1 /*
2 ===============================================================================
3 
4 This C source file is part of the SoftFloat IEC/IEEE Floating-point
5 Arithmetic Package, Release 2.
6 
7 Written by John R. Hauser.  This work was made possible in part by the
8 International Computer Science Institute, located at Suite 600, 1947 Center
9 Street, Berkeley, California 94704.  Funding was partially provided by the
10 National Science Foundation under grant MIP-9311980.  The original version
11 of this code was written as part of a project to build a fixed-point vector
12 processor in collaboration with the University of California at Berkeley,
13 overseen by Profs. Nelson Morgan and John Wawrzynek.  More information
14 is available through the web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
15 arithmetic/softfloat.html'.
16 
17 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE.  Although reasonable effort
18 has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
19 TIMES RESULT IN INCORRECT BEHAVIOR.  USE OF THIS SOFTWARE IS RESTRICTED TO
20 PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
21 AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
22 
23 Derivative works are acceptable, even for commercial purposes, so long as
24 (1) they include prominent notice that the work is derivative, and (2) they
25 include prominent notice akin to these three paragraphs for those parts of
26 this code that are retained.
27 
28 ===============================================================================
29 */
30 
31 #include "fpa11.h"
32 //#include "milieu.h"
33 //#include "softfloat.h"
34 
35 /*
36 -------------------------------------------------------------------------------
37 Floating-point rounding mode, extended double-precision rounding precision,
38 and exception flags.
39 -------------------------------------------------------------------------------
40 */
41 int8 float_rounding_mode = float_round_nearest_even;
42 int8 floatx80_rounding_precision = 80;
43 int8 float_exception_flags;
44 
45 /*
46 -------------------------------------------------------------------------------
47 Primitive arithmetic functions, including multi-word arithmetic, and
48 division and square root approximations.  (Can be specialized to target if
49 desired.)
50 -------------------------------------------------------------------------------
51 */
52 #include "softfloat-macros"
53 
54 /*
55 -------------------------------------------------------------------------------
56 Functions and definitions to determine:  (1) whether tininess for underflow
57 is detected before or after rounding by default, (2) what (if anything)
58 happens when exceptions are raised, (3) how signaling NaNs are distinguished
59 from quiet NaNs, (4) the default generated quiet NaNs, and (5) how NaNs
60 are propagated from function inputs to output.  These details are target-
61 specific.
62 -------------------------------------------------------------------------------
63 */
64 #include "softfloat-specialize"
65 
66 /*
67 -------------------------------------------------------------------------------
68 Takes a 64-bit fixed-point value `absZ' with binary point between bits 6
69 and 7, and returns the properly rounded 32-bit integer corresponding to the
70 input.  If `zSign' is nonzero, the input is negated before being converted
71 to an integer.  Bit 63 of `absZ' must be zero.  Ordinarily, the fixed-point
72 input is simply rounded to an integer, with the inexact exception raised if
73 the input cannot be represented exactly as an integer.  If the fixed-point
74 input is too large, however, the invalid exception is raised and the largest
75 positive or negative integer is returned.
76 -------------------------------------------------------------------------------
77 */
roundAndPackInt32(flag zSign,bits64 absZ)78 static int32 roundAndPackInt32( flag zSign, bits64 absZ )
79 {
80     int8 roundingMode;
81     flag roundNearestEven;
82     int8 roundIncrement, roundBits;
83     int32 z;
84 
85     roundingMode = float_rounding_mode;
86     roundNearestEven = ( roundingMode == float_round_nearest_even );
87     roundIncrement = 0x40;
88     if ( ! roundNearestEven ) {
89         if ( roundingMode == float_round_to_zero ) {
90             roundIncrement = 0;
91         }
92         else {
93             roundIncrement = 0x7F;
94             if ( zSign ) {
95                 if ( roundingMode == float_round_up ) roundIncrement = 0;
96             }
97             else {
98                 if ( roundingMode == float_round_down ) roundIncrement = 0;
99             }
100         }
101     }
102     roundBits = absZ & 0x7F;
103     absZ = ( absZ + roundIncrement )>>7;
104     absZ &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven );
105     z = absZ;
106     if ( zSign ) z = - z;
107     if ( ( absZ>>32 ) || ( z && ( ( z < 0 ) ^ zSign ) ) ) {
108         float_exception_flags |= float_flag_invalid;
109         return zSign ? 0x80000000 : 0x7FFFFFFF;
110     }
111     if ( roundBits ) float_exception_flags |= float_flag_inexact;
112     return z;
113 
114 }
115 
116 /*
117 -------------------------------------------------------------------------------
118 Returns the fraction bits of the single-precision floating-point value `a'.
119 -------------------------------------------------------------------------------
120 */
extractFloat32Frac(float32 a)121 INLINE bits32 extractFloat32Frac( float32 a )
122 {
123 
124     return a & 0x007FFFFF;
125 
126 }
127 
128 /*
129 -------------------------------------------------------------------------------
130 Returns the exponent bits of the single-precision floating-point value `a'.
131 -------------------------------------------------------------------------------
132 */
extractFloat32Exp(float32 a)133 INLINE int16 extractFloat32Exp( float32 a )
134 {
135 
136     return ( a>>23 ) & 0xFF;
137 
138 }
139 
140 /*
141 -------------------------------------------------------------------------------
142 Returns the sign bit of the single-precision floating-point value `a'.
143 -------------------------------------------------------------------------------
144 */
145 #if 0	/* in softfloat.h */
146 INLINE flag extractFloat32Sign( float32 a )
147 {
148 
149     return a>>31;
150 
151 }
152 #endif
153 
154 /*
155 -------------------------------------------------------------------------------
156 Normalizes the subnormal single-precision floating-point value represented
157 by the denormalized significand `aSig'.  The normalized exponent and
158 significand are stored at the locations pointed to by `zExpPtr' and
159 `zSigPtr', respectively.
160 -------------------------------------------------------------------------------
161 */
162 static void
normalizeFloat32Subnormal(bits32 aSig,int16 * zExpPtr,bits32 * zSigPtr)163  normalizeFloat32Subnormal( bits32 aSig, int16 *zExpPtr, bits32 *zSigPtr )
164 {
165     int8 shiftCount;
166 
167     shiftCount = countLeadingZeros32( aSig ) - 8;
168     *zSigPtr = aSig<<shiftCount;
169     *zExpPtr = 1 - shiftCount;
170 
171 }
172 
173 /*
174 -------------------------------------------------------------------------------
175 Packs the sign `zSign', exponent `zExp', and significand `zSig' into a
176 single-precision floating-point value, returning the result.  After being
177 shifted into the proper positions, the three fields are simply added
178 together to form the result.  This means that any integer portion of `zSig'
179 will be added into the exponent.  Since a properly normalized significand
180 will have an integer portion equal to 1, the `zExp' input should be 1 less
181 than the desired result exponent whenever `zSig' is a complete, normalized
182 significand.
183 -------------------------------------------------------------------------------
184 */
packFloat32(flag zSign,int16 zExp,bits32 zSig)185 INLINE float32 packFloat32( flag zSign, int16 zExp, bits32 zSig )
186 {
187 #if 0
188    float32 f;
189    __asm__("@ packFloat32				\n\
190    	    mov %0, %1, asl #31				\n\
191    	    orr %0, %2, asl #23				\n\
192    	    orr %0, %3"
193    	    : /* no outputs */
194    	    : "g" (f), "g" (zSign), "g" (zExp), "g" (zSig)
195    	    : "cc");
196    return f;
197 #else
198     return ( ( (bits32) zSign )<<31 ) + ( ( (bits32) zExp )<<23 ) + zSig;
199 #endif
200 }
201 
202 /*
203 -------------------------------------------------------------------------------
204 Takes an abstract floating-point value having sign `zSign', exponent `zExp',
205 and significand `zSig', and returns the proper single-precision floating-
206 point value corresponding to the abstract input.  Ordinarily, the abstract
207 value is simply rounded and packed into the single-precision format, with
208 the inexact exception raised if the abstract input cannot be represented
209 exactly.  If the abstract value is too large, however, the overflow and
210 inexact exceptions are raised and an infinity or maximal finite value is
211 returned.  If the abstract value is too small, the input value is rounded to
212 a subnormal number, and the underflow and inexact exceptions are raised if
213 the abstract input cannot be represented exactly as a subnormal single-
214 precision floating-point number.
215     The input significand `zSig' has its binary point between bits 30
216 and 29, which is 7 bits to the left of the usual location.  This shifted
217 significand must be normalized or smaller.  If `zSig' is not normalized,
218 `zExp' must be 0; in that case, the result returned is a subnormal number,
219 and it must not require rounding.  In the usual case that `zSig' is
220 normalized, `zExp' must be 1 less than the ``true'' floating-point exponent.
221 The handling of underflow and overflow follows the IEC/IEEE Standard for
222 Binary Floating-point Arithmetic.
223 -------------------------------------------------------------------------------
224 */
roundAndPackFloat32(flag zSign,int16 zExp,bits32 zSig)225 static float32 roundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
226 {
227     int8 roundingMode;
228     flag roundNearestEven;
229     int8 roundIncrement, roundBits;
230     flag isTiny;
231 
232     roundingMode = float_rounding_mode;
233     roundNearestEven = ( roundingMode == float_round_nearest_even );
234     roundIncrement = 0x40;
235     if ( ! roundNearestEven ) {
236         if ( roundingMode == float_round_to_zero ) {
237             roundIncrement = 0;
238         }
239         else {
240             roundIncrement = 0x7F;
241             if ( zSign ) {
242                 if ( roundingMode == float_round_up ) roundIncrement = 0;
243             }
244             else {
245                 if ( roundingMode == float_round_down ) roundIncrement = 0;
246             }
247         }
248     }
249     roundBits = zSig & 0x7F;
250     if ( 0xFD <= (bits16) zExp ) {
251         if (    ( 0xFD < zExp )
252              || (    ( zExp == 0xFD )
253                   && ( (sbits32) ( zSig + roundIncrement ) < 0 ) )
254            ) {
255             float_raise( float_flag_overflow | float_flag_inexact );
256             return packFloat32( zSign, 0xFF, 0 ) - ( roundIncrement == 0 );
257         }
258         if ( zExp < 0 ) {
259             isTiny =
260                    ( float_detect_tininess == float_tininess_before_rounding )
261                 || ( zExp < -1 )
262                 || ( zSig + roundIncrement < 0x80000000 );
263             shift32RightJamming( zSig, - zExp, &zSig );
264             zExp = 0;
265             roundBits = zSig & 0x7F;
266             if ( isTiny && roundBits ) float_raise( float_flag_underflow );
267         }
268     }
269     if ( roundBits ) float_exception_flags |= float_flag_inexact;
270     zSig = ( zSig + roundIncrement )>>7;
271     zSig &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven );
272     if ( zSig == 0 ) zExp = 0;
273     return packFloat32( zSign, zExp, zSig );
274 
275 }
276 
277 /*
278 -------------------------------------------------------------------------------
279 Takes an abstract floating-point value having sign `zSign', exponent `zExp',
280 and significand `zSig', and returns the proper single-precision floating-
281 point value corresponding to the abstract input.  This routine is just like
282 `roundAndPackFloat32' except that `zSig' does not have to be normalized in
283 any way.  In all cases, `zExp' must be 1 less than the ``true'' floating-
284 point exponent.
285 -------------------------------------------------------------------------------
286 */
287 static float32
normalizeRoundAndPackFloat32(flag zSign,int16 zExp,bits32 zSig)288  normalizeRoundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
289 {
290     int8 shiftCount;
291 
292     shiftCount = countLeadingZeros32( zSig ) - 1;
293     return roundAndPackFloat32( zSign, zExp - shiftCount, zSig<<shiftCount );
294 
295 }
296 
297 /*
298 -------------------------------------------------------------------------------
299 Returns the fraction bits of the double-precision floating-point value `a'.
300 -------------------------------------------------------------------------------
301 */
extractFloat64Frac(float64 a)302 INLINE bits64 extractFloat64Frac( float64 a )
303 {
304 
305     return a & LIT64( 0x000FFFFFFFFFFFFF );
306 
307 }
308 
309 /*
310 -------------------------------------------------------------------------------
311 Returns the exponent bits of the double-precision floating-point value `a'.
312 -------------------------------------------------------------------------------
313 */
extractFloat64Exp(float64 a)314 INLINE int16 extractFloat64Exp( float64 a )
315 {
316 
317     return ( a>>52 ) & 0x7FF;
318 
319 }
320 
321 /*
322 -------------------------------------------------------------------------------
323 Returns the sign bit of the double-precision floating-point value `a'.
324 -------------------------------------------------------------------------------
325 */
326 #if 0	/* in softfloat.h */
327 INLINE flag extractFloat64Sign( float64 a )
328 {
329 
330     return a>>63;
331 
332 }
333 #endif
334 
335 /*
336 -------------------------------------------------------------------------------
337 Normalizes the subnormal double-precision floating-point value represented
338 by the denormalized significand `aSig'.  The normalized exponent and
339 significand are stored at the locations pointed to by `zExpPtr' and
340 `zSigPtr', respectively.
341 -------------------------------------------------------------------------------
342 */
343 static void
normalizeFloat64Subnormal(bits64 aSig,int16 * zExpPtr,bits64 * zSigPtr)344  normalizeFloat64Subnormal( bits64 aSig, int16 *zExpPtr, bits64 *zSigPtr )
345 {
346     int8 shiftCount;
347 
348     shiftCount = countLeadingZeros64( aSig ) - 11;
349     *zSigPtr = aSig<<shiftCount;
350     *zExpPtr = 1 - shiftCount;
351 
352 }
353 
354 /*
355 -------------------------------------------------------------------------------
356 Packs the sign `zSign', exponent `zExp', and significand `zSig' into a
357 double-precision floating-point value, returning the result.  After being
358 shifted into the proper positions, the three fields are simply added
359 together to form the result.  This means that any integer portion of `zSig'
360 will be added into the exponent.  Since a properly normalized significand
361 will have an integer portion equal to 1, the `zExp' input should be 1 less
362 than the desired result exponent whenever `zSig' is a complete, normalized
363 significand.
364 -------------------------------------------------------------------------------
365 */
packFloat64(flag zSign,int16 zExp,bits64 zSig)366 INLINE float64 packFloat64( flag zSign, int16 zExp, bits64 zSig )
367 {
368 
369     return ( ( (bits64) zSign )<<63 ) + ( ( (bits64) zExp )<<52 ) + zSig;
370 
371 }
372 
373 /*
374 -------------------------------------------------------------------------------
375 Takes an abstract floating-point value having sign `zSign', exponent `zExp',
376 and significand `zSig', and returns the proper double-precision floating-
377 point value corresponding to the abstract input.  Ordinarily, the abstract
378 value is simply rounded and packed into the double-precision format, with
379 the inexact exception raised if the abstract input cannot be represented
380 exactly.  If the abstract value is too large, however, the overflow and
381 inexact exceptions are raised and an infinity or maximal finite value is
382 returned.  If the abstract value is too small, the input value is rounded to
383 a subnormal number, and the underflow and inexact exceptions are raised if
384 the abstract input cannot be represented exactly as a subnormal double-
385 precision floating-point number.
