[glibc.git] / sysdeps / ieee754 / dbl-64 / s_expm1.c

/* @(#)s_expm1.c 5.1 93/09/24 */
/*
 * ====================================================
 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
 *
 * Developed at SunPro, a Sun Microsystems, Inc. business.
 * Permission to use, copy, modify, and distribute this
 * software is freely granted, provided that this notice
 * is preserved.
 * ====================================================
 */
/* Modified by Naohiko Shimizu/Tokai University, Japan 1997/08/25,
   for performance improvement on pipelined processors.
 */

/* expm1(x)
 * Returns exp(x)-1, the exponential of x minus 1.
 *
 * Method
 *   1. Argument reduction:
 *	Given x, find r and integer k such that
 *
 *               x = k*ln2 + r,  |r| <= 0.5*ln2 ~ 0.34658
 *
 *      Here a correction term c will be computed to compensate
 *	the error in r when rounded to a floating-point number.
 *
 *   2. Approximating expm1(r) by a special rational function on
 *	the interval [0,0.34658]:
 *	Since
 *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
 *	we define R1(r*r) by
 *	    r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 * R1(r*r)
 *	That is,
 *	    R1(r**2) = 6/r *((exp(r)+1)/(exp(r)-1) - 2/r)
 *		     = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
 *		     = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
 *      We use a special Reme algorithm on [0,0.347] to generate
 *	a polynomial of degree 5 in r*r to approximate R1. The
 *	maximum error of this polynomial approximation is bounded
 *	by 2**-61. In other words,
 *	    R1(z) ~ 1.0 + Q1*z + Q2*z**2 + Q3*z**3 + Q4*z**4 + Q5*z**5
 *	where	Q1  =  -1.6666666666666567384E-2,
 *		Q2  =   3.9682539681370365873E-4,
 *		Q3  =  -9.9206344733435987357E-6,
 *		Q4  =   2.5051361420808517002E-7,
 *		Q5  =  -6.2843505682382617102E-9;
 *	(where z=r*r, and the values of Q1 to Q5 are listed below)
 *	with error bounded by
 *	    |                  5           |     -61
 *	    | 1.0+Q1*z+...+Q5*z   -  R1(z) | <= 2
 *	    |                              |
 *
 *	expm1(r) = exp(r)-1 is then computed by the following
 *	specific way which minimize the accumulation rounding error:
 *			       2     3
 *			      r     r    [ 3 - (R1 + R1*r/2)  ]
 *	      expm1(r) = r + --- + --- * [--------------------]
 *			      2     2    [ 6 - r*(3 - R1*r/2) ]
 *
 *	To compensate the error in the argument reduction, we use
 *		expm1(r+c) = expm1(r) + c + expm1(r)*c
 *			   ~ expm1(r) + c + r*c
 *	Thus c+r*c will be added in as the correction terms for
 *	expm1(r+c). Now rearrange the term to avoid optimization
 *	screw up:
 *			(      2                                    2 )
 *			({  ( r    [ R1 -  (3 - R1*r/2) ]  )  }    r  )
 *	 expm1(r+c)~r - ({r*(--- * [--------------------]-c)-c} - --- )
 *			({  ( 2    [ 6 - r*(3 - R1*r/2) ]  )  }    2  )
 *                      (                                             )
 *
 *		   = r - E
 *   3. Scale back to obtain expm1(x):
 *	From step 1, we have
 *	   expm1(x) = either 2^k*[expm1(r)+1] - 1
 *		    = or     2^k*[expm1(r) + (1-2^-k)]
 *   4. Implementation notes:
 *	(A). To save one multiplication, we scale the coefficient Qi
 *	     to Qi*2^i, and replace z by (x^2)/2.
