views:

524

answers:

4

Hello everybody, this is my first message here.

Today I had a problem converting a Long (Int64) to an Integer (Int32). The problem is that my code was always working in 32-bit environments, but when I try THE SAME executable in a 64-bit computer it crashes with a System.OverflowException exception.

I've prepared this test code in VS2008 in a new project with default settings:

Module Module1

    Sub Main()
      Dim alpha As Long = -1
      Dim delta As Integer

      Try
         delta = CInt(alpha And UInteger.MaxValue)
         Console.WriteLine("CINT OK")
         delta = Convert.ToInt32(alpha And UInteger.MaxValue)
         Console.WriteLine("Convert.ToInt32 OK")
      Catch ex As Exception
         Console.WriteLine(ex.GetType().ToString())
      Finally
         Console.ReadLine()
      End Try
   End Sub

End Module

On my 32-bit setups (Windows XP SP3 32-bit and Windows 7 32-bit) it prints up to "CINT OK", but in the 64-bit computer (Windows 7 64-bit) that I've tested THE SAME executable it prints the exception name only.

Is this behavior documented? I tried to find a reference but failed miserably.

For reference I leave the MSIL code too:

.method public static void  Main() cil managed
{
  .entrypoint
  .custom instance void [mscorlib]System.STAThreadAttribute::.ctor() = ( 01 00 00 00 ) 
  // Code size       88 (0x58)
  .maxstack  2
  .locals init ([0] int64 alpha,
           [1] int32 delta,
           [2] class [mscorlib]System.Exception ex)
  IL_0000:  nop
  IL_0001:  ldc.i4.m1
  IL_0002:  conv.i8
  IL_0003:  stloc.0
  IL_0004:  nop
  .try
  {
    .try
    {
      IL_0005:  ldloc.0
      IL_0006:  ldc.i4.m1
      IL_0007:  conv.u8
      IL_0008:  and
      IL_0009:  conv.ovf.i4
      IL_000a:  stloc.1
      IL_000b:  ldstr      "CINT OK"
      IL_0010:  call       void [mscorlib]System.Console::WriteLine(string)
      IL_0015:  nop
      IL_0016:  ldloc.0
      IL_0017:  ldc.i4.m1
      IL_0018:  conv.u8
      IL_0019:  and
      IL_001a:  call       int32 [mscorlib]System.Convert::ToInt32(int64)
      IL_001f:  stloc.1
      IL_0020:  ldstr      "Convert.ToInt32 OK"
      IL_0025:  call       void [mscorlib]System.Console::WriteLine(string)
      IL_002a:  nop
      IL_002b:  leave.s    IL_0055
    }  // end .try
    catch [mscorlib]System.Exception 
    {
      IL_002d:  dup
      IL_002e:  call       void [Microsoft.VisualBasic]Microsoft.VisualBasic.CompilerServices.ProjectData::SetProjectError(class [mscorlib]System.Exception)
      IL_0033:  stloc.2
      IL_0034:  nop
      IL_0035:  ldloc.2
      IL_0036:  callvirt   instance class [mscorlib]System.Type [mscorlib]System.Exception::GetType()
      IL_003b:  callvirt   instance string [mscorlib]System.Type::ToString()
      IL_0040:  call       void [mscorlib]System.Console::WriteLine(string)
      IL_0045:  nop
      IL_0046:  call       void [Microsoft.VisualBasic]Microsoft.VisualBasic.CompilerServices.ProjectData::ClearProjectError()
      IL_004b:  leave.s    IL_0055
    }  // end handler
  }  // end .try
  finally
  {
    IL_004d:  nop
    IL_004e:  call       string [mscorlib]System.Console::ReadLine()
    IL_0053:  pop
    IL_0054:  endfinally
  }  // end handler
  IL_0055:  nop
  IL_0056:  nop
  IL_0057:  ret
} // end of method Module1::Main

I suspect that the instruction that is behaving differently is either conv.ovf.i4 or the ldc.i4.m1/conv.u8 pair.

