diff options
Diffstat (limited to 'sysdeps/ieee754/flt-32/e_powf.c')
-rw-r--r-- | sysdeps/ieee754/flt-32/e_powf.c | 450 |
1 files changed, 208 insertions, 242 deletions
diff --git a/sysdeps/ieee754/flt-32/e_powf.c b/sysdeps/ieee754/flt-32/e_powf.c index c72fe37d3b..ece83f0dd2 100644 --- a/sysdeps/ieee754/flt-32/e_powf.c +++ b/sysdeps/ieee754/flt-32/e_powf.c @@ -1,258 +1,224 @@ -/* e_powf.c -- float version of e_pow.c. - * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com. - */ +/* Single-precision pow function. + Copyright (C) 2017-2018 Free Software Foundation, Inc. + This file is part of the GNU C Library. -/* - * ==================================================== - * 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. - * ==================================================== - */ + 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. -#include <math.h> -#include <math_private.h> + 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. -static const float huge = 1.0e+30, tiny = 1.0e-30; + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <http://www.gnu.org/licenses/>. */ -static const float -bp[] = {1.0, 1.5,}, -dp_h[] = { 0.0, 5.84960938e-01,}, /* 0x3f15c000 */ -dp_l[] = { 0.0, 1.56322085e-06,}, /* 0x35d1cfdc */ -zero = 0.0, -one = 1.0, -two = 2.0, -two24 = 16777216.0, /* 0x4b800000 */ - /* poly coefs for (3/2)*(log(x)-2s-2/3*s**3 */ -L1 = 6.0000002384e-01, /* 0x3f19999a */ -L2 = 4.2857143283e-01, /* 0x3edb6db7 */ -L3 = 3.3333334327e-01, /* 0x3eaaaaab */ -L4 = 2.7272811532e-01, /* 0x3e8ba305 */ -L5 = 2.3066075146e-01, /* 0x3e6c3255 */ -L6 = 2.0697501302e-01, /* 0x3e53f142 */ -P1 = 1.6666667163e-01, /* 0x3e2aaaab */ -P2 = -2.7777778450e-03, /* 0xbb360b61 */ -P3 = 6.6137559770e-05, /* 0x388ab355 */ -P4 = -1.6533901999e-06, /* 0xb5ddea0e */ -P5 = 4.1381369442e-08, /* 0x3331bb4c */ -lg2 = 6.9314718246e-01, /* 0x3f317218 */ -lg2_h = 6.93145752e-01, /* 0x3f317200 */ -lg2_l = 1.42860654e-06, /* 0x35bfbe8c */ -ovt = 4.2995665694e-08, /* -(128-log2(ovfl+.5ulp)) */ -cp = 9.6179670095e-01, /* 0x3f76384f =2/(3ln2) */ -cp_h = 9.6179199219e-01, /* 0x3f763800 =head of cp */ -cp_l = 4.7017383622e-06, /* 0x369dc3a0 =tail of cp_h */ -ivln2 = 1.4426950216e+00, /* 0x3fb8aa3b =1/ln2 */ -ivln2_h = 1.4426879883e+00, /* 0x3fb8aa00 =16b 1/ln2*/ -ivln2_l = 7.0526075433e-06; /* 0x36eca570 =1/ln2 tail*/ +#include <math.h> +#include <stdint.h> +#include <shlib-compat.h> +#include <libm-alias-float.h> +#include "math_config.h" -float -__ieee754_powf(float x, float y) +/* +POWF_LOG2_POLY_ORDER = 5 +EXP2F_TABLE_BITS = 5 + +ULP error: 0.82 (~ 0.5 + relerr*2^24) +relerr: 1.27 * 2^-26 (Relative error ~= 128*Ln2*relerr_log2 + relerr_exp2) +relerr_log2: 1.83 * 2^-33 (Relative error of