386     The input significand `zSig' has its binary point between bits 62
387 and 61, which is 10 bits to the left of the usual location.  This shifted
388 significand must be normalized or smaller.  If `zSig' is not normalized,
389 `zExp' must be 0; in that case, the result returned is a subnormal number,
390 and it must not require rounding.  In the usual case that `zSig' is
391 normalized, `zExp' must be 1 less than the ``true'' floating-point exponent.
392 The handling of underflow and overflow follows the IEC/IEEE Standard for
393 Binary Floating-point Arithmetic.
394 -------------------------------------------------------------------------------
395 */
roundAndPackFloat64(flag zSign,int16 zExp,bits64 zSig)396 static float64 roundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
397 {
398     int8 roundingMode;
399     flag roundNearestEven;
400     int16 roundIncrement, roundBits;
401     flag isTiny;
402 
403     roundingMode = float_rounding_mode;
404     roundNearestEven = ( roundingMode == float_round_nearest_even );
405     roundIncrement = 0x200;
406     if ( ! roundNearestEven ) {
407         if ( roundingMode == float_round_to_zero ) {
408             roundIncrement = 0;
409         }
410         else {
411             roundIncrement = 0x3FF;
412             if ( zSign ) {
413                 if ( roundingMode == float_round_up ) roundIncrement = 0;
414             }
415             else {
416                 if ( roundingMode == float_round_down ) roundIncrement = 0;
417             }
418         }
419     }
420     roundBits = zSig & 0x3FF;
421     if ( 0x7FD <= (bits16) zExp ) {
422         if (    ( 0x7FD < zExp )
423              || (    ( zExp == 0x7FD )
424                   && ( (sbits64) ( zSig + roundIncrement ) < 0 ) )
425            ) {
426             //register int lr = __builtin_return_address(0);
427             //printk("roundAndPackFloat64 called from 0x%08x\n",lr);
428             float_raise( float_flag_overflow | float_flag_inexact );
429             return packFloat64( zSign, 0x7FF, 0 ) - ( roundIncrement == 0 );
430         }
431         if ( zExp < 0 ) {
432             isTiny =
433                    ( float_detect_tininess == float_tininess_before_rounding )
434                 || ( zExp < -1 )
435                 || ( zSig + roundIncrement < LIT64( 0x8000000000000000 ) );
436             shift64RightJamming( zSig, - zExp, &zSig );
437             zExp = 0;
438             roundBits = zSig & 0x3FF;
439             if ( isTiny && roundBits ) float_raise( float_flag_underflow );
440         }
441     }
442     if ( roundBits ) float_exception_flags |= float_flag_inexact;
443     zSig = ( zSig + roundIncrement )>>10;
444     zSig &= ~ ( ( ( roundBits ^ 0x200 ) == 0 ) & roundNearestEven );
445     if ( zSig == 0 ) zExp = 0;
446     return packFloat64( zSign, zExp, zSig );
447 
448 }
449 
450 /*
451 -------------------------------------------------------------------------------
452 Takes an abstract floating-point value having sign `zSign', exponent `zExp',
453 and significand `zSig', and returns the proper double-precision floating-
454 point value corresponding to the abstract input.  This routine is just like
455 `roundAndPackFloat64' except that `zSig' does not have to be normalized in
456 any way.  In all cases, `zExp' must be 1 less than the ``true'' floating-
457 point exponent.
458 -------------------------------------------------------------------------------
459 */
460 static float64
normalizeRoundAndPackFloat64(flag zSign,int16 zExp,bits64 zSig)461  normalizeRoundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
462 {
463     int8 shiftCount;
464 
465     shiftCount = countLeadingZeros64( zSig ) - 1;
466     return roundAndPackFloat64( zSign, zExp - shiftCount, zSig<<shiftCount );
467 
468 }
469 
470 #ifdef FLOATX80
471 
472 /*
473 -------------------------------------------------------------------------------
474 Returns the fraction bits of the extended double-precision floating-point
475 value `a'.
476 -------------------------------------------------------------------------------
477 */
extractFloatx80Frac(floatx80 a)478 INLINE bits64 extractFloatx80Frac( floatx80 a )
479 {
480 
481     return a.low;
482 
483 }
484 
485 /*
486 -------------------------------------------------------------------------------
487 Returns the exponent bits of the extended double-precision floating-point
488 value `a'.
489 -------------------------------------------------------------------------------
490 */
extractFloatx80Exp(floatx80 a)491 INLINE int32 extractFloatx80Exp( floatx80 a )
492 {
493 
494     return a.high & 0x7FFF;
495 
496 }
497 
498 /*
499 -------------------------------------------------------------------------------
500 Returns the sign bit of the extended double-precision floating-point value
501 `a'.
502 -------------------------------------------------------------------------------
503 */
extractFloatx80Sign(floatx80 a)504 INLINE flag extractFloatx80Sign( floatx80 a )
505 {
506 
507     return a.high>>15;
508 
509 }
510 
511 /*
512 -------------------------------------------------------------------------------
513 Normalizes the subnormal extended double-precision floating-point value
514 represented by the denormalized significand `aSig'.  The normalized exponent
515 and significand are stored at the locations pointed to by `zExpPtr' and
516 `zSigPtr', respectively.
517 -------------------------------------------------------------------------------
518 */
519 static void
normalizeFloatx80Subnormal(bits64 aSig,int32 * zExpPtr,bits64 * zSigPtr)520  normalizeFloatx80Subnormal( bits64 aSig, int32 *zExpPtr, bits64 *zSigPtr )
521 {
522     int8 shiftCount;
523 
524     shiftCount = countLeadingZeros64( aSig );
525     *zSigPtr = aSig<<shiftCount;
526     *zExpPtr = 1 - shiftCount;
527 
528 }
529 
530 /*
531 -------------------------------------------------------------------------------
532 Packs the sign `zSign', exponent `zExp', and significand `zSig' into an
533 extended double-precision floating-point value, returning the result.
534 -------------------------------------------------------------------------------
535 */
packFloatx80(flag zSign,int32 zExp,bits64 zSig)536 INLINE floatx80 packFloatx80( flag zSign, int32 zExp, bits64 zSig )
537 {
538     floatx80 z;
539 
540     z.low = zSig;
541     z.high = ( ( (bits16) zSign )<<15 ) + zExp;
542     return z;
543 
544 }
545 
546 /*
547 -------------------------------------------------------------------------------
548 Takes an abstract floating-point value having sign `zSign', exponent `zExp',
549 and extended significand formed by the concatenation of `zSig0' and `zSig1',
550 and returns the proper extended double-precision floating-point value
551 corresponding to the abstract input.  Ordinarily, the abstract value is
552 rounded and packed into the extended double-precision format, with the
553 inexact exception raised if the abstract input cannot be represented
554 exactly.  If the abstract value is too large, however, the overflow and
555 inexact exceptions are raised and an infinity or maximal finite value is
556 returned.  If the abstract value is too small, the input value is rounded to
557 a subnormal number, and the underflow and inexact exceptions are raised if
558 the abstract input cannot be represented exactly as a subnormal extended
559 double-precision floating-point number.
560     If `roundingPrecision' is 32 or 64, the result is rounded to the same
561 number of bits as single or double precision, respectively.  Otherwise, the
562 result is rounded to the full precision of the extended double-precision
563 format.
564     The input significand must be normalized or smaller.  If the input
565 significand is not normalized, `zExp' must be 0; in that case, the result
566 returned is a subnormal number, and it must not require rounding.  The
567 handling of underflow and overflow follows the IEC/IEEE Standard for Binary
568 Floating-point Arithmetic.
569 -------------------------------------------------------------------------------
570 */
571 static floatx80
roundAndPackFloatx80(int8 roundingPrecision,flag zSign,int32 zExp,bits64 zSig0,bits64 zSig1)572  roundAndPackFloatx80(
573      int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
574  )
575 {
576     int8 roundingMode;
577     flag roundNearestEven, increment, isTiny;
578     int64 roundIncrement, roundMask, roundBits;
579 
580     roundingMode = float_rounding_mode;
581     roundNearestEven = ( roundingMode == float_round_nearest_even );
582     if ( roundingPrecision == 80 ) goto precision80;
583     if ( roundingPrecision == 64 ) {
584         roundIncrement = LIT64( 0x0000000000000400 );
585         roundMask = LIT64( 0x00000000000007FF );
586     }
587     else if ( roundingPrecision == 32 ) {
588         roundIncrement = LIT64( 0x0000008000000000 );
589         roundMask = LIT64( 0x000000FFFFFFFFFF );
590     }
591     else {
592         goto precision80;
593     }
594     zSig0 |= ( zSig1 != 0 );
595     if ( ! roundNearestEven ) {
596         if ( roundingMode == float_round_to_zero ) {
597             roundIncrement = 0;
598         }
599         else {
600             roundIncrement = roundMask;
601             if ( zSign ) {
602                 if ( roundingMode == float_round_up ) roundIncrement = 0;
603             }
604             else {
605                 if ( roundingMode == float_round_down ) roundIncrement = 0;
606             }
607         }
608     }
609     roundBits = zSig0 & roundMask;
610     if ( 0x7FFD <= (bits32) ( zExp - 1 ) ) {
611         if (    ( 0x7FFE < zExp )
612              || ( ( zExp == 0x7FFE ) && ( zSig0 + roundIncrement < zSig0 ) )
613            ) {
614             goto overflow;
615         }
616         if ( zExp <= 0 ) {
617             isTiny =
618                    ( float_detect_tininess == float_tininess_before_rounding )
619                 || ( zExp < 0 )
620                 || ( zSig0 <= zSig0 + roundIncrement );
621             shift64RightJamming( zSig0, 1 - zExp, &zSig0 );
622             zExp = 0;
623             roundBits = zSig0 & roundMask;
624             if ( isTiny && roundBits ) float_raise( float_flag_underflow );
625             if ( roundBits ) float_exception_flags |= float_flag_inexact;
626             zSig0 += roundIncrement;
627             if ( (sbits64) zSig0 < 0 ) zExp = 1;
628             roundIncrement = roundMask + 1;
629             if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) {
630                 roundMask |= roundIncrement;
631             }
632             zSig0 &= ~ roundMask;
633             return packFloatx80( zSign, zExp, zSig0 );
634         }
635     }
636     if ( roundBits ) float_exception_flags |= float_flag_inexact;
637     zSig0 += roundIncrement;
638     if ( zSig0 < roundIncrement ) {
639         ++zExp;
640         zSig0 = LIT64( 0x8000000000000000 );
641     }
642     roundIncrement = roundMask + 1;
643     if ( roundNearestEven && ( roundBits<<1 == roundIncrement ) ) {
644         roundMask |= roundIncrement;
645     }
646     zSig0 &= ~ roundMask;
647     if ( zSig0 == 0 ) zExp = 0;
648     return packFloatx80( zSign, zExp, zSig0 );
649  precision80:
650     increment = ( (sbits64) zSig1 < 0 );
651     if ( ! roundNearestEven ) {
652         if ( roundingMode == float_round_to_zero ) {
653             increment = 0;
654         }
655         else {
656             if ( zSign ) {
657                 increment = ( roundingMode == float_round_down ) && zSig1;
658             }
659             else {
660                 increment = ( roundingMode == float_round_up ) && zSig1;
661             }
662         }
663     }
664     if ( 0x7FFD <= (bits32) ( zExp - 1 ) ) {
665         if (    ( 0x7FFE < zExp )
666              || (    ( zExp == 0x7FFE )
667                   && ( zSig0 == LIT64( 0xFFFFFFFFFFFFFFFF ) )
668                   && increment
669                 )
670            ) {
671             roundMask = 0;
672  overflow:
673             float_raise( float_flag_overflow | float_flag_inexact );
674             if (    ( roundingMode == float_round_to_zero )
675                  || ( zSign && ( roundingMode == float_round_up ) )
676                  || ( ! zSign && ( roundingMode == float_round_down ) )
677                ) {
678                 return packFloatx80( zSign, 0x7FFE, ~ roundMask );
679             }
680             return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
681         }
682         if ( zExp <= 0 ) {
683             isTiny =
684                    ( float_detect_tininess == float_tininess_before_rounding )
685                 || ( zExp < 0 )
686                 || ! increment
687                 || ( zSig0 < LIT64( 0xFFFFFFFFFFFFFFFF ) );
688             shift64ExtraRightJamming( zSig0, zSig1, 1 - zExp, &zSig0, &zSig1 );
689             zExp = 0;
690             if ( isTiny && zSig1 ) float_raise( float_flag_underflow );
691             if ( zSig1 ) float_exception_flags |= float_flag_inexact;
692             if ( roundNearestEven ) {
693                 increment = ( (sbits64) zSig1 < 0 );
694             }
695             else {
696                 if ( zSign ) {
697                     increment = ( roundingMode == float_round_down ) && zSig1;
698                 }
699                 else {
700                     increment = ( roundingMode == float_round_up ) && zSig1;
701                 }
702             }
703             if ( increment ) {
704                 ++zSig0;
705                 zSig0 &= ~ ( ( zSig1 + zSig1 == 0 ) & roundNearestEven );
706                 if ( (sbits64) zSig0 < 0 ) zExp = 1;
707             }
708             return packFloatx80( zSign, zExp, zSig0 );
709         }
710     }
711     if ( zSig1 ) float_exception_flags |= float_flag_inexact;
712     if ( increment ) {
713         ++zSig0;
714         if ( zSig0 == 0 ) {
715             ++zExp;
716             zSig0 = LIT64( 0x8000000000000000 );
717         }
718         else {
719             zSig0 &= ~ ( ( zSig1 + zSig1 == 0 ) & roundNearestEven );
720         }
721     }
722     else {
723         if ( zSig0 == 0 ) zExp = 0;
724     }
725 
726     return packFloatx80( zSign, zExp, zSig0 );
727 }
728 
729 /*
730 -------------------------------------------------------------------------------
731 Takes an abstract floating-point value having sign `zSign', exponent
732 `zExp', and significand formed by the concatenation of `zSig0' and `zSig1',
733 and returns the proper extended double-precision floating-point value
734 corresponding to the abstract input.  This routine is just like
735 `roundAndPackFloatx80' except that the input significand does not have to be
736 normalized.
