http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60291
--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so the rest is hash collisions in the mem-attrs hash. We also do useless
work here:
case MEM_REF:
{
addr_space_t as
= TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (TREE_OPERAND (exp, 0))));
...
set_mem_attributes (temp, exp, 0);
set_mem_addr_space (temp, as);
set_mem_attributes already handles MEM_REFs properly:
/* Address-space information is on the base object. */
if (TREE_CODE (base) == MEM_REF
|| TREE_CODE (base) == TARGET_MEM_REF)
as = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (TREE_OPERAND (base,
0))));
so removing that would cut expand time in half. Oh, and expand_debug_expr
even gets addr-spaces wrong (in cfgexpand.c).
We have 1107000 callso to set_mem_attrs but 151459000 calls to
mem_attrs_htab_eq.
The hash function is somewhat ad-hoc apart from hashing MEM_EXPR via
iterative_hash_expr. We hash MEM_REFs via hashing their operands
recursively but we compare them more strictly factoring in properties
of types involved. I don't see an easy way to make the hashing stronger
(given PR58521 and its fix). There is always the old idea of simply
removing that mem_attr sharing completely ...
At -O0 we could simply always force a NULL_TREE MEM_EXPR for example.