logx.) +relerr_exp2: 1.69 * 2^-34 (Relative error of exp2(ylogx).) +*/ + +#define N (1 << POWF_LOG2_TABLE_BITS) +#define T __powf_log2_data.tab +#define A __powf_log2_data.poly +#define OFF 0x3f330000 + +/* Subnormal input is normalized so ix has negative biased exponent. + Output is multiplied by N (POWF_SCALE) if TOINT_INTRINICS is set. */ +static inline double_t +log2_inline (uint32_t ix) { - float z,ax,z_h,z_l,p_h,p_l; - float y1,t1,t2,r,s,t,u,v,w; - int32_t i,j,k,yisint,n; - int32_t hx,hy,ix,iy,is; - - GET_FLOAT_WORD(hx,x); - GET_FLOAT_WORD(hy,y); - ix = hx&0x7fffffff; iy = hy&0x7fffffff; - - /* y==zero: x**0 = 1 */ - if(iy==0) return one; + /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ + double_t z, r, r2, r4, p, q, y, y0, invc, logc; + uint32_t iz, top, tmp; + int k, i; + + /* x = 2^k z; where z is in range [OFF,2*OFF] and exact. + The range is split into N subintervals. + The ith subinterval contains z and c is near its center. */ + tmp = ix - OFF; + i = (tmp >> (23 - POWF_LOG2_TABLE_BITS)) % N; + top = tmp & 0xff800000; + iz = ix - top; + k = (int32_t) top >> (23 - POWF_SCALE_BITS); /* arithmetic shift */ + invc = T[i].invc; + logc = T[i].logc; + z = (double_t) asfloat (iz); + + /* log2(x) = log1p(z/c-1)/ln2 + log2(c) + k */ + r = z * invc - 1; + y0 = logc + (double_t) k; + + /* Pipelined polynomial evaluation to approximate log1p(r)/ln2. */ + r2 = r * r; + y = A[0] * r + A[1]; + p = A[2] * r + A[3]; + r4 = r2 * r2; + q = A[4] * r + y0; + q = p * r2 + q; + y = y * r4 + q; + return y; +} - /* x==+-1 */ - if(x == 1.0) return one; - if(x == -1.0 && isinf(y)) return one; +#undef N +#undef T +#define N (1 << EXP2F_TABLE_BITS) +#define T __exp2f_data.tab +#define SIGN_BIAS (1 << (EXP2F_TABLE_BITS + 11)) + +/* The output of log2 and thus the input of exp2 is either scaled by N + (in case of fast toint intrinsics) or not. The unscaled xd must be + in [-1021,1023], sign_bias sets the sign of the result. */ +static inline double_t +exp2_inline (double_t xd, uint32_t sign_bias) +{ + uint64_t ki, ski, t; + /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ + double_t kd, z, r, r2, y, s; + +#if TOINT_INTRINSICS +# define C __exp2f_data.poly_scaled + /* N*x = k + r with r in [-1/2, 1/2] */ + kd = roundtoint (xd); /* k */ + ki = converttoint (xd); +#else +# define C __exp2f_data.poly +# define SHIFT __exp2f_data.shift_scaled + /* x = k/N + r with r in [-1/(2N), 1/(2N)] */ + kd = (double) (xd + SHIFT); /* Rounding to double precision is required. */ + ki = asuint64 (kd); + kd -= SHIFT; /* k/N */ +#endif + r = xd - kd; + + /* exp2(x) = 2^(k/N) * 2^r ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */ + t = T[ki % N]; + ski = ki + sign_bias; + t += ski << (52 - EXP2F_TABLE_BITS); + s = asdouble (t); + z = C[0] * r + C[1]; + r2 = r * r; + y = C[2] * r + 1; + y = z * r2 + y; + y = y * s; + return y; +} - /* +-NaN return x+y */ - if(__builtin_expect(ix > 0x7f800000 || - iy > 0x7f800000, 0)) - return x+y; +/* Returns 0 if not int, 1 if odd int, 2 if even int. */ +static inline int +checkint (uint32_t iy) +{ + int e = iy >> 23 & 0xff; + if (e < 0x7f) + return 0; + if (e > 0x7f + 23) + return 2; + if (iy & ((1 << (0x7f + 23 - e)) - 1)) + return 0; + if (iy & (1 << (0x7f + 23 - e))) + return 1; + return 2; +} - /* determine if y is an odd int when x < 0 - * yisint = 0 ... y is not an integer - * yisint = 1 ... y is an odd int - * yisint = 2 ... y is an even int - */ - yisint = 0; - if(hx<0) { - if(iy>=0x4b800000) yisint = 2; /* even integer y */ - else if(iy>=0x3f800000) { - k = (iy>>23)-0x7f; /* exponent */ - j = iy>>(23-k); - if((j<<(23-k))==iy) yisint = 2-(j&1); - } - } +static inline int +zeroinfnan (uint32_t ix) +{ + return 2 * ix - 1 >= 2u * 0x7f800000 - 1; +} - /* special value of y */ - if (__builtin_expect(iy==0x7f800000, 0)) { /* y is +-inf */ - if (ix==0x3f800000) - return y - y; /* inf**+-1 is NaN */ - else if (ix > 0x3f800000)/* (|x|>1)**+-inf = inf,0 */ - return (hy>=0)? y: zero; - else /* (|x|<1)**-,+inf = inf,0 */ - return (hy<0)?-y: zero; - } - if(iy==0x3f800000) { /* y is +-1 */ - if(hy<0) return one/x; else return x; - } - if(hy==0x40000000) return x*x; /* y is 2 */ - if(hy==0x3f000000) { /* y is 0.5 */ - if(__builtin_expect(hx>=0, 1)) /* x >= +0 */ - return __ieee754_sqrtf(x); +float +__powf (float x, float y) +{ + uint32_t sign_bias = 0; + uint32_t ix, iy; + + ix = asuint (x); + iy = asuint (y); + if (__glibc_unlikely (ix - 0x00800000 >= 0x7f800000 - 0x00800000 + || zeroinfnan (iy))) + { + /* Either (x < 0x1p-126 or inf or nan) or (y is 0 or inf or nan). */ + if (__glibc_unlikely (zeroinfnan (iy))) + { + if (2 * iy == 0) + return issignalingf_inline (x) ? x + y : 1.0f; + if (ix == 0x3f800000) + return issignalingf_inline (y) ? x + y : 1.0f; + if (2 * ix > 2u * 0x7f800000 || 2 * iy > 2u * 0x7f800000) + return x + y; + if (2 * ix == 2 * 0x3f800000) + return 1.0f; + if ((2 * ix < 2 * 0x3f800000) == !(iy & 0x80000000)) + return 0.0f; /* |x|<1 && y==inf or |x|>1 && y==-inf. */ + return y * y; } - - ax = fabsf(x); - /* special value of x */ - if(__builtin_expect(ix==0x7f800000||ix==0||ix==0x3f800000, 0)){ - z = ax; /*x is +-0,+-inf,+-1*/ - if(hy<0) z = one/z; /* z = (1/|x|) */ - if(hx<0) { - if(((ix-0x3f800000)|yisint)==0) { - z = (z-z)/(z-z); /* (-1)**non-int is NaN */ - } else if(yisint==1) - z = -z; /* (x<0)**odd = -(|x|**odd) */ + if (__glibc_unlikely (zeroinfnan (ix))) + { + float_t x2 = x * x; + if (ix & 