nvals weirdness with 8.0.0 for max-sized Vector about graphblas HOT 6 CLOSED

I return UINT64_MAX if the product of nrows times ncols can cause int64_t overflow, but I might be overly cautious in this case because ncols is 1 for a GrB_Vector.

It would be a good idea to use 1152921504606846976L to make sure that value is a 64-bit constant.

from graphblas.

Gotcha. I actually cause this error from Python via python-graphblas (where types are handled correctly):

v = Vector(int, size=2**60)
v[:] = 1
v.nvals  # bad value!

This worked correctly with previous versions of SuiteSparse:GraphBLAS. v8.0.0 introduced a regression.

from graphblas.

Try the latest push to the master branch. I changed how I handle checking integer overflow in Source/GB_uint64_multiply.h in v8.0.0 and I think I introduced this bug then. The latest push to the master branch should fix this.

from graphblas.

That fixed it. I added a new test (GraphBLAS/Test/GB_mex_test25.c) that fails with v8.0.0 (triggering this error) but works fine for the draft v8.0.1.

from graphblas.

The fix isn't perfect because the integer overflow guard is not exact (it wasn't exact in v7.x either). For example, if you ask for GrB_Matrix_nvals for a matrix of size 2 by 2^30, it will report INT64_MAX, or 'infinity' because that case triggers the overflow guard. That's by design. It's too costly to put in a precise integer overflow guard.

from graphblas.

I confirm that this is now fixed in v8.0.1. Thanks @DrTimothyAldenDavis! Closing.

from graphblas.