TypingError when conditionally shadowing variable

[ricardoV94]

I'm not sure if this is expected/documented behavior, but I see a typing failure when conditionally shadowing an existing name:

import numba
import numpy as np

@numba.njit
def inc_array(x, val):
    # No shadowing
    if x.ndim == val.ndim:
        new_val = val[0]
    else:
        new_val = val
    x[0] += new_val

inc_array(np.zeros((5, 3)), np.array([[2.]]))  # Fine

@numba.njit
def inc_array(x, val):
    # Unconditional shadowing
    val = val[0]
    x[0] += val
    
inc_array(np.zeros((5, 3)), np.array([[2.]]))  # Fine 

@numba.njit
def inc_array(x, val):
    # Conditional shadowing
    if x.ndim == val.ndim:
        val = val[0]
    x[0] += val

inc_array(np.zeros((5, 3)), np.array([[2.]]))  # Fails: Cannot unify array(float64, 1d, C) and float64 for 'val.2'

[sklam]

That is interesting. I'll need to take a deeper look. So far, it seems to be a problem for running SSA after deadbranch pruning. The first version work because deadbranch pruning eliminated the else branch. But I think the third version reuse of val is somehow disabling the pruning.

[sklam]

Conclusion from triage meeting:

This happens because deadbranch pruning is running before SSA. To resolve this bug, we need to make deadbranch pruning SSA aware and then move SSA above it.