Invalid CSE with field seqs

[jakobbotsch]

// Generated by Fuzzlyn v1.1 on 2021-06-12 11:06:35
// Seed: 8276490119048877745
// Reduced from 962.5 KiB to 0.6 KiB in 00:05:32
// Debug: Outputs 0
// Release: Outputs 1
struct S0
{
    public ushort F1;
    public uint F2;
    public byte F3;
    public int F5;
    public S0(byte f3): this()
    {
        F3 = f3;
    }
}

struct S1
{
    public S0 F0;
    public S0 F4;
    public S1(S0 f0): this()
    {
        F0 = f0;
    }
}

struct S2
{
    public S1 F0;
    public S2(S1 f0): this()
    {
        F0 = f0;
    }
}

public class Program
{
    public static void Main()
    {
        S2 vr0 = new S2(new S1(new S0(1)));
        System.Console.WriteLine(vr0.F0.F0.F3);
        System.Console.WriteLine(vr0.F0.F4.F1);
    }
}

We end up CSE'ing the two field seqs and print 1 twice in release.
cc @dotnet/jit-contrib

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[SingleAccretion]

@sandreenko I've been looking at this a bit, here's a bit simpler reproduction:

public static void Problem()
{
    S2 s2;
    S1 s1;
    S0 s0;

    s1 = default;
    s0.F0_UByte = 1;
    s1.F0_S0 = s0;
    s2.F0_S1 = s1;

    Console.WriteLine(s2.F0_S1.F0_S0.F0_UByte);
}

struct S0
{
    public byte F0_UByte;
}

struct S1
{
    public S0 F0_S0;
    public byte Dummy;
}

struct S2
{
    public S1 F0_S1;
}

With JitNoStructPromotion=1.

And the bad VN:

N002 [000029]   LCL_FLD   V00 loc0         u:2[+0] Fseq[F0_S1, F0_S0, F0_UByte] (last use) => $41 {IntCns 0}

***** BB01, STMT00004(after)
N003 ( 18, 11) [000030] --CXG-------              *  CALL      void   System.Console.WriteLine $VN.Void
N002 (  4,  5) [000029] ------------ arg0 in rcx  \--*  LCL_FLD   ubyte  V00 loc0         u:2[+0] Fseq[F0_S1, F0_S0, F0_UByte] (last use) $41

[sandreenko]

Thanks @jakobbotsch and @SingleAccretion for the repros, I think the issue is similar to #49954, the root cause is in this optimization:

runtime/src/coreclr/jit/lclmorph.cpp

Line 1016 in 4aa29f4

// We're not going to produce a TYP_STRUCT LCL_FLD so we don't need the field sequence.

where we can generate S2 = S1 assignments. The closer we get to 6.0 the more I am thinking about disabling it and accept several kb regressions because my former tries to fix it were unsuccessful.

[jakobbotsch]

Not sure if this one is related:

// Generated by Fuzzlyn v1.2 on 2021-07-06 09:46:44
// Seed: 16635934940619066544
// Reduced from 447.5 KiB to 0.6 KiB in 00:02:24
// Debug: Runs successfully
// Release: Throws 'System.NullReferenceException'
struct S0
{
    public uint F1;
    public byte F3;
    public long F4;
    public uint F5;
    public S0(long f4): this()
    {
        F4 = f4;
    }
}

class C0
{
    public S0 F4;
}

struct S1
{
    public C0 F2;
    public S0 F8;
    public S1(C0 f2, S0 f8) : this()
    {
        F2 = f2;
        F8 = f8;
    }
}

struct S2
{
    public S1 F0;
    public S2(S1 f0): this()
    {
        F0 = f0;
    }
}

public class Program
{
    public static void Main()
    {
        S2 vr0 = new S2(new S1(new C0(), new S0(0)));
        M17(ref vr0.F0.F2.F4.F1);
    }

    static void M17(ref uint arg2)
    {
    }
}

[briansull]

Triage Area: CSE