A graph replace that changes the type from dynamic to static can lead to miscompilation

[aseyboldt]

Description

If I replace a tensor with shape (None,) by one with static shape (1,) llvm aborts during codegen for me. Maybe some rewrite doesn't behave properly, and we produce invalid llvm code somewhere?

import pytensor.tensor as pt
import pytensor

x = pt.vector("x")
logp = -(x ** 2).sum()

pytensor.dprint(logp, print_type=True)

# Works if I set the shape to x.shape
#x_known = pt.vector("x_known", shape=x.type.shape)
x_known = pt.vector("x_known", shape=(1,))

logp = pytensor.graph_replace(logp, {x: x_known})

print("after")
pytensor.dprint(logp, print_type=True)
pytensor.function((x_known,), logp, mode=pytensor.compile.NUMBA)

Maybe we should just check that we don't change the type in a graph_replace?

[ricardoV94]

Here is a more direct example:

import pytensor.tensor as pt
import pytensor

x = pt.vector("x", shape=(1,))
pow_node = pytensor.graph.Apply(
    pt.pow,
    [x, pt.as_tensor(2)[None]],
    [pt.vector("x", shape=(None,))]
)
pow_out = pow_node.default_output()

func = pytensor.function((x,), pow_out, mode="NUMBA")
pytensor.dprint(func, print_destroy_map=True)
func([0.5])

Edit: Updated

[aseyboldt]

I also opened a numba issue: https://github.com/numba/numba/issues/9554