On 10/18/17 13:30, Segher Boessenkool wrote: > On Wed, Oct 18, 2017 at 11:10:48AM +0000, Bernd Edlinger wrote: >> A memory clobber would also make rtx_moveable_p return false, >> thru the following case: >> >> case MEM: >> if (type == OP_IN && MEM_READONLY_P (x)) >> return rtx_moveable_p (&XEXP (x, 0), OP_IN); >> return false; >> >> ... >> >> case CLOBBER: >> return rtx_moveable_p (&SET_DEST (x), OP_OUT); >> >> >> because that memory clobber is in a parallel statement >> together with the ASM_OUTPUT. >> >> Right? > > Yes, that looks correct. And a memory clobber of course _should_ make > the insn non-moveable as well -- it's an unknown side effect all by > itself :-) > >
Yeah, that's probably not what they wanted to hear, but supposed this is the only place where the volatility of ASM_OUTPUT is ignored, then adding a memory clobber to the asm stmt is the way to go if they want to play safe, especially when implementing spin-locks. Bernd.
