[Bug tree-optimization/18768] New: Missed ivopts opportunity

[Bug tree-optimization/18768] New: Missed ivopts opportunity

[pthaugen at us dot ibm dot com]

Opening bug report per Zdeneck's request, snippets of email exchange follows:

=========
void f1 (void * coefPtr, double * dd)
{
  int  i,j;

  /* Cast of "coefPtr" results in poor code for this loop (missed strength
     reduction). */
  for (i = 0; i < 16; i++) {
    *(((double *) coefPtr) + i) = +0.0;
  }

  for (j = 0; j < 16; j++) {
    *(dd + j) = +0.0;
  }
}

=========

> this seems to be a problem with the cost function:
>
> Cost of strength reduction of the access =
>  Cost for incrementing the new induction variable:  8
>  Cost for increased register pressure: 4
>  Cost for the memory reference: 1
>
> Cost for expressing the access using i =
>  Cost for multiplication by 8:  12
>  Cost for the memory reference: 1 (addition of the result of
>    multiplications takes place in the address).
>
> The result is that it is not worthwhile to perform strength reduction.
> I will try to do something with the code that estimates cost of memory
> references, since it is quite wrong here.

could you please create a bugreport for this?  The things are a bit more
complicated than what I expected; there is actually no way how the
target could let ivopts know that DFmode address for (reg + reg) is more
expensive than just reg, in the current state.

-- 
           Summary: Missed ivopts opportunity
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pthaugen at us dot ibm dot com
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: powerpc64-linux
  GCC host triplet: powerpc64-linux
GCC target triplet: powerpc64-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768

[Bug tree-optimization/18768] Missed ivopts opportunity

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-12-01 21:22 -------
Confirmed, I think what is really is happening is not actually due with the cost of [reg +reg] but really 
with the additional code of "(double *) coefPtr" which we should really be zero.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2004-12-01 21:22:47
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768

[Bug tree-optimization/18768] Missed ivopts opportunity

[rakdver at gcc dot gnu dot org]

------- Additional Comments From rakdver at gcc dot gnu dot org  2004-12-01 21:44 -------
There is no additional cost counted for "(double *) coefPtr".  The reason why
the result with the cast is different is because PRE creates a (dead) phi node
when moving the cast out of the loop.  This phi node changes estimate of register
pressure used in ivopts, which leads to divergence -- we do not perform strength
reduction.

Which is a problem in DFmode, since [reg + reg] addressing for DF mode object
requires two extra additions:

[reg1 + reg2] = ...   /* lower half */
reg2+=4;
[reg1 + reg2] = ...   /* upper half */
reg2-=4;

Whereas with [reg] addressing mode, this would only be

[reg] = ...
[reg + 4] = ...

Of course when we clean up the unused phi nodes created by PRE, the regression
gets hidden (Daniel is working on this).  But the unrelying problem -- the fact
that we cannot tell that the [reg + reg] addressing for DF mode is expensive --
remains.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768

[Bug tree-optimization/18768] Missed ivopts opportunity

[pthaugen at us dot ibm dot com]

------- Additional Comments From pthaugen at us dot ibm dot com  2004-12-02 21:30 -------
Here's another example of the same thing, minus the cast/double issue. This one
is strength reduced for 3.4 but not for 4.0.

typedef struct {
  union {
    unsigned int Init[9];
    unsigned char Link[37];
  } u;
} BitStreamLinkStruct;
void f2(BitStreamLinkStruct *);

void f1 ()
{
  int  ixBlock;
  BitStreamLinkStruct bitStreamLink;

  for(ixBlock=0; ixBlock<=7; ixBlock++){
    bitStreamLink.u.Init[ixBlock] = 0x1e1e1e1e;
  }

  f2(&bitStreamLink);

}


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768

[Bug tree-optimization/18768] Missed ivopts opportunity

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-12-02 21:56 -------
(In reply to comment #3)
> Here's another example of the same thing, minus the cast/double issue. This one
> is strength reduced for 3.4 but not for 4.0.

I filed PR 18800 for that, it has to do with local arrays, it works with global arrays.
Thanks,
Andrew Pinski

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768

[Bug tree-optimization/18768] Missed ivopts opportunity

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-03-03 03:07 -------
This has no been fixed on the mainline.  I think PRE has changed to not to produce these dead phis.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18768