Thanks for the clarification. You are right, our load-store-opt cannot handle loops.
Dead code elimination technically [0] is O(n). However, it should be called "copying garbage collection" instead.
LibFirms SSA construction is minimal [1], it does not create dead phis. Or course, phis might die due to optimizations, in which case they become unreachable and get garbage collected at some point (see above). However, Firm does not model memory (global variables) as SSA, only registers (local variables).
Minor correction: LibFirm also generates pruned SSA, which is the property that guarantees lack of dead phis. Minimal SSA roughly means "no superfluous live phis".
Let me know if you still do this once you integrate a polyhedral code generator and try to do incremental SSA renaming on the nested loop outputs :)
Note that dead phis are actually useful in some contexts, particularly around loop updates
This is why things like loop-closed ssa (http://gcc.gnu.org/onlinedocs/gccint/LCSSA.html) exist, and phi nodes are generated even if the variable is never initially used outside the loop (makes sinking easier)
Dead code elimination technically [0] is O(n). However, it should be called "copying garbage collection" instead.
LibFirms SSA construction is minimal [1], it does not create dead phis. Or course, phis might die due to optimizations, in which case they become unreachable and get garbage collected at some point (see above). However, Firm does not model memory (global variables) as SSA, only registers (local variables).
[0] https://github.com/MatzeB/libfirm/blob/master/ir/opt/dead_co... [1] https://pp.info.uni-karlsruhe.de/publication.php?id=braun13c...