A branch that's almost always taken

fastPath and slowPath are both moderately large, and in the real production workload the condition is true about 99.9% of the time, which nothing in the source reveals. Compare a plain -O2 build with a PGO build (profile collected from representative traffic, then fed back into a second compilation). What changes?

main.cpp
int fastPath(int x);  // large-ish function bodies, cost borderline for inlining
int slowPath(int x);

int dispatch(int x, bool condition) {
    if (condition) {
        return fastPath(x);
    } else {
        return slowPath(x);
    }
}

25 of 44 in Compilers