URL
https://opencores.org/ocsvn/tinycpu/tinycpu/trunk
Subversion Repositories tinycpu
[/] [tinycpu/] [trunk/] [docs/] [design.md.txt] - Rev 28
Go to most recent revision | Compare with Previous | Blame | View Log
This is the design of TinyCPU. It's goals are as follows:
1. 8-bit registers and operations (8 bit processor)
2. 16-bit address bus
3. fixed 16-bit instruction length
4. use a small amount of "rich" instructions to do powerful things
5. 1 instruction per clock cycle
Relative moves:
In order to provide uesfulness to the segment-carryover feature, there are a few options for moving a "relative" amount to a register, including IP and SP
A relative move differs in most of the opcodes in that the relative factor is treated as a signed value.
so for instance, a
mov r0,50
mov_relative r0, -10
in the ned, r0 will end up being 40. Although this feature won't see much use in general registers, IP and SP are special because of the option of using the
segment-carryover feature. This means that SP and IP, while being 8-bit registers, can function very similar to a 16-bit register, enabling full usage of the available address space.
Register list:
r0-r5 general purpose registers
sp stack pointer (represented as r6)
ip instruction pointer register (represented as r7)
cs, ds, es, ss segment registers (code segment, data segment, extra segment, stack segment)
tr truth register for conditionals
general opcode format
first byte:
first 4 bits: actual instruction
next 3 bits: (target) register
last 1 bit: conditional
second byte:
first 1 bit: second portion of condition (if not immediate) (1 for only if false)
next 1 bit: use extra segment
next 3 bits: other register. If not 3rd register
last 3 bits: extra opcode information or third register. such as for ADD it could be target=source+third_register
...or second byte is immediate value
For opcodes requiring 3 registers but without room, the target opcode is assume to be the second operation. Such as for AND, target=source AND target
short list of instructions: (not final, still planning)
immediates:
1. move reg, immediate
2. move [reg], immediate
3. push and move reg, immediate (or call immediate)
4. move (relative) reg, immediate
mini-group 5. Root opcode is 5, register is to tell which opcode( up to 8). No register room, only immediate
push immedate
XX
XX
XX
XX
XX
XX
XX
groups: (limited to 2 registers and no immediates. each group has 8 opcodes)
group 1:
move(store) [reg],reg
move(load) reg,[reg]
out reg1,reg2 (output to port reg1 value reg2)
in reg1,reg2 (input from port reg2 and store in reg1)
XX
XX
move segmentreg,reg
move reg,segmentreg
group 2:
and reg1,reg2 (reg1=reg1 and reg2)
or reg, reg
xor reg,reg
not reg1,reg2 (reg1=not reg2)
left shift reg,reg
right shift reg,reg
rotate right reg,reg
rotate left reg,reg
group 3: compares
is greater than reg1,reg2 (TR=reg1>reg2)
is greater or equal to reg,reg
is less than reg,reg
is less than or equal to reg,reg
is equal to reg,reg
is not equal to reg,reg
equals 0 reg
not equals 0 reg
group 4:
push segmentreg
pop segmentreg
push and move reg, reg (or call reg)
exchange reg,reg
exchange reg,seg
XX
XX
group 5:
XX
XX
far jmp reg1, reg2 (CS=reg1 and IP=reg2)
far call reg1,reg2
far jmp [reg] (first byte is CS, second byte is IP)
push extended segmentreg, reg (equivalent to push seg; push reg)
pop extended segmentreg, reg (equivalent to pop reg; pop seg)
reset processor (will completely reset the processor to starting state, but not RAM or anything else)
group 6:
set default register bank to 0 (can be condensed to 1 opcode)
set default register bank to 1
push extended reg, reg
pop extended reg,reg
enable carryover seg
disable carryover seg
mov relative reg, reg
exchange reg, reg
super group: Super groups only have room for 1 register argument. Each subgroup has 8 opcodes, capable of 8 subgroups.
subgroup 0:
push reg
pop reg
set TR
reset TR
increment reg
decrement reg
set register bank 0
set register bank 1
subgroup 1:
enable carryover seg
disable carryover seg
3 register instructions:
1. add reg1, reg2, reg3 (reg1=reg2+reg3)
2. sub reg1, reg2, reg3
opcodes used: 14 of 16. 2 more opcodes available. Decide what to do with the room later.
Possible canidates for opcode compression include
* equals 0 and not equals 0 (room for 7 sub-opcodes each) (not doing that because it'd screw with the easy ALU code
conditionals
0 -- always
1 -- only if true
for only if false, there should basically be another compare or if applicable an always afterwards
limitations that shouldn't be passed with instructions
* Doing 2 memory references
* pushing a memory reference (equates to 2 memory references)
Note it is possible however to read and write 16bits at one time to the memory to consecutive addresses that are 16-bit aligned.
segments:
DS is used in all "normal" memory references
SS is used in all push and pop instructions
ES is used when the ExtraSegment bit is set for either push/pop or normal memory references
CS is only used for fetching instructions
Segment carryover:
In order to overcome the limitations of only having a 256 byte segment, there is a workaround option to "pretend" that IP is a 16 bit register.
