Changeset [53df545c3266c3599a9a5bcb12437906162b6722] by Jed Brown

Check alignment in TensorMult_Hex_P4_Q4_D1

16-byte alignment is required for the MOVAPD instructions used
extensively in this function. Prior to a fix in PetscTrMalloc, this
function was crashing with optimized PETSc and -malloc_dump. The reason
for this: PetscTrMallocDefault() was only guaranteeing 8-byte alignment
(rather than respecting PETSC_MEMALIGN). In debug mode, these happened to be equivalent because struct _trSPACE contained a PetscStack member.
In optimized mode, the stack is not stored so we were only getting
8-byte alignment, causing MOVAPD to generate a GP fault.
http://github.com/jedbrown/dohp/commit/53df545c3266c3599a9a5bcb1243...