[Flang][OpenMP][MLIR] Add pass to optimize out allocation for worksharing loop indices

[skatrak]

This pass rewrites this pattern commonly appearing for parallel worksharing loops:

omp.parallel   {
  %0 = fir.alloca i32 {adapt.valuebyref, pinned}
  %c1_i32 = arith.constant 1 : i32
  %c10_i32 = arith.constant 10 : i32
  %c1_i32_0 = arith.constant 1 : i32
  omp.wsloop   for  (%arg0) : i32 = (%c1_i32) to (%c10_i32) inclusive step (%c1_i32_0) {
    fir.store %arg0 to %0 : !fir.ref<i32>
    %c7_i32 = arith.constant 7 : i32
    %1 = fir.load %0 : !fir.ref<i32>
    %2 = fir.convert %1 : (i32) -> i64
    %c1_i64 = arith.constant 1 : i64
    %3 = arith.subi %2, %c1_i64 : i64
    %4 = fir.coordinate_of %0, %3 : (!fir.ref<!fir.array<10xi32>>, i64) -> !fir.ref<i32>
    fir.store %c7_i32 to %4 : !fir.ref<i32>
    omp.yield
  }
  omp.terminator
}

Into this, where allocations for loop index variables have been removed:

omp.parallel   {
  %c1_i32 = arith.constant 1 : i32
  %c10_i32 = arith.constant 10 : i32
  %c1_i32_0 = arith.constant 1 : i32
  omp.wsloop   for  (%arg0) : i32 = (%c1_i32) to (%c10_i32) inclusive step (%c1_i32_0) {
    %c7_i32 = arith.constant 7 : i32
    %2 = fir.convert %arg0 : (i32) -> i64
    %c1_i64 = arith.constant 1 : i64
    %3 = arith.subi %2, %c1_i64 : i64
    %4 = fir.coordinate_of %0, %3 : (!fir.ref<!fir.array<10xi32>>, i64) -> !fir.ref<i32>
    fir.store %c7_i32 to %4 : !fir.ref<i32>
    omp.yield
  }
  omp.terminator
}

[skatrak]

Closing due to creating equivalent upstream PR that won't probably be merged. Another approach will be needed.