1*0a6a1f1dSLionel Sambuc /* This file is distributed under the University of Illinois Open Source
2*0a6a1f1dSLionel Sambuc * License. See LICENSE.TXT for details.
3*0a6a1f1dSLionel Sambuc */
4*0a6a1f1dSLionel Sambuc
5*0a6a1f1dSLionel Sambuc /* long double __gcc_qsub(long double x, long double y);
6*0a6a1f1dSLionel Sambuc * This file implements the PowerPC 128-bit double-double add operation.
7*0a6a1f1dSLionel Sambuc * This implementation is shamelessly cribbed from Apple's DDRT, circa 1993(!)
8*0a6a1f1dSLionel Sambuc */
9*0a6a1f1dSLionel Sambuc
10*0a6a1f1dSLionel Sambuc #include "DD.h"
11*0a6a1f1dSLionel Sambuc
__gcc_qsub(long double x,long double y)12*0a6a1f1dSLionel Sambuc long double __gcc_qsub(long double x, long double y)
13*0a6a1f1dSLionel Sambuc {
14*0a6a1f1dSLionel Sambuc static const uint32_t infinityHi = UINT32_C(0x7ff00000);
15*0a6a1f1dSLionel Sambuc
16*0a6a1f1dSLionel Sambuc DD dst = { .ld = x }, src = { .ld = y };
17*0a6a1f1dSLionel Sambuc
18*0a6a1f1dSLionel Sambuc register double A = dst.s.hi, a = dst.s.lo,
19*0a6a1f1dSLionel Sambuc B = -src.s.hi, b = -src.s.lo;
20*0a6a1f1dSLionel Sambuc
21*0a6a1f1dSLionel Sambuc /* If both operands are zero: */
22*0a6a1f1dSLionel Sambuc if ((A == 0.0) && (B == 0.0)) {
23*0a6a1f1dSLionel Sambuc dst.s.hi = A + B;
24*0a6a1f1dSLionel Sambuc dst.s.lo = 0.0;
25*0a6a1f1dSLionel Sambuc return dst.ld;
26*0a6a1f1dSLionel Sambuc }
27*0a6a1f1dSLionel Sambuc
28*0a6a1f1dSLionel Sambuc /* If either operand is NaN or infinity: */
29*0a6a1f1dSLionel Sambuc const doublebits abits = { .d = A };
30*0a6a1f1dSLionel Sambuc const doublebits bbits = { .d = B };
31*0a6a1f1dSLionel Sambuc if ((((uint32_t)(abits.x >> 32) & infinityHi) == infinityHi) ||
32*0a6a1f1dSLionel Sambuc (((uint32_t)(bbits.x >> 32) & infinityHi) == infinityHi)) {
33*0a6a1f1dSLionel Sambuc dst.s.hi = A + B;
34*0a6a1f1dSLionel Sambuc dst.s.lo = 0.0;
35*0a6a1f1dSLionel Sambuc return dst.ld;
36*0a6a1f1dSLionel Sambuc }
37*0a6a1f1dSLionel Sambuc
38*0a6a1f1dSLionel Sambuc /* If the computation overflows: */
39*0a6a1f1dSLionel Sambuc /* This may be playing things a little bit fast and loose, but it will do for a start. */
40*0a6a1f1dSLionel Sambuc const double testForOverflow = A + (B + (a + b));
41*0a6a1f1dSLionel Sambuc const doublebits testbits = { .d = testForOverflow };
42*0a6a1f1dSLionel Sambuc if (((uint32_t)(testbits.x >> 32) & infinityHi) == infinityHi) {
43*0a6a1f1dSLionel Sambuc dst.s.hi = testForOverflow;
44*0a6a1f1dSLionel Sambuc dst.s.lo = 0.0;
45*0a6a1f1dSLionel Sambuc return dst.ld;
46*0a6a1f1dSLionel Sambuc }
47*0a6a1f1dSLionel Sambuc
48*0a6a1f1dSLionel Sambuc double H, h;
49*0a6a1f1dSLionel Sambuc double T, t;
50*0a6a1f1dSLionel Sambuc double W, w;
51*0a6a1f1dSLionel Sambuc double Y;
52*0a6a1f1dSLionel Sambuc
53*0a6a1f1dSLionel Sambuc H = B + (A - (A + B));
54*0a6a1f1dSLionel Sambuc T = b + (a - (a + b));
55*0a6a1f1dSLionel Sambuc h = A + (B - (A + B));
56*0a6a1f1dSLionel Sambuc t = a + (b - (a + b));
57*0a6a1f1dSLionel Sambuc
58*0a6a1f1dSLionel Sambuc if (local_fabs(A) <= local_fabs(B))
59*0a6a1f1dSLionel Sambuc w = (a + b) + h;
60*0a6a1f1dSLionel Sambuc else
61*0a6a1f1dSLionel Sambuc w = (a + b) + H;
62*0a6a1f1dSLionel Sambuc
63*0a6a1f1dSLionel Sambuc W = (A + B) + w;
64*0a6a1f1dSLionel Sambuc Y = (A + B) - W;
65*0a6a1f1dSLionel Sambuc Y += w;
66*0a6a1f1dSLionel Sambuc
67*0a6a1f1dSLionel Sambuc if (local_fabs(a) <= local_fabs(b))
68*0a6a1f1dSLionel Sambuc w = t + Y;
69*0a6a1f1dSLionel Sambuc else
70*0a6a1f1dSLionel Sambuc w = T + Y;
71*0a6a1f1dSLionel Sambuc
72*0a6a1f1dSLionel Sambuc dst.s.hi = Y = W + w;
73*0a6a1f1dSLionel Sambuc dst.s.lo = (W - Y) + w;
74*0a6a1f1dSLionel Sambuc
75*0a6a1f1dSLionel Sambuc return dst.ld;
76*0a6a1f1dSLionel Sambuc }
77