If the depth is computed in the fragment shader, the fragment can’t be culled before the shader is executed, because the shader can replace the fixed pipeline depth by any value. It would be nice to give a hint that the difference between the new depth computed in the shader and the depth computed by the fixed pipeline lies within a specific range. The typical use is z-correct bumpmapping and the depth is only shifted a small distance. The next polygons behind are far beyond this range and so it should be possibly for the HW to cull even these occluded pixels.