I tend to apply different "rules" to different kinds of device.
For example, an op-amp is a well known device with a universally recognised symbol, so that's what I use, regardless of the physical pin-out of the chip. I'll split the symbol into parts, so a dual op-amp will have three parts to the symbol - two individual amplifiers, and a third symbol which has the power and ground pins.
Power supply components (eg. SMPS controllers) are drawn with whatever pin arrangement makes the schematic clear and easy to read. Usually this means drawing them something like the example schematics in the data sheet.
Microcontrollers, on the other hand, I'll draw with the pins in their physical order, again with power pins separated out to another symbol.
FPGAs are split up into a number of symbols, each of which corresponds to one I/O bank. Since each bank can be powered from a different I/O voltage, splitting up the symbol this way makes it easier to avoid mistakes in designs using multiple I/O voltage levels.
Complex microprocessors are split into multiple symbols, where each symbol corresponds to a different function. So, for example, the DRAM interface will be on one symbol, external Flash interface on another, GPIOs on another, and so on. Chances are the device is a BGA anyway, so placing pins per their physical position would be nonsense.
As an aside: I once found myself working for a company which had a policy on schematics: inputs on the left, outputs on the right, always. So, if a processor had pins for a crystal oscillator, XIN would go on the left, XOUT on the right, and the resulting ugly circuit was unavoidable because symbols would be drawn by 'trusted' component librarians rather than the engineers themselves. I protested, and won a small bonus for my efforts to improve company procedure - but no actual change. To my knowledge, a well-known network equipment manufacturer's Pierce oscillators still look like crap on their schematics to this day.