CS222 Lecture: Other CPU Architectures: 0,1,2 and 3 address machines
                                                        Last revised 2/23/99

Introduction
------------

   A. Recall that, at the start of the course, we drew a distinction between
      computer ARCHITECTURE and computer ORGANIZATION.
 
      1. An architecture is a functional specification of a computer system, 
         generally defined in terms of how the system is seen by an assembly
         language programmer.

      2. An organization a particular way of implementing an architecture.

   B. For most of this semester so far, we have been studying in detail a
      particular architecture: that of the VAX, often contrasting it with
      a very different architecture: that of MIPS.  After spring break, we will
      move to a study of basic principles of computer organization.  Before we 
      do this, though, we will spend some time looking at computer architectures
      other than that of the VAX and MIPS.

      1. It turns out that a computer architect faces a number of fundamental
         choices in designing a CPU architecture.  Different architectures
         represent different combinations of these choice.

         ASK CLASS: If you were designing a new CPU architecture, what
                    basic issues would you need to think about. (Recall that
                    a CPU architecture is basically the structure of a machine
                    as seen by an assembly language programmer.)

         a. The set of data types that are to be directly supported by the
            hardware.  

         b. The size of the logical address space the machine can access.

         c. The set of programmer-visible registers that are used to store and
            manipulate data.

         d. The format of individual machine instructions.

         e. The set of instructions that are provided in hardware.

         f. The modes that are provided for generating the address of the
            operands these instructions work on.

      2. We will look at each choice individually.

I. Hardware data types
-  -------- ---- -----

   A. All CPU's provide for the manipulation of binary integers of some
      specified number of bits (called the word length of the machine).

      1. Word lengths for current architectures vary from 8 bits to 64 bits.

         a. The earliest microprocessors used an 8 bit word, due to
            technological reasons.  Today's integrated circuit technology
            permits 32 and 64-bit word lengths, but 8-bit word length
            microprocessors are still used in applications (e.g. embedded
            control systems) where a longer word length does not justify the 
            increased cost.

            Example: the Z80

         b. 16 bit word length CPU's were once fairly common, both in the
            form of minicomputers (CPU's built from individual chips), and
            as a technological step up from 8 bit microprocessors), but now if 
            the application calls for more than an 8 bit word the usual step 
            is to a 32 bit word length (although 16 bit CPU's are still made.)  
            
            Examples (historic): Intel 8086, 80186, 80286 (microprocessors)

                                 DEC PDP-11 (minicomputer)

            The chip used in the original IBM PC (Intel 8086/88) used a 16-bit 
            word, and prior to Windows 95 Microsoft's operating systems were \
            based on this architecture, even when running on 32-bit chips.

         c. From the 1960's onward, mainframe computers have used a 32-bit
            word size.  32-bit microprocessors became possible during the
            late 1980's.  The most common word size for CPU's today is 32 bits.

         d. There are also some 64-bit CPU's - e.g. DEC's Alpha, high end
            members of the MIPS family, and Intel's Merced project (not yet
            in production).

      2. Many CPU's actually provide for the manipulation of several sizes of 
         binary integers - often 8, 16, 32 and perhaps 64 bits.

         a. As we have noted, one of these sizes is typically designated as the
            natural WORD SIZE of the machine, and governs the width of internal 
            registers, data paths etc. (E.g. 32 bits for the VAX).

         b. Frequently, another size (generally the smallest one) is chosen as 
            the basic ADDRESSABLE UNIT, and represents the smallest piece of 
            information that can be transferred to or from a unique address in 
            memory.

            i. On most modern machines, the addressable unit is the byte

           ii. Many older machines, and some modern ones, are only word 
               addressable 

          iii. Some machines have been built that allow addressing down to the 
               individual bit.

         c. The drawing of a distinction between the addressable unit and the
            word size was an important technical innovation that first appeared
            in the IBM 360 series (early 1960's).  It allowed a single 
            architecture to be used for both business applications (which
            typically work with byte-sized character data) and scientific ones
            (that typically need long words for arithmetic.)

