I've been working on implementing callable words in #uxntal.
Instead of the typical PUSH->POP->JSR sequence each time you want to run a routine, these tokens are not pushing to the stack - they'll run immediately.
They'll be impractical for doing pointer arithmetic, but save 1 byte for every subroutine call, and will save millions of cycles in a project of the size of a text editor or drawing software.
They don't break compatibility with old roms, and make source files a bit more readable.
I'm wondering a little about JCN2… we have a lot more JCNs in the project/ dir's files (they're 18× the abundance of JCN2), so while opcode 40 does save cycles, they take up the same amount of space.
One idea that keeps coming back to me is shoe-horning in a opcode for a conditional JSR2 instead of conditional JMP2… I struggle with what the opcode's name could be, but if we use it as “?square” in Uxntal then that might not be so bad!
@tty yep, they fill the unused opcodes, they follow the same format the LIT opcodes. I'll slowly roll out this feature as I experiment with it, I won't migrate apps to the new opcodes to give time to test this further, but I'll add it to uxnasm and drifblim soon so anyone can start noodling with it.
@alderwick @tty it's true that that one won't save space, but it's still a crutial one in terms of cycles in JCN because it's BY FAR, the most used opcode.
It's tricky to know how much having a conditional JSR would save us, because the workarounds to not having it are so varied. But I shouldn't have gone for JCN as the opcode to eliminate to make room for it! 😅
@alderwick @tty I didn't think you could do that with switch cases! Won't it be slower tho? I really struggled with where to put this stuff. I chucked it into 0x00, but I can't be sure if it wouldn't be faster above the mode checks. 🤔
@tty I can't say which approach in the C code would be the fastest, thanks to compiler magic and so on, but I wouldn't say the situation is dire. My first optimisation would be to replace the if-else with another switch: http://okturing.com/src/14768/body (untested)
POP is pretty rare in your pie chart too, so even if POPs were made slower to implement the extra features, that isn't that bad either.
@tty It's super late for me, but my point is we can make it fast one way or another! It might need more rejigging anyway to get the best performance, so my link is just for illustration. Let's focus on the features we want first, then we can optimise it later.
@alderwick @tty good point 😀 Right now I'm just testing things. I'm not sure if that'd be interesting to you but I've been reading the book Stack Machines: A New Wave, and there's a TON of useful insights in there, maybe you'll get a kick out of it too.
I noticed that the entry for JSI on https://wiki.xxiivv.com/site/uxntal_reference.html says "Pushes the program counter to the return-stack" first, but wouldn't it be pushing PC+2 onto the stack, to skip the address to jump to?
@tty technically yes, I wasn't sure how technical I wanted to get with this. Literals are also doing this sort of thing where they move the PC over the length of the literal, I thought it might be best to keep it light? What do you prefer?
@parnikkapore @ddlyh @tty @makeworld I'll keep both, having gone over a couple of apps now to see how they convert, there's many situations where I can't just use one of the new words. Anything even a tiny bit fancy like UI components, you won't be able to handle with the new opcodes alone.
@ddlyh @tty @makeworld JSR2 would be useful if you want to do some pointer arithmetic (@JumpTableRoot ADD2 JSR2); if it was removed, you will have to STA2 to the address bytes to do this
But yeah, I think it's a better idea to keep only one. Maybe having both would be more "pona", but idk
this feels a lot like the "guix environment" vs "guix shell" situation
Devine Lu Linvega
Unknown parent • • •a0 12 34 = #1234, for reference
20 xx xx = !square, think JMP2
40 xx xx = ?square, think JCN2
60 xx xx = square, think JSR2
Andy Alderwick
in reply to Devine Lu Linvega • • •I'm wondering a little about JCN2… we have a lot more JCNs in the project/ dir's files (they're 18× the abundance of JCN2), so while opcode 40 does save cycles, they take up the same amount of space.
One idea that keeps coming back to me is shoe-horning in a opcode for a conditional JSR2 instead of conditional JMP2… I struggle with what the opcode's name could be, but if we use it as “?square” in Uxntal then that might not be so bad!
Devine Lu Linvega
Unknown parent • • •Devine Lu Linvega
in reply to Andy Alderwick • • •Devine Lu Linvega
in reply to Devine Lu Linvega • • •Andy Alderwick
in reply to Devine Lu Linvega • • •It's tricky to know how much having a conditional JSR would save us, because the workarounds to not having it are so varied. But I shouldn't have gone for JCN as the opcode to eliminate to make room for it! 😅
Devine Lu Linvega
Unknown parent • • •Andy Alderwick
in reply to Andy Alderwick • • •If we break out the handling of ((opcode & 0x1f) == 0x00), we could do that again for ((opcode & 0x1f) == 0x02)
Andy Alderwick
in reply to Devine Lu Linvega • • •POP is pretty rare in your pie chart too, so even if POPs were made slower to implement the extra features, that isn't that bad either.
Devine Lu Linvega
in reply to Andy Alderwick • • •Andy Alderwick
in reply to Devine Lu Linvega • • •Devine Lu Linvega
in reply to Andy Alderwick • • •https://users.ece.cmu.edu/%7Ekoopman/stack_computers/sec4_4.html
Stack Computers: 4.4 ARCHITECTURE OF THE NOVIX NC4016
users.ece.cmu.eduAndy Alderwick
in reply to Devine Lu Linvega • • •🥺 🙏
Please don't be both… not again…
Devine Lu Linvega
in reply to Andy Alderwick • • •Devine Lu Linvega
Unknown parent • • •Kira, feral fox 🦊 🏳️⚧️
Unknown parent • • •XXIIVV — uxntal reference
wiki.xxiivv.comKira, feral fox 🦊 🏳️⚧️
Unknown parent • • •instead of
;function JSR2
which assembles to "LIT $ADDR JSR2"
function
assembles to "JSI $ADDR"
The instruction looks ahead of itself and jumps to that address while saving the proper return address on the r-stack.
Devine Lu Linvega
in reply to Kira, feral fox 🦊 🏳️⚧️ • • •Devine Lu Linvega
Unknown parent • • •Reid D. M.
in reply to Devine Lu Linvega • • •Devine Lu Linvega
Unknown parent • • •~Parnikkapore
Unknown parent • • •But yeah, I think it's a better idea to keep only one. Maybe having both would be more "pona", but idk
this feels a lot like the "guix environment" vs "guix shell" situation
DHeadshot's Alt
in reply to Kira, feral fox 🦊 🏳️⚧️ • • •Doesn't that make JSR2 a redundant legacy opcode there just for compatibility? Turning into an x86 situation...
@makeworld @neauoire