__ockl_lane_u32 should not be moved out of divergent loop

[ssahasra]

This is a copy of the internal issue SWDEV-167691.

In the attached file, the loop condition depends on __ockl_lane_u32() which itself expands to mbcnt_lo and mbcnt_hi. These intrinsics are marked as sources of divergence, but they get moved out of the loop, thereby incorrectly providing a constant value to the loop condition.

waterfall_converged.cpp.txt waterfall_converged-gfx900.isa.txt