Re: [tree-ssa] More aliasing fixes

[Diego Novillo <dnovillo at redhat dot com>]

On Wed, 2003-12-17 at 07:08, Jan Hubicka wrote:

> I see, this makes it more Morgan like.
>
How so?  It was just a bug fix :)

> We still don't have separate tag
> for each load, instead we take advantage of the fact that pointers with
> same properties will have equivalently looking tags, right?
>
Yes.  Pointers whose pointed-to types have the same alias set will have
the same memory tag (if they have been dereferenced).

> I've been considering the idea of making the whole thing more Morgan
> like - ie making may_alias arrays transitive and always emit one VUSE
> for loads (currently we VUSE all aliased tags, right?).  Do you see some
> fundamental problem with this idea?
> 
It's a possibility, yes.  If you want loads to have a single VUSE, you
will cause stores to aliased variables to have 2 VDEFs and loads from
aliased variables to have 2 VDEFs.

foo (int i)
{
  int a, b, c;

  p = (i > 10) ? &a : (i < 20) ? &b : &c;
				Current		  New
  *p = ...		<-- 3 VDEFs		1 VDEF
  a = ...;		<-- 1 VDEF		2 VDEFs
  b = ...;		<-- 1 VDEF		2 VDEFs
  c = a + b;		<-- 1 VDEF, 2 VUSEs	2 VDEFs, 4 VUSEs
  *p = a - c;		<-- 3 VDEFs, 2 VUSEs	1 VDEF, 4 VUSEs

  return c;		<-- 1 VUSE		2 VUSEs

			   14 vops		18 vops
}

With the current implementation we'd generate 14 vops, the new one would
generate 18 vops.  But we would have to experiment with real code.  I
can see definite advantages to the new scheme in pointer-heavy code.


Diego.