gh-125063: Move slice constant-folding to AST

[encukou]

Generate slice constants in AST, rather than in codegen.

This could allow more optimizations in earlier steps, albeit in practice it probably only helps rare cases (e.g. "abc"[:1] -> "a').

Tests for the final result were added in GH-125064.

• Issue: Emit slices as constants in the bytecode compiler #125063

[iritkatriel]

Strictly speaking this test doesn't belong in test_peepholer now that the optimization is not part of the peephole optimiser.

Where are the other const folding tests?

[Eclips4]

You're right, Irit.

This test should be re-written and placed in the test_ast/test_ast.py::ASTOptimiziationTests

[markshannon]

I want to suggest the opposite: remove the AST optimizer altogether.

The flowgraph optimizer has more information and can do a better job.
In fact the AST optimizer prevents the flowgraph optimizer producing the best code by mis-optimizing in some cases.
For example, by doing premature constant folding, This:

   a, b = 1,2

is compiled to

  LOAD_CONST               0 ((1, 2))
  UNPACK_SEQUENCE          2
  STORE_FAST_STORE_FAST    1 (a, b)

but it should be (and would be but for the AST optimizer) compiled to this:

  LOAD_SMALL_INT           1
  LOAD_SMALL_INT           2
  STORE_FAST_STORE_FAST    1 (a, b)

There are a few changes that the AST optimizer does that the flowgraph optimizer doesn't, and they could all be easily add
ed to the flowgraph optimizer.

The AST optimizer converts UNARY_OP("-", CONST(1)) to CONST(-1) which needs to be done before code gen, but other than that everything else should be done in the flowgraph optimizer.

[markshannon]

One more example that is probably more relevant to this PR.

  x[:1]

is currently compiled as:

  LOAD_FAST                0 (x)
  LOAD_CONST               0 (slice(None, 1, None))
  BINARY_SUBSCR

we might want to compile it as:

  LOAD_FAST                0 (x)
  LOAD_CONST               0 (None)
  LOAD_SMALL_INT           1
  BINARY_SLICE

This PR would prevent that.

[Eclips4]

I concur with Mark. I would like to see the elimination of the ast optimizer, and I would be happy to do it.