[Bug fortran/56015] New: Option -ffast-math reveals i*(a+bi) -> -b-bi, a complex multiplication bug

[Bug fortran/56015] New: Option -ffast-math reveals i*(a+bi) -> -b-bi, a complex multiplication bug

[t_nissie at yahoo dot co.jp]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

             Bug #: 56015
           Summary: Option -ffast-math reveals i*(a+bi) -> -b-bi, a
                    complex multiplication bug
    Classification: Unclassified
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: t_nissie@yahoo.co.jp


With gfortran, optimization option "-O3 -ffast-math", more
precisely "-O3 -funsafe-math-optimizations -ffinite-math-only"
reveals a complex multiplication bug of
i*(a+bi) =
Good: -b+ai
Bug:  -b-bi

Source:
! multiplyi.f -*-f90-*-
!!
program multiplyi
  implicit none
  complex*16 p(10)
  p(:) = (0.1d0, 0.2d0)
  p(:) = (0.0d0, 1.0d0) * p(:)
  write(6,'(2f5.1)') p(1)
end program multiplyi

Compilation and execution:
$ gfortran --version
GNU Fortran (GCC) 4.7.2
$ gfortran -ffree-form -O3 -o nofast-math multiplyi.f && ./nofast-math
  -0.2  0.1
$ gfortran -ffree-form -O3 -ffast-math -o fast-math multiplyi.f && ./fast-math
  -0.2 -0.2

Details:
 * I could reproduce this bug with gcc-4.4.6, 4.6.3, and 4.7.2.
 * I could reproduce it on Intel Core i5 and Xeon5650 (64bit)
   and VIA C7-M (32bit).
 * To reproduce this bug, the variable p(10) should not be an
   array in 4.6.3, but it should be an array in 4.7.2.

Tank you for your constant efforts on gfortran,
Takeshi

[Bug middle-end/56015] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-01-17
          Component|fortran                     |middle-end
            Summary|Option -ffast-math reveals  |expand expands p[9] =
                   |i*(a+bi) -> -b-bi, a        |COMPLEX_EXPR
                   |complex multiplication bug  |<-IMAGPART_EXPR <p[9]>,
                   |                            |REALPART_EXPR <p[9]>>;
                   |                            |incorrectly.
     Ever Confirmed|0                           |1
      Known to fail|4.4.6, 4.6.3, 4.7.2         |4.3.5, 4.4.5, 4.8.0

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2013-01-17 08:52:25 UTC ---
Actually it is not the multiplication that goes wrong.  In fact the code looks
correct before expand.  Turning off TER (-fno-tree-ter) will work around this
expand issue though.

[Bug middle-end/56015] [4.6/4.7/4.8 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
                 CC|                            |jakub at gcc dot gnu.org
      Known to work|                            |4.1.2
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |
   Target Milestone|---                         |4.6.4
            Summary|expand expands p[9] =       |[4.6/4.7/4.8 Regression]
                   |COMPLEX_EXPR                |expand expands p[9] =
                   |<-IMAGPART_EXPR <p[9]>,     |COMPLEX_EXPR
                   |REALPART_EXPR <p[9]>>;      |<-IMAGPART_EXPR <p[9]>,
                   |incorrectly.                |REALPART_EXPR <p[9]>>;
                   |                            |incorrectly.

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-17 10:30:11 UTC ---
Works correctly with 4.1.2, where we had
  CR.108 = REALPART_EXPR <p[0]>;
  REALPART_EXPR <p[0]> = -IMAGPART_EXPR <p[0]>;
  IMAGPART_EXPR <p[0]> = CR.108;
in the *.optimized dump.  Thus a regression.

[Bug middle-end/56015] [4.6/4.7/4.8 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-17 10:31:08 UTC ---
Created attachment 29189
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29189
gcc48-pr56015.patch

Untested fix.

[Bug middle-end/56015] [4.6/4.7/4.8 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> 2013-01-17 10:48:40 UTC ---
Hmm, I wonder if it isn't better to disable TER for these cases.  We might
run into similar issues with other loads (BIT_FIELD_REF and vectors for
example).

[Bug middle-end/56015] [4.6/4.7/4.8 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[t_nissie at yahoo dot co.jp]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

--- Comment #5 from Takeshi Nishimatsu <t_nissie at yahoo dot co.jp> 2013-01-17 11:58:24 UTC ---
(In reply to comment #3)
> Created attachment 29189 [details]
> gcc48-pr56015.patch
> 
> Untested fix.

