction and are shown in FIG. 51. Also in the BSR instruction, in the same way as the case of BRA instruction, the branch target address can be calculated at instruction decoding about all cases.
Timings at BSR instruction processing are shown in a timing chart of FIG. 53. As shown in FIG. 53 by reference 53b, the instruction decoding unit 112 decodes the BSR instruction in the C1 cycle, and the jump target address generating unit 1 calculates the branch target address at the D stage 132.
Since the decoded instruction is the BSR instruction, the D stage unconditional jump processing is performed in the C1 cycle. As shown in FIG. 53 by reference 53a, the branch target instruction fetch is performed in the C2 cycle. At this time, a return target address from a subroutine is calculated at the A stage 133. This processing is performed by transferring a value of ATPC 28 in the PC calculation unit 2 via the IX bus 206 and address adder 61. As shown in FIG. 53 by reference 53e, a value of the return target address is registered in the PC stack 201 in the C4 cycle. As shown in FIG. 53 by reference 53f, the return target address is written in a cache and memory in the C5 cycle. At least two clock cycles are necessary for writing in the memory.
As shown in FIG. 53 by reference 53b, at cache hit, decoding of the branch target instruction can be started in the C3 cycle. Thus, the processing of the BSR instruction is finished by two cycles at cache hit.
(6.3) "Bcc Instruction"
Though instruction formats of the Bcc instruction are shown in FIG. 28 and FIG. 29, this instruction is decoded in one cycle. Branch displacement positions on the II bus 204 at Bcc instruction processing are same as the case of the BRA instruction, and are shown in FIG. 51. Also in the Bcc instruction, as same as the case of BRA instruction, the branch target address can be calculated about all cases at instruction decoding.
Timings at Bcc instruction processing are shown in timing charts in FIG. 54A and FIG. 54B. In the C1 cycle, the instruction decoding unit 112 retrieves a branch hysteresis table according to address of the DPC 29, and performs branch prediction of the Bcc instruction to be decoded next. In the first half of the C1 cycle, the address decoded immediately before the Bcc instruction is stored in the DPC 29. That is, branch prediction of the Bcc instruction is performed by the instruction address immediately before the Bcc instruction. Since this processing is performed before decoding the instruction, it is performed at every decoding cycle and valid information is outputted only when the next instruction is the Bcc instruction. Related to the Bcc instruction, it is also possible to set so as to predict statically that it is not branched by the control register in the integer operating unit 117. When branch prediction is not made, a value read out from the branch hysteresis table is not taken but a value showing that the branch prediction result is not forcibly branched is taken.
FIG. 54A shows a timing chart when it is predicted not to branch. As shown in FIG. 54A by reference 54Ab, the instruction decoding unit 112 decodes the Bcc instruction and the branch target calculation unit 1 calculates the branch target address in the C2 cycle. When it is predicted not to branch, as shown in FIG. 54A by reference 54Ab, D stage branch target instruction pre-fetch processing is performed in the C2 cycle. As shown in FIG. 54A by reference 54Ac, the branch target address of the Bcc instruction is calculated at the A stage 133 in the C3 cycle. The branch condition is judged in the C5 cycle, and when the branch prediction is right and the branch does not occur, the instruction fetch unit 111 voids the instruction queue on the branch target sequence side. In this case, the Bcc instruction is processed in one cycle. When the branch prediction is missed, the instruction fetch unit 111 voids the instruction queue on the sequence side succeeding the conditional branch instruction, and the processing sequence is switched. The branch hysteresis table is also updated. In the C5 cycle, the instruction address executed immediately before the Bcc instruction is stored in the OLDPC 36 of the step PC transfer unit 3, and the branch hysteresis table is updated by that value. When the branch target instruction is already fetched in the instruction queue in the C5 cycle, apparently 4 cycles have been spent for processing of the Bcc instruction.
