https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90273
--- Comment #27 from rguenther at suse dot de <rguenther at suse dot de> --- On April 30, 2019 4:27:25 PM GMT+02:00, "aoliva at gcc dot gnu.org" <gcc-bugzi...@gcc.gnu.org> wrote: >https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90273 > >--- Comment #26 from Alexandre Oliva <aoliva at gcc dot gnu.org> --- >I saw the #c11 patch in gcc-patches, and it seemed to have been posted >FTR and >installed. It looked good, so I didn't comment on it. > >I agree about the effects of #c16, though I begin to get a feeling that >it's >working too hard for too little benefit. Ditto trying to optimize >debug temps: >you will get some savings, sure, but how much benefit for such global >analyses? > >Perhaps we'd get a much bigger bang for the buck introducing vector >resets, in >which a single gimple bind stmt would reset several decls at once. If >that's >become as common as it is being made out to be, this could save a >significant >amount of memory. > >Though from Jan's comments on compile times, it doesn't look like we've >got >much slower, which makes me wonder what the new problem really is... >Could it >be that debug binds have always been there, plentiful but under the >radar, and >that the real recent regression (assuming there really is one) lies >elsewhere? >(sorry, I haven't really dug into it myself) The recent regression is we no longer throw them away plentiful during CFG cleanup and now they pile up during inlining. I agree full DCE with liveness will be expensive for usually little gain. Not sure if vector resets will improve things much.