In the past decade, several tools have been developed to automate the floating-point to fixed-point conversion for DSP systems. In the conversion process, a number of scaling shifts are introduced, and they inevitably alter the original code sequence. Recently, we have observed that a compiler can often be adversely affected by this alteration, and consequently fails to generate efficient machine code for its target processor. In this paper, we present an optimization technique that safely migrates scaling shifts to other places within the code so that the compiler can produce better-quality code. We consider our technique to be safe in that it does not introduce new overflows, yet preserving the original SQNR. The experiments on a commercial fixed-point DSP processor exhibit that our technique is effective enough to achieve tangible improvement on code size and speed for a set of benchmarks.