[Bug c++/96733] New: std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[Bug c++/96733] New: std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[soonts at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

            Bug ID: 96733
           Summary: std::clamp for floats and doubles produces worse code
                    than a combo of std::min / std::max
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: soonts at gmail dot com
  Target Milestone: ---

C++/17 introduced std::clamp in the standard library.

Very useful, but the code gcc generates for it is not optimal for doubles and
floats. Source code:

// This version is good in gcc, not so on vc++
double clamp_minmax( double x, double i, double ax )
{
    return std::min( std::max( x, i ), ax );
}

// This version should be identical but is not.
// gcc compiles to scalar comparisons
// vc++ code is outright horrible, compiles into RAM access.
double clamp_builtin( double x, double i, double ax )
{
    return std::clamp( x, i, ax );
} 

Demo there: https://godbolt.org/z/Gnvc1s

[Bug middle-end/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c++                         |middle-end
   Last reconfirmed|                            |2020-08-21
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Without the library dependency:

namespace std
{
  template<typename _Tp>
    constexpr const _Tp&
    min(const _Tp& __a, const _Tp& __b)
    {
      //return __b < __a ? __b : __a;
      if (__b < __a)
        return __b;
      return __a;
    }

  template<typename _Tp>
    constexpr const _Tp&
    max(const _Tp& __a, const _Tp& __b)
    {
      //return  __a < __b ? __b : __a;
      if (__a < __b)
        return __b;
      return __a;
    }

  template<typename _Tp>
    constexpr const _Tp&
    clamp(const _Tp& __val, const _Tp& __lo, const _Tp& __hi)
    {
      return (__val < __lo) ? __lo : (__hi < __val) ? __hi : __val;
    }
}

// This version is good in gcc, not so on vc++
double clamp_minmax( double x, double i, double ax )
{
    return std::min( std::max( x, i ), ax );
}

// This version should be identical but is not.
// gcc compiles to scalar comparisons
// vc++ code is outright horrible, compiles into RAM access.
double clamp_builtin( double x, double i, double ax )
{
    return std::clamp( x, i, ax );
}

[Bug middle-end/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug middle-end/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |redi at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I'm afraid there is nothing that the compiler can do to optimize the
clamp_builtin into similar code to clamp_minmax, because the optimizer doesn't
know an important restriction on std::clamp
- "The behavior is undefined if the value of lo is greater than hi".
In the IL this isn't expressed in any way, and when the compiler sees
(__val < __lo) ? __lo : (__hi < __val) ? __hi : __val
it can only optimize (for -Ofast) the (__hi < __val) ? __hi : __val part into
min (__hi, __val), but as it needs to assume e.g. __lo could be 10 and __hi 5,
it can't emit max.

[Bug middle-end/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Shouldn't this help in the library then?

      if (__hi < __lo)
        __builtin_unreachable();

[Bug middle-end/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Jonathan Wakely from comment #3)
> Shouldn't this help in the library then?
> 
>       if (__hi < __lo)
>         __builtin_unreachable();

For integers, I guess it could if Andrew's/Aldy's stuff gets in and can be used
for that, for floating point types, we don't really have any value range
handling, and without -ffast-math it might be even impossible to optimize away.
Can't std::clamp be implemented using std::min/std::max instead, or at least as
code that can be optimized into that?

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|middle-end                  |libstdc++

--- Comment #5 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Yes, it looks like we might want to do that then. We can have different code
for different types if needed.

Reassigning to libstdc++.

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vittorio.romeo at outlook dot com

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
*** Bug 100008 has been marked as a duplicate of this bug. ***

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

--- Comment #7 from Jonathan Wakely <redi at gcc dot gnu.org> ---
In my GCC fork I just switched clamp to use min and max unconditionally:
https://gitlab.com/jonathan-wakely/gcc/-/commit/ec5ecc2e142fbb7fc532a5ef9527b61675bcbfc4

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jonathan Wakely <redi@gcc.gnu.org>:

https://gcc.gnu.org/g:741c7350c08b0884689466867b6c9e711c7b109e

commit r12-4061-g741c7350c08b0884689466867b6c9e711c7b109e
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Sat Apr 17 22:34:09 2021 +0100

    libstdc++: Implement std::clamp with std::min and std::max [PR 96733]

    The compiler doesn't know about the precondition of std::clamp that
    (hi < lo) is false, and so can't optimize as well as we'd like. By using
    std::min and std::max we help the compiler.

    Signed-off-by: Jonathan Wakely <jwakely@redhat.com>

    libstdc++-v3/ChangeLog:

            PR libstdc++/96733
            * include/bits/stl_algo.h (clamp): Use std::min and std::max.

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #9 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Should be fixed for GCC 12.

[Bug libstdc++/96733] std::clamp for floats and doubles produces worse code than a combo of std::min / std::max

[cosiekvfj at o2 dot pl]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96733

Kacper <cosiekvfj at o2 dot pl> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |cosiekvfj at o2 dot pl

--- Comment #10 from Kacper <cosiekvfj at o2 dot pl> ---
Still not fixed?
https://godbolt.org/z/1ehf9EsEa