Hi Nick, On Wed, Mar 18, 2020 at 08:56:11PM -0400, Nicholas Krause wrote: > I've not sure if I've misunderstanding something in the combine code but > in make_more_copies > for combine.c this seems very odd: > if (!(REG_P (dest) && !HARD_REGISTER_P (dest))) > continue; > > rtx src = SET_SRC (set); > if (!(REG_P (src) && HARD_REGISTER_P (src))) > continue; > > Is there any good reason we are assuming the destination can't both be a > hard register or a regular code here?
The destionation should be a pseudo-register, and the source should be a hard register. If those are true, *then* we make a copy to a new intermediate pseudo. We do not want combine to move the hard register into other instructions (that is the register allocator's job, and it does a much better job of it, combine just does the greedy first-fit solution). But, it turns out that combining a register move with another instruction often is beneficial (in effect just replacing that other instruction with a better one). make_more_copies makes a fresh new register copy to keep the status quo for that (other parts of combine now disallow combining a move from a hard register into other insns). Segher
