Re: [PATCH] Allow tree-ssa Forward Prop to produce VIEW_CONVERT_EXPR in some cases

["Richard Guenther" <richard dot guenther at gmail dot com>]

On Fri, May 2, 2008 at 5:32 AM, Andrew Pinski <pinskia@gmail.com> wrote:
> While looking into PR 26069, I had posted a patch which made
>  fold_indirect_ref create a VIEW_CONVERT_EXPR, but that failed because
>  using VCE on the lhs side has some issues as we could produce a subreg
>  like setting.  So Instead I decided to add to forwprop a way to
>  produce it.  I first did this for the PS3 toolchain which is GCC 4.1.1
>  based and then I redid it for the trunk where the code has changed a
>  lot and actually made it easier to do already.  We now are able to
>  convert *(int*)&float_var into a VIEW_CONVERT_EXPR (we still get a
>  warning about aliasing violation which is why I requested that we kept
>  the warning in the front-end for this case :)) and get the behavior
>  what users would expect.
>
>  OK? Bootstrapped and tested on i686-linux-gnu with no regressions.

It looks like this will now refuse to forwprop

struct X { int a[5]; };
int foo(struct X *q)
{
  int (*p)[5] = &q->a;
  return (*p)[0];
}

?

It also looks it will now forwprop

int i;
int foo(void)
{
  volatile int *p = (volatile int *)&i;
  return *p + *p;
}

to VIEW_CONVERT_EXPR <volatile int>(i) -- I don't think that V_C_E
is designed to carry changes just in CV qualifiers.

Also if you cannot use fold_build1 because of the invalid gimple issue,
won't you run into this problem with fold_stmt as well?  Will 'i' become
non-addressable at any time if it is used in a V_C_E?  Do we fold it
then?

Thanks,
Richard.