git @ Cat's Eye Technologies SixtyPical / 0.20 doc / Design Goals.md
0.20

Tree @0.20 (Download .tar.gz)

Design Goals.md @0.20view markup · raw · history · blame

Design Goals for SixtyPical

(draft)

The intent of SixtyPical is to have a very low-level language that benefits from abstract interpretation.

"Very low-level" means, on a comparable level of abstraction as assembly language.

In the original vision for SixtyPical, SixtyPical instructions mapped nearly 1:1 to 6502 instructions. However, many times when programming in 6502 you're using idioms (e.g. adding a 16-bit constant to a 16-bit value stored in 2 bytes) and it's just massively easier to analyze such actions when they are represented by a single instruction.

So SixtyPical instructions are similar to, inspired by, and have analogous restrictions as 6502 instructions, but in many ways, they are more abstract. For example, copy.

The intent is that programming in SixtyPical is a lot like programming in 6052 assembler, but it's harder to make a stupid error that you have to spend a lot of time debugging.

The intent is not to make it absolutely impossible to make such errors, just harder.

Things it will Not Do

To emphasize the point, the intent is not to make it impossible to make data-usage (and other) errors, just harder.

Here are some things SixtyPical will not try to detect or prevent you from doing:

  • Check that a vector is initialized before it's called.
  • Check that the stack has enough room on it.
  • Prevent bad things happening (e.g. clobbering a static storage location) because of a recursive call. (You can always recursively call yourself through a vector.)
  • Check that reads and writes to a buffer are in bounds. (This may happen someday, but it's difficult. It's more likely that this will happen for tables, than for buffers.)

At one point I wanted to do a call-tree analysis to find sets of routines that would never be called together (i.e. would never be on the call stack at the same time) and allow any static storage locations defined within them to occupy the same addresses, i.e. allow storage to be re-used across these routines. But, in the presence of vectors, this becomes difficult (see "Prevent bad things happening", above.) Also, it would usually only save a few bytes of storage space.

Some Background

The ideas in SixtyPical came from a couple of places.

One major impetus was when I was working on [Shelta][], trying to cram all that code for that compiler into 512 bytes. This involved looking at the x86 registers and thinking hard about which ones were preserved when (and which ones weren't) and making the best use of that. And while doing that, one thing that came to mind was: I Bet The Assembler Could Track This.

Another influence was around 2007 when "Typed Assembly Language" (and "Proof Carrying Code") were all the rage. I haven't heard about them in a while, so I guess they turned out to be research fads? But for a while there, it was all Necula, Necula, Necula. Anyway, I remember at the time looking into TAL and expecting to find something that matched the impression I had pre-formulated about what a "Typed Assembly" might be like. And finding that it didn't match my vision very well.

I don't actually remember what TAL seemed like to me at the time, but what I had in mind was more like SixtyPical.

(I'll also write something about abstract interpretation here at some point, hopefully.)