# Type inference for/ dispatching of function arguments in njit with multiple signatures

**URL:** <https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536>\
**Category:** Support: How do I do ...?\
**Created:** [August 30, 2022, 8:49am UTC](https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536 "2022-08-30T08:49:16Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![drchaieb](https://yyz2.discourse-cdn.com/free1/user_avatar/numba.discourse.group/drchaieb/32/690_2.png) [@drchaieb](https://numba.discourse.group/u/drchaieb)\
**Post date:** [August 30, 2022, 8:49am UTC](https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536/1 "2022-08-30T08:49:16Z")

</div>

I am encountering a (type inference?) problem that can be reproduced by the minimal example below. The problem arises in my setting when passing njit compiled functions as arguments. All calls occur in a controlled manner and without the signatures I had very high JIT-compilation times.

```auto
`from numba import njit, float64, bool_
from numba.core.types import FunctionType, UniTuple, Tuple

@njit([float64(float64)])
def f(x):
    return x

@njit([float64(float64, bool_)])
def g(x, b):
    return x if b else 0.0

@njit([float64(FunctionType(float64(float64)), float64, Tuple(())),
       float64(FunctionType(float64(float64, bool_)), bool_, UniTuple(bool_, 1))])
def h(fn, x, *args):
    return fn(x, *args)

h(g, 1.0, *(True,)) # Will fail
h(f, 1.0, *()) # will fail`

```

I followed the calls into numba and the problem seems to happen in numba.core.types.functions.Dispatcher.can\_convert\_to which at some point tries to compile code for the wrong signature (since type inference did not exclude it) – or in this case a function no longer allowing compilation.

From what I see all types and calls are correct, and dispatching would work if the can\_convert functions just returned False instead of raising RuntimeException.

The problem does not occur when compilation is allowed (which is the problem I am trying to solve) but at the cost of high compilation times and the added signatures in h for very specific functions (i.e. CPUDispatcher bound to specific function, and not as I want it to be bound to the function type, not the pointer)

Am I using the right signatures? Am I using the signatures feature in a wrong manner? Why is the type of an njitted function type(Dispatcher f at mem\_addr) and not the FunctionType by the way?

Any hint would be greatly appreciated!

Amine.

---

<div class="post-metadata">

**Author:** ![gmarkall](https://yyz2.discourse-cdn.com/free1/user_avatar/numba.discourse.group/gmarkall/32/42_2.png) [@gmarkall](https://numba.discourse.group/u/gmarkall)\
**Post date:** [August 30, 2022, 10:58am UTC](https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536/2 "2022-08-30T10:58:58Z")

</div>

> From what I see all types and calls are correct, and dispatching would work if the can\_convert functions just returned False instead of raising RuntimeException.

I think you’re right, the `can_convert_to_` function assumes it’s possible to compile a function when that’s not always the case. I’ll open an issue.

---

<div class="post-metadata">

**Author:** ![gmarkall](https://yyz2.discourse-cdn.com/free1/user_avatar/numba.discourse.group/gmarkall/32/42_2.png) [@gmarkall](https://numba.discourse.group/u/gmarkall)\
**Post date:** [August 30, 2022, 11:06am UTC](https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536/3 "2022-08-30T11:06:20Z")

</div>

Issue: [Checks for function type casting erroneously assume compilation is always possible · Issue #8405 · numba/numba · GitHub](https://github.com/numba/numba/issues/8405)

---

<div class="post-metadata">

**Author:** ![drchaieb](https://yyz2.discourse-cdn.com/free1/user_avatar/numba.discourse.group/drchaieb/32/690_2.png) [@drchaieb](https://numba.discourse.group/u/drchaieb)\
**Post date:** [August 31, 2022, 12:45pm UTC](https://numba.discourse.group/t/type-inference-for-dispatching-of-function-arguments-in-njit-with-multiple-signatures/1536/4 "2022-08-31T12:45:08Z")

</div>

Thank you for creating the issue, which will solve the problem.