bitfiddling decompilation problem

[tamlin-mike]

Expected fnFails to produce the same decompiled code as fnWorks. Only difference is the size of the constants, i.e. ldc.i4+conv.i8 vs ldc.i8.

Not sure what produces an IL i8 less than 2^31 (I tried marking a constant with 'L' suffix, but the compiler seemingly disregarded that), but that makes ilspy fail to include the required argument casts to long.

.method public hidebysig static int64 fnWorks (int32 x, int32 y) cil managed 
{
  .maxstack 8

  ldarg.1
  conv.i8
  ldc.i4 16777215
  conv.i8
  and
  ldc.i4.s 24
  shl
  ldarg.0
  conv.i8
  ldc.i4 16777215
  conv.i8
  and
  or
  ret
}
decompiled OK: return (((long)y & 16777215L) << 24) | ((long)x & 16777215L);


.method public hidebysig static int64 fnFails (int32 x, int32 y) cil managed 
{
  .maxstack 8

  ldarg.1
  conv.i8
  ldc.i8 16777215
  and
  ldc.i4.s 24
  shl
  ldarg.0
  conv.i8
  ldc.i8 16777215
  and
  or
  ret
}
decompiled bad: return ((y & 0xFFFFFF) << 24) | (x & 0xFFFFFF);

Another minor detail here, is that the failing case properly recognize that the constants are of form 2^n-1 and turns them into hex, but the working case (probably due to the conv.i8) does not.

[dgrunwald]

Fixed in ad5ba92.