The SPU Mark II Architecture

Documentation Style

Numbers

Numbers are documented in three ways:

Decimal numbers are written the usual style. Hexadecimal numbers are prefixed by a 0x. Binary numbers are postfixed with a subscript ₂. If both decimal and binary notation are given, the decimal notation is postfixed with a subscript ₁₀ to make the difference clear.

Examples:

• Decimal numbers: 10, 23, 44₁₀
• Hexadecimal: 0x10, 0xFF, 0xCC3D
• Binary: 10₂, 1100101₂
• Mixed binary and decimal: 10₂ (2₁₀), 1100₂ (12₁₀)

Bit Ranges

This document uses some special notations to define bit ranges or indices.

[n] is a placeholder for the nth bit in a word.

[n:m] is a placeholder for the bit range from the nth (highest significant bit) to the mth (lowest significant bit).

So [3] means the the bit with the significance of 2³ and [7:4] is the upper nibble of a 8 bit word.

Wording

Undefined (value)

If a value is defined undefined, its initial value may be any possible value. The value may change between power cycles or even after a reset. Each bit must be considered random.

Undefined behavior

Undefined behavior is used similar to the way the C standard uses this word. It means that if a situation happens where undefined behavior would occur, the results of the operation may be anything. This can be a no-op, any state or memory change or even a CPU hang or hard reset (may even requires a power cycle).

CPU Variants

This document specifies two different variants for the SPU Mark II architecture:

• SPU Mark II
• SPU Mark II-L

The SPU Mark II-L is a reduced version of the default instruction set and features no interrupt handling, thus making an implementation of the ISA much easier.

SPU Mark II Architecture

The SPU Mark II (Stack Processing Unit Mark II) is a 16 bit processor that, in contrary to the most popular CPUs, works primarily with a stack instead of registers. It features a RISC instruction set with highly configurable instructions. Although being a stack processor it still requires some basic registers to work. These registers are either accessed indirectly (like the IP register) or by special instructions (like BP or SP register).

Each instruction is composed of a configuration part and a command part. Commands are the actual function of the instruction like STORE8 or MUL. Each command has two input variables input0 and input1 which contain the two command parameters. A command also has an output value.

In which way the command parameters are filled and the result is processed is defined by the configuration bits of the instruction. These bits allow conditional execution, input parameter modification, affect flags and define how the result of the command is processed.

As each instruction may be conditional, there are no special conditional jump commands. In fact, there isn't even a jump command at all. A jump is done by taking the output of a command to be the next instruction pointer.

Thus, the most simple COPY command can be used already for a whole set of different operations: jmp $imm, push $imm, pop and many more.

Basic Properties

Word Encoding and Signedness

The SPU Mark II uses the little endian encoding, so the less significant byte is at the lower address, the more significant byte at the higher address.

Integer arithmetic uses two-complement signed integers. This allows most arithmetic instructions to be used with signed and unsigned values.

Memory Access

The cpu only allows aligned memory access for word access. Unaligned word access must be programmed manually.

CPU Registers

Stack Pointer (SP)

16 bit register storing the address of the topmost value of the stack. The stack grows downwards, so a push operation decrements the SP by two and then stores a value to the decremented address. A pop or peek operation reads the value from SP, and for pop, SP will be incremented by 2.

Bit Fields	Description
[0]	reserved, must be zero
[15:1]	aligned stack top address

Initial Value: Undefined

Base Pointer (BP)

16 bit register that may be used for indirect addressing via GET and SET commands and may be used as a frame pointer or index register.

Bit Fields	Description
[0]	reserved, must be zero
[15:1]	aligned address

Initial Value: Undefined

Instruction Pointer (IP)

16 bit register pointing to the instruction to be executed next.

Bit Fields	Description
[0]	reserved, must be zero
[15:1]	aligned instruction address

Initial Value: Undefined

Flag Register (FR)

16 bit register saving CPU state and interrupt system

Bit Range	Name	Description
[0]	Z	Zero Flag
[1]	N	Negative Flag
[2]	C	Carry
[3]	CE	Carry Enable
[7:4]	I[3:0]	Interrupt Enabled bitfield for interrupts 4 to 7
[15:8]	-	Reserved, must be 0.

