I was reading "perfect exposure" last night by Hicks and Schultz and it goes over this in great detail. I think understanding the origins of ISO ratings helps. ISO ratings are proportional to 1/exposure required to get an exposure to the first acceptable level of density on the negative, ie have enough energy to get satisfactorily up the toe of the film, ie register anything on the negative. By increasing your personal EI, you just effectively say you are happier with a lower exposure...you will lose shadow detail, but higher levels of exposure on the same pic, will be further out on the density curve and still trigger the silver going.
By changing development time, you are changing the contrast of the negative. So a highlight in a negative (dark part) which is made lighter by reducing exposure, gets compensated for by the extra development. The digital equivalent would be -2 exposure compensation at point of capture, and then +2 compensation in the RAW processing.
An easy way to understand this push process is to look at teh characteristic curves on a datasheet for film:
http://www.fujifilm.co.uk/professional/films/pdfs/neo1600.pdf
if you see the three curves for the first developer (SPD), they show how for increasing development time, you get increased density. So, for example, if you were taking a pic with -2.0 "log10" exposure, and then develop for 2.75 minutes (bottom curve) you get a density of 0.75. Now if you shoot the same scene at -2.5 "log10" exposure (this is like pushing 1.5 stops, as 10^-0.5 is roughly equivalent to 2^-1.5), you will achieve the same density of 0.75 by developing for 6.25 mins.
Now if you are not completely bored by this post...consider the 0.0 exposure point on the curve...the 2.75min development gives density of 1.75. To get that density for a -0.5 underexposure, you now only need a development time of less than 4.25 minutes...you will end up blowing the highlights with your necessary 6.25mins development for the shadows...this is the problem with the curves not being linear. (I dont know what TMAX P3200 looks like, maybe its better. Neopan was on offer for me at 3 quid a roll so I bought it, but I guess these curves mean I am going to end up with contrasty negatives.)
Looking at the -3.5 exposure point on the figure, increasing development time will not change the density for anything. This is why there are limits to how much you can underexpose. If you don't break the energy barrier to put something somewhere into the toe of the film response curve, nothing is going to help you. I've been experiencing this problem recently with TMAX 400, hence my interest in Neopan 1600 at the moment, although I think the ultimate solution is to just open up the lens aperture or use flash.