[Bug c++/34099] New: optimizer problem

[Bug c++/34099] New: optimizer problem

[rwgk at yahoo dot com]

Platform:
  Fedora release 7 (Moonshine)
  Linux idle.lbl.gov 2.6.22.9-91.fc7 #1 SMP Thu Sep 27 20:47:39 EDT 2007 x86_64
x86_64 x86_64 GNU/Linux

% g++ -v
Using built-in specs.
Target: x86_64-unknown-linux-gnu
Configured with: /net/rosie/scratch2/rwgk/gcc_trunk/configure
--prefix=/net/cci-filer1/vol1/tmp/rwgk/gcc_trunk_130159_x86_64_fc7
--enable-languages=c,c++,fortran --with-mpfr=/usr
Thread model: posix
gcc version 4.3.0 20071113 (experimental) (GCC)

I'll attach a small reproducer.

It works with -O0:

% g++ -O0 -Wall multiply_bug.cpp -fno-strict-aliasing && ./a.out
(7,4)

But doesn't with -O1 or higher:

% g++ -O1 -Wall multiply_bug.cpp -fno-strict-aliasing && ./a.out
(0,0)

The reproducer #includes complex. I'm guessing it wouldn't be
hard to reduce the reproducer if necessary.

Ralf


-- 
           Summary: optimizer problem
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: rwgk at yahoo dot com
 GCC build triplet: x86_64-unknown-linux-gnu
  GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: x86_64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug c++/34099] optimizer problem

[rwgk at yahoo dot com]

------- Comment #1 from rwgk at yahoo dot com  2007-11-15 06:56 -------
Created an attachment (id=14554)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14554&action=view)
reproducer


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #2 from rguenth at gcc dot gnu dot org  2007-11-15 10:21 -------
Confirmed.  Shorter testcase:

#include <complex>
#include <iostream>

typedef std::complex<double> NumType;

void
multiply(NumType a, NumType b, unsigned ac, NumType &ab)
{
  NumType s;
  for (unsigned j=0; j<ac; j++)
    s = a * b;
  ab = s;
}

int main()
{
  NumType a(1,2), b(3,-2), c;
  multiply(a, b, 1, c);
  std::cout << c << std::endl;
  return 0;
}

Looks like another SRA problem.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aoliva at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
          Component|c++                         |tree-optimization
     Ever Confirmed|0                           |1
           Keywords|                            |wrong-code
      Known to fail|                            |4.3.0
      Known to work|                            |4.2.2
   Last reconfirmed|0000-00-00 00:00:00         |2007-11-15 10:21:31
               date|                            |
            Summary|optimizer problem           |[4.3 Regression] optimizer
                   |                            |problem
   Target Milestone|---                         |4.3.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2007-11-15 14:56 -------
Uhm, this goes wrong in CCP.  (w/o SRA the failure doesn't trigger though)

Before CCP we have:

void multiply(NumType, NumType, unsigned int, NumType&) (a, b, ac, ab)
{ 
  double s$_M_value$real;
  double s$_M_value$imag;
  double b$_M_value$real;
  double b$_M_value$imag;
  double a$_M_value$imag;
  double a$_M_value$real;
  complex double __t;
  complex double D.26933;
  complex double D.26932;
  double D.26931;
  double D.26929;
  unsigned int j;

<bb 2>:
  # VUSE <SFT.90_33(D)>
  a$_M_value$real_21 = REALPART_EXPR <a._M_value>;
  # VUSE <SFT.91_34(D)>
  a$_M_value$imag_22 = IMAGPART_EXPR <a._M_value>;
  # VUSE <SFT.93_35(D)>
  b$_M_value$imag_23 = IMAGPART_EXPR <b._M_value>;
  # VUSE <SFT.92_36(D)>
  b$_M_value$real_24 = REALPART_EXPR <b._M_value>;
  s$_M_value$real_25 = 0.0;
  s$_M_value$imag_26 = 0.0;
  goto <bb 4>;

<bb 3>:
  D.26929_10 = IMAGPART_EXPR <__t_6(D)>;
  __t_11 = COMPLEX_EXPR <b$_M_value$real_24, D.26929_10>;
  D.26931_13 = REALPART_EXPR <__t_11>;
  __t_14 = COMPLEX_EXPR <D.26931_13, b$_M_value$imag_23>;
  D.26932_15 = COMPLEX_EXPR <a$_M_value$real_21, a$_M_value$imag_22>;
  D.26933_16 = __t_14 * D.26932_15;
  s$_M_value$imag_17 = IMAGPART_EXPR <D.26933_16>;
  s$_M_value$real_18 = REALPART_EXPR <D.26933_16>;
  j_4 = j_1 + 1;

