Regarding your last point, as of now we have four main processor hardware namely CPU, GPU, FPGA and TPU (the new kid on the block). The hard question or problem is how can we program these different animals seamlessly?

Chris Lattner of LLVM and Swift fame is currently working on MLIR to program heterogeneous processors [1].

At the same time ETH Zurich is working on LLHD, to unify hardware design and programming (e.g. Verilog, VHDL) including the testing or testbench that is using conventional software programming (e.g. TCL, Perl and SystemVerilog) [2].

Apparently it seems that the key concept to solving this problem is Static Single Assignment (SSA) form. It is well known that SSA can seamlessly support imperative and functional programming [3].

For the past several years, both Chris (CPU, GPU & TPU) and ETH Zurich (CPU, FPGA and Circuit Design) are independently using SSA to solve this hard problem.

I personally believe SSA concept will empower the compiler community as the quaternion concept empower physicists like Einstein overcome Maxwell Equation and 3D game programmers to overcome gimbal lock.

Only time will tell how this idea will pan out.

[1] https://mlir.llvm.org/

[2] http://llhd.io/

[3] https://www.cs.princeton.edu/~appel/papers/ssafun.pdf