Hello Ian. But is there a chance that between a rtl generation and splitting an insn, a compiler could use that clobbered CC reg in some of intermediate phases, but after a register allocation it turns out that there shouldn't be any clobber at all? Thanks. Dmitry. --- Исходное сообщение --- От кого: "Ian Lance Taylor" <iant@xxxxxxxxxx> Кому: "Dmitry" <mittie@xxxxxxx> Дата: 28 апреля 2011, 17:15:20 Тема: Re: Back end question. > "Dmitry" <mittie@xxxxxxx> writes: > > > I'm currently writing a GCC back end for a custom architecture. And I > > have a RTL related question. If I am, for example, writing an > > implementation for "movhi" insn, and depending on a register class of > > the destination operand it could include a condition code register > > change. How can I write a right implementation for such case? If I use > > "define_expand" RTX destination registers are not allocated yet so > > they could be any pseudo registers. So I can not determine the class > > of that registers during RTL generation phase. Should I use > > "define_split" additionally to analyze operand's register class after > > reload phase? Or maybe there's other way? > > I think you'll have to clobber the conditional code register in the > initial insn, and use a define_split to remove the clobber (or perhaps > change it to a set when appropriate) after register allocation is > complete. > > Ian > >
