Improve hoisting, CSE, LSRA for MD array accesses #71676

As described in the comment #70271 (comment) and those that follow, with change #70271 a regression was introduced in the MDMulMatrix benchmark due to limitations or heuristics in the optimization phases downstream of MD array index expansion.

E.g.,

• LSRA inserts spills and reloads in unfortunate places, including an identified case of un-spilling and re-spilling in the same block, when the value wasn't modified. One issue to investigate this was opened: Investigate the heuristics for creating DummyDef in LSRA #71539
• there are many cases where a GT_MDARR_LOWER_BOUND or GT_MDARR_LENGTH node is loop-invariant, but can't be hoisted due to exception ordering, as these could cause null reference exception. This is due to hoisting having a very simple notion of side-effects and ordering. In the identified case, we already know that the specific null reference exception can't occur because a previous expression would have caused it, so we should be free to hoist up to that location.
• CSE seems to create too many long-lifetime CSE, leading LSRA to spill. For the two mentioned nodes, they are very cheap to compute. If they get assigned a register, it's useful to hoist them, but if they don't, they shouldn't be CSE'ed. Maybe CSE's costing should be improved, or we should implement rematerialization for "undo-ing" some CSE.

category:cq
theme:md-arrays
skill-level:expert
cost:large

impact:medium