SkipStage doesn't understand mux

[abadams]

A mux is equivalent to a select tree, but the skip-stages pass doesn't understand that it means some values are unused. Consider the following:

#include "Halide.h"

using namespace Halide;

int main(int argc, char **argv) {
    Func f, g;
    Var x, c;
    f(x) = sqrt(sqrt(sqrt(sqrt(sqrt(x)))));
    f.compute_root();

    // g(x, c) = select(c == 0, f(x), 0.f);
    g(x, c) = mux(c, {f(x), 0.f});

    g.reorder(x, c).vectorize(x, 16, TailStrategy::RoundUp).bound(c, 0, 2).unroll(c);
    f.compute_at(g, x).vectorize(x);

    g.compile_jit();

    return 0;
}

f is only used for c == 0, but this code computes it in the c == 1 case too:

produce g {
  let t3 = (g.extent.0 + 15)/16
  for (g.s0.x.x, 0, t3) {
   allocate f[float32 * 16]
   produce f {
    f[ramp(0, 1, 16)] = (float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32(float32x16(ramp((g.s0.x.x*16) + g.min.0, 1, 16)))))))
   }
   consume f {
    g[ramp(g.s0.x.x*16, 1, 16) aligned(16, 0)] = f[ramp(0, 1, 16)]
   }
   free f
  }
  let t4 = (g.extent.0 + 15)/16
  for (g.s0.x.x, 0, t4) {
   allocate f[float32 * 16]
   produce f {
    f[ramp(0, 1, 16)] = (float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32((float32x16)sqrt_f32(float32x16(ramp((g.s0.x.x*16) + g.min.0, 1, 16)))))))
   }
   free f
   consume f {
    g[ramp((g.s0.x.x*16) + g.stride.1, 1, 16)] = x16(0.000000f)
   }
  }
 }

This extra copy of f isn't always dead-stripped by LLVM.