I am trying to implent a NIF for simple linear algebra. Here is my internal structure for matrix:
typedef struct la_matrix {
uint rows, columns;
double **data;
} la_matrix;
And here is a "constructor" for it:
la_result
la_matrix_constructor(la_matrix **res,
const uint rows,
const uint columns)
{
if (rows == 0 || columns == 0)
return dimensional_problems;
// allocate memory for meta-structure
*res = malloc(sizeof(la_matrix));
if (*res == NULL)
return null_ptr;
// allocater memory for array of pointers to rows
(*res)->data = malloc(rows * sizeof(double*));
if ((*res)->data == NULL) {
free(*res);
return null_ptr;
}
//allocate memory for each row
uint i = 0;
bool failed = false;
for (; i < rows; i++) {
(*res)->data[i] = malloc(columns * sizeof(double));
if ((*res)->data[i] == NULL) {
failed = true;
break;
}
}
if (failed) {
// one step back, since i-th row wasn't allocated
i -= 1;
for(; i < ~((uint) 0); i--)
free((*res)->data[i]);
free((*res)->data);
free(*res);
return null_ptr;
}
(*res)->rows = rows;
(*res)->columns = columns;
return ok;
}
Then I have two wrappers for NIF -- one for constructor:
static ERL_NIF_TERM
nif_matrix_constructor(ErlNifEnv *env,
int argc,
const ERL_NIF_TERM *argv)
{
uint rows, columns;
enif_get_uint(env, argv[0], &rows);
enif_get_uint(env, argv[1], &columns);
la_matrix **mat_res = enif_alloc_resource(LA_MATRIX_TYPE, sizeof(la_matrix *));
la_matrix *mat_ptr;
la_result result = la_matrix_constructor(&mat_ptr, rows, columns);
if (result != ok)
return enif_make_atom(env, "err");
memcpy((void *) mat_res, (void *) &mat_ptr, sizeof(la_matrix *));
ERL_NIF_TERM term = enif_make_resource(env, mat_res);
enif_release_resource(mat_res);
return term;
}
And one to test if the constructor works correctly:
static ERL_NIF_TERM
nif_matrix_rows(ErlNifEnv *env,
int argc,
const ERL_NIF_TERM *argv)
{
la_matrix *mat_ptr;
if(!enif_get_resource(env, argv[0], LA_MATRIX_TYPE, (void **) &mat_ptr))
return enif_make_atom(env, "err");
return enif_make_uint(env, mat_ptr->rows);
}
It seems that the constructor wrapper works perfectly fine (I've tested it with using printf
), but nif_matrix_rows
returns strange results, e.g.
iex(1)> mat = LinearAlgebra.matrix(2,3)
""
iex(2)> LinearAlgebra.rows(mat)
1677732752
And directly passing LinearAlgebra.matrix(2,3)
to LinearAlgebra.rows
twice results in segfault:
iex(3)> LinearAlgebra.rows(LinearAlgebra.matrix(2,3))
1543520864
iex(4)> LinearAlgebra.rows(LinearAlgebra.matrix(2,3))
zsh: (core dumped) iex -S mix
(Note different results for the "same" matrices).
I was following Andrea Leopardi's tutorial with minor (I don't really sure if they are so) changes to fight gcc
warnings. Most important, IMHO, was this part
la_matrix *mat_ptr;
if(!enif_get_resource(env, argv[0], LA_MATRIX_TYPE, (void **) &mat_ptr))
return enif_make_atom(env, "err");
while Andrea Leopardi uses
db_conn_t **conn_res;
enif_get_resource(env, argv[0], DB_RES_TYPE, (void *) conn_res);
db_conn_t *conn = *conn_res;
But it looks invalid for me since, AFAIR, (void *) conn_res
assumes that conn_res
was initalized.
Here is an error which occurs when I use Andrea's way:
src/nif.c: In function ‘nif_matrix_rows’:
src/nif.c:72:3: warning: ‘mat_res’ is used uninitialized in this function [-Wuninitialized]
enif_get_resource(env, argv[0], LA_MATRIX_TYPE, (void *) mat_res);
And calling LinearAlgebra.rows
from iex
causes segfault.
Could one tell me a proper way for handling structures in NIFs?
P.S. Sorry for C code, I've never wrote something more than bunch of helloworlds.
The problem was indeed in nif_matrix_rows
: with my code Elixir passes a pointer to a pointer to a structure (la_matrix **
) and I assumed that it would be a proper pointer.
So, quick fix is
static ERL_NIF_TERM
nif_matrix_rows(ErlNifEnv *env,
int argc,
const ERL_NIF_TERM *argv)
{
la_matrix const **mat_res;
if(!enif_get_resource(env, argv[0], LA_MATRIX_TYPE,(void **) &mat_res))
return enif_make_atom(env, "err");
la_matrix const *mat_ptr = *mat_res;
return enif_make_uint(env, mat_ptr->rows);
}
However, I'll wait some time for more elegant solution and won't accept this answer so far.