Performance on the SunSpider math-cordic benchmark is very fast, presumably due to some sort of dead code analysis. However, the analysis seems very fragile or narrow. Simple transformations of the benchmark code seem to prevent the analysis from being applied.
Here are two slight changes to the math-cordic code that seem to defeat the analysis in question:
A performance test with those two changes, in addition to the original math-cordic file, is here:
What sorts of code does the analysis work on, other than the exact function included in SunSpider?