737 -------------------------------------------------------------------------------
738 */
739 static floatx80
normalizeRoundAndPackFloatx80(int8 roundingPrecision,flag zSign,int32 zExp,bits64 zSig0,bits64 zSig1)740  normalizeRoundAndPackFloatx80(
741      int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
742  )
743 {
744     int8 shiftCount;
745 
746     if ( zSig0 == 0 ) {
747         zSig0 = zSig1;
748         zSig1 = 0;
749         zExp -= 64;
750     }
751     shiftCount = countLeadingZeros64( zSig0 );
752     shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 );
753     zExp -= shiftCount;
754     return
755         roundAndPackFloatx80( roundingPrecision, zSign, zExp, zSig0, zSig1 );
756 
757 }
758 
759 #endif
760 
761 /*
762 -------------------------------------------------------------------------------
763 Returns the result of converting the 32-bit two's complement integer `a' to
764 the single-precision floating-point format.  The conversion is performed
765 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
766 -------------------------------------------------------------------------------
767 */
int32_to_float32(int32 a)768 float32 int32_to_float32( int32 a )
769 {
770     flag zSign;
771 
772     if ( a == 0 ) return 0;
773     if ( a == 0x80000000 ) return packFloat32( 1, 0x9E, 0 );
774     zSign = ( a < 0 );
775     return normalizeRoundAndPackFloat32( zSign, 0x9C, zSign ? - a : a );
776 
777 }
778 
779 /*
780 -------------------------------------------------------------------------------
781 Returns the result of converting the 32-bit two's complement integer `a' to
782 the double-precision floating-point format.  The conversion is performed
783 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
784 -------------------------------------------------------------------------------
785 */
int32_to_float64(int32 a)786 float64 int32_to_float64( int32 a )
787 {
788     flag aSign;
789     uint32 absA;
790     int8 shiftCount;
791     bits64 zSig;
792 
793     if ( a == 0 ) return 0;
794     aSign = ( a < 0 );
795     absA = aSign ? - a : a;
796     shiftCount = countLeadingZeros32( absA ) + 21;
797     zSig = absA;
798     return packFloat64( aSign, 0x432 - shiftCount, zSig<<shiftCount );
799 
800 }
801 
802 #ifdef FLOATX80
803 
804 /*
805 -------------------------------------------------------------------------------
806 Returns the result of converting the 32-bit two's complement integer `a'
807 to the extended double-precision floating-point format.  The conversion
808 is performed according to the IEC/IEEE Standard for Binary Floating-point
809 Arithmetic.
810 -------------------------------------------------------------------------------
811 */
int32_to_floatx80(int32 a)812 floatx80 int32_to_floatx80( int32 a )
813 {
814     flag zSign;
815     uint32 absA;
816     int8 shiftCount;
817     bits64 zSig;
818 
819     if ( a == 0 ) return packFloatx80( 0, 0, 0 );
820     zSign = ( a < 0 );
821     absA = zSign ? - a : a;
822     shiftCount = countLeadingZeros32( absA ) + 32;
823     zSig = absA;
824     return packFloatx80( zSign, 0x403E - shiftCount, zSig<<shiftCount );
825 
826 }
827 
828 #endif
829 
830 /*
831 -------------------------------------------------------------------------------
832 Returns the result of converting the single-precision floating-point value
833 `a' to the 32-bit two's complement integer format.  The conversion is
834 performed according to the IEC/IEEE Standard for Binary Floating-point
835 Arithmetic---which means in particular that the conversion is rounded
836 according to the current rounding mode.  If `a' is a NaN, the largest
837 positive integer is returned.  Otherwise, if the conversion overflows, the
838 largest integer with the same sign as `a' is returned.
839 -------------------------------------------------------------------------------
840 */
float32_to_int32(float32 a)841 int32 float32_to_int32( float32 a )
842 {
843     flag aSign;
844     int16 aExp, shiftCount;
845     bits32 aSig;
846     bits64 zSig;
847 
848     aSig = extractFloat32Frac( a );
849     aExp = extractFloat32Exp( a );
850     aSign = extractFloat32Sign( a );
851     if ( ( aExp == 0x7FF ) && aSig ) aSign = 0;
852     if ( aExp ) aSig |= 0x00800000;
853     shiftCount = 0xAF - aExp;
854     zSig = aSig;
855     zSig <<= 32;
856     if ( 0 < shiftCount ) shift64RightJamming( zSig, shiftCount, &zSig );
857     return roundAndPackInt32( aSign, zSig );
858 
859 }
860 
861 /*
862 -------------------------------------------------------------------------------
863 Returns the result of converting the single-precision floating-point value
864 `a' to the 32-bit two's complement integer format.  The conversion is
865 performed according to the IEC/IEEE Standard for Binary Floating-point
866 Arithmetic, except that the conversion is always rounded toward zero.  If
867 `a' is a NaN, the largest positive integer is returned.  Otherwise, if the
868 conversion overflows, the largest integer with the same sign as `a' is
869 returned.
870 -------------------------------------------------------------------------------
871 */
float32_to_int32_round_to_zero(float32 a)872 int32 float32_to_int32_round_to_zero( float32 a )
873 {
874     flag aSign;
875     int16 aExp, shiftCount;
876     bits32 aSig;
877     int32 z;
878 
879     aSig = extractFloat32Frac( a );
880     aExp = extractFloat32Exp( a );
881     aSign = extractFloat32Sign( a );
882     shiftCount = aExp - 0x9E;
883     if ( 0 <= shiftCount ) {
884         if ( a == 0xCF000000 ) return 0x80000000;
885         float_raise( float_flag_invalid );
886         if ( ! aSign || ( ( aExp == 0xFF ) && aSig ) ) return 0x7FFFFFFF;
887         return 0x80000000;
888     }
889     else if ( aExp <= 0x7E ) {
890         if ( aExp | aSig ) float_exception_flags |= float_flag_inexact;
891         return 0;
892     }
893     aSig = ( aSig | 0x00800000 )<<8;
894     z = aSig>>( - shiftCount );
895     if ( (bits32) ( aSig<<( shiftCount & 31 ) ) ) {
896         float_exception_flags |= float_flag_inexact;
897     }
898     return aSign ? - z : z;
899 
900 }
901 
902 /*
903 -------------------------------------------------------------------------------
904 Returns the result of converting the single-precision floating-point value
905 `a' to the double-precision floating-point format.  The conversion is
906 performed according to the IEC/IEEE Standard for Binary Floating-point
907 Arithmetic.
908 -------------------------------------------------------------------------------
909 */
float32_to_float64(float32 a)910 float64 float32_to_float64( float32 a )
911 {
912     flag aSign;
913     int16 aExp;
914     bits32 aSig;
915 
916     aSig = extractFloat32Frac( a );
917     aExp = extractFloat32Exp( a );
918     aSign = extractFloat32Sign( a );
919     if ( aExp == 0xFF ) {
920         if ( aSig ) return commonNaNToFloat64( float32ToCommonNaN( a ) );
921         return packFloat64( aSign, 0x7FF, 0 );
922     }
923     if ( aExp == 0 ) {
924         if ( aSig == 0 ) return packFloat64( aSign, 0, 0 );
925         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
926         --aExp;
927     }
928     return packFloat64( aSign, aExp + 0x380, ( (bits64) aSig )<<29 );
929 
930 }
931 
932 #ifdef FLOATX80
933 
934 /*
935 -------------------------------------------------------------------------------
936 Returns the result of converting the single-precision floating-point value
937 `a' to the extended double-precision floating-point format.  The conversion
938 is performed according to the IEC/IEEE Standard for Binary Floating-point
939 Arithmetic.
940 -------------------------------------------------------------------------------
941 */
float32_to_floatx80(float32 a)942 floatx80 float32_to_floatx80( float32 a )
943 {
944     flag aSign;
945     int16 aExp;
946     bits32 aSig;
947 
948     aSig = extractFloat32Frac( a );
949     aExp = extractFloat32Exp( a );
950     aSign = extractFloat32Sign( a );
951     if ( aExp == 0xFF ) {
952         if ( aSig ) return commonNaNToFloatx80( float32ToCommonNaN( a ) );
953         return packFloatx80( aSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
954     }
955     if ( aExp == 0 ) {
956         if ( aSig == 0 ) return packFloatx80( aSign, 0, 0 );
957         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
958     }
959     aSig |= 0x00800000;
960     return packFloatx80( aSign, aExp + 0x3F80, ( (bits64) aSig )<<40 );
961 
962 }
963 
964 #endif
965 
966 /*
967 -------------------------------------------------------------------------------
968 Rounds the single-precision floating-point value `a' to an integer, and
969 returns the result as a single-precision floating-point value.  The
970 operation is performed according to the IEC/IEEE Standard for Binary
971 Floating-point Arithmetic.
972 -------------------------------------------------------------------------------
973 */
float32_round_to_int(float32 a)974 float32 float32_round_to_int( float32 a )
975 {
976     flag aSign;
977     int16 aExp;
978     bits32 lastBitMask, roundBitsMask;
979     int8 roundingMode;
980     float32 z;
981 
982     aExp = extractFloat32Exp( a );
983     if ( 0x96 <= aExp ) {
984         if ( ( aExp == 0xFF ) && extractFloat32Frac( a ) ) {
985             return propagateFloat32NaN( a, a );
986         }
987         return a;
988     }
989     if ( aExp <= 0x7E ) {
990         if ( (bits32) ( a<<1 ) == 0 ) return a;
991         float_exception_flags |= float_flag_inexact;
992         aSign = extractFloat32Sign( a );
993         switch ( float_rounding_mode ) {
994          case float_round_nearest_even:
995             if ( ( aExp == 0x7E ) && extractFloat32Frac( a ) ) {
996                 return packFloat32( aSign, 0x7F, 0 );
997             }
998             break;
999          case float_round_down:
1000             return aSign ? 0xBF800000 : 0;
1001          case float_round_up:
1002             return aSign ? 0x80000000 : 0x3F800000;
1003         }
1004         return packFloat32( aSign, 0, 0 );
1005     }
1006     lastBitMask = 1;
1007     lastBitMask <<= 0x96 - aExp;
1008     roundBitsMask = lastBitMask - 1;
1009     z = a;
1010     roundingMode = float_rounding_mode;
1011     if ( roundingMode == float_round_nearest_even ) {
1012         z += lastBitMask>>1;
1013         if ( ( z & roundBitsMask ) == 0 ) z &= ~ lastBitMask;
1014     }
1015     else if ( roundingMode != float_round_to_zero ) {
1016         if ( extractFloat32Sign( z ) ^ ( roundingMode == float_round_up ) ) {
1017             z += roundBitsMask;
1018         }
1019     }
1020     z &= ~ roundBitsMask;
1021     if ( z != a ) float_exception_flags |= float_flag_inexact;
1022     return z;
1023 
1024 }
1025 
1026 /*
1027 -------------------------------------------------------------------------------
1028 Returns the result of adding the absolute values of the single-precision
1029 floating-point values `a' and `b'.  If `zSign' is true, the sum is negated
1030 before being returned.  `zSign' is ignored if the result is a NaN.  The
1031 addition is performed according to the IEC/IEEE Standard for Binary
1032 Floating-point Arithmetic.
1033 -------------------------------------------------------------------------------
1034 */
addFloat32Sigs(float32 a,float32 b,flag zSign)1035 static float32 addFloat32Sigs( float32 a, float32 b, flag zSign )
1036 {
1037     int16 aExp, bExp, zExp;
1038     bits32 aSig, bSig, zSig;
1039     int16 expDiff;
1040 
1041     aSig = extractFloat32Frac( a );
1042     aExp = extractFloat32Exp( a );
1043     bSig = extractFloat32Frac( b );
1044     bExp = extractFloat32Exp( b );
1045     expDiff = aExp - bExp;
1046     aSig <<= 6;
1047     bSig <<= 6;
1048     if ( 0 < expDiff ) {
1049         if ( aExp == 0xFF ) {
1050             if ( aSig ) return propagateFloat32NaN( a, b );
1051             return a;
1052         }
1053         if ( bExp == 0 ) {
1054             --expDiff;
1055         }
1056         else {
1057             bSig |= 0x20000000;
1058         }
1059         shift32RightJamming( bSig, expDiff, &bSig );
1060         zExp = aExp;
1061     }
1062     else if ( expDiff < 0 ) {
1063         if ( bExp == 0xFF ) {
1064             if ( bSig ) return propagateFloat32NaN( a, b );
1065             return packFloat32( zSign, 0xFF, 0 );
1066         }
1067         if ( aExp == 0 ) {
1068             ++expDiff;
1069         }
1070         else {
1071             aSig |= 0x20000000;
1072         }
1073         shift32RightJamming( aSig, - expDiff, &aSig );
1074         zExp = bExp;
1075     }
1076     else {
1077         if ( aExp == 0xFF ) {
1078             if ( aSig | bSig ) return propagateFloat32NaN( a, b );
1079             return a;
1080         }
1081         if ( aExp == 0 ) return packFloat32( zSign, 0, ( aSig + bSig )>>6 );
1082         zSig = 0x40000000 + aSig + bSig;
1083         zExp = aExp;
1084         goto roundAndPack;
1085     }
1086     aSig |= 0x20000000;
1087     zSig = ( aSig + bSig )<<1;
1088     --zExp;
1089     if ( (sbits32) zSig < 0 ) {
1090         zSig = aSig + bSig;
1091         ++zExp;
1092     }
1093  roundAndPack:
1094     return roundAndPackFloat32( zSign, zExp, zSig );
1095 
1096 }
1097 
1098 /*
1099 -------------------------------------------------------------------------------
1100 Returns the result of subtracting the absolute values of the single-
1101 precision floating-point values `a' and `b'.  If `zSign' is true, the
1102 difference is negated before being returned.  `zSign' is ignored if the
1103 result is a NaN.  The subtraction is performed according to the IEC/IEEE
1104 Standard for Binary Floating-point Arithmetic.
1105 -------------------------------------------------------------------------------
1106 */
subFloat32Sigs(float32 a,float32 b,flag zSign)1107 static float32 subFloat32Sigs( float32 a, float32 b, flag zSign )
1108 {
1109     int16 aExp, bExp, zExp;
1110     bits32 aSig, bSig, zSig;
1111     int16 expDiff;
1112 
1113     aSig = extractFloat32Frac( a );
1114     aExp = extractFloat32Exp( a );
1115     bSig = extractFloat32Frac( b );
1116     bExp = extractFloat32Exp( b );
1117     expDiff = aExp - bExp;
1118     aSig <<= 7;
1119     bSig <<= 7;
1120     if ( 0 < expDiff ) goto aExpBigger;
1121     if ( expDiff < 0 ) goto bExpBigger;
1122     if ( aExp == 0xFF ) {
1123         if ( aSig | bSig ) return propagateFloat32NaN( a, b );
1124         float_raise( float_flag_invalid );
1125         return float32_default_nan;
1126     }
1127     if ( aExp == 0 ) {
1128         aExp = 1;
1129         bExp = 1;
1130     }
1131     if ( bSig < aSig ) goto aBigger;
1132     if ( aSig < bSig ) goto bBigger;
1133     return packFloat32( float_rounding_mode == float_round_down, 0, 0 );
1134  bExpBigger:
1135     if ( bExp == 0xFF ) {
1136         if ( bSig ) return propagateFloat32NaN( a, b );
1137         return packFloat32( zSign ^ 1, 0xFF, 0 );
1138     }
1139     if ( aExp == 0 ) {
1140         ++expDiff;
1141     }
1142     else {
1143         aSig |= 0x40000000;
1144     }
1145     shift32RightJamming( aSig, - expDiff, &aSig );
1146     bSig |= 0x40000000;
1147  bBigger:
1148     zSig = bSig - aSig;
1149     zExp = bExp;
1150     zSign ^= 1;
1151     goto normalizeRoundAndPack;
1152  aExpBigger:
1153     if ( aExp == 0xFF ) {
1154         if ( aSig ) return propagateFloat32NaN( a, b );
1155         return a;
1156     }
1157     if ( bExp == 0 ) {
1158         --expDiff;
1159     }
1160     else {
1161         bSig |= 0x40000000;
1162     }
1163     shift32RightJamming( bSig, expDiff, &bSig );
1164     aSig |= 0x40000000;
1165  aBigger:
1166     zSig = aSig - bSig;
1167     zExp = aExp;
1168  normalizeRoundAndPack:
1169     --zExp;
1170     return normalizeRoundAndPackFloat32( zSign, zExp, zSig );
1171 
1172 }
1173 
1174 /*
1175 -------------------------------------------------------------------------------
1176 Returns the result of adding the single-precision floating-point values `a'
1177 and `b'.  The operation is performed according to the IEC/IEEE Standard for
1178 Binary Floating-point Arithmetic.
1179 -------------------------------------------------------------------------------
1180 */
float32_add(float32 a,float32 b)1181 float32 float32_add( float32 a, float32 b )
1182 {
1183     flag aSign, bSign;
1184 
1185     aSign = extractFloat32Sign( a );
1186     bSign = extractFloat32Sign( b );
1187     if ( aSign == bSign ) {
1188         return addFloat32Sigs( a, b, aSign );
1189     }
1190     else {
1191         return subFloat32Sigs( a, b, aSign );
1192     }
1193 
1194 }
1195 
1196 /*
1197 -------------------------------------------------------------------------------
1198 Returns the result of subtracting the single-precision floating-point values
1199 `a' and `b'.  The operation is performed according to the IEC/IEEE Standard
1200 for Binary Floating-point Arithmetic.