Initial Value: 0x0000

Note: The flags I[3:0] are not available in SPU Mark II-L

Interrupt Register (IR)

16 bit register storing internal interrupt information.

Bit Range	Name	Description
[0]	RST	Reset was triggered
[1]	NMI	Non-maskable interrupt triggered
[2]	BUS	Bus error triggered
[3]	-	Reserved, must be 0
[4]	ARITH	ARITH was triggered
[5]	SOFTWARE	SOFTWARE was triggered
[6]	RESERVED	RESERVED was triggerd
[7]	IRQ	IRQ was triggered
[15:8]	-	Reserved (must be 0)

Initial Value: 0x0001 (Reset interrupt triggered)

Note: Not available in SPU Mark II-L

Instruction Encoding

Instructions use 16 bit opcodes organized in different bit fields defining the behaviour of the instruction.

Bit Range	Description
[2:0]	Execution Conditional
[4:3]	Input 0 Behaviour
[6:5]	Input 1 Behaviour
[7:7]	Flag Modification Behaviour
[8:8]	Output Behaviour
[14:9]	Command
[15:15]	Reserved for future use (must be 0)

Conditional Execution

This field determines when the command is executed or ignored. The execution is dependent on the current state of the flags.

This allows conditional execution of all possible opcodes.

Value	Enumeration	Description
000₂ (0₁₀)	Always	The command is always executed
001₂ (1₁₀)	=0	The command is executed when result is zero (Z=1)
010₂ (2₁₀)	≠0	The command is executed when result is not zero (Z=0)
011₂ (3₁₀)	>0	The command is executed when result is positive (Z=0 and N=0)
100₂ (4₁₀)	<0	The command is executed when result is less than zero (N=1)
101₂ (5₁₀)	≥0	The command is executed when result is zero or positive (Z=1 or N=0)
110₂ (6₁₀)	≤0	The command is executed when result is zero or negative (Z=1 or N=1)
111₂ (7₁₀)	Overflow	The command is executed when Carry is set.

Argument Input 0 and 1

These two fields define what arguments are provided to the executed command.

Value	Enumeration	Description
00₂ (0₁₀)	Zero	The input register will be zero.
01₂ (1₁₀)	Immediate	The input registers value is located after this command.
10₂ (2₁₀)	Peek	The input register will be the stack top.
11₂ (3₁₀)	Pop	The input register will be popped from the stack.

Zero means that the argument will be zero, Immediate means that it will be fetched from the instruction pointer (it is located behind the opcode in memory). Peek will take the argument from the stack top, but won't change the stack and Pop will take the argument from the stack top and decreases the stack pointer.

input0 is fetched before input1 so when both arguments pop a value, input0 receives the stack top and input1 receives the value one below the stack top. Likewise, when both arguments use the Immediate option, the value for input0 must located directly after the opcode, input1 directly after input0.

Flag Modification

When the flag modification is enabled, the current N and Z flags will be overwritten by this command. Otherwise the flags stay as they were before the instruction.

Value	Enumeration	Description
0₂ (0₁₀)	No	The flags won't be modified.
1₂ (1₁₀)	Yes	The flags will be set according to the command output.

The flags are modified according to this table:

Flag	Condition
Z	output[15:0] = 0
N	output[15] = 1
C	unchanged
CE	unchanged
I	unchanged

Output Behaviour

Each command may output a value which can be processed in various ways. The output could be pushed to the stack, the command could be made into a jump or the output could be ignored.

Value	Enumeration	Description
0₂ (0₁₀)	Discard	The command output will be ignored.
1₂ (1₁₀)	Push	The command output will be pushed to the stack.

For Jump Relative, the instruction pointer will point to the next instruction plus output words. output is considered a two-complements signed number. This differs from the Jump behavior which takes an address, not a word offset.

Commands

Commands are what define the core behaviour of the opcode. They allow arithmetics, modification of memory, changing system registers and so on.

Some hints on notation: - MEM16[x] is the 16 bit word at address x - MEM8[x] is the 8 bit word at address x

Executing instructions not defined in the list below are considered undefined behaviour and may do any behaviour.

