Many computers have instruction formats that are complicated to explain, because instructions that refer to memory do not contain complete memory addresses. Instead, a short address is included in the instruction, referring to a memory location within a small area of memory, and another field in the instruction allows that small area to be chosen from among a number of choices; perhaps two, as on a PDP-8, or sixteen, as on an IBM 360.
One type of computer that generally avoided this issue were computers with 24-bit word lengths, as that gave enough room for a 15-bit address, a 6-bit opcode, and three additional bits for such things as indexing and indirect addressing. Although only a few computers were made with 24-bit word lengths, for that reason, that particular word length lends itself well to very straightforward versions of old-style computer architecture.
Even though computers with a 24-bit word length were less common than those with other word lengths, still, if one looks hard enough, one can find quite a few. Among those examined here are the A/S Regencentralen RC4000, the Control Data 924, the ICL 1900, and the Foxboro FOX-1.
The classic group of 24 bit computers, however, consists of such machines as the Harris Series 500 and SLASH 6/7 (and their progenitor, the Datacraft 6024), the Scientific Data Systems 920/930/940 and also the 9300, the Honeywell (or Computer Controls Corporation) DDP-24, DDP-124 and DDP-224, the SCC 660, the ASI 6020 and 6040, and finally the SEL 840, 840A, and 840 MP. These machines share many architectural similarities, and all date from the latter part of the discrete transistor era (although the SEL 840 was one of the earliest machines to be made from monolithic integrated circuits instead).
In addition to having the potential to simplify the layout of instruction words, by allowing at least a modest memory to be fully addressed within each instruction, this length has two other potential benefits.
It lends itself to the use of floating-point variables that are 48 bits long. For a great many scientific purposes, 32-bit floating-point numbers, with a precision of 6 or 7 decimal digits, are not quite precise enough, due to the accumulation of error during calculations. But a 64-bit floating-point number, with a precision of 15 or 16 decimal digits, is usually far in excess of any practical requirement. 48-bit floating-point numbers have a precision of 10 decimal digits, as is provided by typical scientific calculators, which is just about right.
The ability to pack either three 8-bit bytes, or four 6-bit characters, in a 24-bit word, is attractive, since many computer applications don't need to bother with such extravagances as lower-case letters. In general, however, most computers with a 24-bit word did not take advantage of this, but simply used 6-bit characters.
There are also two microprogrammable machines with a 24-bit word length that particularly deserve a special mention: the Burroughs B1700, which hid its word length from the programmer, and the Packard Bell 440, which was perhaps one of the very earliest user-microprogrammable machines.
Incidentally, a somewhat larger number of computers had a 48-bit word length; some of them had 24-bit instructions, others 48-bit instructions, and some had variable-length instructions: the Control Data 1604 and 3600, the Honeywell 400, 800 and 1800 (and their progenitor, the Datamatic D1000), the Philco 2000 (whose central processor could have been the Philco 210, 211, or 212, and which was therefore sometimes referred to as the Philco 2000 Model 212 with a Philco 212 central processor), the RCA 601, the Maniac II, the AN/FSQ-31 and AN/FSQ-32, which were both made by IBM for the U.S. government, the Elliot 4130 (later the I.C.L. 4130), the FACOM 202, the University of Manchester Atlas, and the I.C.T. Orion are computers with that word length. As is the BESM-6 from the Soviet Union, and several computers from the People's Republic of China: the Model 109C, the DJS-6 (also known as the Model 108B), the DJS-8 (also known as the Model 320), and an early prototype machine based on integrated circuits, the Model 111. A later 48-bit integrated circuit machine, the Model 013, was also referred to as a prototype in one source I believe I saw, but more than one of this computer was produced. Photographs of some of those computers are visible here; a larger picture of the Model 013 front panel appeared in an issue of Datamation magazine.
DJS stands for Dianzi Jisuanji, the Putonghua phrase for "computer" (electronic calculator); the Kuo-yü phrase for computer is Tien Nao (electric brain) (in Pinyin, Diannao).