FIG. 54B is a timing chart showing a case where a branch is predicted. As shown in FIG. 54B by reference 54Bb, the instruction decoding unit 112 decodes the Bcc instruction and the jump target address generating unit 1 calculates the branch target address in the C2 cycle. When the branch is predicted, D stage conditional jump processing is performed in the C2 cycle. As shown in FIG. 54A by reference 54Ac, in the C3 cycle, a PC value of the next instruction of the Bcc instruction is calculated at the A stage 133 by transferring the address held in the ATPC 28 at start of the cycle. The branch condition is judged in the C5 cycle, and when the branch prediction is right, the instruction fetch unit 111 voids the instruction queue on the sequence side succeeding the conditional branch instruction. In this case, processing of the Bcc instruction is finished by 2 cycles at cache hit. When the branch prediction is missed and the instruction succeeding the Bcc instruction must be processed, the instruction fetch unit 111 voids the instruction queue on the branch target sequence side and switches the processing sequence. The branch hysteresis table is also updated. Refresh processing is same as the case where it is predicted not to branch but the branch has occurred. When the instruction succeeding the Bcc instruction is fetched already in the instruction queue in the C5 cycle, apparently 4 cycles have been spent for processing of the Bcc instruction.
(6.4) "JMP Instruction"
As shown in FIG. 32, though the JMP instruction is able to designate the jump target address in a general type addressing mode, a timing for performing the pre-jump processing differs by the addressing mode.
First, the case where the addressing mode is a PC relative indirect mode (32 bits displacement) or an absolute mode is described. A schematic view of FIG. 55 shows designated positions of the branch displacement and absolute address on the II bus 204 at processing the JMP instruction. In FIG. 55, references 55a and 55b respectively show the cases where the branch displacement is 16 bits and 32 bits in the PC relative indirect mode, and references 55c and 55d In FIG. 55 respectively show the cases where address is designated by 16 bits and 32 bits in the absolute mode.
In the JMP instruction, different from the BRA instruction, II(6:7) is not allocated as the branch displacement size field. The branch displacement size when the addressing mode is the PC relative indirect mode is allocated to II(14:15) , and in case of 16-bit displacement, it is "01", and in case of 32-bit displacement, it is "10", respectively. In the JMP instruction , since II(6:7) is always "10", a proper branch target address is calculated at the second adder 6 when the branch displacement is 32 bits. But when the branch displacement is 16 bits, a proper jump target address cannot be calculated at the jump target address generating unit 1.
When addressing mode is the absolute mode, the size of the absolute address is designated by II(14:15), and in case of 16-bit displacement, it is "01" and in case of 32-bit displacement, it is "10" respectively. Since the input leach 9 takes in data obtained by making II(16:31) sign-extended to 32 bits when II(15) is "1", and takes in II(16:47) intact when "0", in case of the absolute mode, the proper jump target address is taken the input latch 9 and the output latch 12 respectively.
Accordingly, a branch target address can be calculated at the time of an instruction decoding only in case of the absolute mode and PC relative indirect mode in which a branch displacement is designated by 32 bits.
When the jump target address is shown in the PC relative indirect mode (32 bits displacement) or absolute mode, a timing at JMP instruction processing shows entirely the same processing timing of the BRA instruction shown in FIG. 52. The instruction decoding unit 112 decodes the JMP instruction and the jump target address generating unit 1 calculates the jump target address in C1 cycle at the D stage 132. When the branch displacement is 32 bits in the PC relative indirect mode, the jump target address is calculated by the second adder 6. When it is the absolute mode, the jump target address is held in the output latch 12, as the 32-bit address.
In these cases, the D stage unconditional jump processing is performed in the C1 cycle. A processing method is entirely same as the case of BRA instruction. Processing of the JMP instruction is finished at the D stage 132, and neither the D code 141 nor the A code 142 is output ted. At cache hit, processing of the JMP instruction is finished by two cycles.
Next, the case where the addressing mode is designated by the PC relative indirect mode (32 bits displacement) or mode other than the absolute mode is described. In this case, the jump target address can not be calculated in the jump target address generating unit 1. Thus, in this case, A stage unconditional pre-jump processing is performed.
