https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102150
--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #4) > Created attachment 60465 [details] > Path for the ifcvt case > > Note this patch does NOT change the definition of trapping but it does fix a > costing issue that could show up. speculatively executing an expensive > instruction via an inline-asm is definitely a shocker for ifcvt; this is > unlike moving invariants out of a loop where the cost might be not as bad. > Note I have not looked into invariant motion either because we might not > want to do full on speculatively execution there either (there is some code > on the gimple level to prevent some speculatively execution motion for LIM > which could be expanded to return MOVE_PRESERVE_EXECUTION for inline-asm > too). So looking into the code of movement_possibility_1, it turns out it already returns MOVE_IMPOSSIBLE because inline-asm is NOT a gimple assign. I need to read the RTL level invariant motion pass now.
