|
|
00db10 |
# commit 650ef4bd7976e36831cba22d838b567d3b5f6e8f
|
|
|
00db10 |
# Author: Alan Modra <amodra@gmail.com>
|
|
|
00db10 |
# Date: Sat Aug 17 18:25:51 2013 +0930
|
|
|
00db10 |
#
|
|
|
00db10 |
# PowerPC floating point little-endian [4 of 15]
|
|
|
00db10 |
# http://sourceware.org/ml/libc-alpha/2013-08/msg00084.html
|
|
|
00db10 |
#
|
|
|
00db10 |
# Another batch of ieee854 macros and union replacement. These four
|
|
|
00db10 |
# files also have bugs fixed with this patch. The fact that the two
|
|
|
00db10 |
# doubles in an IBM long double may have different signs means that
|
|
|
00db10 |
# negation and absolute value operations can't just twiddle one sign bit
|
|
|
00db10 |
# as you can with ieee864 style extended double. fmodl, remainderl,
|
|
|
00db10 |
# erfl and erfcl all had errors of this type. erfl also returned +1 for
|
|
|
00db10 |
# large magnitude negative input where it should return -1. The hypotl
|
|
|
00db10 |
# error is innocuous since the value adjusted twice is only used as a
|
|
|
00db10 |
# flag. The e_hypotl.c tests for large "a" and small "b" are mutually
|
|
|
00db10 |
# exclusive because we've already exited when x/y > 2**120. That allows
|
|
|
00db10 |
# some further small simplifications.
|
|
|
00db10 |
#
|
|
|
00db10 |
# [BZ #15734], [BZ #15735]
|
|
|
00db10 |
# * sysdeps/ieee754/ldbl-128ibm/e_fmodl.c (__ieee754_fmodl): Rewrite
|
|
|
00db10 |
# all uses of ieee875 long double macros and unions. Simplify test
|
|
|
00db10 |
# for 0.0L. Correct |x|<|y| and |x|=|y| test. Use
|
|
|
00db10 |
# ldbl_extract_mantissa value for ix,iy exponents. Properly
|
|
|
00db10 |
# normalize after ldbl_extract_mantissa, and don't add hidden bit
|
|
|
00db10 |
# already handled. Don't treat low word of ieee854 mantissa like
|
|
|
00db10 |
# low word of IBM long double and mask off bit when testing for
|
|
|
00db10 |
# zero.
|
|
|
00db10 |
# * sysdeps/ieee754/ldbl-128ibm/e_hypotl.c (__ieee754_hypotl): Rewrite
|
|
|
00db10 |
# all uses of ieee875 long double macros and unions. Simplify tests
|
|
|
00db10 |
# for 0.0L and inf. Correct double adjustment of k. Delete dead code
|
|
|
00db10 |
# adjusting ha,hb. Simplify code setting kld. Delete two600 and
|
|
|
00db10 |
# two1022, instead use their values. Recognise that tests for large
|
|
|
00db10 |
# "a" and small "b" are mutually exclusive. Rename vars. Comment.
|
|
|
00db10 |
# * sysdeps/ieee754/ldbl-128ibm/e_remainderl.c (__ieee754_remainderl):
|
|
|
00db10 |
# Rewrite all uses of ieee875 long double macros and unions. Simplify
|
|
|
00db10 |
# test for 0.0L and nan. Correct negation.
|
|
|
00db10 |
# * sysdeps/ieee754/ldbl-128ibm/s_erfl.c (__erfl): Rewrite all uses of
|
|
|
00db10 |
# ieee875 long double macros and unions. Correct output for large
|
|
|
00db10 |
# magnitude x. Correct absolute value calculation.
|
|
|
00db10 |
# (__erfcl): Likewise.
|
|
|
00db10 |
# * math/libm-test.inc: Add tests for errors discovered in IBM long
|
|
|
00db10 |
# double versions of fmodl, remainderl, erfl and erfcl.
|
|
|
00db10 |
#
|
|
|
00db10 |
diff -urN glibc-2.17-c758a686/math/libm-test.inc glibc-2.17-c758a686/math/libm-test.inc
|
|
|
00db10 |
--- glibc-2.17-c758a686/math/libm-test.inc 2014-05-27 20:02:29.000000000 -0500