Thank you for your fix.
I add one more test multiplying -i.

program pr56015
  implicit none
  complex*16 m(10),p(10)
  m(:) = (0.3d0, 0.4d0)
  m(:) = (0.0d0,-1.0d0) * m(:)
  call bar (m)
  p(:) = (0.1d0, 0.2d0)
  p(:) = (0.0d0, 1.0d0) * p(:)
  call foo (p)
contains
  subroutine bar (m)
    complex*16 m(10)
    if (any (m .ne. ( 0.4d0,-0.3d0))) call abort
  end subroutine
  subroutine foo (p)
    complex*16 p(10)
    if (any (p .ne. (-0.2d0, 0.1d0))) call abort
  end subroutine
end program pr56015

[Bug middle-end/56015] [4.6/4.7/4.8 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-18 17:15:17 UTC ---
Author: jakub
Date: Fri Jan 18 17:15:07 2013
New Revision: 195301

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195301
Log:
    PR middle-end/56015
    * expr.c (expand_expr_real_2) <case COMPLEX_EXPR>: Handle
    the case where writing real complex part of target modifies
    op1.

    * gfortran.dg/pr56015.f90: New test.

Added:
    trunk/gcc/testsuite/gfortran.dg/pr56015.f90
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/expr.c
    trunk/gcc/testsuite/ChangeLog

[Bug middle-end/56015] [4.6/4.7 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.8.0
            Summary|[4.6/4.7/4.8 Regression]    |[4.6/4.7 Regression] expand
                   |expand expands p[9] =       |expands p[9] = COMPLEX_EXPR
                   |COMPLEX_EXPR                |<-IMAGPART_EXPR <p[9]>,
                   |<-IMAGPART_EXPR <p[9]>,     |REALPART_EXPR <p[9]>>;
                   |REALPART_EXPR <p[9]>>;      |incorrectly.
                   |incorrectly.                |
      Known to fail|4.8.0                       |

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-18 17:25:46 UTC ---
Fixed on the trunk so far.

[Bug middle-end/56015] [4.6/4.7 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-02-01 14:11:03 UTC ---
Author: jakub
Date: Fri Feb  1 14:10:48 2013
New Revision: 195657

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195657
Log:
    Backported from mainline
    2013-01-18  Jakub Jelinek  <jakub@redhat.com>

    PR middle-end/56015
    * expr.c (expand_expr_real_2) <case COMPLEX_EXPR>: Handle
    the case where writing real complex part of target modifies
    op1.

    * gfortran.dg/pr56015.f90: New test.

Added:
    branches/gcc-4_7-branch/gcc/testsuite/gfortran.dg/pr56015.f90
Modified:
    branches/gcc-4_7-branch/gcc/ChangeLog
    branches/gcc-4_7-branch/gcc/expr.c
    branches/gcc-4_7-branch/gcc/testsuite/ChangeLog

[Bug middle-end/56015] [4.6 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |4.7.3
            Summary|[4.6/4.7 Regression] expand |[4.6 Regression] expand
                   |expands p[9] = COMPLEX_EXPR |expands p[9] = COMPLEX_EXPR
                   |<-IMAGPART_EXPR <p[9]>,     |<-IMAGPART_EXPR <p[9]>,
                   |REALPART_EXPR <p[9]>>;      |REALPART_EXPR <p[9]>>;
                   |incorrectly.                |incorrectly.

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-02-01 14:32:05 UTC ---
Fixed for 4.7.3+ too.

[Bug middle-end/56015] [4.6 Regression] expand expands p[9] = COMPLEX_EXPR <-IMAGPART_EXPR <p[9]>, REALPART_EXPR <p[9]>>; incorrectly.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56015

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

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

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-04-03 18:20:25 UTC ---
Author: jakub
Date: Wed Apr  3 18:01:51 2013
New Revision: 197448

URL: http://gcc.gnu.org/viewcvs?rev=197448&root=gcc&view=rev
Log:
    Backported from mainline
    2013-01-18  Jakub Jelinek  <jakub@redhat.com>

    PR middle-end/56015
    * expr.c (expand_expr_real_2) <case COMPLEX_EXPR>: Handle
    the case where writing real complex part of target modifies
    op1.

    * gfortran.dg/pr56015.f90: New test.

Added:
    branches/gcc-4_6-branch/gcc/testsuite/gfortran.dg/pr56015.f90
Modified:
    branches/gcc-4_6-branch/gcc/ChangeLog
    branches/gcc-4_6-branch/gcc/expr.c
    branches/gcc-4_6-branch/gcc/testsuite/ChangeLog