0x80000000 && checkint (iy) == 1) + { + x2 = -x2; + sign_bias = 1; } - return z; - } - - /* (x<0)**(non-int) is NaN */ - if(__builtin_expect(((((u_int32_t)hx>>31)-1)|yisint)==0, 0)) - return (x-x)/(x-x); - - /* |y| is huge */ - if(__builtin_expect(iy>0x4d000000, 0)) { /* if |y| > 2**27 */ - /* over/underflow if x is not close to one */ - if(ix<0x3f7ffff8) return (hy<0)? huge*huge:tiny*tiny; - if(ix>0x3f800007) return (hy>0)? huge*huge:tiny*tiny; - /* now |1-x| is tiny <= 2**-20, suffice to compute - log(x) by x-x^2/2+x^3/3-x^4/4 */ - t = ax-1; /* t has 20 trailing zeros */ - w = (t*t)*((float)0.5-t*((float)0.333333333333-t*(float)0.25)); - u = ivln2_h*t; /* ivln2_h has 16 sig. bits */ - v = t*ivln2_l-w*ivln2; - t1 = u+v; - GET_FLOAT_WORD(is,t1); - SET_FLOAT_WORD(t1,is&0xfffff000); - t2 = v-(t1-u); - } else { - float s2,s_h,s_l,t_h,t_l; - /* Avoid internal underflow for tiny y. The exact value - of y does not matter if |y| <= 2**-32. */ - if (iy < 0x2f800000) - SET_FLOAT_WORD (y, (hy & 0x80000000) | 0x2f800000); - n = 0; - /* take care subnormal number */ - if(ix<0x00800000) - {ax *= two24; n -= 24; GET_FLOAT_WORD(ix,ax); } - n += ((ix)>>23)-0x7f; - j = ix&0x007fffff; - /* determine interval */ - ix = j|0x3f800000; /* normalize ix */ - if(j<=0x1cc471) k=0; /* |x|<sqrt(3/2) */ - else if(j<0x5db3d7) k=1; /* |x|<sqrt(3) */ - else {k=0;n+=1;ix -= 0x00800000;} - SET_FLOAT_WORD(ax,ix); - - /* compute s = s_h+s_l = (x-1)/(x+1) or (x-1.5)/(x+1.5) */ - u = ax-bp[k]; /* bp[0]=1.0, bp[1]=1.5 */ - v = one/(ax+bp[k]); - s = u*v; - s_h = s; - GET_FLOAT_WORD(is,s_h); - SET_FLOAT_WORD(s_h,is&0xfffff000); - /* t_h=ax+bp[k] High */ - SET_FLOAT_WORD (t_h, - ((((ix>>1)|0x20000000)+0x00400000+(k<<21)) - & 0xfffff000)); - t_l = ax - (t_h-bp[k]); - s_l = v*((u-s_h*t_h)-s_h*t_l); - /* compute log(ax) */ - s2 = s*s; - r = s2*s2*(L1+s2*(L2+s2*(L3+s2*(L4+s2*(L5+s2*L6))))); - r += s_l*(s_h+s); - s2 = s_h*s_h; - t_h = (float)3.0+s2+r; - GET_FLOAT_WORD(is,t_h); - SET_FLOAT_WORD(t_h,is&0xfffff000); - t_l = r-((t_h-(float)3.0)-s2); - /* u+v = s*(1+...) */ - u = s_h*t_h; - v = s_l*t_h+t_l*s; - /* 2/(3log2)*(s+...) */ - p_h = u+v; - GET_FLOAT_WORD(is,p_h); - SET_FLOAT_WORD(p_h,is&0xfffff000); - p_l = v-(p_h-u); - z_h = cp_h*p_h; /* cp_h+cp_l = 2/(3*log2) */ - z_l = cp_l*p_h+p_l*cp+dp_l[k]; - /* log2(ax) = (s+..)