I’ve just done a quick and nasty comparison of runtimes using GNU Fortran and the recommended compiler flags from here. I used a CFD code mentioned earlier.

Compiling as before (gfortran euler.f) gave a runtime (using time) of:

CALCULATION CONVERGED IN 3490 ITERATIONS
TO A CONVERGENCE LIMIT OF 4.99999987E-06

real 1m6.156s
user 1m4.030s
sys 0m0.300s

Next, I compiled using gfortran -march=armv6 -mfpu=vfp -mfloat-abi=hard euler.f with the following result:

CALCULATION CONVERGED IN 3490 ITERATIONS
TO A CONVERGENCE LIMIT OF 4.99999987E-06

real 1m5.127s
user 1m3.800s
sys 0m0.330s

So a slight speed increase is shown. This difference would probably be amplified when more complex code is compiled. I had none to hand, alas.