> The following testcase is miscompiled on i686-linux at -O3.
> The bug is in DSE record_store, which for group_id < 0 uses mem_addr
> set to result of get_addr (base->val_rtx) (plus optional offset),
> which is fine for canon_true_dependence with other MEMs in that function,
> but we also store that address in store_info. The problem is if later on
> e.g. some read uses the same e.g. hard register as get_addr returned, but
> that register contains at that later point a different value.
> canon_true_dependence then happily returns the read does not alias the
> store, although it might.
> The fix is to store the VALUE (plus optional offset) into
> store_info->mem_addr instead, then at some later insn when get_addr is
> called on it it will either return the same register or expression (if it
> has not changed), or some different one otherwise.
I presume that the origin of the bug is:
/* get_addr can only handle VALUE but cannot handle expr like:
VALUE + OFFSET, so call get_addr to get original addr for
mem_addr before plus_constant. */
mem_addr = get_addr (mem_addr);
if (offset)
mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset);
both in record_store and check_mem_read_rtx, so I wonder if we shouldn't bite
the bullet and try enhancing get_addr since it's a mainline-only regression.
--
Eric Botcazou
