I know now a days FPGAs have very good floating point performance as reported here for example . But then why it is said that Floating point is non-synthesizable in verilog? How FPGA handles float then?
It is not that they can't be synthesized (they can, of course!), but tools won't do it combinationally (like they would a fixed point adder) because the resource usage would be unfeasibly large. So it is done sequentially in multiple steps, and there are many, many ways of doing it with several trade-offs to consider, with division being quite complex. So your floating point operation needs to be treated like any other module, you can design it yourself or license it from a vendor, and a tool won't synthesize c = a/b for you, for the same reasons it won't synthesize an integral or an equation solver, but it doesn't mean it can't be done in an FPGA!
I know you asked about Verilog, but as an adjunct to apalopohapa's answer, VHDL has a synthesisable floating point library, which does not require instantiation of modules to perform floating-point operations, merely the use of ordinary function calls.
Other than the elegant use of VHDL's bit numbering to separate the exponent and significand, I don't think there's anything to stop someone writing a similar set of functions for Verilog to give the same capability.