As author of that benchmark, I want to say that the differences I observed between FPC and LLVM were pretty modest for x86-64. Therefore, I only report the values for FPC and FPC-LLVM for the ARM-based aarch64. This might reflect the fact that FPC is not yet mature on aarch64, with releases focusing on compatibility and not performance. Indeed, since I wrote that article, J. Gareth Moreton has submitted a lot of aarch64 optimizations to the FPC trunk. I encourage anyone who wants to see FPC's aarch64 improved to consider sponsoring his terrific work:
https://www.patreon.com/curiouskitIn general, I find the performance of FPC more than sufficient for my needs, and for the computationally expensive and common routines I have tried to use the GPU (GLSL, Metal) or SIMD (SSE, Neon) for acceleration. Therefore, my most critical routines tend not to evaluate the code emitted by FPC vs FPC-LLVM and I do not have benchmarks that are likely to generalize.
For x86-64 code, I think Jonas provided realistic expectations for typical benefits, e.g. around 18% if you enable link-time optimization. However, as he notes it really depends on the algorithm.
https://wiki.freepascal.org/LLVM