Segfault when assigning to a complex Scalar

[eriknw]

I get a segfault when I run this code:

import suitesparse_graphblas as ssgb

ssgb.initialize()
s = ssgb.ffi.new("GrB_Scalar*")
assert ssgb.lib.GrB_Scalar_new(s, ssgb.lib.GxB_FC64) == 0
assert ssgb.lib.GxB_Scalar_setElement_FC64(s[0], 1j) == 0  # <-- segfault here

Any thoughts? I guess we should try running this in C.

[eriknw]

Here's some spooky action: the following doesn't crash!

import suitesparse_graphblas as ssgb

ssgb.initialize()
for attr in dir(ssgb.lib):
    getattr(ssgb.lib, attr)
s = ssgb.ffi.new("GrB_Scalar*")
assert ssgb.lib.GrB_Scalar_new(s, ssgb.lib.GxB_FC64) == 0
assert ssgb.lib.GxB_Scalar_setElement_FC64(s[0], 1j) == 0

[DrTimothyAldenDavis]

I just tried a simple stand-alone C test and it worked fine. Give this a try:

setEl.zip

[DrTimothyAldenDavis]

What is the type of 1j? Is it complex, or a pure imaginary value (with no real part)? Does Python have such scalar types?

[DrTimothyAldenDavis]

Here is the output I get from "make" and everything is fine:

$ make
gcc -g -fopenmp -I../GraphBLAS/Include setEl.c -c -o setEl.o
gcc setEl.o -L../GraphBLAS/build -Wl,-rpath,../GraphBLAS/build -lgraphblas -lgomp -lpthread -lm -ldl -o setEl
ldd setEl
	linux-vdso.so.1 (0x00007fff68ff6000)
	libgraphblas.so.7 => /home/faculty/d/davis/master/GraphBLAS/build/libgraphblas.so.7 (0x00007f40a86e3000)
	libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f40a84d3000)
	libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f40a8384000)
	libgomp.so.1 => /lib/x86_64-linux-gnu/libgomp.so.1 (0x00007f40a8342000)
	libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f40a831f000)
	/lib64/ld-linux-x86-64.so.2 (0x00007f40b27c4000)
	libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f40a8319000)
./setEl

library: SuiteSparse:GraphBLAS 7.0.3 (Apr 8, 2022)
include: SuiteSparse:GraphBLAS 7.0.3 (Apr 8, 2022)

  1x1 GraphBLAS double complex scalar, sparse by col
  s, no entries, memory: 272 bytes


  1x1 GraphBLAS double complex scalar, full by col
  s, 1 entry, memory: 272 bytes
  iso value:    0 + 1i

    (0,0)    0 + 1i

[eriknw]

Thanks for giving this a try @DrTimothyAldenDavis! Yeah, Python has builtin complex numbers, and x = 1j has Re(x) == 0 and Im(x) == 1.

Upon further stress testing, adding

for attr in dir(ssgb.lib):
    getattr(ssgb.lib, attr)

reliably fixes the segfault for me.

How does the joke go?

Patient: doctor, it hurts with I do this

Doctor: so don't do that!

I have no idea why the above trick works (is cffi doing something weird under the hood?), but it does (apparently), so I guess we should use it. PR forthcoming!