WO1983001324A1

WO1983001324A1 - Flexible architecture for digital computers

Info

Publication number: WO1983001324A1
Application number: PCT/GB1981/000206
Authority: WO
Inventors: Corporation Burroughs; Machines Limited Burroughs; Hanan Potash; Lee Levin Burton; Melvyn Ernest Genter
Original assignee: Burroughs Corp; Burroughs Machines Ltd
Priority date: 1981-09-29
Filing date: 1981-09-29
Publication date: 1983-04-14

Abstract

Flexible architecture for digital computers which can be adapted to meet the many different functional requirements of several computer models. Each model is comprised of an array of sequential logic units. These units include respective control memories for storing commands, means for sequentially fetching and executing selectable sequences of the commands, and soft functional structures for performing customized functions in response to the commands. Included within the soft functional structures are a plurality of selectable electrical contacts which customize the functional response of the structures to the commands. Except for these contacts and the content of the respective control memories, the units are substantially identical. All of the units in the array execute respective command sequences from their control memory to perform a single instruction for the computer model.

Description

TITLE

FLEXIBLE ARCHITECTURE FOR DIGITAL COMPUTERS

BACKGROUND OF THE INVENTION

This invention relates to the architecture of digital information processing equipment; and more particularly to the architecture of digital computers.

Typically, a manufacturer of digital computers produces not just one computer type; but instead, it produces several different varieties, or models. These models vary substantially in processing power and price. Basically the various models are needed because customer requirements vary widely. Thus, IBM has produced system 360 models 20, 30, 40, 50, 65, 75, and 90, and now produces system 370 models 125, 135, 138, 145, 148, 155, 158, and 168. All other major manufacturers of digital computers also produce several computer models.

In the past, the various computer models of any one particular manufacturer differed substantially from each other in their architecture. Compare for example, the architecture of the above IBM models 40 and 50 as illustrated at pages 221 and 297 of Microprogramming

OMPI Principles and Practice, Sa ir S. Husson, Prentice-Hall Inc., 1970. Compare also the architecture of the NCR Century 100, 200, 3.00, or the architecture of the Burroughs 4800 and 6800. Each model has its own unique data paths, its own unique functional logic, etc.

From a design engineer's point of view these differences in architecture pose no problem; all that is relevant is whether each model meets its own functional requirements. But from a manufacturing point of view, each model essentially is a separate ensemble of unique parts. Thus,- little or no economy is achieved through commonality in design, in fabrication, or in inventory of parts for the various models.

This dissimilarity between models can place very severe strains on a manufacturer's financial resources; because essentially separate design cycles, separate production facilities, and separate inventory need to be provided for each model. And typically, for^" any one model, these items can cost several million dollars.

Further, from an integrated circuit (Υ manufacturer's point of view, the problem is even more severe. It must supply IC's which meet the diverse functional requirements of the computer models from several computer manuf cturers. And typically, no commonality in architecture exists between computer models from different manuf cturers. Compare for example, the IBM 370 models to the Burroughs 6800.

Accordingly, a primary object of^■ this invention is to provide an improved architecture for digital information processing equipment which uses "standardized" parts and is flexible enough to adapt to the different functional requirements of any computer model. BRIEF SUMMARY OF THE INVENTION

These and other objects are accomplished in accordance with .the invention by an architecture for multiple digital computer models wherein each model is comprised of an array of sequential logic units. These units include respective control memories for storing commands, means for sequentially fetching and executing selectable sequences of the commands, and soft functional structures for performing customized functions in" response to the commands. Included within ; the soft functional structures are a plurality of selectable electrical contacts which customize the functional response of the structures to the commands. Except for these contacts and the content of the respective control memories, the units are substantially identical. All of the units in the array execute respective command sequences from their control memory to perform a single instruction for the computer model. BRIEF DESCRIPTION OF THE DRAWINGS Various features and advantages of the invention will best be understood by reference to the following detailed description and accompanying drawings wherein:

Figure 1 illustrates an embodiment of one digital computer model which is constructed in accordance with the invention.

Figure 2 illustrates an embodiment of another digital computer model which is constructed in accordance with the invention. Figure .3 is a timing diagram illustrating the sequence of which the Figure 1 embodiment performs a B6800 STOD instruction. Figures -4a-4b are timing diagrams illustrating the sequence by which the Figure 2 embodiment performs a B4800 ADD instruction.

Figure 5 illustrates one preferred embodiment of a sequential logic unit for incorporation within the digital computer models of Figures 1 and 2.

Figure 6 illustrates another preferred embodiment of a sequential logic unit for incorporation within the digital computer models of Figures 1 and 2. Figure 7 is a detailed circuit diagram of a programmable multiplexor in. the sequential logic unit of Figure 5.

Figure 8 is a detailed circuit diagram of a programmable memory in the sequential logic units of Figures 5 and 6.

Figure 9 is a detailed circuit diagram of a programmable interconnect matrix in the sequential logic unit^*of Figure 6.

Figure 10 is a chart illustrating the relation between the circuits of Figures 7, 8, and 9, and the timing diagrams of Figures 3 and 4. DETAILED DESCRIPTION

Referring now to Figures 1 and 2, the details of two digital computer models which are constructed in accordance with the invention will be described. One of the models is indicated in Figure 1 by reference numeral 10; and the other model is indicated in Figure 2 by reference numeral 20.