*2/(3*log2) = n + dp_h + z_h + z_l */ - t = (float)n; - t1 = (((z_h+z_l)+dp_h[k])+t); - GET_FLOAT_WORD(is,t1); - SET_FLOAT_WORD(t1,is&0xfffff000); - t2 = z_l-(((t1-t)-dp_h[k])-z_h); - } - - s = one; /* s (sign of result -ve**odd) = -1 else = 1 */ - if(((((u_int32_t)hx>>31)-1)|(yisint-1))==0) - s = -one; /* (-ve)**(odd int) */ - - /* split up y into y1+y2 and compute (y1+y2)*(t1+t2) */ - GET_FLOAT_WORD(is,y); - SET_FLOAT_WORD(y1,is&0xfffff000); - p_l = (y-y1)*t1+y*t2; - p_h = y1*t1; - z = p_l+p_h; - GET_FLOAT_WORD(j,z); - if (__builtin_expect(j>0x43000000, 0)) /* if z > 128 */ - return s*huge*huge; /* overflow */ - else if (__builtin_expect(j==0x43000000, 0)) { /* if z == 128 */ - if(p_l+ovt>z-p_h) return s*huge*huge; /* overflow */ +#if WANT_ERRNO + if (2 * ix == 0 && iy & 0x80000000) + return __math_divzerof (sign_bias); +#endif + return iy & 0x80000000 ? 1 / x2 : x2; } - else if (__builtin_expect((j&0x7fffffff)>0x43160000, 0))/* z <= -150 */ - return s*tiny*tiny; /* underflow */ - else if (__builtin_expect((u_int32_t) j==0xc3160000, 0)){/* z == -150*/ - if(p_l<=z-p_h) return s*tiny*tiny; /* underflow */ + /* x and y are non-zero finite. */ + if (ix & 0x80000000) + { + /* Finite x < 0. */ + int yint = checkint (iy); + if (yint == 0) + return __math_invalidf (x); + if (yint == 1) + sign_bias = SIGN_BIAS; + ix &= 0x7fffffff; } - /* - * compute 2**(p_h+p_l) - */ - i = j&0x7fffffff; - k = (i>>23)-0x7f; - n = 0; - if(i>0x3f000000) { /* if |z| > 0.5, set n = [z+0.5] */ - n = j+(0x00800000>>(k+1)); - k = ((n&0x7fffffff)>>23)-0x7f; /* new k for n */ - SET_FLOAT_WORD(t,n&~(0x007fffff>>k)); - n = ((n&0x007fffff)|0x00800000)>>(23-k); - if(j<0) n = -n; - p_h -= t; + if (ix < 0x00800000) + { + /* Normalize subnormal x so exponent becomes negative. */ + ix = asuint (x * 0x1p23f); + ix &= 0x7fffffff; + ix -= 23 << 23; } - t = p_l+p_h; - GET_FLOAT_WORD(is,t); - SET_FLOAT_WORD(t,is&0xfffff000); - u = t*lg2_h; - v = (p_l-(t-p_h))*lg2+t*lg2_l; - z = u+v; - w = v-(z-u); - t = z*z; - t1 = z - t*(P1+t*(P2+t*(P3+t*(P4+t*P5)))); - r = (z*t1)/(t1-two)-(w+z*w); - z = one-(r-z); - GET_FLOAT_WORD(j,z); - j += (n<<23); - if((j>>23)<=0) /* subnormal output */ - { - z = __scalbnf (z, n); - float force_underflow = z * z; - math_force_eval (force_underflow); - } - else SET_FLOAT_WORD(z,j); - return s*z; + } + double_t logx = log2_inline (ix); + double_t ylogx = y * logx; /* Note: cannot overflow, y is single prec. */ + if (__glibc_unlikely ((asuint64 (ylogx) >> 47 & 0xffff) + >= asuint64 (126.0 * POWF_SCALE) >> 47)) + { + /* |y*log(x)| >= 126. */ + if (ylogx > 0x1.fffffffd1d571p+6 * POWF_SCALE) + return __math_oflowf (sign_bias); + if (ylogx <= -150.0 * POWF_SCALE) + return __math_uflowf (sign_bias); +#if WANT_ERRNO_UFLOW + if (ylogx < -149.0 * POWF_SCALE) + return __math_may_uflowf (sign_bias); +#endif + } + return (float) exp2_inline (ylogx, sign_bias); } -strong_alias (__ieee754_powf, __powf_finite) +#ifndef __powf +strong_alias (__powf, __ieee754_powf) +strong_alias (__powf, __powf_finite) +versioned_symbol (libm, __powf, powf, GLIBC_2_27); +libm_alias_float_other (__pow, pow) +#endif |