If you reference phase to the beginning of your FFT aperture, then any sinusoid that is not exactly integer periodic within that window width will be discontinuous across the two window edges, thus providing a phase (phase of a discontinuity?) that doesn't really make much intuitive sense at that measurement point, as well as alternating in phase between successive FFT result bins. Also, most common "hump" shaped windows reduce the signal to nothing, or nearly so, at the window edges, which again doesn't make much sense, as the measurement of the phase of almost nothing at the reference point can be quite noisy.
To prevent this seeming nonsense, do an FFTshift, which moves the phase reference to the center of your signal window. You can so this by rotating the data (and window) so that the peak of the window hump is at the FFT's phase measurement point (time index 0), and the discontinuity and/or window-edge-attenuation is far from the phase reference measurement point. This FFTshift or data rotation will also prevent the phase of a non-integer-periodic sinusoid from alternating between FFT result bins, thus allowing simpler phase interpolation between FFT result bins.