|
|
|
00db10 |
+++ glibc-2.17-c758a686/math/libm-test.inc 2014-05-27 20:09:59.000000000 -0500
|
|
|
00db10 |
@@ -4040,6 +4040,10 @@
|
|
|
00db10 |
TEST_f_f (erf, 2.0L, 0.995322265018952734162069256367252929L);
|
|
|
00db10 |
TEST_f_f (erf, 4.125L, 0.999999994576599200434933994687765914L);
|
|
|
00db10 |
TEST_f_f (erf, 27.0L, 1.0L);
|
|
|
00db10 |
+#if defined TEST_LDOUBLE && LDBL_MANT_DIG >= 54
|
|
|
00db10 |
+ /* The input is not exactly representable as a double. */
|
|
|
00db10 |
+ TEST_f_f (erf, -0x1.fffffffffffff8p-2L, -0.5204998778130465132916303345518417673509L);
|
|
|
00db10 |
+#endif
|
|
|
00db10 |
|
|
|
00db10 |
END (erf);
|
|
|
00db10 |
}
|
|
|
00db10 |
@@ -4071,6 +4075,10 @@
|
|
|
00db10 |
TEST_f_f (erfc, 0x1.ffa002p+2L, 1.233585992097580296336099501489175967033e-29L);
|
|
|
00db10 |
TEST_f_f (erfc, 0x1.ffffc8p+2L, 1.122671365033056305522366683719541099329e-29L);
|
|
|
00db10 |
#ifdef TEST_LDOUBLE
|
|
|
00db10 |
+# if LDBL_MANT_DIG >= 54
|
|
|
00db10 |
+ /* The input is not exactly representable as a double. */
|
|
|
00db10 |
+ TEST_f_f (erfc, -0x1.fffffffffffff8p-2L, 1.52049987781304651329163033455184176735L);
|
|
|
00db10 |
+# endif
|
|
|
00db10 |
/* The result can only be represented in long double. */
|
|
|
00db10 |
# if LDBL_MIN_10_EXP < -319
|
|
|
00db10 |
TEST_f_f (erfc, 27.0L, 0.523704892378925568501606768284954709e-318L);
|
|
|
00db10 |
@@ -5634,6 +5642,13 @@
|
|
|
00db10 |
#if defined TEST_LDOUBLE && LDBL_MIN_EXP <= -16381
|
|
|
00db10 |
TEST_ff_f (fmod, 0x0.fffffffffffffffep-16382L, 0x1p-16445L, plus_zero);
|
|
|
00db10 |
#endif
|
|
|
00db10 |
+#if defined TEST_LDOUBLE && LDBL_MANT_DIG >= 56
|
|
|
00db10 |
+ TEST_ff_f (fmod, -0x1.00000000000004p+0L, 0x1.fffffffffffff8p-1L, -0x1p-53L);
|
|
|
00db10 |
+ TEST_ff_f (fmod, 0x1.fffffffffffffap-1L, 0x1.fffffffffffff8p-1L, 0x1p-56L);
|
|
|
00db10 |
+ TEST_ff_f (fmod, -0x1.fffffffffffffap-1L, 0x1.fffffffffffff8p-1L, -0x1p-56L);
|
|
|
00db10 |
+ TEST_ff_f (fmod, 0x1.fffffffffffffap-1L, -0x1.fffffffffffff8p-1L, 0x1p-56L);
|
|
|
00db10 |
+ TEST_ff_f (fmod, -0x1.fffffffffffffap-1L, -0x1.fffffffffffff8p-1L, -0x1p-56L);
|
|
|
00db10 |
+#endif
|
|
|
00db10 |
|
|
|
00db10 |
END (fmod);
|
|
|
00db10 |
}
|
|
|
00db10 |
@@ -8642,6 +8657,9 @@
|
|
|
00db10 |
TEST_ff_f (remainder, -1.625, -1.0, 0.375);
|
|
|
00db10 |
TEST_ff_f (remainder, 5.0, 2.0, 1.0);
|
|
|
00db10 |
TEST_ff_f (remainder, 3.0, 2.0, -1.0);
|
|
|
00db10 |
+#if defined TEST_LDOUBLE && LDBL_MANT_DIG >= 56
|
|
|
00db10 |
+ TEST_ff_f (remainder, -0x1.80000000000002p1L, 2.0, 0x1.fffffffffffff8p-1L);
|
|
|
00db10 |
+#endif
|
|
|
00db10 |
|
|
|
00db10 |
END (remainder);
|
|
|
00db10 |
}
|
|
|
00db10 |
diff -urN glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_fmodl.c glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_fmodl.c
|
|
|
00db10 |
--- glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_fmodl.c 2014-05-27 20:02:27.000000000 -0500
|
|
|
00db10 |
+++ glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_fmodl.c 2014-05-27 20:04:08.000000000 -0500
|
|
|
00db10 |
@@ -27,76 +27,83 @@
|
|
|
00db10 |
long double
|
|
|
00db10 |
__ieee754_fmodl (long double x, long double y)
|
|
|
00db10 |
{
|
|
|
00db10 |
