Avoing tupling/untupling

[toelli-msft]

Cgen sometimes generates code like the below

tuple<tuple<tensor<2, double>,tensor<1, double>,tensor<1, double>,int>,int> s_k
    = std::make_tuple(s_karg1,s_karg2);
auto [s, k] = s_k;

We would expect a good compiler to be able to optimise away such redundancy. However, I believe that a C++ compiler cannot because of technical reasons to do with std::tuple's historical properties.

See also https://github.com/microsoft/knossos-ksc/issues/528 and https://github.com/microsoft/knossos-ksc/issues/531.

Part of the cause is the way that we unpack tuple arguments in Cgen. Here is a synthesis of ideas from @dcrc2 and @simonpj to avoid the problem.

• Make it a precondition of Cgen that a Def has a TupPat not a VarPat
• To achieve the precondition of Cgen run (a variation of) params_withPackedParamsPat on each Def.
• Run optimization after converting each Def to a TupPat not before (or also before, but that's probably redundant). Running optimization after means that optLets has an opportunity to remove this redundancy.

[toelli-msft]

I recommend waiting for https://github.com/microsoft/knossos-ksc/issues/531 to be resolved before implementing this strategy though. It may not be needed at all.

[toelli-msft]

Oh, and in Cgen we can also flatten nested tuple types as function arguments, if we think that would help.