@cyborgyn Is this operator valid negative? (Equals method is same as?)

since String implements IComparable, it's not a ReferenceEquals compare anymore defined on Object. It has 3 possible return values of -1, 0, 1, where 0 indicates match. In case the two strings xname and yname don't match, we do a second try, with explicit interface naming match. There were cases observed, where the Equals specifically was used, and even though it should have been match, it didn't get to the point to validate a parameter list, and returned false. The same logic was adapted to the sorting aid, Compare() method.

I look to seem this expression makes short circuit evaluation:

var rn = xname.CompareTo(yname);    // (a)
if (rn != 0)
{
    if (x.DeclaringType.IsInterface)
    {
          xname = $"{x.DeclaringType.FriendlyName}.{xname}";   // (b-1)
    }
    if (y.DeclaringType.IsInterface)
    {
          yname = $"{y.DeclaringType.FriendlyName}.{yname}";  // (b-2)
    }
    rn = xname.CompareTo(yname);   // (c)
    if (rn != 0)
    {
        // ...

At (a) xname value and yname value is different (not 0), and contains string of last part at (b-1) and (b-2), so final result at (c) is always different...? (always not 0)

Think about the following scenario:

namespace NS {
  interface Ib {
    void Method();
  }
  class A : Ib {
    void NS.Ib.Method() { ... }
  }
}

The trouble was, for the class MethodInformation.Name was "NS.Ib.Method()", but in the interface definition it was just "Method()". So when searching for interface implementation on the class, it didn't find it. One of them's DeclaringType is the class, the other's is the interface. So, the Equals() always returned false, and in the CallConverters.cs, at line 100 it throw Exception at this place:

                    var implementationMethod = allDeclaredMethods.First(
                        dm => MetadataUtilities.VirtualMethodSignatureComparer.Equals(dm, m));

So, what this thing does, is just give an other chance, putting the full "namespace.interfaceTypeName." before the appropriate method name (x, or y), and does an other compare. A side effect in the case for ordering, is that all interface implementation would be packed together after each other, I guess, but it is not used for things like this, as I am aware. In the case of Equals() there is no side effect.

I was insight it and finally understood what is problem for ;)

See branch feature/implement-ilcodes-r1 / 19bb2ea

There was nothing regression test for this code fragment, it's my badness....

The callvirt opcode devirtualization function try it a few conditions:

// Invoke interface method with the class-type instance and
// IL2C detected here for can cast statically.
else if (method.DeclaringType.IsInterface &&
     arg0.TargetType.IsClass &&                // (a)
     method.DeclaringType.IsAssignableFrom(arg0.TargetType))

At (a), this expression will test arg0 (this) is class type (analysis by IL2C), but this case couldn't be realize when using C# compiler. (Maybe) C# compiler never generate these conditions.

So I added some test written IL (b159b49, 65ebbbe). These test will cause making (a) condition with true and false. And I understood your explain.

Then I thought of a way to make your method better. In this case, we can only check for method matches by symbolic name, only methods on class/value type.

For example, if the F# compiler generate an explicit interface implementation method, the method names will not necessarily match. To make sure that the methods match exactly, check if they are included in IMethodInformation.Overrides property.

So, I made the following changes to make sure the unit test passes.

// Try strictly matching when both class/value type member method
// and explicitly implemented interface method are same.
if (!x.DeclaringType.IsInterface && y.DeclaringType.IsInterface)
{
    if (x.Overrides.Contains(y))
    {
        return true;
    }
}
if (x.DeclaringType.IsInterface && !y.DeclaringType.IsInterface)
{
    if (y.Overrides.Contains(x))
    {
        return true;
    }
}

@cyborgyn The IsInt32StackFriendlyType property means "Can place evaluation stack these types, that's compatible int32 type on CLR".

Therefore use IsInt32Type, IsInt64Type, IsIntPtrType and IsUIntPtrType. By the way, can use TargetType directly onto decodeContext.PushStack(...) argument.

We need regression test for not converter, require these topics:

• Apply not opcode each types int32 int64 and IntPtr.
• Asserts equality not values.

Thank you! This way, it will be much nicer. I am getting to know more and more of your code, and how parts are interacting with each other. I need to note, when I uncomment IntPtr unit tests, and also insert those into the IL, a UnitTestFramework fails to discover any tests, even though the produced DLL can be verified to be OK with DotPeek.

@cyborgyn Likely not opcode's integer type limitation.

ECMA-335: III.3.58 shl - shift integer left section is written:

The shl instruction shifts value (int32, int64 or native int) left by the number of bits
specified by shiftAmount. shiftAmount is of type int32 or native int. 

shiftAmount (si1) tricky specification, it's excepted int64.

We need regression test for shl and shr converter, require these topics:

• Apply shl/shr opcode each types int32 int64 and IntPtr combination with shiftAmount by int32 and IntPtr.
• Asserts equality constant values.

