Some memories need to be explicitly enabled.
In terms of initializing content, the GNU/GCC compiler / linker expect you to manage the allocation, and staging on content so that the code in startup.s can either zero, or copy content out of staging areas in FLASH. You have to determine the beginning/end of each section with symbols, and also the place in FLASH it gets placed.
IAR and KEIL deal with this better, the linker creates tables per the scatter file, and the scatter loader processes that automatically. Arduino's GNU implementation is also better, and similarly table driven.
Code you have in SystemInit() is also expected to bring up clocks, memory interfaces, and pins so the run time initialization code in startup.s can fully do it's job, and the memory is viable when the C++ constructors are called.
SDRAM is typically the slowest, longest latency memory available. Generally not good for code or stacks, but the caching can hide some of the ugliness. Internal FLASH is also rather slow, but has very wide fetch lines, and prefetch / branch-prediction can hide that, via ST's ART
Watch most of the _sbrk implementation's ST has been using assume the stack/heap reside in the same memory, and the SP is used as a ceiling for the heap. Better implementation might use symbols from the linker.
