[Bug middle-end/35305] Speculative PRE support missing

xinliangli at gmail dot com gcc-bugzilla@gcc.gnu.org
Tue Oct 30 18:58:00 GMT 2012


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35305

--- Comment #4 from davidxl <xinliangli at gmail dot com> 2012-10-30 18:57:18 UTC ---
The suggested transformation can be useful in some cases, but won't be general
enough. The listed example is an extreme case. For instance, the second a+b
instance does not have to be in the hot trace but still still hot enough to be
PREed. Or there is no one dominating traces available -- the expression is
available in all incoming paths except for one rare path.

switch (a)
{
   case 1:
     g[1] = a + b;
     break;
   case 2:
     g[2] = a + b;
     break;
   case 3: 
     g[3] = a + b;
   ...
   default:
     g[0] = 0;
 }

switch (b)
{
   case 1:
     ... a + b; // partially redundant
   case 2:
      ... a + b; // redundant

   default:
     // does nothing
     break;
}


Regarding handling dereferences, the availability and down-safety analysis
needs to be extended to to recognize safe speculation candidates.


Example 1:  dereference of same pointer fully available -- 

int g1, g2;
struct A
{
  int a;
  int b;
};

void foo(struct A* ap, int k,int m)
{

   if (__builtin_expect (k, 1))
      g1 = ap->b;
   else
      g2 = ap->a;


   if (__builtin_expect (m, 1))
   {
       g2 = ap->b;        // Good safe speculative PRE candidate
   }
}


Example 2:  deference of ap fully anticipated



int g1, g2;
struct A
{
  int a;
  int b;
};

void foo(struct A* ap, int k,int m)
{

   if (__builtin_expect (k, 1))
      g1 = ap->b;


   if (__builtin_expect (m, 1))
       g2 = ap->b;             // Safe to speculatively hoist across the branch
into the else of the previous branch
   else
       g1 = ap->a;
}


David


(In reply to comment #3)
> Wouldn't this be a candidate for forming a superblock from hot traces of
> a function?  Thus in the testcase
> 
>   if (k && m)
>     {
>       g1 = a + b;
>       g2 = a + b;
>     }
>   else
>     {
> ... old code
>     }
> 
> which would also handle the case where we cannot speculatively move code
> (like dereferences)?



More information about the Gcc-bugs mailing list