I have already added quite an amount of tests especially for shr & shr.un, as they were not that intuitive when they sign extend, and when not. I found this:
// Arithmetic Shift right (with sign)
// Warning!!! MSIL always operates for bit storage on Int32, or Int64
// and as it seems, smaller unsigned values (uint8 & uint16) will never
// be able to set the int32 highest bit => they will operate as logical shift right.
// On the other hand: uint32 and uint64 are still operated as arithmetic shift.

Will try to add the native int shift parameter as test, good point, thank you!

Unfortunately I needed to comment them out again.

OK, I will take a look on my environment, could you tell me explicit commit id reproducible?

Is it valid on newest commit on branch feature/implement-ilcodes (ca5706c) ?

Commented: #104 (comment)

@cyborgyn The test case result expression (TestCase attribute first argument) is object type, so (comparing) value can makes final cast operator.

Result with C# interactive window:

> (~(sbyte)0x12).GetType().FullName
"System.Int32"

Therefore:

[TestCase((sbyte)~(sbyte)0x12, "Not_int8", (sbyte)0x12)]

IntPtr couldn't make any constant value on .NET attribute. Because .NET attribute has to serialize argument into binary format, but IntPtr type is having to call constructor from any constant value.

In these cases, test framework will apply automatically conversion between IntPtr/UIntPtr and integer types at this then this.

Therefore:

[TestCase(~0, "not_intptr", 0)]

But IntPtr value be going to cause different result on runtime environment (test case on 32/64bit, native test is only 32bit gcc).

Unfortunately all test case disabled when IntPtr was used... I wrote this case:

// IL2C.BasicTypes.System_IntPtr
[TestCase(4, "SizeOf", Assert = TestCaseAsserts.IgnoreValidateInvokeResult)]    // Unit test environment is unknown, gcc is 32bit

IgnoreValidateInvokeResult flag will ignore regression test execution on CLR. Then test case result value can make from 32bit value.

@cyborgyn Ouch, I forgot signed integer flag :)

shr opcode will understand signed bit on MSB:

shr 10000010
result: 11000001

It case, we can use C language shift operator with signed integer type (int32_t, int64_t and intptr_t).
But shr.un opcode:

shr.un 10000010
result: 01000001

Strictly ignores MSB bit and insert zero. ECMA-335 III.3.60 shr.un - shift integer right, unsigned section:

shr.un inserts a zero bit on each shift.

On IL2C side, The C language expression is having to write temporary casting unsigned integers explicitly, like:

stack0 = (int32_t)(((uint32_t)stack0) >> shiftAmount);

We need regression test for shr.un (and shr) converter, require these topics:

• Positive and negative value makes valid shift result.

I did the temporary casting, if you look closely, and the tests are also walking through all of those tricky cases. One thing I forget, is negative shift (I never tried it, never even thought about it on any platform/language before). :) For the shift left, no magic happens, this is why it is omitted there.

Sorry, I found it on branch latest commit. (I took and read commit from bottom of branch...)

I approved it!

Bit off topic: The CLR (.NET Framework CLR and Core CLR) behavior is funny. Because shr and shr.un opcode behavior isn't defined with any int32, int64 and IntPtr on ECMA-335.

The shr.un instruction shifts value (int32, int 64 or native int) right by the number of bits
specified by shiftAmount.

But real regression test on CLR is passed...

@cyborgyn The 'neg' opcode calculus is serious with floating point type...

ECMA-335 III.3.50 neg - negate :

Negation of integral values is standard twos-complement negation. In particular, negating the
most negative number (which does not have a positive counterpart) yields the most negative
number. To detect this overflow use the sub.ovf instruction instead (i.e., subtract from 0).
Negating a floating-point number cannot overflow; negating NaN returns NaN.

If we write the same expression in C language, I feel that the result may not be the same between C and IL (C#). Also, the current IL2C doesn't consider NaN value at all, so it seems that it's time to think about it....

At this point, I think it's okay to include TODO comments and write with comment out regression test case.

I wrote ckfinite(float/double) opcode implementation in C to conform with these tests https://github.com/dotnet/runtime/blob/f9a2449/src/tests/JIT/IL_Conformance/Old/Base/ckfinite.il#L52.

Here is the C version: https://godbolt.org/z/vxMYrvb8d passing all conformance tests. coreclr uses the same approach by checking if all exponent bits are 1 to find out if float/double number is infinite or NaN, and emits ArithmeticException.

Was looking at the place to implement it in IL2C. Note: in this case, we need the raw single/double number in binary storage variable (e.g. uint32_t and uint64_t respectively) as it is represented in IL (IEEE 754 floating-point); without converting it to C float where it may or may not conform to IEEE 754 fp (ISO C99 is relaxed on the format...).

Feel free to implement ckfinite too, as this one is pretty simple. Otherwise, I'll take a stab at it; once this PR is merged. 🙂

Good news! I will include NaN handler and ckfinite implementation when after merged this PR. (Memoized: #112)

Thank you inserted details!