fix(ssa): mem2reg to treat block input references as alias

[aakoshh]

Description

Problem*

Resolves #6443

Summary*

The mem2reg SSA pass eliminates stores that have no use, but as testing with different inliner aggressiveness showed it also eliminated stores into mutable references passed to a function. This is an effect that is expected to be visible outside, after the call. It turns out that PerFuncContext::is_store_alias_used does consider the input reference parameters of the function as stopping conditions for removing a store to an address, but it does so only if there is an alias for the address which appears in the input parameters. This wasn't the case, and the removal went ahead.

I added a method that adds a self-alias to all block parameters if they are reference types and don't already have aliases from the merge of the predecessor blocks.

Additional Context

See investigation in the ticket, e.g. #6443 (comment) shows a minimal failing form.

After merging this into #6429 I only have these tests failing:

failures:
    tests::execution_success::test_eddsa::true_0_expects
    tests::execution_success::test_eddsa::true_i64_min_expects
    tests::execution_success::test_slice_dynamic_index::true_i64_min_expects

Which means it potentially affects #6440 and #6442 as well.

Documentation*

Check one:

• No documentation needed.
• Documentation included in this PR.
• [For Experimental Features] Documentation to be submitted in a separate PR.

PR Checklist*

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

[github-actions]

Changes to Brillig bytecode sizes

Generated at commit: 12208dc0dd3d0b829407c8d52a8328cde9726efe, compared to commit: 8b5afec57725f68a069342067dd42a79a2e866c3

🧾 Summary (10% most significant diffs)

Program	Brillig opcodes (+/-)	%
uhashmap	-82 ✅	-0.49%
array_sort	-10 ✅	-3.27%

---

Full diff report 👇

Program	Brillig opcodes (+/-)	%
hashmap	27,063 (-20)	-0.07%
uhashmap	16,526 (-82)	-0.49%
array_sort	296 (-10)	-3.27%

[github-actions]

Changes to number of Brillig opcodes executed

Generated at commit: 12208dc0dd3d0b829407c8d52a8328cde9726efe, compared to commit: 8b5afec57725f68a069342067dd42a79a2e866c3

🧾 Summary (10% most significant diffs)

Program	Brillig opcodes (+/-)	%
uhashmap	-2,736 ✅	-1.76%
array_sort	-44 ✅	-8.06%

---

Full diff report 👇

Program	Brillig opcodes (+/-)	%
hashmap	57,191 (-198)	-0.35%
uhashmap	152,484 (-2,736)	-1.76%
array_sort	502 (-44)	-8.06%

[jfecher]

LGTM. I'll test to see if this fixes my mem2reg issues in my bug fix PR as well.

Is there any reason this is a draft?

[aakoshh]

Is there any reason this is a draft?

Thanks for checking! I left it was draft just because:

• I wanted to quickly run my tests locally on a merge into an "integ" branch that contains the brillig_cow fix and the inlining test matrix, to see if anything else broke
• I was looking at the unit tests, wondering if I should add one - I probably should, but if you're okay with it as it is, covered by the test_programs, then I will move on. NB @asterite had the idea of writing a parser for the display format of SSA to make it easier to create tests.

[TomAFrench]

NB @asterite had the idea of writing a parser for the display format of SSA to make it easier to create tests.

I think this would be very helpful tbh. It would be a lot easier for us to move full integration tests into unit tests for individual SSA passes.

[asterite]

Awesome! I was about to starting working on that a couple of times yesterday but wasn't sure if it would be wanted, and also wanted to start working on #6453... and the SSA language isn't that small, though it's definitely much smaller than Noir. I'll start working on this in parallel with other things after I finish #6453. I think it will also help quickly write a test by copying some output you get from --show-ssa, putting it into a test and then further checking things.