Here are some performances with higher optimization parameters and with the Intel C Compiler. All run with a Haswell Xeon (E5-2698 v3). Using icc, for the larger problem sizes you get performance speedup that very similar to the speedup from the hard-coded asm.

With same arguments as blog post:

gcc -O1 -o fft-test fft-test.c fft-portable.c -lm:

...

   262144    min=15721397  mean=15748901  sd=0.12%
   524288    min=34474976  mean=34546878  sd=0.18%
  1048576    min=99814826  mean=100391926  sd=0.26%

...

Adding -march=native and -O2:

...

   262144    min=14668178  mean=14701671  sd=0.14%
   524288    min=31297081  mean=31395468  sd=0.14%
  1048576    min=94027867  mean=94196746  sd=0.13%

...

With Intel C Compiler (-O2 -xHost):

...

   262144    min=14258167  mean=14297606  sd=0.16%
   524288    min=31126755  mean=31212153  sd=0.23%
  1048576    min=92893600  mean=93491419  sd=0.46%

...

Can provide more numbers if people want them. EDIT: looks like I was using the less optimized C code, will re-run with the better performing C code when I get the chance.