Fix VN of an ASG involving struct retyping.

[sandreenko]

Fixes #42517, replaces logic from dotnet/coreclr#24482 .

Diffs are small as expected:
x64 crossgen - no diffs;
x64 pmi - +113 (1 improved, 3 regressed);
arm64 crossgen - +8 (1 regressed);
arm64 pmi - +40 (1 improved, 3 regressed);
and similar for x86.

An example diff (from arm64 win pmi):

N005 ( 10, 10) [000097] -A------R---              *  ASG       struct (copy)
N004 (  6,  7) [000096] n-----------              +--*  BLK       struct<StackFrame, 16>
N003 (  3,  5) [000095] ------------              |  \--*  ADDR      byref 
N002 (  3,  4) [000084] U------N----              |     \--*  LCL_FLD   struct V07 tmp3         ud:2->3[+16] Fseq[_frame]
N001 (  3,  2) [000093] -------N----              \--*  LCL_VAR   struct<StackFrame, 16> V11 tmp7         u:1 (last use)

--- "a/D:\\Sergey\\logs\\PushTag\\old.txt"
+++ "b/D:\\Sergey\\logs\\PushTag\\new.txt"
@@ -5,7 +5,7 @@ 
 N003 (  3,  5) [000095] ------------              |  \--*  ADDR      byref
 N002 (  3,  4) [000084] U------N----              |     \--*  LCL_FLD   struct V07 tmp3         ud:2->3[+16] Fseq[_frame]
 N001 (  3,  2) [000093] -------N----              \--*  LCL_VAR   struct<StackFrame, 16> V11 tmp7         u:1 (last use)

-N001 [000093]   LCL_VAR   V11 tmp7         u:1 (last use) => $388 {388}
+N001 [000093]   LCL_VAR   V11 tmp7         u:1 (last use) => $387 {387}
   VNApplySelectors:
     VNForHandle(_frame) is $1ca, fieldType is struct, size = 16
       AX2: $1ca != $1c8 ==> select([$342]store($1, $1c8, $80), $1ca) ==> select($1, $1ca).
@@ -19,13 +19,13 @@
 N001 [000093]   LCL_VAR   V11 tmp7         u:1 (last use) => $388 {388}
 N002 [000084]   LCL_FLD   V07 tmp3         ud:2->3[+16] Fseq[_frame] => <l:$252 {252}, c:$253 {253}>
     FieldSeq {_frame} is $214
 N003 [000095]   ADDR      => $185 {PtrToLoc($45, $214)}
-  VNApplySelectorsAssign:
-    VNForHandle(_frame) is $1ca, fieldType is struct
-    VNForMapStore($342, $1ca, $388):struct returns $500 {$342[$1ca := $388]}
-  VNApplySelectorsAssign:
-    VNForHandle(_frame) is $1ca, fieldType is struct
-    VNForMapStore($342, $1ca, $388):struct returns $500 {$342[$1ca := $388]}
-Tree [000097] assigned VN to local var V07/3: $500 {$342[$1ca := $388]}
+Generate a unique VN for  on V11 because of struct reinterpretation
+
+
+
+
+
+Tree [000097] assigned VN to local var V07/3: new uniq $388 {388}
 N005 [000097]   ASG       => $VN.Void

 ***** BB03, STMT00021(after)
@@ -33,4 +33,4 @@ N005 ( 10, 10) [000097] -A------R---              *  ASG       struct (copy) $VN
 N004 (  6,  7) [000096] n-----------              +--*  BLK       struct<StackFrame, 16>
 N003 (  3,  5) [000095] ------------              |  \--*  ADDR      byref  $185
 N002 (  3,  4) [000084] U------N----              |     \--*  LCL_FLD   struct V07 tmp3         ud:2->3[+16] Fseq[_frame] <l:$252, c:$253>
-N001 (  3,  2) [000093] -------N----              \--*  LCL_VAR   struct<StackFrame, 16> V11 tmp7         u:1 (last use) $388
+N001 (  3,  2) [000093] -------N----              \--*  LCL_VAR   struct<StackFrame, 16> V11 tmp7         u:1 (last use) $387

it looks like this change is valid because if 000084 type was not <StackFrame, 16> it could confuse VN optimizations. For example, if the type was NotAStackFrame and we access its field using an VN from StackField we will get the same bug as this PR is fixing.

[sandreenko]

there were no overlapping fields there, the flag was used to mark the local var as "bad" for optimizations.

[briansull]

Reinterpreting a stuct of one type as a struct of a different type, essentially is the same as a C++ union where any and all fields of the structs overlap with each other.

[briansull]

So such as case is always "bad for optimizations"

[sandreenko]

This is not a C++ union because in this case we copy memory, so we don't address the same bytes using different handles.
What you are talking about in our model is represented as "address exposed" in my understanding.

[sandreenko]

for some static vars we are getting field seq that ends with a pseudo field, not sure why. It could be another issue.
x86.failure.txt

[sandreenko]

I am looking for a better way to handle this.

[briansull]

Pseudo fields are things like the first element of an array

[sandreenko]

PTAL @briansull, cc @dotnet/jit-contrib

[sandreenko]

I am trying to understand some VN principles and can't find documentation, the idea is described in JIT Architecture Plan (2009), but do we have description of the actual implementation?

[briansull]

These field maps was one of the last things added by Dave Detlefs before he left.
I don't know of additional documentation on this.