CarpetLib::poison_new_memory when aligning memory

[Roland Haas]

looking at the poison code in CarpetLib's mem.cc it seems to me that the poisoning cdoe uses the wrong data pointer ie uses line 144 of mem.cc:

    total_allocated_bytes += nbytes;
    max_allocated_bytes = max (max_allocated_bytes, total_allocated_bytes);
    if (poison_new_memory) {
      memset (storage_, poison_value, nbytes);
    }

when in fact I think it should be storage_base_ since this is the pointer to nbytes bytes of memory returned by malloc.

Attached please find a patch which tries to address this. Note that this only matters on systems where the vector lenght is large enough so that the "natural" alignment provided by malloc (at least 4 bytes I'd assume more likely actually 8 bytes).

Keyword:

[Erik Schnetter]

storage_base_ is the pointer returned from operator new. It is only needed when operator delete is called. There may be additional unused bytes between storage_base_ and storage_, or after storage_+nbytes. These are never used, and are thus also not memset.

It would be correct to

memset(storage_base_, poison_value, nbytes + alignment - 1)

but as is, the patch leaves some bytes at the end of the used storage space unset.

[Roland Haas]

Oh, right. I had apparently not properly checked what the argument to malloc is.