 *	(B). To achieve maximum accuracy, we compute expm1(x) by
 *	  (i)   if x < -56*ln2, return -1.0, (raise inexact if x!=inf)
 *	  (ii)  if k=0, return r-E
 *	  (iii) if k=-1, return 0.5*(r-E)-0.5
 *        (iv)	if k=1 if r < -0.25, return 2*((r+0.5)- E)
 *		       else	     return  1.0+2.0*(r-E);
 *	  (v)   if (k<-2||k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
 *	  (vi)  if k <= 20, return 2^k((1-2^-k)-(E-r)), else
 *	  (vii) return 2^k(1-((E+2^-k)-r))
 *
 * Special cases:
 *	expm1(INF) is INF, expm1(NaN) is NaN;
 *	expm1(-INF) is -1, and
 *	for finite argument, only expm1(0)=0 is exact.
 *
 * Accuracy:
 *	according to an error analysis, the error is always less than
 *	1 ulp (unit in the last place).
 *
 * Misc. info.
 *	For IEEE double
 *	    if x >  7.09782712893383973096e+02 then expm1(x) overflow
 *
 * Constants:
 * The hexadecimal values are the intended ones for the following
 * constants. The decimal values may be used, provided that the
 * compiler will convert from decimal to binary accurately enough
 * to produce the hexadecimal values shown.
 */

#include <errno.h>
#include <float.h>
#include <math.h>
#include <math-barriers.h>
#include <math_private.h>
#include <math-underflow.h>
#include <libm-alias-double.h>
#define one Q[0]
static const double
  huge = 1.0e+300,
  tiny = 1.0e-300,
  o_threshold = 7.09782712893383973096e+02,  /* 0x40862E42, 0xFEFA39EF */
  ln2_hi = 6.93147180369123816490e-01,       /* 0x3fe62e42, 0xfee00000 */
  ln2_lo = 1.90821492927058770002e-10,       /* 0x3dea39ef, 0x35793c76 */
  invln2 = 1.44269504088896338700e+00,       /* 0x3ff71547, 0x652b82fe */
/* scaled coefficients related to expm1 */
  Q[] = { 1.0, -3.33333333333331316428e-02, /* BFA11111 111110F4 */
	  1.58730158725481460165e-03, /* 3F5A01A0 19FE5585 */
	  -7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
	  4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
	  -2.01099218183624371326e-07 }; /* BE8AFDB7 6E09C32D */

#ifndef SECTION
# define SECTION
#endif

SECTION
double
__expm1 (double x)
{
  double y, hi, lo, c, t, e, hxs, hfx, r1, h2, h4, R1, R2, R3;
  int32_t k, xsb;
  uint32_t hx;

  GET_HIGH_WORD (hx, x);
  xsb = hx & 0x80000000;                /* sign bit of x */
  if (xsb == 0)
    y = x;
  else
    y = -x;                             /* y = |x| */
  hx &= 0x7fffffff;                     /* high word of |x| */

  /* filter out huge and non-finite argument */
  if (hx >= 0x4043687A)                         /* if |x|>=56*ln2 */
    {
      if (hx >= 0x40862E42)                     /* if |x|>=709.78... */
	{
	  if (hx >= 0x7ff00000)
	    {
	      uint32_t low;
	      GET_LOW_WORD (low, x);
	      if (((hx & 0xfffff) | low) != 0)
		return x + x;            /* NaN */
	      else
		return (xsb == 0) ? x : -1.0;    /* exp(+-inf)={inf,-1} */
	    }
	  if (x > o_threshold)
	    {
	      __set_errno (ERANGE);
	      return huge * huge;   /* overflow */
	    }
	}
      if (xsb != 0)      /* x < -56*ln2, return -1.0 with inexact */
	{
	  math_force_eval (x + tiny);           /* raise inexact */
	  return tiny - one;            /* return -1 */
	}
    }

  /* argument reduction */
