On Sun, Aug 28, 2016 at 7:59 PM, Jan Hubicka <hubi...@ucw.cz> wrote:
> Hi,
> I have noticed that ivopts build MEM_REF (NOP_EXPR (address))
> where NOP_EXPR converts one pointer type to another. Later it tries
> to expand it that on strict alignment targets punt on determining
> the alignment and winds expensive memory reference.
>
> It seems to make no sense to have NOP_EXPR here, but in addition
> I think we want to somehow preserve alignment from original MEM_REF?
As you noticed the type carries semantics thus if you want to fix it this
way then you need to make build_simple_mem_ref strip the nops _after_
remembering the type to use for the MEM_REF (and the aliasing).
After all you are changing *(int *)&large-struct to *&large-struct with
your patch - certainly not what you intended?
Sth like
@@ -4618,6 +4623,7 @@ build_simple_mem_ref_loc (location_t loc
HOST_WIDE_INT offset = 0;
tree ptype = TREE_TYPE (ptr);
tree tem;
+ STRIP_NOPS (ptr);
/* For convenience allow addresses that collapse to a simple base
and offset. */
if (TREE_CODE (ptr) == ADDR_EXPR
Richard.
> Honza
>
> * tree-ssa-lop-ivopts.c (get_computation_cost_at):
> strip nops before building mem ref
> diff --git a/gcc/tree-ssa-loop-ivopts.c b/gcc/tree-ssa-loop-ivopts.c
> index 62ba71b..bd92031 100644
> --- a/gcc/tree-ssa-loop-ivopts.c
> +++ b/gcc/tree-ssa-loop-ivopts.c
> @@ -5058,7 +5058,10 @@ fallback:
> return infinite_cost;
>
> if (address_p)
> - comp = build_simple_mem_ref (comp);
> + {
> + STRIP_NOPS (comp);
> + comp = build_simple_mem_ref (comp);
> + }
>
> cost = comp_cost (computation_cost (comp, speed), 0);
>
