Fix the same issue as #646 (MR \!1000) but now in the allocation of extra gpart

src/space_extras.c

@@ -158,7 +158,6 @@ void space_allocate_extras(struct space *s, int verbose) {
       if (swift_memalign("gparts", (void **)&gparts_new, gpart_align,
                          sizeof(struct gpart) * size_gparts) != 0)
         error("Failed to allocate new gpart data");
-      const ptrdiff_t delta = gparts_new - s->gparts;
       memcpy(gparts_new, s->gparts, sizeof(struct gpart) * s->size_gparts);
       swift_free("gparts", s->gparts);
       s->gparts = gparts_new;
@@ -166,23 +165,10 @@ void space_allocate_extras(struct space *s, int verbose) {
       /* Update the counter */
       s->size_gparts = size_gparts;
 
-      /* We now need to reset all the part and spart pointers */
-      for (size_t i = 0; i < nr_parts; ++i) {
-        if (s->parts[i].time_bin != time_bin_not_created)
-          s->parts[i].gpart += delta;
-      }
-      for (size_t i = 0; i < nr_sinks; ++i) {
-        if (s->sinks[i].time_bin != time_bin_not_created)
-          s->sinks[i].gpart += delta;
-      }
-      for (size_t i = 0; i < nr_sparts; ++i) {
-        if (s->sparts[i].time_bin != time_bin_not_created)
-          s->sparts[i].gpart += delta;
-      }
-      for (size_t i = 0; i < nr_bparts; ++i) {
-        if (s->bparts[i].time_bin != time_bin_not_created)
-          s->bparts[i].gpart += delta;
-      }
+      /* We now need to correct all the pointers of the other particle arrays */
+      part_relink_all_parts_to_gparts(gparts_new, s->nr_gparts, s->parts,
+                                      s->sinks, s->sparts, s->bparts,
+                                      &s->e->threadpool);
     }
 
     /* Turn some of the allocated spares into particles we can use */

[Matthieu Schaller]

@folkert and @lhausammann.

That solves the mystery. Taking a ptrdiff_t between two pointers and using this as "delta" in another array is actually a so-called undefined behaviour. GCC implements this as we expect. ICC does not. (And both are correct as the standard does not say what to do)

This means that extra gpart likely never worked with the intel compiler until now.

[Matthieu Schaller]

@folkert I also pushed the exact same fix to your branch.

[Loic Hausammann]

Hi,

Thanks for fixing that, I think I would have never been able to figure out. I am a bit surprised that this is undefined.

[Matthieu Schaller]

It's lucky that I had spent days scratching my head about !1000 (merged) in the past and that I remembered. This is really an obscure part of C...

[Folkert Nobels]

Hi Matthieu, thanks for fixing this!