Retro68/gcc/libgcc/config/libbid/bid64_sqrt.c
2014-09-21 19:33:12 +02:00

553 lines
14 KiB
C

/* Copyright (C) 2007-2014 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC 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 General Public License
for more details.
Under Section 7 of GPL version 3, you are granted additional
permissions described in the GCC Runtime Library Exception, version
3.1, as published by the Free Software Foundation.
You should have received a copy of the GNU General Public License and
a copy of the GCC Runtime Library Exception along with this program;
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
<http://www.gnu.org/licenses/>. */
/*****************************************************************************
* BID64 square root
*****************************************************************************
*
* Algorithm description:
*
* if(exponent_x is odd)
* scale coefficient_x by 10, adjust exponent
* - get lower estimate for number of digits in coefficient_x
* - scale coefficient x to between 31 and 33 decimal digits
* - in parallel, check for exact case and return if true
* - get high part of result coefficient using double precision sqrt
* - compute remainder and refine coefficient in one iteration (which
* modifies it by at most 1)
* - result exponent is easy to compute from the adjusted arg. exponent
*
****************************************************************************/
#include "bid_internal.h"
#include "bid_sqrt_macros.h"
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
#include <fenv.h>
#define FE_ALL_FLAGS FE_INVALID|FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW|FE_INEXACT
#endif
extern double sqrt (double);
#if DECIMAL_CALL_BY_REFERENCE
void
bid64_sqrt (UINT64 * pres,
UINT64 *
px _RND_MODE_PARAM _EXC_FLAGS_PARAM _EXC_MASKS_PARAM
_EXC_INFO_PARAM) {
UINT64 x;
#else
UINT64
bid64_sqrt (UINT64 x _RND_MODE_PARAM _EXC_FLAGS_PARAM
_EXC_MASKS_PARAM _EXC_INFO_PARAM) {
#endif
UINT128 CA, CT;
UINT64 sign_x, coefficient_x;
UINT64 Q, Q2, A10, C4, R, R2, QE, res;
SINT64 D;
int_double t_scale;
int_float tempx;
double da, dq, da_h, da_l, dqe;
int exponent_x, exponent_q, bin_expon_cx;
int digits_x;
int scale;
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
fexcept_t binaryflags = 0;
#endif
#if DECIMAL_CALL_BY_REFERENCE
#if !DECIMAL_GLOBAL_ROUNDING
_IDEC_round rnd_mode = *prnd_mode;
#endif
x = *px;
#endif
// unpack arguments, check for NaN or Infinity
if (!unpack_BID64 (&sign_x, &exponent_x, &coefficient_x, x)) {
// x is Inf. or NaN or 0
if ((x & INFINITY_MASK64) == INFINITY_MASK64) {
res = coefficient_x;
if ((coefficient_x & SSNAN_MASK64) == SINFINITY_MASK64) // -Infinity
{
res = NAN_MASK64;
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
}
#ifdef SET_STATUS_FLAGS
if ((x & SNAN_MASK64) == SNAN_MASK64) // sNaN
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
BID_RETURN (res & QUIET_MASK64);
}
// x is 0
exponent_x = (exponent_x + DECIMAL_EXPONENT_BIAS) >> 1;
res = sign_x | (((UINT64) exponent_x) << 53);
BID_RETURN (res);
}
// x<0?
