cyclic control flow for loops

[carljm]

Currently we don't even model control flow back edges in loops (because they will often lead to inference cycles). Once we have fixpoint iteration support in Salsa, we need to add these back edges and tests for them, including cases where we have to fallback to avoid runaway fixpoint iteration, e.g.:

x = 0
for _ in range(y):
    x += 1

[sharkdp]

I've seen multiple real world examples where missing support for loop control flow leads to an incorrect reachability analysis for large blocks of code within loops, because we observe a type that is too narrow. For example:

first_iteration = True

for x in xs:
    if not first_iteration:
        # this whole block is considered unreachable

    # more code

    first_iteration = False

[Gankra]

I just encountered the exact same problem as David:

https://play.ty.dev/9214d74c-f4de-46df-80d0-44b3262d1ff4

some_list: list[int] = [0, 1]

first_part = True
for j in range(0, 10):
    if first_part:
        first_part = False
        continue
    # ty infers `val: Never` here!?
    # (many other results now infer to Unknown)
    val = some_list[0]

And because we don't respect first_part: bool there's no (non-tedious) way to tell ty to chill.

[sharkdp]

And because we don't respect first_part: bool there's no (non-tedious) way to tell ty to chill.

Obviously not ideal, but a cast always helps as a workaround in these situations (first_part = cast(bool, True)).