The TI's DSP-processors support 64bit native float, but 40bit float is enough for control systems. So Sharc is good, but support of 40bit float is bad in C++ of VDSP++. It's to do the user, but not compiler.
I understand that for supporting 40bit float I must turn on RND32 and place my variables in 40-bit internal memory and save them to external memory the special method. But I want that it's native.