Ctrl flow to data flow

[tancheng]

To represent ctrl flow to data flow:

• Leverage predicate-based data type: https://mpslab-asu.github.io/publications/papers/Hamzeh2014DAC.pdf
• For unstructured for loop (in MLIR.LLVM), Phi node added at the beginning of each basic block
  • Pattern matching/fusion can be applied later to optimize the inter-iteration dependency @MeowMJ @HardyVon @yo96 @HobbitQia
• On the other hand, if the loop is represented as a high level for i in N with stride S (e.g., scf/cf/affine), we can directly lower it to a special operation (mimic what happen in HLS) or the fused Phi-related operation @ShangkunLi
• Interpreter should be able to run/execute both ctrl-flow IR and data-flow IR @Yfeng-44

[tancheng]

#29 inserts phi into each basic block (except the entry one) and removes bb. Note that the phi op can take a variable number of operands, so either HW should support this or we need another pass to decompose the phi with more than 2 operands.

[tancheng]

@ShangkunLi I will next enable the backward ctrl flow like:

reserve a
phi out, a, b
...
add z, x, y
ctrl_mov z, a

Different from data_mov, the ctrl_mov has no output, it just serves as an indicator to connect the phi and add. This is to respect SSA. (I would rename ctrl_mov to backward_mov.)