If you know what is going on here please let me know

Thanks

PS: Forgot to mention, Convert.ToInt32(long) fails in both environment, it is only CInt(Long) which is behaving differently.

A: 

I don't know of any real reference as such, but if you go to this page:

http://msdn.microsoft.com/en-us/library/system.int32.aspx

You can see in the sample where they use CInt they do wrap it in a OverflowException handler (try searching for CInt on that page to find it). So at least they say implicitly that CInt can throw that in certain circumstances.

If you do not want the exceptions being thrown you can change the Remove integer overflow checks setting on the Advanced Compile Options page.

ho1
ho, yes CInt can throw an exception, VB.NET checks for overflow pretty much everything (it can be disabled, but besides the project-wide setting, I don't know of any method to disable at class-, method- or sentence-level).The problem here is that the same executable with the same inputs (constants here), runs perfectly in 32-bit environments but fails in 64-bit environments. Fortunately I solved this issue in my code by using UInteger instead so this error helped me to remove unneeded intermediate steps (I was trying to split a Long into two UIntegers)Thanks for your answer anyway.
LocoDelAssembly
A: 

Try to change build platform target from “Any CPU” to "x86".

Dmitri Kouminov
Now it prints "CINT OK" and "System.OverflowException" as in the 32-bit environments. With Any CPU and x64 it prints "System.OverflowException" only.The question is, why this happens? Shouldn't it behave the same way in both architectures?Thanks for the tip.
LocoDelAssembly
+2  A: 

Unfortunately, the 64-bit version is accurate. It really is an overflow, the result of the expression is a long with the value &hffffffff. The sign bit is AND-ed off the value, it is no longer negative. The resulting value cannot be converted to an integer, the maximum integer value is &h7fffffff. You can see this by adding this code to your snippet:

 Dim value As Long = alpha And UInteger.MaxValue
 Console.WriteLine(value)

Output: 4294967295

The x64 jitter uses an entirely different way to check for overflows, it doesn't rely on the CPU overflow exception but explicitly compares the values to Integer.MaxValue and Integer.MinValue. The x86 jitter gets it wrong, it optimizes the code too much and ends up making an unsigned operation that doesn't trip the CPU exception.

Filing a bug report at connect.microsoft.com is probably not worth the effort, fixing this for the x86 jitter would be a drastically breaking change. You'll have to rework this logic. Not sure how, I don't see what you are trying to do.

Hans Passant
Hans, many thanks for your answer. Yes, I did find strange that VB.NET was not complaining about the sign bit after ANDing but it did for other stuff (I was working in a Int64x64 to Int128 multiplication routine). Doing alpha = alpha And UInteger.MaxValue : delta = CInt(alpha) makes both architectures behave the same way (overflow). The problem fortunately is gone for me because now I split the Long parameters converting them to two UInteger straight (I really don't know why I was making the intermediate step).Thanks again!
LocoDelAssembly
Hans, I've submitted this issue to Microsoft Connect. Lets see what happens... You can see it here: http://connect.microsoft.com/VisualStudio/feedback/details/557734/conv-ovf-i4-behaves-different-in-x86-and-x64-architectures
LocoDelAssembly
I've just received an automated mail notifying me about the changes.This is what they say:[quote]We have confirmed that this is a bug in the CLR JIT. It will be fixed in a future release.Thank you for your report![/quote]
LocoDelAssembly
A: 

Just to complete the documentation of this issue I made this:

Imports System.Runtime.InteropServices

Module Module1
   <DllImport("KERNEL32.DLL", EntryPoint:="DebugBreak", _
        SetLastError:=False, CharSet:=CharSet.Unicode, _
        ExactSpelling:=True, _
        CallingConvention:=CallingConvention.StdCall)> _
        Public Sub DebugBreak()
   End Sub

