With 2.7 I encountered a bug where a simple function that returned a bool was always returning true. The function worked fine in 2.4. Looking into it appears that only the least significant byte of EAX was being set. So if EAX contained, say, 0xcccccccc and your function returned false it would now contain 0xcccccc00. RetQW then gets set to 0xcccccc00 as does eventually context->register1. The fix for this was to just copy over the last byte when returning a bool. Looking at the way the return values where set revealed another bug, the register isn't cleared before setting, so like EAX it may contain garbage since it's 64bit and typical return values are 32bit. Also, we must now check to see if we're returning a reference since GetSizeInMemoryBytes seems to return the size of the object being referred to (1 byte in the case of a bool) not the size of a reference (a 32bit pointer). Replacing the code around line 450 in as_callfunc_x86.cpp with the following fixes things:

	else
	{
		context->register1 = 0;
		// Store value in register1 register
		if( sysFunc->hostReturnFloat )
		{
			if( sysFunc->hostReturnSize == 1 )
				*(asDWORD*)&context->register1 = GetReturnedFloat();
			else
				context->register1 = GetReturnedDouble();
		}
		else if( descr->returnType.IsReference() )
		{	// pointer returned
			*(asDWORD*)&context->register1 = (asDWORD)retQW;
		}
		else if( descr->returnType.GetSizeInMemoryBytes() > 0 )
		{	// BUGFIX: if a bool is returned only the low byte of retQW
			// is set and the rest might be garbage so only take the
			// needed bytes over.
			if( descr->returnType.GetSizeInMemoryBytes() == 1 )
				*(asBYTE*)&context->register1 = (asBYTE)retQW;
			else if( descr->returnType.GetSizeInMemoryBytes() == 2 )
				*(asWORD*)&context->register1 = (asWORD)retQW;
			else// if( descr->returnType.GetSizeInMemoryBytes() == 4 )
				*(asDWORD*)&context->register1 = (asDWORD)retQW;
		}
		else
			context->register1 = retQW;

	}

No Deyja. That's not true. MSVC has very strict rules for type size and alignments. If this wasn't the case it wouldn't be able to properly link across translation units. You can change these rules using compiler switches or things like #pragma pack but the compiler doesn't just pick whatever it wants.

Also, using ints won't save you from the above bug because EAX/context->register1 is using a 64bit integer while an int only takes up 32bits (assuming a 32bit OS) so you'll still potentially have uncleared garbage.

Yeah I figured my fix wasn't the prettiest way to do it which is why I tried to explain exactly what was being fixed. That code is actually something that I backported from a PowerPC version of callfunc I got working (and I'll give you this week sometime - just need to clean up the file). It was needed to take care of some endian issues.

I did run your test_feature code on it to check if I'd broken anything else and all the tests did pass.