glibc/sysdeps/ieee754/flt-32/lgamma_negf.c
Szabolcs Nagy 347a5b592c math: Fix float conversion regressions with gcc-12 [BZ #28713]
Converting double precision constants to float is now affected by the
runtime dynamic rounding mode instead of being evaluated at compile
time with default rounding mode (except static object initializers).

This can change the computed result and cause performance regression.
The known correctness issues (increased ulp errors) are already fixed,
this patch fixes remaining cases of unnecessary runtime conversions.

Add float M_* macros to math.h as new GNU extension API.  To avoid
conversions the new M_* macros are used and instead of casting double
literals to float, use float literals (only required if the conversion
is inexact).

The patch was tested on aarch64 where the following symbols had new
spurious conversion instructions that got fixed:

  __clog10f
  __gammaf_r_finite@GLIBC_2.17
  __j0f_finite@GLIBC_2.17
  __j1f_finite@GLIBC_2.17
  __jnf_finite@GLIBC_2.17
  __kernel_casinhf
  __lgamma_negf
  __log1pf
  __y0f_finite@GLIBC_2.17
  __y1f_finite@GLIBC_2.17
  cacosf
  cacoshf
  casinhf
  catanf
  catanhf
  clogf
  gammaf_positive

Fixes bug 28713.

Reviewed-by: Paul Zimmermann <Paul.Zimmermann@inria.fr>
2022-01-10 14:27:17 +00:00