In operation, model 10 performs all of the instructions of the Burroughs B6800 digital computer.

These instructions are described in the Burroughs B6800 Information Processing System Reference Manual; and all of the information there contained is herein incorporated by reference.

OMPI By comparison, model 20 performs all of the instructions of the Burroughs B4800 digital computer. Those' instructions- are described in the Burroughs B4800 Information Processing System Reference Manual; and the information there contained is also herein incorporated by reference.

These two instruction sets (i.e. - the instruction sets of the B6800 and the B4800) are totally different and unrelated to each other. For example, the B6800 is a stack oriented processor; while the B4800 is a three address .machine. Most of the B6800 operands are received from the stack; and most of its arithmetic results are stored back onto the stack. By comparison, the B4800 receives its operands from and stores its arithmetic results in main memory.

Also, the B6800 executes many of its instructions in conjunction with a variety of "information words" which are specially defined to aid in the processing of ALGOL statements. These information words are termed Data Descriptors (DD) ,

Indexed W^rord ^'Data -Descriptors (IWDD),- Ijidexed String Data Descriptors (ISDD) , Normal Indexed Reference Words (NIRW) , and Stuffed Index Reference Words (SIRW) . By comparison, no such information words are processed by the B4800 processor.

Furthermore, in the B6800 processor, many fields are formatted in 48 bit words that have a 3 bit identification tag and a parity bit appended to them. But in the B4800 processor, fields are formatted in groups of six-"4-bit digits and they contain no tag.

Now, ^' in accordance with the present invention, the B6800 and B4800 instruction sets are processed by

OMPI WIPO respective arrays^" of sequential logic units. Computer model 10 contains sequential logic units lOa-lOk; and computer model 20 contains sequential logic^' units 20a-20d. These units are interconnected via busses as illustrated in Figures 1 and 2.

Each sequential logic unit includes a control memory for -storing commands, and a means for sequentially fetching and executing selectable sequences of those commands. Also^", each unit includes thousands of selectable electrical contacts which define various "soft functional structures" .that give the unit its personality. That is, they define the type of functions which the unit can perform; and they also define the interconnecting data paths within the unit. These soft functional structures will be described in great detail in conjunction with Figures 5-10. For the moment however, suffice it to say that by appropriately choosing the selectable electrical contacts, each sequential logic unit can have its function performing capability tailored to particular tasks; and this tailoring enables the unit to perform its tasks very quickly.

Also, except for the content of their respective control memory and the selectable electrical contacts, each of the units in computer models 10 and 20 are identical to each other. This, of course, is highly desirable because it achieves tremendous economy in design, fabrication, and inventory.

In computer model 10, sequential logic units lOa-IOk are assigned the following tasks. Unit 10a basically operates as a stack simulator. It keeps the top two words of the stack in its registers; and it stores- all other stack words in memory. But,

^ E ,

OMPI as far as the rest of the array is concerned, all of the stack "resides" in unit 10a. Each word is 52 bits in length; and one word is partitioned^' into several registers. Unit 10a also performs simple operations on the top of stack items. For example, it "pops" items off of the stack and sends them on the S-bus to other units. Also, it receives items from other units and "pushes" them onto the stack. Further, since -only the top two words of the stack are held within unit 20a," it performs memory reads and .writes to add and delete stack items.

Unit 10b by comparison, performs condition checking on the action taken by unit 10a. For example, unit 10b contains top of stack and bottom of stack address registers; and it utilizes the contents of those registers to determine if the unit 10a memory requests are within bounds. If those bounds are exceeded, a stack overflow or a stack underflow results; and unit 10b signals this condition to unit 10a.

Unit 10c basically operates to evaluate all "Descriptors". That is, it evaluates Data Descriptors, Indexed Word Tja-ta Descriptors, .and Indexed String Descriptors Typically, this evaluation involves some arithmetic operations on various bits in the descriptor and results in generating a memory address. Then, unit 10c utilizes the memory address to read or write an item from memory.

Unit lOd checks for various conditions which are associated with the Descriptors. For example, it determines whether the memory addresses which unit 10c forms are within the memory bounds. Also, it determines whether the items which are read from main

- fRiXt

OMPI memory are tagged^" as expected. For example, an item should be tagged as data if the B6800 instruction that is being performed" is an arithmetic instruction.

Unit lOe evaluates all of the above mentioned Reference Words. These include Normal Indirect Reference Words, and Stuffed Indirect Reference Words. Also, it evaluates an Address Couple (AC) in the Value Call instruction. To make these evaluations, unit lOe also keeps, track of the current lexographical level in the program being run. Also,- it contains a plurality of D registers which keep track of the beginning of procedures in the program. Unit lOe also reads and writes items in memory which are addressed by the Reference Words. Unit lOf performs most of the condition checking associated with the Reference Words. For example, it checks address bounds for addresses which are generated by NIRW's, SIRW's, and AC's as described above. Also, it checks the tag on the data which unit lOe reads from memory to determine if it is the expected type.

