XPOL, the first correlation polarimeter at a large millimeter telescope, uses a flexible digital correlator to measure all four Stokes parameters simultaneously, i.e. the total power I, the linear polarization components Q and U, and the circular polarization V. The versatility of the backend provides adequate bandwidth for efficient continuum observations as well as sufficient spectral resolution (40 kHz) for observations of narrow lines. We demonstrate that the polarimetry specific calibrations are handled with sufficient precision, in particular the relative phase between the Observatory's two orthogonally linearly polarized receivers. The many facets of instrumental polarization are studied at 3mm wavelength in all Stokes parameters: on-axis with point sources and off-axis with beam maps. Stokes Q which is measured as the power difference between the receivers is affected by instrumental polarization at the 1.5% level. Stokes U and V which are measured as cross correlations are very little affected (maximum sidelobes 0.6% (U) and 0.3% (V)). These levels critically depend on the precision of the receiver alignment. They reach these minimum levels set by small ellipticities of the feed horns when alignment is optimum (<~ 0.3"). A second critical prerequisite for low polarization sidelobes turned out to be the correct orientation of the polarization splitter grid. Its cross polarization properties are modeled in detail. XPOL observations are therefore limited only by receiver noise in Stokes U and V even for extended sources. Systematic effects set in at the 1.5% level in observations of Stokes Q. With proper precautions, this limitation can be overcome for point sources. Stokes Q observations of extended sources are the most difficult with XPOL.