2 * The implementations contained in this file are heavily based on the
3 * implementations found in the Berkeley SoftFloat library. As such, they are
4 * licensed under the same 3-clause BSD license:
6 * License for Berkeley SoftFloat Release 3e
11 * The following applies to the whole of SoftFloat Release 3e as well as to
12 * each source file individually.
14 * Copyright 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018 The Regents of the
15 * University of California. All rights reserved.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are met:
20 * 1. Redistributions of source code must retain the above copyright notice,
21 * this list of conditions, and the following disclaimer.
23 * 2. Redistributions in binary form must reproduce the above copyright
24 * notice, this list of conditions, and the following disclaimer in the
25 * documentation and/or other materials provided with the distribution.
27 * 3. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS "AS IS", AND ANY
32 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
33 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
34 * DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY
35 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
36 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
37 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
38 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
40 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44 #extension GL_ARB_gpu_shader_int64 : enable
45 #extension GL_ARB_shader_bit_encoding : enable
46 #extension GL_EXT_shader_integer_mix : enable
47 #extension GL_MESA_shader_integer_functions : enable
51 /* Software IEEE floating-point rounding mode.
52 * GLSL spec section "4.7.1 Range and Precision":
53 * The rounding mode cannot be set and is undefined.
54 * But here, we are able to define the rounding mode at the compilation time.
56 #define FLOAT_ROUND_NEAREST_EVEN 0
57 #define FLOAT_ROUND_TO_ZERO 1
58 #define FLOAT_ROUND_DOWN 2
59 #define FLOAT_ROUND_UP 3
60 #define FLOAT_ROUNDING_MODE FLOAT_ROUND_NEAREST_EVEN
62 /* Absolute value of a Float64 :
66 __fabs64(uint64_t __a)
68 uvec2 a = unpackUint2x32(__a);
70 return packUint2x32(a);
73 /* Returns 1 if the double-precision floating-point value `a' is a NaN;
74 * otherwise returns 0.
77 __is_nan(uint64_t __a)
79 uvec2 a = unpackUint2x32(__a);
80 return (0xFFE00000u <= (a.y<<1)) &&
81 ((a.x != 0u) || ((a.y & 0x000FFFFFu) != 0u));
84 /* Negate value of a Float64 :
88 __fneg64(uint64_t __a)
90 uvec2 a = unpackUint2x32(__a);
92 return packUint2x32(a);
96 __fsign64(uint64_t __a)
98 uvec2 a = unpackUint2x32(__a);
101 retval.y = mix((a.y & 0x80000000u) | 0x3FF00000u, 0u, (a.y << 1 | a.x) == 0u);
102 return packUint2x32(retval);
105 /* Returns the fraction bits of the double-precision floating-point value `a'.*/
107 __extractFloat64FracLo(uint64_t a)
109 return unpackUint2x32(a).x;
113 __extractFloat64FracHi(uint64_t a)
115 return unpackUint2x32(a).y & 0x000FFFFFu;
118 /* Returns the exponent bits of the double-precision floating-point value `a'.*/
120 __extractFloat64Exp(uint64_t __a)
122 uvec2 a = unpackUint2x32(__a);
123 return int((a.y>>20) & 0x7FFu);
127 __feq64_nonnan(uint64_t __a, uint64_t __b)
129 uvec2 a = unpackUint2x32(__a);
130 uvec2 b = unpackUint2x32(__b);
131 return (a.x == b.x) &&
132 ((a.y == b.y) || ((a.x == 0u) && (((a.y | b.y)<<1) == 0u)));
135 /* Returns true if the double-precision floating-point value `a' is equal to the
136 * corresponding value `b', and false otherwise. The comparison is performed
137 * according to the IEEE Standard for Floating-Point Arithmetic.
140 __feq64(uint64_t a, uint64_t b)
142 if (__is_nan(a) || __is_nan(b))
145 return __feq64_nonnan(a, b);
148 /* Returns true if the double-precision floating-point value `a' is not equal
149 * to the corresponding value `b', and false otherwise. The comparison is
150 * performed according to the IEEE Standard for Floating-Point Arithmetic.
153 __fne64(uint64_t a, uint64_t b)
155 if (__is_nan(a) || __is_nan(b))
158 return !__feq64_nonnan(a, b);
161 /* Returns the sign bit of the double-precision floating-point value `a'.*/
163 __extractFloat64Sign(uint64_t a)
165 return unpackUint2x32(a).y & 0x80000000u;
168 /* Returns true if the 64-bit value formed by concatenating `a0' and `a1' is less
169 * than the 64-bit value formed by concatenating `b0' and `b1'. Otherwise,
173 lt64(uint a0, uint a1, uint b0, uint b1)
175 return (a0 < b0) || ((a0 == b0) && (a1 < b1));
179 __flt64_nonnan(uint64_t __a, uint64_t __b)
181 uvec2 a = unpackUint2x32(__a);
182 uvec2 b = unpackUint2x32(__b);
183 uint aSign = __extractFloat64Sign(__a);
184 uint bSign = __extractFloat64Sign(__b);
186 return (aSign != 0u) && ((((a.y | b.y)<<1) | a.x | b.x) != 0u);
188 return mix(lt64(a.y, a.x, b.y, b.x), lt64(b.y, b.x, a.y, a.x), aSign != 0u);
191 /* Returns true if the double-precision floating-point value `a' is less than
192 * the corresponding value `b', and false otherwise. The comparison is performed
193 * according to the IEEE Standard for Floating-Point Arithmetic.
196 __flt64(uint64_t a, uint64_t b)
198 if (__is_nan(a) || __is_nan(b))
201 return __flt64_nonnan(a, b);
204 /* Returns true if the double-precision floating-point value `a' is greater
205 * than or equal to * the corresponding value `b', and false otherwise. The
206 * comparison is performed * according to the IEEE Standard for Floating-Point
210 __fge64(uint64_t a, uint64_t b)
212 if (__is_nan(a) || __is_nan(b))
215 return !__flt64_nonnan(a, b);
219 __fsat64(uint64_t __a)
221 if (__flt64(__a, 0ul))
224 if (__fge64(__a, 0x3FF0000000000000ul /* 1.0 */))
225 return 0x3FF0000000000000ul;
230 /* Adds the 64-bit value formed by concatenating `a0' and `a1' to the 64-bit
231 * value formed by concatenating `b0' and `b1'. Addition is modulo 2^64, so
232 * any carry out is lost. The result is broken into two 32-bit pieces which
233 * are stored at the locations pointed to by `z0Ptr' and `z1Ptr'.
236 __add64(uint a0, uint a1, uint b0, uint b1,
242 z0Ptr = a0 + b0 + uint(z1 < a1);
246 /* Subtracts the 64-bit value formed by concatenating `b0' and `b1' from the
247 * 64-bit value formed by concatenating `a0' and `a1'. Subtraction is modulo
248 * 2^64, so any borrow out (carry out) is lost. The result is broken into two
249 * 32-bit pieces which are stored at the locations pointed to by `z0Ptr' and
253 __sub64(uint a0, uint a1, uint b0, uint b1,
258 z0Ptr = a0 - b0 - uint(a1 < b1);
261 /* Shifts the 64-bit value formed by concatenating `a0' and `a1' right by the
262 * number of bits given in `count'. If any nonzero bits are shifted off, they
263 * are "jammed" into the least significant bit of the result by setting the
264 * least significant bit to 1. The value of `count' can be arbitrarily large;
265 * in particular, if `count' is greater than 64, the result will be either 0
266 * or 1, depending on whether the concatenation of `a0' and `a1' is zero or
267 * nonzero. The result is broken into two 32-bit pieces which are stored at
268 * the locations pointed to by `z0Ptr' and `z1Ptr'.
