It might be necessary to move data back and forth between a general purpose register (EAX for instance) to an MMX register. This can be done by:
movd eax,mm0
movd mm0,eax
pinsrw mm0,eax,0
pextrw eax,mm0,0
movmskb eax,mm0
These instructions should however be avoided as much as possible since they are executing very slow. It seems like there are severe penalties when moving data between the MMX registers and the general purpose registers. Most of these instructions take more than 16 cycles to execute, which is more than 8 times the cyclecount of normal MMX instructions.
In many cases it is actually faster to write the data to a temporary memory location, using movd, and move the data back into the registry. For much data it is actually often faster to process the data one line at the time, and store all data in a temporary space, and process the data in a separate loop.
phaeron writes:
Note that the Microsoft Visual C++ compiler will give you trouble here if you are using MMX intrinsics, as it is unable to directly generate movd mmreg,mem and movd mem,mmreg instructions for the _m_to_int() and _m_from_int() intrinsics. Instead, the values trampoline off of the general purpose registers, triggering the undesirable performance behavior described above. This is true even with the Visual Studio .NET 2003 compiler (VC7.1), which fixed most of the MMX code generation embarrassments from the VC6PP and VC7 compilers, most notably MOVQ hell. As such, you're still better off sticking with inline or explicit assembly than using MMX intrinsics with MSVC.
