Re: Instruction Parcel Size

On 2025-03-08 8:43 p.m., MitchAlsup1 wrote:

On Sun, 9 Mar 2025 1:15:14 +0000, Robert Finch wrote:
 

Recently started Q+2 development.
>
Trying to get code density closer to something like the 68k or VAX.
Sounds like My66000 also has good code density using 32-bit parcels. If
32-bit parcels work well, I have thought to try 24-bit parcels.

 Ok, go all Quadriblock on me .... see if I care !
 

Decided to stay away from other odd sized parcels which create
addressing issues. Currently 3 sizes of instructions: 24 / 48 and
96-bit. The 96-bit instructions are usually for encoding a 64-bit
immediate. The first two opcode bits determine the size. 00=24 bit,
01=48 bit, 10=96 bit, 11 (reserved)

 You might find 72-bit instructions useful in carrying a 32-bit
immediate.

Yeah, I was sure I wanted to support 23-bit immediates (in the 48-bit format) and 64-bit immediates (96-bit format). But had not decided on supporting other sizes perhaps 128-bit immediates.A 72-bit format would allow over 40 bits for immediates.

 

ADD, AND, OR, EOR, CMP have 24-bit instruction forms:
iiiii-aaaaa-ttttt-ooooooo-00    <- immediate
bbbbb-aaaaa-ttttt-ooooooo-00 <- register
24-bit instruction forms allow using only the first 32 registers.

 A 6-bit OpCode might let 1 more bit into immediate, or similar
sign control over register operand. Remember, this is the highly
used OpCode category. So, we have 32 Imm5(6) OpCodes.

Yeah, I have been scratching my head over how to free up another opcode bit. But the decode is simple. ooooooo always refers to the same opcode even if longer/shorter forms do not make sense.
There are only about 23 opcodes free out of 128. That is 105 instructions. Sometimes the opcode bits are used to determine the precision of the operation. So, ooooooo is sometimes ooooopp. The data type is also sometimes encoded in ooooooo (ooooTTT).

 

ADD, AND, OR, EOR, CMP 48-bit instruction forms:
fffffff-oooo-ccccccc-bbbbbbb-aaaaaaa-ttttttt-ooooooo-01
48-bit instruction forms support a sign control bit on the register
spec, along with 64 registers.

 It really looks like you are forming 2×24-bit instructions into
(wait for it) 2×24-bit containers. Sign control on operands is
useful 5-register operands may not be so. I am not against this
format, but I think you are wasting a lot of entropy here.

It is almost two instructions, but there is only a single target register. That way in some cases up to 8 ops per clock can be processed instead of just 4. The compiler does not do a good job of making use of the dual ops yet.
There are only four register operands. fffffff is a function code. oooo
is a second operation between the result of (a op b) and c. The seven bit register spec includes sign control (there are only 64 GPRs). aaaaaaa is really saaaaaa.
Register ops do not needs anything beyond 48 bits so there is a lot wasted as the length field could also be 10 or 11 both of which are not used.

 

LOAD / STORE word size have 24-bit forms
ddddd-aaaaa-ttttt-ooooooo-00

 Where do you get stack and structure displacements ?? This is one
place My 66000 ISA is significantly better than RISC-V.

ddddd is the displacement which is multiplied by 8 (word size).

 

Conditional Branches (compare and branch) are 48-bit
pp-R-TTTTTTTTTTTTTTTTTTT-aaaaaa-bbbbbb-A-ffff-ooooooo-01

 Careful choice of oooooo may allow it to contain the condition
in the ffff field expanding the displacement to 25-effective
bits.

There is a whole row (8 opcodes) in the opcode table dedicated to the branch data type. The data type could be integer, unsigned integer, float, decimal float, posit or capability.
ooooooo = 0101TTT (the same breakdown is used for loads and stores).
TTT
000 - unsigned
001 - signed
010 - float
011 - decimal float
100 - posit
101 - capability
1000TTT = loads
1001TTT = stores
It may be possible to gain a bit. I did not think it was critical. There are 19 displacement bits. 'A' indicates relative or absolute addressing for the target. Only relative addressing is really needed so a bit could be gained there as well. 'R' indicates to use register 'c' as the target, another bit that may not be needed. But it allows branch to register / conditional return.

 

With load / store / basic arithmetic as 24-bit and 48-bit
compare-and-branch a good portion of instructions should occupy the same
or less storage space than a 32-bit ISA.

 Questions remain wrt floating point constants and large integer
constants.

iiiiiiiiiiiiiiiiiiiiiii-pp-saaaaaa-stttttt-ooooooo-01    <- immediate
iii--65 bits--iii-pp-saaaaaa-stttttt-ooooooo-10    <- immediate
Immediate postfixes may override register specs. Useful for encoding float constants too.
iiiiiiiiiiiii-SS-1111100-00 (18-bit immediate postfix - 24 bit insn
iii--36 bits--iii-SS-1111100-01 (41-bit immediate postfix - 48 bit insn
iii--79 bits--iii-SS-1111100-10 (89-bit immediate postfix - 96 bit insn
SS selects which register spec to override.

 

I have not written the assembler yet, so nothing to measure.

 The data will be interesting.