[Bug middle-end/110711] New: possible missed optimization for std::max with -march=znver2

[Bug middle-end/110711] New: possible missed optimization for std::max with -march=znver2

[mrks2023 at proton dot me]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

            Bug ID: 110711
           Summary: possible missed optimization for std::max with
                    -march=znver2
           Product: gcc
           Version: 13.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mrks2023 at proton dot me
  Target Milestone: ---

I think I found a missed optimization involving std::max() for -march=znver2
(sorry if it was already reported, but I didn't find anything related in the
bug tracker).

I have two functions that compute the maximum element of an array:

- function k_std_max uses std::max() and is never vectorized
- function k_max uses conditional assignment and is vectorized, when the
optimization flags allow for it

The code (also https://godbolt.org/z/hW49nbqMY):

#include <cassert>
#include <algorithm>

double k_std_max(size_t n_els, double * a)
{
    assert(n_els > 0);
    double m = a[0];
    #ifdef _OPENMP
        #pragma omp simd reduction(max:m)
    #endif
    for (size_t i = 1; i < n_els; ++i) {
        m = std::max(m, a[i]);
    }

    return m;
}

double k_max(size_t n_els, double * a)
{
    assert(n_els > 0);
    double m = a[0];
    #ifdef _OPENMP
        #pragma omp simd reduction(max:m)
    #endif
    for (size_t i = 1; i < n_els; ++i) {
        m = m < a[i] ? a[i] : m;
    }

    return m;
}

Compiling with "-O3 -fopenmp -march=znver2 -Wall -Wextra -DNDEBUG" vectorizes
k_max:

.L19:
        vmovupd ymm3, YMMWORD PTR [rax+8]
        add     rax, 32
        vmaxpd  ymm1, ymm3, ymm1
        cmp     rax, rdx
        jne     .L19

but for k_std_max still scalar instructions are used:

.L3:
        vmovsd  xmm0, QWORD PTR [rax]
        add     rax, 8
        vmaxsd  xmm0, xmm0, xmm1
        cmp     rdx, rax
        jne     .L5

Note that I had to use -fopenmp as using only -fopenmp-simd did not vectorize
k_max.

Even when I use "-Ofast" or "-Ofast -fopenmp" instead of "-O3" k_std_max is not
vectorized:

.L3:
        vmaxsd  xmm0, xmm0, QWORD PTR [rax]
        add     rax, 8
        cmp     rdx, rax
        jne     .L3

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

Hongtao.liu <crazylht at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |crazylht at gmail dot com

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
You need to use -ffast-math, w/o it, operands order matters for floating point
max/min, they're not commutative.

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #1)
> You need to use -ffast-math, w/o it, operands order matters for floating
> point max/min, they're not commutative.

Sorry, too fast to reply, ignore this comment.

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization, openmp

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Both are vectorized but somehow the OMP simd setup discards the vectorized
variant of k_std_max.  The .GOMP_SIMD_VF (simduid.3_12(D)) seems to be
statically zero?!

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-07-18
             Status|UNCONFIRMED                 |NEW

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
  <bb 18> [local count: 105119324]:
  _6 = .LOOP_VECTORIZED (1, 2);
  if (_6 != 0)
    goto <bb 19>; [100.00%]
  else
    goto <bb 20>; [100.00%]

  <bb 19> [local count: 105119324]:

  <bb 5> [local count: 955630227]:
  # i_21 = PHI <i_16(10), 1(19)>
  # prephitmp_19 = PHI <prephitmp_24(10), _1(19)>
  _2 = i_21 * 8;
  _3 = a_10(D) + _2;
  _17 = MEM[(const double &)_3];
  _29 = MAX_EXPR <_17, prephitmp_19>;
  prephitmp_24 = _29;
  i_16 = i_21 + 1;
  if (n_els_7(D) > i_16)
    goto <bb 10>; [89.00%]
  else
    goto <bb 13>; [11.00%]

  <bb 13> [local count: 105119324]:
  # prephitmp_15 = PHI <_29(5), prephitmp_27(16)>
  goto <bb 8>; [100.00%]

  <bb 10> [local count: 850510903]:
  goto <bb 5>; [100.00%]


test.C:8:26: note:   Analyze phi: prephitmp_19 = PHI <prephitmp_24(10), _1(19)>
test.C:8:26: note:   reduction path: prephitmp_24 _29 prephitmp_19 
test.C:8:26: note:   reduction: unknown pattern
test.C:8:26: missed:   Unknown def-use cycle pattern.
test.C:8:26: note:   === vect_determine_precisions ===

It looks like there's extra move generate by ifcvt which make vectorizer think
it's not a reduction. prephitmp_24 can be replaced with _29 since it's not used
elsewhere.

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

--- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
Looks like it's GCC13 regression, GCC12.3 successfully vectorizes k_std_max

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

--- Comment #6 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #5)
> Looks like it's GCC13 regression, GCC12.3 successfully vectorizes k_std_max

https://godbolt.org/z/6111MP354

[Bug middle-end/110711] possible missed optimization for std::max with -march=znver2

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110711

--- Comment #7 from Hongtao.liu <crazylht at gmail dot com> ---
git diff gcc12_171t.ifcvt gcc13.173.ifcvt

   <bb 5> [local count: 955630227]:
-  # i_22 = PHI <i_16(10), 1(19)>
+  # i_21 = PHI <i_16(10), 1(19)>
   # prephitmp_19 = PHI <prephitmp_24(10), _1(19)>
-  _2 = i_22 * 8;
+  _2 = i_21 * 8;
   _3 = a_10(D) + _2;
   _17 = MEM[(const double &)_3];
-  prephitmp_24 = MAX_EXPR <_17, prephitmp_19>;
-  i_16 = i_22 + 1;
+  _29 = MAX_EXPR <_17, prephitmp_19>;
+  prephitmp_24 = _29;
+  i_16 = i_21 + 1;
   if (n_els_7(D) > i_16)
     goto <bb 10>; [89.00%]
   else
     goto <bb 13>; [11.00%]

   <bb 13> [local count: 105119324]:
-  # prephitmp_13 = PHI <prephitmp_24(5), prephitmp_12(16)>
+  # prephitmp_15 = PHI <_29(5), prephitmp_27(16)>
   goto <bb 8>; [100.00%]