Do not generate TYP_STRUCT LCL_FLD nodes

[SingleAccretion]

Today, the morpher can fold IND(struct ADDR(LCL)) into LCL_FLD TYP_STRUCT layout-less nodes, these end up being locations, always (well -- there is one case where that's not true, such a node can end up under a return of a small struct later to be retyped by lowering, but this is rare).

These will be a significant impediment to actual, proper TYP_STRUCT local field nodes, as it would mean handling the case, everywhere in code, where the layout would be missing, that we don't want to allow in the first place.

This change fixed that by introducing the folding that the creation of these location nodes achieves at the address level, where we can choose the type of the location node arbitrarily and freely.

Along the way, IsLocalAddrExpr was fixed to not forget to look at zero-offset sequences attached to ADDRs, to avoid regressions.

We are expecting some positive diffs here, from the IsLocalAddrExpr change (enabling more precise VNs) and the folding change itself (the old code path would, ever so rarely, miss out on some things).

We are also expecting one small regression due to call retyping of reg-sized IND struct(ADDR(LCL)) nodes losing the field sequence. I did not fix it because a) it was small, and rare, b) I do not want to permeate the zero-offset code any more than is strictly required.

Diffs.

[ghost]

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

Issue Details

---

Today, the morpher can fold IND(struct ADDR(LCL)) into LCL_FLD TYP_STRUCT layout-less nodes, these end up being locations, always (well -- there is one case where that's not true, such a node can end up under a return of a small struct later to be retyped by lowering, but this is rare).

These will be a significant impediment to actual, proper TYP_STRUCT local field nodes, as it would mean handling the case, everywhere in code, where the layout would be missing, that we don't want to allow in the first place.

This change fixed that by introducing the folding that the creation of these location nodes achieves at the address level, where we can choose the type of the location node arbitrarily and freely.

Along the way, IsLocalAddrExpr was fixed to not forget to look at zero-offset sequences attached to ADDRs, to avoid regressions.

We are expecting some positive diffs here, from the IsLocalAddrExpr change (enabling more precise VNs) and the folding change itself (the old code path would, ever so rarely, miss out on some things).

We are also expecting one small regression due to call retyping of reg-sized IND struct(ADDR(LCL)) nodes losing the field sequence. I did not fix it because a) it was small, and rare, b) I do not want to permeate the zero-offset code any more than is strictly required.

Blocked on #66247.

Author:	SingleAccretion
Assignees:	-
Labels:	area-CodeGen-coreclr, community-contribution
Milestone:	-

[SingleAccretion]

@dotnet/jit-contrib

[kunalspathak]

Couple of comments.

[kunalspathak]

LGTM

[kunalspathak]

The tasks seems to be complete but not sure why it is not reflecting in GH. I will go ahead and merge this.