Experiments We compared several implementation methods. shuJIT Scaling can be performed with ... fscale instruction multiplication (fmul instruction) Pre-loading of the scales load the scales to registers when a strictfp method is called. load when scaling. BulletTrain 1.4.0 An Ahead-of-Time compiler, which translates a Java bytecode into a x86 native code. uses the fscale instruction, without pre-loading. We also developed small benchmark programs in C and modified the derived assembly code to implement the strict semantics.