Unit lOg has the task of fetching instructions from memory, reformatting them so they are directly usable by the other units, and broadcasting the reformatted instruction to all of the units. Also, unit lOg keeps track of the memory address for the next instruction.

It also performs some branch instructions directly; and it keeps track of miscellaneous processor states (such as a carry flip-flop, a true/false flip-flop etc.) which are changed upon the execution of an Enter or Exit instruction.'

Unit lOh carries out most of the arithmetic steps in an arithmetic instruction. For example, it performs a plurality of add and subtract steps in executing a B6800 Multiply or Divide instruction. Operands for those instructions are supplied by units 10a, 10c, or lOe. •

Unit lOi checks the action which is taken by 5 unit lOh. For example, it determines if the data which unit lOh is being sent as an operand is double precision or single precision. Single precision operands are used much more frequently than double precision operands; and unit lOh can perform its arithmetic operations 10 ^■ much more quickly if it simply assumes that it is operating on single precision, operands. Unit lOi then signals unit lOh if the operands are double precision.

Unit lOj performs substantially all of the bit oriented instructions. These include Bit Set/Reset, 15 Transfer While Greater or Equal, Transfer While Greater Destructive, etc. Thus, the soft functional structures within unit lOj are tailored to perform bit manipulations such as shifting and rotating.

Unit 10k provides masks and literals which 20 are used in conjunction with the execution of a variety of B6800 instructions. These instructions include for example, all Field Transfers and Field Inserts.

From the above description, it should be evident that units lOa-lOk each have totally different 2.5 functional requirements. And these different requirements can only be performed efficiently by tailoring the functional capability of each unit.

Also, the functional requirements of units lOa-lOk are totally .different than the functional 0 requirements of units 20a-20d in computer model 20. In that model,- unit 20a has the task of fetching instructions from the memory. Then, it alternately sends the fetched instruction to either unit 20b or 20c. Also,

^S A

OMPI ^" ΪPO^" -_*V these B4800 instructions vary in length; and thus during the instruction fetching operation, unit 30a must determine each instruction's format.

Units 20b and 20c compute the operand addresses for the instructions that they receive. These address calculations can involve index registers, or indirect addressing. ^' In either case, memory reads need to be performed. Also, units 20b and 20c determine the length of the operands; and this determination may also involve addressing the memory.

Finally, unit 20d alternately receives instruction opcodes and operand addresses from units 20b and unit 20c; and it performs arithmetic operations on the addressed operands. Unit 20d also reads the addressed operands from memory and stores the result back in memory.

Turning now to Figure 3, the manner in which logic units lOa-lOk perform B6800 instructions will be described in further detail. In particular, that Figure illustrates the command sequences which logic units lOa-lOk simultaneously perform to execute one B6800 STORE DESTRUCTIVE (STOD) instruction.

However, in order to even begin to understand the illustrated command sequences, and the functions which each unit is performing, it first is necessary to further explain Data Descriptor, Indexed Word Data Descriptor, Indexed String Data Descriptor, Normal Indirect Reference Word, and Stuffed Indirect Reference Word. Each of these items are fifty-two bits long; and they vary the particular action that is taken during the instruction's execution.

Basically, a Data Descriptor defines an array of data. Bits 19-0 specify the base addresss of the

^r - MPI array; and bits 39-20 specify the number of items in the array. Each of these items can be either a single precision word, a double precision word, a Hexidecimal character, or an EBCDIC character. These are respectively specified by bits 42-40 being equal to 0, 1, 2, or 4. Bit^' 43 being equal to 1 indicates the array can only be ^'read from but not written into. Bit 47 being equal to 1 indicates the array is in main memory. And bits 50, 49., 48, and 45 are a code which identifies the Data Descriptor.

An^': Indexed Word Data Descriptor points to one particular item in an array of single precision or double precision words. Bits 19-0, 43, and 47 are as defined above. Bits 32-20 specify the number of the referenced item in the array relative to the base address. Bits 50, 49, 48, 45, 42, 41, and 40 are codes which identifies--._.the -Indexed

and its content as being either single or double precision. Similarly, an Indexed String Data Descriptor points to one particular item in an array of Hexidecimal or EBCDIC characters. Bits 50, 49, 48, 45-, 42, 41, and 40 are codes which identify the indexed string data descriptor and its content as being either Hexidecimal or EBCDIC. And bits 19-0, 43, 47, and 32-20 are as defined for tτie indexed word descriptor.

A Normal Indirect Reference Word, and a Stuffed Indirect Reference Word both specify a memory address. They are identified by a code of bits 50, 49, 48, and 46. .In a Normal Indirect Reference Word, the memory address is the content of a "D" register (as specified by one portion of bits 13-0) plus an offset (which is specified by another portion of bits 13-0) .

- REAyf

OMPI One "D" register exists for each lexographical level in the program to be executed.

With a Stuffed Indirect Reference" Word, the memory address is the content of a base register plus an offset value plus a displacement value. Bits 12-0 and 35-16 respectively define the offset and displacement.

