On Thu, Mar 16, 2023 at 10:04 AM HAO CHEN GUI <guih...@linux.ibm.com> wrote: > > Hi Richard, > > 在 2023/3/16 15:57, Richard Biener 写道: > > So this is one way around the lack of CSE/PRE of constant operands. I'd > > argue that a better spot for this _might_ be LRA (split the constant out if > > there's a free register available), postreload-[g]cse (CSE the constants) > > and > > then maybe cprop_hardreg to combine back single-use constants? > > > > I'm not sure if careful constraints massaging like adding magic letters to > > alternatives with constants to pessimize them for LRA, making them > > more expensive than spilling the constant to a register but avoid > > secondary reloads with spilling a register to the stack to make room > > for the constant, is possible - but in theory a special constraint modifier > > for this purpose could be invented. > > Thanks so much for your advice. > > cse/gcse doesn't take cost of constant set (the def insn of the constant) into > consideration. So it won't replace the register with a constant as it costs 1 > insn with the register and costs 2 insn with the constant.
I think it does (and should) cost the constant set (IIRC we had some improvements there, or at least proposed, during this stage1). But sure - this is why your "trick" works. > Finally, the single- > use constants can't be back to 2 insn. And that's because of the issue you point out above? > Not sure if I understand it correctly. > Looking forward to your advice. My main point is that CSEing constants has impacts on register pressure and thus should probably be done after or within register allocation. RTL expansion itself is probably a bad time to pro-actively split out constants even more if, as you say, nothing puts them back. Richard. > Thanks > Gui Haochen
