Static field addresses are not deterministic when doing SPMI replay

[SingleAccretion]

Reproduction steps:

1. Construct a trivial app that accesses a static:

public class Program
{
    static int _global;

    [MethodImpl(MethodImplOptions.AggressiveOptimization)]
    private static int Main()
    {
        return _global;
    }
}

2. Do an SPMI collection for that app.
3. Add some printf logging for the static address' value to morph.cpp here (or use -jitoption JitDump=*, but it did not work for me for some reason).
4. Replay using superpmi.exe, twice, log the output to two files.
5. git diff them.
6. You should see something like the following:

-Static field address is 2902599013844, method is Main
+Static field address is 2220318337444, method is Main

This can cause spurious diffs, because CSE uses these values, as was seen in #53748. The commit I reproduced this on is 44f050a.

cc @BruceForstall

category:eng-sys
theme:super-pmi
skill-level:expert
cost:medium
impact:medium

[SingleAccretion]

Debugging it some more, it looks like the root cause is this code:

runtime/src/coreclr/ToolBox/superpmi/superpmi-shared/methodcontext.cpp

Lines 3592 to 3602 in 41af30c

	if (value.fieldValue != (DWORD)-1)
	{
	temp = (void*)GetFieldAddress->GetBuffer(value.fieldValue);
	cr->recAddressMap((void*)value.fieldAddress, temp, toCorInfoSize(repGetFieldType(field, nullptr, nullptr)));
	}
	else
	{
	temp = (void*)value.fieldAddress;
	}
	return temp;

SPMI tries to record values of static fields (all of them, even though only readonly ones are looked at by the Jit), so the value.fieldValue != (DWORD)-1 branch is hit practically always, and the changed temp returned.

This is necessary, since we must have the Jit look for the value now stored in the SPMI's map (replay_address), the original address will almost certainly make no sense in the replay process (in other words, dereferencing it will cause a segfault).

Looks like solving this might require extending the Jit-EE interface.

[BruceForstall]

I'll have to think about/debug this further. But typically what we would have to do is improve the recording side, in MethodContext::recGetFieldAddress().

[BruceForstall]

It seems like this is a problem in CSE for using the field address for any purpose that might cause diffs, not SuperPMI. E.g., it seems like multiple non-SuperPMI runs could also see asm diffs if the actual value of the address could affect codegen.

So I don't think this is a SuperPMI problem. SuperPMI does generally return the same addresses to the JIT as during record, assuming the JIT won't deref them, but things like the target code buffer address obviously must be able to change.