Permutations IP and PC-1 are near-transpositions, and play no cryptographic role. IP-1 is simply IP reversed. The best theory about why they are here is: as a technical by-product of the 8-bit interface used by early DES ICs, translated into the same formalism as the rest when writing the DES standard. They make wiring of hardware implementations simple when using eight 8-bit shift registers for data input and output, and eight 7-bit shift registers for key input (the later being re-grouped as two 28-bit shift registers after key loading).
Expansion E (from 32 to 48 bits, before the XOR with keys bits leading to S-boxes inputs) is very regular; it is here to compensate the S-boxes going from 6 to 4 bits. That geometry of S-boxes is itself a compromise between security and silicon area.
Permutation P (of 32 bits, at the output of S-boxes) is quite irregular, plays a central role in the diffusion from round to round, and was carefully optimized for that purpose (the combination of E, P, and the S-tables makes DES quite resistant to differential cryptanalysis). Notice that the additional silicon cost of such arbitrary permutation is low (especially when compared to alternatives giving some diffusion, and considering 1970's technology). In software, E has a sizable performance cost, especially on less than 32-bit CPUs: typical implementations use S-tables with 32-bit wide output, requiring either more memory than would otherwise be needed, or a masking operation per S-box, with 8 of these per round.
Permutation PC-2 (between key shift registers and the XOR before the S-boxes) plays a role in making the 16 DES subkeys "good", that is without too much exploitable correlation from one round to another. It is such that the 8 bits (out of the 56 actually used key bits) that are not used in a round subkey, are used in the subkey for the next round. Other than that, it is rather arbitrary, as far as I know.
I do not buy the theory that the permutations in DES have been introduced for the purpose of making software implementations slow: IP, IP-1, and PC-1 represent a small fraction of a software DES computation time, since they are outside of the round loop; and are not arbitrary, so much that some software implementations of these three tables take advantage of the fact that they are near-transpositions. Other bit massaging is either as simple as can be (E), or with a cryptographic role: a major one for P; a lesser one for PC-2, but the later introduces very low software overhead for bulk encryption. Further, I know no historical evidence supporting that theory; when evidence has already surfaced (see end of this answer) supporting that the 56-bit key size of DES was deliberately introduced to weaken it, and result from a compromise between the designers willing to have 64 or 128 bits (as in some Lucifer variants); and US authorities pushing for 48 bits.