summaryrefslogtreecommitdiff
path: root/sysdeps/ieee754/dbl-64/slowpow.c
diff options
context:
space:
mode:
authorThomas Schwinge <thomas@codesourcery.com>2013-05-23 23:37:00 +0200
committerThomas Schwinge <thomas@codesourcery.com>2013-05-23 23:37:00 +0200
commitf9e888643115b4b2f28853ebd1733f4410fb8839 (patch)
tree58c69f6cef623679080e8933b6c79880bfbd7cb8 /sysdeps/ieee754/dbl-64/slowpow.c
parentd78eef6ebc008f784f501ce208bef12c6eafda27 (diff)
parentb934acf0e93c5a220551ed6e686bb9d45a24a8cc (diff)
Merge branch 'baseline' into refs/top-bases/tschwinge/Roger_Whittaker
Diffstat (limited to 'sysdeps/ieee754/dbl-64/slowpow.c')
-rw-r--r--sysdeps/ieee754/dbl-64/slowpow.c96
1 files changed, 65 insertions, 31 deletions
diff --git a/sysdeps/ieee754/dbl-64/slowpow.c b/sysdeps/ieee754/dbl-64/slowpow.c
index c303eaa5af..a379728b14 100644
--- a/sysdeps/ieee754/dbl-64/slowpow.c
+++ b/sysdeps/ieee754/dbl-64/slowpow.c
@@ -38,42 +38,76 @@
# define SECTION
#endif
-void __mpexp(mp_no *x, mp_no *y, int p);
-void __mplog(mp_no *x, mp_no *y, int p);
-double ulog(double);
-double __halfulp(double x,double y);
+void __mpexp (mp_no *x, mp_no *y, int p);
+void __mplog (mp_no *x, mp_no *y, int p);
+double ulog (double);
+double __halfulp (double x, double y);
double
SECTION
-__slowpow(double x, double y, double z) {
- double res,res1;
- mp_no mpx, mpy, mpz,mpw,mpp,mpr,mpr1;
- static const mp_no eps = {-3,{1.0,4.0}};
+__slowpow (double x, double y, double z)
+{
+ double res, res1;
+ mp_no mpx, mpy, mpz, mpw, mpp, mpr, mpr1;
+ static const mp_no eps = {-3, {1.0, 4.0}};
int p;
- res = __halfulp(x,y); /* halfulp() returns -10 or x^y */
- if (res >= 0) return res; /* if result was really computed by halfulp */
- /* else, if result was not really computed by halfulp */
- p = 10; /* p=precision */
- __dbl_mp(x,&mpx,p);
- __dbl_mp(y,&mpy,p);
- __dbl_mp(z,&mpz,p);
- __mplog(&mpx, &mpz, p); /* log(x) = z */
- __mul(&mpy,&mpz,&mpw,p); /* y * z =w */
- __mpexp(&mpw, &mpp, p); /* e^w =pp */
- __add(&mpp,&eps,&mpr,p); /* pp+eps =r */
- __mp_dbl(&mpr, &res, p);
- __sub(&mpp,&eps,&mpr1,p); /* pp -eps =r1 */
- __mp_dbl(&mpr1, &res1, p); /* converting into double precision */
- if (res == res1) return res;
+ /* __HALFULP returns -10 or X^Y. */
+ res = __halfulp (x, y);
- p = 32; /* if we get here result wasn't calculated exactly, continue */
- __dbl_mp(x,&mpx,p); /* for more exact calculation */
- __dbl_mp(y,&mpy,p);
- __dbl_mp(z,&mpz,p);
- __mplog(&mpx, &mpz, p); /* log(c)=z */
- __mul(&mpy,&mpz,&mpw,p); /* y*z =w */
- __mpexp(&mpw, &mpp, p); /* e^w=pp */
- __mp_dbl(&mpp, &res, p); /* converting into double precision */
+ /* Return if the result was computed by __HALFULP. */
+ if (res >= 0)
+ return res;
+
+ /* Compute pow as long double. This is currently only used by powerpc, where
+ one may get 106 bits of accuracy. */
+#ifdef USE_LONG_DOUBLE_FOR_MP
+ long double ldw, ldz, ldpp;
+ static const long double ldeps = 0x4.0p-96;
+
+ ldz = __ieee754_logl ((long double) x);
+ ldw = (long double) y *ldz;
+ ldpp = __ieee754_expl (ldw);
+ res = (double) (ldpp + ldeps);
+ res1 = (double) (ldpp - ldeps);
+
+ /* Return the result if it is accurate enough. */
+ if (res == res1)
+ return res;
+#endif
+
+ /* Or else, calculate using multiple precision. P = 10 implies accuracy of
+ 240 bits accuracy, since MP_NO has a radix of 2^24. */
+ p = 10;
+ __dbl_mp (x, &mpx, p);
+ __dbl_mp (y, &mpy, p);
+ __dbl_mp (z, &mpz, p);
+
+ /* z = x ^ y
+ log (z) = y * log (x)
+ z = exp (y * log (x)) */
+ __mplog (&mpx, &mpz, p);
+ __mul (&mpy, &mpz, &mpw, p);
+ __mpexp (&mpw, &mpp, p);
+
+ /* Add and subtract EPS to ensure that the result remains unchanged, i.e. we
+ have last bit accuracy. */
+ __add (&mpp, &eps, &mpr, p);
+ __mp_dbl (&mpr, &res, p);
+ __sub (&mpp, &eps, &mpr1, p);
+ __mp_dbl (&mpr1, &res1, p);
+ if (res == res1)
+ return res;
+
+ /* If we don't, then we repeat using a higher precision. 768 bits of
+ precision ought to be enough for anybody. */
+ p = 32;
+ __dbl_mp (x, &mpx, p);
+ __dbl_mp (y, &mpy, p);
+ __dbl_mp (z, &mpz, p);
+ __mplog (&mpx, &mpz, p);
+ __mul (&mpy, &mpz, &mpw, p);
+ __mpexp (&mpw, &mpp, p);
+ __mp_dbl (&mpp, &res, p);
return res;
}