Default (conservative) behaviour of FENCE.I is to flush the pipeline, all uncoherent instruction caches and anything else related.
fence.i also didn't allocate whole 22.125 bits of opcode just to waste it.
IMPORT.I - is just an isb like pipeline flush that is already subset of fence.i. Why can't we use those 12bits of imm to somehow encode fence.ipipe instruction? In this way it degrades to regular fence.i on "legacy" cores.
There is also an likely need for fence.ipipe instruction for different purposes than importing new instructions, so fence.ipipe and import.i probably should be encoded with a different bit pattern as the latter might not need to perform whole pipeline flush.
about id-consistency-proposal.pdf v0.9.1
Default (conservative) behaviour of FENCE.I is to flush the pipeline, all uncoherent instruction caches and anything else related.
fence.ialso didn't allocate whole 22.125 bits of opcode just to waste it.IMPORT.I- is just anisblike pipeline flush that is already subset offence.i. Why can't we use those 12bits of imm to somehow encodefence.ipipeinstruction? In this way it degrades to regularfence.ion "legacy" cores.There is also an likely need for
fence.ipipeinstruction for different purposes than importing new instructions, sofence.ipipeandimport.iprobably should be encoded with a different bit pattern as the latter might not need to perform whole pipeline flush.