AVM2 verifier

[Lord-McSweeney]

This is #14563 but with state merging removed, since I couldn't figure it out. It also includes some basic optimizations, and it keeps verification of local registers in (since that's not dependent on the control flow).

TODO:

• Something in try..catch blocks is broken (I keep losing working reproductions, but FP is throwing verify errors where Ruffle is not) Not relevant anymore
  • Also only verify exception target if there are any potentially error-throwing opcodes present in the from..to Fixed
  • All error-throwing opcodes in the from..to should act as a potential jump to the exception target (just like iftrue, iffalse, etc)- example: verify.zip Fixed, like above
• More brokenness Not relevant anymore
• Add many two tests
• Store parsed_code on the method itself rather than in AbcMethodBody
• Add a custom Op type in AVM2 that stores already-resolved constant pool entries
• Storing an entire verified AbcMethodBody seems unnecessary, and is mostly leftover from Adrian's original work

The optimization framework that this PR implements should work well with #7920.

Supersedes #14563.

Future optimizations:

• Remove Nop opcodes in code and modify jump offsets
• Record more local types from parameters (relies on [WIP] avm2: Store resolved function param types #11954)
• Optimize setproperty/initproperty and callproperty/callpropvoid

[evilpie]

Can we unnest this a bit? (And definitely rename previous_op_some)

e.g.

let previous_op = if let Some(previous_op) = previous_op {
  previous_op
} else {
  previous_op = Some(op.clone());
}

In theory we could even do something like

previous_op = code.first();
for (...) in code.iter_mut().enumerate().skip(1) {

[Lord-McSweeney]

I can't do let Some(previous_op) = previous_op because the optimizer needs to set the Option value of previous_op. I'll try unnesting it.

[Lord-McSweeney]

I tried unnesting it but I realized that there are likely going to be optimizations that we can do that would work despite previous_op being None, and/or those that would work even when the next op is on a jump target.

[evilpie]

We could write this as previous_op = mem::replace(op, Op::Nop), but I am not sure if that is really worth it.

[Lord-McSweeney]

I think this current implementation is more readable.

[adrian17]

can we move these to script.rs? It already has a bunch of similar functions, so it'd be more consistent (and decrease verify.rs size a bit too).
Also, any chance this could be unified with abc_int() (and friends)? As they are pretty much duplicates, sans 0 treatment. (this'd also probably make function with invalid default values throw a correct VerifyError "for free"). Can be left for a follow-up though.

[Lord-McSweeney]

I'm going to leave this for a follow-up to also fix more issues with namespace and string pooling