[SCFToCalyx] Issue with lowering scf::parallel to Calyx when there is scf::if in it

[jiahanxie353]

We lower scf::parallel in SCFToCalyx by assuming constant loop bounds and strides, which enables us to unroll the parallel loop manually: #7830

The issue is that, if scf::parallel contains scf::if, whose condition variable depends on some constants and scf::parallel's loop induction variables, those condition variables will be eventually evaluated to constant true/false automatically even if we never implement this sort of "canonicalization" pass manually in SCFToCalyx.cpp (I guess it's done by some underling MLIR mechanisms?).

For example, if we begin with:

func.func() {
    %c0 = arith.constant 0 : index
    scf.parallel (%arg1, %arg2) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
      %1 = arith.remsi %arg2, %c2 : index
      %2 = arith.cmpi slt, %1, %c0 : index
      %3 = scf.if %2 -> (f32) {
        scf.yield ...
      }
    }
}

will turn into:

func.func() {
    %c0 = arith.constant 0 : index
    scf.parallel (%arg1, %arg2) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
    ^bb0:
      %1 = arith.remsi "value 0", %c2 : index
      %2 = arith.cmpi slt, %1, %c0 : index
      %3 = scf.if %2 -> (f32) {
        scf.yield ...
      }
    ^bb1:
      %1 = arith.remsi "value 1", %c2 : index
      %2 = arith.cmpi slt, %1, %c0 : index
      %3 = scf.if %2 -> (f32) {
        scf.yield ...
      }
    ^bb2:
    ...
    }
}

And in each basic block:

%1 = arith.remsi "value 0", %c2 : index
%2 = arith.cmpi slt, %1, %c0 : index

will be folded to constants automatically, turning the scf::if op to:

%3 = scf.if true/false {
 ...
}

If they were to be lowered to Calyx, it'll become:

calyx.if 1'd1/1'd0 {
...
}

but it's not allowed by Calyx' grammar: https://github.com/calyxir/calyx/blob/2e18d555ff00339f46c99cc9d8b698ee27ce25ad/calyx-frontend/src/syntax.pest#L348

Any idea on how to tackle this? @rachitnigam @cgyurgyik @andrewb1999 @mikeurbach Thanks!

[cgyurgyik]

Can't we just simplify scf.if true { body } => body?