Segfault in np.power involving integers and fastmath

[aseyboldt]

The following code segfaults on my machine:

import numba
import numpy as np

@numba.jit("float64(float64, int8)", fastmath=True)
def foo(x, y):
    return np.power(x, y)

foo(np.float64(4.), np.int8(2))

It generates the following llvm code:

define i32 @_ZN8__main__3fooB2v3B60c8tJTIeFIjxB2IKSgI4CrvQClYb5wBbdC9XqICn1Wk1gKBZBVGsCAA_3d_3dEda(double* noalias nocapture writeonly %retptr, { i8*, i32, i8*, i8*, i32 }** noalias nocapture readnone %excinfo, double %arg.x, i8 %arg.y) local_unnamed_addr #0 {
for.end:
  %0 = sext i8 %arg.y to i32
  %1 = tail call fast double @llvm.powi.f64.i32(double %arg.x, i32 %0)
  store double %1, double* %retptr, align 8
  ret i32 0
}

Without the fastmath it uses @llvm.pow.f64 with no problems.

The generated assembly calls __powidf2

	pushq	%rbx
	movq	%rdi, %rbx
	movsbl	%dl, %edi
	movabsq	$__powidf2, %rax
	callq	*%rax
	vmovsd	%xmm0, (%rbx)
	xorl	%eax, %eax
	popq	%rbx
	retq

Maybe this is just a variant of this issue, but for some reason I see a segfault instead?

A user also reported seeing the missing symbol error in a related case.

I'm on arch linux, and I see the same problem both with numba from the numba conda channel and from numba installed systemwide from the arch package.

• I have tried using the latest released version of Numba (most recent is
  visible in the release notes
  (https://numba.readthedocs.io/en/stable/release-notes-overview.html).
• I have included a self contained code sample to reproduce the problem.
  i.e. it's possible to run as 'python bug.py'.

[kc611]

Hi @aseyboldt, Thanks for reporting the issue. I can confirm the same locally on a Ubuntu Linux too. Not sure right now from where the error is originating, marking this as needs triage.

Could you please provide some additional information like the CPU info,OS version, libm version, whether SVML was installed or not along with the code/logical path that produces the said error.