read_mtx() invalid read due to loop condition

[brian-kelley]

read_mtx() in KokkosKernels_IOUtils.hpp:998 loads a list of (row,col) pairs, sorts them, and then inserts into the CRS structure one row at a time. The loop over entries in the current row only checked if the next entry matches the current row, but NOT that the end of the pair array
has been reached. This caused an invalid read just after inserting the last entry in the matrix.

To fix this, I added another loop condition that makes sure eind < nE before reading edges[eind].

Valgrind output demonstrating the bug:

==59918== Invalid read of size 4
==59918==    at 0x43E819: int KokkosKernels::Impl::read_mtx<int, unsigned long, double>(char const*, int*, unsigned long*, unsigned long**, int**, double**, bool, bool, bool) (KokkosKernels_IOUtils.hpp:998)
==59918==    by 0x43F2FD: void KokkosKernels::Impl::read_matrix<int, unsigned long, double>(int*, unsigned long*, unsigned long**, int**, double**, char const*) (KokkosKernels_IOUtils.hpp:1019)
==59918==    by 0x43F480: KokkosSparse::CrsMatrix<double, int, Kokkos::Cuda, void, unsigned long> KokkosKernels::Impl::read_kokkos_crst_matrix<KokkosSparse::CrsMatrix<double, int, Kokkos::Cuda, void, unsigned long> >(char const*) (KokkosKernels_IOUtils.hpp:1055)
==59918==    by 0x405810: main (spgemm.cpp:55)
==59918==  Address 0xe169e50 is 0 bytes after a block of size 178,496 alloc'd
==59918==    at 0x4C2A1E3: operator new(unsigned long) (vg_replace_malloc.c:334)
==59918==    by 0x43DF9B: allocate (new_allocator.h:111)
==59918==    by 0x43DF9B: allocate (alloc_traits.h:436)
==59918==    by 0x43DF9B: _M_allocate (stl_vector.h:172)
==59918==    by 0x43DF9B: _M_create_storage (stl_vector.h:187)
==59918==    by 0x43DF9B: _Vector_base (stl_vector.h:138)
==59918==    by 0x43DF9B: vector (stl_vector.h:283)
==59918==    by 0x43DF9B: int KokkosKernels::Impl::read_mtx<int, unsigned long, double>(char const*, int*, unsigned long*, unsigned long**, int**, double**, bool, bool, bool) (KokkosKernels_IOUtils.hpp:933)
==59918==    by 0x43F2FD: void KokkosKernels::Impl::read_matrix<int, unsigned long, double>(int*, unsigned long*, unsigned long**, int**, double**, char const*) (KokkosKernels_IOUtils.hpp:1019)
==59918==    by 0x43F480: KokkosSparse::CrsMatrix<double, int, Kokkos::Cuda, void, unsigned long> KokkosKernels::Impl::read_kokkos_crst_matrix<KokkosSparse::CrsMatrix<double, int, Kokkos::Cuda, void, unsigned long> >(char const*) (KokkosKernels_IOUtils.hpp:1055)
==59918==    by 0x405810: main (spgemm.cpp:55)

[brian-kelley]

Fix ready in #452, running spot check now.

[ndellingwood]

@brian-kelley just for reference, was this related to reading one/some of the matrices as reported in #402 during issues with spgemm?

[brian-kelley]

@ndellingwood I was running your spgemm driver with those matrices when I discovered this, but it would have happened with any matrix.