[Bug tree-optimization/31146] New: forwprop does not look through casts

[Bug tree-optimization/31146] New: forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

The following two testcases should produce the same code after forwprop1:

struct foo { int i[3]; };

void bar (void)
{
  struct foo Foo;
  int i;

  for (i=0; i<3; ++i)
    {
      void *p = &Foo.i[i];
      int *pi = (int *)p;
      if (pi != 0)
        {
          *pi = 0;
        }
    }
}

---

struct foo { int i[3]; };

void bar (void)
{
  struct foo Foo;
  int i;

  for (i=0; i<3; ++i)
    {
      void *p = &Foo.i[i];
      int *pi = (int *)p;
      if (p != 0)
        {
          *pi = 0;
        }
    }
}


(the difference is testing p != 0 vs. pi != 0)

This happens in tramp3d-v4 via placement new:

  for (i=0; i<3; ++i)
    new (&a[i]) double;

where placement new expands to the "wrong" test sequence.

Mine.


-- 
           Summary: forwprop does not look through casts
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: rguenth at gcc dot gnu dot org
        ReportedBy: rguenth at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2007-03-12 14:59 -------
forwprop should NOT, I repeat should NOT look through casts for conditional if
the SSA_NAME that defines the cast is still used after the conditional.

Really this non zero optimization should be done by VRP.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2007-03-12 15:09 -------
I should mention, unless it can produce a constant result.
For an example this function:

void bar (void * a)
{
  struct foo Foo;
  int i;

  for (i=0; i<3; ++i)
    {
      void *p = a;
      int *pi = (int *)p;
      if (pi != 0)
        {
          *pi = 0;
        }
    }
}

Should not be transformed into:

void bar (void * a)
{
  struct foo Foo;
  int i;

  for (i=0; i<3; ++i)
    {
      void *p = a;
      int *pi = (int *)p;
      if (p != 0)
        {
          *pi = 0;
        }
    }
}

As VRP is not as smart as there is no back tracking of ranges.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2007-03-12 15:35 -------
The problem is that for example FRE value numbers

  void *p_4 = &a[0];
  int *q_1 = (int *)p_4;

p_4 with void* type (even if &a[0] is of int* type) and so re-generates the
conversion to int* even though it is about to replace (int *)p_4 with p_4
(as tree_ssa_useless_type_conversion_1 is false for converting void* to int*).

At the same time tree-combining the conversion (int *)&a[0] leads to
&a[0] for which we have a value already, the void* p_4 :/

Maybe the new VN will fix this testcase as well.  With the series of forwprop
patches in my queue (which you obviously don't like at all, but I blame you
for that) we get for

  (void) (TARGET_EXPR <D.2605, (double *) operator new (8, (void *) &a[i])>;,
D.2605 != 0B ? try
    {
      *D.2605 = 0.0;
    }
  catch
    {
      operator delete ((void *) D.2605, (void *) &a[i]);
    }, D.2605 : D.2605;)

instead of

  __p_5 = &a[i_3];
  iftmp.3_6 = (double *) __p_5;
  if (iftmp.3_6 != 0B) goto <L1>; else goto <L2>;

<L1>:;
  *iftmp.3_6 = 0.0;

<L2>:;

just

  a[i_3] = 0.0;


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2007-03-12 15:41 -------
I think really this comes down to our type system, I remember posting a patch
for the LTO branch which adds back the explict cast to "void*", I can test
that, it should fix this testcase also :).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #5 from rguenth at gcc dot gnu dot org  2007-03-12 16:01 -------
Well sure - if we then would get

 int *p_1 = &a[0];
 void *q_1 = (void *)p_1;
 int *r_1 = (int *)q_1;
 if (r_1 != NULL)
   ...

then FRE would figure out that r_1 == p_1 and the forwprop pass after FRE will
fold the comparison.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #6 from rguenth at gcc dot gnu dot org  2007-03-16 10:11 -------
Subject: Bug 31146

Author: rguenth
Date: Fri Mar 16 10:11:14 2007
New Revision: 122985

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=122985
Log:
2007-03-16  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/31146
        * tree-ssa-forwprop.c (forward_propagate_addr_expr_1): Restructure
        to allow recursion of forward_propagate_addr_expr.
        (forward_propagate_addr_into_variable_array_index): Likewise.
        (forward_propagate_addr_expr): Likewise.
        (tree_ssa_forward_propagate_single_use_vars): Likewise.
        (forward_propagate_addr_expr_1): Recurse on simple copies
        instead of propagating into them.  Do so for useless conversions
        as well.
        (forward_propagate_addr_expr): Clean up unused statements after
        recursion.

        * g++.dg/tree-ssa/pr31146.C: New testcase.

Added:
    trunk/gcc/testsuite/g++.dg/tree-ssa/pr31146.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-forwprop.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #7 from rguenth at gcc dot gnu dot org  2007-03-20 11:05 -------
A slightly related testcase is now properly optimized.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[mueller at gcc dot gnu dot org]

------- Comment #8 from mueller at gcc dot gnu dot org  2007-03-20 12:48 -------
*** Bug 30864 has been marked as a duplicate of this bug. ***


-- 

mueller at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dcb314 at hotmail dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #9 from rguenth at gcc dot gnu dot org  2007-05-02 10:13 -------
Subject: Bug 31146

Author: rguenth
Date: Wed May  2 09:12:49 2007
New Revision: 124349

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=124349
Log:
2007-05-02  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/31146
        * tree-ssa-forwprop.c (forward_propagate_addr_expr_1): New
        argument, single_use_p.  If we have a single use that is
        a conversion to the definition rhs type, propagate that rhs.
        (forward_propagate_addr_expr): Pass single_use_p argument
        to forward_propagate_addr_expr_1.

        * g++.dg/tree-ssa/pr31146-2.C: New testcase.

Added:
    trunk/gcc/testsuite/g++.dg/tree-ssa/pr31146-2.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-forwprop.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146

[Bug tree-optimization/31146] forwprop does not look through casts

[rguenth at gcc dot gnu dot org]

------- Comment #10 from rguenth at gcc dot gnu dot org  2007-05-02 10:17 -------
Fixed.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.3.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31146