I don't know about MIPS, but ARM & SH compilers indeed inline the function if count param for memcpy is constant and small enough (and intrinsic option is enabled).
"Small enough" is like this:
For ARM: count <= 8
For SH: count < 16
But generated code copies data byte by byte (ldrb/strb or mov.b), i.e. compiler makes no assumptions about src/dest alignment.
SC