      3. Of course, only one size is ESSENTIAL.  Other sizes could be 
         synthesized in software by combining operations on adjacent data items
         (to get a bigger size) or packing several smaller items into one word.
         But doing such operations in software requires multiple instructions,
         and so performance considerations generally dictate offering multiple 
         sizes in hardware.

   B. Some CPU's provide for other data types - including some or all of the
      following:

      1. Binary bit vectors of varying length (not necessarily multiples of the
         addressable unit.)

      2. Decimal integers, represented using packed binary coded decimal (BCD)
         or some other scheme.

      3. Floating point numbers. (Sometimes provided for by a separate - and
         optional - coprocessor chip in microprocessor-based systems.)

      4. Character strings (arbitrary-length strings of bytes).

         Example: The VAX supports all of the above, plus others.

   C. The architect's choice of data types to be supported by the hardware has
      a profound impact on the rest of the architecture.

      1. The tendency in the 1980's was to move toward ever-richer sets of 
         hardware-supported data types.  The VAX is really the epitomy of
         this trend.

      2. RISC's tend to support only the basic integer types (8, 16, 32
         and maybe 64 bit) with a floating point coprocessor used for
         floating point numbers and all other data types managed by software
         using these basic types.

II. Size of the Address Space
--  ---- -- --- ------- -----

   A. As we have seen, many instructions have to refer to one or more operands
      contained in memory.  One key architectural question is the length, in
      bits, of the address of an item in memory, because this determines the
      size of the logical address space the CPU can address, and hence how much
      memory can be used.

      Example: the IBM 360/370 series was originally designed to use a 24
               bit address, limiting the address space to 16MB. At the time the
               architecture was designed (early 1960's), this seemed an 
               exceedingly generous figure; but it has since become a key 
               limitation that has led to a need to significantly modify the 
               architecture.

   B. Often, the address size is the same as the word size, because addresses
      are often calculated using integer arithmetic operations.

      1. Example: PDP-11 - 16 bit word, 16-bit address
        
      2. Example: VAX, MIPS, many others: 32 bit word, 32-bit address

   C. However, some machines have address sizes smaller than the word size or
      larger than the word size.  The latter requires some tricky mechanisms to
      allow the formation of an address.

      1. Example: IBM 360/370 - 32 bit word, but 24 bit address

      2. Example: Intel 8086/8088 etc. - 16 bit word, but 20 bit address.  (All
         addresses are formed using a 16 bit offset relative to a 16 bit
         segment register that is internally shifted left 4 places.)

III. Register sets
---  -------- ----

   A. Early computers - like the VonNeumann machine - generally provided only
      a single register that could be directly manipulated by the assembly 
      language programmer - normally called the accumulator.

   B. However, as hardware costs dropped it became obvious that having more
      than one such register would be advantageous.

      1. Data that is stored in a register can be accessed much more quickly 
         than data stored in memory, so having multiple registers allows the 
         subset of a program's data items that is currently being manipulated 
         to be accessible more quickly.

      2. Modern CPU's typically have anywhere from a half dozen to 32 more or
         less general purpose registers (plus other special purpose ones.)

      3. In fact, some CPU's have been built with as many as 256 registers!

      4. However, there are some microprocessors (e.g. the 6502) that have
         the one accumulator register structure inherited directly from
         the VonNeumann machine.

   C. One question is how functionality should be apportioned among the
      programmer-visible registers.  Here, several approaches can be taken.

      1. Multiple registers, each with a specialized function - e.g.:

            - AC (accumulator)
            - MQ (multiplier-quotient: extends the AC for multiply/divide)
            - Base register - one or several
            - Index register - one or several

         Examples: Many micros

            - Z80: A register = accumulator; B..E = temporary storage;
                   IX,IY = index.
            - 8086/88: AX = accumulator; BX, CX, DX can also be used as
                       accumulators, but not as flexibly; BX, BP can be used as
                       base registers; SI, DI - index registers. [IBM PC]

      2. Multiple general-purpose registers - often designated R0, R1 ....

         Example: Many modern machines including the VAX and MI{S

      3. A hybrid approach, with multiple general registers for arithmetic, etc.
         and multiple address registers for forming addresses.