- int64_t n,hx,hy,hz,ix,iy,sx, i;
|
|
|
00db10 |
- u_int64_t lx,ly,lz;
|
|
|
00db10 |
- int temp;
|
|
|
00db10 |
-
|
|
|
00db10 |
- GET_LDOUBLE_WORDS64(hx,lx,x);
|
|
|
00db10 |
- GET_LDOUBLE_WORDS64(hy,ly,y);
|
|
|
00db10 |
+ int64_t hx, hy, hz, sx, sy;
|
|
|
00db10 |
+ uint64_t lx, ly, lz;
|
|
|
00db10 |
+ int n, ix, iy;
|
|
|
00db10 |
+ double xhi, xlo, yhi, ylo;
|
|
|
00db10 |
+
|
|
|
00db10 |
+ ldbl_unpack (x, &xhi, &xlo;;
|
|
|
00db10 |
+ EXTRACT_WORDS64 (hx, xhi);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (lx, xlo);
|
|
|
00db10 |
+ ldbl_unpack (y, &yhi, &ylo);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (hy, yhi);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (ly, ylo);
|
|
|
00db10 |
sx = hx&0x8000000000000000ULL; /* sign of x */
|
|
|
00db10 |
- hx ^=sx; /* |x| */
|
|
|
00db10 |
- hy &= 0x7fffffffffffffffLL; /* |y| */
|
|
|
00db10 |
+ hx ^= sx; /* |x| */
|
|
|
00db10 |
+ sy = hy&0x8000000000000000ULL; /* sign of y */
|
|
|
00db10 |
+ hy ^= sy; /* |y| */
|
|
|
00db10 |
|
|
|
00db10 |
/* purge off exception values */
|
|
|
00db10 |
- if(__builtin_expect((hy|(ly&0x7fffffffffffffff))==0 ||
|
|
|
00db10 |
+ if(__builtin_expect(hy==0 ||
|
|
|
00db10 |
(hx>=0x7ff0000000000000LL)|| /* y=0,or x not finite */
|
|
|
00db10 |
(hy>0x7ff0000000000000LL),0)) /* or y is NaN */
|
|
|
00db10 |
return (x*y)/(x*y);
|
|
|
00db10 |
- if(__builtin_expect(hx<=hy,0)) {
|
|
|
00db10 |
- if((hx
|
|
|
00db10 |
- if(lx==ly)
|
|
|
00db10 |
- return Zero[(u_int64_t)sx>>63]; /* |x|=|y| return x*0*/
|
|
|
00db10 |
+ if (__builtin_expect (hx <= hy, 0))
|
|
|
00db10 |
+ {
|
|
|
00db10 |
+ /* If |x| < |y| return x. */
|
|
|
00db10 |
+ if (hx < hy)
|
|
|
00db10 |
+ return x;
|
|
|
00db10 |
+ /* At this point the absolute value of the high doubles of
|
|
|
00db10 |
+ x and y must be equal. */
|
|
|
00db10 |
+ /* If the low double of y is the same sign as the high
|
|
|
00db10 |
+ double of y (ie. the low double increases |y|)... */
|
|
|
00db10 |
+ if (((ly ^ sy) & 0x8000000000000000LL) == 0
|
|
|
00db10 |
+ /* ... then a different sign low double to high double
|
|
|
00db10 |
+ for x or same sign but lower magnitude... */
|
|
|
00db10 |
+ && (int64_t) (lx ^ sx) < (int64_t) (ly ^ sy))
|
|
|
00db10 |
+ /* ... means |x| < |y|. */
|
|
|
00db10 |
+ return x;
|
|
|
00db10 |
+ /* If the low double of x differs in sign to the high
|
|
|
00db10 |
+ double of x (ie. the low double decreases |x|)... */
|
|
|
00db10 |
+ if (((lx ^ sx) & 0x8000000000000000LL) != 0
|
|
|
00db10 |
+ /* ... then a different sign low double to high double
|
|
|
00db10 |
+ for y with lower magnitude (we've already caught
|
|
|
00db10 |
+ the same sign for y case above)... */
|
|
|
00db10 |
+ && (int64_t) (lx ^ sx) > (int64_t) (ly ^ sy))
|
|
|
00db10 |
+ /* ... means |x| < |y|. */
|
|
|
00db10 |
+ return x;
|
|
|
00db10 |
+ /* If |x| == |y| return x*0. */
|
|
|
00db10 |
+ if ((lx ^ sx) == (ly ^ sy))
|
|
|
00db10 |
+ return Zero[(uint64_t) sx >> 63];
|
|
|
00db10 |
}
|
|
|
00db10 |
|
|
|
00db10 |
- /* determine ix = ilogb(x) */
|
|
|
00db10 |
- if(__builtin_expect(hx<0x0010000000000000LL,0)) { /* subnormal x */
|
|
|
00db10 |
- if(hx==0) {
|
|
|
00db10 |
- for (ix = -1043, i=lx; i>0; i<<=1) ix -=1;