   Sub Main()
      Dim alpha As Long = -1
      Dim delta As Integer

      DebugBreak() ' To call OllyDbg
      ' Needed to prevent the jitter from raising the overflow exception in the second CInt without really doing the convertion first
      alpha = alpha Xor Environment.TickCount
      Console.WriteLine(alpha)

      delta = CInt(alpha And UInteger.MaxValue)
      Console.WriteLine(delta)

      alpha = alpha And UInteger.MaxValue
      delta = CInt(alpha)
      Console.WriteLine(delta)

      Console.ReadLine()
   End Sub
End Module

Using OllyDbg I got this:

CPU Disasm
Address   Hex dump          Command                                  Comments
00D10070    55              PUSH EBP
00D10071    8BEC            MOV EBP,ESP
00D10073    57              PUSH EDI
00D10074    56              PUSH ESI
00D10075    53              PUSH EBX
00D10076    E8 A1BFC7FF     CALL 0098C01C
00D1007B    E8 A18C1879     CALL <JMP.&KERNEL32.GetTickCount>        ; Jump to KERNEL32.GetTickCount
00D10080    99              CDQ
00D10081    F7D0            NOT EAX
00D10083    F7D2            NOT EDX
00D10085    8BF0            MOV ESI,EAX
00D10087    8BFA            MOV EDI,EDX
00D10089    E8 62D25D78     CALL 792ED2F0                            ; Called everytime Console is referenced here
00D1008E    57              PUSH EDI
00D1008F    56              PUSH ESI
00D10090    8BC8            MOV ECX,EAX
00D10092    8B01            MOV EAX,DWORD PTR DS:[ECX]
00D10094    FF90 C4000000   CALL DWORD PTR DS:[EAX+0C4]              ; Console.WriteLine(Int64)
00D1009A    8BDE            MOV EBX,ESI                              ; Note: EDI:ESI holds alpha variable
00D1009C    83E3 FF         AND EBX,FFFFFFFF                         ; delta = CInt(alpha And UInteger.MaxValue)
00D1009F    E8 4CD25D78     CALL 792ED2F0
00D100A4    8BC8            MOV ECX,EAX
00D100A6    8BD3            MOV EDX,EBX
00D100A8    8B01            MOV EAX,DWORD PTR DS:[ECX]
00D100AA    FF90 BC000000   CALL DWORD PTR DS:[EAX+0BC]              ; Console.WriteLine(Int32)
00D100B0    33FF            XOR EDI,EDI                              ; alpha = alpha And UInteger.MaxValue
00D100B2    85F6            TEST ESI,ESI                             ; delta = CInt(alpha) [Begins here]
00D100B4    7C 06           JL SHORT 00D100BC
00D100B6    85FF            TEST EDI,EDI
00D100B8    75 2B           JNE SHORT 00D100E5
00D100BA    EB 05           JMP SHORT 00D100C1
00D100BC    83FF FF         CMP EDI,-1
00D100BF    75 24           JNE SHORT 00D100E5
00D100C1    8BDE            MOV EBX,ESI                              ; delta = CInt(alpha) [Ends here]
00D100C3    E8 28D25D78     CALL 792ED2F0
00D100C8    8BC8            MOV ECX,EAX
00D100CA    8BD3            MOV EDX,EBX
00D100CC    8B01            MOV EAX,DWORD PTR DS:[ECX]
00D100CE    FF90 BC000000   CALL DWORD PTR DS:[EAX+0BC]              ; Console.WriteLine(Int32)
00D100D4    E8 1B1AA878     CALL 79791AF4
00D100D9    8BC8            MOV ECX,EAX
00D100DB    8B01            MOV EAX,DWORD PTR DS:[ECX]
00D100DD    FF50 64         CALL DWORD PTR DS:[EAX+64]
00D100E0    5B              POP EBX
00D100E1    5E              POP ESI
00D100E2    5F              POP EDI
00D100E3    5D              POP EBP
00D100E4    C3              RETN

As you can see the second CInt sentence is much more complex than just ANDing (which it could actually be suppressed as EBX won't change and the EFLAGS are not consumed anywhere). The probable origin of this problem can be seen in Hans' answer

LocoDelAssembly