If a register has a subscript ₀, it means that this refers to the value before the instruction is executed. If there is a subscript ₁, it means that this refers to the value after the instruction was executed. Note: ₀ still

Value	Name	Pseudo-Code
000000₂ (0₁₀)	COPY	output = input0
000001₂ (1₁₀)	reserved
000010₂ (2₁₀)	GET	output = MEM16[BP + 2 * input0]
000011₂ (3₁₀)	SET	output = input1; MEM16[BP + 2 * input0] = input1
000100₂ (4₁₀)	STORE8	output = input0 & 0xFF; MEM8[input1] = input0
000101₂ (5₁₀)	STORE16	output = input0; MEM16[input1] = input0
000110₂ (6₁₀)	LOAD8	output = MEM8[input0]
000111₂ (7₁₀)	LOAD16	output = MEM16[input0]
001000₂ (8₁₀)	CPUID	See below
001001₂ (9₁₀)	HALT	See below
001010₂ (10₁₀)	FRGET	output = (FR & ~input1)
001011₂ (11₁₀)	FRSET	output = FR₀; FR₁ = (input0 & ~input1) \| (FR₀ & input1)
001100₂ (12₁₀)	BPGET	output = BP
001101₂ (13₁₀)	BPSET	output = BP₀; BP₁ = input0
001110₂ (14₁₀)	SPGET	output = SP
001111₂ (15₁₀)	SPSET	output = SP₀; SP₁ = input0
010000₂ (16₁₀)	ADD	output = input0 + input1 + (C & CE)
010001₂ (17₁₀)	SUB	output = input0 - input1 - (C & CE)
010010₂ (18₁₀)	MUL	output = input0 * input1
010011₂ (19₁₀)	DIV	output = input0 / input1
010100₂ (20₁₀)	MOD	output = input0 % input1
010101₂ (21₁₀)	AND	output = input0 & input1
010110₂ (22₁₀)	OR	output = input0 \| input1
010111₂ (23₁₀)	XOR	output = input0 ^ input1
011000₂ (24₁₀)	NOT	output = ~input0
011001₂ (25₁₀)	SIGNEXT	output = if(input[7] = 1) (0xFF00 \| input0) else (input0 & 0xFF)
011010₂ (26₁₀)	ROL	output = concat(input0[14:0], input0[15])
011011₂ (27₁₀)	ROR	output = concat(input0[0], input0[15:1])
011100₂ (28₁₀)	BSWAP	output = concat(input0[7:0], input0[15:8])
011101₂ (29₁₀)	ASR	output = concat(input0[15], input0[15:1])
011110₂ (30₁₀)	LSL	output = concat(input0[14:0], '0')
011111₂ (31₁₀)	LSR	output = concat('0', input0[15:1])
100000₂ (32₁₀)	SETIP	output = IP₀; IP₁ = input0; FR₁ = FR₀ \| input1
100001₂ (33₁₀)	ADDIP	output = IP₀; IP₁ = IP₀ + input0; FR₁ = FR₀ \| input1
100010₂ (34₁₀)	INTR	output = input0; IR = IR \| input0
100011₂ (35₁₀)	reserved
100100₂ (36₁₀)	reserved
100101₂ (37₁₀)	reserved
100110₂ (38₁₀)	reserved
100111₂ (39₁₀)	reserved
101000₂ (40₁₀)	reserved
101001₂ (41₁₀)	reserved
101010₂ (42₁₀)	reserved
101011₂ (43₁₀)	reserved
101100₂ (44₁₀)	reserved
101101₂ (45₁₀)	reserved
101110₂ (46₁₀)	reserved
101111₂ (47₁₀)	reserved
110000₂ (48₁₀)	reserved
110001₂ (49₁₀)	reserved
110010₂ (50₁₀)	reserved
110011₂ (51₁₀)	reserved
110100₂ (52₁₀)	reserved
110101₂ (53₁₀)	reserved
110110₂ (54₁₀)	reserved
110111₂ (55₁₀)	reserved
111000₂ (56₁₀)	reserved
111001₂ (57₁₀)	reserved
111010₂ (58₁₀)	reserved
111011₂ (59₁₀)	reserved
111100₂ (60₁₀)	reserved
111101₂ (61₁₀)	reserved
111110₂ (62₁₀)	reserved
111111₂ (63₁₀)	reserved

MUL, DIV and MOD use signed values as input and output. It is not possible to get the upper 16 bit of the multiplication result.