271 __shift64RightJamming(uint a0,
279 int negCount = (-count) & 31;
281 z0 = mix(0u, a0, count == 0);
282 z0 = mix(z0, (a0 >> count), count < 32);
284 z1 = uint((a0 | a1) != 0u); /* count >= 64 */
285 uint z1_lt64 = (a0>>(count & 31)) | uint(((a0<<negCount) | a1) != 0u);
286 z1 = mix(z1, z1_lt64, count < 64);
287 z1 = mix(z1, (a0 | uint(a1 != 0u)), count == 32);
288 uint z1_lt32 = (a0<<negCount) | (a1>>count) | uint ((a1<<negCount) != 0u);
289 z1 = mix(z1, z1_lt32, count < 32);
290 z1 = mix(z1, a1, count == 0);
295 /* Shifts the 96-bit value formed by concatenating `a0', `a1', and `a2' right
296 * by 32 _plus_ the number of bits given in `count'. The shifted result is
297 * at most 64 nonzero bits; these are broken into two 32-bit pieces which are
298 * stored at the locations pointed to by `z0Ptr' and `z1Ptr'. The bits shifted
299 * off form a third 32-bit result as follows: The _last_ bit shifted off is
300 * the most-significant bit of the extra result, and the other 31 bits of the
301 * extra result are all zero if and only if _all_but_the_last_ bits shifted off
302 * were all zero. This extra result is stored in the location pointed to by
303 * `z2Ptr'. The value of `count' can be arbitrarily large.
304 * (This routine makes more sense if `a0', `a1', and `a2' are considered
305 * to form a fixed-point value with binary point between `a1' and `a2'. This
306 * fixed-point value is shifted right by the number of bits given in `count',
307 * and the integer part of the result is returned at the locations pointed to
308 * by `z0Ptr' and `z1Ptr'. The fractional part of the result may be slightly
309 * corrupted as described above, and is returned at the location pointed to by
313 __shift64ExtraRightJamming(uint a0, uint a1, uint a2,
322 int negCount = (-count) & 31;
324 z2 = mix(uint(a0 != 0u), a0, count == 64);
325 z2 = mix(z2, a0 << negCount, count < 64);
326 z2 = mix(z2, a1 << negCount, count < 32);
328 z1 = mix(0u, (a0 >> (count & 31)), count < 64);
329 z1 = mix(z1, (a0<<negCount) | (a1>>count), count < 32);
331 a2 = mix(a2 | a1, a2, count < 32);
332 z0 = mix(z0, a0 >> count, count < 32);
333 z2 |= uint(a2 != 0u);
335 z0 = mix(z0, 0u, (count == 32));
336 z1 = mix(z1, a0, (count == 32));
337 z2 = mix(z2, a1, (count == 32));
338 z0 = mix(z0, a0, (count == 0));
339 z1 = mix(z1, a1, (count == 0));
340 z2 = mix(z2, a2, (count == 0));
346 /* Shifts the 64-bit value formed by concatenating `a0' and `a1' left by the
347 * number of bits given in `count'. Any bits shifted off are lost. The value
348 * of `count' must be less than 32. The result is broken into two 32-bit
349 * pieces which are stored at the locations pointed to by `z0Ptr' and `z1Ptr'.
352 __shortShift64Left(uint a0, uint a1,
358 z0Ptr = mix((a0 << count | (a1 >> ((-count) & 31))), a0, count == 0);
361 /* Packs the sign `zSign', the exponent `zExp', and the significand formed by
362 * the concatenation of `zFrac0' and `zFrac1' into a double-precision floating-
363 * point value, returning the result. After being shifted into the proper
364 * positions, the three fields `zSign', `zExp', and `zFrac0' are simply added
365 * together to form the most significant 32 bits of the result. This means
366 * that any integer portion of `zFrac0' will be added into the exponent. Since
367 * a properly normalized significand will have an integer portion equal to 1,
368 * the `zExp' input should be 1 less than the desired result exponent whenever
369 * `zFrac0' and `zFrac1' concatenated form a complete, normalized significand.
372 __packFloat64(uint zSign, int zExp, uint zFrac0, uint zFrac1)
376 z.y = zSign + (uint(zExp) << 20) + zFrac0;
378 return packUint2x32(z);
381 /* Takes an abstract floating-point value having sign `zSign', exponent `zExp',
382 * and extended significand formed by the concatenation of `zFrac0', `zFrac1',
383 * and `zFrac2', and returns the proper double-precision floating-point value
384 * corresponding to the abstract input. Ordinarily, the abstract value is
385 * simply rounded and packed into the double-precision format, with the inexact
386 * exception raised if the abstract input cannot be represented exactly.
387 * However, if the abstract value is too large, the overflow and inexact
388 * exceptions are raised and an infinity or maximal finite value is returned.
389 * If the abstract value is too small, the input value is rounded to a
390 * subnormal number, and the underflow and inexact exceptions are raised if the
391 * abstract input cannot be represented exactly as a subnormal double-precision
392 * floating-point number.
393 * The input significand must be normalized or smaller. If the input
394 * significand is not normalized, `zExp' must be 0; in that case, the result
395 * returned is a subnormal number, and it must not require rounding. In the
396 * usual case that the input significand is normalized, `zExp' must be 1 less
397 * than the "true" floating-point exponent. The handling of underflow and
398 * overflow follows the IEEE Standard for Floating-Point Arithmetic.
