Spurious panics in var-to-phi promotion

[fabianschuiki]

The vtpp transformation generates spurious panics of the following form, when running on the complex tests in the moore repository (test/svlog/complex/*):

thread '<unnamed>' panicked at 'invalid value', [...]/llhd-0.13.0/src/ir/unit.rs:367:35

The backtrace would generally look something like this:

[...]
12: llhd::ir::unit::Unit::value_type
      at src/ir/unit.rs:367
13: llhd::ir::inst::InstBuilder::value_type
      at src/ir/inst.rs:585
14: llhd::ir::inst::InstBuilder::phi
      at src/ir/inst.rs:471
15: llhd::pass::vtpp::materialize_value
      at src/pass/vtpp.rs:168
16: <llhd::pass::vtpp::VarToPhiPromotion as llhd::opt::pass::Pass>::run_on_cfg
      at src/pass/vtpp.rs:83
17: llhd::opt::pass::Pass::run_on_unit
      at src/opt/pass.rs:23
18: llhd::opt::pass::Pass::run_on_module::{{closure}}
      at src/opt/pass.rs:17
[...]

The issue occurs only sporadically, so is probably either dependent on the rayon parallelization for some reason, or some randomness implied by HashSet/HashMap key ordering.

Possibly a phi node is being inserted for a value that was already removed, or should have been collected by DCE?