Translating out of SSA sometimes requires inserting copies and
constant initializations on edges to eliminate PHI nodes.
In some cases those copies and constant initializations are
redundant because the target already has the value on the
RHS of the assignment.
We previously tried to catch these cases after translating
out of SSA form. However, that code often missed cases. Worse
yet, the cases it missed were also often missed by the RTL
optimizers. Thus the resulting code had redundant instructions.
This pass attempts to detect these situations before translating
out of SSA form.
The key concept that this pass is built upon is that these
redundant copies and constant initializations often occur
due to constant/copy propagating equivalences resulting from
COND_EXPRs and SWITCH_EXPRs.
We want to do those propagations as they can sometimes allow
the SSA optimizers to do a better job. However, in the cases
where such propagations do not result in further optimization,
we would like to "undo" the propagation to avoid the redundant
copies and constant initializations.
This pass works by first associating equivalences with edges in
the CFG. For example, the edge leading from a SWITCH_EXPR to
its associated CASE_LABEL will have an equivalency between
SWITCH_COND and the value in the case label.
Once we have found the edge equivalences, we proceed to walk
the CFG in dominator order. As we traverse edges we record
equivalences associated with those edges we traverse.
When we encounter a PHI node, we walk its arguments to see if we
have an equivalence for the PHI argument. If so, then we replace
the argument.
Equivalences are looked up based on their value (think of it as
the RHS of an assignment). A value may be an SSA_NAME or an
invariant. We may have several SSA_NAMEs with the same value,
so with each value we have a list of SSA_NAMEs that have the
same value.
Main structure for recording equivalences into our hash table.