401 __roundAndPackFloat64(uint zSign,
407 bool roundNearestEven;
410 roundNearestEven = FLOAT_ROUNDING_MODE == FLOAT_ROUND_NEAREST_EVEN;
411 increment = int(zFrac2) < 0;
412 if (!roundNearestEven) {
413 if (FLOAT_ROUNDING_MODE == FLOAT_ROUND_TO_ZERO) {
417 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN) &&
420 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP) &&
426 if ((0x7FD < zExp) ||
428 (0x001FFFFFu == zFrac0 && 0xFFFFFFFFu == zFrac1) &&
430 if ((FLOAT_ROUNDING_MODE == FLOAT_ROUND_TO_ZERO) ||
431 ((zSign != 0u) && (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP)) ||
432 ((zSign == 0u) && (FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN))) {
433 return __packFloat64(zSign, 0x7FE, 0x000FFFFFu, 0xFFFFFFFFu);
435 return __packFloat64(zSign, 0x7FF, 0u, 0u);
438 __shift64ExtraRightJamming(
439 zFrac0, zFrac1, zFrac2, -zExp, zFrac0, zFrac1, zFrac2);
441 if (roundNearestEven) {
442 increment = zFrac2 < 0u;
445 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN) &&
448 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP) &&
455 __add64(zFrac0, zFrac1, 0u, 1u, zFrac0, zFrac1);
456 zFrac1 &= ~((zFrac2 + uint(zFrac2 == 0u)) & uint(roundNearestEven));
458 zExp = mix(zExp, 0, (zFrac0 | zFrac1) == 0u);
460 return __packFloat64(zSign, zExp, zFrac0, zFrac1);
464 __roundAndPackUInt64(uint zSign, uint zFrac0, uint zFrac1, uint zFrac2)
466 bool roundNearestEven;
468 uint64_t default_nan = 0xFFFFFFFFFFFFFFFFUL;
470 roundNearestEven = FLOAT_ROUNDING_MODE == FLOAT_ROUND_NEAREST_EVEN;
472 if (zFrac2 >= 0x80000000u)
475 if (!roundNearestEven) {
477 if ((FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN) && (zFrac2 != 0u)) {
481 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP) &&
487 __add64(zFrac0, zFrac1, 0u, 1u, zFrac0, zFrac1);
488 if ((zFrac0 | zFrac1) != 0u)
489 zFrac1 &= ~(1u) + uint(zFrac2 == 0u) & uint(roundNearestEven);
491 return mix(packUint2x32(uvec2(zFrac1, zFrac0)), default_nan,
492 (zSign != 0u && (zFrac0 | zFrac1) != 0u));
496 __roundAndPackInt64(uint zSign, uint zFrac0, uint zFrac1, uint zFrac2)
498 bool roundNearestEven;
500 int64_t default_NegNaN = -0x7FFFFFFFFFFFFFFEL;
501 int64_t default_PosNaN = 0xFFFFFFFFFFFFFFFFL;
503 roundNearestEven = FLOAT_ROUNDING_MODE == FLOAT_ROUND_NEAREST_EVEN;
505 if (zFrac2 >= 0x80000000u)
508 if (!roundNearestEven) {
510 increment = ((FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN) &&
513 increment = (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP) &&
519 __add64(zFrac0, zFrac1, 0u, 1u, zFrac0, zFrac1);
520 if ((zFrac0 | zFrac1) != 0u)
521 zFrac1 &= ~(1u) + uint(zFrac2 == 0u) & uint(roundNearestEven);
524 int64_t absZ = mix(int64_t(packUint2x32(uvec2(zFrac1, zFrac0))),
525 -int64_t(packUint2x32(uvec2(zFrac1, zFrac0))),
527 int64_t nan = mix(default_PosNaN, default_NegNaN, zSign != 0u);
528 return mix(absZ, nan, ((zSign != 0u) != (absZ < 0)) && bool(absZ));
531 /* Returns the number of leading 0 bits before the most-significant 1 bit of
532 * `a'. If `a' is zero, 32 is returned.
535 __countLeadingZeros32(uint a)
537 return 31 - findMSB(a);
540 /* Takes an abstract floating-point value having sign `zSign', exponent `zExp',
541 * and significand formed by the concatenation of `zSig0' and `zSig1', and
542 * returns the proper double-precision floating-point value corresponding
543 * to the abstract input. This routine is just like `__roundAndPackFloat64'
544 * except that the input significand has fewer bits and does not have to be
545 * normalized. In all cases, `zExp' must be 1 less than the "true" floating-
549 __normalizeRoundAndPackFloat64(uint zSign,
563 shiftCount = __countLeadingZeros32(zFrac0) - 11;
564 if (0 <= shiftCount) {
566 __shortShift64Left(zFrac0, zFrac1, shiftCount, zFrac0, zFrac1);
568 __shift64ExtraRightJamming(
569 zFrac0, zFrac1, 0u, -shiftCount, zFrac0, zFrac1, zFrac2);
572 return __roundAndPackFloat64(zSign, zExp, zFrac0, zFrac1, zFrac2);
575 /* Takes two double-precision floating-point values `a' and `b', one of which
576 * is a NaN, and returns the appropriate NaN result.
579 __propagateFloat64NaN(uint64_t __a, uint64_t __b)
581 bool aIsNaN = __is_nan(__a);
582 bool bIsNaN = __is_nan(__b);
583 uvec2 a = unpackUint2x32(__a);
584 uvec2 b = unpackUint2x32(__b);
588 return packUint2x32(mix(b, mix(a, b, bvec2(bIsNaN, bIsNaN)), bvec2(aIsNaN, aIsNaN)));
591 /* Returns the result of adding the double-precision floating-point values
592 * `a' and `b'. The operation is performed according to the IEEE Standard for
593 * Floating-Point Arithmetic.
596 __fadd64(uint64_t a, uint64_t b)
598 uint aSign = __extractFloat64Sign(a);
599 uint bSign = __extractFloat64Sign(b);
600 uint aFracLo = __extractFloat64FracLo(a);
601 uint aFracHi = __extractFloat64FracHi(a);
602 uint bFracLo = __extractFloat64FracLo(b);
603 uint bFracHi = __extractFloat64FracHi(b);
604 int aExp = __extractFloat64Exp(a);
605 int bExp = __extractFloat64Exp(b);
608 int expDiff = aExp - bExp;
609 if (aSign == bSign) {
612 bool orig_exp_diff_is_zero = (expDiff == 0);
614 if (orig_exp_diff_is_zero) {
616 bool propagate = (aFracHi | aFracLo | bFracHi | bFracLo) != 0u;
617 return mix(a, __propagateFloat64NaN(a, b), propagate);
619 __add64(aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1);
621 return __packFloat64(aSign, 0, zFrac0, zFrac1);
623 zFrac0 |= 0x00200000u;
625 __shift64ExtraRightJamming(
626 zFrac0, zFrac1, zFrac2, 1, zFrac0, zFrac1, zFrac2);
627 } else if (0 < expDiff) {
629 bool propagate = (aFracHi | aFracLo) != 0u;
630 return mix(a, __propagateFloat64NaN(a, b), propagate);
633 expDiff = mix(expDiff, expDiff - 1, bExp == 0);
634 bFracHi = mix(bFracHi | 0x00100000u, bFracHi, bExp == 0);
635 __shift64ExtraRightJamming(
636 bFracHi, bFracLo, 0u, expDiff, bFracHi, bFracLo, zFrac2);
638 } else if (expDiff < 0) {
640 bool propagate = (bFracHi | bFracLo) != 0u;
641 return mix(__packFloat64(aSign, 0x7ff, 0u, 0u), __propagateFloat64NaN(a, b), propagate);
643 expDiff = mix(expDiff, expDiff + 1, aExp == 0);
644 aFracHi = mix(aFracHi | 0x00100000u, aFracHi, aExp == 0);
645 __shift64ExtraRightJamming(
646 aFracHi, aFracLo, 0u, - expDiff, aFracHi, aFracLo, zFrac2);
649 if (!orig_exp_diff_is_zero) {
650 aFracHi |= 0x00100000u;
651 __add64(aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1);
653 if (!(zFrac0 < 0x00200000u)) {
