Canonicalization: Remove unused primal arguments for region op

[vimarsh6739]

It is safe to remove unused primal arguments for AutoDiffRegionOp when the primal has no uses within the function and is enzyme_const

[wsmoses]

It's safe to remove any unused arg, not just const

[vimarsh6739]

There is one caveat, we assume that the only use of primals in the region are the block arguments of the entry basic block(and not the passed arguments to the op). Should we add a verifier for this? Currently, this is a manual check which is repeated for both outlining and #2562

[wsmoses]

e assume that the only use of primals in the region are the block arguments of the entry basic block(and not the passed arguments to the op). Should we add a verifier for this? Currently, this is a manual check which is repeated for both out

No that's a limitation that should be fixed by the requisite users, as that is legal IR that ought be supported [tho perhaps not as well optimized]

[vimarsh6739]

@wsmoses thinking a bit more about it, the trigger conditions for args with activity which aren't enzyme_const is dicey. Consider the following example:

func.func @foo(%x : f64, %dx : f64) -> f64 {
  %cst2 = arith.constant 2.0 : f64
  %x2 = arith.mulf %cst2, %x : f64
  %dout = enzyme.fwddiff_region(%x, %dx){
  ^bb0(%xx : f64):
      %y = arith.mulf %x2, %x2 : f64
      enzyme.yield %y : f64 
  } attributes { activity = [#enzyme<activity enzyme_dup>], ret_activity=[#enzyme<activity enzyme_dupnoneed>] }: (f64,f64) -> f64
  return %dout : f64
}

If we remove (%xx == {%x, %dx}) (since it is never used, we would run into issues, since %x2 is effectively re-added always as enzyme_const when outlining the region and lowering).

This also holds for the reverse mode case(in the below example, %x2 will be considered enzyme_const)

func.func @foo(%x : f64, %dout : f64) -> f64 {
  %cst2 = arith.constant 2.0 : f64
  %x2 = arith.mulf %cst2, %x : f64
  %dx = enzyme.autodiff_region(%x, %dout){
  ^bb0(%xx : f64):
      %y = arith.mulf %x2, %x2 : f64
      enzyme.yield %y : f64 
  } attributes { activity = [#enzyme<activity enzyme_active>], ret_activity=[#enzyme<activity enzyme_activenoneed>] }: (f64,f64) -> f64
  return %dx : f64
}

The other option is to leave it as a convention and make a note somewhere to users that this would const out computations.

[wsmoses]

ove (%xx == {%x, %dx}) (since it is never used, we would run into issues, since %x2 is effectively re-added always a

I don't understand this case, x2 is always const here. in fact the entire region is constant? %xx is literally unused in the region so it is always safe to remove, definitionally.