         Example: Motorola 680x0 

      4. A very large number of general registers, organized into smaller sets -
         only one of which is "visible" at a time.  (This allows each procedure 
         in a sequence of procedure calls to have its own private set of
         registers).

         Example: Berkeley RISC

      5. At the opposite extreme, it is possible to design a machine with NO
         programmer visible registers per se, by using a stack architecture
         in which all instructions operate on a stack maintained by the 
         hardware. 

         Example: Java Virtual Machine

   D. In deciding on a register set architecture, there is an important trade
      off.  When a machine has multiple registers, each instruction must
      somehow designate which register(s) is/are to be used.  Thus, going from
      one AC to (say) 8 registers adds 3-9 bits to each machine language 
      instruction - depending on how many register operands must be specified.

IV. Instruction Formats
--  ----------- -------

   A. Different computer architectures differ quite significantly in the
      format of instructions.  Broadly speaking, they can be classified into
      perhaps four categories, by looking at the format of a typical
      computational instruction (e.g. ADD).  To compare these architectures,
      we will look at how they implement two HLL statements:

         Z := X + Y

         Z := (X + Y)/(Q - Z)
 
   B. Memory-memory architectures: all operands of a computational instruction 
      are contained in memory.
        
      1. Three address architecture:  op source1 source2 destination
                                      (or op destination source1 source2)
         Z := X + Y

                ADD     X, Y, Z


         Z := (X + Y)/(Q - Z)

                ADD     X, Y, T1
                SUB     Z, Q, T2
                DIV     T2, T1, Z

         Example: VAX ...3 instructions

      2. Two address architecture: op source destination
                                   (or op destination source)

         (destination serves as both the second source and the destination)

         Z := X + Y

                MOV     X, Z
                ADD     Y, Z

         Z := (X + Y)/(Q - Z)

                MOV     Q, T1
                SUB     Z, T1
                MOV     X, Z
                ADD     Y, Z
                DIV     T1, Z

         Example: VAX ...2 instructions

      3. Note that the VAX is unusual in having both formats

      4. Multiple register machines generally allow registers to be
         used in place of memory locations, as on the VAX.

   B. Memory-register architectures: one operand of an instruction is 
      in memory; the other must be in a register.  Note: These are often 
      called "one address" architectures.

      1. Multiple register machine: op source register

         (The designated register serves as both the second source and
          the destination; on some machines (e.g. Intel 80x86), the
          memory location may also serve as the destination.)

         (Multiple register machines generally allow a register to be used
          instead of the memory operand, as well)
        
         Z := X + Y

                LOAD    X, R1
                ADD     Y, R1
                STORE   R0, Z

         Z := (X + Y)/(Q - Z)

                LOAD    X, R1
                ADD     Y, R1
                LOAD    Q, R2
                SUB     Z, R2
                DIV     R2, R1
                STORE   R1, Z

          Example: Intel 80x86

      2. Single accumulator machine:    op source

         (AC serves as both the second source and destination)

         Z := X + Y

                LOAD    X
                ADD     Y
                STORE   Z

         Z := (X + Y)/(Q - Z)

                LOAD    Q
                SUB     Z
                STORE   T1
                LOAD    X
                ADD     Y
                DIV     T1
                STORE   Z

         Example: The Z80


   C. Load-store architecture: All operands of a computational instruction
      must be in registers.  Separate load and store instructions are
      provided for moving a value from memory to a register (load) or
      from a register to memory (store).