A timing chart at JMP instruction processing where the jump target address is designated by the mode neither than the PC relative indirect mode (32 bits displacement) nor the absolute mode is shown in FIG. 56.
As shown in FIG. 56 by reference 56b, at the D stage 132, the instruction decoding unit 112 decodes the JMP instruction in the C1 cycle. As shown in FIG. 56 by reference 56c, the jump target address is calculated and the A stage unconditional pre-jump processing is performed in the C2 cycle. Processing of the JMP instruction is finished at the A stage 133 nd neither the R code 143 nor the F code 144 are outputted. As shown in FIG. 56 by reference 56a, the branch Larger instruction fetch is performed at the IF stage 131 in C3 cycle, and at cache hit, as shown in FIG. 56 by reference b, branch target instruction decoding is started in the C4 cycle. That is, the JMP instruction processing is finished by three cycles at cache hit. When the jump target address is designated by the chained addressing mode, since the A stage unconditional jump processing is performed in the cycle where the jump target address calculation is finished, the processing time of the JMP instruction is prolonged.
(6.5) "JSR Instruction"
In the same way as the JMP instruction, the JSR instruction is also able to designate the jump target address in a general type addressing mode, a timing for performing the pre-jump processing differs according to the addressing mode.
First, the case where the addressing mode is the PC relative indirect mode (32 bits displacement) or the absolute mode is described. In these cases, in the same way of the JMP instruction, the branch target address can be calculated at instruction decoding.
Designated position of a branch displacement and an absolute address on the II bus 204 at processing of the JSR instruction whose jump target address is indicated by the PC relative indirect mode or the absolute mode is entirely as same as at processing of the JMP instruction shown in FIG. 55.
When the jump target address is shown in the PC relative indirect mode (32 bits displacement) or the absolute mode, a timing at JSR instruction processing is entirely same as the BSR instruction processing timing shown in FIG. 53.
At the D stage 132, the instruction decoding unit 112 decodes the JSR instruction and the jump target address generating unit 1 calculates the jump target address in the C1 cycle. When the branch displacement is 32 bits in the PC relative indirect mode, the jump target address is calculated at the second adder 6. When it is absolute mode, the jump target address is held in the output latch 12.
D stage unconditional jump processing is performed in the C1 cycle. The branch target instruction fetch is performed in the C2 cycle. At this time, return target address from the subroutine is calculated at the A stage 133. Following processings are also entirely same as the case of BSR instruction. The JSR instruction processing is finished by two cycles at cache hit.
Next, the case where the addressing mode is designated by the mode other than the PC relative indirect mode (32 bits displacement) or absolute mode is described. In this case, the jump target address can not be calculated by the jump target address generating unit 1. Thus, in this case, A stage unconditional pre-jump processing is performed. In this case, also it is divided into two step codes for processing till the A stage 133, wherein it is merged in one step code and outputted.
A timing chart at JSR instruction processing where the jump target address is designated by the mode other than the PC relative indirect mode (32 bits displacement) or absolute mode is shown in FIG. 57.
As shown in FIG. 57 by reference 57b, at the D stage 132, the instruction decoding unit 112 decodes the JSR instruction on in the C1 cycle. As shown in FIG. 57 by reference 57c, the jump target address is calculated and A stage unconditional pre-jump processing is performed in the C2 cycle. In the C2 cycle, the R code 143 and F code 144 are not outputted. At the D stage, a control signal for calculating return target address in the C2 cycle is generated. As shown in FIG. 57 by reference 57c, at the A stage 133, the return target address from the subroutine is calculated in the C3 cycle is calculated, and the R code 143 and F code 144 are outputted. When the odd address jump is detected in the C2 cycle, its information is reflected to the R code 143 outputted in the C3 cycle. Following processings are entirely same as the case of the BSR information.
As shown in FIG. 57 by reference 57a, in the C3 cycle, the branch target instruction fetch is performed in the IF stage 131, and at cache hit, as shown in FIG. 57 by reference 57b, the branch target instruction decoding is started in the C4 cycle. That is, at cache hit, the JSR instruction processing is finished by three cycles. When the jump target address is designated by the chained mode, since the A stage unconditional jump processing is performed in a cycle where the jump target address calculation is finished, the processing time of the JSR instruction is prolonged.
