LXM.Pure GC bug

[bagnalla]

Hi, I've been using the LXM.Pure interface and have encountered some strange/erroneous behavior (described in detail here).

Basically, the sequence of random values produced by the PRNG is sometimes incorrect, and this happens when the GC runs a minor collection. After some experimentation, I managed to fix the example linked above in my local environment by changing pringo_LXM_copy and pringo_LXM_init_unboxed from

CAMLprim value pringo_LXM_copy(value v)
{
  value res = caml_alloc_small(Wsizeof(struct LXM_state), Abstract_tag);
  memcpy(LXM_val(res), LXM_val(v), sizeof(struct LXM_state));
  return res;
}

CAMLprim value pringo_LXM_init_unboxed(uint64_t i1, uint64_t i2,
                                       uint64_t i3, uint64_t i4)
{
  value v = caml_alloc_small(Wsizeof(struct LXM_state), Abstract_tag);
  struct LXM_state * st = LXM_val(v);
  st->a = i1 | 1;    /* must be odd */
  st->x[0] = i2 != 0 ? i2 : 1; /* must be nonzero */
  st->x[1] = i3 != 0 ? i3 : 2; /* must be nonzero */
  st->s = i4;
  return v;
}

to

CAMLprim value pringo_LXM_copy(value v)
{
  CAMLparam1(v);
  value res = caml_alloc_small(Wsizeof(struct LXM_state), Abstract_tag);
  memcpy(LXM_val(res), LXM_val(v), sizeof(struct LXM_state));
  CAMLreturn(res);
}

CAMLprim value pringo_LXM_init_unboxed(uint64_t i1, uint64_t i2,
                                       uint64_t i3, uint64_t i4)
{
  CAMLparam0();
  CAMLlocal1(v);
  v = caml_alloc_small(Wsizeof(struct LXM_state), Abstract_tag);
  struct LXM_state * st = LXM_val(v);
  st->a = i1 | 1;    /* must be odd */
  st->x[0] = i2 != 0 ? i2 : 1; /* must be nonzero */
  st->x[1] = i3 != 0 ? i3 : 2; /* must be nonzero */
  st->s = i4;
  CAMLreturn(v);
}

I don't fully understand what the problem is, but it seems that maybe the original versions of these functions do not properly register the objects they create with the GC/runtime, so those objects end up getting garbled after the GC runs (perhaps being written over by subsequent allocations?).

I also noticed that the OCaml stdlib Random module uses the same LXM implementation, but it does all of its allocations on the OCaml side so I don't think it should suffer from this problem.

Side note: setting OCAMLRUNPARAM='v=0x02' to print minor collection messages seems to not work on OCaml 5.1.1... I don't know if I'm doing something wrong or if that's a bug that should be reported on the main OCaml repo.

-- Alex

[xavierleroy]

I agree this is a bug. I'll fix it in a few weeks, when I have more time for library maintenance. Thanks for your feedback.

[xavierleroy]

pringo_LXM_init_unboxed doesn't need a root registration, but pringo_LXM_copy definitely does !

Fixed as proposed by commit b94865d