Re: Memory aliasing and stack slots

Re: Memory aliasing and stack slots

[Geert Bosch]

On Sun, 28 May 00 08:19:18 EDT, Richard Kenner wrote:

  Presumably, at least in some languages, X and Y will have different alias
  sets, so they won't conflict.  But that means that we might move load of X
  until after a store of Y, which is wrong.

If X and Y share the same stack slot, it seems to me that they should be
in the same alias set. Do you know if it is possible to determine
that X and Y share the same stack slot?

-Geert

Re: Memory aliasing and stack slots

[Richard Henderson]

On Sun, May 28, 2000 at 08:07:28AM -0400, Richard Kenner wrote:
> Presumably, at least in some languages, X and Y will have different alias
> sets, so they won't conflict.  But that means that we might move load of X
> until after a store of Y, which is wrong.

See assign_stack_temp_for_type -- X and Y will not share a
stack slot if they do not have compatible alias sets.


r~

Re: Memory aliasing and stack slots

[Richard Kenner]

    If X and Y share the same stack slot, it seems to me that they should be
    in the same alias set. 

The alias set of an object is determined by its type and two objects of
different types can be assigned the same stack slot at different times.

    Do you know if it is possible to determine that X and Y share the
    same stack slot?

In general it's not.  Consider taking the address and passing it
through a const function that returns a pointer to the stucture.

Memory aliasing and stack slots

[Richard Kenner]

I have a concern with the reuse of stack slots and memory aliasing.

Suppose we have a case where there are two lexical blocks within the same
basic block and we have something like:

	{
	  struct bar x;

	  x = ...
	  ... x ...
	}
	{
	  struct foo y;

	  y = ...
	  ... y ...
	}

Presumably, at least in some languages, X and Y will have different alias
sets, so they won't conflict.  But that means that we might move load of X
until after a store of Y, which is wrong.

We have exactly the same problem with having read-only stack slots, but it
will be much rarer.

Is there any good suggestion to fix this?  Perhaps we could put some note in
the insn stream when a stack slot dies, but that's a lot of notes.