[Bug tree-optimization/56365] New: Missed opportunities for smin/smax standard name patterns

[Bug tree-optimization/56365] New: Missed opportunities for smin/smax standard name patterns

[olegendo at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

             Bug #: 56365
           Summary: Missed opportunities for smin/smax standard name
                    patterns
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: olegendo@gcc.gnu.org


While working on a patch for PR 55303 to add signed / unsigned clipping insns
for the SH2A target, I've noticed the following (tested with -O2 on 196091 for
SH and ARM cross configs):

int test_03 (int a, int b)
{
  int r = a + b;
  if (r > 127)
    r = 127;
  else if (r < -128)
    r = -128;
  return r;
}

This will utilize smin / smax standard name patterns.

The following equivalent (if I'm not mistaken), however:

static inline int min (int a, int b) { return a < b ? a : b; }
static inline int max (int a, int b) { return a < b ? b : a; }

int test_04 (int a, int b)
{
  return max (-128, min (127, a));
}

will not expand to smin / smax patterns.


Another case is:

int test_05 (int a)
{
  if (127 <= a)
    a = 127;
  else if (a <= -128)
    a = -128;
  return a;
}

For integers this could also be done with smin / smax, but it isn't.

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-02-18
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-18 10:58:13 UTC ---
I see, at -O2, on x86_64 in 070t.phiopt:

test_04 (int a, int b)
{
  int D.1744;
  int D.1741;
  int _3;
  int _4;

  <bb 2>:
  _3 = MIN_EXPR <a_1(D), 127>;
  _4 = MAX_EXPR <_3, -128>;
  return _4;

}

for the other cases you run into the issue that the tree-level phiopt
can be confused by phi-merging:

test_05 (int a)
{
  <bb 2>:
  if (a_2(D) > 126)
    goto <bb 5>;
  else
    goto <bb 3>;

  <bb 3>:
  if (a_2(D) < -127)
    goto <bb 5>;
  else
    goto <bb 4>;

  <bb 4>:

  <bb 5>:
  # a_1 = PHI <127(2), a_2(D)(4), -128(3)>
  return a_1;

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns

[olegendo at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

--- Comment #2 from Oleg Endo <olegendo at gcc dot gnu.org> 2013-02-18 20:04:40 UTC ---
(In reply to comment #1)
> I see, at -O2, on x86_64 in 070t.phiopt:
> 
> test_04 (int a, int b)
> {
>   int D.1744;
>   int D.1741;
>   int _3;
>   int _4;
> 
>   <bb 2>:
>   _3 = MIN_EXPR <a_1(D), 127>;
>   _4 = MAX_EXPR <_3, -128>;
>   return _4;
> 
> }

Ah yes, now I see that here, too.  I don't know where or how I was looking,
sorry.

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns when compiling as C++

[olegendo at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

Oleg Endo <olegendo at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Missed opportunities for    |Missed opportunities for
                   |smin/smax standard name     |smin/smax standard name
                   |patterns                    |patterns when compiling as
                   |                            |C++

--- Comment #3 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #2)
> (In reply to comment #1)
> > I see, at -O2, on x86_64 in 070t.phiopt:
> > 
> > test_04 (int a, int b)
> > {
> >   int D.1744;
> >   int D.1741;
> >   int _3;
> >   int _4;
> > 
> >   <bb 2>:
> >   _3 = MIN_EXPR <a_1(D), 127>;
> >   _4 = MAX_EXPR <_3, -128>;
> >   return _4;
> > 
> > }
> 
> Ah yes, now I see that here, too.  I don't know where or how I was looking,
> sorry.

If compiled as C (-x c -std=gnu99) the min/max patterns work.  Compiling as C++
(-x c++ -std=c++11) does not work.  Probably that was reason for my original
confusion -- usually I compile such small test snippets as C++ ...

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns when compiling as C++

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

--- Comment #4 from Oleg Endo <olegendo at gcc dot gnu.org> ---
As of r213381 this problem still exists.

compiled as C 003t.original:

;; Function min (null)
;; enabled by -tree-original


{
  return MIN_EXPR <b, a>;
}


;; Function max (null)
;; enabled by -tree-original


{
  return MAX_EXPR <a, b>;
}


;; Function test_04 (null)
;; enabled by -tree-original


{
  return max (-128, min (127, a));
}


compiled as C++ 003t.original:


;; Function int min(int, int) (null)
;; enabled by -tree-original


return <retval> = a < b ? a : b;


;; Function int max(int, int) (null)
;; enabled by -tree-original


return <retval> = a < b ? b : a;


;; Function int test_04(int, int) (null)
;; enabled by -tree-original


<<cleanup_point return <retval> = max (-128, min (127, a))>>;

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns when compiling as C++

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

Oleg Endo <olegendo at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org

--- Comment #5 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Richard, any chance this might get fixed with the match-and-simplify branch?

[Bug tree-optimization/56365] Missed opportunities for smin/smax standard name patterns when compiling as C++

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56365

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
No, that doesn't handle PHI nodes.  phiopt needs to be improved to handle
merged PHIs for this case.