<bb 4>:
  # s$_M_value$real_20 = PHI <s$_M_value$real_25(2), s$_M_value$real_18(3)>
  # s$_M_value$imag_19 = PHI <s$_M_value$imag_26(2), s$_M_value$imag_17(3)>
  # j_1 = PHI <0(2), j_4(3)>
  if (j_1 < ac_3(D))
    goto <bb 3>;
  else
    goto <bb 5>;

<bb 5>:
  # SMT.98_38 = VDEF <SMT.98_37(D)>
  IMAGPART_EXPR <ab_5(D)->_M_value> = s$_M_value$imag_19;
  # SMT.98_39 = VDEF <SMT.98_38>
  REALPART_EXPR <ab_5(D)->_M_value> = s$_M_value$real_20;
  return;

} 

where I can see nothing wrong.  Now CCP propagates s$_M_value$real_25 and
s$_M_value$imag_26 into the PHI nodes, but somehow does not simulate BB3:

Visiting PHI node: s$_M_value$real_20 = PHI <s$_M_value$real_25(2),
s$_M_value$real_18(3)>
    Argument #0 (2 -> 4 executable)
        s$_M_value$real_25      Value: CONSTANT 0.0

    Argument #1 (3 -> 4 executable)
        s$_M_value$real_18      Value: UNDEFINED

    PHI node value: CONSTANT 0.0


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dnovillo at google dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #4 from rguenth at gcc dot gnu dot org  2007-11-15 15:08 -------
This looks like this old bug that evaluate_stmt() sets results to UNDEFINED. 
As
we visit

D.26933_16 = __t_14 * D.26932_15;

the result should become VARYING, but we make it UNDEFINED.  Because also
likely_value () returns UNDEFINED for it.  Because in the process of setting
__t we use undefined operands (but those get overwritten).

  D.26929_10 = IMAGPART_EXPR <__t_6(D)>;
  __t_11 = COMPLEX_EXPR <b$_M_value$real_24, D.26929_10>;
  D.26931_13 = REALPART_EXPR <__t_11>;
  __t_14 = COMPLEX_EXPR <D.26931_13, b$_M_value$imag_23>;
  D.26932_15 = COMPLEX_EXPR <a$_M_value$real_21, a$_M_value$imag_22>;
  D.26933_16 = __t_14 * D.26932_15;

so at least

  has_constant_operand = false;
  FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
    {
      prop_value_t *val = get_value (use);

      if (val->lattice_val == UNDEFINED)
        return UNDEFINED; 

      if (val->lattice_val == CONSTANT)
        has_constant_operand = true;
    }

doesn't look conservatively correct, because we pro-actively prefer UNDEFINED
here.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #5 from rguenth at gcc dot gnu dot org  2007-11-15 15:40 -------
That is, a certain class of operations (like COMPLEX_EXPR) do not fulfil the
constraint that if one operand is UNDEFINED the result is UNDEFINED as well.

For example MIN_EXPR <INT_MIN, UNDEFINED>, or MAX_EXPR <INT_MAX, UNDEFINED>
are of such kind as well.  No idea if we want a white-list here instead.
BIT_AND_EXPR <UNDEFINED, 0> is also not UNDEFINED.

(all of the above not with actual constants as defined arg, but VARYING ones)

I have a patch.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rguenth at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2007-11-15 10:21:31         |2007-11-15 15:40:13
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[ebotcazou at gcc dot gnu dot org]

------- Comment #6 from ebotcazou at gcc dot gnu dot org  2007-11-15 15:44 -------
> That is, a certain class of operations (like COMPLEX_EXPR) do not fulfil the
> constraint that if one operand is UNDEFINED the result is UNDEFINED as well.

Is the problem somehow related to PR middle-end/33088?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenther at suse dot de]

------- Comment #7 from rguenther at suse dot de  2007-11-15 15:51 -------
Subject: Re:  [4.3 Regression] optimizer problem

On Thu, 15 Nov 2007, ebotcazou at gcc dot gnu dot org wrote:

> ------- Comment #6 from ebotcazou at gcc dot gnu dot org  2007-11-15 15:44 -------
> > That is, a certain class of operations (like COMPLEX_EXPR) do not fulfil the
> > constraint that if one operand is UNDEFINED the result is UNDEFINED as well.
> 
> Is the problem somehow related to PR middle-end/33088?

It looks like so, but I cannot reproduce the problem in PR33088.

