[Bug tree-optimization/15524] [4.0 Regression] jump threading on trees is slow with switch statements with large # of cases

["law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From law at redhat dot com  2004-09-24 16:04 -------
Subject: Re:  [4.0 Regression] jump threading
	on trees is slow with switch statements with large # of cases

On Fri, 2004-09-24 at 04:29, dnovillo at redhat dot com wrote:
> ------- Additional Comments From dnovillo at redhat dot com  2004-09-24 10:29 -------
> Subject: Re:  [4.0 Regression] jump threading
> 	on trees is slow with switch statements with large # of cases
> 
> On Fri, 2004-09-24 at 02:13, law at redhat dot com wrote:
> 
> > We can turn that into
> > 
> >   switch (x)
> >     {
> >       case 42:
> >         x = 42;
> >         whatever
> >     }
> > 
> This is what I'm working with as part of VRP.  The idea is to insert
> ASSERT_EXPR that act as special assignments.  I have a WIP patch if you
> are interested, but I at the moment I'm paying more attention to fixes
> for 4.0.
But for the case of a simple x == 42 or a case statement such as
I've shown above, an ASSERT_EXPR is, well, silly.  Just issue
x = 42 in the appropriate place and let's get on with real work :-)


Of course given a conditional such as

if (x == 42)
  goto BB2
else
  goto BB4

Using ASSERT_EXPR I would think you'd really want

if (x == 42)
  x = 42;
  goto BB2
else
  x = ASSERT_EXPR (x != 42)

ie, you use the ASSERT_EXPR to assert something about the value of
an expression, not the value of objects within the expression.  And
if that is the case, then the LHS of these assert expressions ought
to be a new formal temporary.  And if I've understood all this
correctly, you can get precisely the same behavior without using
ASSERT_EXPR.

t = (x == 42)
if (t)
  t = 1;
  goto BB2;
else
  t = 0;
  goto BB4;

You set up an equivalency between t and (x == 42) when you hit the
assignment, then you add the constant true to the equivalency list
in the true arm (removing it of course when you're done with the
true arm).  Then add the constant false to the equivalancy list
in the false arm.

This is a perfect example of why I want to have an equivalency list,
much like the tree-vn code provides.

So before you go too much further with ASSERT_EXPR I think we need
to reach an agreement that ASSERT_EXPR is actually useful.  I believe
you can do the same thing without ASSERT_EXPR using facilities that
already exist today.

Jeff



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15524