saladdais / lummao Goto Github PK

Not many sane scripts actually do this, but required for things like

jump foo;
integer bar = 1;
@foo;
llOwnerSay((string)bar);  // should print "0"

to work "correctly". Heap types should be declared with Optional[...] wrapper around their type and a default value of None, and hoisted declarations for quaternions should be <0, 0, 0, 0> rather than ZERO_ROTATION.

Probably further down the road, since I'd have to figure out how to rewrite the desugaring logic for PyOptimizer as well, but it'd be nice to get rid of the native code if we don't really need it.

From https://docs.python.org/3.11/whatsnew/3.11.html#cpython-bytecode-changes:

Replaced JUMP_IF_NOT_EXC_MATCH by CHECK_EXC_MATCH which performs the check but does not jump.

Replaced JUMP_IF_NOT_EG_MATCH by CHECK_EG_MATCH which performs the check but does not jump.

Replaced JUMP_ABSOLUTE by the relative JUMP_BACKWARD.

Added JUMP_BACKWARD_NO_INTERRUPT, which is used in certain loops where it is undesirable to handle interrupts.

Replaced POP_JUMP_IF_TRUE and POP_JUMP_IF_FALSE by the relative POP_JUMP_FORWARD_IF_TRUE, POP_JUMP_BACKWARD_IF_TRUE, POP_JUMP_FORWARD_IF_FALSE and POP_JUMP_BACKWARD_IF_FALSE.

Added POP_JUMP_FORWARD_IF_NOT_NONE, POP_JUMP_BACKWARD_IF_NOT_NONE, POP_JUMP_FORWARD_IF_NONE and POP_JUMP_BACKWARD_IF_NONE opcodes to speed up conditional jumps.

Looks like the most important change is that absolute jumps no longer exist, and we have to use separate forward / backward relative jump types now. Will need separate codegen paths for 3.11+.

We use LSL-PyOptimizer's typecast function for runtime casts, but it's failing on a couple of cases from the second LSL language conformance test:

>>> lslopt.lslfuncs.typecast(lslopt.lslfuncs.F32(1.4e-45), str)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/user/source/LSL-PyOptimizer/lslopt/lslbasefuncs.py", line 649, in typecast
    return InternalTypecast(val, out, InList, f32)
  File "/home/user/source/LSL-PyOptimizer/lslopt/lslbasefuncs.py", line 470, in InternalTypecast
    if out == unicode: return f2s(val, 6)
  File "/home/user/source/LSL-PyOptimizer/lslopt/lslbasefuncs.py", line 391, in f2s
    while s[i] in u'0.':
IndexError: string index out of range

because it gets formatted as 0.00000 internally, but its float val is not equal to 0.0, making that loop explode.

also:

>>> lslopt.lslfuncs.typecast("-0.0", float)
0.0

when it should be -0.0 seems to be due to a missing math.copysign() call when replacing things equivalent to 0.0 with 0.0 itself.

LSL-PyOptimizer itself doesn't run into either of these issues when optimizing an LSL script. Need to dig in more, things might be special-cased further up in the folding code.

Would be great to be able to do script_cls = exec(lummao.lsl_to_python(lsl_text)).Script or something like that to dynamically convert LSL to Python at runtime. Then we could just build binary wheels that include everything people need instead of needing to do compile a bunch of stuff and run CLI commands separately.

Depends on SaladDais/tailslide#24.

Lummao is annoying to set up right now, and an action would be helpful since some people are going to want to test their LSL scripts in CI.

Depends on #1 . It was mentioned that the cibuildwheel Github action may be helpful for doing the cross-compiling as well.

We need to do symbol name mangling to enforce LSL's shadowing rules, as well as prevent people from shadowing Python or Lummao builtins from their scripts.