[Bug fortran/19619] New: LAPACK optimisation error

[Bug fortran/19619] New: LAPACK optimisation error

[billingd at gcc dot gnu dot org]

I have an LAPACK testuite failure with 20050124 gfortran on i686-pc-cygwin.

gcd.out has the failure
 CGV drivers: 64 out of   1092 tests failed to pass the threshold
with -O2 but not with -O0.

I have narrowed down the input file (below) to find a case that fails for a 
2x2 matrix.  I will work on isolating the bug and producing a self-contained 
test case.

To reproduce:
 - build LAPACK with -O2
 - ./xeigtstc <  cgd_bug.in > cgd_bug.out

####### cgd_bug.in #####
CGV               Data for the Complex Nonsymmetric Eigenvalue Driver
1                 Number of matrix dimensions
2                 Matrix dimensions  
1 1 1 2 1         Parameters NB, NBMIN, NXOVER, NS, NBCOL
10                Threshold for test ratios
.FALSE.           Put T to test the error exits
0                 Code to interpret the seed
CGV 26            Test all 26 matrix types

#### cgd_bug.out ###

 Tests of the Generalized Nonsymmetric Eigenvalue Problem Driver CGGEV

 LAPACK VERSION 3.0, released June 30, 1999 

 The following parameter values will be used:
    M:         2
    N:         2
    NB:        1
    NBMIN:     1
    NX:        1
    NS:        2
    MAXB:      1

 Relative machine underflow is taken to be    0.117549E-37
 Relative machine overflow  is taken to be    0.340282E+39
 Relative machine precision is taken to be    0.596046E-07

 Routines pass computational tests if test ratio is less than   10.00


 CGV -- Complex Generalized eigenvalue problem driver
 Matrix types (see CDRGEV for details): 
 Special Matrices:                       (J'=transposed Jordan block)
   1=(0,0)  2=(I,0)  3=(0,I)  4=(I,I)  5=(J',J')  6=(diag(J',I), diag(I,J'))
 Diagonal Matrices:  ( D=diag(0,1,2,...) )
   7=(D,I)   9=(large*D, small*I)  11=(large*I, small*D)  13=(large*D, large*I)
   8=(I,D)  10=(small*D, large*I)  12=(small*I, large*D)  14=(small*D, small*I)
  15=(D, reversed D)
 Matrices Rotated by Random Orthogonal Matrices U, V:
  16=Transposed Jordan Blocks             19=geometric alpha, beta=0,1
  17=arithm. alpha&beta                   20=arithmetic alpha, beta=0,1
  18=clustered alpha, beta=0,1            21=random alpha, beta=0,1
 Large & Small Matrices:
  22=(large, small)   23=(small,large)    24=(small,small)    25=(large,large)
  26=random O(1) matrices.

 Tests performed:    
 1 = max | ( b A - a B )'*l | / const.,
 2 = | |VR(i)| - 1 | / ulp,
 3 = max | ( b A - a B )*r | / const.
 4 = | |VL(i)| - 1 | / ulp,
 5 = 0 if W same no matter if r or l computed,
 6 = 0 if l same no matter if l computed,
 7 = 0 if r same no matter if r computed,
 
 Matrix order=    2, type=20, seed=1914,1677,3228,1473, result  5 is 8.389E+06
 Matrix order=    2, type=22, seed=2020,2295,3604,2829, result  5 is 8.389E+06
 Matrix order=    2, type=24, seed=3140, 353,1345, 653, result  5 is 8.389E+06
 Matrix order=    2, type=25, seed=3352,1849,3502,3661, result  5 is 8.389E+06
 Matrix order=    2, type=26, seed=2749,1691, 257,2573, result  5 is 8.389E+06
 CGV drivers:      5 out of    182 tests failed to pass the threshold

 -----------------------------------------------------------------------


 End of tests
 Total time used =         0.00 seconds

-- 
           Summary: LAPACK optimisation error
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: fortran
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: billingd at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-cygwin
  GCC host triplet: i686-pc-cygwin
GCC target triplet: i686-pc-cygwin
OtherBugsDependingO 5900
             nThis:


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[billingd at gcc dot gnu dot org]

------- Additional Comments From billingd at gcc dot gnu dot org  2005-01-25 03:37 -------
The problem is in optimization of routine chgeqz.f

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-25 03:52 -------
(In reply to comment #1)
> The problem is in optimization of routine chgeqz.f

Hmm, could this be like the old PR 1645? (that was reduced from the same routine :) ).

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-25 03:55 -------
(In reply to comment #2)
> (In reply to comment #1)
> > The problem is in optimization of routine chgeqz.f
> 
> Hmm, could this be like the old PR 1645? (that was reduced from the same routine :) ).
Never mind because that testcase is in the testsuite :)

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-25 03:59 -------
Hmm (even though you might not remember this, google does):
http://gcc.gnu.org/ml/gcc-bugs/2001-05/msg00754.html

so does -ffloat-store fix the problem?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-25 04:09 -------
The quotes from that email really makes it sound like this "bug".
"The failures depend on the optimisation level on file chgeqz.f, independent
of the optimisation used on the rest of LAPACK, and start at -O2

I then isolated and analysed a failure on a 2x2 matrix. For this case the
failure is due to a change of order 1.0e-16 in a quantity of order 1.0e-8,
at the level of round off for single precision floating point.  "



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[billingd at gcc dot gnu dot org]

------- Additional Comments From billingd at gcc dot gnu dot org  2005-01-25 05:02 -------
-ffloat-store make one more case fail.

Thanks for the reference.  I remember it now.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[billingd at gcc dot gnu dot org]

------- Additional Comments From billingd at gcc dot gnu dot org  2005-03-02 03:56 -------
Gone away.  Probably fixed by complex division algorithm change.

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619

[Bug fortran/19619] LAPACK optimisation error

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19619