Precise liveness for LCL_FLD-s

[sandreenko]

For the following test:

 struct S
    {
        public int i1;
        public int i2;
        public int i3;
        public int i4;
        public int i5;
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    static void Test(S s)
    {}


    [MethodImpl(MethodImplOptions.NoInlining)]
    public static int Main()
    {
        S s;//= new S();
        s.i1 = 1;
        s.i2 = 2;
        s.i3 = 3;
        s.i4 = 4;
        s.i5 = 5;
        Test(s);
        return 100;
    }

we will have must-init flag on 's' because our liveness (in fgPerNodeLocalVarLiveness) does not see

STORE fld1
STORE fld2
STORE fld3
STORE fld4
STORE fld5

as a full def, so JIT thinks that s value is used before defined:

BB01
use def in out
{V00}
{V00 V03}
{V00}
{}
Var=Reg beg of BB01: none
Var=Reg end of BB01: none

so "conservative" liveness prevents some optimizations, like mentioned zero-init elimination, dead store removals etc.
dumpForIssue.txt

I don't see a cheap solution, maybe mark "good" structs (no holes, overlaps, not address taken, not too many fields) with a new flag and track their liveness by fields using some "pseudo" numbers in fgCurUseSet instead of the parent lcl.

Marking as future unless we see benchmarks where it could improve CQ, cc @dotnet/jit-contrib

category:cq
theme:liveness
skill-level:intermediate
cost:medium
impact:medium

[sandreenko]

@erozenfeld this is my analysis that we discussed for #48797

[jakobbotsch]

I think physical promotion closes this gap.