[RF] Use checkpointing for loops in codegen

This reduces gradient JIT time by 25 % for Higgs combination models and
also makes the generated gradient faster.

Author: guitargeek

roofit/codegen/src/CodegenImpl.cxx

@@ -629,6 +629,7 @@ void codegenImpl(RooRealIntegral &arg, CodegenContext &ctx)
       << "   const int n = 1000; // number of sampling points\n"
       << "   double d = " << intVar.getMax(arg.intRange()) << " - " << intVar.getMin(arg.intRange()) << ";\n"
       << "   double eps = d / n;\n"
+      << "   #pragma clad checkpoint loop\n"
       << "   for (int i = 0; i < n; ++i) {\n"
       << "      " << obsName << "[0] = " << intVar.getMin(arg.intRange()) << " + eps * i;\n"
       << "      double tmpA = " << funcName << "(params, " << obsName << ", xlArr);\n"

roofit/roofitcore/inc/RooFit/Detail/MathFuncs.h

@@ -177,6 +177,9 @@ template <typename DoubleArray>
 double constraintSum(DoubleArray comp, unsigned int compSize)
 {
    double sum = 0;
+   #ifdef __CLING__
+   #pragma clad checkpoint loop
+   #endif
    for (unsigned int i = 0; i < compSize; i++) {
       sum -= std::log(comp[i]);
    }
@@ -364,6 +367,9 @@ double flexibleInterp(unsigned int code, ParamsArray params, unsigned int n, Dou
                       double boundary, double nominal, int doCutoff)
 {
    double total = nominal;
+   #ifdef __CLING__
+   #pragma clad checkpoint loop
+   #endif
    for (std::size_t i = 0; i < n; ++i) {
       total += flexibleInterpSingle(code, low[i], high[i], boundary, nominal, params[i], total);
    }

roofit/roofitcore/src/RooFit/CodegenContext.cxx

@@ -186,6 +186,7 @@ std::unique_ptr<CodegenContext::LoopScope> CodegenContext::beginLoop(RooAbsArg c
    }
 
    // Make sure that the name of this variable doesn't clash with other stuff
+   addToCodeBody(in, "#pragma clad checkpoint loop\n");
    addToCodeBody(in, "for(int " + idx + " = 0; " + idx + " < " + std::to_string(numEntries) + "; " + idx + "++) {\n");
 
    return std::make_unique<LoopScope>(*this, std::move(vars));