1201 -------------------------------------------------------------------------------
1202 */
float32_sub(float32 a,float32 b)1203 float32 float32_sub( float32 a, float32 b )
1204 {
1205     flag aSign, bSign;
1206 
1207     aSign = extractFloat32Sign( a );
1208     bSign = extractFloat32Sign( b );
1209     if ( aSign == bSign ) {
1210         return subFloat32Sigs( a, b, aSign );
1211     }
1212     else {
1213         return addFloat32Sigs( a, b, aSign );
1214     }
1215 
1216 }
1217 
1218 /*
1219 -------------------------------------------------------------------------------
1220 Returns the result of multiplying the single-precision floating-point values
1221 `a' and `b'.  The operation is performed according to the IEC/IEEE Standard
1222 for Binary Floating-point Arithmetic.
1223 -------------------------------------------------------------------------------
1224 */
float32_mul(float32 a,float32 b)1225 float32 float32_mul( float32 a, float32 b )
1226 {
1227     flag aSign, bSign, zSign;
1228     int16 aExp, bExp, zExp;
1229     bits32 aSig, bSig;
1230     bits64 zSig64;
1231     bits32 zSig;
1232 
1233     aSig = extractFloat32Frac( a );
1234     aExp = extractFloat32Exp( a );
1235     aSign = extractFloat32Sign( a );
1236     bSig = extractFloat32Frac( b );
1237     bExp = extractFloat32Exp( b );
1238     bSign = extractFloat32Sign( b );
1239     zSign = aSign ^ bSign;
1240     if ( aExp == 0xFF ) {
1241         if ( aSig || ( ( bExp == 0xFF ) && bSig ) ) {
1242             return propagateFloat32NaN( a, b );
1243         }
1244         if ( ( bExp | bSig ) == 0 ) {
1245             float_raise( float_flag_invalid );
1246             return float32_default_nan;
1247         }
1248         return packFloat32( zSign, 0xFF, 0 );
1249     }
1250     if ( bExp == 0xFF ) {
1251         if ( bSig ) return propagateFloat32NaN( a, b );
1252         if ( ( aExp | aSig ) == 0 ) {
1253             float_raise( float_flag_invalid );
1254             return float32_default_nan;
1255         }
1256         return packFloat32( zSign, 0xFF, 0 );
1257     }
1258     if ( aExp == 0 ) {
1259         if ( aSig == 0 ) return packFloat32( zSign, 0, 0 );
1260         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
1261     }
1262     if ( bExp == 0 ) {
1263         if ( bSig == 0 ) return packFloat32( zSign, 0, 0 );
1264         normalizeFloat32Subnormal( bSig, &bExp, &bSig );
1265     }
1266     zExp = aExp + bExp - 0x7F;
1267     aSig = ( aSig | 0x00800000 )<<7;
1268     bSig = ( bSig | 0x00800000 )<<8;
1269     shift64RightJamming( ( (bits64) aSig ) * bSig, 32, &zSig64 );
1270     zSig = zSig64;
1271     if ( 0 <= (sbits32) ( zSig<<1 ) ) {
1272         zSig <<= 1;
1273         --zExp;
1274     }
1275     return roundAndPackFloat32( zSign, zExp, zSig );
1276 
1277 }
1278 
1279 /*
1280 -------------------------------------------------------------------------------
1281 Returns the result of dividing the single-precision floating-point value `a'
1282 by the corresponding value `b'.  The operation is performed according to the
1283 IEC/IEEE Standard for Binary Floating-point Arithmetic.
1284 -------------------------------------------------------------------------------
1285 */
float32_div(float32 a,float32 b)1286 float32 float32_div( float32 a, float32 b )
1287 {
1288     flag aSign, bSign, zSign;
1289     int16 aExp, bExp, zExp;
1290     bits32 aSig, bSig, zSig;
1291 
1292     aSig = extractFloat32Frac( a );
1293     aExp = extractFloat32Exp( a );
1294     aSign = extractFloat32Sign( a );
1295     bSig = extractFloat32Frac( b );
1296     bExp = extractFloat32Exp( b );
1297     bSign = extractFloat32Sign( b );
1298     zSign = aSign ^ bSign;
1299     if ( aExp == 0xFF ) {
1300         if ( aSig ) return propagateFloat32NaN( a, b );
1301         if ( bExp == 0xFF ) {
1302             if ( bSig ) return propagateFloat32NaN( a, b );
1303             float_raise( float_flag_invalid );
1304             return float32_default_nan;
1305         }
1306         return packFloat32( zSign, 0xFF, 0 );
1307     }
1308     if ( bExp == 0xFF ) {
1309         if ( bSig ) return propagateFloat32NaN( a, b );
1310         return packFloat32( zSign, 0, 0 );
1311     }
1312     if ( bExp == 0 ) {
1313         if ( bSig == 0 ) {
1314             if ( ( aExp | aSig ) == 0 ) {
1315                 float_raise( float_flag_invalid );
1316                 return float32_default_nan;
1317             }
1318             float_raise( float_flag_divbyzero );
1319             return packFloat32( zSign, 0xFF, 0 );
1320         }
1321         normalizeFloat32Subnormal( bSig, &bExp, &bSig );
1322     }
1323     if ( aExp == 0 ) {
1324         if ( aSig == 0 ) return packFloat32( zSign, 0, 0 );
1325         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
1326     }
1327     zExp = aExp - bExp + 0x7D;
1328     aSig = ( aSig | 0x00800000 )<<7;
1329     bSig = ( bSig | 0x00800000 )<<8;
1330     if ( bSig <= ( aSig + aSig ) ) {
1331         aSig >>= 1;
1332         ++zExp;
1333     }
1334     zSig = ( ( (bits64) aSig )<<32 ) / bSig;
1335     if ( ( zSig & 0x3F ) == 0 ) {
1336         zSig |= ( ( (bits64) bSig ) * zSig != ( (bits64) aSig )<<32 );
1337     }
1338     return roundAndPackFloat32( zSign, zExp, zSig );
1339 
1340 }
1341 
1342 /*
1343 -------------------------------------------------------------------------------
1344 Returns the remainder of the single-precision floating-point value `a'
1345 with respect to the corresponding value `b'.  The operation is performed
1346 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
1347 -------------------------------------------------------------------------------
1348 */
float32_rem(float32 a,float32 b)1349 float32 float32_rem( float32 a, float32 b )
1350 {
1351     flag aSign, bSign, zSign;
1352     int16 aExp, bExp, expDiff;
1353     bits32 aSig, bSig;
1354     bits32 q;
1355     bits64 aSig64, bSig64, q64;
1356     bits32 alternateASig;
1357     sbits32 sigMean;
1358 
1359     aSig = extractFloat32Frac( a );
1360     aExp = extractFloat32Exp( a );
1361     aSign = extractFloat32Sign( a );
1362     bSig = extractFloat32Frac( b );
1363     bExp = extractFloat32Exp( b );
1364     bSign = extractFloat32Sign( b );
1365     if ( aExp == 0xFF ) {
1366         if ( aSig || ( ( bExp == 0xFF ) && bSig ) ) {
1367             return propagateFloat32NaN( a, b );
1368         }
1369         float_raise( float_flag_invalid );
1370         return float32_default_nan;
1371     }
1372     if ( bExp == 0xFF ) {
1373         if ( bSig ) return propagateFloat32NaN( a, b );
1374         return a;
1375     }
1376     if ( bExp == 0 ) {
1377         if ( bSig == 0 ) {
1378             float_raise( float_flag_invalid );
1379             return float32_default_nan;
1380         }
1381         normalizeFloat32Subnormal( bSig, &bExp, &bSig );
1382     }
1383     if ( aExp == 0 ) {
1384         if ( aSig == 0 ) return a;
1385         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
1386     }
1387     expDiff = aExp - bExp;
1388     aSig |= 0x00800000;
1389     bSig |= 0x00800000;
1390     if ( expDiff < 32 ) {
1391         aSig <<= 8;
1392         bSig <<= 8;
1393         if ( expDiff < 0 ) {
1394             if ( expDiff < -1 ) return a;
1395             aSig >>= 1;
1396         }
1397         q = ( bSig <= aSig );
1398         if ( q ) aSig -= bSig;
1399         if ( 0 < expDiff ) {
1400             q = ( ( (bits64) aSig )<<32 ) / bSig;
1401             q >>= 32 - expDiff;
1402             bSig >>= 2;
1403             aSig = ( ( aSig>>1 )<<( expDiff - 1 ) ) - bSig * q;
1404         }
1405         else {
1406             aSig >>= 2;
1407             bSig >>= 2;
1408         }
1409     }
1410     else {
1411         if ( bSig <= aSig ) aSig -= bSig;
1412         aSig64 = ( (bits64) aSig )<<40;
1413         bSig64 = ( (bits64) bSig )<<40;
1414         expDiff -= 64;
1415         while ( 0 < expDiff ) {
1416             q64 = estimateDiv128To64( aSig64, 0, bSig64 );
1417             q64 = ( 2 < q64 ) ? q64 - 2 : 0;
1418             aSig64 = - ( ( bSig * q64 )<<38 );
1419             expDiff -= 62;
1420         }
1421         expDiff += 64;
1422         q64 = estimateDiv128To64( aSig64, 0, bSig64 );
1423         q64 = ( 2 < q64 ) ? q64 - 2 : 0;
1424         q = q64>>( 64 - expDiff );
1425         bSig <<= 6;
1426         aSig = ( ( aSig64>>33 )<<( expDiff - 1 ) ) - bSig * q;
1427     }
1428     do {
1429         alternateASig = aSig;
1430         ++q;
1431         aSig -= bSig;
1432     } while ( 0 <= (sbits32) aSig );
1433     sigMean = aSig + alternateASig;
1434     if ( ( sigMean < 0 ) || ( ( sigMean == 0 ) && ( q & 1 ) ) ) {
1435         aSig = alternateASig;
1436     }
1437     zSign = ( (sbits32) aSig < 0 );
1438     if ( zSign ) aSig = - aSig;
1439     return normalizeRoundAndPackFloat32( aSign ^ zSign, bExp, aSig );
1440 
1441 }
1442 
1443 /*
1444 -------------------------------------------------------------------------------
1445 Returns the square root of the single-precision floating-point value `a'.
1446 The operation is performed according to the IEC/IEEE Standard for Binary
1447 Floating-point Arithmetic.
1448 -------------------------------------------------------------------------------
1449 */
float32_sqrt(float32 a)1450 float32 float32_sqrt( float32 a )
1451 {
1452     flag aSign;
1453     int16 aExp, zExp;
1454     bits32 aSig, zSig;
1455     bits64 rem, term;
1456 
1457     aSig = extractFloat32Frac( a );
1458     aExp = extractFloat32Exp( a );
1459     aSign = extractFloat32Sign( a );
1460     if ( aExp == 0xFF ) {
1461         if ( aSig ) return propagateFloat32NaN( a, 0 );
1462         if ( ! aSign ) return a;
1463         float_raise( float_flag_invalid );
1464         return float32_default_nan;
1465     }
1466     if ( aSign ) {
1467         if ( ( aExp | aSig ) == 0 ) return a;
1468         float_raise( float_flag_invalid );
1469         return float32_default_nan;
1470     }
1471     if ( aExp == 0 ) {
1472         if ( aSig == 0 ) return 0;
1473         normalizeFloat32Subnormal( aSig, &aExp, &aSig );
1474     }
1475     zExp = ( ( aExp - 0x7F )>>1 ) + 0x7E;
1476     aSig = ( aSig | 0x00800000 )<<8;
1477     zSig = estimateSqrt32( aExp, aSig ) + 2;
1478     if ( ( zSig & 0x7F ) <= 5 ) {
1479         if ( zSig < 2 ) {
1480             zSig = 0xFFFFFFFF;
1481         }
1482         else {
1483             aSig >>= aExp & 1;
1484             term = ( (bits64) zSig ) * zSig;
1485             rem = ( ( (bits64) aSig )<<32 ) - term;
1486             while ( (sbits64) rem < 0 ) {
1487                 --zSig;
1488                 rem += ( ( (bits64) zSig )<<1 ) | 1;
1489             }
1490             zSig |= ( rem != 0 );
1491         }
1492     }
1493     shift32RightJamming( zSig, 1, &zSig );
1494     return roundAndPackFloat32( 0, zExp, zSig );
1495 
1496 }
1497 
1498 /*
1499 -------------------------------------------------------------------------------
1500 Returns 1 if the single-precision floating-point value `a' is equal to the
1501 corresponding value `b', and 0 otherwise.  The comparison is performed
1502 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
1503 -------------------------------------------------------------------------------
1504 */
float32_eq(float32 a,float32 b)1505 flag float32_eq( float32 a, float32 b )
1506 {
1507 
1508     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1509          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1510        ) {
1511         if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) {
1512             float_raise( float_flag_invalid );
1513         }
1514         return 0;
1515     }
1516     return ( a == b ) || ( (bits32) ( ( a | b )<<1 ) == 0 );
1517 
1518 }
1519 
1520 /*
1521 -------------------------------------------------------------------------------
1522 Returns 1 if the single-precision floating-point value `a' is less than or
1523 equal to the corresponding value `b', and 0 otherwise.  The comparison is
1524 performed according to the IEC/IEEE Standard for Binary Floating-point
1525 Arithmetic.
1526 -------------------------------------------------------------------------------
1527 */
float32_le(float32 a,float32 b)1528 flag float32_le( float32 a, float32 b )
1529 {
1530     flag aSign, bSign;
1531 
1532     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1533          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1534        ) {
1535         float_raise( float_flag_invalid );
1536         return 0;
1537     }
1538     aSign = extractFloat32Sign( a );
1539     bSign = extractFloat32Sign( b );
1540     if ( aSign != bSign ) return aSign || ( (bits32) ( ( a | b )<<1 ) == 0 );
1541     return ( a == b ) || ( aSign ^ ( a < b ) );
1542 
1543 }
1544 
1545 /*
1546 -------------------------------------------------------------------------------
1547 Returns 1 if the single-precision floating-point value `a' is less than
1548 the corresponding value `b', and 0 otherwise.  The comparison is performed
1549 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
1550 -------------------------------------------------------------------------------
1551 */
float32_lt(float32 a,float32 b)1552 flag float32_lt( float32 a, float32 b )
1553 {
1554     flag aSign, bSign;
1555 
1556     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1557          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1558        ) {
1559         float_raise( float_flag_invalid );
1560         return 0;
1561     }
1562     aSign = extractFloat32Sign( a );
1563     bSign = extractFloat32Sign( b );
1564     if ( aSign != bSign ) return aSign && ( (bits32) ( ( a | b )<<1 ) != 0 );
1565     return ( a != b ) && ( aSign ^ ( a < b ) );
1566 
1567 }
1568 
1569 /*
1570 -------------------------------------------------------------------------------
1571 Returns 1 if the single-precision floating-point value `a' is equal to the
1572 corresponding value `b', and 0 otherwise.  The invalid exception is raised
1573 if either operand is a NaN.  Otherwise, the comparison is performed
1574 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
1575 -------------------------------------------------------------------------------
1576 */
float32_eq_signaling(float32 a,float32 b)1577 flag float32_eq_signaling( float32 a, float32 b )
1578 {
1579 
1580     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1581          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1582        ) {
1583         float_raise( float_flag_invalid );
1584         return 0;
1585     }
1586     return ( a == b ) || ( (bits32) ( ( a | b )<<1 ) == 0 );
1587 
1588 }
1589 
1590 /*
1591 -------------------------------------------------------------------------------
1592 Returns 1 if the single-precision floating-point value `a' is less than or
1593 equal to the corresponding value `b', and 0 otherwise.  Quiet NaNs do not
1594 cause an exception.  Otherwise, the comparison is performed according to the
