------- Comment #6 from jakub at gcc dot gnu dot org 2007-08-23 20:36 -------
It is unfortunate that gimple can only initialize the whole array, unless
__builtin_memcpy is used. Unfortunately __builtin_memcpy has a different
drawback - the var will be forced to be TREE_ADDRESSABLE even when it otherwise
wouldn't have to be.
So, we have the option to add some flag to count_type_elements etc., so that
they will behave one way for expr.c etc. and one way for the gimplifier - there
it would need to count whole arrays as units if they are initialized from
string
literals, etc.
The other option is IMHO to hint during expansion expand_assignment that
the memory is already cleared by previous var = {} and so it can avoid storing
zeros. Or teach some RTL pass to remove the redundant zero stores. If
the big clear_storage is done through clearing by pieces, DSE will already
handle this to some extent (will remove redundant stores from the big
clear_storage), but if it is done through a special insn (like rep stosb) or
through memset call, nothing will remove it.
While looking at this, I have noticed that a simple
struct A { char c[10]; };
void
foo (void)
{
struct A a = { "abcdefghi" };
baz (&a);
}
void
bar (void)
{
struct A a;
__builtin_strcpy (&a.c[0], "abcdefghi");
baz (&a);
}
is not as efficient as it ought to be in the foo function - while strcpy/memcpy
can and in this case will use store_by_pieces, store_expr will always emit the
string literal to memory and then emit a block move from it to target followed
optionally by clear_storage of the rest of the memory.
I'm testing a patch to improve this ATM.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31150
