Conclusion

We have implemented the strict (and default) floating-point semantics into our JIT compiler.
We have studied efficient implementation.
We could save setting the rounding precision.
We measured the performance of several methods in shuJIT, BulletTrain, and C & assembly code implementations, and compared them.
More analysis is necessary to obtain the useful guideline.
We should make investigation into the scheduling in the processor in detail.

Future work
More analysis.
Experiments on the other x86 processors including Athlon, Pentium.
Implementing the language factor like `strictfp' in other languages (C or Fortran?) is interesting.