Re: [PATCH] fmax(), fmaxf(), fmaxl(), fmin(), fminf(), fminl() simplified

["Stefan Kanthak" <stefan.kanthak@nexgo.de>]

"Szabolcs Nagy" <nsz@port70.net> wrote:


>* Stefan Kanthak <stefan.kanthak@nexgo.de> [2019-12-11 13:33:44 +0100]:
>> "Szabolcs Nagy" <nsz@port70.net> wrote:
>> >* Stefan Kanthak <stefan.kanthak@nexgo.de> [2019-12-11 10:55:29 +0100]:
>> > these two are not equivalent for snan input, but we dont care
>> > about snan, nor the compiler by default, so the compiler can
>> > optimize one to the other (although musl uses explicit int
>> > arithmetics instead of __builtin_isnan so it's a bit harder).
>> 
>> The latter behaviour was my reason to use (x != x) here: I attempt
>> to replace as many function calls as possible with "normal" code,
>> and also try to avoid transfers to/from FPU/SSE registers to/from
>> integer registers if that does not result in faster/shorter code.
> 
> why not just change the definition of isnan then?

Because I did not want to introduce such a global change; until now my
patches are just local (peephole) optimisations.

> #if __GNUC__ > xxx
> #define isnan(x) sizeof(x)==sizeof(float) ? __builtin_isnanf(x) : ...

This is better than my proposed change, as it also avoids the side-
effect of (x != x) which can raise exceptions, and gets rid of the
explicit transfer to integer registers, which can hurt performance.

The macros isinf(), isnormal(), isfinite(), signbit() should of
course be implemented in a similar way too, and the (internal only?)
functions __FLOAT_BITS() and __DOUBLE_BITS() removed completely!

regards
Stefan

PS: the following is just a "Gedankenspiel", extending the idea to
    avoid transfers from/to SSE registers.
    On x86-64, functions like isunordered(), copysign() etc. may be
    implemented using SSE intrinsics _mm_*() as follows:

#include <immintrin.h>

int signbit(double argument)
{
    return /* 1 & */ _mm_movemask_pd(_mm_set_sd(argument));
}

int isunordered(double a, double b)
{
#if 0
    return _mm_comieq_sd(_mm_cmp_sd(_mm_set_sd(a), _mm_set_sd(b), _CMP_ORD_Q), _mm_set_sd(0.0));
#elif 0
    return _mm_comineq_sd(_mm_set_sd(a), _mm_set_sd(a))
        || _mm_comineq_sd(_mm_set_sd(b), _mm_set_sd(b));
#else
    return /* 1 & */ _mm_movemask_pd(_mm_cmp_sd(_mm_set_sd(a), _mm_set_sd(b), _CMP_UNORD_Q));
#endif
}

uint32_t lrint(double argument)
{
    return _mm_cvtsd_si32(_mm_set_sd(argument));
}

uint64_t llrint(double argument)
{
    return _mm_cvtsd_si64(_mm_set_sd(argument));
}

double copysign(double magnitude, double sign)
{
    return _mm_cvtsd_f64(_mm_or_pd(_mm_and_pd(_mm_set_sd(-0.0), _mm_set_sd(sign)),
                                   _mm_andnot_pd(_mm_set_sd(-0.0), _mm_set_sd(magnitude))));
}

double fdim(double x, double y)
{
    return _mm_cvtsd_f64(_mm_and_pd(_mm_cmp_sd(_mm_set_sd(x), _mm_set_sd(y), _CMP_NLE_US),
                                    _mm_sub_sd(_mm_set_sd(x), _mm_set_sd(y))));
}

double fmax(double x, double y)
{
    __m128d mask = _mm_cmp_sd(_mm_set_sd(x), _mm_set_sd(x), _CMP_ORD_Q);

    return _mm_cvtsd_f64(_mm_or_pd(_mm_and_pd(mask, _mm_max_sd(_mm_set_sd(y), _mm_set_sd(x))),
                                   _mm_andnot_pd(mask, _mm_set_sd(y))));
}

double fmin(double x, double y)
{
    __m128d mask = _mm_cmp_sd(_mm_set_sd(x), _mm_set_sd(x), _CMP_ORD_Q);

    return _mm_cvtsd_f64(_mm_or_pd(_mm_and_pd(mask, _mm_min_sd(_mm_set_sd(y), _mm_set_sd(x))),
                                   _mm_andnot_pd(mask, _mm_set_sd(y))));
}

    Although the arguments and results are all held in SSE registers,
    there's no way to use them directly; it's but necessary to
    transfer them using _mm_set_sd() and _mm_cvtsd_f64(), which may
    result in superfluous instructions emitted by the compiler.

    If you but cheat and "hide" these functions from the compiler
    by placing them in a library, you can implement them as follows:

__m128d fmin(__m128d x, __m128d y)
{
    __m128d mask = _mm_cmp_sd(x, x, _CMP_ORD_Q);

    return _mm_or_pd(_mm_and_pd(mask, _mm_min_sd(y, x)),
                     _mm_andnot_pd(mask, y));
}

        .code   ; Intel syntax
fmin    proc    public
        movsd   xmm2, xmm0     ; xmm2 = x
        cmpsd   xmm2, xmm0, 7  ; xmm2 = (x != NAN) ? -1 : 0
        movsd   xmm3, xmm2
        andnpd  xmm3, xmm1     ; xmm3 = (x != NAN) ? 0.0 : y
        minsd   xmm1, xmm0     ; xmm1 = (x < y) ? x : y
                               ;      = min(x, y)
        andpd   xmm2, xmm1     ; xmm2 = (x != NAN) ? min(x, y) : 0.0
        orpd    xmm2, xmm3     ; xmm2 = (x != NAN) ? min(x, y) : y
        movsd   xmm0, xmm2     ; xmm0 = fmin(x, y)
        ret
fmin    endp