On Thu, Nov 13, 2014 at 01:14:41PM +0100, Richard Biener wrote: > It might be that RTL alias analysis / CSE give up too early here > (we don't optimize across asm() on the GIMPLE level at all ... heh). > > I didn't look where it gives up (even though appearantly it does).
The cse1 pass doesn't carry the store forward in the "3" case, while it does in the "4" case. So yes it is just a missed optimisation. Segher