1595 IEC/IEEE Standard for Binary Floating-point Arithmetic.
1596 -------------------------------------------------------------------------------
1597 */
float32_le_quiet(float32 a,float32 b)1598 flag float32_le_quiet( float32 a, float32 b )
1599 {
1600     flag aSign, bSign;
1601     //int16 aExp, bExp;
1602 
1603     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1604          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1605        ) {
1606         if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) {
1607             float_raise( float_flag_invalid );
1608         }
1609         return 0;
1610     }
1611     aSign = extractFloat32Sign( a );
1612     bSign = extractFloat32Sign( b );
1613     if ( aSign != bSign ) return aSign || ( (bits32) ( ( a | b )<<1 ) == 0 );
1614     return ( a == b ) || ( aSign ^ ( a < b ) );
1615 
1616 }
1617 
1618 /*
1619 -------------------------------------------------------------------------------
1620 Returns 1 if the single-precision floating-point value `a' is less than
1621 the corresponding value `b', and 0 otherwise.  Quiet NaNs do not cause an
1622 exception.  Otherwise, the comparison is performed according to the IEC/IEEE
1623 Standard for Binary Floating-point Arithmetic.
1624 -------------------------------------------------------------------------------
1625 */
float32_lt_quiet(float32 a,float32 b)1626 flag float32_lt_quiet( float32 a, float32 b )
1627 {
1628     flag aSign, bSign;
1629 
1630     if (    ( ( extractFloat32Exp( a ) == 0xFF ) && extractFloat32Frac( a ) )
1631          || ( ( extractFloat32Exp( b ) == 0xFF ) && extractFloat32Frac( b ) )
1632        ) {
1633         if ( float32_is_signaling_nan( a ) || float32_is_signaling_nan( b ) ) {
1634             float_raise( float_flag_invalid );
1635         }
1636         return 0;
1637     }
1638     aSign = extractFloat32Sign( a );
1639     bSign = extractFloat32Sign( b );
1640     if ( aSign != bSign ) return aSign && ( (bits32) ( ( a | b )<<1 ) != 0 );
1641     return ( a != b ) && ( aSign ^ ( a < b ) );
1642 
1643 }
1644 
1645 /*
1646 -------------------------------------------------------------------------------
1647 Returns the result of converting the double-precision floating-point value
1648 `a' to the 32-bit two's complement integer format.  The conversion is
1649 performed according to the IEC/IEEE Standard for Binary Floating-point
1650 Arithmetic---which means in particular that the conversion is rounded
1651 according to the current rounding mode.  If `a' is a NaN, the largest
1652 positive integer is returned.  Otherwise, if the conversion overflows, the
1653 largest integer with the same sign as `a' is returned.
1654 -------------------------------------------------------------------------------
1655 */
float64_to_int32(float64 a)1656 int32 float64_to_int32( float64 a )
1657 {
1658     flag aSign;
1659     int16 aExp, shiftCount;
1660     bits64 aSig;
1661 
1662     aSig = extractFloat64Frac( a );
1663     aExp = extractFloat64Exp( a );
1664     aSign = extractFloat64Sign( a );
1665     if ( ( aExp == 0x7FF ) && aSig ) aSign = 0;
1666     if ( aExp ) aSig |= LIT64( 0x0010000000000000 );
1667     shiftCount = 0x42C - aExp;
1668     if ( 0 < shiftCount ) shift64RightJamming( aSig, shiftCount, &aSig );
1669     return roundAndPackInt32( aSign, aSig );
1670 
1671 }
1672 
1673 /*
1674 -------------------------------------------------------------------------------
1675 Returns the result of converting the double-precision floating-point value
1676 `a' to the 32-bit two's complement integer format.  The conversion is
1677 performed according to the IEC/IEEE Standard for Binary Floating-point
1678 Arithmetic, except that the conversion is always rounded toward zero.  If
1679 `a' is a NaN, the largest positive integer is returned.  Otherwise, if the
1680 conversion overflows, the largest integer with the same sign as `a' is
1681 returned.
1682 -------------------------------------------------------------------------------
1683 */
float64_to_int32_round_to_zero(float64 a)1684 int32 float64_to_int32_round_to_zero( float64 a )
1685 {
1686     flag aSign;
1687     int16 aExp, shiftCount;
1688     bits64 aSig, savedASig;
1689     int32 z;
1690 
1691     aSig = extractFloat64Frac( a );
1692     aExp = extractFloat64Exp( a );
1693     aSign = extractFloat64Sign( a );
1694     shiftCount = 0x433 - aExp;
1695     if ( shiftCount < 21 ) {
1696         if ( ( aExp == 0x7FF ) && aSig ) aSign = 0;
1697         goto invalid;
1698     }
1699     else if ( 52 < shiftCount ) {
1700         if ( aExp || aSig ) float_exception_flags |= float_flag_inexact;
1701         return 0;
1702     }
1703     aSig |= LIT64( 0x0010000000000000 );
1704     savedASig = aSig;
1705     aSig >>= shiftCount;
1706     z = aSig;
1707     if ( aSign ) z = - z;
1708     if ( ( z < 0 ) ^ aSign ) {
1709  invalid:
1710         float_exception_flags |= float_flag_invalid;
1711         return aSign ? 0x80000000 : 0x7FFFFFFF;
1712     }
1713     if ( ( aSig<<shiftCount ) != savedASig ) {
1714         float_exception_flags |= float_flag_inexact;
1715     }
1716     return z;
1717 
1718 }
1719 
1720 /*
1721 -------------------------------------------------------------------------------
1722 Returns the result of converting the double-precision floating-point value
1723 `a' to the 32-bit two's complement unsigned integer format.  The conversion
1724 is performed according to the IEC/IEEE Standard for Binary Floating-point
1725 Arithmetic---which means in particular that the conversion is rounded
1726 according to the current rounding mode.  If `a' is a NaN, the largest
1727 positive integer is returned.  Otherwise, if the conversion overflows, the
1728 largest positive integer is returned.
1729 -------------------------------------------------------------------------------
1730 */
float64_to_uint32(float64 a)1731 int32 float64_to_uint32( float64 a )
1732 {
1733     flag aSign;
1734     int16 aExp, shiftCount;
1735     bits64 aSig;
1736 
1737     aSig = extractFloat64Frac( a );
1738     aExp = extractFloat64Exp( a );
1739     aSign = 0; //extractFloat64Sign( a );
1740     //if ( ( aExp == 0x7FF ) && aSig ) aSign = 0;
1741     if ( aExp ) aSig |= LIT64( 0x0010000000000000 );
1742     shiftCount = 0x42C - aExp;
1743     if ( 0 < shiftCount ) shift64RightJamming( aSig, shiftCount, &aSig );
1744     return roundAndPackInt32( aSign, aSig );
1745 }
1746 
1747 /*
1748 -------------------------------------------------------------------------------
1749 Returns the result of converting the double-precision floating-point value
1750 `a' to the 32-bit two's complement integer format.  The conversion is
1751 performed according to the IEC/IEEE Standard for Binary Floating-point
1752 Arithmetic, except that the conversion is always rounded toward zero.  If
1753 `a' is a NaN, the largest positive integer is returned.  Otherwise, if the
1754 conversion overflows, the largest positive integer is returned.
1755 -------------------------------------------------------------------------------
1756 */
float64_to_uint32_round_to_zero(float64 a)1757 int32 float64_to_uint32_round_to_zero( float64 a )
1758 {
1759     flag aSign;
1760     int16 aExp, shiftCount;
1761     bits64 aSig, savedASig;
1762     int32 z;
1763 
1764     aSig = extractFloat64Frac( a );
1765     aExp = extractFloat64Exp( a );
1766     aSign = extractFloat64Sign( a );
1767     shiftCount = 0x433 - aExp;
1768     if ( shiftCount < 21 ) {
1769         if ( ( aExp == 0x7FF ) && aSig ) aSign = 0;
1770         goto invalid;
1771     }
1772     else if ( 52 < shiftCount ) {
1773         if ( aExp || aSig ) float_exception_flags |= float_flag_inexact;
1774         return 0;
1775     }
1776     aSig |= LIT64( 0x0010000000000000 );
1777     savedASig = aSig;
1778     aSig >>= shiftCount;
1779     z = aSig;
1780     if ( aSign ) z = - z;
1781     if ( ( z < 0 ) ^ aSign ) {
1782  invalid:
1783         float_exception_flags |= float_flag_invalid;
1784         return aSign ? 0x80000000 : 0x7FFFFFFF;
1785     }
1786     if ( ( aSig<<shiftCount ) != savedASig ) {
1787         float_exception_flags |= float_flag_inexact;
1788     }
1789     return z;
1790 }
1791 
1792 /*
1793 -------------------------------------------------------------------------------
1794 Returns the result of converting the double-precision floating-point value
1795 `a' to the single-precision floating-point format.  The conversion is
1796 performed according to the IEC/IEEE Standard for Binary Floating-point
1797 Arithmetic.
1798 -------------------------------------------------------------------------------
1799 */
float64_to_float32(float64 a)1800 float32 float64_to_float32( float64 a )
1801 {
1802     flag aSign;
1803     int16 aExp;
1804     bits64 aSig;
1805     bits32 zSig;
1806 
1807     aSig = extractFloat64Frac( a );
1808     aExp = extractFloat64Exp( a );
1809     aSign = extractFloat64Sign( a );
1810     if ( aExp == 0x7FF ) {
1811         if ( aSig ) return commonNaNToFloat32( float64ToCommonNaN( a ) );
1812         return packFloat32( aSign, 0xFF, 0 );
1813     }
1814     shift64RightJamming( aSig, 22, &aSig );
1815     zSig = aSig;
1816     if ( aExp || zSig ) {
1817         zSig |= 0x40000000;
1818         aExp -= 0x381;
1819     }
1820     return roundAndPackFloat32( aSign, aExp, zSig );
1821 
1822 }
1823 
1824 #ifdef FLOATX80
1825 
1826 /*
1827 -------------------------------------------------------------------------------
1828 Returns the result of converting the double-precision floating-point value
1829 `a' to the extended double-precision floating-point format.  The conversion
1830 is performed according to the IEC/IEEE Standard for Binary Floating-point
1831 Arithmetic.
1832 -------------------------------------------------------------------------------
1833 */
float64_to_floatx80(float64 a)1834 floatx80 float64_to_floatx80( float64 a )
1835 {
1836     flag aSign;
1837     int16 aExp;
1838     bits64 aSig;
1839 
1840     aSig = extractFloat64Frac( a );
1841     aExp = extractFloat64Exp( a );
1842     aSign = extractFloat64Sign( a );
1843     if ( aExp == 0x7FF ) {
1844         if ( aSig ) return commonNaNToFloatx80( float64ToCommonNaN( a ) );
1845         return packFloatx80( aSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
1846     }
1847     if ( aExp == 0 ) {
1848         if ( aSig == 0 ) return packFloatx80( aSign, 0, 0 );
1849         normalizeFloat64Subnormal( aSig, &aExp, &aSig );
1850     }
1851     return
1852         packFloatx80(
1853             aSign, aExp + 0x3C00, ( aSig | LIT64( 0x0010000000000000 ) )<<11 );
1854 
1855 }
1856 
1857 #endif
1858 
1859 /*
1860 -------------------------------------------------------------------------------
1861 Rounds the double-precision floating-point value `a' to an integer, and
1862 returns the result as a double-precision floating-point value.  The
1863 operation is performed according to the IEC/IEEE Standard for Binary
1864 Floating-point Arithmetic.
1865 -------------------------------------------------------------------------------
1866 */
float64_round_to_int(float64 a)1867 float64 float64_round_to_int( float64 a )
1868 {
1869     flag aSign;
1870     int16 aExp;
1871     bits64 lastBitMask, roundBitsMask;
1872     int8 roundingMode;
1873     float64 z;
1874 
1875     aExp = extractFloat64Exp( a );
1876     if ( 0x433 <= aExp ) {
1877         if ( ( aExp == 0x7FF ) && extractFloat64Frac( a ) ) {
1878             return propagateFloat64NaN( a, a );
1879         }
1880         return a;
1881     }
1882     if ( aExp <= 0x3FE ) {
1883         if ( (bits64) ( a<<1 ) == 0 ) return a;
1884         float_exception_flags |= float_flag_inexact;
1885         aSign = extractFloat64Sign( a );
1886         switch ( float_rounding_mode ) {
1887          case float_round_nearest_even:
1888             if ( ( aExp == 0x3FE ) && extractFloat64Frac( a ) ) {
1889                 return packFloat64( aSign, 0x3FF, 0 );
1890             }
1891             break;
1892          case float_round_down:
1893             return aSign ? LIT64( 0xBFF0000000000000 ) : 0;
1894          case float_round_up:
1895             return
1896             aSign ? LIT64( 0x8000000000000000 ) : LIT64( 0x3FF0000000000000 );
1897         }
1898         return packFloat64( aSign, 0, 0 );
1899     }
1900     lastBitMask = 1;
1901     lastBitMask <<= 0x433 - aExp;
1902     roundBitsMask = lastBitMask - 1;
1903     z = a;
1904     roundingMode = float_rounding_mode;
1905     if ( roundingMode == float_round_nearest_even ) {
1906         z += lastBitMask>>1;
1907         if ( ( z & roundBitsMask ) == 0 ) z &= ~ lastBitMask;
1908     }
1909     else if ( roundingMode != float_round_to_zero ) {
1910         if ( extractFloat64Sign( z ) ^ ( roundingMode == float_round_up ) ) {
1911             z += roundBitsMask;
1912         }
1913     }
1914     z &= ~ roundBitsMask;
1915     if ( z != a ) float_exception_flags |= float_flag_inexact;
1916     return z;
1917 
1918 }
1919 
1920 /*
1921 -------------------------------------------------------------------------------
