[Relay] Improve build error when no lowered funcs are produced

[weberlo]

If you accidentally create a module with no lowered funcs via relay.build, interaction with the module down the road segfaults without an error message. For example, the annotated line in the program below would segfault, because I have accidentally passed params to relay.build, causing the entire program to be const-folded away.

x = relay.var('x', shape=(1, 16, 64, 64), dtype='float32')
conv_expr = relay.nn.conv2d(
        x, relay.var('w'),
        kernel_size=(3, 3),
        padding=(1, 1),
        channels=16)
func = relay.Function(relay.analysis.free_vars(conv_expr), conv_expr)
mod = relay.Module.from_expr(func)

x_data = np.random.uniform(size=(1, 16, 64, 64)).astype('float32')
w_data = np.random.uniform(size=(16, 16, 3, 3)).astype('float32')
params = { 'x': x_data, 'w': w_data }

with tvm.build_config(disable_vectorize=True):
    graph, c_mod, params = relay.build(mod, target="c", params=params)
print(c_mod.get_source())  # <-- crashes here

This PR adds an error message to prevent the generation of modules without lowered funcs.

CC @jroesch @zhiics @tqchen

[zhiics]

Hmm, this is a good point. But I am not sure if we really want to signal fatal here. We might just have an empty lowered_funcs in some circumstances. For example, we could just have a constant and invoke build. Since there is no primitive, we would not have the module, but it should still be valid.

[weberlo]

I see. I wasn't sure if there were cases where one would want to use a build result without a defined module. Maybe the right change would be to modify the ModuleNode* operator-> methods to check that the shared ptr isn't null?

[zhiics]

I personally think returning a nullptr is okay. Dereferencing a nullptr should crash. I would recommend logging a warning here so that users would know the module is empty.

[weberlo]

Yeah. A warning seems like a good compromise.

[tqchen]

@weberlo please look into the ci error

[weberlo]

I've switched the check in build_module.cc from fatal to a warning, as @zhiics suggested. @tqchen Do you think modifying the ModuleNode* operator-> methods to include a runtime nullptr check is a good idea?

I'm not sure what other code paths would generate null modules, but I think if we're generating them anywhere in our APIs, we should at least give users the tools to understand why things are crashing (i.e., by adding a helpful error message).

[zhiics]

LGTM