  if (hx > 0x3fd62e42)                  /* if  |x| > 0.5 ln2 */
    {
      if (hx < 0x3FF0A2B2)              /* and |x| < 1.5 ln2 */
	{
	  if (xsb == 0)
	    {
	      hi = x - ln2_hi; lo = ln2_lo;  k = 1;
	    }
	  else
	    {
	      hi = x + ln2_hi; lo = -ln2_lo;  k = -1;
	    }
	}
      else
	{
	  k = invln2 * x + ((xsb == 0) ? 0.5 : -0.5);
	  t = k;
	  hi = x - t * ln2_hi;          /* t*ln2_hi is exact here */
	  lo = t * ln2_lo;
	}
      x = hi - lo;
      c = (hi - x) - lo;
    }
  else if (hx < 0x3c900000)             /* when |x|<2**-54, return x */
    {
      math_check_force_underflow (x);
      t = huge + x;     /* return x with inexact flags when x!=0 */
      return x - (t - (huge + x));
    }
  else
    k = 0;

  /* x is now in primary range */
  hfx = 0.5 * x;
  hxs = x * hfx;
  R1 = one + hxs * Q[1]; h2 = hxs * hxs;
  R2 = Q[2] + hxs * Q[3]; h4 = h2 * h2;
  R3 = Q[4] + hxs * Q[5];
  r1 = R1 + h2 * R2 + h4 * R3;
  t = 3.0 - r1 * hfx;
  e = hxs * ((r1 - t) / (6.0 - x * t));
  if (k == 0)
    return x - (x * e - hxs);                   /* c is 0 */
  else
    {
      e = (x * (e - c) - c);
      e -= hxs;
      if (k == -1)
	return 0.5 * (x - e) - 0.5;
      if (k == 1)
	{
	  if (x < -0.25)
	    return -2.0 * (e - (x + 0.5));
	  else
	    return one + 2.0 * (x - e);
	}
      if (k <= -2 || k > 56)         /* suffice to return exp(x)-1 */
	{
	  uint32_t high;
	  y = one - (e - x);
	  GET_HIGH_WORD (high, y);
	  SET_HIGH_WORD (y, high + (k << 20));  /* add k to y's exponent */
	  return y - one;
	}
      t = one;
      if (k < 20)
	{
	  uint32_t high;
	  SET_HIGH_WORD (t, 0x3ff00000 - (0x200000 >> k));    /* t=1-2^-k */
	  y = t - (e - x);
	  GET_HIGH_WORD (high, y);
	  SET_HIGH_WORD (y, high + (k << 20));  /* add k to y's exponent */
	}
      else
	{
	  uint32_t high;
	  SET_HIGH_WORD (t, ((0x3ff - k) << 20));       /* 2^-k */
	  y = x - (e + t);
	  y += one;
	  GET_HIGH_WORD (high, y);
	  SET_HIGH_WORD (y, high + (k << 20));  /* add k to y's exponent */
	}
    }
  return y;
}
#ifndef __expm1
libm_alias_double (__expm1, expm1)
#endif
Commit	Line	Data
f7eac6eb RM	1	/* @(#)s_expm1.c 5.1 93/09/24 */
	2	/*
	3	* ====================================================
	4	* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
	5	*
	6	* Developed at SunPro, a Sun Microsystems, Inc. business.
	7	* Permission to use, copy, modify, and distribute this
cccda09f	8	* software is freely granted, provided that this notice
f7eac6eb RM	9	* is preserved.
	10	* ====================================================
	11	*/
923609d1 UD	12	/* Modified by Naohiko Shimizu/Tokai University, Japan 1997/08/25,
923609d1 UD	13	for performance improvement on pipelined processors.
c5d5d574	14	*/
f7eac6eb	15
f7eac6eb RM	16	/* expm1(x)
	17	* Returns exp(x)-1, the exponential of x minus 1.
	18	*
	19	* Method
	20	* 1. Argument reduction:
	21	* Given x, find r and integer k such that
	22	*
cccda09f	23	* x = kln2 + r, \|r\| <= 0.5ln2 ~ 0.34658
f7eac6eb	24	*
cccda09f	25	* Here a correction term c will be computed to compensate
f7eac6eb RM	26	* the error in r when rounded to a floating-point number.
