On Tue, Jul 6, 2021 at 8:03 AM Jeff Law via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > > > On 7/5/2021 5:30 PM, Segher Boessenkool wrote: > > Hi! > > > > I ran into this in shrink-wrap.c today. > > > > On Thu, Jun 03, 2021 at 02:54:07PM +0800, liuhongt via Gcc-patches wrote: > >> Use "used" flag for CALL_INSN to indicate it's a fake call. If it's a > >> fake call, it won't have its own function stack. > > Could you document somewhere what a "fake call" *is*? Including what > > that means to RTL, how this is expected to be used, etc.? In rtl.h is > > fine with me, but as it is, no one can know when to use this. What does > > "its own function stack" mean in the description here? You can only put > > FAKE_CALL on functions that do not have a stack frame? But that is > > never true on x86, so that cannot be it, unless there isn't a call > > instruction at all? But then, why use an RTL call insn for this? > > > > Other targets simply do not use an RTL "call" when they want to hide > > such an instruction, why can't you do that here, wouldn't that work much > > better? There are many more insns that you may want to hide. The > > traditional solution is to use unspecs, which very directly hides all > > details. > It reminds me a bit of millicode calls on the PA or calls to special > routines in libgcc. They're calls to functions, but those functions do > not follow the standard ABI. I'd like to remove > INSN_REFERENCES_ARE_DELAYED and instead use the new fake call mechanism, > but I haven't tried it or even looked at the fake call bits enough to > know if that's possible. Fake call is used for TARGET_INSN_CALLEE_ABI which is used for vzeroupper in i386. vzeroupper clobber high part of ymm registers but leave low part unchanged, define it and call_insn with special callee ABI so that RA/CSE knows this instruction kills high parts of ymm registers, and can still optimize with lowpart. I didn't handle FAKE_CALL_P thoroughly in the RTL, but only changed the necessary parts so that I could get my patch to survive the regression test(also fix some optimization issues I observed). n through the tests> > jeff
-- BR, Hongtao
