[WIP] Removes unnecessary shift count masking

[mikedn]

The C# compiler translates code like x << y to x << (y & 31) because the behavior of IL shift operations is unspecified if the shift count is greater than or equal to the bit width of the shifted value. However, x86/x64 shift instructions do mask the shift count so the & 31 is redundant.

Contributes to #1741

[mikedn]

FX diff:

Total bytes of diff: -515 (-0.01 % of base)
    diff is an improvement.
Total byte diff includes 0 bytes from reconciling methods
        Base had    0 unique methods,        0 unique bytes
        Diff had    0 unique methods,        0 unique bytes
Top file improvements by size (bytes):
        -109 : System.Private.CoreLib.dasm (0.00 % of base)
        -105 : System.Reflection.Metadata.dasm (-0.03 % of base)
         -79 : Microsoft.CodeAnalysis.dasm (-0.01 % of base)
         -74 : System.Runtime.Numerics.dasm (-0.13 % of base)
         -64 : System.Linq.Expressions.dasm (-0.01 % of base)
13 total files with size differences.
Top method regessions by size (bytes):
           4 : Microsoft.CodeAnalysis.dasm - Microsoft.CodeAnalysis.RealParser:RightShiftWithRounding(long,int,bool):long
           1 : Microsoft.CodeAnalysis.dasm - ArityEnumerator[__Canon][System.__Canon]:MoveNext():bool:this
           1 : System.Runtime.Numerics.dasm - System.Numerics.BigIntegerCalculator:Divide(long,int,long,int,long,int)
Top method improvements by size (bytes):
         -24 : Microsoft.CodeAnalysis.CSharp.dasm - Microsoft.CodeAnalysis.CSharp.Binder:FoldNeverOverflowBinaryOperators(int,ref,ref):ref
         -19 : System.Private.CoreLib.dasm - System.Text.UTF7Encoding:GetChars(long,int,long,int,ref):int:this
         -18 : Microsoft.CodeAnalysis.dasm - FloatingPointType:AssembleFloatingPointValue(long,int,bool,byref):int:this
         -18 : System.Runtime.Numerics.dasm - System.Numerics.BigInteger:op_RightShift(struct,int):struct
         -15 : System.Runtime.Numerics.dasm - System.Numerics.BigInteger:.ctor(double):this
120 total methods with size differences.

Considering that the eliminated and is usually 3 bytes long this means that there are probably some 170 shift count masking occurrences in the framework.

The few size regressions are a bit odd, it seems that in some cases the elimination of and somehow affected register allocation and we ended up with additional spills/reloads.

[mikedn]

Closing for now, I can't convince myself that this belongs in morph. Perhaps after #8117.