1922 Returns the result of adding the absolute values of the double-precision
1923 floating-point values `a' and `b'.  If `zSign' is true, the sum is negated
1924 before being returned.  `zSign' is ignored if the result is a NaN.  The
1925 addition is performed according to the IEC/IEEE Standard for Binary
1926 Floating-point Arithmetic.
1927 -------------------------------------------------------------------------------
1928 */
addFloat64Sigs(float64 a,float64 b,flag zSign)1929 static float64 addFloat64Sigs( float64 a, float64 b, flag zSign )
1930 {
1931     int16 aExp, bExp, zExp;
1932     bits64 aSig, bSig, zSig;
1933     int16 expDiff;
1934 
1935     aSig = extractFloat64Frac( a );
1936     aExp = extractFloat64Exp( a );
1937     bSig = extractFloat64Frac( b );
1938     bExp = extractFloat64Exp( b );
1939     expDiff = aExp - bExp;
1940     aSig <<= 9;
1941     bSig <<= 9;
1942     if ( 0 < expDiff ) {
1943         if ( aExp == 0x7FF ) {
1944             if ( aSig ) return propagateFloat64NaN( a, b );
1945             return a;
1946         }
1947         if ( bExp == 0 ) {
1948             --expDiff;
1949         }
1950         else {
1951             bSig |= LIT64( 0x2000000000000000 );
1952         }
1953         shift64RightJamming( bSig, expDiff, &bSig );
1954         zExp = aExp;
1955     }
1956     else if ( expDiff < 0 ) {
1957         if ( bExp == 0x7FF ) {
1958             if ( bSig ) return propagateFloat64NaN( a, b );
1959             return packFloat64( zSign, 0x7FF, 0 );
1960         }
1961         if ( aExp == 0 ) {
1962             ++expDiff;
1963         }
1964         else {
1965             aSig |= LIT64( 0x2000000000000000 );
1966         }
1967         shift64RightJamming( aSig, - expDiff, &aSig );
1968         zExp = bExp;
1969     }
1970     else {
1971         if ( aExp == 0x7FF ) {
1972             if ( aSig | bSig ) return propagateFloat64NaN( a, b );
1973             return a;
1974         }
1975         if ( aExp == 0 ) return packFloat64( zSign, 0, ( aSig + bSig )>>9 );
1976         zSig = LIT64( 0x4000000000000000 ) + aSig + bSig;
1977         zExp = aExp;
1978         goto roundAndPack;
1979     }
1980     aSig |= LIT64( 0x2000000000000000 );
1981     zSig = ( aSig + bSig )<<1;
1982     --zExp;
1983     if ( (sbits64) zSig < 0 ) {
1984         zSig = aSig + bSig;
1985         ++zExp;
1986     }
1987  roundAndPack:
1988     return roundAndPackFloat64( zSign, zExp, zSig );
1989 
1990 }
1991 
1992 /*
1993 -------------------------------------------------------------------------------
1994 Returns the result of subtracting the absolute values of the double-
1995 precision floating-point values `a' and `b'.  If `zSign' is true, the
1996 difference is negated before being returned.  `zSign' is ignored if the
1997 result is a NaN.  The subtraction is performed according to the IEC/IEEE
1998 Standard for Binary Floating-point Arithmetic.
1999 -------------------------------------------------------------------------------
2000 */
subFloat64Sigs(float64 a,float64 b,flag zSign)2001 static float64 subFloat64Sigs( float64 a, float64 b, flag zSign )
2002 {
2003     int16 aExp, bExp, zExp;
2004     bits64 aSig, bSig, zSig;
2005     int16 expDiff;
2006 
2007     aSig = extractFloat64Frac( a );
2008     aExp = extractFloat64Exp( a );
2009     bSig = extractFloat64Frac( b );
2010     bExp = extractFloat64Exp( b );
2011     expDiff = aExp - bExp;
2012     aSig <<= 10;
2013     bSig <<= 10;
2014     if ( 0 < expDiff ) goto aExpBigger;
2015     if ( expDiff < 0 ) goto bExpBigger;
2016     if ( aExp == 0x7FF ) {
2017         if ( aSig | bSig ) return propagateFloat64NaN( a, b );
2018         float_raise( float_flag_invalid );
2019         return float64_default_nan;
2020     }
2021     if ( aExp == 0 ) {
2022         aExp = 1;
2023         bExp = 1;
2024     }
2025     if ( bSig < aSig ) goto aBigger;
2026     if ( aSig < bSig ) goto bBigger;
2027     return packFloat64( float_rounding_mode == float_round_down, 0, 0 );
2028  bExpBigger:
2029     if ( bExp == 0x7FF ) {
2030         if ( bSig ) return propagateFloat64NaN( a, b );
2031         return packFloat64( zSign ^ 1, 0x7FF, 0 );
2032     }
2033     if ( aExp == 0 ) {
2034         ++expDiff;
2035     }
2036     else {
2037         aSig |= LIT64( 0x4000000000000000 );
2038     }
2039     shift64RightJamming( aSig, - expDiff, &aSig );
2040     bSig |= LIT64( 0x4000000000000000 );
2041  bBigger:
2042     zSig = bSig - aSig;
2043     zExp = bExp;
2044     zSign ^= 1;
2045     goto normalizeRoundAndPack;
2046  aExpBigger:
2047     if ( aExp == 0x7FF ) {
2048         if ( aSig ) return propagateFloat64NaN( a, b );
2049         return a;
2050     }
2051     if ( bExp == 0 ) {
2052         --expDiff;
2053     }
2054     else {
2055         bSig |= LIT64( 0x4000000000000000 );
2056     }
2057     shift64RightJamming( bSig, expDiff, &bSig );
2058     aSig |= LIT64( 0x4000000000000000 );
2059  aBigger:
2060     zSig = aSig - bSig;
2061     zExp = aExp;
2062  normalizeRoundAndPack:
2063     --zExp;
2064     return normalizeRoundAndPackFloat64( zSign, zExp, zSig );
2065 
2066 }
2067 
2068 /*
2069 -------------------------------------------------------------------------------
2070 Returns the result of adding the double-precision floating-point values `a'
2071 and `b'.  The operation is performed according to the IEC/IEEE Standard for
2072 Binary Floating-point Arithmetic.
2073 -------------------------------------------------------------------------------
2074 */
float64_add(float64 a,float64 b)2075 float64 float64_add( float64 a, float64 b )
2076 {
2077     flag aSign, bSign;
2078 
2079     aSign = extractFloat64Sign( a );
2080     bSign = extractFloat64Sign( b );
2081     if ( aSign == bSign ) {
2082         return addFloat64Sigs( a, b, aSign );
2083     }
2084     else {
2085         return subFloat64Sigs( a, b, aSign );
2086     }
2087 
2088 }
2089 
2090 /*
2091 -------------------------------------------------------------------------------
2092 Returns the result of subtracting the double-precision floating-point values
2093 `a' and `b'.  The operation is performed according to the IEC/IEEE Standard
2094 for Binary Floating-point Arithmetic.
2095 -------------------------------------------------------------------------------
2096 */
float64_sub(float64 a,float64 b)2097 float64 float64_sub( float64 a, float64 b )
2098 {
2099     flag aSign, bSign;
2100 
2101     aSign = extractFloat64Sign( a );
2102     bSign = extractFloat64Sign( b );
2103     if ( aSign == bSign ) {
2104         return subFloat64Sigs( a, b, aSign );
2105     }
2106     else {
2107         return addFloat64Sigs( a, b, aSign );
2108     }
2109 
2110 }
2111 
2112 /*
2113 -------------------------------------------------------------------------------
2114 Returns the result of multiplying the double-precision floating-point values
2115 `a' and `b'.  The operation is performed according to the IEC/IEEE Standard
2116 for Binary Floating-point Arithmetic.
2117 -------------------------------------------------------------------------------
2118 */
float64_mul(float64 a,float64 b)2119 float64 float64_mul( float64 a, float64 b )
2120 {
2121     flag aSign, bSign, zSign;
2122     int16 aExp, bExp, zExp;
2123     bits64 aSig, bSig, zSig0, zSig1;
2124 
2125     aSig = extractFloat64Frac( a );
2126     aExp = extractFloat64Exp( a );
2127     aSign = extractFloat64Sign( a );
2128     bSig = extractFloat64Frac( b );
2129     bExp = extractFloat64Exp( b );
2130     bSign = extractFloat64Sign( b );
2131     zSign = aSign ^ bSign;
2132     if ( aExp == 0x7FF ) {
2133         if ( aSig || ( ( bExp == 0x7FF ) && bSig ) ) {
2134             return propagateFloat64NaN( a, b );
2135         }
2136         if ( ( bExp | bSig ) == 0 ) {
2137             float_raise( float_flag_invalid );
2138             return float64_default_nan;
2139         }
2140         return packFloat64( zSign, 0x7FF, 0 );
2141     }
2142     if ( bExp == 0x7FF ) {
2143         if ( bSig ) return propagateFloat64NaN( a, b );
2144         if ( ( aExp | aSig ) == 0 ) {
2145             float_raise( float_flag_invalid );
2146             return float64_default_nan;
2147         }
2148         return packFloat64( zSign, 0x7FF, 0 );
2149     }
2150     if ( aExp == 0 ) {
2151         if ( aSig == 0 ) return packFloat64( zSign, 0, 0 );
2152         normalizeFloat64Subnormal( aSig, &aExp, &aSig );
2153     }
2154     if ( bExp == 0 ) {
2155         if ( bSig == 0 ) return packFloat64( zSign, 0, 0 );
2156         normalizeFloat64Subnormal( bSig, &bExp, &bSig );
2157     }
2158     zExp = aExp + bExp - 0x3FF;
2159     aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<10;
2160     bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11;
2161     mul64To128( aSig, bSig, &zSig0, &zSig1 );
2162     zSig0 |= ( zSig1 != 0 );
2163     if ( 0 <= (sbits64) ( zSig0<<1 ) ) {
2164         zSig0 <<= 1;
2165         --zExp;
2166     }
2167     return roundAndPackFloat64( zSign, zExp, zSig0 );
2168 
2169 }
2170 
2171 /*
2172 -------------------------------------------------------------------------------
2173 Returns the result of dividing the double-precision floating-point value `a'
2174 by the corresponding value `b'.  The operation is performed according to
2175 the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2176 -------------------------------------------------------------------------------
2177 */
float64_div(float64 a,float64 b)2178 float64 float64_div( float64 a, float64 b )
2179 {
2180     flag aSign, bSign, zSign;
2181     int16 aExp, bExp, zExp;
2182     bits64 aSig, bSig, zSig;
2183     bits64 rem0, rem1;
2184     bits64 term0, term1;
2185 
2186     aSig = extractFloat64Frac( a );
2187     aExp = extractFloat64Exp( a );
2188     aSign = extractFloat64Sign( a );
2189     bSig = extractFloat64Frac( b );
2190     bExp = extractFloat64Exp( b );
2191     bSign = extractFloat64Sign( b );
2192     zSign = aSign ^ bSign;
2193     if ( aExp == 0x7FF ) {
2194         if ( aSig ) return propagateFloat64NaN( a, b );
2195         if ( bExp == 0x7FF ) {
2196             if ( bSig ) return propagateFloat64NaN( a, b );
2197             float_raise( float_flag_invalid );
2198             return float64_default_nan;
2199         }
2200         return packFloat64( zSign, 0x7FF, 0 );
2201     }
2202     if ( bExp == 0x7FF ) {
2203         if ( bSig ) return propagateFloat64NaN( a, b );
2204         return packFloat64( zSign, 0, 0 );
2205     }
2206     if ( bExp == 0 ) {
2207         if ( bSig == 0 ) {
2208             if ( ( aExp | aSig ) == 0 ) {
2209                 float_raise( float_flag_invalid );
2210                 return float64_default_nan;
2211             }
2212             float_raise( float_flag_divbyzero );
2213             return packFloat64( zSign, 0x7FF, 0 );
2214         }
2215         normalizeFloat64Subnormal( bSig, &bExp, &bSig );
2216     }
2217     if ( aExp == 0 ) {
2218         if ( aSig == 0 ) return packFloat64( zSign, 0, 0 );
2219         normalizeFloat64Subnormal( aSig, &aExp, &aSig );
2220     }
2221     zExp = aExp - bExp + 0x3FD;
2222     aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<10;
2223     bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11;
2224     if ( bSig <= ( aSig + aSig ) ) {
2225         aSig >>= 1;
2226         ++zExp;
2227     }
2228     zSig = estimateDiv128To64( aSig, 0, bSig );
2229     if ( ( zSig & 0x1FF ) <= 2 ) {
2230         mul64To128( bSig, zSig, &term0, &term1 );
2231         sub128( aSig, 0, term0, term1, &rem0, &rem1 );
2232         while ( (sbits64) rem0 < 0 ) {
2233             --zSig;
2234             add128( rem0, rem1, 0, bSig, &rem0, &rem1 );
2235         }
2236         zSig |= ( rem1 != 0 );
2237     }
2238     return roundAndPackFloat64( zSign, zExp, zSig );
2239 
2240 }
2241 
2242 /*
2243 -------------------------------------------------------------------------------
2244 Returns the remainder of the double-precision floating-point value `a'
2245 with respect to the corresponding value `b'.  The operation is performed
2246 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2247 -------------------------------------------------------------------------------
2248 */
float64_rem(float64 a,float64 b)2249 float64 float64_rem( float64 a, float64 b )
2250 {
2251     flag aSign, bSign, zSign;
2252     int16 aExp, bExp, expDiff;
2253     bits64 aSig, bSig;
2254     bits64 q, alternateASig;
2255     sbits64 sigMean;
2256 
2257     aSig = extractFloat64Frac( a );
2258     aExp = extractFloat64Exp( a );
2259     aSign = extractFloat64Sign( a );
2260     bSig = extractFloat64Frac( b );
2261     bExp = extractFloat64Exp( b );
2262     bSign = extractFloat64Sign( b );
2263     if ( aExp == 0x7FF ) {
2264         if ( aSig || ( ( bExp == 0x7FF ) && bSig ) ) {
2265             return propagateFloat64NaN( a, b );
2266         }
2267         float_raise( float_flag_invalid );
2268         return float64_default_nan;
2269     }
2270     if ( bExp == 0x7FF ) {
2271         if ( bSig ) return propagateFloat64NaN( a, b );
2272         return a;
2273     }
2274     if ( bExp == 0 ) {
2275         if ( bSig == 0 ) {
2276             float_raise( float_flag_invalid );
2277             return float64_default_nan;
2278         }
2279         normalizeFloat64Subnormal( bSig, &bExp, &bSig );
2280     }
2281     if ( aExp == 0 ) {
2282         if ( aSig == 0 ) return a;
2283         normalizeFloat64Subnormal( aSig, &aExp, &aSig );
2284     }
2285     expDiff = aExp - bExp;
2286     aSig = ( aSig | LIT64( 0x0010000000000000 ) )<<11;
2287     bSig = ( bSig | LIT64( 0x0010000000000000 ) )<<11;
2288     if ( expDiff < 0 ) {
2289         if ( expDiff < -1 ) return a;
2290         aSig >>= 1;
2291     }
2292     q = ( bSig <= aSig );
2293     if ( q ) aSig -= bSig;
2294     expDiff -= 64;
2295     while ( 0 < expDiff ) {
2296         q = estimateDiv128To64( aSig, 0, bSig );
2297         q = ( 2 < q ) ? q - 2 : 0;
2298         aSig = - ( ( bSig>>2 ) * q );
2299         expDiff -= 62;
2300     }
2301     expDiff += 64;
2302     if ( 0 < expDiff ) {
2303         q = estimateDiv128To64( aSig, 0, bSig );
2304         q = ( 2 < q ) ? q - 2 : 0;
2305         q >>= 64 - expDiff;
2306         bSig >>= 2;
2307         aSig = ( ( aSig>>1 )<<( expDiff - 1 ) ) - bSig * q;
2308     }
2309     else {
2310         aSig >>= 2;
2311         bSig >>= 2;
2312     }
2313     do {
2314         alternateASig = aSig;
2315         ++q;
2316         aSig -= bSig;
2317     } while ( 0 <= (sbits64) aSig );
2318     sigMean = aSig + alternateASig;
2319     if ( ( sigMean < 0 ) || ( ( sigMean == 0 ) && ( q & 1 ) ) ) {
2320         aSig = alternateASig;
2321     }
2322     zSign = ( (sbits64) aSig < 0 );
2323     if ( zSign ) aSig = - aSig;
2324     return normalizeRoundAndPackFloat64( aSign ^ zSign, bExp, aSig );
2325 
2326 }
2327 
2328 /*
2329 -------------------------------------------------------------------------------
