https://bugs.freedesktop.org/show_bug.cgi?id=83570
--- Comment #9 from Roland Scheidegger <srol...@vmware.com> --- (In reply to comment #8) > for the llvm version I was going to return bld->undef. I figure I might as > well change tgsi_exec.c version to 0xffffffff for consistency...but I don't > have strong feelings either way. I'm not sure how your code is going to look like, but I can't really see a way to return bld->undef since you will be required to replace the input vector with something sensible (similar to udiv_emit_cpu()) (*). That said, I guess using the same code as udiv_emit_cpu() but omitting the final "or" would also work - in which case the result for the idiv for x / 0 would be x / -1, why not... I'm not really sure it makes sense to return 0xffffffff just for consistency, for uint div this value sort of makes sense, but for signed div it does not - MAX_INT or MIN_INT probably would make more sense. But since d3d10 doesn't support it, and glsl doesn't care, I think whatever is cheapest is ok (another cheap option would be to use a "andnot" instead of the "or" udiv does, in which case you'd get zero - MIN_INT/MAX_INT are going to be slightly more complicated). (*) Because sse doesn't have int div, llvm actually breaks down the vector div to ordinary multiple int divs anyway (and yes it's bound to be very slow), thus breaking it down manually and do per element selection wouldn't really be too bad. Still wouldn't really make sense imho however. -- You are receiving this mail because: You are the assignee for the bug.
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