(6.6) "ACB, SCB Instructions"
Since the ACB instruction and SCB instruction are used as a primitive of loop control, in these instruction it is predicted to branch statically to perform D stage conditional pre-jump processing. In the ACB and SCB instructions, though a number of formats are included as shown in FIG. 34 to FIG. 41, it can be roughly classified into a contracted type (Q, R formats) and a general type (G, E formats). In case of contracted type formats, decoding is finished by one cycle, but, in case of general type formats, two-cycle or three-cycle time is required for decoding one instruction even when the chained mode is not used in designating the operand. In this case, branch displacement fetch and conditional pre-jump processing are performed in the last decoding cycle.
In FIG. 58, references 58a, 58b, 58c, 58d, 58e and 58f show respectively branch displacement positions on the II bus 204 at the last decoding cycle of the ACB and SCB instructions.
In the short type format, the branch displacement size field is allocated not to II(6:7) but bit (22:23). However, in this case, the II(6:7) is always "11". As it is apparent by comparing with FIG. 48, though the branch target address can be calculated by the branch displacement field when the displacement size is 8 bits. Also, in the abovementioned instruction, that the bit (6:7) is "11" has not significance especially. Thus, in this embodiment, a possibility that the second adder 6 can calculate a branch target address is increased by performing processing in which "11" is given significance to be that the branch displacement is 8.
In case of the general format, II(6:7) is a branch displacement size field. But since a position of the branch displacement field in the case where a branch displacement is designated by eight bits is different from that of the BRA instruction or the like, a branch target address cannot be calculated in an instruction decoding cycle. Therefore, only in case of short type and eight bits of the displacement size, and of general type and 16 bits or 32 bits of the displacement size, a branch target address can be calculated in the branch displacement fetch cycle. In other cases, a branch target address is calculated using a branch displacement cut out in instruction decoding after a branch instruction is decoded.
In such a way, ways of processing differ from each other responsive to the format types and the branch displacement sizes.
In case of general type, since only the operand saving processing is performed in the processings except at the last step, the contracted type instruction is described.
First, the former case is described.
A timing chart of the ACB instruction processing when the branch displacement is 8 bits in the contracted type is shown in FIG. 59A.
As shown in FIG. 59A by reference 59Ab, the instruction decoding unit 112 decodes the ACB instruction and the jump target address generating unit 1 calculates the branch target address in the C1 cycle. D stage conditional jump processing is performed in the C1 cycle. A PC value of the next instruction of the ACB instruction is calculated in the C2 cycle. As shown in FIG. 59A by reference 59Ae, operation is performed in the C4 cycle, and the branch condition is judged in the C5 cycle. When the branch prediction is right, the instruction fetch unit 111 voids the instruction queue of the sequence on the side succeeding the ACB instruction. In this case, the ACB instruction processing is finished by two cycles at cache hit. When the branch prediction is missed and the instruction succeeding the ACB instruction must be processed, the processing sequence is returned to the origin in the C5 cycle. When the instruction succeeding the ACB instruction is already fetched in the instruction queue in the C5 cycle, apparently five cycles have been spent for processing the ACB instruction.
Next, the latter case is described. A timing chart of the ACB instruction processing when the branch displacement is 16 bits or 32 bits in the contracted type is shown in FIG. 59B.
