On 1.7.2011 16:24, Laurent Desnogues wrote: > On Fri, Jul 1, 2011 at 4:11 PM, Jakub Jermar <ja...@jermar.eu> wrote: > [...] >> Actually, the testcase can be further reduced into: >> >> .global _start >> >> .text >> >> .space 0x20 >> >> _start: >> set 110393, %i1 >> set 0x40, %i2 >> >> cmp %i1, %i2 >> udivx %g0, 1, %g0 >> movgu %xcc, %i2, %i1 >> cmp %i1, 512 >> bgu %xcc, 0f >> nop >> >> succ: >> ta 0 >> >> fail: >> 0: >> ta 1 >> >> The presence of the `udivx` instruction seems to be essential. Even >> though it has no effect on the computation, removing it will make the >> testcase non-reproducible. > > Could you try to replace udivx with sdivx? It looks wrong too.
Yeah, `sdivx` behaves the same wrt. the testcase. Jakub