|
|
|
00db10 |
- } else {
|
|
|
00db10 |
- for (ix = -1022, i=(hx<<11); i>0; i<<=1) ix -=1;
|
|
|
00db10 |
- }
|
|
|
00db10 |
- } else ix = (hx>>52)-0x3ff;
|
|
|
00db10 |
-
|
|
|
00db10 |
- /* determine iy = ilogb(y) */
|
|
|
00db10 |
- if(__builtin_expect(hy<0x0010000000000000LL,0)) { /* subnormal y */
|
|
|
00db10 |
- if(hy==0) {
|
|
|
00db10 |
- for (iy = -1043, i=ly; i>0; i<<=1) iy -=1;
|
|
|
00db10 |
- } else {
|
|
|
00db10 |
- for (iy = -1022, i=(hy<<11); i>0; i<<=1) iy -=1;
|
|
|
00db10 |
- }
|
|
|
00db10 |
- } else iy = (hy>>52)-0x3ff;
|
|
|
00db10 |
-
|
|
|
00db10 |
/* Make the IBM extended format 105 bit mantissa look like the ieee854 112
|
|
|
00db10 |
bit mantissa so the following operations will give the correct
|
|
|
00db10 |
result. */
|
|
|
00db10 |
- ldbl_extract_mantissa(&hx, &lx, &temp, x);
|
|
|
00db10 |
- ldbl_extract_mantissa(&hy, &ly, &temp, y);
|
|
|
00db10 |
+ ldbl_extract_mantissa(&hx, &lx, &ix, x);
|
|
|
00db10 |
+ ldbl_extract_mantissa(&hy, &ly, &iy, y);
|
|
|
00db10 |
|
|
|
00db10 |
- /* set up {hx,lx}, {hy,ly} and align y to x */
|
|
|
00db10 |
- if(__builtin_expect(ix >= -1022, 1))
|
|
|
00db10 |
- hx = 0x0001000000000000LL|(0x0000ffffffffffffLL&hx;;
|
|
|
00db10 |
- else { /* subnormal x, shift x to normal */
|
|
|
00db10 |
- n = -1022-ix;
|
|
|
00db10 |
- if(n<=63) {
|
|
|
00db10 |
- hx = (hx<<n)|(lx>>(64-n));
|
|
|
00db10 |
- lx <<= n;
|
|
|
00db10 |
- } else {
|
|
|
00db10 |
- hx = lx<<(n-64);
|
|
|
00db10 |
- lx = 0;
|
|
|
00db10 |
- }
|
|
|
00db10 |
- }
|
|
|
00db10 |
- if(__builtin_expect(iy >= -1022, 1))
|
|
|
00db10 |
- hy = 0x0001000000000000LL|(0x0000ffffffffffffLL&hy);
|
|
|
00db10 |
- else { /* subnormal y, shift y to normal */
|
|
|
00db10 |
- n = -1022-iy;
|
|
|
00db10 |
- if(n<=63) {
|
|
|
00db10 |
- hy = (hy<<n)|(ly>>(64-n));
|
|
|
00db10 |
- ly <<= n;
|
|
|
00db10 |
- } else {
|
|
|
00db10 |
- hy = ly<<(n-64);
|
|
|
00db10 |
- ly = 0;
|
|
|
00db10 |
- }
|
|
|
00db10 |
- }
|
|
|
00db10 |
+ if (__builtin_expect (ix == -IEEE754_DOUBLE_BIAS, 0))
|
|
|
00db10 |
+ {
|
|
|
00db10 |
+ /* subnormal x, shift x to normal. */
|
|
|
00db10 |
+ while ((hx & (1LL << 48)) == 0)
|
|
|
00db10 |
+ {
|
|
|
00db10 |
+ hx = (hx << 1) | (lx >> 63);
|
|
|
00db10 |
+ lx = lx << 1;
|
|
|
00db10 |
+ ix -= 1;
|
|
|
00db10 |
+ }
|
|
|
00db10 |
+ }
|
|
|
00db10 |
+
|
|
|
00db10 |
+ if (__builtin_expect (iy == -IEEE754_DOUBLE_BIAS, 0))
|
|
|
00db10 |
+ {
|
|
|
00db10 |
+ /* subnormal y, shift y to normal. */
|
|
|
00db10 |
+ while ((hy & (1LL << 48)) == 0)
|
|
|
00db10 |
+ {
|
|
|
00db10 |
+ hy = (hy << 1) | (ly >> 63);
|
|
|
00db10 |
+ ly = ly << 1;
|
|
|
00db10 |
+ iy -= 1;
|
|
|
00db10 |
+ }
|
|
|
00db10 |
+ }
|
|
|
00db10 |
|
|
|
00db10 |
/* fix point fmod */
|
|
|
00db10 |
n = ix - iy;
|
|
|
00db10 |
@@ -104,7 +111,7 @@
|
|
|
00db10 |
hz=hx-hy;lz=lx-ly; if(lx
|
|
|
00db10 |
if(hz<0){hx = hx+hx+(lx>>63); lx = lx+lx;}
|
|
|
00db10 |
else {
|
|
|
00db10 |
- if((hz|(lz&0x7fffffffffffffff))==0) /* return sign(x)*0 */
|
|
|
00db10 |
+ if((hz|lz)==0) /* return sign(x)*0 */
|
|
|
00db10 |
return Zero[(u_int64_t)sx>>63];
|
|
|
00db10 |
hx = hz+hz+(lz>>63); lx = lz+lz;
|
|
|
00db10 |
}
|
|
|
00db10 |
@@ -113,7 +120,7 @@
|
|
|
00db10 |