654 __shift64ExtraRightJamming(zFrac0, zFrac1, zFrac2, 1, zFrac0, zFrac1, zFrac2);
658 return __roundAndPackFloat64(aSign, zExp, zFrac0, zFrac1, zFrac2);
663 __shortShift64Left(aFracHi, aFracLo, 10, aFracHi, aFracLo);
664 __shortShift64Left(bFracHi, bFracLo, 10, bFracHi, bFracLo);
667 bool propagate = (aFracHi | aFracLo) != 0u;
668 return mix(a, __propagateFloat64NaN(a, b), propagate);
670 expDiff = mix(expDiff, expDiff - 1, bExp == 0);
671 bFracHi = mix(bFracHi | 0x40000000u, bFracHi, bExp == 0);
672 __shift64RightJamming(bFracHi, bFracLo, expDiff, bFracHi, bFracLo);
673 aFracHi |= 0x40000000u;
674 __sub64(aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1);
677 return __normalizeRoundAndPackFloat64(aSign, zExp - 10, zFrac0, zFrac1);
681 bool propagate = (bFracHi | bFracLo) != 0u;
682 return mix(__packFloat64(aSign ^ 0x80000000u, 0x7ff, 0u, 0u), __propagateFloat64NaN(a, b), propagate);
684 expDiff = mix(expDiff, expDiff + 1, aExp == 0);
685 aFracHi = mix(aFracHi | 0x40000000u, aFracHi, aExp == 0);
686 __shift64RightJamming(aFracHi, aFracLo, - expDiff, aFracHi, aFracLo);
687 bFracHi |= 0x40000000u;
688 __sub64(bFracHi, bFracLo, aFracHi, aFracLo, zFrac0, zFrac1);
690 aSign ^= 0x80000000u;
692 return __normalizeRoundAndPackFloat64(aSign, zExp - 10, zFrac0, zFrac1);
695 bool propagate = (aFracHi | aFracLo | bFracHi | bFracLo) != 0u;
696 return mix(0xFFFFFFFFFFFFFFFFUL, __propagateFloat64NaN(a, b), propagate);
698 bExp = mix(bExp, 1, aExp == 0);
699 aExp = mix(aExp, 1, aExp == 0);
700 bool zexp_normal = false;
702 if (bFracHi < aFracHi) {
703 __sub64(aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1);
706 else if (aFracHi < bFracHi) {
707 __sub64(bFracHi, bFracLo, aFracHi, aFracLo, zFrac0, zFrac1);
711 else if (bFracLo < aFracLo) {
712 __sub64(aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1);
715 else if (aFracLo < bFracLo) {
716 __sub64(bFracHi, bFracLo, aFracHi, aFracLo, zFrac0, zFrac1);
720 zExp = mix(bExp, aExp, blta);
721 aSign = mix(aSign ^ 0x80000000u, aSign, blta);
722 uint64_t retval_0 = __packFloat64(uint(FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN) << 31, 0, 0u, 0u);
723 uint64_t retval_1 = __normalizeRoundAndPackFloat64(aSign, zExp - 11, zFrac0, zFrac1);
724 return mix(retval_0, retval_1, zexp_normal);
728 /* Multiplies the 64-bit value formed by concatenating `a0' and `a1' to the
729 * 64-bit value formed by concatenating `b0' and `b1' to obtain a 128-bit
730 * product. The product is broken into four 32-bit pieces which are stored at
731 * the locations pointed to by `z0Ptr', `z1Ptr', `z2Ptr', and `z3Ptr'.
734 __mul64To128(uint a0, uint a1, uint b0, uint b1,
747 umulExtended(a1, b1, z2, z3);
748 umulExtended(a1, b0, z1, more2);
749 __add64(z1, more2, 0u, z2, z1, z2);
750 umulExtended(a0, b0, z0, more1);
751 __add64(z0, more1, 0u, z1, z0, z1);
752 umulExtended(a0, b1, more1, more2);
753 __add64(more1, more2, 0u, z2, more1, z2);
754 __add64(z0, z1, 0u, more1, z0, z1);
761 /* Normalizes the subnormal double-precision floating-point value represented
762 * by the denormalized significand formed by the concatenation of `aFrac0' and
763 * `aFrac1'. The normalized exponent is stored at the location pointed to by
764 * `zExpPtr'. The most significant 21 bits of the normalized significand are
765 * stored at the location pointed to by `zFrac0Ptr', and the least significant
766 * 32 bits of the normalized significand are stored at the location pointed to
770 __normalizeFloat64Subnormal(uint aFrac0, uint aFrac1,
776 uint temp_zfrac0, temp_zfrac1;
777 shiftCount = __countLeadingZeros32(mix(aFrac0, aFrac1, aFrac0 == 0u)) - 11;
778 zExpPtr = mix(1 - shiftCount, -shiftCount - 31, aFrac0 == 0u);
780 temp_zfrac0 = mix(aFrac1<<shiftCount, aFrac1>>(-shiftCount), shiftCount < 0);
781 temp_zfrac1 = mix(0u, aFrac1<<(shiftCount & 31), shiftCount < 0);
783 __shortShift64Left(aFrac0, aFrac1, shiftCount, zFrac0Ptr, zFrac1Ptr);
785 zFrac0Ptr = mix(zFrac0Ptr, temp_zfrac0, aFrac0 == 0);
786 zFrac1Ptr = mix(zFrac1Ptr, temp_zfrac1, aFrac0 == 0);
789 /* Returns the result of multiplying the double-precision floating-point values
790 * `a' and `b'. The operation is performed according to the IEEE Standard for
791 * Floating-Point Arithmetic.
794 __fmul64(uint64_t a, uint64_t b)
802 uint aFracLo = __extractFloat64FracLo(a);
803 uint aFracHi = __extractFloat64FracHi(a);
804 uint bFracLo = __extractFloat64FracLo(b);
805 uint bFracHi = __extractFloat64FracHi(b);
806 int aExp = __extractFloat64Exp(a);
807 uint aSign = __extractFloat64Sign(a);
808 int bExp = __extractFloat64Exp(b);
809 uint bSign = __extractFloat64Sign(b);
810 uint zSign = aSign ^ bSign;
812 if (((aFracHi | aFracLo) != 0u) ||
813 ((bExp == 0x7FF) && ((bFracHi | bFracLo) != 0u))) {
814 return __propagateFloat64NaN(a, b);
816 if ((uint(bExp) | bFracHi | bFracLo) == 0u)
817 return 0xFFFFFFFFFFFFFFFFUL;
818 return __packFloat64(zSign, 0x7FF, 0u, 0u);
821 if ((bFracHi | bFracLo) != 0u)
822 return __propagateFloat64NaN(a, b);
823 if ((uint(aExp) | aFracHi | aFracLo) == 0u)
824 return 0xFFFFFFFFFFFFFFFFUL;
825 return __packFloat64(zSign, 0x7FF, 0u, 0u);
828 if ((aFracHi | aFracLo) == 0u)
829 return __packFloat64(zSign, 0, 0u, 0u);
830 __normalizeFloat64Subnormal(aFracHi, aFracLo, aExp, aFracHi, aFracLo);
833 if ((bFracHi | bFracLo) == 0u)
834 return __packFloat64(zSign, 0, 0u, 0u);
835 __normalizeFloat64Subnormal(bFracHi, bFracLo, bExp, bFracHi, bFracLo);
837 zExp = aExp + bExp - 0x400;
838 aFracHi |= 0x00100000u;
839 __shortShift64Left(bFracHi, bFracLo, 12, bFracHi, bFracLo);
841 aFracHi, aFracLo, bFracHi, bFracLo, zFrac0, zFrac1, zFrac2, zFrac3);
842 __add64(zFrac0, zFrac1, aFracHi, aFracLo, zFrac0, zFrac1);
843 zFrac2 |= uint(zFrac3 != 0u);
844 if (0x00200000u <= zFrac0) {
845 __shift64ExtraRightJamming(
846 zFrac0, zFrac1, zFrac2, 1, zFrac0, zFrac1, zFrac2);
849 return __roundAndPackFloat64(zSign, zExp, zFrac0, zFrac1, zFrac2);
853 __ffma64(uint64_t a, uint64_t b, uint64_t c)
855 return __fadd64(__fmul64(a, b), c);
858 /* Shifts the 64-bit value formed by concatenating `a0' and `a1' right by the
859 * number of bits given in `count'. Any bits shifted off are lost. The value
860 * of `count' can be arbitrarily large; in particular, if `count' is greater
861 * than 64, the result will be 0. The result is broken into two 32-bit pieces
862 * which are stored at the locations pointed to by `z0Ptr' and `z1Ptr'.
