Allow more obj(addr(lcl_var) foldings.

[sandreenko]

Diffs on libraries:

x64 win pmi: -18832 (-0.037% of base)
arm64 win pmi: -6532 (-0.012% of base)
x64 unix pmi: -10302 (-0.020% of base)

x64 win crossgen: -1011 (-0.003% of base)
arm64 win crossgen: -352 (-0.001% of base)
x64 unix crossgen: -527 (-0.001% of base)

a different number of improved methods, no regression.

x64 win pmi of System.Net.Http benefits the most:

       -6867 : System.Net.Http.dasm (-0.932% of base)

a typical improvement comes from copy propagation when we can delete an unnecessary block copy:

@@ -1,16 +1,7 @@
-***** BB01
-STMT00003 (IL 0x000...  ???)
-N005 ( 13, 12) [000017] -A------R---              *  ASG       struct (copy)
-N004 (  9,  9) [000016] n-----------              +--*  OBJ       struct<Microsoft.CodeAnalysis.SyntaxTriviaList, 40>
-N003 (  3,  5) [000015] ------------              |  \--*  ADDR      byref
-N002 (  3,  4) [000012] D------N----              |     \--*  LCL_FLD   struct V03 tmp1         d:1[+0] Fseq[_list]
-N001 (  3,  2) [000013] ------------              \--*  LCL_VAR   struct<Microsoft.CodeAnalysis.SyntaxTriviaList, 40> V04 tmp2         u:1 (last use)
-
-***** BB01
 STMT00001 (IL 0x00B...  ???)
 N004 ( 11,  8) [000008] -A-X--------              *  ASG       struct (copy)
 N002 (  7,  5) [000007] D--X--------              +--*  OBJ       struct<Reversed, 40>
 N001 (  1,  1) [000006] ------------              |  \--*  LCL_VAR   byref  V01 RetBuf       u:1
-N003 (  3,  2) [000005] ------------              \--*  LCL_VAR   struct<Reversed, 40> V03 tmp1         u:1 (last use)
+N003 (  3,  2) [000005] ------------              \--*  LCL_VAR   struct<Microsoft.CodeAnalysis.SyntaxTriviaList, 40> V04 tmp2         u:1 (last use)

-Method code size: 136
\ No newline at end of file
+Method code size: 81
\ No newline at end of file

[sandreenko]

It is a continuation for #42320 in a preparation for arm64 apple changes.
PTAL @CarolEidt @dotnet/jit-contrib

[jashook]

Why did the asserts move down? Seems strange to allow clsHnd2 == NO_CLASS_HANDLE for the above checks

[sandreenko]

We could have a situation when we have GT_BLOCK <40>(GT_ADDR ref (LCL_VAR struct<40>)). GenTreeBlock nodes do not have a class handle, they represent a memory chunk. Such chunk is compatible with a struct if they have the same size and the struct does not have GC pointers.

[CarolEidt]

LGTM - thanks! I'm liking these incremental improvements!

[briansull]

Regresses libraries testing:

Issue #42517 Test failure: System.Buffers.Tests.ArrayBufferWriterTests_Char.Advance