Because the Burroughs B5500 and B6700 computers used a word containing 48 bits of data, but also extra bits that described the type of the data which played an important role in programming the machine, I would consider that series sufficiently out-of-the-ordinary to attempt to discuss here; the same applies to the English Electric KDF9 computer, which was stack-oriented like the Burroughs machine. The Telefunken TR440 computer also added two descriptor bits to each word with 48 bits of data, but these seem to only serve to protect the machine against accidentally interpreting instructions as data, or integers as floating-point quantities: its instructions were 24 bits long, with an 8-bit opcode and either a 16-bit address or two 8-bit indexed addresses, but I have not yet been able to derive the structure of an 8-bit address from the available references.
The instruction formats of some of these computers are shown in the diagram below:
The DDP-24 was the first machine of its group. Later came the larger DDP-224, for which optional floating-point hardware was available. Both of them were implemented with transistors. The DDP-124, which came later, was implemented with integrated circuits, and could trap on unimplemented opcodes to simulate the floating-point instructions of the DDP-224.
On the SDS 920, 930, and 940, half the opcode space was given over to user-defined instructions, and the first bit of the word was unused in the machine (except in some operating modes of the 940), although the object program format used it to indicate relocatable addresses. The SDS 9300, on the other hand, had a two-bit indexing field to indicate either that no indexing was taking place, or which of three index registers would be used with an instruction.
The 930 could have up to 32K words of memory. When it was expanded in this fashion, the first bit of the 14-bit address field indicated, if zero, that the address referred to the first 8K words of memory, and if one, that the second bit would select from one of two three-bit page registers which allowed pages of 4K words to be accessed anywhere in memory.
As the 940 was intended for timesharing, in user mode the first bit of the word was used to indicate a supervisor call instruction, and the first three bits of the address would select from one of eight five-bit page values (plus a bit indicating a read-only page). This meant that the 940 could have up to 64K words of memory. In monitor mode, the first bit of the word, if a 1, indicated that the instruction would use the user memory map instead of the compatible 930-style memory extension, thus allowing all 64K of memory to be accessed while in monitor mode.
Although the SDS 9300, the SEL 840, and the DDP-224 used the same general instruction format, they were not compatible: the SDS 9300 used two's complement arithmetic, while both the DDP-224 and the SEL 840 represented numbers in sign-magnitude form, for example; the DDP-224 and the SEL 840 were distinct from each other in other ways, one of which was that in the SEL 840, the shift instructions had a special format. Similarly, the ASI 6020 computer had the same general instruction format as the Datacraft 6024 computer and its compatible successors by Harris while having a different floating-point format, and being incompatible in other ways, one of which was that instead of having three index registers, it used memory locations 1 through 3 for this purpose. (The UNIVAC 9140 computer was also envisaged as having the same general instruction format as the Harris computers, but I am not yet certain if that computer had actually existed.)
In the SDS 9300, index registers were 24 bits in length, but addresses were only 15 bits wide; the index registers contained both a nine bit signed increment field and a 15 bit address.
The P bit also stoood for a programmed operation on the SCC 655, 660 and 670 computers as well, making them very similar to the SDS 920, 930, and 940 except for also using an indirect bit and having a different instruction set. Another similarity is that just as the SDS 9300 computer was similar, but not fully compatible, to the SDS 920, 930,and 940, there was also the SCC 6700 computer, a larger scale time-sharing computer in which the size of the address field was reduced from 15 bits to 14 bits, so that the P bit could be moved, and another modifier bit, shown here as the S bit, could be added to indicate instruction words that functioned as system calls.
Just as the SDS 92 had a 12-bit instruction word, so did the SCC 650.
The ICL 1900 had a similar feature; however, when an instruction addressed 6-bit characters, the increment field was shortened, and its first two bits were instead used to indicate which of the four bits in a word was pointed to by the effective address. It also had a 22-bit addressing mode, in which the first two bits of a word could still serve this purpose, but the increment field was not used. Since addresses in instructions were only 12 bits long, the index registers actually functioned either as base registers, or, when access to arrays was desired, or when their values were being constantly changed thanks to the increment field, as pointer registers. Since the address could never be longer than 12 bits, unlike the case with the Philco 2000 computers, they were needed as base registers.
The Four-Phase Systems IV/70 had some interesting peculiarities.
Index register 3 could not be used with indirect addressing; instead, if the X and I bits were all ones, this indicated an instruction that worked on character strings.