	27	*
	28	* 2. Approximating expm1(r) by a special rational function on
	29	* the interval [0,0.34658]:
	30	* Since
	31	* r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
	32	* we define R1(r*r) by
	33	* r(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 R1(r*r)
	34	* That is,
	35	* R1(r*2) = 6/r ((exp(r)+1)/(exp(r)-1) - 2/r)
	36	* = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
	37	* = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
cccda09f	38	* We use a special Reme algorithm on [0,0.347] to generate
d7826aa1	39	* a polynomial of degree 5 in r*r to approximate R1. The
cccda09f	40	* maximum error of this polynomial approximation is bounded
f7eac6eb RM	41	* by 2**-61. In other words,
f7eac6eb RM	42	* R1(z) ~ 1.0 + Q1z + Q2z*2 + Q3z*3 + Q4z*4 + Q5z**5
d7826aa1 UD	43	* where Q1 = -1.6666666666666567384E-2,
	44	* Q2 = 3.9682539681370365873E-4,
	45	* Q3 = -9.9206344733435987357E-6,
	46	* Q4 = 2.5051361420808517002E-7,
	47	* Q5 = -6.2843505682382617102E-9;
	48	* (where z=r*r, and the values of Q1 to Q5 are listed below)
f7eac6eb RM	49	* with error bounded by
f7eac6eb RM	50	* \| 5 \| -61
cccda09f	51	* \| 1.0+Q1z+...+Q5z - R1(z) \| <= 2
f7eac6eb	52	* \| \|
cccda09f UD	53	*
cccda09f UD	54	* expm1(r) = exp(r)-1 is then computed by the following
d7826aa1	55	* specific way which minimize the accumulation rounding error:
f7eac6eb RM	56	* 2 3
	57	* r r [ 3 - (R1 + R1*r/2) ]
	58	* expm1(r) = r + --- + --- * [--------------------]
d7826aa1	59	* 2 2 [ 6 - r(3 - R1r/2) ]
cccda09f	60	*
f7eac6eb	61	* To compensate the error in the argument reduction, we use
cccda09f UD	62	* expm1(r+c) = expm1(r) + c + expm1(r)*c
cccda09f UD	63	* ~ expm1(r) + c + r*c
f7eac6eb	64	* Thus c+r*c will be added in as the correction terms for
cccda09f	65	* expm1(r+c). Now rearrange the term to avoid optimization
d7826aa1 UD	66	* screw up:
	67	* ( 2 2 )
	68	* ({ ( r [ R1 - (3 - R1*r/2) ] ) } r )
f7eac6eb	69	* expm1(r+c)~r - ({r(--- [--------------------]-c)-c} - --- )
d7826aa1	70	* ({ ( 2 [ 6 - r(3 - R1r/2) ] ) } 2 )
f7eac6eb	71	* ( )
cccda09f	72	*
f7eac6eb RM	73	* = r - E
	74	* 3. Scale back to obtain expm1(x):
	75	* From step 1, we have
	76	* expm1(x) = either 2^k*[expm1(r)+1] - 1
	77	* = or 2^k*[expm1(r) + (1-2^-k)]
	78	* 4. Implementation notes:
	79	* (A). To save one multiplication, we scale the coefficient Qi
	80	* to Qi*2^i, and replace z by (x^2)/2.
	81	* (B). To achieve maximum accuracy, we compute expm1(x) by
	82	* (i) if x < -56*ln2, return -1.0, (raise inexact if x!=inf)
	83	* (ii) if k=0, return r-E
	84	* (iii) if k=-1, return 0.5*(r-E)-0.5
	85	* (iv) if k=1 if r < -0.25, return 2*((r+0.5)- E)
d7826aa1	86	* else return 1.0+2.0*(r-E);
f7eac6eb RM	87	* (v) if (k<-2\|\|k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
f7eac6eb RM	88	* (vi) if k <= 20, return 2^k((1-2^-k)-(E-r)), else
cccda09f	89	* (vii) return 2^k(1-((E+2^-k)-r))
f7eac6eb RM	90	*
	91	* Special cases:
	92	* expm1(INF) is INF, expm1(NaN) is NaN;
	93	* expm1(-INF) is -1, and
	94	* for finite argument, only expm1(0)=0 is exact.
