With linear address space in an ARM, you may trade code or data space by storing boolean variables as bits and then potentially having extra code to mask out the bits.

With the multi-memory solution in an 8051, you have to count references or experiment to decide which variables that should be pushed to XDATA.

I often start the programming using a big brother with way more memory and pins. I can use a prototype board and extract the required processor signals for interfacing with some preliminary electronics. This will quickly let me know how the requirements I will have on amount of code and data space and any timing issues (suggested clock frequency).

When the initial prototype PCBs are ordered, I already know quite well that I have a bit of safety margin. Either with the selected chip, or that there exists a larger chip that can be used (pin-compatible and with a price I can live with if the chip is needed). The existence of a suitably large safety margin means that even if some product manager/customer somewhere have f..ed up about their requirements, there will be reasonable changes to save the day. I never want to start a new design with the biggest of it's class unless there are very, very strong reasons to do so, and there are strong probabilities that the chip/programmer combination will manage the task.

Normally, you design a product with some form of emergency backup plan. Can an older chip with higher current consumption be used if in panic? Can the board quickly be redesigned with a chip with more pins? Can external memory be added? Can the processor be run at a higher frequency? Can any data be precomputed? Can any precomputed data be replaced with real-time computations? Can an external EEPROM be used? Is it feasible with a slave processor?