On Sat, Dec 22, 2012 at 01:44:49PM +0200, Andriy Gapon wrote: > on 22/12/2012 13:21 Konstantin Belousov said the following: > > This is due to the vtoslab() returning NULL. Since slabref is dereferenced > > later, clang tries to be helpful as usual and converts the !(p->flags & > > PG_SLAB) case from vtoslab() into the jump to un2 instruction if vtoslab() > > result is NULL. > > > > So instead of KASSERT triggering the next line, you see this improvement. > > Interesting. Thank you for the explanation. > > But looking at the code I think that slabref->us_keg access _before_ KASSERT > is the culprit? I.e. even with GCC we could get a page fault before the > KASSERT is reached (modulo reordering)?
May be, but I do not think it is matter. Because KASSERT() now can return, even if you reorder the assert and deref, I think that compiler authors still find an excuse.
pgp8ueyRkyG3w.pgp
Description: PGP signature
