One other thing to check: Go into the Linker options under the Project Options, in the 'General' section there is an option to "Individually map functions and data items".
Without this set, the Linker would try to find space to map ALL data of the same section name to a single memory block. So if you have an 8K input, 8K output and 8K temp buffer for your FFT, there needs to be 24K of 32-bit words contiguous space to map this.
If you enable this option, though, the Linker can split these buffers across different memory sections (e.g. maybe it finds space for two of them in seg_dmda, and sticks the 3rd in seg_pmda).
That might help you find space in internal memory for all your buffers...
Regards,
Craig.