2330 Returns the square root of the double-precision floating-point value `a'.
2331 The operation is performed according to the IEC/IEEE Standard for Binary
2332 Floating-point Arithmetic.
2333 -------------------------------------------------------------------------------
2334 */
float64_sqrt(float64 a)2335 float64 float64_sqrt( float64 a )
2336 {
2337     flag aSign;
2338     int16 aExp, zExp;
2339     bits64 aSig, zSig;
2340     bits64 rem0, rem1, term0, term1; //, shiftedRem;
2341     //float64 z;
2342 
2343     aSig = extractFloat64Frac( a );
2344     aExp = extractFloat64Exp( a );
2345     aSign = extractFloat64Sign( a );
2346     if ( aExp == 0x7FF ) {
2347         if ( aSig ) return propagateFloat64NaN( a, a );
2348         if ( ! aSign ) return a;
2349         float_raise( float_flag_invalid );
2350         return float64_default_nan;
2351     }
2352     if ( aSign ) {
2353         if ( ( aExp | aSig ) == 0 ) return a;
2354         float_raise( float_flag_invalid );
2355         return float64_default_nan;
2356     }
2357     if ( aExp == 0 ) {
2358         if ( aSig == 0 ) return 0;
2359         normalizeFloat64Subnormal( aSig, &aExp, &aSig );
2360     }
2361     zExp = ( ( aExp - 0x3FF )>>1 ) + 0x3FE;
2362     aSig |= LIT64( 0x0010000000000000 );
2363     zSig = estimateSqrt32( aExp, aSig>>21 );
2364     zSig <<= 31;
2365     aSig <<= 9 - ( aExp & 1 );
2366     zSig = estimateDiv128To64( aSig, 0, zSig ) + zSig + 2;
2367     if ( ( zSig & 0x3FF ) <= 5 ) {
2368         if ( zSig < 2 ) {
2369             zSig = LIT64( 0xFFFFFFFFFFFFFFFF );
2370         }
2371         else {
2372             aSig <<= 2;
2373             mul64To128( zSig, zSig, &term0, &term1 );
2374             sub128( aSig, 0, term0, term1, &rem0, &rem1 );
2375             while ( (sbits64) rem0 < 0 ) {
2376                 --zSig;
2377                 shortShift128Left( 0, zSig, 1, &term0, &term1 );
2378                 term1 |= 1;
2379                 add128( rem0, rem1, term0, term1, &rem0, &rem1 );
2380             }
2381             zSig |= ( ( rem0 | rem1 ) != 0 );
2382         }
2383     }
2384     shift64RightJamming( zSig, 1, &zSig );
2385     return roundAndPackFloat64( 0, zExp, zSig );
2386 
2387 }
2388 
2389 /*
2390 -------------------------------------------------------------------------------
2391 Returns 1 if the double-precision floating-point value `a' is equal to the
2392 corresponding value `b', and 0 otherwise.  The comparison is performed
2393 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2394 -------------------------------------------------------------------------------
2395 */
float64_eq(float64 a,float64 b)2396 flag float64_eq( float64 a, float64 b )
2397 {
2398 
2399     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2400          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2401        ) {
2402         if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) {
2403             float_raise( float_flag_invalid );
2404         }
2405         return 0;
2406     }
2407     return ( a == b ) || ( (bits64) ( ( a | b )<<1 ) == 0 );
2408 
2409 }
2410 
2411 /*
2412 -------------------------------------------------------------------------------
2413 Returns 1 if the double-precision floating-point value `a' is less than or
2414 equal to the corresponding value `b', and 0 otherwise.  The comparison is
2415 performed according to the IEC/IEEE Standard for Binary Floating-point
2416 Arithmetic.
2417 -------------------------------------------------------------------------------
2418 */
float64_le(float64 a,float64 b)2419 flag float64_le( float64 a, float64 b )
2420 {
2421     flag aSign, bSign;
2422 
2423     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2424          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2425        ) {
2426         float_raise( float_flag_invalid );
2427         return 0;
2428     }
2429     aSign = extractFloat64Sign( a );
2430     bSign = extractFloat64Sign( b );
2431     if ( aSign != bSign ) return aSign || ( (bits64) ( ( a | b )<<1 ) == 0 );
2432     return ( a == b ) || ( aSign ^ ( a < b ) );
2433 
2434 }
2435 
2436 /*
2437 -------------------------------------------------------------------------------
2438 Returns 1 if the double-precision floating-point value `a' is less than
2439 the corresponding value `b', and 0 otherwise.  The comparison is performed
2440 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2441 -------------------------------------------------------------------------------
2442 */
float64_lt(float64 a,float64 b)2443 flag float64_lt( float64 a, float64 b )
2444 {
2445     flag aSign, bSign;
2446 
2447     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2448          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2449        ) {
2450         float_raise( float_flag_invalid );
2451         return 0;
2452     }
2453     aSign = extractFloat64Sign( a );
2454     bSign = extractFloat64Sign( b );
2455     if ( aSign != bSign ) return aSign && ( (bits64) ( ( a | b )<<1 ) != 0 );
2456     return ( a != b ) && ( aSign ^ ( a < b ) );
2457 
2458 }
2459 
2460 /*
2461 -------------------------------------------------------------------------------
2462 Returns 1 if the double-precision floating-point value `a' is equal to the
2463 corresponding value `b', and 0 otherwise.  The invalid exception is raised
2464 if either operand is a NaN.  Otherwise, the comparison is performed
2465 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2466 -------------------------------------------------------------------------------
2467 */
float64_eq_signaling(float64 a,float64 b)2468 flag float64_eq_signaling( float64 a, float64 b )
2469 {
2470 
2471     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2472          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2473        ) {
2474         float_raise( float_flag_invalid );
2475         return 0;
2476     }
2477     return ( a == b ) || ( (bits64) ( ( a | b )<<1 ) == 0 );
2478 
2479 }
2480 
2481 /*
2482 -------------------------------------------------------------------------------
2483 Returns 1 if the double-precision floating-point value `a' is less than or
2484 equal to the corresponding value `b', and 0 otherwise.  Quiet NaNs do not
2485 cause an exception.  Otherwise, the comparison is performed according to the
2486 IEC/IEEE Standard for Binary Floating-point Arithmetic.
2487 -------------------------------------------------------------------------------
2488 */
float64_le_quiet(float64 a,float64 b)2489 flag float64_le_quiet( float64 a, float64 b )
2490 {
2491     flag aSign, bSign;
2492     //int16 aExp, bExp;
2493 
2494     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2495          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2496        ) {
2497         if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) {
2498             float_raise( float_flag_invalid );
2499         }
2500         return 0;
2501     }
2502     aSign = extractFloat64Sign( a );
2503     bSign = extractFloat64Sign( b );
2504     if ( aSign != bSign ) return aSign || ( (bits64) ( ( a | b )<<1 ) == 0 );
2505     return ( a == b ) || ( aSign ^ ( a < b ) );
2506 
2507 }
2508 
2509 /*
2510 -------------------------------------------------------------------------------
2511 Returns 1 if the double-precision floating-point value `a' is less than
2512 the corresponding value `b', and 0 otherwise.  Quiet NaNs do not cause an
2513 exception.  Otherwise, the comparison is performed according to the IEC/IEEE
2514 Standard for Binary Floating-point Arithmetic.
2515 -------------------------------------------------------------------------------
2516 */
float64_lt_quiet(float64 a,float64 b)2517 flag float64_lt_quiet( float64 a, float64 b )
2518 {
2519     flag aSign, bSign;
2520 
2521     if (    ( ( extractFloat64Exp( a ) == 0x7FF ) && extractFloat64Frac( a ) )
2522          || ( ( extractFloat64Exp( b ) == 0x7FF ) && extractFloat64Frac( b ) )
2523        ) {
2524         if ( float64_is_signaling_nan( a ) || float64_is_signaling_nan( b ) ) {
2525             float_raise( float_flag_invalid );
2526         }
2527         return 0;
2528     }
2529     aSign = extractFloat64Sign( a );
2530     bSign = extractFloat64Sign( b );
2531     if ( aSign != bSign ) return aSign && ( (bits64) ( ( a | b )<<1 ) != 0 );
2532     return ( a != b ) && ( aSign ^ ( a < b ) );
2533 
2534 }
2535 
2536 #ifdef FLOATX80
2537 
2538 /*
2539 -------------------------------------------------------------------------------
2540 Returns the result of converting the extended double-precision floating-
2541 point value `a' to the 32-bit two's complement integer format.  The
2542 conversion is performed according to the IEC/IEEE Standard for Binary
2543 Floating-point Arithmetic---which means in particular that the conversion
2544 is rounded according to the current rounding mode.  If `a' is a NaN, the
2545 largest positive integer is returned.  Otherwise, if the conversion
2546 overflows, the largest integer with the same sign as `a' is returned.
2547 -------------------------------------------------------------------------------
2548 */
floatx80_to_int32(floatx80 a)2549 int32 floatx80_to_int32( floatx80 a )
2550 {
2551     flag aSign;
2552     int32 aExp, shiftCount;
2553     bits64 aSig;
2554 
2555     aSig = extractFloatx80Frac( a );
2556     aExp = extractFloatx80Exp( a );
2557     aSign = extractFloatx80Sign( a );
2558     if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) aSign = 0;
2559     shiftCount = 0x4037 - aExp;
2560     if ( shiftCount <= 0 ) shiftCount = 1;
2561     shift64RightJamming( aSig, shiftCount, &aSig );
2562     return roundAndPackInt32( aSign, aSig );
2563 
2564 }
2565 
2566 /*
2567 -------------------------------------------------------------------------------
2568 Returns the result of converting the extended double-precision floating-
2569 point value `a' to the 32-bit two's complement integer format.  The
2570 conversion is performed according to the IEC/IEEE Standard for Binary
2571 Floating-point Arithmetic, except that the conversion is always rounded
2572 toward zero.  If `a' is a NaN, the largest positive integer is returned.
2573 Otherwise, if the conversion overflows, the largest integer with the same
2574 sign as `a' is returned.
2575 -------------------------------------------------------------------------------
2576 */
floatx80_to_int32_round_to_zero(floatx80 a)2577 int32 floatx80_to_int32_round_to_zero( floatx80 a )
2578 {
2579     flag aSign;
2580     int32 aExp, shiftCount;
2581     bits64 aSig, savedASig;
2582     int32 z;
2583 
2584     aSig = extractFloatx80Frac( a );
2585     aExp = extractFloatx80Exp( a );
2586     aSign = extractFloatx80Sign( a );
2587     shiftCount = 0x403E - aExp;
2588     if ( shiftCount < 32 ) {
2589         if ( ( aExp == 0x7FFF ) && (bits64) ( aSig<<1 ) ) aSign = 0;
2590         goto invalid;
2591     }
2592     else if ( 63 < shiftCount ) {
2593         if ( aExp || aSig ) float_exception_flags |= float_flag_inexact;
2594         return 0;
2595     }
2596     savedASig = aSig;
2597     aSig >>= shiftCount;
2598     z = aSig;
2599     if ( aSign ) z = - z;
2600     if ( ( z < 0 ) ^ aSign ) {
2601  invalid:
2602         float_exception_flags |= float_flag_invalid;
2603         return aSign ? 0x80000000 : 0x7FFFFFFF;
2604     }
2605     if ( ( aSig<<shiftCount ) != savedASig ) {
2606         float_exception_flags |= float_flag_inexact;
2607     }
2608     return z;
2609 
2610 }
2611 
2612 /*
2613 -------------------------------------------------------------------------------
2614 Returns the result of converting the extended double-precision floating-
2615 point value `a' to the single-precision floating-point format.  The
2616 conversion is performed according to the IEC/IEEE Standard for Binary
2617 Floating-point Arithmetic.
2618 -------------------------------------------------------------------------------
2619 */
floatx80_to_float32(floatx80 a)2620 float32 floatx80_to_float32( floatx80 a )
2621 {
2622     flag aSign;
2623     int32 aExp;
2624     bits64 aSig;
2625 
2626     aSig = extractFloatx80Frac( a );
2627     aExp = extractFloatx80Exp( a );
2628     aSign = extractFloatx80Sign( a );
2629     if ( aExp == 0x7FFF ) {
2630         if ( (bits64) ( aSig<<1 ) ) {
2631             return commonNaNToFloat32( floatx80ToCommonNaN( a ) );
2632         }
2633         return packFloat32( aSign, 0xFF, 0 );
2634     }
2635     shift64RightJamming( aSig, 33, &aSig );
2636     if ( aExp || aSig ) aExp -= 0x3F81;
2637     return roundAndPackFloat32( aSign, aExp, aSig );
2638 
2639 }
2640 
2641 /*
2642 -------------------------------------------------------------------------------
2643 Returns the result of converting the extended double-precision floating-
2644 point value `a' to the double-precision floating-point format.  The
2645 conversion is performed according to the IEC/IEEE Standard for Binary
2646 Floating-point Arithmetic.
2647 -------------------------------------------------------------------------------
2648 */
floatx80_to_float64(floatx80 a)2649 float64 floatx80_to_float64( floatx80 a )
2650 {
2651     flag aSign;
2652     int32 aExp;
2653     bits64 aSig, zSig;
2654 
2655     aSig = extractFloatx80Frac( a );
2656     aExp = extractFloatx80Exp( a );
2657     aSign = extractFloatx80Sign( a );
2658     if ( aExp == 0x7FFF ) {
2659         if ( (bits64) ( aSig<<1 ) ) {
2660             return commonNaNToFloat64( floatx80ToCommonNaN( a ) );
2661         }
2662         return packFloat64( aSign, 0x7FF, 0 );
2663     }
2664     shift64RightJamming( aSig, 1, &zSig );
2665     if ( aExp || aSig ) aExp -= 0x3C01;
2666     return roundAndPackFloat64( aSign, aExp, zSig );
2667 
2668 }
2669 
2670 /*
2671 -------------------------------------------------------------------------------
2672 Rounds the extended double-precision floating-point value `a' to an integer,
2673 and returns the result as an extended quadruple-precision floating-point
2674 value.  The operation is performed according to the IEC/IEEE Standard for
2675 Binary Floating-point Arithmetic.
2676 -------------------------------------------------------------------------------
2677 */
floatx80_round_to_int(floatx80 a)2678 floatx80 floatx80_round_to_int( floatx80 a )
2679 {
2680     flag aSign;
2681     int32 aExp;
2682     bits64 lastBitMask, roundBitsMask;
2683     int8 roundingMode;
2684     floatx80 z;
2685 
2686     aExp = extractFloatx80Exp( a );
2687     if ( 0x403E <= aExp ) {
2688         if ( ( aExp == 0x7FFF ) && (bits64) ( extractFloatx80Frac( a )<<1 ) ) {
2689             return propagateFloatx80NaN( a, a );
2690         }
2691         return a;
2692     }
2693     if ( aExp <= 0x3FFE ) {
2694         if (    ( aExp == 0 )
2695              && ( (bits64) ( extractFloatx80Frac( a )<<1 ) == 0 ) ) {
2696             return a;
2697         }
2698         float_exception_flags |= float_flag_inexact;
2699         aSign = extractFloatx80Sign( a );
2700         switch ( float_rounding_mode ) {
2701          case float_round_nearest_even:
2702             if ( ( aExp == 0x3FFE ) && (bits64) ( extractFloatx80Frac( a )<<1 )
2703                ) {
2704                 return
2705                     packFloatx80( aSign, 0x3FFF, LIT64( 0x8000000000000000 ) );
2706             }
2707             break;
2708          case float_round_down:
2709             return
2710                   aSign ?