Now in the STOD instruction itself, a reference chain (which begins in the stack) is evaluated in order to store .some item from the stack (the store object) into a data word targ.et location in main memory. The initial reference chain item is either an IRW chain or an IWDD; and the data word target location in main memory is either a data type operand, tag 4 word, or an uninitialized operand. An IRW chain, in turn, can point to an IWDD, a PCW or a target item. Evaluation of an IWDD will result in either another IWDD or a target item. Evaluation of a PCW can again point to an Initial Reference item, which must then be evaluated as described above. \-

The Initial Reference is assumed to be the top item of the stack and the store object the second item. But if the top item is a Data Word, (a word with tag bits 50, 49, and 48 being even) the second item is assumed to be the Initial Reference. Note that a Data Word store object and the Initial Reference may be in either order, but if the store object has an odd tag, the Initial Reference list will be the top item and the store object the second item.^" If the -top of stack item is not a Data

Word or an Initial Reference, or if the top item is a Data Word and" the second item is not an Initial Reference, an Invalid Stack Argument interrupt is

OMPI generated. If any^" reference evaluation produces a tag 3 item or an IWDD is marked read-only, a Memory Protect interrupt is generated. If reference evaluation produces an item otherwise not a valid result according to the above chain evaluation rules, an Invalid Reference Chain interrupt is generated.

The store object is written into the target location. Note that normal store evaluation operators will not write ^'into^"a"^'Location containing an odd tagged word. Both,the. Initial Reference and the store object are deleted from the stack. <

Type conversion between double precision operands and single word items (single precision operands, tag 4 words, uninitialized operands) depends on the type of the store object (the store type) and the type associated with the target location (the target type). The target type is determined as follows: if one or more IWDD's are evaluated, the target type is the element size value of the last IWDD; otherwise, the target type is single word if a single word item is currently stored in the target location and double precision if a double precision operand is in the target location.

If the store type Is double precision and the target type is single word, a SNGL (set to single precision, rounded) operation is performed on the store object, and the resultant single precision operand is stored into the target location.

If the store type and the target type are double precision,' both words of the store object are stored into the target locations. If the store type is single word and the target type is double precision, the single word store object is extended to double precision by changing its tag and appending a second word initialized to zero. Both words of the pair are stored into the target locations. Where two double precision words are written, if the second (adjoining) target location contains an odd tagged word, a Memory Protect interrupt is generated.

Consider now the command sequence of Figure 3. In that figure, columns lOa-l through lOk-1 respectively indicate the command sequences which units lOa-lOk p.erform in .executing the STOD instruction. Commands shown In brackets ar.e executed by units 10b, lOd, and lOf. Time intervals tl-tl3 indicate the sequences of these commands.

During time interval tl, unit lOg determines that an instruction execution should begin. That instruction is then broadcast on the system bus during time interval t2; and from there it is received by all of the units.

Thereafter, during time interval t3, all of the receiving units decode the STOD instruction to determine which instruction it is. Based on this decode, units lOh, lOi, 10 and 10k determine that they are not involved with the" execution of this particular Instruction; and they -therefore suspend further command execution until the next instruction is broadcast by unit lOg on the system bus.

By comparison, units lOa-lOf each determine that they have further command sequences to perform in the execution of this instruction. In particular, unit 10a has the-task of sending the top two words in the stack to units lOc-lOf. These two words contain the initial reference and the store object as was described above. Thus, during time intervals t4, t5, - _/_* t6, and t7, unit 10b uses the S bus to transmit the least significant half of the top of stack (TOS), the most significant half of the TOS, the least significant half of the TOS-1, and the most significant half of the TOS-1.

Units lOc-lOf receive this data from the S bus during time intervals t5-t8. If the TOS contains the store object and the TOS-1 contains the data descriptor, then this instruction is performed by units lOe and lOf.- However, if the TOS contains an IRW^' and the TOS-1 contains the store object, then this instruction will be performed by units 10c and lOd.

Also, unit 10c executes its commands without first checking to see if the store object is in the TOS and the data descriptor is in TOS-1. Instead, the check for those conditions is performed by unit lOd. If they do not exist, then unit lOd interrupts the command sequence of unit lOe. By this mechanism, the execution time of the instruction is shortened over that of which it would be if^" the checks were made before the taking of any substantive action.

Similarly, unit lOe performs its command sequence under the assumption that the TOS contains an IRW and the TOS-1 contains the store object. At - the same time, unit lOf performs various checks to determine whether this condition in fact exists. If it does not, then unit lOf interrupts the command sequence of unit lOe.

Thus, during time interval t5, unit lOe partitions the least significant half of the TOS word as if it were an address couple. Then during time interval t6, unit lOe sends a read/write command on the M bus to the memory. That memory address

OMPI WIPO equals the D register (as specified by the A portion of the above address couple) plus the portion of that address couple. Then during time intervals' t7 and t8, unit lOe sends the TOS-1 word as write data to the memory.

All of.these actions, however, of unit lOe are dependent upon the checking that is performed by unit lOf. This checking is indicated in Figure 3 within brackets. Thus, during time interval t6, unit lOf checks the .tag bits o.f the top of stack word to determine if in fact it is a .store ob ect. And during time interval t7, unit lOf checks the TOS-1 word to determine if in fact it is a data descriptor. If they are not, then unit lOe is interrupted and another command flow sequence (not shown) is executed.