When CS carryover is enabled, when IP rollover from 255 to 0 or whatever, CS will be incremented. This makes it so that if you start at address 0:0.
you can continue as far as needed into the address space without having to do ugly far jumps at each of the borders.
Carryover can only be done on CS and SS. The required circuitry is not implemented for DS or ES due to an extreme level of complexity required for it, also
it would only lead to unncessarily complex code
Also of note is that `move relative` implements a "carryover" component. This component will work on either IP or SP, and uses CS and SS respectively.
If used on other registers, there will be no carry over functionality, though it can be used as an easy way to add or subtract an immediate from a register.
States needed:
0. reset
1. decode current instruction (All without memory capable within 1 clock cycle)
2. increment IP(and SP if needed) and fetch next instruction
3. Write 1 register to memory
4. Read 1 register from memory
5. Write 2 registers to memory
6. Read 2 registers from memory
7. Write 1 register to memory and setup increment of sp
8. Write 2 registers to memory and setup double increment of sp
9. Read 1 register from memory and setup decrement of sp
10. Read 2 registers from memory and setup double decrement of sp
11.
registerfile map:
0000: general r0
0001: general r1
0010: general r2
0011: general r3
0100: general r4
0101: general r5
0110: SP (r6)
0111: IP (r7)
1000: second bank r0
1001: second bank r1
1010: second bank r2
1011: second bank r3
1100: CS
1101: DS
1110: ES
1111: SS
Banking works like if(regnumber(2) = '0') then regnumber(3)=regbank; end if;
ALU operations
00000 and reg1,reg2 (reg1=reg1 and reg2)
00001 or reg, reg
00010 xor reg,reg
00011 not reg1,reg2 (reg1=not reg2)
00100 left shift reg,reg (logical)
00101 right shift reg,reg (logical)
00110 rotate right reg,reg
00111 rotate left reg,reg
01000 is greater than reg1,reg2 (TR=reg1>reg2)
01001 is greater or equal to reg,reg
01010 is less than reg,reg
01011 is less than or equal to reg,reg
01100 is equal to reg,reg
01101 is not equal to reg,reg
01110 equals 0 reg
01111 not equals 0 reg
10000 Set TR
10001 Reset TR
10011 Increment
10010 Decrement
10100 Add
10101 Subtract
Alignment restrictions:
In general, their is very few times that a full 16-bit read or 16-bit write is done. These are the times:
* Extended push
* Extended pop
* instruction fetch
Because of this, and because I want for 2 clock cycles to be the longest instruction, I must place some alignment restrictions on the CPU
So, IP must be aligned to a 16-bit address (must be an even number). And SP must also be aligned to a 16-bit address.
Though I don't plan on putting any "real" restriction to setting it to an odd address, nothing will actually work right.
Stack Details:
Because of the need for 16-bit writes and reads of the stack, even though we're usually only using 8-bit values, we end up pushing 2 bytes at one time always.
Stack is oppositely done from the 8086. push X will move X to SS:SP and then increment SP by 2.
Let's take an example program:
--SS is 0
mov sp, 10
push 0xff
after this, 0x00FF will be moved to SS:SP (0x0010) and then sp will be incremented by 2. If we push an 8-bit value, the value is put in the least-significant byte, and the MSB is 0
On Reset:
On reset, all general registers are set to 0
CS is set to 1, IP is set to 0. SS is set to 2 and SP is set to 0.
Carryover is set on CS and not set on SS. DS and ES is 0. TR is false.
Register bank 0 is selected.
Electrical operation:
On power-on, RESET should be high for at least 2 clock cycles. HOLD can optionally be high as well after these two clock cycles.
When HOLD is no longer needed, it should just be turned low and an extra clock cycle should be waited on for it to return to RESET state
When RESET is held low, the processor will execute. It takes 3 clock cycles for the processor to "catch up" to actually executing instructions
Implemented opcode list:
legend:
r = register choice
C = conditional portion
s = segment register choice
i = immediate data
N = not used
o = opcode choice (for groups)
0000 rrrC iiii iiii
mov reg, immediate
0001 rrrC iiii iiii
mov [reg], immediate
group 3 comparions
0011 rrrC Crrr Nooo
opcode choices
000: is greater than reg1,reg2 (TR=reg1>reg2)
001: is greater or equal to reg,reg
010: is less than reg,reg
011: is less than or equal to reg,reg
100: is equal to reg,reg
101: is not equal to reg,reg
110: equals 0 reg
111: not equals 0 reg
group 4 bitwise
0100 rrrC Crrr Nooo
opcode choices
000: and reg1,reg2 (reg1=reg1 and reg2)
001: or reg, reg
010: xor reg,reg
011: not reg1,reg2 (reg1=not reg2)
100: left shift reg,reg
101: right shift reg,reg
110: rotate right reg,reg
111: rotate left reg,reg
Go to most recent revision | Compare with Previous | Blame | View Log