Work around issue with truncated Float32 representation of lwork in

sgesdd by using nextfloat. Now with comments and test so this commit
supersedes ad59ceb

See

http://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=13&t=4587&p=11036&hilit=sgesdd#p11036

and

scipy/scipy#5401

Fixes #15784

base/linalg/lapack.jl

@@ -1572,7 +1572,12 @@ for (geev, gesvd, gesdd, ggsvd, elty, relty) in
                 end
                 chklapackerror(info[])
                 if i == 1
-                    lwork = BlasInt(real(work[1]))
+                    # Work around issue with truncated Float32 representation of lwork in
+                    # sgesdd by using nextfloat. See
+                    # http://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=13&t=4587&p=11036&hilit=sgesdd#p11036
+                    # and
+                    # https://github.com/scipy/scipy/issues/5401
+                    lwork = round(BlasInt, nextfloat(real(work[1])))
                     work = Array($elty, lwork)
                 end
             end

test/perf/lapack/perf.jl

@@ -4,3 +4,10 @@ include("../perfutil.jl")
 
 include("eig.jl")
 include("factorizations.jl")
+
+# The test below is not a perf test but it takes too long for the normal test suite.
+# This tests a work around for a bug in LAPACK where the work space is truncated
+# to a too small number because the number is too big to be represented as a Float32.
+# See the calculation for lwork in our sgesdd wrapper in lapack.jl.
+println("testing work space bug in sgesdd")
+svd(rand(Float32, 9537, 9537))