Based on the analysis of various programs, the probabilities are very low that the conditions for which unit lOf checks will occur. Thus in normal case, unit lOe will not be interrupted, and therefore the STOD Instruction will be executed relatively quickly. Unit lOf also must perform its checking for abnormal conditions relatively quickly. But those STOD instruction variations which unit lOf checks for are actually performed by another flow sequence in a much slower fashion.

Continuing on now with the normal flow sequence of unit lOe, that unit waits for the data which was previously stored at the memory location to which it. wrote. That data is sent hγ the memory control unit on the M bus during time Intervals tlO and til-. Then during time interval til, unit IQe compares the tag of the data which it wrote to the tag of the data which is read to determine whether or

OMPI not any "type conversion" as described above is required. Also, a determination is made as to whether, the tag of- the store location Is valid. If these conditions are met, unit lOe sends an op complete code to unit lOg during time interval tl2. Unit lOg then responds to this op complete message by broadcasting the next high level instruction to all of the units as has been described above.. That next instruction is formatted during time intervals..t2-t4. A .memory read is also performed by unit lOg during those time intervals if it is needed to obtain the next instruction.

Turning now to the operation of units 10c and lOd, Figure 3 shows that during time interval t5, unit 10c loads the least significant half of the TOS into a register TEMPI. Then during time interval t6, unit 10c loads the most significant half of the TOS into another register TEMP2, Next, during time interval t7, unit 10c checks the most significant half of the TOS-1 word to determine if it is a data descriptor. Also, unit lOd checks TEMPI to determine if it is a valid store object. In the illustrated example, these conditions do not exist, and thus units 10c.and lOd suspend further operation until the next instruction is broadcast by unit lOg.

Next, referring to Figure 4, the detailed microcode flow for one of the instructions which the Figure 2 embodiment performs will be- described. That instruction is a B4800 ADD instruction; and its format is as follows. Digits 1 and 2 are the OPCODE; digits

3 and 4 are an AF field; digits 5 and 6 are a- BF field; digits 7-12 are an A address field; and digits 15-18 are a B address field.

OMPI Basically, in response to this instruction, the contents of memory at the A address are added to the contents of memory at the B address; and the result is stored in memory at the B address. However, several variations on this basic operation are possible. For example, if the two most significant bits of digit 7 equal 01, 1.0, or 11, then the contents of index register 1, 2, or 3 respectively must be added to the A address. No indexing occurs if those bits equal 00. Also, ^'if the two least significant bits of digit 7 equal 00, then the data at the A address is treated as unsigned four bit data; but if those same two bits equal 01, then the data at the A address is treated as signed four bit data; and if they equal to 10, then the data at the A address Is treated as unsigned eight bit data.

Further, if the two least significant bits of digit 7 are equal to 11, then the data at the A address is not an operand; but instead, it is the address of n operand. This is called indirect addressing; and it can be repeated to any. level.

Similar variations also occur on the £ address. That is, the two most significant bits and the two least significant bits of digit 13 are interpreted as described above, but only with reference to the B address.

Also, the AF field modifies the meaning of the A operand. Normally, digits 3 and 4 are decimal numbers which range between 0 and 9; and in that case they specify the length of the A operand. ^' However, If the two most significant bits of digit 3 equal 11, then an indirect- field length is specified; and the address • of that indirect field length is formed by adding the two least significant bits of digit 3 (as a ten's digit) _#•

with digit 4 (as -a units digit) and with the contents of a base register.

Further,, if digit 3 equals 1010, 'then the A address field is not interpreted as a memory address; but instead, it is interpreted as a literal. In that case, digits 7-12 form A-operand of the ADD instruction. Also, as before, the BF field is interpreted similarly to the AF field,- but only with reference to the B operand. Now, to understand how this instruction is performed by the array of Figure 2, consider the chart of Figure 4. In that chart, column 20a-l lists the microcommands which unit 20a performs. One microcommand is performed during time interval tl; another microcommand is performed^'during time interval t2; etc. Similarly, columns 20b-l, 20c-l, and 20d-l respectively indicate the microcommand sequences that are performed by units 20b, 20c, and 20d.

To begin the ADD instruction, unit 20a issues a fetch command to the memory controller. This occurs during time interval tl. In response thereto, the memory controller fetches and sends the instruction back to unit 20a on memory bus 1. Time intervals t2-t6 are utilized by the memory controller to perform this operation.

During time interval t5, unit 20a receives the OP, AF, and BF portions of the instruction. Also in that time interval, unit 20a sends these fields to address calculator unit 20b; and it further decodes these fields to determine the instruction's format.

Next, during time interval t6, unit 20a receives the A-address field and passes it to address calculator unit 20b. Then during time interval t7,

OMPI /,. WIPO ,« unit 20a receives- the B-address field, and sends it to address calculator unit 20b. Also during time interval t7, unit 20a updates the instruction address pointer based on the ADD Instruction's format. Then, during the next time interval, unit 20a begins to fetch the next instruction. iri sequence.

Unit 20b then begins its execution of the ADD instruction during time Interval t7. During that time interval unit 20b receives the OP, AF, and BF portions of the. instruction. Also, it decodes those instruction portions to determine the Instruction's format, and branches to a corresponding routine.

