99% sure that you are fetching instructions from external memory directly without realizing it. Please use e-objdump -D "your.elf" to see where your code is placed.
For example, with the constants you used, pretty sure that they get called in as doubles without a n f specifier. i.e. 1.0f If this is the case, then there will be lots of messy double/float conversion (and this code is read from DRAM!)
Andreas