if (sign_x && coefficient_x) {
res = NAN_MASK64;
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
BID_RETURN (res);
}
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fegetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
//--- get number of bits in the coefficient of x ---
tempx.d = (float) coefficient_x;
bin_expon_cx = ((tempx.i >> 23) & 0xff) - 0x7f;
digits_x = estimate_decimal_digits[bin_expon_cx];
// add test for range
if (coefficient_x >= power10_index_binexp[bin_expon_cx])
digits_x++;
A10 = coefficient_x;
if (exponent_x & 1) {
A10 = (A10 << 2) + A10;
A10 += A10;
}
dqe = sqrt ((double) A10);
QE = (UINT32) dqe;
if (QE * QE == A10) {
res =
very_fast_get_BID64 (0, (exponent_x + DECIMAL_EXPONENT_BIAS) >> 1,
QE);
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fesetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
BID_RETURN (res);
}
// if exponent is odd, scale coefficient by 10
scale = 31 - digits_x;
exponent_q = exponent_x - scale;
scale += (exponent_q & 1); // exp. bias is even
CT = power10_table_128[scale];
__mul_64x128_short (CA, coefficient_x, CT);
// 2^64
t_scale.i = 0x43f0000000000000ull;
// convert CA to DP
da_h = CA.w[1];
da_l = CA.w[0];
da = da_h * t_scale.d + da_l;
dq = sqrt (da);
Q = (UINT64) dq;
// get sign(sqrt(CA)-Q)
R = CA.w[0] - Q * Q;
R = ((SINT64) R) >> 63;
D = R + R + 1;
exponent_q = (exponent_q + DECIMAL_EXPONENT_BIAS) >> 1;
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INEXACT_EXCEPTION);
#endif
#ifndef IEEE_ROUND_NEAREST
#ifndef IEEE_ROUND_NEAREST_TIES_AWAY
if (!((rnd_mode) & 3)) {
#endif
#endif
// midpoint to check
Q2 = Q + Q + D;
C4 = CA.w[0] << 2;
// get sign(-sqrt(CA)+Midpoint)
R2 = Q2 * Q2 - C4;
R2 = ((SINT64) R2) >> 63;
// adjust Q if R!=R2
Q += (D & (R ^ R2));
#ifndef IEEE_ROUND_NEAREST
#ifndef IEEE_ROUND_NEAREST_TIES_AWAY
} else {
C4 = CA.w[0];
Q += D;
if ((SINT64) (Q * Q - C4) > 0)
Q--;
if (rnd_mode == ROUNDING_UP)
Q++;
}
#endif
#endif
res = fast_get_BID64 (0, exponent_q, Q);
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fesetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
BID_RETURN (res);
}
TYPE0_FUNCTION_ARG1 (UINT64, bid64q_sqrt, x)
UINT256 M256, C4, C8;
UINT128 CX, CX2, A10, S2, T128, CS, CSM, CS2, C256, CS1,
mul_factor2_long = { {0x0ull, 0x0ull} }, QH, Tmp, TP128, Qh, Ql;
UINT64 sign_x, Carry, B10, res, mul_factor, mul_factor2 = 0x0ull, CS0;
SINT64 D;
int_float fx, f64;
int exponent_x, bin_expon_cx, done = 0;
int digits, scale, exponent_q = 0, exact = 1, amount, extra_digits;
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
fexcept_t binaryflags = 0;
#endif
// unpack arguments, check for NaN or Infinity
if (!unpack_BID128_value (&sign_x, &exponent_x, &CX, x)) {
res = CX.w[1];
// NaN ?
if ((x.w[1] & 0x7c00000000000000ull) == 0x7c00000000000000ull) {
#ifdef SET_STATUS_FLAGS
if ((x.w[1] & 0x7e00000000000000ull) == 0x7e00000000000000ull) // sNaN
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
Tmp.w[1] = (CX.w[1] & 0x00003fffffffffffull);
Tmp.w[0] = CX.w[0];
TP128 = reciprocals10_128[18];
__mul_128x128_full (Qh, Ql, Tmp, TP128);
amount = recip_scale[18];
__shr_128 (Tmp, Qh, amount);
res = (CX.w[1] & 0xfc00000000000000ull) | Tmp.w[0];
BID_RETURN (res);
}
// x is Infinity?
if ((x.w[1] & 0x7800000000000000ull) == 0x7800000000000000ull) {
if (sign_x) {
// -Inf, return NaN
res = 0x7c00000000000000ull;
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
}
BID_RETURN (res);
}
// x is 0 otherwise