if(hz>=0) {hx=hz;lx=lz;}
|
|
|
00db10 |
|
|
|
00db10 |
/* convert back to floating value and restore the sign */
|
|
|
00db10 |
- if((hx|(lx&0x7fffffffffffffff))==0) /* return sign(x)*0 */
|
|
|
00db10 |
+ if((hx|lx)==0) /* return sign(x)*0 */
|
|
|
00db10 |
return Zero[(u_int64_t)sx>>63];
|
|
|
00db10 |
while(hx<0x0001000000000000LL) { /* normalize x */
|
|
|
00db10 |
hx = hx+hx+(lx>>63); lx = lx+lx;
|
|
|
00db10 |
diff -urN glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_hypotl.c glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_hypotl.c
|
|
|
00db10 |
--- glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_hypotl.c 2014-05-27 20:02:27.000000000 -0500
|
|
|
00db10 |
+++ glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_hypotl.c 2014-05-27 20:04:08.000000000 -0500
|
|
|
00db10 |
@@ -45,76 +45,84 @@
|
|
|
00db10 |
#include <math.h>
|
|
|
00db10 |
#include <math_private.h>
|
|
|
00db10 |
|
|
|
00db10 |
-static const long double two600 = 0x1.0p+600L;
|
|
|
00db10 |
-static const long double two1022 = 0x1.0p+1022L;
|
|
|
00db10 |
-
|
|
|
00db10 |
long double
|
|
|
00db10 |
__ieee754_hypotl(long double x, long double y)
|
|
|
00db10 |
{
|
|
|
00db10 |
- long double a,b,t1,t2,y1,y2,w,kld;
|
|
|
00db10 |
+ long double a,b,a1,a2,b1,b2,w,kld;
|
|
|
00db10 |
int64_t j,k,ha,hb;
|
|
|
00db10 |
+ double xhi, yhi, hi, lo;
|
|
|
00db10 |
|
|
|
00db10 |
- GET_LDOUBLE_MSW64(ha,x);
|
|
|
00db10 |
+ xhi = ldbl_high (x);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (ha, xhi);
|
|
|
00db10 |
+ yhi = ldbl_high (y);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (hb, yhi);
|
|
|
00db10 |
ha &= 0x7fffffffffffffffLL;
|
|
|
00db10 |
- GET_LDOUBLE_MSW64(hb,y);
|
|
|
00db10 |
hb &= 0x7fffffffffffffffLL;
|
|
|
00db10 |
if(hb > ha) {a=y;b=x;j=ha; ha=hb;hb=j;} else {a=x;b=y;}
|
|
|
00db10 |
a = fabsl(a); /* a <- |a| */
|
|
|
00db10 |
b = fabsl(b); /* b <- |b| */
|
|
|
00db10 |
- if((ha-hb)>0x780000000000000LL) {return a+b;} /* x/y > 2**120 */
|
|
|
00db10 |
+ if((ha-hb)>0x0780000000000000LL) {return a+b;} /* x/y > 2**120 */
|
|
|
00db10 |
k=0;
|
|
|
00db10 |
kld = 1.0L;
|
|
|
00db10 |
if(ha > 0x5f30000000000000LL) { /* a>2**500 */
|
|
|
00db10 |
if(ha >= 0x7ff0000000000000LL) { /* Inf or NaN */
|
|
|
00db10 |
- u_int64_t low;
|
|
|
00db10 |
w = a+b; /* for sNaN */
|
|
|
00db10 |
- GET_LDOUBLE_LSW64(low,a);
|
|
|
00db10 |
- if(((ha&0xfffffffffffffLL)|(low&0x7fffffffffffffffLL))==0)
|
|
|
00db10 |
+ if(ha == 0x7ff0000000000000LL)
|
|
|
00db10 |
w = a;
|
|
|
00db10 |
- GET_LDOUBLE_LSW64(low,b);
|
|
|
00db10 |
- if(((hb^0x7ff0000000000000LL)|(low&0x7fffffffffffffffLL))==0)
|
|
|
00db10 |
+ if(hb == 0x7ff0000000000000LL)
|
|
|
00db10 |
w = b;
|
|
|
00db10 |
return w;
|
|
|
00db10 |
}
|
|
|
00db10 |
/* scale a and b by 2**-600 */
|
|
|
00db10 |
- ha -= 0x2580000000000000LL; hb -= 0x2580000000000000LL; k += 600;
|
|
|
00db10 |
- a /= two600;
|
|
|
00db10 |
- b /= two600;
|
|
|
00db10 |
- k += 600;
|
|
|
00db10 |
- kld = two600;
|
|
|
00db10 |
+ a *= 0x1p-600L;
|
|
|
00db10 |
+ b *= 0x1p-600L;
|
|
|
00db10 |
+ k = 600;
|
|
|
00db10 |
+ kld = 0x1p+600L;
|
|
|
00db10 |
}
|
|
|
00db10 |
- if(hb < 0x23d0000000000000LL) { /* b < 2**-450 */
|
|
|
00db10 |
