Loop index type inferred as iList in squin kernel

[david-pl]

MWE:

@squin.kernel
def main():
    q = squin.qubit.new(3)
    x = squin.op.x()
    for i in range(3):
        squin.qubit.apply(x, [q[i]])

main.print()

The resulting IR is shows that the loop variable is an IList for some reason:

%x_1, %q_1, %x_2, %q_2 = scf.for %i : !py.IList[!py.int, Literal(3,int)] in %1 -> !py.Op, !py.IList[!py.Qubit, !Any], !py.Op, !py.IList[!py.Qubit, !Any]

Tested on bloqade-circuit main branch with kirin v0.17.6.

[weinbe58]

@Roger-luo This might be an issue in Kirin?

[david-pl]

Just tried on the latest kirin main. Here's a kirin only MWE:

def test_loop_index():
    @python_basic.union(
        [func, scf, py.unpack, lowering.func, ilist, lowering.range.ilist]
    )
    def main():
        ls = [1,2,3]
        for i in range(3):
            ls[i]

    main.print()

You also get

%ls_1, %ls_2 = scf.for %i in %6 -> !py.IList[!py.int, Literal(3,int)], !py.IList[!py.int, Literal(3,int)]

[david-pl]

It seems this may be already fixed on the latest version of kirin.

With this example:

from kirin.prelude import structural_no_opt
from kirin.passes import TypeInfer


@structural_no_opt
def main():
    x = 0
    for i in range(2):
        x = i

    return x

TypeInfer(main.dialects)(main)

main.print()

on main I get

func.func @main() -> Literal(1,int) {
  ^0(%main_self):
  │   %x = py.constant.constant 0 : !py.int
  │   %0 = py.constant.constant 0 : !py.int
  │   %1 = py.constant.constant 2 : !py.int
  │   %2 = py.constant.constant 1 : !py.int
  │   %3 = py.ilist.range(start=%0, stop=%1, step=%2) : !py.IList[!py.int, Literal(2,int)]
  │ %x_1 = scf.for %i : !py.int in %3 -> !py.int
  │        │ iter_args(%x_2 : !py.int = %x) {
  │        │ %x_3 = py.assign.alias x = %i : !py.int
  │        │        scf.yield %x_3
  │        } -> purity=True
  │        func.return %x_1
} // func.func main

vs on v0.17.10

func.func main() -> Literal(1,Union(types=frozenset({int, IList[int, Literal(data=2, type=int)]}))) {
  ^0(%main_self):
  │   %x = py.constant.constant 0 : !py.int
  │   %0 = py.constant.constant 0 : !py.int
  │   %1 = py.constant.constant 2 : !py.int
  │   %2 = py.constant.constant 1 : !py.int
  │   %3 = py.ilist.range(start=%0, stop=%1, step=%2) : !py.IList[!py.int, Literal(2,int)]
  │ %x_1 = scf.for %i : !py.IList[!py.int, Literal(2,int)] in %3 -> !Union[!py.int, !py.IList[!py.int, Literal(2,int)]]
  │        │ iter_args(%x_2 : !py.int = %x) {
  │        │ %x_3 = py.assign.alias x = %i : !py.IList[!py.int, Literal(2,int)]
  │        │        scf.yield %x_3
  │        } -> purity=True
  │        func.return %x_1
} // func.func main

[david-pl]

Also, FYI, the fact that we get an ilist as index messes with the qasm2 IndexingDesugarpass.

@qasm2.extended
def main():
    q = qasm2.qreg(2)

    for i in range(2):
        qasm2.x(q[i])


main.print()

You get

  │ %q_1, %q_2 = scf.for %i : !Bottom in %1 -> !py.IList[!py.Qubit, !Any], !py.IList[!py.Qubit, !Any]

I.e. the loop variable is inferred as Bottom.

The problem is that when rewriting the statements, we create a new statement (QRegGet | CRegGet) (right here), which expects an integer as index. Since the type is wrong, we get Bottom in the end.

So I'd expect this to also be fixed once we update bloqade-circuit to the new kirin version.