------- Comment #15 from jakub at gcc dot gnu dot org 2008-11-13 15:28 -------
This is a regression from 3.4 still present in 4.3/4.4.
At the tree level, we only fold memcpy if the size of the source and
destination
elements is the same and length is tgat size (say
int i = 6;
float f;
memcpy (&f, &i, sizeof (int));
) into an assignment. As this isn't the case here, before expansion we have:
main ()
{
int arg1;
int arg2;
int arg3;
int arg4;
int arg5;
int D.1602;
<bb 2>:
D.1602 = *(int *) "abcd";
arg1 = D.1602;
arg2 = D.1602;
arg3 = D.1602;
arg4 = D.1602;
arg5 = D.1602;
__builtin_memcpy (&buf, &arg1, 4);
__builtin_memcpy (&buf[4], &arg2, 4);
__builtin_memcpy (&buf[8], &arg3, 4);
__builtin_memcpy (&buf[12], &arg4, 4);
__builtin_memcpy (&buf[16], &arg5, 4);
buf[20] = 0;
__builtin_puts (&buf);
return 0;
}
and all of arg{1,2,3,4,5} are TREE_ADDRESSABLE. In 3.4 ADDRESSOF took care of
optimizing out these dead stores. But ADDRESSOF is gone, and RTL DSE is
apparently unable to kill these. No idea why dse1 doesn't kill it, dse2
doesn't
at least on x86_64 because pro_epilogue_adjust_stack_rex64 has (clobber
(mem:BLK (scratch))) in it.
On the tree level, we could perhaps optimize this by adding special builtins
and transforming __builtin_memcpy into it if possible. Say have
void *__builtin_memstore2 (char *, HImode);, void *__builtin_memstore4 (char *,
SImode); etc., and transform say
__builtin_memcpy (buffer + 8, &arg2, sizeof (arg2)) into
__builtin_memstore4 (buffer + 8, arg2); (or for non-integral types, say arg2
being float, __builtin_memstore4 (buffer + 8, VCE<SImode, arg2>). At expansion
time these builtins would check if the second argument must live in memory.
If yes, it would just expand it as __builtin_memcpy. Otherwise it would just
do a store (with alias set 0). Or we could try to play with
TYPE_REF_CAN_ALIAS_ALL pointers.
That said, it would be good to know why DSE doesn't eliminate these dead
stores.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29215