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #8 from rguenth at gcc dot gnu dot org  2007-11-15 16:04 -------
Which doesn't work :(  Unassigning.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|rguenth at gcc dot gnu dot  |unassigned at gcc dot gnu
                   |org                         |dot org
             Status|ASSIGNED                    |NEW


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #9 from rguenth at gcc dot gnu dot org  2007-11-15 16:06 -------
Created an attachment (id=14557)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14557&action=view)
broken patch

It miscompiles gengtype.  I remember problems with changing likely_value in
similar ways back in time.  The other option is to not ignore UNDEFINED lattice
values in ccp_lattice_meet but do

                any  M UNDEFINED   = VARYING

instead of

                any  M UNDEFINED   = any


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #10 from rguenth at gcc dot gnu dot org  2007-11-15 16:08 -------
Created an attachment (id=14558)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14558&action=view)
new broken patch

Err, that was an old patch.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #14557|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #11 from rguenth at gcc dot gnu dot org  2007-11-15 16:38 -------
Oh, trivial error.  Looking again tomorrow.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rguenth at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2007-11-15 15:40:13         |2007-11-15 16:38:30
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #12 from rguenth at gcc dot gnu dot org  2007-11-15 17:10 -------
Trivial testcase we get wrong with the current scheme:

int foo (int b, int c)
{
  int x;
  if (b)
    return x & c;
  else
    return 1;
}
extern void abort (void);
int main()
{
  if (foo(1, 0) != 0)
    abort ();
  return 0;
}

because we say that x & c is UNDEFINED (and so ignore its value in the
otherwise CONSTANT PHI).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #13 from rguenth at gcc dot gnu dot org  2007-11-15 17:17 -------
I think this has been caused by the fix for PR29738.

Author: rakdver
Date: Thu Nov  9 00:09:43 2006
New Revision: 118602

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=118602
Log:
        PR tree-optimization/29738
        * tree-ssa-ccp.c: Remove UNKNOWN_VAL from comments.
        (ccp_lattice_t): Remove UNKNOWN_VAL.
        (dump_lattice_value, ccp_lattice_meet, ccp_visit_phi_node):
        Do not handle UNKNOWN_VAL.
        (get_default_value): Set initial value of virtual operands to
        VARYING.
        (get_value): Always use get_default_value on uninitialized
        operands.
        (set_value_varying, surely_varying_stmt_p): New functions.
        (set_lattice_value): Do not pass argument to get_value.
        Do not handle UNKNOWN_VAL.
        (likely_value): Follow the semantics described in the comment.
        (ccp_initialize): Use surely_varying_stmt_p.  Do not mark
        phi nodes DONT_SIMULATE_AGAIN.
        (ccp_fold): Do not pass argument to get_value.
        (fold_const_aggregate_ref, visit_assignment): Ditto.  Do not
        handle UNKNOWN_VAL.

        * gcc.dg/tree-ssa/ssa-ccp-14.c: New test.
        * gcc.dg/tree-ssa/ssa-ccp-15.c: New test.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
OtherBugsDependingO|                            |29738
              nThis|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #14 from rguenth at gcc dot gnu dot org  2007-11-16 10:10 -------
Subject: Bug 34099

Author: rguenth
Date: Fri Nov 16 10:10:05 2007
New Revision: 130222

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130222
Log:
2007-11-16  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/34099
        * tree-ssa-ccp.c (likely_value): Use a whitelist for operators
        that produce UNDEFINED result if at least one of its operands
        is UNDEFINED.  By default the result is only UNDEFINED if all
        operands are UNDEFINED.

        * g++.dg/torture/pr3499.C: New testcase.
        * gcc.c-torture/execute/pr34099.c: Likewise.

Added:
    trunk/gcc/testsuite/g++.dg/torture/pr34099.C
    trunk/gcc/testsuite/gcc.c-torture/execute/pr34099.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-ccp.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #15 from rguenth at gcc dot gnu dot org  2007-11-16 10:36 -------
Fixed.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #16 from rguenth at gcc dot gnu dot org  2007-11-19 12:52 -------
Subject: Bug 34099

Author: rguenth
Date: Mon Nov 19 12:52:09 2007
New Revision: 130289

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130289
Log:
2007-11-19  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/34099
        * tree-ssa-ccp.c (likely_value): Exclude all but PLUS_EXPR,
        MINUS_EXPR and POINTER_PLUS_EXPR from handling as UNDEFINED
        if only one operand is undefined.

        * gcc.c-torture/execute/pr34099-2.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.c-torture/execute/pr34099-2.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-ssa-ccp.c


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099

[Bug tree-optimization/34099] [4.3 Regression] optimizer problem

[rguenth at gcc dot gnu dot org]

------- Comment #17 from rguenth at gcc dot gnu dot org  2007-11-19 12:53 -------
Subject: Bug 34099

Author: rguenth
Date: Mon Nov 19 12:52:58 2007
New Revision: 130290

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130290
Log:
2007-11-19  Richard Guenther  <rguenther@suse.de>

        PR tree-optimization/34099
        * tree-ssa-ccp.c (likely_value): Exclude all but PLUS_EXPR,
        MINUS_EXPR and POINTER_PLUS_EXPR from handling as UNDEFINED
        if only one operand is undefined.

        * gcc.c-torture/execute/pr34099-2.c: New testcase.

Modified:
    trunk/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34099