+ else if(hb < 0x23d0000000000000LL) { /* b < 2**-450 */
|
|
|
00db10 |
if(hb <= 0x000fffffffffffffLL) { /* subnormal b or 0 */
|
|
|
00db10 |
- u_int64_t low;
|
|
|
00db10 |
- GET_LDOUBLE_LSW64(low,b);
|
|
|
00db10 |
- if((hb|(low&0x7fffffffffffffffLL))==0) return a;
|
|
|
00db10 |
- t1=two1022; /* t1=2^1022 */
|
|
|
00db10 |
- b *= t1;
|
|
|
00db10 |
- a *= t1;
|
|
|
00db10 |
- k -= 1022;
|
|
|
00db10 |
- kld = kld / two1022;
|
|
|
00db10 |
+ if(hb==0) return a;
|
|
|
00db10 |
+ a *= 0x1p+1022L;
|
|
|
00db10 |
+ b *= 0x1p+1022L;
|
|
|
00db10 |
+ k = -1022;
|
|
|
00db10 |
+ kld = 0x1p-1022L;
|
|
|
00db10 |
} else { /* scale a and b by 2^600 */
|
|
|
00db10 |
- ha += 0x2580000000000000LL; /* a *= 2^600 */
|
|
|
00db10 |
- hb += 0x2580000000000000LL; /* b *= 2^600 */
|
|
|
00db10 |
- k -= 600;
|
|
|
00db10 |
- a *= two600;
|
|
|
00db10 |
- b *= two600;
|
|
|
00db10 |
- kld = kld / two600;
|
|
|
00db10 |
+ a *= 0x1p+600L;
|
|
|
00db10 |
+ b *= 0x1p+600L;
|
|
|
00db10 |
+ k = -600;
|
|
|
00db10 |
+ kld = 0x1p-600L;
|
|
|
00db10 |
}
|
|
|
00db10 |
}
|
|
|
00db10 |
/* medium size a and b */
|
|
|
00db10 |
w = a-b;
|
|
|
00db10 |
if (w>b) {
|
|
|
00db10 |
- SET_LDOUBLE_WORDS64(t1,ha,0);
|
|
|
00db10 |
- t2 = a-t1;
|
|
|
00db10 |
- w = __ieee754_sqrtl(t1*t1-(b*(-b)-t2*(a+t1)));
|
|
|
00db10 |
+ ldbl_unpack (a, &hi, &lo);
|
|
|
00db10 |
+ a1 = hi;
|
|
|
00db10 |
+ a2 = lo;
|
|
|
00db10 |
+ /* a*a + b*b
|
|
|
00db10 |
+ = (a1+a2)*a + b*b
|
|
|
00db10 |
+ = a1*a + a2*a + b*b
|
|
|
00db10 |
+ = a1*(a1+a2) + a2*a + b*b
|
|
|
00db10 |
+ = a1*a1 + a1*a2 + a2*a + b*b
|
|
|
00db10 |
+ = a1*a1 + a2*(a+a1) + b*b */
|
|
|
00db10 |
+ w = __ieee754_sqrtl(a1*a1-(b*(-b)-a2*(a+a1)));
|
|
|
00db10 |
} else {
|
|
|
00db10 |
a = a+a;
|
|
|
00db10 |
- SET_LDOUBLE_WORDS64(y1,hb,0);
|
|
|
00db10 |
- y2 = b - y1;
|
|
|
00db10 |
- SET_LDOUBLE_WORDS64(t1,ha+0x0010000000000000LL,0);
|
|
|
00db10 |
- t2 = a - t1;
|
|
|
00db10 |
- w = __ieee754_sqrtl(t1*y1-(w*(-w)-(t1*y2+t2*b)));
|
|
|
00db10 |
+ ldbl_unpack (b, &hi, &lo);
|
|
|
00db10 |
+ b1 = hi;
|
|
|
00db10 |
+ b2 = lo;
|
|
|
00db10 |
+ ldbl_unpack (a, &hi, &lo);
|
|
|
00db10 |
+ a1 = hi;
|
|
|
00db10 |
+ a2 = lo;
|
|
|
00db10 |
+ /* a*a + b*b
|
|
|
00db10 |
+ = a*a + (a-b)*(a-b) - (a-b)*(a-b) + b*b
|
|
|
00db10 |
+ = a*a + w*w - (a*a - 2*a*b + b*b) + b*b
|
|
|
00db10 |
+ = w*w + 2*a*b
|
|
|
00db10 |
+ = w*w + (a1+a2)*b
|
|
|
00db10 |
+ = w*w + a1*b + a2*b
|
|
|
00db10 |
+ = w*w + a1*(b1+b2) + a2*b
|
|
|
00db10 |
+ = w*w + a1*b1 + a1*b2 + a2*b */
|
|
|
00db10 |
+ w = __ieee754_sqrtl(a1*b1-(w*(-w)-(a1*b2+a2*b)));
|
|
|
00db10 |
}
|
|
|
00db10 |
if(k!=0)
|
|
|
00db10 |
return w*kld;
|
|
|
00db10 |
diff -urN glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_remainderl.c glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_remainderl.c
|
|
|
00db10 |
--- glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_remainderl.c 2014-05-27 20:02:27.000000000 -0500
|
|
|
00db10 |
+++ glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/e_remainderl.c 2014-05-27 20:04:08.000000000 -0500
|
|
|
00db10 |
@@ -33,18 +33,22 @@
|
|
|
00db10 |
int64_t hx,hp;
|
|
|
00db10 |
u_int64_t sx,lx,lp;
|
|
|
00db10 |
long double p_half;
|
|
|
00db10 |
+ double xhi, xlo, phi, plo;