Then during time Interval t8, unit 20b receives the A-address field of the instruction; and during time interval t9, it receives the B-address field of the instruction. Also during time interval t9, unit 20b decodes the AF field and the digit 7 of the A-address field. Based on that decode, a determination is made that indexing is required; so the specified index register is read from memory.

Similarly, during time interval tlO, unit 20b decodes the BF and digit 13 of the B-address field. Based on that decode, a determination is made that an indirect field length is specified. Accordingly, unit 20b forms a memory address and performs a memory read to fetch the actual length of. the B field.

During time interval til, the index register is received; and during time interval tl2, the actual B field length is received. Also during time interval tl2, unit 20b adds the index register to the A-address field. .-

Next, during time interval tl3, unit 20b sends the address of the A operand to the memory controller. Then ^"during time interval tl4, unit 20b sends the address of the B operand to the memory controller. These -addresses are subsequently utilized by the memory controller in response to commands from 5 execution unit 20e.

At this point, unit 20b waits for a request for data from the execution unit 20e. In the illustrated example, this waiting occurs during time intervals tl7-t!9. Then during time interval tl9, execution unit 10 20d completes the execution of the preceding instruction; and so it requests more data ^"from unit 20a.^" In response to this request, unit 20b sends the Opcode, A field length and data type, and B field length and data type to the execution unit. This occurs during time 15 intervals t20 and t21. Also, during time interval t21, unit 20b signals the instruction fetch unit 20a that it is free to start address calculations on another instruction.

Next, beginning at time interval t21, unit 20 20d begins its execution of the ADD instruction.

During time interval t21, unit 20e receives the OP, A length, and A type. Then during time interval t22, unit 20a receives the B-Length and B-type information. Also, it tests the field lengths and data types; and 2.5 based on that test, branches to a specialized routine. Thereafter, during time intervals t23 and t24, operands at the A address and B address are read from the memory. ^• This reading is initiated by commands from unit 20d. Thereafter, the A operand is received 0 by unit 20d during time interval t27; and the B operand is received- during time interval t28. Also during time interval t28, unit 20d adds the two operands together.

OMPI Then during time interval t29, uni.t 20d stores the result of the ADD operation in main memory; and sets 'appropriate status bits, such as carry indicator. Finally, during time Interval t30, unit 20d signals address calculator unit 20c that it is ready to begin processing the next instruction.

Now, in order for the arrays of Figures 1 and 2 to perform the various operations depicted in Figur.es 3 and 4 respectively, it is necessary that each sequential logic unit have a customized function performing capability. That.is, each unit must be able to perform a variety of unique tasks in only one time cycle. For example, by inspection of column 20e-l, time t5 of Figure 4, it is evident that in one cycle, unit 20e must be able to partition the address couple field into its component parts ~/( and <& .

Alternatively, this partitioning could be achieved by a series of standard one-bit shifting and masking steps; but that would Increase the execution time of the instruction being performed. A problem then is how to customize the one-cycle task performing capability of each of the units; and at the same time, provide for substantial commonality among the units - thereby achieving economy in design, fabrication, and inventory.

In the present invention, this problem is solved by constructing all of the sequential logic units of Figures 1 and 2 in accordance with Figures 5 and 6. Basically, the sequential logic unit of Figure 5 has a sophisticated computation performing capability; whereas the^" sequential logic unit of Figure 6 has a sophisticated parallel processing capability. But these capabilities are implemented in a very "soft" architecture which enables them to be readily customized to a particular task.

In regard to the Figure 5 and Figure 6 embodiments, all of the- teachings of copending United States Patent Application S.N. 087,666, entitled

"Digital Computer Having Programmable Structure", by Hanan Potash and Bud Levin filed October 24, 1979, and copending United States Patent Application S.N. 162,0^7, entitled "Digital Device with Interconnect Matrix", by.Hanan Potash and Melvyn Genter filed June 23, 1980 are^' herein incorporated by reference. Those applications describe Figures 5 and 6 in detail as individual stand-alone data processors.

For the purposes of this system-application, only those functional structures within Figures 5 and 6 which provide a level of "softness" or flexibility need be expanded upon here. These soft functional- structures are indicated in Figure 5 by reference numerals 30, 31, 32, and 33; and are indicated in Figure 6 by reference numerals 40, 41, 42, and 43. Structures 32 and 33 are identical to structures 42 and 43; so the latter are only illustrated as a single box.

Structure 30 is comprised of a plurality of memories as illustrated. These memories can be read- write or read-only memories. Each memory has two address inputs labeled C1-C4 and C5-C8. Data bits of like power are applied to address inputs C1-C4; while a single set of control signals is applied in parallel to address inputs C5-C8. By this structure, any type of arithmetic or 'logical transformation can be performed on the bits^' applied to address inputs C1-C4. Each particular transformation is specified by the content of the memories 30; and in the special case of read-only memories, that content translates to a set of electrical contacts.

Structure 31 is a means for enabling any bit of an input word I¥ to be transposed onto any other bit or bits of an output word OW. Thus it enables various fields to be coήcatinated or partitioned. Each particular type of transposition to be done is customized by specifying a plurality of selectable electrical contacts within structure 31. Then, during operation, those transpositions .are selectively performed in response to a set of control signals CS.

