Revert PRs 4515 and 4520 (restore second, dsecnd)

[frjohnst]

It turns out that PRs 4515 and 4520 break xlf builds of tests under
lapack-netlib/TESTING which require SECOND and DSECND. IBM
has decided this is a bigger problem than the conflict
between lapack second_ and the xlf run time.

[martin-frbg]

Strange - unless I'm missing something, your original PRs were conditional on the compiler being xlf, and the value returned by second or dsecnd is only ever used to provide performance data that does not impact the validity of test results.

[frjohnst]

A number of tests under lapack-netlib/TESTING/LIN and EIG call SECOND and DSECND and won't build
without these routines. PRs 4515 and 4520 removed those routines from builds with xlf compiler, as was mentioned,
but IBM decided it is important to be able to build these tests, as downloaded from the OpenBLAS repository,
with the xlf compiler and wants to restore this capability. Thanks.

grep -i '= *second' /.f /.F
LIN/cchkrfp.f: S1 = SECOND( )
LIN/cchkrfp.f: S2 = SECOND( )
LIN/schkrfp.f: S1 = SECOND( )
LIN/schkrfp.f: S2 = SECOND( )
EIG/cchkee.F: S1 = SECOND( )
EIG/cchkee.F: S2 = SECOND( )
EIG/schkee.F: S1 = SECOND( )
EIG/schkee.F: S2 = SECOND( )
LIN/cchkaa.F: S1 = SECOND( )
LIN/cchkaa.F: S2 = SECOND( )
LIN/schkaa.F: S1 = SECOND( )
LIN/schkaa.F: S2 = SECOND( )

grep -i '= *dsecnd' /.f /.F
LIN/dchkab.f: S1 = DSECND( )
LIN/dchkab.f: S2 = DSECND( )
LIN/dchkrfp.f: S1 = DSECND( )
LIN/dchkrfp.f: S2 = DSECND( )
LIN/zchkab.f: S1 = DSECND( )
LIN/zchkab.f: S2 = DSECND( )
LIN/zchkrfp.f: S1 = DSECND( )
LIN/zchkrfp.f: S2 = DSECND( )
EIG/dchkee.F: S1 = DSECND( )
EIG/dchkee.F: S2 = DSECND( )
EIG/zchkee.F: S1 = DSECND( )
EIG/zchkee.F: S2 = DSECND( )
LIN/dchkaa.F: S1 = DSECND( )
LIN/dchkaa.F: S2 = DSECND( )
LIN/zchkaa.F: S1 = DSECND( )
LIN/zchkaa.F: S2 = DSECND( )

[frjohnst]

I should have said that PRs 4515 and 4520 break xlf builds of some tests under lapack-netlib/TESTING ,
not the function of the tests themselves. (Just tried to update that.)

[martin-frbg]

Right, but my understanding was that these routines were conflicting with equivalent implementations provided by xlf ? I have no problem with reverting the changes, I am only wondering if there might be a better solution for this issue (that ultimately affects Reference-LAPACK as well, from where OpenBLAS' lapack-netlib tree is copied)

[frjohnst]

I understand your point, but we have had several discussions within IBM, and it
was decided that not being able to build the tests as downloaded from OpenBLAS
with xlf was a bigger problem than the conflict between lapack's second_ and the
xlf run time. We will investigate whether a better solution might be possible in the
future, but in the short term we need to be able to easily build these tests with xlf.

So in the short term we would simply like to revert the changes in PRs 4515 and 4520. Thanks.