Also, single-precision floating point numbers consisted of a one-word mantissa followed by a one-word exponent, which had to be aligned on a 48-bit boundary. Double-precision floating-point numbers had the additional least significant word of the mantissa preceding a single-precision floating-point number; thus, they had to be aligned so that they began on an odd word address; thus, arrays of these numbers would contain empty space.
The A/S Regencentralen RC4000 computer from Denmark allowed itself a mere 12-bit address field, like an IBM 360. Addresses could be program-relative, and if indirect they were pre-indexed. This machine is remembered for being used by Per Brinch Hansen, himself famous for the language Concurrent Pascal, for some of his early work on operating systems. Also, it is unusual in that its addresses were based on dividing memory into areas 12 bits in length.
The 12-bit address field of the RCA 110 had a simpler explanation; it had only a 4K-word core memory, but this was supplemented by a drum which could store 32K words of data.
The Packard-Bell 440 was one of the earliest machines that was user-microprogrammable. Microprograms were contained in a memory of up to 4,096 words in length made of special fast biaxial core. Microcode instructions were 12 bits long, two to a 24-bit word.
Although microprogrammable, the design favored the format shown above for regular instructions, in which the opcode came at the beginning of the word, followed by a field indicating an index register; and the machine came with normal software (such as a FORTRAN compiler) which worked under a standard instruction set. It could have up to 32,768 24-bit words of normal memory (2 usec access, 5 usec cycle), and was made from discrete transistors.
Possibly it could only have had up to 28,670 words of conventional memory, since apparently the regular memory and the biaxial shared the same address space, unless biaxial memory was not used in a given configuration.
Its microprogram instructions had the following format:
The second field could indicate the source for the second operand in an operation, or it could indicate the source of a memory address. The destination to which results from the operation would be sent is implicit in the opcode.
It had four 24-bit registers, A, B, C, and D, and two 15-bit registers, P and L. P was the microprogram counter, and L was usually used as the program counter. An eight-bit register, N, was also present, to which the exponent part of a 24-bit register could be sent; the exponent part consisted of the second through ninth bits from the left, and so the floating-point format was also determined.
The Control Data 1604 also began the instruction with a 6-bit opcode field, and used the following 3 bits to indicate indexing; but it had six instead of seven index registers: a value of 7 in this field indicated indirect addressing.
The Control Data 3600 was largely compatible with the 1604, but two unused opcodes (00 and 77), and three other opcodes used for the operations input, output, and external function were used to add new capabilities to the machine. One new capability was an instruction format in which the 15-bit address field was divided into three 5-bit register addresses; this allowed eight instructions which acted on data in the machine's various registers; these instructions allowed ease in exchanging information between the accumulator, the Q register (equivalent to the multiplier-quotient register), the six index registers, and the like.
Thus, the 3600 was not a machine that did the majority of its arithmetic using three-address instructions involving a set of 32 general registers or a set of 32 floating-point registers, although its instruction formats might suggest the idea of designing such a machine.
As well, the 3600 extended the 1604 architecture by allowing up to eight banks of 32,768 words of memory; two three-bit registers indicated the current bank used for instructions and data respectively. And among the new instructions were some which were 48 bits in length, not shown in the diagram above.
The 24-bit Control Data 924 also used the same general instruction format as the 1604, a six-bit opcode field followed by a three-bit index field, which in the case of the 924 as well as the 1604 and 3600, selected one of six index registers if it indicated that indexing was to be performed, followed by a 15-bit address.
On the other hand, the Control Data 3300 was a machine with a 24-bit word that had a somewhat different instruction format than that of the Control Data 3600. It only had three index registers, making available an extra bit for indirect addressing. Floating-point was an option on this machine, and character string manipulation instructions were also optional for it. These included BCD addition instructions on unpacked BCD numbers, six bits per digit, although there were also instructions to convert to and from packed BCD. The format of this type of instruction, which was three words long, is shown following that of the normal memory reference instruction.
When I originally heard of the instruction format of the Philco 212 computer, it seemed to me that with a 32,768-word memory, and a 12-bit address field, the index register field in the instruction should more properly be referred to as a base register field. I subsequently learned, from a private communication, that the selector bit was used to choose between a 15-bit address, and a 12-bit address with a 3-bit index register field.
