Clang Performance Fix

Hi,

Last night I saw your post on ycombinator and I was really surprised to see that Clang took about 12 hours to complete the entire test suite whereas GCC was about 8 hours. I knew something was up because in my experience Clang's virtual function implementation is about 20% faster than GCC's. So I decided to dig in and found out that Clang's uniform_real_distribution was the main cause behind the slowdown. These findings are unrelated to the final keyword being present or not.

Clang had a lot more calls to libm compared to GCC when ran through perf, with the majority of the time spent calculating logl. And suspiciously _GeneralizedRandomGenerator showing up there.

GCC's perf chart:

Looking through Clang's decompiled binary, all of the functions which have a call to RandomGenerator.get_real() were calling logl. Some Googling around and I found some other people experiencing the same thing: https://old.reddit.com/r/cpp/comments/17ite2b/highperformance_alternative_to_stduniform_real/k6x6d6a/

I have no idea how to add a patch/commit to GitLab so I've uploaded it to pastebin if you want to further investigate: https://pastebin.com/ByCmu2J6

Here's the results I got from running book2::final_scene with default settings on a Ryzen 5 7600X running Arch Linux 6.8.7.

Results were generated running: time for i in $(seq 1 10); do ./build/PSRayTracing ; done and using the time from the program.

Clang 17.0.6 unpatched	Clang 17.0.6 patched	GCC 13.2.1 unpatched	GCC 13.2.1 patched
11.776	5.209	6.518	6.361
11.725	5.211	6.509	6.363
11.721	5.216	6.513	6.366
11.775	5.212	6.516	6.367
11.773	5.214	6.566	6.369
11.872	5.211	6.518	6.37
11.723	5.258	6.511	6.364
11.771	5.216	6.513	6.364
11.767	5.212	6.509	6.361
11.72	5.211	6.513	6.367

Clang's performance has doubled and beats GCC's by a solid second, and GCC has obtained a small performance increase of about 150ms.

For reference, here's an image from Clang with my fix applied:

And one from Clang without the patch: