Add read-uninitialised locals check

[agle]

• adds read-uninitialised locals invariant check: this is a correctness requirement of the parameter transform
• for this to hold for all procedures param analysis needs interproc liveness results for every procedure, not just those reachable from main. Still need a better approach for this for all IDE solver calls than currently implemented
• fix interproc liveness propagating live local vars across returns
• make copyprop throw on subst cycle with useful error message

[github-actions]

scalafmt check passed.

last updated at 90ebef7 (logs).

[agle]

In test IntervalDSATest/cntlm globals - local phase 4 more procedures collapse now:

      "hlist_subcmp_all",
      "headers_recv",
      "scanner_hook",
      "config_open"

Not sure which change causes this, it doesnt look like the inlining fix or the change to liveness alone cause it.

[agle]

Seems fixed now, idk why I couldn't get the tail-recursive find op to work, probably a skill issue.

[l-kent]

Is the read-uninitialised invariant check supposed to catch the sort of case we had with cntlm_noduk?

If I revert the fix for the inter-procedural live variables analysis and run the cntlm-no-duk InvervalDSATest, the infinite loop in the copy prop is identified and an exception is thrown, but the read-uninitialised invariant doesn't seem to catch any of that?

[l-kent]

The read-uninitialised invariant does seem to fail for some of the cases in MemoryTransformTests, although it isn't causing the tests to fail

[agle]

It detects it when placing the invariant check in the simplify pipeline before the copyprop pass, immediately after DSA, I'll commit a check there too.

[agle]

The read-uninitialised invariant does seem to fail for some of the cases in MemoryTransformTests, although it isn't causing the tests to fail

In runutils after the memory transform I run the check but don't assert it passes as it seems to not conform; I expect the transform would also need to update parameter lists in some cases but I'm not sure the exact reasons it fails.

[l-kent]

Was this just for debugging purposes?

[katrinafyi]

looks good and it has worked in tv.

is it worth a comment that read-uninit only checks in RPO order rather than dominance? because of this, it is possible for it to miss some uninitialised reads.