The funny thing about GraalVM is that some workloads are actually slower because of some of the dynamic optimisations the JVM applies. Picking the right garbage collector can make a Java program execute faster than when precompiled to native instructions.

I'm sure the same would happen when writing C for the JVM. Tricks like dynamically rearranging/compressing pointers and altering structs to better suit the cache lines and data locality properties of the system code is running on can probably boost any major C program, but that's an incredibly complex change to a C program that the JVM just gives you for free.