Not an issue but here my results. A bit more consistent as opposed to what you mentioned in the talk.

Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz (governor performance)
Linux 5.1.11-300.fc30.x86_64
gcc 9.1.1

P.S. You should add the link to your talk to the Readme, and maybe instruction on how to build.
This worked for me (even without using clion):

git clone https://github.com/mattgodbolt/pt-three-ways.git
cd pt-three-ways
mkdir cmake-build-release
cd cmake-build-release
cmake .. -DCMAKE_BUILD_TYPE=Release
cd ..
scripts/bench-6t-cornell.sh

EDIT: suzanne st and 6t values were. Maybe we should use same sampling?

I very much liked the talk.
I would be really interested in how these three techniques would compare performance-wise, if a bounding volume hierarchy, like a k-d tree were to be introduced.
My expectation is that all three versions will speed up dramatically; however I wonder which one of the three is able to benefit from it the most.

When building with clang, the Unpredictable::any workaround will still cause multiple branches, playing havoc with the branch predictor.
Clang has a way to define a branch as being unpredictable and using that yields much better code and on the suzanne scene, my render time dropped from 81s -> 66s
The hacked way i did this was

#if __has_builtin(__builtin_unpredictable)
#define UNPREDICTABLE(x) __builtin_unpredictable(x)
#else
#define UNPREDICTABLE(x) x
#endif
    if (UNPREDICTABLE(u < 0.0) || UNPREDICTABLE(u > 1.0)
        || UNPREDICTABLE(v < 0.0) || UNPREDICTABLE((u + v) > 1))
      continue;

For the case when that isn't available, I tried using this(which works for gcc too), but clang still emitted branches here

#elif __has_builtin(__builtin_expect_with_probability)
#define UNPREDICTABLE(x) __builtin_expect_with_probability(x, 1, 0.5)