         Z := X + Y

                LOAD    X, R1
                LOAD    Y, R2
                ADD     R1, R2, R1
                STORE   Z, R1

         Z := (X + Y)/(Q - Z)
 
                LOAD    X, R1
                LOAD    Y, R2
                ADD     R1, R2, R1
                LOAD    Q, R2
                LOAD    Z, R3
                SUB     R2, R2, R3
                DIV     R1, R1, R2
                STORE   R1, Z

      Example: RISCs, including MIPS

   D. Stack architecture: All operands of a computational instruction
      are popped from the runtime stack, and the result is pushed back on
      the stack.  Separate push and pop instructions are provided for moving
      a value from memory to the stack (push), or from the stack to a
      register (pop).

         Z := X + Y

                PUSH    X
                PUSH    Y
                ADD
                POP     Z

         Z := (X + Y)/(Q - Z)
 
                PUSH    X
                PUSH    Y
                ADD
                PUSH    Q
                PUSH    Z
                SUB
                DIV
                POP     Z

      Example: Java Virtual Machine

   E. As the examples above have noted, many machines allow the programmer
      to choose between several formats.

      1. VAX: Three address versus two address formats; and any operand can
         be either a register or memory

      2. IBM 360/370, 80x86: one operand can be either memory or a register;
         the other must be a register.

      3. Of course, this requires one or more additional bits in the op-code
         to specify the format!

   F. There appears to be a trend in terms of program length in the above
      examples.

      1. The example programs for the 3 address machine used the fewest
         instructions, and those for the load-store and stack machines used
         the most, with the other architectures in between.  (This is a general 
         pattern.)

      2. However, the program that is shortest in terms of number of
         instructions may not be shortest in terms of total number of BITS -
         because encoding an address in an instruction requires a significant
         number of bits.

      3. You will investigate this further on a homework set.

V. Instruction Sets
-  ----------- ----

   A. Every machine instruction includes an operation code (op-code) that
      specifies what operation is to be performed.  The set of all such
      possible operations constitutes the INSTRUCTION SET of a machine.

   B. Early computers tended to have very limited instruction sets: data        
      movement, basic integer arithmetic operations (+, -, *, and / - but 
      sometimes only + and -); integer comparison; bitwise logical and/or; 
      conditional and unconditional branch, subroutine call/return, IO
      operations and the like.

   C. Introducing multiple data types increases the size of the instruction
      set, since there must be a version of each operation for each (relevant)
      data type.

   D. In thhe late 1970's and 1980's machines moved in the direction of 
      including some very complex and high-power instructions in the 
      instruction set.  The VAX is one of the best examples of this trend, 
      with single machine-language instructions for operations like:

      1. Character string operations that copy, compare, or search an entire
         character string of arbitrary length using a single instruction.

      2. Direct support for higher-level language constructs like CASE, FOR,
         etc.

      3. Complicated arithmetic operations such as polynomial evaluation.

      4. Queue manipulations to or remove an item from a queue.

      etc.

   E. RISC architectures arose from a questioning of the wisdom of this trend,
      leading to much simpler instruction sets.

      1. The existence of complex instructions imposes a performance penalty
         on all the instructions in the instruction set - for reasons that
         will become more evident later.

      2. On the other hand, if the instruction set is kept simple, the door
         is opened to significant speed-up by the use of pipelining - i.e.
         executing portions of two or more instructions simultaneously.

VI. Addressing Modes
--  ---------- -----

   A. All architectures must provide some instructions that access data in
      memory - either as part of a computational instruction, or via separate
      load/store or push/pop instructions.  A final architectural issue is
      how is the address of a memory operand to be specified?

   B. The set of addressing modes on the VAX represents the most inclusive
      answer to this question.  Most machines include only a subset of all
      the possible modes.  The following are quite common:

        Immediate
        Absolute
        PC Relative
        Register indirect (deferred)
        Register + displacement

   C. In keeping with the RISC philosophy, RISCs typically have a smaller
      subset:

        Immediate - often for just some instructions
        Register + displacement (yields register indirect with displacement = 0)
        PC Relative (some RISCs - but only for branches on MIPS)

Copyright ©1999 - Russell C. Bjork