On Wed, Nov 16, 2011 at 6:54 AM, Olivier Hainque <hain...@adacore.com> wrote: > > On Nov 9, 2011, at 18:15 , Olivier Hainque wrote: >> I'm not convinced that the potential gain is worth the extra >> complexity and potential risk of running into another subtle >> subcase, with hard to track sporadic runtime failures for >> starters. I don't have numbers though. >> >> That's a port maintainer call, I guess ? > > David, opinion on this point ? > > My understanding is that we have two options > > 1) try to preserve the current attempt at maximizing > optimization opportunities with precise stack/frame > tie insns, > > 2) simplify for a slightly more brutal option, with > a strong (mem:blk scratch) barrier instead > > My feeling is that 2 would be a sensible option. > > Trying to get the precise insns right has caused multiple > issues (several PRs about hard to track silent wrong code > generated, trickier implementation), and I'm not convinced > that the legitimate code efficiency gains are worth the > trouble. > > As I wrote, I don't have numbers to backup the latter point > though.
We can try (2), but we will have to benchmark it to determine the impact on performance. Thanks, David