ADD and SUB will add/subtract 1 more if the Carry and Carry Enabled flag are both set.

ADD, SUB, MUL will modify the Carry flag, even if Carry Enabled flag or Modify Flags instruction field is disabled:

• When ADD is overflowing and setting a virtual 17th bit, carry will be set
• When SUB is overflowing and setting a virtual 17th bit, carry will be set
• When MUL is overflowig and setting any bits in the upper half of the virtual 32 bit result, carry will be set.

In all other cases when one of the carry modifying command is invoked, carry is cleared. Other commands then those specified above will not modify carry in any way.

CPUID

This instruction is meant to return the current set of CPU features or possible future extensions. For now, this instruction requires both input0 and input1 to be zero, the output will be zero as well.

HALT

Stops execution of instructions until an interrupt happens and interrupts are enabled.

Memory Access

Only 2-aligned access to memory is possible with code or data. Only exception are the STORE8 and LOAD8 commands which allow 1-aligned memory access.

When accessing memory with alignment, the least significant address bit is reserved and must be 0. If the bit is 1, the behavior is undefined.

Fetch-Execute-Cycle

This pseudocode documents what the CPU does in detail when execution a single instruction.

if (IR & 1) != 0:
    IP := fetch(0x0000)
    FR := 0x0000
    IR := 0x0000

while IR != 0:
    intr_bit := indexOfHighestSetBit(IR)
    IR &= ~(1<<intr_bit)         # clear "interrupt triggered"
    if intr_bit < 4 or (FR & (1<<intr_bit)) != 0: # if interrupt is not masked
        push(1<<intr_bit)
        push(IP)
        IP := fetch(2 * intr_bit)   # fetch ISR handler
        if intr_bit >= 4:           # if interrupt is maskable
            FR &= ~(1<<intr_bit)      # unmask interrupt

instruction := fetch(IP)
IP += 2

if isExecuted(instruction, FR):
    input0 := fetchInput(instruction.input0)
    input1 := fetchInput(instruction.input1)

    output := instruction.command(input0, input1)

    select instruction.output:
        when 'push':
            push(output)
        when 'discard':
            # ignore
        when 'jmp':
            IP := output
        when 'rjmp':        
            IP += 2 * output

    if instruction.modifyFlags:
        FR.Z = (output == 0x0000)
        FR.N = (output >= 0x8000)
else
    if instruction.input0 == IMM:
        IP += 2
    if instruction.input1 == IMM:
        IP += 2

TODO: Add handling of BUS fault!

For the non-interrupt version, the state machine is simpler:

if RST is high:
    IP := 0x0000
    FR := 0x0000

instruction := fetch(IP)
IP += 2

if isExecuted(instruction, FR):
    input0 := fetchInput(instruction.input0)
    input1 := fetchInput(instruction.input1)

    output := instruction.command(input0, input1)

    select instruction.output:
        when 'push':
            push(output)
        when 'discard':
            # ignore

    if instruction.modifyFlags:
        FR.Z = (output == 0x0000)
        FR.N = (output >= 0x8000)
else
    if instruction.input0 == IMM:
        IP += 2
    if instruction.input1 == IMM:
        IP += 2

Interrupt handling

Note: Interrupt handling is not available in SPU Mark II-L

The SPU Mark II provides 8 possible interrupts, four unmasked and four masked interrupts.

When an interrupt is triggered the CPU pushes the current instruction pointer to the stack and fetches the new instruction Pointer from the interrupt table. Then the flag in the Interrupt Register is cleared as well as the mask bit in the Flag Register (if applicable).

The reset interrupt is a special interrupt that does not push the return address to the stack. It also resets the Interrupt Register and the Flag Register.

Masking

If an interrupt is masked via the Flag Register (corresponding bit is 0) it won't be triggered (the Interrupt Register bit can't be set).

