[PatternMatcher] Support matching tuples, call nodes, and functions with variable numbers of inputs

[mbrookhart]

Also fixes partitioning functions whose bodies are matched fuzzily.

cc @jroesch @masahi @slyubomirsky @ekalda

Also see discuss forums here:
https://discuss.tvm.apache.org/t/tuplepattern-for-any-number-of-inputs-to-the-tuple/8926
https://discuss.tvm.apache.org/t/pattern-language-matching-a-tuple-with-variable-size/9497

[ekalda]

Looks great! Thank you for doing this! :)

[ekalda]

Why the use of None? What would happen if the brackets were left empty (similarly to is_constant())? (Just asking for enlightenment)

[mbrookhart]

I wanted to keep the API consistent between tuple, call, and function. There are many places were we may have functions or ops with no arguments, so calling an op with op() yields an empty input list, which is a concrete pattern we want to keep matching. I choose None to easily identify the fuzzy case against the empty input case.

[mbaret]

I agree the use of None here is a bit confusing, I think I'd naturally interpret it the other way around actually. Specifically, None seems like you're explicitly requesting that there be no arguments rather than an unknown number. Just as a suggestion, would a syntax like is_tuple(args()) be clearer where we explicitly introduce a new pattern which matches a variable number of arguments?

[mbrookhart]

The problem is that we're not actually talking about a pattern, which explicitly represents an AST node. Instead, we're trying to represent an attribute of an AST node, particularly an attribute of the type Array<Expr>, that's what needs to be passed in. In the backend I need to be able to distinguish between a valid array of size 0 that I want to match on, and an invalid fuzzy array I don't want to match on. The simplest way to do that is to store a nullptr, and when I pass python None to the backend, that's what I get.

I could introduce something like an arg+ pattern that goes into the argument list, but that screws with the locality of the pattern matcher and creates a need to do a lot of look-ahead to determine how to match arguments so I don't accidentally index into an array I shouldn't or reject the match because the arguments don't match. It's feasible, but messy.

[masahi]

@ekalda @mbaret do we agree with the interface?

[ekalda]

I'm fine with it, I think all alternatives to None are non-obvious in their own way. Maybe worth adding an example to https://tvm.apache.org/docs/langref/relay_pattern.html# so that people would be aware of this functionality and the interface?

[mbaret]

I agree with @ekalda, in the absence of a more self-documenting interface explicit documentation is the next best option.

[mbrookhart]

This makes sense, thanks! I'll add this to the docs (fix my spelling mistake :D) And re-push.

[masahi]

thanks @mbrookhart @ekalda @mbaret