This is a fun one.

In the RTS, `cas` expects a pointer to StgWord which will translate to
unsigned long (8 bytes under LP64.) But we had previously declared
token_locked as *StgBool* - which evaluates to 'int' (4 bytes under
LP64.) That means we fail to provide enough storage for the cas
primitive, causing it to corrupt memory on a 64bit platform.

Hilariously, this somehow did not affect little-endian platforms (ARM,
x86, etc) before. That's because to clear our lock token, we would say:

    token_locked = 0;

But because token_locked is 32bits technically, this only writes to
half of the 64bit quantity. On a Big-Endian machine, this won't do
anything. That is, token_locked starts as 0:

 / token_locked
 |
 v
 0x00000000

and the first cas modifies the memory to:

 / valid    / corrupted
 |          |
 v          v
 0x00000000 0x00000001

We then clear token_locked, but this doesn't change the corrupted 4
bytes of memory. And then we try to lock the token again, spinning until
it is released - clearly a deadlock.

Related: Windows (amd64) doesn't follow LP64, but LLP64, where both
int and long are 4 bytes, so this shouldn't change anything on these
platforms.

Thanks to Reid Barton for helping the diagnosis. Also, thanks to Jens
Peterson who confirmed this also fixes building GHC on Fedora/ppc64 and
Fedora/s390x.

Authored-by: Gustavo Luiz Duarte <gustavold@linux.vnet.ibm.com>
Signed-off-by: Austin Seipp <austin@well-typed.com>