In this case, although it took some reflection for me to realize this, the index registers are properly named. A normal memory access would use a 15-bit address, but when an array is referenced, then an index register would need to be specified. Since the address field is short in that case, the index register would have to be loaded with a pointer to the target array, although the displacement field could be used when one is dealing with two or more arrays which all fit in the same 4,096-word area of memory.
The addressing modes of the 68000 microprocessor, before the 68020 added additional addressing modes to that architecture, had a similar limitation, so this was not unique. Using a pointer register instead of indexing has the advantage that the addition of the location of an array to a displacement is done once at the beginning of a loop instead of with every instruction; even if one has to increment multiple pointers with each loop iteration, there is still a net savings; but this seems to be lost when one has a short displacement field anyways. It had occurred to me that if one were to put the optional index register field at the end, instead of the beginning, of the address field, one could locate arrays anywhere in memory, subject only to an eight-word or eight-byte alignment restriction. However, this wouldn't have been practical on the Philco 2000 computers for another reason, of which I learned when the Philco 212 manual became available online, thanks to the tireless work of Al Kossow.
The index registers of the Philco 212 were versatile, however. In addition to a 15-bit value, there were two additional control bits. The index registers could be set with those bits to autoincrement by one, or to act as pointer registers, with the 12-bit address field in the instruction serving instead to increment or decrement that pointer; this mode would not be too adversely affected by the shorter address field in indexed instructions.
In the case of the Transac S-2000 computer at least, the number of index registers was variable; a machine could have 4, 8, 16, and perhaps 32 index registers, and the index register field would be only as large as required for the number of index registers a particular machine had, leaving the remaining bits for the displacement.
The first bit of the opcode field was called the F bit. For arithmetic instructions, it was used to indicate floating-point instructions. For jump instructions, it indicated branches to the second 24-bit instruction in a 48-bit word in memory. To simplify how the assembler worked, branches to the right-half instruction were given separate mnemonics, even though, in effect, the F bit was really extending the length of the address field by one bit.
The bit marked P in the instruction layout of the BESM-6 is used to indicate whether an instruction refers to an address on the current page, or on page zero of the current half of memory. This is not entirely unlike the similar bit on a PDP-8, as when the memory on that computer exceeded 4K words, with the aid of its memory extension feature, page zero was in the current 4K words, not the first page of the entire potential 32K address space.
Some of the computers listed here are old enough that their data formats departed from some conventions now in universal use.
The SCC 660, as well as the SDS 900 and SDS 9300, used two's complement notation for negative numbers, as is the modern practice, but the multiplication instruction treated numbers as fixed-point fractions in [0,1) rather than integers.
The DDP 24 and the ASI 6020 both used sign-magnitude notation, and a multiply instruction placed the sign of the result in the sign position of both halves, so that the result was aligned for use as either the product of integers or the product of fractions.
The Datacraft 6024 and its successors from Harris used two's complement notation, and the result of a multiply also skipped the first bit of the second word of a result, but this bit was always filled with a zero, even if the result was negative.
The CDC 924 computer used one's complement notation. Multiplication produced a 48-bit integer result. (The CDC 1604 was also a one's complement machine.)
Thus, of the 24-bit machines that had a 15-bit address (rather than a 12-bit address) in the instruction word, none followed the modern convention of both two's complement arithmetic and an integer multiplication result with no skipped bit.
I find this a pity because a 24-bit word with a 15-bit address allows the use of a reasonably-sized memory without complications such as a page bit or base registers, and I was looking for a machine with the smallest possible word length (rather than 32 or 36 bits) to provide this, but not a smaller word length that was harder to program (such as 12 or 16 bits), but which was a real historically-used architecture as being suitable to a project for a modern "toy" computer with front panel lights and switches.
The General Electric 425 and 435 computers also had a 24 bit word length.
The basic instruction format of these computers was very similar to that of the Control Data 1604 computer; there was an index register field in the same location that could be zero for no indexing, and 1 to 6 to indicate an index register.
However, when that field contained a 7, instead of indicating an indirect address, it indicated that the instruction would be an extra word long, to allow the address to be modified.
The modifier word contained a class field, which, when equal to 01 or 10, indicated that it pointed to a word to use as a modifier word (in the case of 10, the indirect bit of the modifier word would be ignored). If it was 00, the modifier word would produce an effective address to be added to the address of the original instruction.
