The histogram statistics printing utility
HistoStats.cpp in GaudiUtils defines the follow default 'bad' value
const double s_bad = -1 * std::numeric_limits<double>::max() ;
which is used in the various methods when the calculations cannot be performed for whatever reason.
The problem with this is it leads to floating point exceptions when subsequent calculations are performed. For example in
% ( 100 * HistoStats::overflowEntriesFracErr ( histo ) ) // 22) error on 21
This generates an overflow FPE when
s_bad, as this value cannot be multiplied by 100.
This MR proposes the follow fix, which is to change the 'bad' value to something smaller than the minimum possible value a double can represent
const double s_bad = std::numeric_limits<float>::lowest() ;
-3.40282e+38 instead of
-1.79769e+308 ... My assertion is, as far as satisfying the requirement of a 'bad' value, either is good enough, and by using the float value subsequent calculations performed in double precision are safe.
Of course, it does not have to be
std::numeric_limits<float>::lowest(). Any large negative value,
-1e100, whatever serves the same purpose...
I have also made a number of other small cleanups, to prevent type conversions (to double) until absolutely required, to allow some checks to be performed on the original native types (sometimes int).