On Tue, Jan 03, 2017 at 01:43:01PM +0000, Richard Sandiford wrote: > An alternative would be to add a new macro to control this block in > general_operand: > > #ifdef INSN_SCHEDULING > /* On machines that have insn scheduling, we want all memory > reference to be explicit, so outlaw paradoxical SUBREGs. > However, we must allow them after reload so that they can > get cleaned up by cleanup_subreg_operands. */ > if (!reload_completed && MEM_P (sub) > && GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (sub))) > return 0; > #endif > > The default would still be INSN_SCHEDULING, but AVR could override it > to 1 and reject the same subregs.
Or you can define INSN_SCHEDULING (by defining a trivial automaton for your port: a define_automaton, a define_cpu_unit for that automaton, and a define_insn_reservation for that unit). It would be nice if we could separate "no subregs of memory" from "has instruction scheduling". Or ideally subregs of memory would just go away completely. > That would still be a hack, but at least it would be taking things in > a good direction. Having different rules depending on whether targets > define a scheduler is just a horrible wart that no-one's ever had chance > to fix. If using the code above works well on AVR then that'd be a big > step towards making the code unconditional. Ugh, it sounds like I am volunteering. Oh well. > It'd definitely be worth checking how it affects code quality though. It shouldn't be too bad, if ports that do have instruction scheduling can live without it. > (Although the same goes for the current patch, since C_C_M_C is a pretty > big hammer.) Yeah, and the current patch disallows much more than is needed here, even. Segher
