Re: [PATCH, PR 44133] Preserve sane warnings when deleting loads in SRA

[Richard Guenther <rguenther at suse dot de>]

On Mon, 17 May 2010, Martin Jambor wrote:

> Hi,
> 
> On Mon, May 17, 2010 at 11:28:05AM +0200, Richard Guenther wrote:
> >
> 
> ...
> 
> > 
> > Are you sure that it helps even when you add another memory store
> > before the access in the testcase, like
> > 
> > + /* { dg-do compile } */
> > + /* { dg-options "-O2 -Wall" } */
> > +
> > + struct S { int i, j; };
> > +
> > int x;
> > + int foo (int l)
> > + {
> > +   struct S s;
> > x = 0;
> > +   s.j = l - 22;   /* { dg-warning ".s\.i. is used uninitialized" } */
> > +   return s.i + s.j;
> > + }
> 
> Yes, it still emits the same warning (still at the wrong line).  Why
> should this matter?
> 
> > 
> > ?  How does the function look like after SRA with your patch?
> 
> Before early SRA it looks like:
> 
> foo (int l)
> {
>   struct S s;
>   int D.2726;
>   int D.2725;
>   int D.2724;
>   int D.2723;
> 
> <bb 2>:
>   x = 0;
>   D.2723_2 = l_1(D) + -22;
>   s.j = D.2723_2;
>   D.2725_3 = s.i;   // <-- this statement is removed and def-defs propagated
>   D.2726_4 = s.j;
>   D.2724_5 = D.2725_3 + D.2726_4;
>   return D.2724_5;
> 
> }
> 
> The current (unpatched) early SRA transforms it into:
> 
> foo (int l)
> {
>   int s$j;
>   struct S s;
>   int D.2726;
>   int D.2725;
>   int D.2724;
>   int D.2723;
> 
> <bb 2>:
>   x = 0;
>   D.2723_2 = l_1(D) + -22;
>   s$j_9 = D.2723_2;
>   D.2726_4 = s$j_9;
>   D.2724_5 = D.2726_4 + D.2725_8(D);  // <-- this def-def is the problem
>   return D.2724_5;
> 
> }
>  
> Whereas early SRA with the proposed patch produces:
> 
> foo (int l)
> {
>   int s$i;
>   int s$j;
>   struct S s;
>   int D.2726;
>   int D.2725;
>   int D.2724;
>   int D.2723;
> 
> <bb 2>:
>   x = 0;
>   D.2723_2 = l_1(D) + -22;
>   s$j_9 = D.2723_2;
>   D.2726_4 = s$j_9;
>   D.2724_5 = D.2726_4 + s$i_8(D);
>   return D.2724_5;
> 
> }

Ah, ok.  I was confused by the non-renaming stuff.

The patch is ok.

Thanks,
Richard.