	95	*
	96	* Accuracy:
	97	* according to an error analysis, the error is always less than
	98	* 1 ulp (unit in the last place).
	99	*
	100	* Misc. info.
cccda09f	101	* For IEEE double
f7eac6eb RM	102	* if x > 7.09782712893383973096e+02 then expm1(x) overflow
	103	*
	104	* Constants:
cccda09f UD	105	* The hexadecimal values are the intended ones for the following
cccda09f UD	106	* constants. The decimal values may be used, provided that the
f7eac6eb RM	107	* compiler will convert from decimal to binary accurately enough
	108	* to produce the hexadecimal values shown.
	109	*/
	110
f0e3c47f	111	#include <errno.h>
554edb23	112	#include <float.h>
1ed0291c	113	#include <math.h>
b4d5b8b0	114	#include <math-barriers.h>
1ed0291c	115	#include <math_private.h>
8f5b00d3	116	#include <math-underflow.h>
1e2bffd0	117	#include <libm-alias-double.h>
923609d1	118	#define one Q[0]
f7eac6eb	119	static const double
c5d5d574 OB	120	huge = 1.0e+300,
	121	tiny = 1.0e-300,
	122	o_threshold = 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */
	123	ln2_hi = 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
	124	ln2_lo = 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */
	125	invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
	126	/* scaled coefficients related to expm1 */
	127	Q[] = { 1.0, -3.33333333333331316428e-02, /* BFA11111 111110F4 */
	128	1.58730158725481460165e-03, /* 3F5A01A0 19FE5585 */
	129	-7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
	130	4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
	131	-2.01099218183624371326e-07 }; /* BE8AFDB7 6E09C32D */
f7eac6eb	132
1b214630 L	133	#ifndef SECTION
	134	# define SECTION
	135	#endif
	136
	137	SECTION
d7826aa1	138	double
c5d5d574	139	__expm1 (double x)
f7eac6eb	140	{
c5d5d574 OB	141	double y, hi, lo, c, t, e, hxs, hfx, r1, h2, h4, R1, R2, R3;
c5d5d574 OB	142	int32_t k, xsb;
24ab7723	143	uint32_t hx;
f7eac6eb	144
c5d5d574 OB	145	GET_HIGH_WORD (hx, x);
	146	xsb = hx & 0x80000000; /* sign bit of x */
	147	if (xsb == 0)
	148	y = x;
	149	else
	150	y = -x; /* y = \|x\| */
	151	hx &= 0x7fffffff; /* high word of \|x\| */
f7eac6eb	152
c5d5d574 OB	153	/* filter out huge and non-finite argument */
	154	if (hx >= 0x4043687A) /* if \|x\|>=56ln2 /
	155	{
	156	if (hx >= 0x40862E42) /* if \|x\|>=709.78... */
	157	{
	158	if (hx >= 0x7ff00000)
	159	{
24ab7723	160	uint32_t low;
c5d5d574 OB	161	GET_LOW_WORD (low, x);
	162	if (((hx & 0xfffff) \| low) != 0)
	163	return x + x; /* NaN */
	164	else
	165	return (xsb == 0) ? x : -1.0; /* exp(+-inf)={inf,-1} */
f7eac6eb	166	}
c5d5d574 OB	167	if (x > o_threshold)
	168	{
	169	__set_errno (ERANGE);