865 __shift64Right(uint a0, uint a1,
872 int negCount = (-count) & 31;
875 z0 = mix(z0, (a0 >> count), count < 32);
876 z0 = mix(z0, a0, count == 0);
878 z1 = mix(0u, (a0 >> (count & 31)), count < 64);
879 z1 = mix(z1, (a0<<negCount) | (a1>>count), count < 32);
880 z1 = mix(z1, a0, count == 0);
886 /* Returns the result of converting the double-precision floating-point value
887 * `a' to the unsigned integer format. The conversion is performed according
888 * to the IEEE Standard for Floating-Point Arithmetic.
891 __fp64_to_uint(uint64_t a)
893 uint aFracLo = __extractFloat64FracLo(a);
894 uint aFracHi = __extractFloat64FracHi(a);
895 int aExp = __extractFloat64Exp(a);
896 uint aSign = __extractFloat64Sign(a);
898 if ((aExp == 0x7FF) && ((aFracHi | aFracLo) != 0u))
901 aFracHi |= mix(0u, 0x00100000u, aExp != 0);
903 int shiftDist = 0x427 - aExp;
905 __shift64RightJamming(aFracHi, aFracLo, shiftDist, aFracHi, aFracLo);
907 if ((aFracHi & 0xFFFFF000u) != 0u)
908 return mix(~0u, 0u, aSign != 0u);
912 __shift64Right(aFracHi, aFracLo, 12, zero, z);
914 uint expt = mix(~0u, 0u, aSign != 0u);
916 return mix(z, expt, (aSign != 0u) && (z != 0u));
920 __uint_to_fp64(uint a)
925 int shiftDist = __countLeadingZeros32(a) + 21;
929 int negCount = (- shiftDist) & 31;
931 aHigh = mix(0u, a<< shiftDist - 32, shiftDist < 64);
933 aHigh = mix(aHigh, 0u, shiftDist == 0);
934 aLow = mix(aLow, a, shiftDist ==0);
935 aHigh = mix(aHigh, a >> negCount, shiftDist < 32);
936 aLow = mix(aLow, a << shiftDist, shiftDist < 32);
938 return __packFloat64(0u, 0x432 - shiftDist, aHigh, aLow);
942 __uint64_to_fp64(uint64_t a)
947 uvec2 aFrac = unpackUint2x32(a);
948 uint aFracLo = __extractFloat64FracLo(a);
949 uint aFracHi = __extractFloat64FracHi(a);
951 if ((aFracHi & 0x80000000u) != 0u) {
952 __shift64RightJamming(aFracHi, aFracLo, 1, aFracHi, aFracLo);
953 return __roundAndPackFloat64(0, 0x433, aFracHi, aFracLo, 0u);
955 return __normalizeRoundAndPackFloat64(0, 0x432, aFrac.y, aFrac.x);
960 __fp64_to_uint64(uint64_t a)
962 uint aFracLo = __extractFloat64FracLo(a);
963 uint aFracHi = __extractFloat64FracHi(a);
964 int aExp = __extractFloat64Exp(a);
965 uint aSign = __extractFloat64Sign(a);
967 uint64_t default_nan = 0xFFFFFFFFFFFFFFFFUL;
969 aFracHi = mix(aFracHi, aFracHi | 0x00100000u, aExp != 0);
970 int shiftCount = 0x433 - aExp;
972 if ( shiftCount <= 0 ) {
973 if (shiftCount < -11 && aExp == 0x7FF) {
974 if ((aFracHi | aFracLo) != 0u)
975 return __propagateFloat64NaN(a, a);
976 return mix(default_nan, a, aSign == 0u);
978 __shortShift64Left(aFracHi, aFracLo, -shiftCount, aFracHi, aFracLo);
980 __shift64ExtraRightJamming(aFracHi, aFracLo, zFrac2, shiftCount,
981 aFracHi, aFracLo, zFrac2);
983 return __roundAndPackUInt64(aSign, aFracHi, aFracLo, zFrac2);
987 __fp64_to_int64(uint64_t a)
990 uint aFracLo = __extractFloat64FracLo(a);
991 uint aFracHi = __extractFloat64FracHi(a);
992 int aExp = __extractFloat64Exp(a);
993 uint aSign = __extractFloat64Sign(a);
994 int64_t default_NegNaN = -0x7FFFFFFFFFFFFFFEL;
995 int64_t default_PosNaN = 0xFFFFFFFFFFFFFFFFL;
997 aFracHi = mix(aFracHi, aFracHi | 0x00100000u, aExp != 0);
998 int shiftCount = 0x433 - aExp;
1000 if (shiftCount <= 0) {
1001 if (shiftCount < -11 && aExp == 0x7FF) {
1002 if ((aFracHi | aFracLo) != 0u)
1003 return default_NegNaN;
1004 return mix(default_NegNaN, default_PosNaN, aSign == 0u);
1006 __shortShift64Left(aFracHi, aFracLo, -shiftCount, aFracHi, aFracLo);
1008 __shift64ExtraRightJamming(aFracHi, aFracLo, zFrac2, shiftCount,
1009 aFracHi, aFracLo, zFrac2);
1012 return __roundAndPackInt64(aSign, aFracHi, aFracLo, zFrac2);
1016 __fp32_to_uint64(float f)
1018 uint a = floatBitsToUint(f);
1019 uint aFrac = a & 0x007FFFFFu;
1020 int aExp = int((a>>23) & 0xFFu);
1021 uint aSign = a & 0x80000000u;
1025 uint64_t default_nan = 0xFFFFFFFFFFFFFFFFUL;
1026 int shiftCount = 0xBE - aExp;
1028 if (shiftCount <0) {
1033 aFrac = mix(aFrac, aFrac | 0x00800000u, aExp != 0);
1034 __shortShift64Left(aFrac, 0, 40, zFrac0, zFrac1);
1036 if (shiftCount != 0) {
1037 __shift64ExtraRightJamming(zFrac0, zFrac1, zFrac2, shiftCount,
1038 zFrac0, zFrac1, zFrac2);
1041 return __roundAndPackUInt64(aSign, zFrac0, zFrac1, zFrac2);
1045 __fp32_to_int64(float f)
1047 uint a = floatBitsToUint(f);
1048 uint aFrac = a & 0x007FFFFFu;
1049 int aExp = int((a>>23) & 0xFFu);
1050 uint aSign = a & 0x80000000u;
1054 int64_t default_NegNaN = -0x7FFFFFFFFFFFFFFEL;
1055 int64_t default_PosNaN = 0xFFFFFFFFFFFFFFFFL;
1056 int shiftCount = 0xBE - aExp;
1058 if (shiftCount <0) {
1059 if (aExp == 0xFF && aFrac != 0u)
1060 return default_NegNaN;
1061 return mix(default_NegNaN, default_PosNaN, aSign == 0u);
1064 aFrac = mix(aFrac, aFrac | 0x00800000u, aExp != 0);
1065 __shortShift64Left(aFrac, 0, 40, zFrac0, zFrac1);
1067 if (shiftCount != 0) {