The AMS field could contain a 7 to indicate that there would be yet another modifier word, but it could not have a value from 1 to 6 to allow more than one index register to be added.
Also, some instructions were two-address instructions. In such a case, the amount of modification that could be applied to the second address was also restricted, only allowing the two class bits (rather than the I bit) to specify that the address points to the word actually containing the address, and not allowing indexing or more than one modifier.
It might be noted that the University of Manchester Atlas was significant as the first computer to embody what we now refer to as "virtual memory", although it was called a "single-level store" in connection with that computer.
Here are the instruction formats of some computers that used 48-bit instructions:
The Datamatic 1000, and the Honeywell 400, 1400, 800 and 1800 computers used a 48-bit instruction word with a three-address format.
The Datamatic 1000 was a vacuum-tube computer, and the others were solid-state. The Honeywell 400 had the 1400 as its fully compatible and faster successor, and the Honeywell 800 was succeeded by the 1800. The 400 and 1400 had three index registers, and the 800 and 1800 had eight, and so the 800 was a computer with added features.
The Honeywell 800 and 1800 in particular had some unusual and advanced architectural features.
It could operate both on binary and packed decimal numbers. Because it used sign-magnitude notation, and it used the same four-bit sign symbols for binary and decimal numbers, it had the one peculiar weakness that a signed binary number consisted of four bits of sign plus 44 bits of precision, but it could also handle 48-bit unsigned binary integers.
The computer had eight complete sets of registers, so that up to eight tasks could be processed concurrently without context-switching overhead. Each task had 32 registers; only eight of them were general registers, although another eight were usually available as general registers as well.
Register indirect addressing was possible, and instructions that used it also provided for a five-bit increment field in the instruction, so that a pointer in a register could be advanced after use. (When a register was used as a direct source or operand location, the increment field was also applied to its value after use if that value was nonzero, but in that case, the increment could not be used if the register was the destination of the instruction.)
As the diagram above illustrates, an indexed memory address includes a three-bit field indicating the index register, and a displacement field of eight bits. When a field indicating a register address was indexed, the index register contained an increment field and a register field instead of a memory displacement, allowing a type of indirect indexing.
But the most unique aspect of the H800 architecture is that each process had two program counters, the main one called the Sequence Counter, and an auxilliary one, called the Cosequence Counter. The first bit of each instruction indicated which program counter would be used to locate the next instruction to be executed. This unusual feature may have only been used for some quite specialized applications, and the number of programmers who understood its implications and effective use may have been limited. When used, it would have avoided branching and context-switching overheads. One possible use, perhaps even the most likely, would be in code that made frequent calls to a subroutine that behaved like a state machine. Switching to the cosequence counter could then plunge directly into the code for the current state; on a conventional architecture, first there would be a subroutine call, and then the subroutine would begin by doing a computed go to (or a pre-indexed indirect branch) to the code applicable to the current state, and so both these overheads would be avoided with this feature.
Upon reflection, it seems to me that the practical use to which this feature could be put, in order to explain its existence, is most likely something to do with optimizing merge sort routines.
Considerably more information about the Honeywell 800 computer is available at this site.
The next instruction format shown is that of the I.C.T. Atlas computer, famous for introducing the concept of virtual memory under the name "single-level store".
Next is shown the instruction formats of the NCR 315 computer. This commercial computer is notable for having a memory addressable in units of 12 bits, which it called slabs, that could contain either two six-bit characters or four binary-coded decimal digits, and for a large-capacity random-access storage system, providing rapid selection of a strip of magnetic tape, known as CRAM, for card random-access memory. IBM's Data Cell is another example of a storage device in this class. Since other computers with 24-bit instructions did not also have a two-word instruction format, its instructions are shown in this diagram rather than in the preceding one.