283 lines
7.9 KiB
C

/* lgammaf expanding around zeros.
Copyright (C) 2015-2022 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<https://www.gnu.org/licenses/>. */
#include <float.h>
#include <math.h>
#include <math-narrow-eval.h>
#include <math_private.h>
#include <fenv_private.h>
static const float lgamma_zeros[][2] =
{
{ -0x2.74ff94p+0f, 0x1.3fe0f2p-24f },
{ -0x2.bf682p+0f, -0x1.437b2p-24f },
{ -0x3.24c1b8p+0f, 0x6.c34cap-28f },
{ -0x3.f48e2cp+0f, 0x1.707a04p-24f },
{ -0x4.0a13ap+0f, 0x1.e99aap-24f },
{ -0x4.fdd5ep+0f, 0x1.64454p-24f },
{ -0x5.021a98p+0f, 0x2.03d248p-24f },
{ -0x5.ffa4cp+0f, 0x2.9b82fcp-24f },
{ -0x6.005ac8p+0f, -0x1.625f24p-24f },
{ -0x6.fff3p+0f, 0x2.251e44p-24f },
{ -0x7.000dp+0f, 0x8.48078p-28f },
{ -0x7.fffe6p+0f, 0x1.fa98c4p-28f },
{ -0x8.0001ap+0f, -0x1.459fcap-28f },
{ -0x8.ffffdp+0f, -0x1.c425e8p-24f },
{ -0x9.00003p+0f, 0x1.c44b82p-24f },
{ -0xap+0f, 0x4.9f942p-24f },
{ -0xap+0f, -0x4.9f93b8p-24f },
{ -0xbp+0f, 0x6.b9916p-28f },
{ -0xbp+0f, -0x6.b9915p-28f },
{ -0xcp+0f, 0x8.f76c8p-32f },
{ -0xcp+0f, -0x8.f76c7p-32f },
{ -0xdp+0f, 0xb.09231p-36f },
{ -0xdp+0f, -0xb.09231p-36f },
{ -0xep+0f, 0xc.9cba5p-40f },
{ -0xep+0f, -0xc.9cba5p-40f },
{ -0xfp+0f, 0xd.73f9fp-44f },
};
static const float e_hi = 0x2.b7e15p+0f, e_lo = 0x1.628aeep-24f;
/* Coefficients B_2k / 2k(2k-1) of x^-(2k-1) in Stirling's
approximation to lgamma function. */
static const float lgamma_coeff[] =
{
0x1.555556p-4f,
-0xb.60b61p-12f,
0x3.403404p-12f,
};
#define NCOEFF (sizeof (lgamma_coeff) / sizeof (lgamma_coeff[0]))
/* Polynomial approximations to (|gamma(x)|-1)(x-n)/(x-x0), where n is
the integer end-point of the half-integer interval containing x and
x0 is the zero of lgamma in that half-integer interval. Each
polynomial is expressed in terms of x-xm, where xm is the midpoint
of the interval for which the polynomial applies. */
static const float poly_coeff[] =
{
/* Interval [-2.125, -2] (polynomial degree 5). */
-0x1.0b71c6p+0f,
-0xc.73a1ep-4f,
-0x1.ec8462p-4f,
-0xe.37b93p-4f,
-0x1.02ed36p-4f,
-0xe.cbe26p-4f,
/* Interval [-2.25, -2.125] (polynomial degree 5). */
-0xf.29309p-4f,
-0xc.a5cfep-4f,
0x3.9c93fcp-4f,
-0x1.02a2fp+0f,
0x9.896bep-4f,
-0x1.519704p+0f,
/* Interval [-2.375, -2.25] (polynomial degree 5). */
-0xd.7d28dp-4f,
-0xe.6964cp-4f,
0xb.0d4f1p-4f,
-0x1.9240aep+0f,
0x1.dadabap+0f,
-0x3.1778c4p+0f,
/* Interval [-2.5, -2.375] (polynomial degree 6). */
-0xb.74ea2p-4f,
-0x1.2a82cp+0f,
0x1.880234p+0f,
-0x3.320c4p+0f,
0x5.572a38p+0f,
-0x9.f92bap+0f,
0x1.1c347ep+4f,
/* Interval [-2.625, -2.5] (polynomial degree 6). */
-0x3.d10108p-4f,
0x1.cd5584p+0f,
0x3.819c24p+0f,
0x6.84cbb8p+0f,
0xb.bf269p+0f,
0x1.57fb12p+4f,
0x2.7b9854p+4f,
/* Interval [-2.75, -2.625] (polynomial degree 6). */
-0x6.b5d25p-4f,
0x1.28d604p+0f,
0x1.db6526p+0f,
0x2.e20b38p+0f,
0x4.44c378p+0f,
0x6.62a08p+0f,
0x9.6db3ap+0f,
/* Interval [-2.875, -2.75] (polynomial degree 5). */
-0x8.a41b2p-4f,
0xc.da87fp-4f,
0x1.147312p+0f,
0x1.7617dap+0f,
0x1.d6c13p+0f,
0x2.57a358p+0f,
/* Interval [-3, -2.875] (polynomial degree 5). */
-0xa.046d6p-4f,
0x9.70b89p-4f,
0xa.a89a6p-4f,
0xd.2f2d8p-4f,
0xd.e32b4p-4f,
0xf.fb741p-4f,
};
static const size_t poly_deg[] =
{
5,
5,
5,
6,
6,
6,
5,
5,
};
static const size_t poly_end[] =
{
5,
11,
17,
24,
31,
38,
44,
50,
};
/* Compute sin (pi * X) for -0.25 <= X <= 0.5. */
