[Bug c++/107458] New: std::fma generates slow scalar-call

[Bug c++/107458] New: std::fma generates slow scalar-call

[g.peterhoff@t-online.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107458

            Bug ID: 107458
           Summary: std::fma generates slow scalar-call
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: g.peterhoff@t-online.de
  Target Milestone: ---

Please see https://godbolt.org/z/bxxc9ezeM

thx
Gero

[Bug c++/107458] std::fma generates slow scalar-call

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107458

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 53790
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53790&action=edit
testcase -march=x86-64-v2 -O3 -std=c++20 -Wall -mno-vzeroupper

Please attach the testcase next time instead of just linking to godbolt.

[Bug c++/107458] std::fma generates slow scalar-call

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107458

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
fmaf80_manually, fmaf64_manually, and fmaf32_manually are not exactly fma.

at least for x86-64-v2 due not doing the add before the rounding of the
multiply.

[Bug c++/107458] std::fma generates slow scalar-call

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107458

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

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

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Your definition of fma_manually is broken really.

The following is closer to being correct though there is still an extra
rounding step so it is still not 100%.

inline constexpr float fma_manually(const float a, const float b, const float
c)  noexcept
{
    double
        res = a;

    res *= b;
    res += c;
    return res;
}
inline constexpr double fma_manually(const double a, const double b, const
double 
 c)  noexcept
{
    long double
        res = a;

    res *= b;
    res += c;
    return res;
}

But GCC does the correct thing for fma and fmaf with x86-64-v3 and use the fma
instruction.

So closing as invalid.

[Bug c++/107458] std::fma generates slow scalar-call

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107458

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The C++ standard defines fma as doing (a*b) + c in infinite precision (and then
round), this is why GCC cannot inline if your target does not have a FMA
instruction.

There is still no bug here.