|
|
|
00db10 |
|
|
|
00db10 |
- GET_LDOUBLE_WORDS64(hx,lx,x);
|
|
|
00db10 |
- GET_LDOUBLE_WORDS64(hp,lp,p);
|
|
|
00db10 |
+ ldbl_unpack (x, &xhi, &xlo;;
|
|
|
00db10 |
+ EXTRACT_WORDS64 (hx, xhi);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (lx, xlo);
|
|
|
00db10 |
+ ldbl_unpack (p, &phi, &plo;;
|
|
|
00db10 |
+ EXTRACT_WORDS64 (hp, phi);
|
|
|
00db10 |
+ EXTRACT_WORDS64 (lp, plo);
|
|
|
00db10 |
sx = hx&0x8000000000000000ULL;
|
|
|
00db10 |
hp &= 0x7fffffffffffffffLL;
|
|
|
00db10 |
hx &= 0x7fffffffffffffffLL;
|
|
|
00db10 |
|
|
|
00db10 |
/* purge off exception values */
|
|
|
00db10 |
- if((hp|(lp&0x7fffffffffffffff))==0) return (x*p)/(x*p); /* p = 0 */
|
|
|
00db10 |
+ if(hp==0) return (x*p)/(x*p); /* p = 0 */
|
|
|
00db10 |
if((hx>=0x7ff0000000000000LL)|| /* x not finite */
|
|
|
00db10 |
- ((hp>=0x7ff0000000000000LL)&& /* p is NaN */
|
|
|
00db10 |
- (((hp-0x7ff0000000000000LL)|lp)!=0)))
|
|
|
00db10 |
+ (hp>0x7ff0000000000000LL)) /* p is NaN */
|
|
|
00db10 |
return (x*p)/(x*p);
|
|
|
00db10 |
|
|
|
00db10 |
|
|
|
00db10 |
@@ -64,8 +68,8 @@
|
|
|
00db10 |
if(x>=p_half) x -= p;
|
|
|
00db10 |
}
|
|
|
00db10 |
}
|
|
|
00db10 |
- GET_LDOUBLE_MSW64(hx,x);
|
|
|
00db10 |
- SET_LDOUBLE_MSW64(x,hx^sx);
|
|
|
00db10 |
+ if (sx)
|
|
|
00db10 |
+ x = -x;
|
|
|
00db10 |
return x;
|
|
|
00db10 |
}
|
|
|
00db10 |
strong_alias (__ieee754_remainderl, __remainderl_finite)
|
|
|
00db10 |
diff -urN glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/s_erfl.c glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/s_erfl.c
|
|
|
00db10 |
--- glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/s_erfl.c 2014-05-27 20:02:27.000000000 -0500
|
|
|
00db10 |
+++ glibc-2.17-c758a686/sysdeps/ieee754/ldbl-128ibm/s_erfl.c 2014-05-27 20:04:08.000000000 -0500
|
|
|
00db10 |
@@ -760,16 +760,16 @@
|
|
|
00db10 |
__erfl (long double x)
|
|
|
00db10 |
{
|
|
|
00db10 |
long double a, y, z;
|
|
|
00db10 |
- int32_t i, ix, sign;
|
|
|
00db10 |
- ieee854_long_double_shape_type u;
|
|
|
00db10 |
+ int32_t i, ix, hx;
|
|
|
00db10 |
+ double xhi;
|
|
|
00db10 |
|
|
|
00db10 |
- u.value = x;
|
|
|
00db10 |
- sign = u.parts32.w0;
|
|
|
00db10 |
- ix = sign & 0x7fffffff;
|
|
|
00db10 |
+ xhi = ldbl_high (x);
|
|
|
00db10 |
+ GET_HIGH_WORD (hx, xhi);
|
|
|
00db10 |
+ ix = hx & 0x7fffffff;
|
|
|
00db10 |
|
|
|
00db10 |
if (ix >= 0x7ff00000)
|
|
|
00db10 |
{ /* erf(nan)=nan */
|
|
|
00db10 |
- i = ((sign & 0xfff00000) >> 31) << 1;
|
|
|
00db10 |
+ i = ((uint32_t) hx >> 31) << 1;
|
|
|
00db10 |
return (long double) (1 - i) + one / x; /* erf(+-inf)=+-1 */
|
|
|
00db10 |
}
|
|
|
00db10 |
|
|
|
00db10 |
@@ -778,7 +778,7 @@
|
|
|
00db10 |
if (ix >= 0x4039A0DE)
|
|
|
00db10 |
{
|
|
|
00db10 |
/* __erfcl (x) underflows if x > 25.6283 */
|
|
|
00db10 |
- if (sign)
|
|
|
00db10 |
+ if ((hx & 0x80000000) == 0)
|
|
|
00db10 |
return one-tiny;
|
|
|
00db10 |
else
|
|
|
00db10 |
return tiny-one;
|
|
|
00db10 |
@@ -789,8 +789,9 @@
|
|
|
00db10 |
return (one - y);
|
|
|
00db10 |
}
|
|
|
00db10 |
}
|
|
|
00db10 |
- u.parts32.w0 = ix;
|
|
|
00db10 |
- a = u.value;
|
|
|
00db10 |
+ a = x;
|
|
|
00db10 |
+ if ((hx & 0x80000000) != 0)
|
|
|
00db10 |
+ a = -a;
|
|
|
00db10 |
z = x * x;
|
