stmts clustering feature to speed up code generation:
if the code is:
for i
S0
S1
the generated isl ast will be
for i
S0
then when the AST is parsed into HalideIR, S1 is added just after S0 to have the correct code.
In short, fused stmts will be represented by 1 stmt only for isl polyhedral code generation, then when it is time to parse into HalideIR we add back the previous statements using the same schedule in the same node as the Representative statements (except last beta which equivalent to say that it needs to be scheduled just after.
stmts clustering feature to speed up code generation: if the code is: for i S0 S1