	170	return huge * huge; /* overflow */
f7eac6eb RM	171	}
f7eac6eb RM	172	}
c5d5d574 OB	173	if (xsb != 0) /* x < -56ln2, return -1.0 with inexact /
	174	{
	175	math_force_eval (x + tiny); /* raise inexact */
	176	return tiny - one; /* return -1 */
	177	}
	178	}
f7eac6eb	179
c5d5d574 OB	180	/* argument reduction */
	181	if (hx > 0x3fd62e42) /* if \|x\| > 0.5 ln2 */
	182	{
	183	if (hx < 0x3FF0A2B2) /* and \|x\| < 1.5 ln2 */
	184	{
	185	if (xsb == 0)
	186	{
	187	hi = x - ln2_hi; lo = ln2_lo; k = 1;
	188	}
	189	else
	190	{
	191	hi = x + ln2_hi; lo = -ln2_lo; k = -1;
f7eac6eb	192	}
cccda09f	193	}
c5d5d574 OB	194	else
	195	{
	196	k = invln2 * x + ((xsb == 0) ? 0.5 : -0.5);
	197	t = k;
	198	hi = x - t * ln2_hi; /* tln2_hi is exact here /
	199	lo = t * ln2_lo;
f7eac6eb	200	}
c5d5d574 OB	201	x = hi - lo;
	202	c = (hi - x) - lo;
	203	}
	204	else if (hx < 0x3c900000) /* when \|x\|<2*-54, return x /
	205	{
d96164c3	206	math_check_force_underflow (x);
c5d5d574 OB	207	t = huge + x; /* return x with inexact flags when x!=0 */
	208	return x - (t - (huge + x));
	209	}
	210	else
	211	k = 0;
f7eac6eb	212
c5d5d574 OB	213	/* x is now in primary range */
	214	hfx = 0.5 * x;
	215	hxs = x * hfx;
	216	R1 = one + hxs * Q[1]; h2 = hxs * hxs;
	217	R2 = Q[2] + hxs * Q[3]; h4 = h2 * h2;
	218	R3 = Q[4] + hxs * Q[5];
	219	r1 = R1 + h2 * R2 + h4 * R3;
	220	t = 3.0 - r1 * hfx;
	221	e = hxs * ((r1 - t) / (6.0 - x * t));
	222	if (k == 0)
	223	return x - (x * e - hxs); /* c is 0 */
	224	else
	225	{
	226	e = (x * (e - c) - c);
	227	e -= hxs;
	228	if (k == -1)
	229	return 0.5 * (x - e) - 0.5;
	230	if (k == 1)
	231	{
	232	if (x < -0.25)
	233	return -2.0 * (e - (x + 0.5));
	234	else
	235	return one + 2.0 * (x - e);
	236	}
	237	if (k <= -2 \|\| k > 56) /* suffice to return exp(x)-1 */
	238	{
24ab7723	239	uint32_t high;
c5d5d574 OB	240	y = one - (e - x);
	241	GET_HIGH_WORD (high, y);
	242	SET_HIGH_WORD (y, high + (k << 20)); /* add k to y's exponent */
	243	return y - one;
	244	}
	245	t = one;
	246	if (k < 20)
	247	{
24ab7723	248	uint32_t high;
c5d5d574 OB	249	SET_HIGH_WORD (t, 0x3ff00000 - (0x200000 >> k)); /* t=1-2^-k */
	250	y = t - (e - x);
	251	GET_HIGH_WORD (high, y);
	252	SET_HIGH_WORD (y, high + (k << 20)); /* add k to y's exponent */
	253	}
	254	else
	255	{
24ab7723	256	uint32_t high;
c5d5d574 OB	257	SET_HIGH_WORD (t, ((0x3ff - k) << 20)); /* 2^-k */
	258	y = x - (e + t);
	259	y += one;
	260	GET_HIGH_WORD (high, y);
	261	SET_HIGH_WORD (y, high + (k << 20)); /* add k to y's exponent */
f7eac6eb	262	}
c5d5d574 OB	263	}
c5d5d574 OB	264	return y;
f7eac6eb	265	}
1b214630	266	#ifndef __expm1
1e2bffd0	267	libm_alias_double (__expm1, expm1)
1b214630	268	#endif