Regression in 0.53.0rc1 due to issue in temp variable assignment removal

[stuartarchibald]

This fails type inference on 0.53.0rc1, worked fine on 0.52.0:

from numba import njit
import numpy as np

@njit
def foo():
    increment = 0.0
    foo_sum = 0.0
    i = 2

    while foo_sum < 100:
        a = [np.random.normal(1) for _ in range(i)]
        a = np.array(a)
        increment = np.sum(a)
        increment = np.abs(increment)
        foo_sum += increment
        i += 1

    return (foo_sum, i)

foo()

Debug log:

$ cat di58.py
from numba import njit
import numpy as np

@njit
def foo():
    increment = 0.0
    foo_sum = 0.0
    i = 2

    while foo_sum < 100:
        a = [np.random.normal(1) for _ in range(i)]
        a = np.array(a)
        increment = np.sum(a)
        increment = np.abs(increment)
        foo_sum += increment
        i += 1

    return (foo_sum, i)


$ HEAD=`git rev-parse HEAD`

$ git show-ref --tags -d|grep $HEAD
90e7cc74e1c3775727289162eade941c49a0db4f refs/tags/0.53.0rc1.post1^{}

$ python di58.py
Traceback (most recent call last):
  File "di58.py", line 20, in <module>
    foo()
  <snip>
  File "/numba/core/typeinfer.py", line 1174, in check_var
    raise TypingError(msg % (var, val, loc), loc)
numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
Type of variable 'a' cannot be determined, operation: call $42load_method.9(a, func=$42load_method.9, args=[Var(a, di58.py:11)], kws=(), vararg=None), location: di58.py (12)

File "di58.py", line 12:
def foo():
    <source elided>
        a = [np.random.normal(1) for _ in range(i)]
        a = np.array(a)
        ^

$ git revert -m 1 d938288
[release0.53 9ddebc9] Revert "Merge pull request #6575 from ehsantn/ehsan/avoid_tmp_vars"
 11 files changed, 46 insertions(+), 179 deletions(-)

$ python di58.py # works fine

d938288 i.e. merge of #6575 seems to be the cause.

CC @ehsantn

[sklam]

A smaller reproducer:

from numba import njit
import numpy as np

@njit
def foo():
    for i in range(3):
        a = (lambda j: j)(i)
        a = np.array(a)

    return

foo()

The temp variable assignment pass is causing the _inline_arraycall to misbehave. This can be checked by toggling

numba/numba/core/inline_closurecall.py

Line 44 in 199798e

enable_inline_arraycall = True

.

[ehsantn]

There is probably an assumption about extra assignments somewhere. I can look at this if no one has started yet.

[sklam]

@ehsantn, it's all yours =)