exponent_x =
((exponent_x - DECIMAL_EXPONENT_BIAS_128) >> 1) +
DECIMAL_EXPONENT_BIAS;
if (exponent_x < 0)
exponent_x = 0;
if (exponent_x > DECIMAL_MAX_EXPON_64)
exponent_x = DECIMAL_MAX_EXPON_64;
//res= sign_x | (((UINT64)exponent_x)<<53);
res = get_BID64 (sign_x, exponent_x, 0, rnd_mode, pfpsf);
BID_RETURN (res);
}
if (sign_x) {
res = 0x7c00000000000000ull;
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INVALID_EXCEPTION);
#endif
BID_RETURN (res);
}
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fegetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
// 2^64
f64.i = 0x5f800000;
// fx ~ CX
fx.d = (float) CX.w[1] * f64.d + (float) CX.w[0];
bin_expon_cx = ((fx.i >> 23) & 0xff) - 0x7f;
digits = estimate_decimal_digits[bin_expon_cx];
A10 = CX;
if (exponent_x & 1) {
A10.w[1] = (CX.w[1] << 3) | (CX.w[0] >> 61);
A10.w[0] = CX.w[0] << 3;
CX2.w[1] = (CX.w[1] << 1) | (CX.w[0] >> 63);
CX2.w[0] = CX.w[0] << 1;
__add_128_128 (A10, A10, CX2);
}
C256.w[1] = A10.w[1];
C256.w[0] = A10.w[0];
CS.w[0] = short_sqrt128 (A10);
CS.w[1] = 0;
mul_factor = 0;
// check for exact result
if (CS.w[0] < 10000000000000000ull) {
if (CS.w[0] * CS.w[0] == A10.w[0]) {
__sqr64_fast (S2, CS.w[0]);
if (S2.w[1] == A10.w[1]) // && S2.w[0]==A10.w[0])
{
res =
get_BID64 (0,
((exponent_x - DECIMAL_EXPONENT_BIAS_128) >> 1) +
DECIMAL_EXPONENT_BIAS, CS.w[0], rnd_mode, pfpsf);
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fesetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
BID_RETURN (res);
}
}
if (CS.w[0] >= 1000000000000000ull) {
done = 1;
exponent_q = exponent_x;
C256.w[1] = A10.w[1];
C256.w[0] = A10.w[0];
}
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INEXACT_EXCEPTION);
#endif
exact = 0;
} else {
B10 = 0x3333333333333334ull;
__mul_64x64_to_128_full (CS2, CS.w[0], B10);
CS0 = CS2.w[1] >> 1;
if (CS.w[0] != ((CS0 << 3) + (CS0 << 1))) {
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INEXACT_EXCEPTION);
#endif
exact = 0;
}
done = 1;
CS.w[0] = CS0;
exponent_q = exponent_x + 2;
mul_factor = 10;
mul_factor2 = 100;
if (CS.w[0] >= 10000000000000000ull) {
__mul_64x64_to_128_full (CS2, CS.w[0], B10);
CS0 = CS2.w[1] >> 1;
if (CS.w[0] != ((CS0 << 3) + (CS0 << 1))) {
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INEXACT_EXCEPTION);
#endif
exact = 0;
}
exponent_q += 2;
CS.w[0] = CS0;
mul_factor = 100;
mul_factor2 = 10000;
}
if (exact) {
CS0 = CS.w[0] * mul_factor;
__sqr64_fast (CS1, CS0)
if ((CS1.w[0] != A10.w[0]) || (CS1.w[1] != A10.w[1])) {
#ifdef SET_STATUS_FLAGS
__set_status_flags (pfpsf, INEXACT_EXCEPTION);
#endif
exact = 0;
}
}
}
if (!done) {
// get number of digits in CX
D = CX.w[1] - power10_index_binexp_128[bin_expon_cx].w[1];
if (D > 0
|| (!D && CX.w[0] >= power10_index_binexp_128[bin_expon_cx].w[0]))
digits++;
// if exponent is odd, scale coefficient by 10
scale = 31 - digits;
exponent_q = exponent_x - scale;
scale += (exponent_q & 1); // exp. bias is even
T128 = power10_table_128[scale];
__mul_128x128_low (C256, CX, T128);
CS.w[0] = short_sqrt128 (C256);
}
//printf("CS=%016I64x\n",CS.w[0]);
exponent_q =
((exponent_q - DECIMAL_EXPONENT_BIAS_128) >> 1) +
DECIMAL_EXPONENT_BIAS;
if ((exponent_q < 0) && (exponent_q + MAX_FORMAT_DIGITS >= 0)) {
extra_digits = -exponent_q;
exponent_q = 0;
// get coeff*(2^M[extra_digits])/10^extra_digits
__mul_64x64_to_128 (QH, CS.w[0], reciprocals10_64[extra_digits]);
// now get P/10^extra_digits: shift Q_high right by M[extra_digits]-128
