A common difference between bytecode and native programs is that the native compiler computes the liveness of variables in the stack, so that they are not scanned anymore afterwards (or coalesced with other variables). For example, if you have a data structure that is allocated before a computation, but not used in the computation, the structure might be kept alive in the bytecode, but garbage collected in native code.

Le mar. 28 mai 2019 à 19:23, Ivan Gotovchits <ivg@ieee.org> a écrit :

My (common) suspect is Zarith. It uses a different backend when compiled to bytecode, so I would start my investigation from this point :)

Hope it helps,
Ivan

On Tue, May 28, 2019 at 8:27 AM Jon French <jf451@cam.ac.uk> wrote:

Hi all,

I'm wondering if anyone on this list might have encountered this issue before, since it's got me pretty stumped.

We have an OCaml project ( https://github.com/rems-project/sail ) which tries to allocate huge amounts of memory, but only when compiled to bytecode. The native version works perfectly fine.

strace for an example run ends in:

openat(AT_FDCWD, "/home/ojno/work/sail2/lib/vector_dec.sail", O_RDONLY|O_CLOEXEC) = 3
lseek(3, 0, SEEK_CUR)                   = 0
read(3, "$ifndef _VECTOR_DEC\n$define _VEC"..., 65536) = 7031
mmap(NULL, 1965819695104, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = -1 ENOMEM (Cannot allocate memory)
brk(0x5729f32c4000)                     = 0x55603f2f4000
mmap(NULL, 1965819826176, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = -1 ENOMEM (Cannot allocate memory)
write(2, "Fatal error: out of memory.\n", 28Fatal error: out of memory.
) = 28
exit_group(2)                           = ?
+++ exited with 2 +++

which naturally fails since I don't have 2 TB of memory available.

The exact amount of the allocation varies from input to input and even run to run, but is always that approximate size. There's no characteristic pattern of heap or stack running out of control if I turn on memory debugging info in $OCAMLRUNPARAM, just that one huge allocation. When run in ocamldebug, it also doesn't appear to be failing at a location which is sensible to be allocating such amounts of memory -- and the location also varies with the input. And on tiny inputs it doesn't seem to try and make the allocation at all.

Are we actually seeing a compiler/runtime bug here? Is there something I'm missing?

Thanks,

Jon French
Computer Laboratory
University of Cambridge