[Bug target/100021] New: [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[Bug target/100021] New: [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[nok.raven at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100021

            Bug ID: 100021
           Summary: [9/10/11 Regression] std::clamp unprofitable
                    vectorization on -march=nehalem/.../broadwell
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: nok.raven at gmail dot com
  Target Milestone: ---

//#include <algorithm>
namespace std
{
    template<typename _Tp>
    constexpr const _Tp&
    clamp(const _Tp& __val, const _Tp& __lo, const _Tp& __hi)
    {
      return (__val < __lo) ? __lo : (__hi < __val) ? __hi : __val;
    }
}

int foo(int x, int y) {
    return std::clamp(x - y, -1, 1);
}

https://godbolt.org/z/6534c1ff1 either GCC makes unprofitable vectorization or
LLVM MCA calculation is wrong.

Affected targets are
-march=nehalem/westmere/sandybridge/ivybridge/haswell/broadwell.

I do not know why it is vectorized on -O2 in the first place; could not find a
switch which triggers it, and doing -O1 + everything -O2 supposed to be
enabling does not reproduce the -O2 behavior.

[Bug target/100021] [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100021

--- Comment #1 from Uroš Bizjak <ubizjak at gmail dot com> ---
This is not vectorization, but the compiler uses vector registers to perform
scalar operations. This is STV (scalar-to-vector) pass in action, you can use
-mno-stv to avoid transformation.

The transformation is used to avoid CMOV instruction.

[Bug target/100021] [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100021

--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
Also, you are passing -march=sandybridge, but the profiler seems to show
Skylake (SKX) target. The STV pass heavily depends on target costs, and when
-march=skylake is passed, the conversion is avoided.

[Bug target/100021] [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100021

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
             Target|                            |x86_64-*-* i?86-*-*
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-04-12

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'm quite sure the LLVM MCA calculation is "wrong".  The transform speeds up a
SPEC benchmark by >10% on a Haswell CPU.

So please show an actual runtime testcase that regresses.

[Bug target/100021] [9/10/11 Regression] std::clamp unprofitable vectorization on -march=nehalem/.../broadwell

[nok.raven at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100021

Nikita Kniazev <nok.raven at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|WAITING                     |RESOLVED

--- Comment #4 from Nikita Kniazev <nok.raven at gmail dot com> ---
> Also, you are passing -march=sandybridge, but the profiler seems to show Skylake (SKX) target.

I indeed missed that Compiler Explorer does not pass -march flag to MCA
automatically.