Guard Canonicalization is invalid

[rachitnigam]

Consider the following two sets of assignments:

w.in = a.out ? 1'd1;
w.in = b.out ? 1'd1;

and

w.in = a.out;
w.in = b.out;

In the first program, we're saying something like: if a.out or b.out is active, forward 1'd1 to w.in, otherwise forward nothing. In the second program, we're saying unconditionally forward both a.out and b.out to w.in which is an obvious runtime conflict.

The first program can be optimized into:

w.in = a.out | b.out ? 1'd1;

However, the canonicalization pass currently transforms program (1) to (2), altering the semantics. We should disable rewrites of the form a = g ? 1'd1 -> a = g and instead only allow the canonicalizer to perform a = 1'd1 ? b to a = b.

[sampsyo]

It's a good example, and I think it clearly depends on nailing down semantics for undefinedness as suggested in #922… namely, I don't think it's a foregone conclusion that this is always a conflict:

w.in = a.out;
w.in = b.out;

We don't right now, but we could decide that this is perfectly safe if a.out and b.out are never defined simultaneously.

[rachitnigam]

Right, in which case we need to revisit the meaning of direct assignments and make sure that compilation and other passes don't change it