1068 __shift64ExtraRightJamming(zFrac0, zFrac1, zFrac2, shiftCount,
1069 zFrac0, zFrac1, zFrac2);
1072 return __roundAndPackInt64(aSign, zFrac0, zFrac1, zFrac2);
1076 __int64_to_fp64(int64_t a)
1081 uint64_t absA = mix(uint64_t(a), uint64_t(-a), a < 0);
1082 uint aFracHi = __extractFloat64FracHi(absA);
1083 uvec2 aFrac = unpackUint2x32(absA);
1084 uint zSign = uint(unpackInt2x32(a).y) & 0x80000000u;
1086 if ((aFracHi & 0x80000000u) != 0u) {
1087 return mix(0ul, __packFloat64(0x80000000u, 0x434, 0u, 0u), a < 0);
1090 return __normalizeRoundAndPackFloat64(zSign, 0x432, aFrac.y, aFrac.x);
1093 /* Returns the result of converting the double-precision floating-point value
1094 * `a' to the 32-bit two's complement integer format. The conversion is
1095 * performed according to the IEEE Standard for Floating-Point Arithmetic---
1096 * which means in particular that the conversion is rounded according to the
1097 * current rounding mode. If `a' is a NaN, the largest positive integer is
1098 * returned. Otherwise, if the conversion overflows, the largest integer with
1099 * the same sign as `a' is returned.
1102 __fp64_to_int(uint64_t a)
1104 uint aFracLo = __extractFloat64FracLo(a);
1105 uint aFracHi = __extractFloat64FracHi(a);
1106 int aExp = __extractFloat64Exp(a);
1107 uint aSign = __extractFloat64Sign(a);
1110 uint aFracExtra = 0u;
1111 int shiftCount = aExp - 0x413;
1113 if (0 <= shiftCount) {
1115 if ((aExp == 0x7FF) && bool(aFracHi | aFracLo))
1117 return mix(0x7FFFFFFF, 0x80000000, aSign != 0u);
1119 __shortShift64Left(aFracHi | 0x00100000u, aFracLo, shiftCount, absZ, aFracExtra);
1124 aFracHi |= 0x00100000u;
1125 aFracExtra = ( aFracHi << (shiftCount & 31)) | aFracLo;
1126 absZ = aFracHi >> (- shiftCount);
1129 int z = mix(int(absZ), -int(absZ), aSign != 0u);
1130 int nan = mix(0x7FFFFFFF, 0x80000000, aSign != 0u);
1131 return mix(z, nan, ((aSign != 0u) != (z < 0)) && bool(z));
1134 /* Returns the result of converting the 32-bit two's complement integer `a'
1135 * to the double-precision floating-point format. The conversion is performed
1136 * according to the IEEE Standard for Floating-Point Arithmetic.
1139 __int_to_fp64(int a)
1144 return __packFloat64(0u, 0, 0u, 0u);
1145 uint zSign = uint(a) & 0x80000000u;
1146 uint absA = mix(uint(a), uint(-a), a < 0);
1147 int shiftCount = __countLeadingZeros32(absA) - 11;
1148 if (0 <= shiftCount) {
1149 zFrac0 = absA << shiftCount;
1152 __shift64Right(absA, 0u, -shiftCount, zFrac0, zFrac1);
1154 return __packFloat64(zSign, 0x412 - shiftCount, zFrac0, zFrac1);
1158 __fp64_to_bool(uint64_t a)
1160 return !__feq64_nonnan(__fabs64(a), 0ul);
1164 __bool_to_fp64(bool a)
1166 return packUint2x32(uvec2(0x00000000u, uint(-int(a) & 0x3ff00000)));
1169 /* Packs the sign `zSign', exponent `zExp', and significand `zFrac' into a
1170 * single-precision floating-point value, returning the result. After being
1171 * shifted into the proper positions, the three fields are simply added
1172 * together to form the result. This means that any integer portion of `zSig'
1173 * will be added into the exponent. Since a properly normalized significand
1174 * will have an integer portion equal to 1, the `zExp' input should be 1 less
1175 * than the desired result exponent whenever `zFrac' is a complete, normalized
1179 __packFloat32(uint zSign, int zExp, uint zFrac)
1181 return uintBitsToFloat(zSign + (uint(zExp)<<23) + zFrac);
1184 /* Takes an abstract floating-point value having sign `zSign', exponent `zExp',
1185 * and significand `zFrac', and returns the proper single-precision floating-
1186 * point value corresponding to the abstract input. Ordinarily, the abstract
1187 * value is simply rounded and packed into the single-precision format, with
1188 * the inexact exception raised if the abstract input cannot be represented
1189 * exactly. However, if the abstract value is too large, the overflow and
1190 * inexact exceptions are raised and an infinity or maximal finite value is
1191 * returned. If the abstract value is too small, the input value is rounded to
1192 * a subnormal number, and the underflow and inexact exceptions are raised if
1193 * the abstract input cannot be represented exactly as a subnormal single-
1194 * precision floating-point number.
1195 * The input significand `zFrac' has its binary point between bits 30
1196 * and 29, which is 7 bits to the left of the usual location. This shifted
1197 * significand must be normalized or smaller. If `zFrac' is not normalized,
1198 * `zExp' must be 0; in that case, the result returned is a subnormal number,
1199 * and it must not require rounding. In the usual case that `zFrac' is
1200 * normalized, `zExp' must be 1 less than the "true" floating-point exponent.
1201 * The handling of underflow and overflow follows the IEEE Standard for
1202 * Floating-Point Arithmetic.