Structures 32 and 33 are a means for providing flexibility in the unit's testing and branching capability. In particular, structure 32 provides a means to horizontally translate any field or fields Into an address; and structure 33 provides a means to modify that address by any type of arithmetic or logical operation. Each particular type of transformation which structures 32 and 33 perform is selectable hγ a memory content; which again in the special case of a read-only memory, is implemented by a set of electrical contacts.

A control memory 34 is provided to store commands which direct the operation of structures 30, 31, 32, and 33."_. These commands can be vertical (I.e. - a single encoded field) or can be horizontal (i.e. -have several independent fields). One command specifies the operations to be performed in one cycle time. Similarly, structure 43 is a control memory which directs the operations to be performed by structures 40, 41-, and 42.

Structure 40 provides a means for customizing interconnecting paths among a variety of other "black boxes". For example, during one time cycle, the RAM

i SLEA , output can pass through the arithmetic logic unit (ALU), then through the shifter, and back to the RAM; while. during the next time cycle, input register #1, can pass through the shifter, then through the ALU then to the RAM. Each of the interconnecting paths is customized hγ a set of selectable electrical contacts.

In one..preferred. embodiment, the ALU and: shifter^■of ^• Figure 6 also have the "soft" structure as described above in Figure- 5. In that case, the ALU is constructed as structure- 30; and the shifter is constructed as structure 31. This composite unit is then customized to meet the functional requirements of units lOa-lOk in Figure 1, and units 20a-20d in Figure 2. Alternatively, those units can be made to differ from each other in the particular soft structures 30, 31, 32, 33 and

40 which they incorporate. For example, units 101-lOk and lOa-lOg may be constructed as Illustrated in Figures 5 and 6 respectively.

Preferably, each sequential logic unit is constructed on a single semiconductor substrate. That substrate optionally consists of either one chip or one wafer. In either case however, all of the above described flexibility is achieved by merely altering a single mask. Accordingly, the simultaneous goals of customized sequential logic units, and commonality in design-fabrication-and inventory are achieved.

To further elaborate on this point, consider now the physical layout diagrams of Figures 7, 8, and 9. Basically,Figure 7 shows a physical layout for a portion of structure 31; Figure 8 shows a physical layout for a portion of the memory structures; and

OMPI Figure 9 shows a physical layout for a portion of structure 40.

In these- Figures, all dashed lines are patterned diffusions lying at one surface of the semiconductor substrate; all solid horizontal lines are patterned polysilicon lying on an insulating layer over the one surface; and all solid vertical lines are patterned metal lying on another insulating layer over the polysilicon. All of_.the functions which these circuits perform are customized by choosing the plurality of electrical contacts 51, 52, and 53. Also, eac-h-of these contacts -occur between a -metal line and a diffusion. Thus, only a single mask - the one which defines the holes in the insulating layers between the diffusions and metal lines where contacts are to be made - needs to be specially made in order to completely define a logic unit's personality.

Turning now to Figure 10, the details of making the transition from the code of Figures 3 and

4 to the contacts of Figures 7, 8, and 9 will be described. To begin, assume that the code for the most frequently used instructions is complete. That Is, assume a chart similar to Figures 3 or 4 has been made for each instruction whose execution time will significantly affect the overall performance of the array.

At that point, all of the functions which a particular unit must perform In executing those instructions are tabulated as illustrated by the two left-most columns of Figure 10. In that Figure, some of the functions performed hγ unit lOe are tabulated as an example.

REΛ Next, separate columns are provided to the right of the above two for listing the data paths through connect matrix 40 and the operation of arithmetic unit 30, shifter 31, memory 42, and memory 43, Here it is assumed that the particular sequential logic unit being considered (i.e. - unit lOe) has incorporated all of these soft -functional structures.

Next, these columns are filled in to describe how the specified functions are performed by the soft structures i the unit. For example, consider row 1 of Figure 10. It indicates that the "Decode Op" function will be performed in unit lOe by having connect matrix 40 form a data path of IR1—^M42—^M43—?CM. Also, memory 42 will be required to input the opcode and output a relative address for each opcode. Further, memories 43 will be required to add the above relative address to the present address PA of the control memory. An arbitrary control code is then assigned to specify each of these functions. This process is repeated for each function that the unit under consideration must perform in order to execute the more frequently used instructions. As a further example, row two describes how the Partition Address couple function is performed. In carrying out this step, previously assigned control codes for matrix 40, AU30, SH31, M42, and M43 are used whenever possible; otherwise new control codes are assigned.

Following this step, all of the different control codes and corresponding tasks for each soft structure are tabulated. Then, those tabulated tasks are implemented by a set of contacts in the soft structure. Also, at this point, the^' actual bits in the control memory can be specified.

OMPI ~ In some cases, the number of specialized tasks that are tabulated for a particular soft functional structure may be too large. That is, ,given various chip size -and layout constraints, each soft functional structure will be limited to some finite number of specialized tasks. If that number is exceeded, then those specialized tasks which are used least, are eliminated; and the corresponding code flow is replaced hγ a sequence of code. All of the less frequently used Instructions are also implemented in a similar manner. That is, they are implemented by using the special purpose codes that are dictated by the more heavily weighted instructions. In this manner, the performance penalty Is minimal.

