When working on tvm::runtime::Array, we found that CHECK causes significant regression in stack utilization. This is not acceptable especially in compilers written in recursive visitor pattern, which crashes on comparably deep neural networks.

As diving deeper into this case, we figured out that dmlc::LogMessageFatal is the cause - it is super sophisticated and throws inside the destructor.

So we would love to patch dmlc-core, move dmlc::LogMessageFatal from stack to heap. A simple trick is to wrap it with std::unique_ptr - it is doable because dmlc-core has been updated to C++ 11.

Besides patching dmlc-core, we can change the implementation of tvm::runtime::Array as well as follows:
0) no change (using CHECK)

1. throw dmlc::Error instead of using CHECK
2. throw std::runtime_error instead of using CHECK
3. remove and disable the CHECK

Below are the benchmarks we did. We use g++ -fstack-usage to dump the stack usage information for the target function.

Settings:

• Compiler: g++ 5.4
• OS: Ubuntu 18.04
• Stack size: ulimit -s 8192
• Target function: virtual tvm::relay::Expr tvm::relay::CommonSubexprEliminator::VisitExpr_(const tvm::relay::CallNode*), which originally crashed this PR on a very deep network

Results:

Side notes: It is unclear right now why throwing dmlc::Error takes more stack space than std::runtime_error, although they intent to be identical.

See also: apache/tvm#5585 (comment)