|
|
00db10 |
if (ix < 0x3fec0000) /* a < 0.875 */
|
|
|
00db10 |
{
|
|
|
00db10 |
@@ -814,7 +815,7 @@
|
|
|
00db10 |
y = erf_const + neval (a, TN2, NTN2) / deval (a, TD2, NTD2);
|
|
|
00db10 |
}
|
|
|
00db10 |
|
|
|
00db10 |
- if (sign & 0x80000000) /* x < 0 */
|
|
|
00db10 |
+ if (hx & 0x80000000) /* x < 0 */
|
|
|
00db10 |
y = -y;
|
|
|
00db10 |
return( y );
|
|
|
00db10 |
}
|
|
|
00db10 |
@@ -824,18 +825,18 @@
|
|
|
00db10 |
__erfcl (long double x)
|
|
|
00db10 |
{
|
|
|
00db10 |
long double y, z, p, r;
|
|
|
00db10 |
- int32_t i, ix, sign;
|
|
|
00db10 |
- ieee854_long_double_shape_type u;
|
|
|
00db10 |
-
|
|
|
00db10 |
- u.value = x;
|
|
|
00db10 |
- sign = u.parts32.w0;
|
|
|
00db10 |
- ix = sign & 0x7fffffff;
|
|
|
00db10 |
- u.parts32.w0 = ix;
|
|
|
00db10 |
+ int32_t i, ix;
|
|
|
00db10 |
+ uint32_t hx;
|
|
|
00db10 |
+ double xhi;
|
|
|
00db10 |
+
|
|
|
00db10 |
+ xhi = ldbl_high (x);
|
|
|
00db10 |
+ GET_HIGH_WORD (hx, xhi);
|
|
|
00db10 |
+ ix = hx & 0x7fffffff;
|
|
|
00db10 |
|
|
|
00db10 |
if (ix >= 0x7ff00000)
|
|
|
00db10 |
{ /* erfc(nan)=nan */
|
|
|
00db10 |
/* erfc(+-inf)=0,2 */
|
|
|
00db10 |
- return (long double) (((u_int32_t) sign >> 31) << 1) + one / x;
|
|
|
00db10 |
+ return (long double) ((hx >> 31) << 1) + one / x;
|
|
|
00db10 |
}
|
|
|
00db10 |
|
|
|
00db10 |
if (ix < 0x3fd00000) /* |x| <1/4 */
|
|
|
00db10 |
@@ -846,7 +847,8 @@
|
|
|
00db10 |
}
|
|
|
00db10 |
if (ix < 0x3ff40000) /* 1.25 */
|
|
|
00db10 |
{
|
|
|
00db10 |
- x = u.value;
|
|
|
00db10 |
+ if ((hx & 0x80000000) != 0)
|
|
|
00db10 |
+ x = -x;
|
|
|
00db10 |
i = 8.0 * x;
|
|
|
00db10 |
switch (i)
|
|
|
00db10 |
{
|
|
|
00db10 |
@@ -891,7 +893,7 @@
|
|
|
00db10 |
y += C20a;
|
|
|
00db10 |
break;
|
|
|
00db10 |
}
|
|
|
00db10 |
- if (sign & 0x80000000)
|
|
|
00db10 |
+ if (hx & 0x80000000)
|
|
|
00db10 |
y = 2.0L - y;
|
|
|
00db10 |
return y;
|
|
|
00db10 |
}
|
|
|
00db10 |
@@ -899,10 +901,11 @@
|
|
|
00db10 |
if (ix < 0x405ac000)
|
|
|
00db10 |
{
|
|
|
00db10 |
/* x < -9 */
|
|
|
00db10 |
- if ((ix >= 0x40220000) && (sign & 0x80000000))
|
|
|
00db10 |
+ if (hx >= 0xc0220000)
|
|
|
00db10 |
return two - tiny;
|
|
|
00db10 |
|
|
|
00db10 |
- x = fabsl (x);
|
|
|
00db10 |
+ if ((hx & 0x80000000) != 0)
|
|
|
00db10 |
+ x = -x;
|
|
|
00db10 |
z = one / (x * x);
|
|
|
00db10 |
i = 8.0 / x;
|
|
|
00db10 |
switch (i)
|
|
|
00db10 |
@@ -933,21 +936,17 @@
|
|
|
00db10 |
p = neval (z, RNr8, NRNr8) / deval (z, RDr8, NRDr8);
|
|
|
00db10 |
break;
|
|
|
00db10 |
}
|
|
|
00db10 |
- u.value = x;
|
|
|
00db10 |
- u.parts32.w3 = 0;
|
|
|
00db10 |
- u.parts32.w2 = 0;
|
|
|
00db10 |
- u.parts32.w1 &= 0xf8000000;
|
|
|
00db10 |
- z = u.value;
|
|
|
00db10 |
+ z = (float) x;
|
|
|
00db10 |
r = __ieee754_expl (-z * z - 0.5625) *
|
|
|
00db10 |
__ieee754_expl ((z - x) * (z + x) + p);
|
|
|
00db10 |
- if ((sign & 0x80000000) == 0)
|
|
|
00db10 |
+ if ((hx & 0x80000000) == 0)
|
|
|
00db10 |
return r / x;
|
|
|
00db10 |
else
|
|
|
00db10 |
return two - r / x;
|
|
|
00db10 |
}
|
|
|
00db10 |
else
|
|
|
00db10 |
{
|
|
|
00db10 |
- if ((sign & 0x80000000) == 0)
|
|
|
00db10 |
+ if ((hx & 0x80000000) == 0)
|
|
|
00db10 |
return tiny * tiny;
|
|
|
00db10 |
else
|
|
|
00db10 |
return two - tiny;
|