As shown in FIG. 59B by reference 59Bb, in the C1 cycle, the instruction decoding unit 112 decodes the ACB instruction and cuts out a branch displacement field. In the C2 cycle, a PC value of the next instruction of the ACB instruction is calculated at the A stage 133. At the D stage 132, the branch target address is calculated in the jump target address generating unit 1 on the basis of the cut-out branch displacement so as to perform the D stage conditional jump processing. As shown in FIG. 59B by reference 59Be, operation is performed in the C4 cycle and the branch condition is judged in the C5 cycle. When the branch prediction is right, the instruction fetch unit 111 voids the instruction queue of the sequence on the succeeding side of the ACB instruction. In this case, the ACB instruction processing is finished by three cycles at cache hit. When the branch prediction is missed and the instruction succeeding the ACB instruction must be processed, the processing sequence is returned to the origin in the C5 cycle. When the instruction on succeeding the ACB instruction is already fetched in the instruction queue in the C5 cycle, apparently five cycles have been spent for processing the ACB instruction.
In either of the cases, only when the branch target address is the odd number and the branch occurs practically, the odd address jump trap processing is started after finishing the ACB instruction processing.
(6.7) "FBcc Instruction"
Though the FBcc instruction is decoded in 1 cycle, as the case, where sy="00": 8 bits, sy="01": 16 bits and sy="10": 32 bits, is shown in schematic views of FIG. 60 by references 60a, 60b and 60c, regardless of the branch displacement size, the branch target address can not be calculated in the decoding cycle. Processing is performed by predicting that there is no static branch in this instruction. In the conditional branch instruction, since a rapid conditional judgment is important and processing is performed by predicting not to branch, the step code is advanced to perform the branch target instruction prefetch processing at A stage 133.
FIG. 61 shows a timing chart of the FBcc instruction processing. As shown in FIG. 61 by reference 61b, at the D stage, the instruction decoding unit 112 decodes the FBcc instruction in the C1 cycle. As shown in FIG. 61 by reference 61c, in the C2 cycle, the branch target address is calculated and A stage branch target instruction pre-fetch processing is performed at the A stage 133. The branch condition is judged in the C4 cycle. When the branch prediction is right, the instruction fetch unit 111 voids the instruction queue of the branch target sequence. In this case, the FBcc instruction is processed in one cycle. When the branch prediction is missed and the branch occurs, the processing sequence is switched in the C4 cycle. When the branch target instruction is already fetched in the instruction queue in the C4 cycle, apparently four cycles have been spent for processing the FBcc instruction.
In either of the cases, only when the branch target address is the odd number and the branch occurs practically, the odd address jump trap processing is started after finishing the FBcc instruction processing.
In the aforementioned embodiment, two adders for calculating a branch target address at the time of the instruction decoding, that is, the first adder 5 and the second adder 6 are implemented so that the BRA instruction, BSR instruction and Bcc instruction which are frequently used especially are main objects to be processed at the data processor, in order to improve the performance with smaller hardware cost. But it is a mater of course that such a configuration that more adders are implemented corresponding to the JMP instruction, ACB instruction, SCB instruction of 16 bits displacement, or a branch displacement of the FBcc instruction is also allowed to be adopted. And such a configuration that only one of the first adder 5 or the second adder 6 is implemented is also allowed to be adopted.
And in the aforementioned embodiment, it is so configuration that, a calculation of a branch target address of the case where the branch target address cannot be calculated in parallel with branch instruction decoding, however, it is also a matter of course that the calculation is performed in the second adder 6. In addition, the DISP bus 205 is connected to the input latch 8 not to the input latch 7, and is selectively latched by the input latch 8.
As aforementioned, according to the first and second inventions of the data processor of the present invention, since calculations of branch target addresses of many branch instructions can be performed with smaller quantity of hardware when compared with the conventional data processor in parallel with instruction decoding, a data processor of high performance capable of pipeline processing branch instructions at high speed can be obtained at low cost.
According to the third invention of the data processor of the present invention, since absolute address fields of plural jump instructions by which jump target addresses are designated by absolute addresses can be cut out with smaller quantity of hardware when compared with the conventional data processor in parallel with decoding an instruction, a data processor of high performance capable of pipeline processing of a jump instruction at high speed can be obtained at low cost.
As this invention may be embodied in several forms without departing from the spirit of essential characteristics thereof, the present embodiment is therefore illustrative and not restrictive, since the scope of the invention is defined by the appended claims rather than by the description preceding them, and all changes that fall within meets and bounds of the claims, or equivalence of such meets and bounds thereof are therefore intended to be embraced by the claims.
