[Bug middle-end/53100] Optimize __int128 with range information

["marc.glisse at normalesup dot org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53100

--- Comment #3 from Marc Glisse <marc.glisse at normalesup dot org> 2012-05-01 12:47:03 UTC ---
(In reply to comment #2)
> and not to introduce them just before an optimization that removes them.

Usually, doing (long)num1*(__int128)(long)num2 does the right thing. I tried in
the example here replacing the plain __int128 multiplications with:

inline bool g1(__int128 x){
  //return(x<=LONG_MAX)&&(x>=LONG_MIN);
  //on 2 lines because of PR30318, unless you apply the patch I posted there
  bool b1 = x<=LONG_MAX;
  bool b2 = x>=LONG_MIN;
  return b1&&b2;
}
inline __int128 mul(__int128 a,__int128 b){
  bool B=g1(a)&&g1(b);
  if(__builtin_constant_p(B)&&B)
    return (long)a*(__int128)(long)b;
  return a*b;
}

__builtin_constant_p does detect we are in the right case, however, because of
bad timing between the various optimizations, the double cast
(__int128)(long)(u-x) is simplified to just (u-x) before it gets a chance to
help. I need to replace the subtraction instead (or in addition) to the
multiplication:

inline __int128 sub(__int128 a,__int128 b){
  bool B=g1(a)&&g1(b)&& g1(a-b);
  if(__builtin_constant_p(B)&&B)
    return (long)a-(long)b;
  return a-b;
}

But it would fit better inside the compiler than as a fragile use of
__builtin_constant_p.