mirror of
https://github.com/catseye/SixtyPical.git
synced 2024-11-22 17:32:01 +00:00
101 lines
4.0 KiB
Markdown
101 lines
4.0 KiB
Markdown
SixtyPical
|
|
==========
|
|
|
|
_Version 0.13. Work-in-progress, everything is subject to change._
|
|
|
|
SixtyPical is a very low-level programming language, similar to 6502 assembly,
|
|
with static analysis through abstract interpretation.
|
|
|
|
In practice, this means it catches things like
|
|
|
|
* you forgot to clear carry before adding something to the accumulator
|
|
* a subroutine that you call trashes a register you thought was preserved
|
|
* you tried to write the address of something that was not a routine, to
|
|
a jump vector
|
|
|
|
and suchlike. It also provides some convenient operations and abstractions
|
|
based on common machine-language programming idioms, such as
|
|
|
|
* copying values from one register to another (via a third register when
|
|
there are no underlying instructions that directly support it)
|
|
* explicit tail calls
|
|
* indirect subroutine calls
|
|
|
|
The reference implementation can analyze and compile SixtyPical programs to
|
|
6502 machine code.
|
|
|
|
Documentation
|
|
-------------
|
|
|
|
* [Design Goals](doc/Design%20Goals.md)
|
|
* [SixtyPical specification](doc/SixtyPical.md)
|
|
* [SixtyPical revision history](HISTORY.md)
|
|
* [Literate test suite for SixtyPical syntax](tests/SixtyPical%20Syntax.md)
|
|
* [Literate test suite for SixtyPical analysis](tests/SixtyPical%20Analysis.md)
|
|
* [Literate test suite for SixtyPical compilation](tests/SixtyPical%20Compilation.md)
|
|
* [6502 Opcodes used/not used in SixtyPical](doc/6502%20Opcodes.md)
|
|
|
|
TODO
|
|
----
|
|
|
|
### `for`-like loop
|
|
|
|
We have range-checking in the abstract analysis now, but we lack practical ways
|
|
to use it.
|
|
|
|
We can `and` a value to ensure it is within a certain range. However, in the 6502
|
|
ISA the only register you can `and` is `A`, while loops are done with `X` or `Y`.
|
|
Insisting this as the way to do it would result in a lot of `TXA`s and `TAX`s.
|
|
|
|
What would be better is a dedicated `for` loop, like
|
|
|
|
for x in 0 to 15 {
|
|
// in here, we know the range of x is exactly 0-15 inclusive
|
|
// also in here: we are disallowed from changing x
|
|
}
|
|
|
|
However, this is slightly restrictive, and hides a lot.
|
|
|
|
However however, options which do not hide a lot, require a lot of looking at
|
|
(to ensure: did you increment the loop variable? only once? etc.)
|
|
|
|
The leading compromise so far is an "open-faced for loop", like
|
|
|
|
ld x, 15
|
|
for x downto 0 {
|
|
// same as above
|
|
}
|
|
|
|
This makes it a little more explicit, at least, even though the loop
|
|
decrementation is still hidden.
|
|
|
|
### Save registers on stack
|
|
|
|
This preserves them, so that, semantically, they can be used later even though they
|
|
are trashed inside the block.
|
|
|
|
### Re-order routines and optimize tail-calls to fallthroughs
|
|
|
|
Not because it saves 3 bytes, but because it's a neat trick. Doing it optimally
|
|
is probably NP-complete. But doing it adequately is probably not that hard.
|
|
|
|
### Different preludes for different architectures
|
|
|
|
`--prelude=c64-basic`
|
|
|
|
### And at some point...
|
|
|
|
* `low` and `high` address operators - to turn `word` type into `byte`.
|
|
* `const`s that can be used in defining the size of tables, etc.
|
|
* Tests, and implementation, ensuring a routine can be assigned to a vector of "wider" type
|
|
* Related: can we simply view a (small) part of a buffer as a byte table? If not, why not?
|
|
* Check that the buffer being read or written to through pointer, appears in approporiate inputs or outputs set.
|
|
(Associate each pointer with the buffer it points into.)
|
|
* `static` pointers -- currently not possible because pointers must be zero-page, thus `@`, thus uninitialized.
|
|
* Question the value of the "consistent initialization" principle for `if` statement analysis.
|
|
* `interrupt` routines -- to indicate that "the supervisor" has stored values on the stack, so we can trash them.
|
|
* Add absolute addressing in shl/shr, absolute-indexed for add, sub, etc.
|
|
* Automatic tail-call optimization (could be tricky, w/constraints?)
|
|
* Possibly `ld x, [ptr] + y`, possibly `st x, [ptr] + y`.
|
|
* Maybe even `copy [ptra] + y, [ptrb] + y`, which can be compiled to indirect LDA then indirect STA!
|