Claims
What is claimed is:
1. A data processor for processing i. a first branch instruction including an operation code field and a first branch displacement field and branching to an address obtained by adding a value of said first branch displacement field to a head address thereof, ii. a second branch instruction including an operation code field and a second branch displacement field, different from said first branch displacement field, and branching to an address obtained by adding a value of said second branch displacement field to a head address thereof, and iii. a third branch instruction including an operation code field and a third branch displacement field and branching to an address obtained by adding a value of said third branch displacement field to a head address thereof, comprising: a memory for storing a plurality of instructions including said first branch instruction, said second branch instruction and said third branch instruction; an instruction fetch unit for fetching instructions from said memory; an instruction decoding unit for decoding instructions fetched by said instruction fetch unit; an instruction execution unit for executing the instructions decoded by said instruction decoding unit; a program counter unit for holding a head address of an instruction being decoded by said instruction decoding unit; and a jump target address generating unit for determining a branch target address every time an instruction is decoded, wherein an instruction transmitted from said instruction fetch unit, an instruction of PC value transmitted from said program counter unit, an instruction for generating branch target in said jump target address generating unit, and an instruction now being decoded in said instruction decoding unit are identical, said jump target address generating unit including first adding means for calculating a first address by adding a value of one of said first branch displacement field and said second branch displacement field from said instruction fetch unit according to a value of a part of the field of the operation code field transferred from said instruction fetch unit, and a head address of said instruction from said program counter unit, in parallel with decoding processing of the instruction by said instruction decoding unit, second adding means for calculating a second address by adding a value of said third branch displacement field from said instruction fetch unit and a head address of said instruction from said program counter unit in parallel with decoding processing of the instruction by said instruction decoding unit, and transferring means for transferring one of the first address and the second address to said instruction fetch unit, wherein when an instruction being decoded at said instruction decoding unit is one of said first branch instruction and said second branch instruction, said jump target address generating unit transfers the first address calculated by said first adding means to said instruction fetch unit through said transferring means, and said instruction fetch unit fetches an instruction of the first address transferred from said jump target address generating unit. when an instruction being decoded at said instruction decoding unit is said third branch instruction, said jump target address generating unit transfers the second address calculated by said second adding means to said instruction fetch unit through said transferring means, and said instruction fetch unit fetches an instruction of the second address transferred from said jump target address generating unit.
2. A data processor for processing a first jump instruction designating a jump target by a first absolute address field and a second jump instruction designating a jump target by a second absolute address field, comprising: a memory for storing a plurality of instructions including said first and second jump instructions; an instruction fetch unit for fetching instructions from said memory; an instruction decoding unit for decoding instructions fetched by said instruction fetch unit; an instruction execution unit for executing the instructions decoded by said instruction decoding unit; and a jump target address generating unit for determining a jump target address every time an instruction is decoded, wherein an instruction transmitted from said instruction fetch unit, an instruction for generating jump target in said jump target address generating unit, and an instruction now being decoded in said instruction decoding unit are identical, said jump target address generating unit including field cutting out means for generating an address by executing a processing of selectively cutting out one of a value of said first absolute address field and a value of said second absolute address field from an instruction data transferred from said instruction fetch unit according to a value of a part of the field of the operation code field transferred from said instruction fetch unit, in parallel with decoding processing of the instruction by said instruction decoding unit, and transferring means for transferring the address cut out by said field cutting out means to said instruction fetch unit, wherein when an instruction being decoded at said instruction decoding unit is one of said first jump instruction and said second jump instruction, said jump target address generating unit transfers the address generated by said field cutting out means to said instruction fetch unit through said transferring means, and said instruction fetch unit fetches an instruction of the address transferred from said jump target address generating unit.
Patent Citations (6)
Non-Patent Literature (1)
- The GMicro/100 32-Bit Microprocessor, IEEE Micro vol. 11, No. 4, pp. 20-23, 62-72, Aug. 1991.