Inconsistent rfactor

[naoyam]

In this fusion, the rfactor tensor gets an inconsistent rfactor domain.

TEST_F(NVFuserTest, TMP) {
  Fusion fusion;
  FusionGuard fg(&fusion);

  auto tv0 = makeSymbolicTensor(2);
  fusion.addInput(tv0);

  auto tv1 = sum(tv0, {0, 1});

  fusion.addOutput(tv1);

  tv1->split(1, 4);
  tv1->split(0, 3);
  tv1->merge(1, 2);
  auto rf = tv1->rFactor({-1});

  std::cout << ir_utils::toString(rf->getRootDomain()) << std::endl;
  std::cout << ir_utils::toString(rf->getMaybeRFactorDomain()) << std::endl;

  GpuLower gpu_lower(&fusion);
}

The rfactor domain of tensor rf: iS9{i1} iS15{( 3 * ( ceilDiv(i2, 4) ) )}rf rS14{4}rf. Specifically, 3 comes from i1, but it's also included in the second ID.

Not sure if that's the root cause, but the fusion results in an error when building the computeAt map:

* thread #1, name = 'test_jit', stop reason = breakpoint 1.1
    frame #0: 0x00007fffd5fd910b libstdc++.so.6`__cxxabiv1::__cxa_throw(obj=0x0000000002d9c720, tinfo=0x00007fffd60927a8, dest=(libstdc++.so.6`std::out_of_range::~out_of_range() at stdexcept.cc:65:33))(void *)) at eh_throw.cc:77:1
Process 13953 launched: '/home/nmaruyama/pytorch/debug3/build/bin/test_jit' (x86_64)
(lldb) bt
* thread #1, name = 'test_jit', stop reason = breakpoint 1.1
  * frame #0: 0x00007fffd5fd910b libstdc++.so.6`__cxxabiv1::__cxa_throw(obj=0x0000000002d9c720, tinfo=0x00007fffd60927a8, dest=(libstdc++.so.6`std::out_of_range::~out_of_range() at stdexcept.cc:65:33))(void *)) at eh_throw.cc:77:1
    frame #1: 0x00007fffd5fd6087 libstdc++.so.6`std::__throw_out_of_range(__s="_Map_base::at") at functexcept.cc:82:5
    frame #2: 0x00007ffff6f5c9ce libtorch_cuda.so`std::__detail::_Map_base<torch::jit::fuser::cuda::IterDomain*, std::pair<torch::jit::fuser::cuda::IterDomain* const, torch::jit::fuser::cuda::VectorOfUniqueEntries<torch::jit::fuser::cuda::IterDomain*, std::hash<torch::jit::fuser::cuda::IterDomain*> > >, std::allocator<std::pair<torch::jit::fuser::cuda::IterDomain* const, torch::jit::fuser::cuda::VectorOfUniqueEntries<torch::jit::fuser::cuda::IterDomain*, std::hash<torch::jit::fuser::cuda::IterDomain*> > > >, std::__detail::_Select1st, std::equal_to<torch::jit::fuser::cuda::IterDomain*>, std::hash<torch::jit::fuser::cuda::IterDomain*>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, std::__detail::_Hashtable_traits<false, false, true>, true>::at(this=0x0000000002d98418, __k=0x00007fffffffbe38) at hashtable_policy.h:760:2
    frame #3: 0x00007ffff6f55fdd libtorch_cuda.so`std::unordered_map<torch::jit::fuser::cuda::IterDomain*, torch::jit::fuser::cuda::VectorOfUniqueEntries<torch::jit::fuser::cuda::IterDomain*, std::hash<torch::jit::fuser::cuda::IterDomain*> >, std::hash<torch::jit::fuser::cuda::IterDomain*>, std::equal_to<torch::jit::fuser::cuda::IterDomain*>, std::allocator<std::pair<torch::jit::fuser::cuda::IterDomain* const, torch::jit::fuser::cuda::VectorOfUniqueEntries<torch::jit::fuser::cuda::IterDomain*, std::hash<torch::jit::fuser::cuda::IterDomain*> > > > >::at(this=0x0000000002d98418, __k=0x00007fffffffbe38) at unordered_map.h:991:21
    frame #4: 0x00007ffff6f52a3b libtorch_cuda.so`torch::jit::fuser::cuda::IterDomainGraph::build(this=0x0000000002d982f0, fusion=0x0000000002d28140) at compute_at_map.cpp:180:22

[zasdfgbnm]

Assign myself to take a look