glibc/sysdeps/ieee754/flt-32/e_jnf.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

237 lines
5.6 KiB
C

/* e_jnf.c -- float version of e_jn.c.
*/
/*
* ====================================================
* 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.
* ====================================================
*/
#include <errno.h>
#include <float.h>
#include <math.h>
#include <math-narrow-eval.h>
#include <math_private.h>
#include <fenv_private.h>
#include <math-underflow.h>
#include <libm-alias-finite.h>
static const float
two = 2.0000000000e+00, /* 0x40000000 */
one = 1.0000000000e+00; /* 0x3F800000 */
static const float zero = 0.0000000000e+00;
float
__ieee754_jnf(int n, float x)
{
float ret;
{
int32_t i,hx,ix, sgn;
float a, b, temp, di;
float z, w;
/* J(-n,x) = (-1)^n * J(n, x), J(n, -x) = (-1)^n * J(n, x)
* Thus, J(-n,x) = J(n,-x)
*/
GET_FLOAT_WORD(hx,x);
ix = 0x7fffffff&hx;
/* if J(n,NaN) is NaN */
if(__builtin_expect(ix>0x7f800000, 0)) return x+x;
if(n<0){
n = -n;
x = -x;
hx ^= 0x80000000;
}
if(n==0) return(__ieee754_j0f(x));
if(n==1) return(__ieee754_j1f(x));
sgn = (n&1)&(hx>>31); /* even n -- 0, odd n -- sign(x) */
x = fabsf(x);
SET_RESTORE_ROUNDF (FE_TONEAREST);
if(__builtin_expect(ix==0||ix>=0x7f800000, 0)) /* if x is 0 or inf */
return sgn == 1 ? -zero : zero;
else if((float)n<=x) {
/* Safe to use J(n+1,x)=2n/x *J(n,x)-J(n-1,x) */
a = __ieee754_j0f(x);
b = __ieee754_j1f(x);
for(i=1;i<n;i++){
temp = b;
b = b*((double)(i+i)/x) - a; /* avoid underflow */
a = temp;
}
} else {
if(ix<0x30800000) { /* x < 2**-29 */
/* x is tiny, return the first Taylor expansion of J(n,x)
* J(n,x) = 1/n!*(x/2)^n - ...
*/
if(n>33) /* underflow */
b = zero;
else {
temp = x*(float)0.5; b = temp;
for (a=one,i=2;i<=n;i++) {
a *= (float)i; /* a = n! */
b *= temp; /* b = (x/2)^n */
}
b = b/a;
}
} else {
/* use backward recurrence */
/* x x^2 x^2
* J(n,x)/J(n-1,x) = ---- ------ ------ .....
* 2n - 2(n+1) - 2(n+2)
*
* 1 1 1
* (for large x) = ---- ------ ------ .....
* 2n 2(n+1) 2(n+2)
* -- - ------ - ------ -
* x x x
*
* Let w = 2n/x and h=2/x, then the above quotient
* is equal to the continued fraction:
* 1
* = -----------------------
* 1
* w - -----------------
* 1
* w+h - ---------
* w+2h - ...
*
* To determine how many terms needed, let
* Q(0) = w, Q(1) = w(w+h) - 1,
* Q(k) = (w+k*h)*Q(k-1) - Q(k-2),
* When Q(k) > 1e4 good for single
* When Q(k) > 1e9 good for double
* When Q(k) > 1e17 good for quadruple
*/
/* determine k */
float t,v;
float q0,q1,h,tmp; int32_t k,m;
w = (n+n)/(float)x; h = (float)2.0/(float)x;
q0 = w; z = w+h; q1 = w*z - (float)1.0; k=1;
while(q1<(float)1.0e9) {
k += 1; z += h;
tmp = z*q1 - q0;
q0 = q1;
q1 = tmp;
}
m = n+n;
for(t=zero, i = 2*(n+k); i>=m; i -= 2) t = one/(i/x-t);
a = t;
b = one;
/* estimate log((2/x)^n*n!) = n*log(2/x)+n*ln(n)
* Hence, if n*(log(2n/x)) > ...
* single 8.8722839355e+01
* double 7.09782712893383973096e+02
* long double 1.1356523406294143949491931077970765006170e+04
* then recurrent value may overflow and the result is
* likely underflow to zero
*/
tmp = n;
v = two/x;
tmp = tmp*__ieee754_logf(fabsf(v*tmp));
if(tmp<8.8721679688e+01f) {
for(i=n-1,di=(float)(i+i);i>0;i--){
temp = b;
b *= di;
b = b/x - a;
a = temp;
di -= two;
}
} else {
for(i=n-1,di=(float)(i+i);i>0;i--){
temp = b;
b *= di;
b = b/x - a;
a = temp;
di -= two;
/* scale b to avoid spurious overflow */
if(b>(float)1e10) {
a /= b;
t /= b;
b = one;
}
}
}
/* j0() and j1() suffer enormous loss of precision at and
* near zero; however, we know that their zero points never
* coincide, so just choose the one further away from zero.
*/
z = __ieee754_j0f (x);
w = __ieee754_j1f (x);
if (fabsf (z) >= fabsf (w))
b = (t * z / b);
else
b = (t * w / a);
}
}
if(sgn==1) ret = -b; else ret = b;
ret = math_narrow_eval (ret);
}
if (ret == 0)
{
ret = math_narrow_eval (copysignf (FLT_MIN, ret) * FLT_MIN);
__set_errno (ERANGE);
}
else
math_check_force_underflow (ret);
return ret;
}
libm_alias_finite (__ieee754_jnf, __jnf)
float
__ieee754_ynf(int n, float x)
{
float ret;
{
int32_t i,hx,ix;
uint32_t ib;
int32_t sign;
float a, b, temp;
GET_FLOAT_WORD(hx,x);
ix = 0x7fffffff&hx;
/* if Y(n,NaN) is NaN */
if(__builtin_expect(ix>0x7f800000, 0)) return x+x;
sign = 1;
if(n<0){
n = -n;
sign = 1 - ((n&1)<<1);
}
if(n==0) return(__ieee754_y0f(x));
if(__builtin_expect(ix==0, 0))
return -sign/zero;
if(__builtin_expect(hx<0, 0)) return zero/(zero*x);
SET_RESTORE_ROUNDF (FE_TONEAREST);
if(n==1) {
ret = sign*__ieee754_y1f(x);
goto out;
}
if(__builtin_expect(ix==0x7f800000, 0)) return zero;
a = __ieee754_y0f(x);
b = __ieee754_y1f(x);
/* quit if b is -inf */
GET_FLOAT_WORD(ib,b);
for(i=1;i<n&&ib!=0xff800000;i++){
temp = b;
b = ((double)(i+i)/x)*b - a;
GET_FLOAT_WORD(ib,b);
a = temp;
}
/* If B is +-Inf, set up errno accordingly. */
if (! isfinite (b))
__set_errno (ERANGE);
if(sign>0) ret = b; else ret = -b;
}
out:
if (isinf (ret))
ret = copysignf (FLT_MAX, ret) * FLT_MAX;
return ret;
}
libm_alias_finite (__ieee754_ynf, __ynf)