This removes the indirect call to svm_eval_nodes, which is a significantly taxing function for the compiler, from the kernel_direct_lighting kernel, and speeds up kernel build times by 4-5 seconds. Render time on AMD GPU is slower by ~1%, CPU render times seem unaffected.
Check the split-kernel-faster-building branch for individual commits.
The branch also has kernel_indirect_background split in the same manner, with a 2-3 second improvement in build times on top of this patch. It should be possible to continue with this for a few other kernels, but things become problematic for kernels containing functions that call svm_eval_nodes in a loop. I'm not quite sure what we could do about that just yet.