Because the dassl code does not work with i387 FPU

What about SSE3/4 on modern CPU (Intel i7) (gcc -Ofast -march=native -mfpmath=sse ) ?

SSE3/4 is not disabled by setting -mfpmath=sse. The only valid options to that are: 387, sse, or both. -O3 is disabled by default because it typically then takes longer to compile the code than it takes to compile+run the code using -O0. But you can set cflags manually to override it.

I'm first concerned by -msse2 flag mainly, locking down gcc sse codegen to limited sse command set,
but now I think that this selection was done to circumvent some low-level incompatibility in some library or solver, like dassl/x387 said before.

By default, GCC is tuned to the generic architecture, so -msse2 enables instructions. (I think it was only necessary to specify -mfpmath=sse -msse2 on 32-bit x86 though.)

It does not limit SSE3/4 from being used (unless -march=native, etc is given earlier on the command-line). But again, if you want faster code, appending for example -O3 -march=native to the CFLAGS will enable everything your CPU supports.