1205 __roundAndPackFloat32(uint zSign, int zExp, uint zFrac)
1207 bool roundNearestEven;
1211 roundNearestEven = FLOAT_ROUNDING_MODE == FLOAT_ROUND_NEAREST_EVEN;
1212 roundIncrement = 0x40;
1213 if (!roundNearestEven) {
1214 if (FLOAT_ROUNDING_MODE == FLOAT_ROUND_TO_ZERO) {
1217 roundIncrement = 0x7F;
1219 if (FLOAT_ROUNDING_MODE == FLOAT_ROUND_UP)
1222 if (FLOAT_ROUNDING_MODE == FLOAT_ROUND_DOWN)
1227 roundBits = int(zFrac & 0x7Fu);
1228 if (0xFDu <= uint(zExp)) {
1229 if ((0xFD < zExp) || ((zExp == 0xFD) && (int(zFrac) + roundIncrement) < 0))
1230 return __packFloat32(zSign, 0xFF, 0u) - float(roundIncrement == 0);
1232 bool zexp_lt0 = zExp < 0;
1233 uint zFrac_lt0 = mix(uint(zFrac != 0u), (zFrac>>count) | uint((zFrac<<((-count) & 31)) != 0u), (-zExp) < 32);
1234 zFrac = mix(zFrac, zFrac_lt0, zexp_lt0);
1235 roundBits = mix(roundBits, int(zFrac) & 0x7f, zexp_lt0);
1236 zExp = mix(zExp, 0, zexp_lt0);
1238 zFrac = (zFrac + uint(roundIncrement))>>7;
1239 zFrac &= ~uint(((roundBits ^ 0x40) == 0) && roundNearestEven);
1241 return __packFloat32(zSign, mix(zExp, 0, zFrac == 0u), zFrac);
1244 /* Returns the result of converting the double-precision floating-point value
1245 * `a' to the single-precision floating-point format. The conversion is
1246 * performed according to the IEEE Standard for Floating-Point Arithmetic.
1249 __fp64_to_fp32(uint64_t __a)
1251 uvec2 a = unpackUint2x32(__a);
1255 uint aFracLo = __extractFloat64FracLo(__a);
1256 uint aFracHi = __extractFloat64FracHi(__a);
1257 int aExp = __extractFloat64Exp(__a);
1258 uint aSign = __extractFloat64Sign(__a);
1259 if (aExp == 0x7FF) {
1260 __shortShift64Left(a.y, a.x, 12, a.y, a.x);
1261 float rval = uintBitsToFloat(aSign | 0x7FC00000u | (a.y>>9));
1262 rval = mix(__packFloat32(aSign, 0xFF, 0u), rval, (aFracHi | aFracLo) != 0u);
1265 __shift64RightJamming(aFracHi, aFracLo, 22, allZero, zFrac);
1266 zFrac = mix(zFrac, zFrac | 0x40000000u, aExp != 0);
1267 return __roundAndPackFloat32(aSign, aExp - 0x381, zFrac);
1271 __uint64_to_fp32(uint64_t __a)
1273 uvec2 aFrac = unpackUint2x32(__a);
1274 int shiftCount = mix(__countLeadingZeros32(aFrac.y) - 33,
1275 __countLeadingZeros32(aFrac.x) - 1,
1278 if (0 <= shiftCount)
1279 __shortShift64Left(aFrac.y, aFrac.x, shiftCount, aFrac.y, aFrac.x);
1281 __shift64RightJamming(aFrac.y, aFrac.x, -shiftCount, aFrac.y, aFrac.x);
1283 return __roundAndPackFloat32(0u, 0x9C - shiftCount, aFrac.x);
1287 __int64_to_fp32(int64_t __a)
1289 uint aSign = uint(unpackInt2x32(__a).y) & 0x80000000u;
1290 uint64_t absA = mix(uint64_t(__a), uint64_t(-__a), __a < 0);
1291 uvec2 aFrac = unpackUint2x32(absA);
1292 int shiftCount = mix(__countLeadingZeros32(aFrac.y) - 33,
1293 __countLeadingZeros32(aFrac.x) - 1,
1296 if (0 <= shiftCount)
1297 __shortShift64Left(aFrac.y, aFrac.x, shiftCount, aFrac.y, aFrac.x);
1299 __shift64RightJamming(aFrac.y, aFrac.x, -shiftCount, aFrac.y, aFrac.x);
1301 return __roundAndPackFloat32(aSign, 0x9C - shiftCount, aFrac.x);
1304 /* Returns the result of converting the single-precision floating-point value
1305 * `a' to the double-precision floating-point format.
1308 __fp32_to_fp64(float f)
1310 uint a = floatBitsToUint(f);
1311 uint aFrac = a & 0x007FFFFFu;
1312 int aExp = int((a>>23) & 0xFFu);
1313 uint aSign = a & 0x80000000u;
1321 __shift64Right(nanHi, nanLo, 12, nanHi, nanLo);
1322 nanHi |= aSign | 0x7FF80000u;
1323 return packUint2x32(uvec2(nanLo, nanHi));
1325 return __packFloat64(aSign, 0x7FF, 0u, 0u);
1330 return __packFloat64(aSign, 0, 0u, 0u);
1331 /* Normalize subnormal */
1332 int shiftCount = __countLeadingZeros32(aFrac) - 8;
1333 aFrac <<= shiftCount;
1334 aExp = 1 - shiftCount;
1338 __shift64Right(aFrac, 0u, 3, zFrac0, zFrac1);
1339 return __packFloat64(aSign, aExp + 0x380, zFrac0, zFrac1);
1342 /* Adds the 96-bit value formed by concatenating `a0', `a1', and `a2' to the
1343 * 96-bit value formed by concatenating `b0', `b1', and `b2'. Addition is
1344 * modulo 2^96, so any carry out is lost. The result is broken into three
1345 * 32-bit pieces which are stored at the locations pointed to by `z0Ptr',
1346 * `z1Ptr', and `z2Ptr'.
1349 __add96(uint a0, uint a1, uint a2,
1350 uint b0, uint b1, uint b2,
1356 uint carry1 = uint(z2 < a2);
1358 uint carry0 = uint(z1 < a1);
1361 z0 += uint(z1 < carry1);
1368 /* Subtracts the 96-bit value formed by concatenating `b0', `b1', and `b2' from
1369 * the 96-bit value formed by concatenating `a0', `a1', and `a2'. Subtraction
1370 * is modulo 2^96, so any borrow out (carry out) is lost. The result is broken
1371 * into three 32-bit pieces which are stored at the locations pointed to by
1372 * `z0Ptr', `z1Ptr', and `z2Ptr'.
1375 __sub96(uint a0, uint a1, uint a2,
1376 uint b0, uint b1, uint b2,
1382 uint borrow1 = uint(a2 < b2);
1384 uint borrow0 = uint(a1 < b1);
1386 z0 -= uint(z1 < borrow1);
1394 /* Returns an approximation to the 32-bit integer quotient obtained by dividing
1395 * `b' into the 64-bit value formed by concatenating `a0' and `a1'. The
1396 * divisor `b' must be at least 2^31. If q is the exact quotient truncated
1397 * toward zero, the approximation returned lies between q and q + 2 inclusive.
1398 * If the exact quotient q is larger than 32 bits, the maximum positive 32-bit
1399 * unsigned integer is returned.
