Invalid temporary allocation

[naoyam]

🐛 Bug

Non-static memory allocation used to be rejected at the lowering time by the check in the constructor of Allocate. It isn't now. E.g.,

   Fusion fusion;
    FusionGuard fg(&fusion);

    TensorView* tv0 = makeDummyTensor(1);
    fusion.addInput(tv0);
    TensorView* tv1 = add(tv0, new Float(0));
    TensorView* tv2 = add(tv1, new Float(0));
    fusion.addOutput(tv2);

    torch::jit::fuser::cuda::FusionExecutor fe;
    fe.compileFusion(&fusion);

This is invalid as we can't allocate a buffer for tv1, and that was detected before. Now it passes the lowering phase and fails when the generated code is compiled.

__global__ void kernel1(Tensor<float, 1> T0, Tensor<float, 1> T2){
  float T1[T2.size[0]];
  for(size_t i12 = 0; i12 < T2.size[0]; ++i12 ) {
    T1[ i12 ]
       = T0[ ( i12 * T0.stride[0] ) ]
       + float(0);
  }
  for(size_t i13 = 0; i13 < T2.size[0]; ++i13 ) {
    T2[ ( i13 * T2.stride[0] ) ]
       = T1[ i13 ]
       + float(0);
  }
}

[naoyam]

This first appeared at c94c67e.