| |
oziphantom
Registered: Oct 2014
Posts: 490 |
Packer vs Code VM vs Code Generator
Has anybody experimented with/know of code VMs or Generators.
Just thinking to get code size down they might perform better than exomiser et al.
I know of Sweet16 by Woz any others? |
|
... 80 posts hidden. Click here to view all posts....
|
| |
TNT
Account closed
Registered: Oct 2004
Posts: 189 |
@Krill: did you try separate streams for immediate/low byte operands and high bytes? I found out that high bytes compress better when separated from the low byte.
@soci: no need to add special cases for RTI/RTS - just make preprocessor warn if the following byte (fake operand) is non-zero. Additionally the preprocessor can replace the zero with any other value if that improves compression. JSR needs to be handled separately of course. |
| |
Krill
Registered: Apr 2002
Posts: 2980 |
TNT: I don't remember trying that. Also the RTS/RTI padding is a good idea. Certainly something to play around with for the next 4K of mine! :D |
| |
Krill
Registered: Apr 2002
Posts: 2980 |
Quoting Oswaldyou could just set some flag to tell the irq to stop scheduling the loader in, insead of that supper scifi method? :) Why is the obvious solution of simply splitting the stack (TSX/TXS exists) the "super sci-fi" method and painstakingly copying it back and forth not? :) |
| |
JackAsser
Registered: Jun 2002
Posts: 2014 |
Quote: Quoting Oswaldyou could just set some flag to tell the irq to stop scheduling the loader in, insead of that supper scifi method? :) Why is the obvious solution of simply splitting the stack (TSX/TXS exists) the "super sci-fi" method and painstakingly copying it back and forth not? :)
He was referring to implementing yield vs pause/resume thread |
| |
Krill
Registered: Apr 2002
Posts: 2980 |
Ah, thanks. I should not morning-post before the third coffee. :D |
| |
lft
Registered: Jul 2007
Posts: 369 |
Quoting sociFrom the processor's viewpoint the least significant 5 bits of opcode fully determines the length of the addressing mode.
Still it seems that BRK/JSR/RTI/RTS are exceptions. But that's only because the modification of PC is covering up that they're actually 2 bytes long.
I think if you reduce the table size the result will be shorter even with the special handling added in for these opcodes.
Also remember that this doesn't need to be 100% accurate. It's just a heuristic, and things will work as long as the encoder and the decoder make exactly the same errors. So treating jsr as having a one-byte operand (or brk/rti/rts/ldy#/cpy#/cpx# as having two-byte operands) will throw things out of sync for a while, but the compression ratio might not increase very much, and may even accidentally decrease for some input data. Meanwhile the decoder would be greatly simplified. |
| |
Krill
Registered: Apr 2002
Posts: 2980 |
Yes, but i'd try to get things back in sync again as quickly as possible. In an environment where every byte counts, "tainting" the streams by switching their data may (or may not) result in some losses.
And the question would be how often do the streams get in and out of alignment? Surely you want to keep that number low, otherwise the entire exercise is moot.
Now, i really like the idea of padding in zeros for simplified decoding, which may avoid unaligned streams altogether and have no drawbacks in comparison.
But i guess the goal would be to make the op-code lengths as regular as possible, possibly even getting rid of the table altogether in favour of a simple expression, while keeping the streams as well-aligned as possible by padding and/or allowing transient alignment errors only on rare operations (prefer brk/rti/rts/jsr over lda#/cmp#/etc.) with a quick recovery (also via some sort of padding?). |
| |
Krill
Registered: Apr 2002
Posts: 2980 |
Oh, and of course, most illegals and also quite a few regular operations aren't used much (or can be replaced), especially in size-constrained productions. This may also have a significant influence on the "regularity" of the instruction lengths. |
| |
TBH
Registered: Mar 2010
Posts: 21 |
I've played around with what might be considered code VM systems for the C64, with varying results:
1) Event Management System
The use of Events allows code to be deferred ("fire and forget") within an imperative program design. It's basically a set of tables, pointers and indices tied together with a few routines.
The basic features of such a system are:
* Subscribe To An Event - Add a subroutine address to an Event's list of Handlers.
* Unsubscribe From An Event - Remove a subscription.
* Raise An Event - Notify the Event system that an Event of a certain type occurred.
* Event Manager - (called from main loop) Executes the subscribed Handlers for each raised event in its buffer. Events can be queued and prioritised, eg animation Events handled once every 8 frames.
2) Reactive (Event-Driven) Extended Finite State machine
A state machine compatible with OMG's UML2 specification. Relatively slow, but can handle complex sets of transitions. However, it's probably overkill for most purposes.
3) Decision Table
Similar to a lookup table, but with more features and a routine that (optionally) compiles the data and executes the logic expressed by the data and structure.
A table consists of a set of Rules. Each Rule has a set of Actions that are executed if a set of Conditions are met.
The benefits:
a) Rules are expressed explicity when tabularised, instead of (often) implicitly when in code.
b) No IF-THEN code is required.
Many different tables can exist, but only routine used to execute them. This minimises code and the payoff can be big in a program that needs to evaluate a lot of conditions. Alternatively, a code generator can generate a code representation of the table.
4) Inline Function Parameters
This is simple but useful in that it remains slightly more readable than using a data table outside the code. Implementation can vary, but involve manipulating stack return values. The result is code like this:
; PrintAt routine params: Column, Row, Colour, @-terminated string
JSR PrintAt
.byt 10, 2, LtBlue
.asc "The quick brown fox@"
JSR PrintAt
.byt 10, 3, White
.asc "jumps over the lazy dog.@" |
| |
lft
Registered: Jul 2007
Posts: 369 |
I've implemented the Z-machine for C64+REU. Zeugma 1.2 |
| Previous - 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 - Next |
