feat(ssa): Switch mem2reg pass to be per function rather than per block

[vezenovm]

Description

Update mem2reg to be ran per function rather than per block. This enables mem2reg to be easily reordered amongst the SSA pass, such as being placed before loop unrolling. The previous per block mem2reg would break if placed before the flattening of the CFG.

Problem*

Resolves #2218

Summary*

The main change in this PR is that we now have to analyze the CFG when we want to find a load to substitute and fetch the respective load value to be substituted. Three new methods have been made:

fetch_load_value_to_substitute

• This method fetches the already save load values to substitute for a given address
  The search starts at the block supplied as a parameter. If there is not a load to substitute
  the CFG is analyzed to determine whether a predecessor block has a load value to substitute.
  If there is no load value to substitute the original address is returned.

find_load_to_substitute

• This method determines which loads should be substituted.
  Starting at the block supplied as a parameter, we check whether a store has occurred with the given address.
  If no store has occurred in the supplied block, the CFG is analyzed to determine whether
  a predecessor block has a store at the given address.

has_common_successor

• This method checks whether the block we are examining shares a common successor with its predecessor block. This check is necessary to prevent removing stores across blocks whose execution is dependent upon a runtime if statement. Without this check we could potentially remove a store and replace its load for only one block of a jmp if. We should be waiting for flattening to handle both scenarios. Thus, the mem2reg pass should also be ran after flattening to handle any runtime if statements.

We maintain two maps for loads to substitute. Per function, and per block. As when it comes time to actually substitute load result values and delete load instructions we want it to be done in the correct block order. We could most likely come-up with a strategy to use the per function loads to substitute map for replacing load result values, but simply maintaining a separate map felt like a simpler approach.

Documentation

• This PR requires documentation updates when merged.

  • I will submit a noir-lang/docs PR.
  • I will request for and support Dev Rel's help in documenting this PR.

Additional Context

PR Checklist*

• I have tested the changes locally.
• I have formatted the changes with Prettier and/or cargo fmt on default settings.

[jfecher]

I haven't finished a full review yet - I'll have to look at this more in depth tomorrow. I'd like to test it on some examples as well.

[jfecher]

I'm still not 100% confidant in this pass honestly since it is getting harder to keep the corner cases in my head as the pass gets more complex but my thoughts are we should merge it (after some minor comments) since it passes all the tests and flag any issues for fixing later. I have to make some changes to mem2reg to support alias analysis now as well so I'll be making some other large changes to the pass in the future.

[jfecher]

👍

[vezenovm]

@jfecher Could you approve once more?