Regression with Clad `master` in case of num-diff fallback in specific conditions about clad HOT 6 CLOSED

While the reproducer in the original post works with the fix by @PetroZarytskyi, the underlying issue doesn't seem to really fixed because adding another trivial if statement after defining the constant breaks Clad again:

double wrapper(double *params)
{
   if (true) {
      double constant = 11.;
      if (true) { // NEW
         return foo(params[0], constant, 1.);
      } // NEW
   }

   return 0.0;
}

from clad.

@guitargeek, can you provide us with the values of params and the expected result of wrapper?

from clad.

You can just take the code from the initial post in this issue, and replace the wrapper there with the new one. The reproducer provides the params values and computes the expected gradient numerically

from clad.

Do we care what foo does?

from clad.

Mathematically no, the only requirement to trigger the problem is that de definition of foo is not known to Clad. That's why I put it into a separate shared library in the reproducer

from clad.

Thanks for fixing things up to this point! The current failures with RooFit can be reduced to this function, which can be plugged in to the full reproducer at the beginning of the issue:

double wrapper(double *params)
{
   double obs[]{
      110,
   };

   double result = 0.0;
   double hist[1] = {50.};
   double tmp6[] = {params[0], 1.};

   for (int j = 0; j < 1; j++) {
      double signalshapes = hist[j];
      double background1shapes = hist[j];
      double tmp0[] = {signalshapes, background1shapes};
      double tmp9 = 0.;
      for (int i = 0; i < 2; i++) {
         tmp9 += tmp0[i] * tmp6[i];
      }
      result += std::log(tmp9) - std::lgamma(obs[j] + 1.);
   }

   return result;
}

It seems to be related to double-loops again

from clad.