Interrupt Table

It is possible to assign each interrupt another handler address. The entry points for those handlers are stored in the Interrupt Table at memory location 0x0000:

#	Interrupt	Routine Pointer
0	Reset	0x00
1	NMI	0x02
2	BUS	0x04
3	RESERVED	0x06
4	ARITH	0x08
5	SOFTWARE	0x0A
6	RESERVED	0x0C
7	IRQ	0x0E

Reset

This interrupt resets the CPU and defines the program entry point.

NMI

This interrupt is the non-maskable external interrupt. If the NMI pin is signalled, the interrupt will be triggered.

BUS

This interrupt is a non-maskable external interrupt.

The BUS interrupt is meant for an MMU interface and will prevent the currently executed instruction from having any side effects.

Remarks: It is required that the BUS interrupt happens while doing a memory operation. If after a memory read or write cycle the BUS pin is signalled, the CPU will assume as bus error and will trigger this interrupt.

ARITH

This interrupt is triggered when an error happens in the ALU. The reasons may be:

• Division by zero

SOFTWARE

This interrupt is meant to be triggered by executing CPUCTRL and will never be triggered by either peripherial hardware or internal circumstances.

IRQ

This interrupt is the maskable external interrupt. If the IRQ pin is signalled, the interrupt will be triggered.

Priorities

Before execution of each instruction the cpu checks if any interrupt is triggered. The handler for the lowest interrupt triggered will then be activated.

Chip I/Os

• 16 output address lanes
• 16 in/out data lanes
• output WE, RE signal
• input HOLD signal
• input CLK signal
• input RST signal
• input NMI signal
• input BUS signal
• input IRQ signal
• output TYPE signal (tells if memory access is for code or data)

Changelog

v1.11

• Removes CPUCTRL instruction
• Removes output behaviour jump and rjmp
• Changes instruction encoding, Command now has 6 bits, Output Behaviour has now 1 bit.
• Introduces new HALT command (replaces CPUCTRL)
• Swaps input semantics for STORE8 and STORE16 (reusing the address is more common than reusing the value)
• Introduces new SETPC and ADDPC command (replaces the previous Output Behaviour jump and rjmp)
• Introduces new command INTR

TODO: Document HALT behaviour/state as well as interrupt resumption

TODO: Reintroduce a way to trigger software interrupts

TODO: Change interrupt behaviour - Interrupts now push FR, IP, then jump to the ISR => iret = "setpc [i0:pop] [i1:pop] - Interrupts will disable all interrupts when entering the ISR => no nested/recursive interrupts => iret will restore the interrupt enable state

TODO: Improve description of the state machine + transitions

TODO: Make shifts take input1 the number of bits to shift.

v1.10

• Introduces CPUID and CPUCTRL
• Introduces SOFTWARE interrupt.

v1.9

• Introduces the concept of Carry
• Adds new execution modifier Overflow that allows checking for carry
• Changes ADD and SUB to respect the Carry and Carry Enable bit
• Changes ADD, SUB and MUL to change the Carry bit

v1.8

• Removes NEG command
• Replaces free command slot with SIGNEXT
• Changes pseudo code for Fetch-Execute-Cycle examples to a python-like language
• Improves register description for BP
• Improves register description for SP
• Adds some clarification on the pseudo code in the Command table

v1.7

• Makes frset, spset, bpset return the previous value instead of the newly set. This allows improved or shorting handling of safed parameters.

v1.6

• Fixed typo in IP description

v1.5

• Textual architecture description
• Improves overall document quality.
• Introduces SPU Mark II-L (Version with no interrupt handling)
• FRGET now also provides a masked register access

v1.4

• Reorderes IRQ table
• FIX: Missing interrupt 6 from IRQ table
• Adjusts IR, FR
• Makes BUS nonmaskable
• Allows FRSET to have a masked register access

v1.3

• Defines unaligned memory access undefined behaviour.
• Packs all IRQs into a single one. Requires use of a memory mapped interrupt controller, but is ultimately less limiting.
• Reduces interrupt count to 8 (still two reserved interrupts)
• Adds interrupt descriptions.
• Makes BUS interrupt an external interrupt for a MMU
• Makes BUS prevent all effects from the current instruction

v1.2

• Adds preliminary I/O description

v1.1

• Adds FR, BP, SP, IP register names to register description
• Adds FRGET, FRSET instruction
• Introduces interrupt handling description

v1.0

• Initial version