static float
lg_sinpi (float x)
{
if (x <= 0.25f)
return __sinf (M_PIf * x);
else
return __cosf (M_PIf * (0.5f - x));
}
/* Compute cos (pi * X) for -0.25 <= X <= 0.5. */
static float
lg_cospi (float x)
{
if (x <= 0.25f)
return __cosf (M_PIf * x);
else
return __sinf (M_PIf * (0.5f - x));
}
/* Compute cot (pi * X) for -0.25 <= X <= 0.5. */
static float
lg_cotpi (float x)
{
return lg_cospi (x) / lg_sinpi (x);
}
/* Compute lgamma of a negative argument -15 < X < -2, setting
*SIGNGAMP accordingly. */
float
__lgamma_negf (float x, int *signgamp)
{
/* Determine the half-integer region X lies in, handle exact
integers and determine the sign of the result. */
int i = floorf (-2 * x);
if ((i & 1) == 0 && i == -2 * x)
return 1.0f / 0.0f;
float xn = ((i & 1) == 0 ? -i / 2 : (-i - 1) / 2);
i -= 4;
*signgamp = ((i & 2) == 0 ? -1 : 1);
SET_RESTORE_ROUNDF (FE_TONEAREST);
/* Expand around the zero X0 = X0_HI + X0_LO. */
float x0_hi = lgamma_zeros[i][0], x0_lo = lgamma_zeros[i][1];
float xdiff = x - x0_hi - x0_lo;
/* For arguments in the range -3 to -2, use polynomial
approximations to an adjusted version of the gamma function. */
if (i < 2)
{
int j = floorf (-8 * x) - 16;
float xm = (-33 - 2 * j) * 0.0625f;
float x_adj = x - xm;
size_t deg = poly_deg[j];
size_t end = poly_end[j];
float g = poly_coeff[end];
for (size_t j = 1; j <= deg; j++)
g = g * x_adj + poly_coeff[end - j];
return __log1pf (g * xdiff / (x - xn));
}
/* The result we want is log (sinpi (X0) / sinpi (X))
+ log (gamma (1 - X0) / gamma (1 - X)). */
float x_idiff = fabsf (xn - x), x0_idiff = fabsf (xn - x0_hi - x0_lo);
float log_sinpi_ratio;
if (x0_idiff < x_idiff * 0.5f)
/* Use log not log1p to avoid inaccuracy from log1p of arguments
close to -1. */
log_sinpi_ratio = __ieee754_logf (lg_sinpi (x0_idiff)
/ lg_sinpi (x_idiff));
else
{
/* Use log1p not log to avoid inaccuracy from log of arguments
close to 1. X0DIFF2 has positive sign if X0 is further from
XN than X is from XN, negative sign otherwise. */
float x0diff2 = ((i & 1) == 0 ? xdiff : -xdiff) * 0.5f;
float sx0d2 = lg_sinpi (x0diff2);
float cx0d2 = lg_cospi (x0diff2);
log_sinpi_ratio = __log1pf (2 * sx0d2
* (-sx0d2 + cx0d2 * lg_cotpi (x_idiff)));
}
float log_gamma_ratio;
float y0 = math_narrow_eval (1 - x0_hi);
float y0_eps = -x0_hi + (1 - y0) - x0_lo;
float y = math_narrow_eval (1 - x);
float y_eps = -x + (1 - y);
/* We now wish to compute LOG_GAMMA_RATIO
= log (gamma (Y0 + Y0_EPS) / gamma (Y + Y_EPS)). XDIFF
accurately approximates the difference Y0 + Y0_EPS - Y -
Y_EPS. Use Stirling's approximation. */
float log_gamma_high
= (xdiff * __log1pf ((y0 - e_hi - e_lo + y0_eps) / e_hi)
+ (y - 0.5f + y_eps) * __log1pf (xdiff / y));
/* Compute the sum of (B_2k / 2k(2k-1))(Y0^-(2k-1) - Y^-(2k-1)). */
float y0r = 1 / y0, yr = 1 / y;
float y0r2 = y0r * y0r, yr2 = yr * yr;
float rdiff = -xdiff / (y * y0);
float bterm[NCOEFF];
float dlast = rdiff, elast = rdiff * yr * (yr + y0r);
bterm[0] = dlast * lgamma_coeff[0];
for (size_t j = 1; j < NCOEFF; j++)
{
float dnext = dlast * y0r2 + elast;
float enext = elast * yr2;
bterm[j] = dnext * lgamma_coeff[j];
dlast = dnext;
elast = enext;
}
float log_gamma_low = 0;
for (size_t j = 0; j < NCOEFF; j++)
log_gamma_low += bterm[NCOEFF - 1 - j];
log_gamma_ratio = log_gamma_high + log_gamma_low;
return log_sinpi_ratio + log_gamma_ratio;
}