doc and impl of RewriteArrayExprs are inconsistent

def test_single_array_expr():
    def _test_impl(array):
        ret = array >= 0.1
        return ret

    jitted = numba.jit()(_test_impl)
    arr = np.random.rand(10)
    jit_ret = jitted(arr)

Code above will be rewritten by RewriteArrayExprs Pass. But array >= 0.1 only has one operator. Is it necessary to rewrite this expression, since loop fusion and shortcut deforestation optimizations don’t seem to work? And in numba developer doc, it wrote (https://numba.readthedocs.io/en/latest/developer/rewrites.html):

RewriteArrayExprs.match(): The rewrite pass looks for two or more array operations that form an array expression.

Is it inconsistent between doc and impl?

If this can be confirmed as a real inconsistency, I would like to contribute some code to fix it.

[sklam]

My first guess is that the doc is wrong and arrayexpr is needed to map array operations to ufunc. But the code was written years ago and I need to look at it to confirm.

[gmarkall]

IR is not in SSA form, but SSA IR is in SSA form. For example, if you have:

from numba import njit

@njit('int32(boolean)')
def f(c):
    if c:
        a = 1
    else:
        a = 2
    return a

Then the dumps are:

$ NUMBA_DUMP_IR=1 NUMBA_DUMP_SSA=1 python repro.py 
-----------------------------------IR DUMP: f-----------------------------------
label 0:
    c = arg(0, name=c)                       ['c']
    bool4 = global(bool: <class 'bool'>)     ['bool4']
    $4pred = call bool4(c, func=bool4, args=(Var(c, repro.py:6),), kws=(), vararg=None, target=None) ['$4pred', 'bool4', 'c']
    branch $4pred, 6, 14                     ['$4pred']
label 6:
    a = const(int, 1)                        ['a']
    $12return_value.2 = cast(value=a)        ['$12return_value.2', 'a']
    return $12return_value.2                 ['$12return_value.2']
label 14:
    a = const(int, 2)                        ['a']
    $20return_value.2 = cast(value=a)        ['$20return_value.2', 'a']
    return $20return_value.2                 ['$20return_value.2']

---------------------------------SSA IR DUMP: f---------------------------------
label 0:
    c = arg(0, name=c)                       ['c']
    bool4 = global(bool: <class 'bool'>)     ['bool4']
    $4pred = call bool4(c, func=bool4, args=(Var(c, repro.py:6),), kws=(), vararg=None, target=None) ['$4pred', 'bool4', 'c']
    branch $4pred, 6, 14                     ['$4pred']
label 6:
    a = const(int, 1)                        ['a']
    $12return_value.2 = cast(value=a)        ['$12return_value.2', 'a']
    return $12return_value.2                 ['$12return_value.2']
label 14:
    a.1 = const(int, 2)                      ['a.1']
    $20return_value.2 = cast(value=a.1)      ['$20return_value.2', 'a.1']
    return $20return_value.2                 ['$20return_value.2']

The IR is not SSA form because there are multiple assignments to a - in the SSA form, the second assignment is to a new variable, a.1. The introductory paragraphs of the Wikipedia article provide a reasonable summary: https://en.wikipedia.org/wiki/Static_single_assignment_form