amount = short_recip_scale[extra_digits];
CS0 = QH.w[1] >> amount;
#ifdef SET_STATUS_FLAGS
if (exact) {
if (CS.w[0] != CS0 * power10_table_128[extra_digits].w[0])
exact = 0;
}
if (!exact)
__set_status_flags (pfpsf, UNDERFLOW_EXCEPTION | INEXACT_EXCEPTION);
#endif
CS.w[0] = CS0;
if (!mul_factor)
mul_factor = 1;
mul_factor *= power10_table_128[extra_digits].w[0];
__mul_64x64_to_128 (mul_factor2_long, mul_factor, mul_factor);
if (mul_factor2_long.w[1])
mul_factor2 = 0;
else
mul_factor2 = mul_factor2_long.w[1];
}
// 4*C256
C4.w[1] = (C256.w[1] << 2) | (C256.w[0] >> 62);
C4.w[0] = C256.w[0] << 2;
#ifndef IEEE_ROUND_NEAREST
#ifndef IEEE_ROUND_NEAREST_TIES_AWAY
if (!((rnd_mode) & 3)) {
#endif
#endif
// compare to midpoints
CSM.w[0] = (CS.w[0] + CS.w[0]) | 1;
//printf("C256=%016I64x %016I64x, CSM=%016I64x %016I64x %016I64x\n",C4.w[1],C4.w[0],CSM.w[1],CSM.w[0], CS.w[0]);
if (mul_factor)
CSM.w[0] *= mul_factor;
// CSM^2
__mul_64x64_to_128 (M256, CSM.w[0], CSM.w[0]);
//__mul_128x128_to_256(M256, CSM, CSM);
if (C4.w[1] > M256.w[1] ||
(C4.w[1] == M256.w[1] && C4.w[0] > M256.w[0])) {
// round up
CS.w[0]++;
} else {
C8.w[0] = CS.w[0] << 3;
C8.w[1] = 0;
if (mul_factor) {
if (mul_factor2) {
__mul_64x64_to_128 (C8, C8.w[0], mul_factor2);
} else {
__mul_64x128_low (C8, C8.w[0], mul_factor2_long);
}
}
// M256 - 8*CSM
__sub_borrow_out (M256.w[0], Carry, M256.w[0], C8.w[0]);
M256.w[1] = M256.w[1] - C8.w[1] - Carry;
// if CSM' > C256, round up
if (M256.w[1] > C4.w[1] ||
(M256.w[1] == C4.w[1] && M256.w[0] > C4.w[0])) {
// round down
if (CS.w[0])
CS.w[0]--;
}
}
#ifndef IEEE_ROUND_NEAREST
#ifndef IEEE_ROUND_NEAREST_TIES_AWAY
} else {
CS.w[0]++;
CSM.w[0] = CS.w[0];
C8.w[0] = CSM.w[0] << 1;
if (mul_factor)
CSM.w[0] *= mul_factor;
__mul_64x64_to_128 (M256, CSM.w[0], CSM.w[0]);
C8.w[1] = 0;
if (mul_factor) {
if (mul_factor2) {
__mul_64x64_to_128 (C8, C8.w[0], mul_factor2);
} else {
__mul_64x128_low (C8, C8.w[0], mul_factor2_long);
}
}
//printf("C256=%016I64x %016I64x, CSM=%016I64x %016I64x %016I64x\n",C256.w[1],C256.w[0],M256.w[1],M256.w[0], CS.w[0]);
if (M256.w[1] > C256.w[1] ||
(M256.w[1] == C256.w[1] && M256.w[0] > C256.w[0])) {
__sub_borrow_out (M256.w[0], Carry, M256.w[0], C8.w[0]);
M256.w[1] = M256.w[1] - Carry - C8.w[1];
M256.w[0]++;
if (!M256.w[0]) {
M256.w[1]++;
}
if ((M256.w[1] > C256.w[1] ||
(M256.w[1] == C256.w[1] && M256.w[0] > C256.w[0]))
&& (CS.w[0] > 1)) {
CS.w[0]--;
if (CS.w[0] > 1) {
__sub_borrow_out (M256.w[0], Carry, M256.w[0], C8.w[0]);
M256.w[1] = M256.w[1] - Carry - C8.w[1];
M256.w[0]++;
if (!M256.w[0]) {
M256.w[1]++;
}
if (M256.w[1] > C256.w[1] ||
(M256.w[1] == C256.w[1] && M256.w[0] > C256.w[0]))
CS.w[0]--;
}
}
}
else {
/*__add_carry_out(M256.w[0], Carry, M256.w[0], C8.w[0]);
M256.w[1] = M256.w[1] + Carry + C8.w[1];
M256.w[0]++;
if(!M256.w[0])
{
M256.w[1]++;
}
CS.w[0]++;
if(M256.w[1]<C256.w[1] ||
(M256.w[1]==C256.w[1] && M256.w[0]<=C256.w[0]))
{
CS.w[0]++;
}*/
CS.w[0]++;
}
//printf("C256=%016I64x %016I64x, CSM=%016I64x %016I64x %016I64x %d\n",C4.w[1],C4.w[0],M256.w[1],M256.w[0], CS.w[0], exact);
// RU?
if (((rnd_mode) != ROUNDING_UP) || exact) {
if (CS.w[0])
CS.w[0]--;
}
}
#endif
#endif
//printf("C256=%016I64x %016I64x, CSM=%016I64x %016I64x %016I64x %d\n",C4.w[1],C4.w[0],M256.w[1],M256.w[0], CS.w[0], exact);
res = get_BID64 (0, exponent_q, CS.w[0], rnd_mode, pfpsf);
#ifdef UNCHANGED_BINARY_STATUS_FLAGS
(void) fesetexceptflag (&binaryflags, FE_ALL_FLAGS);
#endif
BID_RETURN (res);
}