On 01/27/2017 02:19 AM, Segher Boessenkool wrote:
But what is "insn cost"? Latency is no good at all -- we *want* insns with higher latency to be earlier. fsqrt is not pipelined, and that is what makes it so costly. (This isn't modeled in the scheduling description btw: that would make the automata sizes explode, the usual problem).
On other machines sqrt might be pipelined, so the patch as-is clearly isn't suitable. rtx_cost/insn_cost probably also won't do, since it doesn't model that property either.
Maybe instead of a hook you could have an insn attribute that the scheduler looks for.
Bernd