Intel® Fortran Compiler
Build applications that can scale for the future with optimized code designed for Intel® Xeon® and compatible processors.

IFX optimization bug suspected

mecej4
Black Belt
243 Views

The program listed below involves only integers, and is deterministic. When compiled with the IFX 20211123 build with /Od on Windows 11, the output is

863862222379

Ifort and other Fortran compilers give the same result. When  IFX is used with /O2, etc., the output is 

863862222378

The program does cause integer overflow, but the code is part of a random number generator, performing lots of shifts and bit operations, so the overflow should be handled as usual. 

 

Here is the program source.

Module sk64_m
   Implicit None
   Integer, Parameter :: i8 = selected_int_kind(18)
   Integer (Kind=i8) :: q = 5797281239641574547_i8, carry = 36243678541_i8

Contains

   subroutine refill()
      Implicit None
      Integer (Kind=i8) :: z

      z = ishft(ishft(q,41),-1) + ishft(ishft(q,39),-1) + ishft(carry,-1)
      carry = ishft(q, -23) + ishft(q, -25) + ishft(z, -63)
      print '(i21)',carry
      Return
   End subroutine refill

End Module sk64_m

Program tk64
   Use sk64_m
   call refill()
End Program tk64
0 Kudos
3 Replies
Steve_Lionel
Black Belt Retired Employee
230 Views

You can't rely on results that overflow - the compiler is allowed to make assumptions that won't happen.

cryptogram
Novice
203 Views

Since everything is known ahead of time,  it would be interesting to see how much code was removed by the optimization.

In theory, it could have precomputed the final value of carry, and just printed out the precomputed value.  The difference would

then be in how the compiler itself handled the computation rather than how it would have worked if all the computations were performed in compiled code.

jimdempseyatthecove
Black Belt
177 Views

Does (your version) of ifx generate the correct result with optimizations disabled?

The ishft should be treating the numbers as unsigned (even though Fortran only supports signed INTEGERs).

Look at the disassembly code to verify that the shift instructions are unsigned.

Secondly, you have a value (q) that is shifted left nn positions immediately followed by shift right 1 position. Check to see if the optimization contains both shifts as opposed to shifting left nn-1 positions. This are not the same as the correct way will always result in 0 in the msb position and the incorrect way will not.

Jim Dempsey

Reply