1402 __estimateDiv64To32(uint a0, uint a1, uint b)
1415 z = (b0<<16 <= a0) ? 0xFFFF0000u : (a0 / b0)<<16;
1416 umulExtended(b, z, term0, term1);
1417 __sub64(a0, a1, term0, term1, rem0, rem1);
1418 while (int(rem0) < 0) {
1421 __add64(rem0, rem1, b0, b1, rem0, rem1);
1423 rem0 = (rem0<<16) | (rem1>>16);
1424 z |= (b0<<16 <= rem0) ? 0xFFFFu : rem0 / b0;
1429 __sqrtOddAdjustments(int index)
1469 __sqrtEvenAdjustments(int index)
1508 /* Returns an approximation to the square root of the 32-bit significand given
1509 * by `a'. Considered as an integer, `a' must be at least 2^31. If bit 0 of
1510 * `aExp' (the least significant bit) is 1, the integer returned approximates
1511 * 2^31*sqrt(`a'/2^31), where `a' is considered an integer. If bit 0 of `aExp'
1512 * is 0, the integer returned approximates 2^31*sqrt(`a'/2^30). In either
1513 * case, the approximation returned lies strictly within +/-2 of the exact
1517 __estimateSqrt32(int aExp, uint a)
1521 int index = int(a>>27 & 15u);
1522 if ((aExp & 1) != 0) {
1523 z = 0x4000u + (a>>17) - __sqrtOddAdjustments(index);
1524 z = ((a / z)<<14) + (z<<15);
1527 z = 0x8000u + (a>>17) - __sqrtEvenAdjustments(index);
1529 z = (0x20000u <= z) ? 0xFFFF8000u : (z<<15);
1531 return uint(int(a)>>1);
1533 return ((__estimateDiv64To32(a, 0u, z))>>1) + (z>>1);
1536 /* Returns the square root of the double-precision floating-point value `a'.
1537 * The operation is performed according to the IEEE Standard for Floating-Point
1541 __fsqrt64(uint64_t a)
1546 uint doubleZFrac0 = 0u;
1555 uint64_t default_nan = 0xFFFFFFFFFFFFFFFFUL;
1557 uint aFracLo = __extractFloat64FracLo(a);
1558 uint aFracHi = __extractFloat64FracHi(a);
1559 int aExp = __extractFloat64Exp(a);
1560 uint aSign = __extractFloat64Sign(a);
1561 if (aExp == 0x7FF) {
1562 if ((aFracHi | aFracLo) != 0u)
1563 return __propagateFloat64NaN(a, a);
1569 if ((uint(aExp) | aFracHi | aFracLo) == 0u)
1574 if ((aFracHi | aFracLo) == 0u)
1575 return __packFloat64(0u, 0, 0u, 0u);
1576 __normalizeFloat64Subnormal(aFracHi, aFracLo, aExp, aFracHi, aFracLo);
1578 int zExp = ((aExp - 0x3FF)>>1) + 0x3FE;
1579 aFracHi |= 0x00100000u;
1580 __shortShift64Left(aFracHi, aFracLo, 11, term0, term1);
1581 zFrac0 = (__estimateSqrt32(aExp, term0)>>1) + 1u;
1583 zFrac0 = 0x7FFFFFFFu;
1584 doubleZFrac0 = zFrac0 + zFrac0;
1585 __shortShift64Left(aFracHi, aFracLo, 9 - (aExp & 1), aFracHi, aFracLo);
1586 umulExtended(zFrac0, zFrac0, term0, term1);
1587 __sub64(aFracHi, aFracLo, term0, term1, rem0, rem1);
1588 while (int(rem0) < 0) {
1591 __add64(rem0, rem1, 0u, doubleZFrac0 | 1u, rem0, rem1);
1593 zFrac1 = __estimateDiv64To32(rem1, 0u, doubleZFrac0);
1594 if ((zFrac1 & 0x1FFu) <= 5u) {
1597 umulExtended(doubleZFrac0, zFrac1, term1, term2);
1598 __sub64(rem1, 0u, term1, term2, rem1, rem2);
1599 umulExtended(zFrac1, zFrac1, term2, term3);
1600 __sub96(rem1, rem2, 0u, 0u, term2, term3, rem1, rem2, rem3);
1601 while (int(rem1) < 0) {
1603 __shortShift64Left(0u, zFrac1, 1, term2, term3);
1605 term2 |= doubleZFrac0;
1606 __add96(rem1, rem2, rem3, 0u, term2, term3, rem1, rem2, rem3);
1608 zFrac1 |= uint((rem1 | rem2 | rem3) != 0u);
1610 __shift64ExtraRightJamming(zFrac0, zFrac1, 0u, 10, zFrac0, zFrac1, zFrac2);
1611 return __roundAndPackFloat64(0u, zExp, zFrac0, zFrac1, zFrac2);
1615 __ftrunc64(uint64_t __a)
1617 uvec2 a = unpackUint2x32(__a);
1618 int aExp = __extractFloat64Exp(__a);
1622 int unbiasedExp = aExp - 1023;
1623 int fracBits = 52 - unbiasedExp;
1624 uint maskLo = mix(~0u << fracBits, 0u, fracBits >= 32);
1625 uint maskHi = mix(~0u << (fracBits - 32), ~0u, fracBits < 33);
1629 zLo = mix(zLo, 0u, unbiasedExp < 0);
1630 zHi = mix(zHi, 0u, unbiasedExp < 0);
1631 zLo = mix(zLo, a.x, unbiasedExp > 52);
1632 zHi = mix(zHi, a.y, unbiasedExp > 52);
1633 return packUint2x32(uvec2(zLo, zHi));
1637 __ffloor64(uint64_t a)
1639 bool is_positive = __fge64(a, 0ul);
1640 uint64_t tr = __ftrunc64(a);
1642 if (is_positive || __feq64(tr, a)) {
1645 return __fadd64(tr, 0xbff0000000000000ul /* -1.0 */);
1650 __fround64(uint64_t __a)
1652 uvec2 a = unpackUint2x32(__a);
1653 int unbiasedExp = __extractFloat64Exp(__a) - 1023;
1657 if (unbiasedExp < 20) {
1658 if (unbiasedExp < 0) {
1659 if ((aHi & 0x80000000u) != 0u && aLo == 0u) {
1663 if ((a.y & 0x000FFFFFu) == 0u && a.x == 0u) {
1665 return packUint2x32(uvec2(aLo, aHi));
1667 aHi = mix(aHi, (aHi | 0x3FF00000u), unbiasedExp == -1);
1670 uint maskExp = 0x000FFFFFu >> unbiasedExp;
1671 uint lastBit = maskExp + 1;
1672 aHi += 0x00080000u >> unbiasedExp;
1673 if ((aHi & maskExp) == 0u)
1678 } else if (unbiasedExp > 51 || unbiasedExp == 1024) {
1681 uint maskExp = 0xFFFFFFFFu >> (unbiasedExp - 20);
1682 if ((aLo & maskExp) == 0u)
1684 uint tmp = aLo + (1u << (51 - unbiasedExp));
1691 return packUint2x32(uvec2(aLo, aHi));
1695 __fmin64(uint64_t a, uint64_t b)
1697 if (__is_nan(a)) return b;
1698 if (__is_nan(b)) return a;
1700 if (__flt64_nonnan(a, b)) return a;
1705 __fmax64(uint64_t a, uint64_t b)
1707 if (__is_nan(a)) return b;
1708 if (__is_nan(b)) return a;
1710 if (__flt64_nonnan(a, b)) return b;
1715 __ffract64(uint64_t a)
1717 return __fadd64(a, __fneg64(__ffloor64(a)));