Re: [C PATCH] Discard P - (P + CST) optimization in pointer_diff (PR c/61240)

[Jeff Law <law at redhat dot com>]

On 08/05/14 08:36, Marek Polacek wrote:
> On Mon, Aug 04, 2014 at 02:04:36PM +0200, Richard Biener wrote:
> > > Looks like .fre can optimize "q - (q - 1)" into 1:
> > >     <bb 2>:
> > >     q.0_3 = (long int) &MEM[(void *)&i + 4B];
> > >     _5 = (long int) &i;
> > > -  _6 = q.0_3 - _5;
> > > -  t.1_7 = _6 /[ex] 4;
> > > -  t ={v} t.1_7;
> > > +  t ={v} 1;
> > >     i ={v} {CLOBBER};
> > >     return;
> > >
> > > But associate_plusminus doesn't optimize it:
> > >            else if (code == MINUS_EXPR
> > >                     && CONVERT_EXPR_CODE_P (def_code)
> > >                     && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
> > >                     && TREE_CODE (rhs2) == SSA_NAME)
> > >              {
> > >                /* (T)(P + A) - (T)P -> (T)A.  */
> > > becase gimple_assign_rhs1 (def_stmt) is not an SSA_NAME, but ADDR_EXPR (it's
> > > &MEM[(void *)&i + 4B]).  Then there's transformation A - (A +- B) -> -+ B
> > > below, but that doesn't handle casts.
> > >
> > > So - should I try to handle it in associate_plusminus?
> >
> > Yes please, with a (few) testcase(s).
>
> Ok, so the following adds the (T)P - (T)(P + A) -> (T)-A
> transformation.  It is based on code hunk that does the
> (T)(P + A) - (T)P -> (T)A transformation.  The difference it makes is
> in the .optimized dump something like:
>
>   int fn2(int, int) (int p, int i)
>   {
> -  unsigned int p.2_2;
> +  unsigned int _3;
>     int _4;
> -  unsigned int _5;
>     unsigned int _6;
> -  int _7;
>
>     <bb 2>:
> -  p.2_2 = (unsigned int) p_1(D);
> -  _4 = p_1(D) + i_3(D);
> -  _5 = (unsigned int) _4;
> -  _6 = p.2_2 - _5;
> -  _7 = (int) _6;
> -  return _7;
> +  _6 = (unsigned int) i_2(D);
> +  _3 = -_6;
> +  _4 = (int) _3;
> +  return _4;
>
> i.e., the PLUS_EXPR and MINUS_EXPR are gone, and NEGATE_EXPR is used instead.
> During bootstrap with --enable-languages=c,c++ this optimization triggered 238 times.
>
> Bootstrapped/regtested on x86_64-linux, ok for trunk?
>
> 2014-08-05  Marek Polacek  <polacek@redhat.com>
>
> 	PR c/61240
> 	* tree-ssa-forwprop.c (associate_plusminus): Add (T)P - (T)(P + A)
> 	-> (T)-A transformation.
> c/
> 	* c-typeck.c (pointer_diff): Remove P - (P + CST) optimization.
> testsuite/
> 	* gcc.dg/pr61240.c: New test.
> 	* gcc.dg/tree-ssa/forwprop-29.c: New test.
So I'm all for delaying folding when possible, so I'm comfortable with 
the general direction this is going.

My concern is the code we're removing discusses the need to simplify 
when these expressions are in static initializers.  What's going to 
ensure that we're still simplifying instances which appear in static 
initializers?  I don't see anything which tests that.   And does it 
still work for targets which utilize PSImode?

Jeff