Forbid creation of non-faulting null-check nodes.

[sandreenko]

This is a follow-up for #44059.

There are 2 methods that create GT_NULLCHECK: gtNewNullCheck and gtChangeOperToNullCheck.
We have added an assert assert(fgAddrCouldBeNull(addr)) to gtNewNullCheck and it could be profitable to add the same assert to gtChangeOperToNullCheck to delete unnecessary null-checks earlier and sometimes allow the later phase to generate better code.

The difficulty is that gtChangeOperToNullCheck is used in different phases and in some of them it is not easy to delete an unnecessary node. An easier solution would be to replace the node with a nop and allow later phases to delete it also hit some asserts (prototype ),

for example in

runtime\artifacts\tests\coreclr\Windows_NT.x64.Checked\Tests\Core_Root\CoreRun.exe "jitutils\bin\pmi.dll" PrepOne  runtime\artifacts\tests\coreclr\Windows_NT.x64.Release\Tests\Core_Root\mscorlib.dll 5480

we try to hoist such expression:

               [000048] n----+------                 |  +--*  OBJ       simd32<System.Numerics.Vector`1[Single]> // note that `n` in the flags means non-faulting.
               [000044] -----+-N----                 |  |  \--*  LCL_VAR   byref  V01 arg1

and we create

               [000092] ----G-------              *  COMMA     void
     (  7,  5) [000089] n---G--H----              +--*  OBJ       simd32<System.Numerics.Vector`1[Single]> <l:$341, c:$382>
     (  1,  1) [000090] -------N----              |  \--*  LCL_VAR   byref  V01 arg1         u:1 $81
               [000091] ------------              \--*  NOP       void

and call gtExtractSideEffList on 000089, gtExtractSideEffList can't drop the tree because it is marked as GTF_MAKE_CSE so it will try to call gtChangeOperToNullCheck on an OBJ that can't fail.

The resulting tree with the current master is:

***** BB06
STMT00012 (IL   ???...  ???)
N004 (  2,  2) [000092] ----G-------              *  COMMA     void
N002 (  2,  2) [000089] ----G--H----              +--*  NULLCHECK int
N001 (  1,  1) [000090] -------N----              |  \--*  LCL_VAR   byref  V01 arg1         u:1 $81
N003 (  0,  0) [000091] ------------              \--*  NOP       void

it is not clear how would we like to transform it, because if the tree should be just deleted why did we choose this node for loop hoisting?

Note: I have run asm diffs with the prototype and have not seen asm diffs except for failing methods, so fixing this issue could be a clean-up only, rather than CQ, but it is not certain.

cc @AndyAyersMS

category:cq
theme:null-checks