Even with today's ludicrously bloated kernels, I'm skeptical about the idea that speculative execution _in the kernel_ seriously helps computer performance. Has anyone measured overall slowdown from recompiling kernel branches to use LFENCE with new (fully serializing) microcode?
-
Particularly bad - speculative execution in context of accessing device control registers, reading reset-on-read status registers, etc...
This is why we have MTRRs (and PATs) to disable speculative reads for MMIO. Otherwise of course everything would break. This has been a thing since, well, speculative reads have been a thing.
