JIT: Change VN's representation for phi definitions

[jakobbotsch]

Replace VNF_PhiDef and VNF_MemoryPhiDef by new explicit representations that represent all phi args directly.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[jakobbotsch]

A few diffs I want to look at the cause of. I probably won't get to that until after the snap.

[jakobbotsch]

The diffs look related to ordering in VNs. In the baseline we have:

N001 [000862]   CNS_INT(h) 0x400000000042cea8 static => $208 {Hnd const: 0x400000000042CEA8 GTF_ICON_STATIC_HDL}
N002 [000863]   CNS_INT   0 Fseq[<unknown field>] => $242 {LngCns 0}
N003 [000864]   ADD       => $380 {ADD($208, $242)}

where we didn't fold this handle + constant. The handle has the lower raw VN here. In the diff we have

N001 [000862]   CNS_INT(h) 0x400000000042cea8 static => $282 {Hnd const: 0x400000000042CEA8 GTF_ICON_STATIC_HDL}
N002 [000863]   CNS_INT   0 Fseq[<unknown field>] => $202 {LngCns 0}
N003 [000864]   ADD       => $282 {Hnd const: 0x400000000042CEA8 GTF_ICON_STATIC_HDL}

where we did fold it. The handle has the higher raw VN here. We ended up with the operands to VNF_ADD in the other order and then it seems like EvalUsingMathIdentity kicked in in the diff.

[jakobbotsch]

cc @dotnet/jit-contrib PTAL @EgorBo

Diffs. Minor diffs due to the above. Some TP improvements otherwise.
Code wise it looks like more code, but it's mainly boilerplate from adding the two new CEA_PhiDef and CEA_MemoryPhiDef. The actual logic that reasons about phi defs/memory phi defs definitely looks simpler now.

[EgorBo]

I'd probably use a single struct and unify memory with regular phi. It seems like we often probe both anyway, but LGTM as is too

[jakobbotsch]

Hmm, it would require introducing some "kind" type in the unified struct anyway that you will need to check. To actually use the SSA args you need to differentiate between whether it's memory or local, so I'm not sure it simplifies much. I think I'll leave it like this and then we can introduce a helper if there is a good way to unify it.