Next, the above sequence of steps is repeated for another sequential logic unit. As an example, one row in Figure 10 is illustrated as applying to unit 20b. Clearly, the special purpose functions which the soft structures must perform in this unit will be totally different than those that are performed in unit lOe. But, as described above, this total difference in personality between the units is achieved hγ changing only a single mask. Various preferred, embodiments of the invention have now been described in detail. In addition however, many changes and modifications can be made to these details without departing from the nature and spirit of the invention. Therefore, it is to be^'understood that the invention is not limited to said details but as defined .by the appended claims .

Claims

WHAT IS CLAIMED IS: .

1. ^' A digital computer for performing any selectable set of instructions, and being comprised of: an array of sequential logic units; said units including respective control memories for storing commands, means for sequentially fetching and executing selectable sequences of said commands, and soft functional structures for performing customized functions in response to said commands; said soft functional structures including a plurality of selectable electrical contacts which customize the particular functional response of the structures to said commands; said units being substantially identical except for the content of their respective control memory and the selection of their respective contacts; said units being adapted to execute respective command sequences from their control memory which^' together perform an instruction in said set.

-^OREAlT

2. A digital computer according to Claim 1 wherein said soft functional structures include a means"for selectively defining data paths through the unit.

3. A digital computer according to Claim 1 wherein said soft functional structures include a means for performing selectable arithmetic and logical transformations^'on corresponding bits of a plurality of data words. ...

4. A digital computer according to Claim 1 wherein said soft functional structures include a.means for performing selectable arithmetic and logical transformations on a plurality of control memory addresses.

5. A digital computer according to Claim 1 wherein said sequential logic units are individually packaged on a single semiconductor chip.

6. A digital computer according to Claim 1 wherein said sequential logic units are individually packaged on a single semiconductor wafer.

7. A digital computer according to Claim 1 wherein said control memory in said sequential logic units is a read-write memory.

8. A digital computer according to Claim 1 wherein said control memory in said sequential logic units is a read-only memory.

9. A digital computer according to Claim 1 wherein one of said units is adapted to perform conditional operations without first determining if those conditions exist; and another of said units is adapted to determine whether said conditions exist and alert said one unit if they do.

10. A digital computer according to Claim 1 wherein one of said units is coupled to transmit instructions, to. multiple -other ones of said units in parallel, and wherein said multiple units are responsive to each instruction to fetch and execute different sequences of commands in their respective control memories which together constitute the execution of that instruction.

11. A digital computer according to Claim 1 wherein said units are intercoupled in series, and are adapted to perform respective portions of each instruction in a serial fashion.

12. For use in a plurality of digital computer models, each model being adapted to perform a dissimilar set of instructions; a digital computer comprised of: a sequential logic unit including at least two input-output interfaces, a control memory for storing commands, means for sequentially fetching and executing selectable sequences of said commands, and soft functional structures for performing customized functions in response to said commands; multiple ones of said unit being .intercoupled in an array to communicate with each other and with a memory over multiple buses which selectively couple to said input-output interfaces; said soft functional structures Including a plurality of selectable electrical contacts which customize the particular functional response of the structures in each unit to .their respective commands.

13. A digital computer according to Claim 12 wherein said soft functional structures include a means 'for selectively defining data paths through the unit.

14. A digital computer according to Claim 12 wherein said soft functional, structures include a means for performing selectable arithmetic and logical transformations on corresponding bits of a plurality of data words. ..

15. A digital computer according to Claim 12 wherein said soft functional structures include a means for performing selectable arithmetic and logical transformations on a plurality of control memory addresses.

16. A digital computer according to Claim 12 wherein said sequential logic units^' are individually packaged on a single semiconductor chip.

17. A digital computer according to Claim 12 wherein said sequential logic units are individually packaged on a single semiconductor wafer.

18. A digital computer according to Claim 12 wherein said control memory in said sequential logic units is a read-write memory.

19. A-digital computer according to Claim 12 wherein said control memory in said sequential logic units is a read-only memory. -- -. f

20. A digital -computer according to Claim 12 wherein one of said units is adapted to perform conditional operations without first determining if those conditions exist; and another of said units is adapted to determine whether said conditions exist and alert said one unit if they do.

21. A digital computer according to Claim 12 wherein one .of ^"said units is coupled to transmit instructions to multiple other ones of said units in parallel, and wherein said multiple units are responsive to each instruction to fetch and execute different sequences of commands in their respective control memories whic together constitute the execution of that Instruction,

22. A digital computer according to Claim 12^" wherein said units are intercoupled in series, and are adapted to perform respective portions of each instruction in a serial fashion.

23. A digital computer for performing any selectable set of instructions, and being comprised of: ^{• •} a plurality of sequential logic units; each unit including a control memory for storing commands, means for sequentially fetching and executing selectable sequences of said commands, and a soft functional structure for performing customized functions in response to said commands; wherein said soft functional structure includes a memory means for performing selectable arithmetic and logical transformations on corresponding bits of a plurality of data words.