The instruction format of the AN/FSQ-32 computer is described as well. The first of the five lines giving its instruction formats shows those for ordinary integer arithmetic instructions as well as load and store instructions. The bit marked as K indicates an immediate operand (in a floating-point instruction, this bit is instead used to indicate an unnormalized operation). The displacement field is used to indicate that the source operand will be shifted by the multiple of six bits indicated prior to the operation. The mode field indicates whether 48-bit arithmetic is performed, or 24-bit arithmetic on both halves of the accumulator, and in addition indicates in the latter case if only one half of the accumulator is used to perform arithmetic, or if the least significant 24 bits of the source operand will be duplicated prior to the operation, called twinning. The activity field enables or disables individual six-bit areas of the accumulator. Character displacement, twinning, and the effects of the activity field are performed in that order. The S bit is used for sign extension (the AN/FSQ-32 performed one's complement arithmetic).
The I bit indicated indirect addressing, and the D bit indicated something called double indexing: a special register indicated which register could be used as a second index register in an instruction. The sX field indicated either one of eight possible index registers, or program-counter relative addressing, or using the accumulator as an index.
In the second line, we see the format of logical instructions. The four-bit field marked connective gives the different bit values to be placed in the result for each of the four combinations of an 0 or 1 bit respectively in the source and destination operands. The activity field, indicating which six-bit characters are affected, is also present, but if the M bit is set, the B register is used as a mask for the operation instead.
In the third line, we see the format of the convert instructions, which perform character translation. The dXR field indicates an index register that can be changed by the instruction, and the R bit will cause one of the convert instructions to proceed from right to left.
In the fourth line, we see the format of a shift instruction. The K bit allows the index registers to modify the shift count, the S bit indicates the direction of the shift, and the R bit indicates rounding for left shifts.
The fifth line indicates the formats of the instructions of the branch decrement type, reminiscent of the IBM 704 architecture, but offering a more complete set of instructions.
Not shown are the instruction formats of the Philco 213 computer, although they were described in a supplement to the proceedings of a Fall Joint Computer Conference with papers on large computers, also including the Control Data 6600 and the IBM System/360 model 92 (which, of course, became the model 91 before becoming available for sale), as this computer did not reach the stage of manufacture.
The English Electric KDF9 is also significant in the history of computing, playing a significant role in the development of the computer language Algol. Its instructions were composed of from one to three 'instruction syllables', eight bits in length. Such short instructions were possible because the machine was stack-oriented, like a series of Burroughs machines whose first member was released almost contemporaneously with the KDF9. (During its lifetime, it became the English Electric Leo KDF9 and then the English Electric Leo Marconi KDF9!)
The ICL 1900 computer was previously the ICT 1900 computer; behind the Iron Curtain, the ODRA 1304 computer was made in imitation of it, but, according to at least one web site, under license.
It may also be noted that the Philco 2000 computer was advertised as the first commercially-available computer made with transistors instead of vacuum tubes, and it was closely followed by the Control Data 1604 computer. However, they both appear to have been preceded by the Metro-Vick (or Metropolitan-Vickers) MV950 computer in Britain.
I remember, long ago, reading about how an Armenian in the U.S., despite feeling no sympathy for the Soviet system, felt a twinge of pride at reading a news item about a computer being designed and built in Soviet Armenia. I had imagined, as I read this, that the computer he was reading of would likely have had a 24-bit word length. I was probably wrong; the Nairi computer from Armenia was an imitation of the 16-bit Digital Equipment Corporation PDP-11, and, afterwards, Armenia produced several of the EC series of computers, compatible with the IBM System/360. But then, there was the Hrazdan-3 and its predecessors; perhaps they had a 24-bit word length, but I have not been able to determine this. In any event, I have found the possible reason why, subconsciously, I associate 24-bit architectures with small, obscure countries.
The ICL 1900 was closely derived from the Ferranti-Packard 6000 computer, which was designed in... Canada! (It apparently has, from the Web, ancestors in both a Canadian military computer, the DATAR, and an Australian computer called the Cirrus.) Its instructions were composed of a 3-bit destination register field, followed by a 7-bit opcode, and then a 2-bit index register field, with 12 bits left at the end for the address.
One notable thing about the DATAR that I discovered in my web surfing for more information on 24-bit computers in general, and then the ICL 1900 in particular, was that it came equipped with an early trackball... which used, for the ball itself, a bowling ball! It was, however, the slightly smaller bowling ball used for five-pin bowling, which is also better suited for trackball use in that it does not have the three finger holes of ordinary 10-pin bowling balls. (Sorry to disappoint you!)