2711                       packFloatx80( 1, 0x3FFF, LIT64( 0x8000000000000000 ) )
2712                 : packFloatx80( 0, 0, 0 );
2713          case float_round_up:
2714             return
2715                   aSign ? packFloatx80( 1, 0, 0 )
2716                 : packFloatx80( 0, 0x3FFF, LIT64( 0x8000000000000000 ) );
2717         }
2718         return packFloatx80( aSign, 0, 0 );
2719     }
2720     lastBitMask = 1;
2721     lastBitMask <<= 0x403E - aExp;
2722     roundBitsMask = lastBitMask - 1;
2723     z = a;
2724     roundingMode = float_rounding_mode;
2725     if ( roundingMode == float_round_nearest_even ) {
2726         z.low += lastBitMask>>1;
2727         if ( ( z.low & roundBitsMask ) == 0 ) z.low &= ~ lastBitMask;
2728     }
2729     else if ( roundingMode != float_round_to_zero ) {
2730         if ( extractFloatx80Sign( z ) ^ ( roundingMode == float_round_up ) ) {
2731             z.low += roundBitsMask;
2732         }
2733     }
2734     z.low &= ~ roundBitsMask;
2735     if ( z.low == 0 ) {
2736         ++z.high;
2737         z.low = LIT64( 0x8000000000000000 );
2738     }
2739     if ( z.low != a.low ) float_exception_flags |= float_flag_inexact;
2740     return z;
2741 
2742 }
2743 
2744 /*
2745 -------------------------------------------------------------------------------
2746 Returns the result of adding the absolute values of the extended double-
2747 precision floating-point values `a' and `b'.  If `zSign' is true, the sum is
2748 negated before being returned.  `zSign' is ignored if the result is a NaN.
2749 The addition is performed according to the IEC/IEEE Standard for Binary
2750 Floating-point Arithmetic.
2751 -------------------------------------------------------------------------------
2752 */
addFloatx80Sigs(floatx80 a,floatx80 b,flag zSign)2753 static floatx80 addFloatx80Sigs( floatx80 a, floatx80 b, flag zSign )
2754 {
2755     int32 aExp, bExp, zExp;
2756     bits64 aSig, bSig, zSig0, zSig1;
2757     int32 expDiff;
2758 
2759     aSig = extractFloatx80Frac( a );
2760     aExp = extractFloatx80Exp( a );
2761     bSig = extractFloatx80Frac( b );
2762     bExp = extractFloatx80Exp( b );
2763     expDiff = aExp - bExp;
2764     if ( 0 < expDiff ) {
2765         if ( aExp == 0x7FFF ) {
2766             if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b );
2767             return a;
2768         }
2769         if ( bExp == 0 ) --expDiff;
2770         shift64ExtraRightJamming( bSig, 0, expDiff, &bSig, &zSig1 );
2771         zExp = aExp;
2772     }
2773     else if ( expDiff < 0 ) {
2774         if ( bExp == 0x7FFF ) {
2775             if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
2776             return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
2777         }
2778         if ( aExp == 0 ) ++expDiff;
2779         shift64ExtraRightJamming( aSig, 0, - expDiff, &aSig, &zSig1 );
2780         zExp = bExp;
2781     }
2782     else {
2783         if ( aExp == 0x7FFF ) {
2784             if ( (bits64) ( ( aSig | bSig )<<1 ) ) {
2785                 return propagateFloatx80NaN( a, b );
2786             }
2787             return a;
2788         }
2789         zSig1 = 0;
2790         zSig0 = aSig + bSig;
2791         if ( aExp == 0 ) {
2792             normalizeFloatx80Subnormal( zSig0, &zExp, &zSig0 );
2793             goto roundAndPack;
2794         }
2795         zExp = aExp;
2796         goto shiftRight1;
2797     }
2798 
2799     zSig0 = aSig + bSig;
2800 
2801     if ( (sbits64) zSig0 < 0 ) goto roundAndPack;
2802  shiftRight1:
2803     shift64ExtraRightJamming( zSig0, zSig1, 1, &zSig0, &zSig1 );
2804     zSig0 |= LIT64( 0x8000000000000000 );
2805     ++zExp;
2806  roundAndPack:
2807     return
2808         roundAndPackFloatx80(
2809             floatx80_rounding_precision, zSign, zExp, zSig0, zSig1 );
2810 
2811 }
2812 
2813 /*
2814 -------------------------------------------------------------------------------
2815 Returns the result of subtracting the absolute values of the extended
2816 double-precision floating-point values `a' and `b'.  If `zSign' is true,
2817 the difference is negated before being returned.  `zSign' is ignored if the
2818 result is a NaN.  The subtraction is performed according to the IEC/IEEE
2819 Standard for Binary Floating-point Arithmetic.
2820 -------------------------------------------------------------------------------
2821 */
subFloatx80Sigs(floatx80 a,floatx80 b,flag zSign)2822 static floatx80 subFloatx80Sigs( floatx80 a, floatx80 b, flag zSign )
2823 {
2824     int32 aExp, bExp, zExp;
2825     bits64 aSig, bSig, zSig0, zSig1;
2826     int32 expDiff;
2827     floatx80 z;
2828 
2829     aSig = extractFloatx80Frac( a );
2830     aExp = extractFloatx80Exp( a );
2831     bSig = extractFloatx80Frac( b );
2832     bExp = extractFloatx80Exp( b );
2833     expDiff = aExp - bExp;
2834     if ( 0 < expDiff ) goto aExpBigger;
2835     if ( expDiff < 0 ) goto bExpBigger;
2836     if ( aExp == 0x7FFF ) {
2837         if ( (bits64) ( ( aSig | bSig )<<1 ) ) {
2838             return propagateFloatx80NaN( a, b );
2839         }
2840         float_raise( float_flag_invalid );
2841         z.low = floatx80_default_nan_low;
2842         z.high = floatx80_default_nan_high;
2843         return z;
2844     }
2845     if ( aExp == 0 ) {
2846         aExp = 1;
2847         bExp = 1;
2848     }
2849     zSig1 = 0;
2850     if ( bSig < aSig ) goto aBigger;
2851     if ( aSig < bSig ) goto bBigger;
2852     return packFloatx80( float_rounding_mode == float_round_down, 0, 0 );
2853  bExpBigger:
2854     if ( bExp == 0x7FFF ) {
2855         if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
2856         return packFloatx80( zSign ^ 1, 0x7FFF, LIT64( 0x8000000000000000 ) );
2857     }
2858     if ( aExp == 0 ) ++expDiff;
2859     shift128RightJamming( aSig, 0, - expDiff, &aSig, &zSig1 );
2860  bBigger:
2861     sub128( bSig, 0, aSig, zSig1, &zSig0, &zSig1 );
2862     zExp = bExp;
2863     zSign ^= 1;
2864     goto normalizeRoundAndPack;
2865  aExpBigger:
2866     if ( aExp == 0x7FFF ) {
2867         if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b );
2868         return a;
2869     }
2870     if ( bExp == 0 ) --expDiff;
2871     shift128RightJamming( bSig, 0, expDiff, &bSig, &zSig1 );
2872  aBigger:
2873     sub128( aSig, 0, bSig, zSig1, &zSig0, &zSig1 );
2874     zExp = aExp;
2875  normalizeRoundAndPack:
2876     return
2877         normalizeRoundAndPackFloatx80(
2878             floatx80_rounding_precision, zSign, zExp, zSig0, zSig1 );
2879 
2880 }
2881 
2882 /*
2883 -------------------------------------------------------------------------------
2884 Returns the result of adding the extended double-precision floating-point
2885 values `a' and `b'.  The operation is performed according to the IEC/IEEE
2886 Standard for Binary Floating-point Arithmetic.
2887 -------------------------------------------------------------------------------
2888 */
floatx80_add(floatx80 a,floatx80 b)2889 floatx80 floatx80_add( floatx80 a, floatx80 b )
2890 {
2891     flag aSign, bSign;
2892 
2893     aSign = extractFloatx80Sign( a );
2894     bSign = extractFloatx80Sign( b );
2895     if ( aSign == bSign ) {
2896         return addFloatx80Sigs( a, b, aSign );
2897     }
2898     else {
2899         return subFloatx80Sigs( a, b, aSign );
2900     }
2901 
2902 }
2903 
2904 /*
2905 -------------------------------------------------------------------------------
2906 Returns the result of subtracting the extended double-precision floating-
2907 point values `a' and `b'.  The operation is performed according to the
2908 IEC/IEEE Standard for Binary Floating-point Arithmetic.
2909 -------------------------------------------------------------------------------
2910 */
floatx80_sub(floatx80 a,floatx80 b)2911 floatx80 floatx80_sub( floatx80 a, floatx80 b )
2912 {
2913     flag aSign, bSign;
2914 
2915     aSign = extractFloatx80Sign( a );
2916     bSign = extractFloatx80Sign( b );
2917     if ( aSign == bSign ) {
2918         return subFloatx80Sigs( a, b, aSign );
2919     }
2920     else {
2921         return addFloatx80Sigs( a, b, aSign );
2922     }
2923 
2924 }
2925 
2926 /*
2927 -------------------------------------------------------------------------------
2928 Returns the result of multiplying the extended double-precision floating-
2929 point values `a' and `b'.  The operation is performed according to the
2930 IEC/IEEE Standard for Binary Floating-point Arithmetic.
2931 -------------------------------------------------------------------------------
2932 */
floatx80_mul(floatx80 a,floatx80 b)2933 floatx80 floatx80_mul( floatx80 a, floatx80 b )
2934 {
2935     flag aSign, bSign, zSign;
2936     int32 aExp, bExp, zExp;
2937     bits64 aSig, bSig, zSig0, zSig1;
2938     floatx80 z;
2939 
2940     aSig = extractFloatx80Frac( a );
2941     aExp = extractFloatx80Exp( a );
2942     aSign = extractFloatx80Sign( a );
2943     bSig = extractFloatx80Frac( b );
2944     bExp = extractFloatx80Exp( b );
2945     bSign = extractFloatx80Sign( b );
2946     zSign = aSign ^ bSign;
2947     if ( aExp == 0x7FFF ) {
2948         if (    (bits64) ( aSig<<1 )
2949              || ( ( bExp == 0x7FFF ) && (bits64) ( bSig<<1 ) ) ) {
2950             return propagateFloatx80NaN( a, b );
2951         }
2952         if ( ( bExp | bSig ) == 0 ) goto invalid;
2953         return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
2954     }
2955     if ( bExp == 0x7FFF ) {
2956         if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
2957         if ( ( aExp | aSig ) == 0 ) {
2958  invalid:
2959             float_raise( float_flag_invalid );
2960             z.low = floatx80_default_nan_low;
2961             z.high = floatx80_default_nan_high;
2962             return z;
2963         }
2964         return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
2965     }
2966     if ( aExp == 0 ) {
2967         if ( aSig == 0 ) return packFloatx80( zSign, 0, 0 );
2968         normalizeFloatx80Subnormal( aSig, &aExp, &aSig );
2969     }
2970     if ( bExp == 0 ) {
2971         if ( bSig == 0 ) return packFloatx80( zSign, 0, 0 );
2972         normalizeFloatx80Subnormal( bSig, &bExp, &bSig );
2973     }
2974     zExp = aExp + bExp - 0x3FFE;
2975     mul64To128( aSig, bSig, &zSig0, &zSig1 );
2976     if ( 0 < (sbits64) zSig0 ) {
2977         shortShift128Left( zSig0, zSig1, 1, &zSig0, &zSig1 );
2978         --zExp;
2979     }
2980     return
2981         roundAndPackFloatx80(
2982             floatx80_rounding_precision, zSign, zExp, zSig0, zSig1 );
2983 
2984 }
2985 
2986 /*
2987 -------------------------------------------------------------------------------
2988 Returns the result of dividing the extended double-precision floating-point
2989 value `a' by the corresponding value `b'.  The operation is performed
2990 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
2991 -------------------------------------------------------------------------------
2992 */
floatx80_div(floatx80 a,floatx80 b)2993 floatx80 floatx80_div( floatx80 a, floatx80 b )
2994 {
2995     flag aSign, bSign, zSign;
2996     int32 aExp, bExp, zExp;
2997     bits64 aSig, bSig, zSig0, zSig1;
2998     bits64 rem0, rem1, rem2, term0, term1, term2;
2999     floatx80 z;
3000 
3001     aSig = extractFloatx80Frac( a );
3002     aExp = extractFloatx80Exp( a );
3003     aSign = extractFloatx80Sign( a );
3004     bSig = extractFloatx80Frac( b );
3005     bExp = extractFloatx80Exp( b );
3006     bSign = extractFloatx80Sign( b );
3007     zSign = aSign ^ bSign;
3008     if ( aExp == 0x7FFF ) {
3009         if ( (bits64) ( aSig<<1 ) ) return propagateFloatx80NaN( a, b );
3010         if ( bExp == 0x7FFF ) {
3011             if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
3012             goto invalid;
3013         }
3014         return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
3015     }
3016     if ( bExp == 0x7FFF ) {
3017         if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
3018         return packFloatx80( zSign, 0, 0 );
3019     }
3020     if ( bExp == 0 ) {
3021         if ( bSig == 0 ) {
3022             if ( ( aExp | aSig ) == 0 ) {
3023  invalid:
3024                 float_raise( float_flag_invalid );
3025                 z.low = floatx80_default_nan_low;
3026                 z.high = floatx80_default_nan_high;
3027                 return z;
3028             }
3029             float_raise( float_flag_divbyzero );
3030             return packFloatx80( zSign, 0x7FFF, LIT64( 0x8000000000000000 ) );
3031         }
3032         normalizeFloatx80Subnormal( bSig, &bExp, &bSig );
3033     }
3034     if ( aExp == 0 ) {
3035         if ( aSig == 0 ) return packFloatx80( zSign, 0, 0 );
3036         normalizeFloatx80Subnormal( aSig, &aExp, &aSig );
3037     }
3038     zExp = aExp - bExp + 0x3FFE;
3039     rem1 = 0;
3040     if ( bSig <= aSig ) {
3041         shift128Right( aSig, 0, 1, &aSig, &rem1 );
3042         ++zExp;
3043     }
3044     zSig0 = estimateDiv128To64( aSig, rem1, bSig );
3045     mul64To128( bSig, zSig0, &term0, &term1 );
3046     sub128( aSig, rem1, term0, term1, &rem0, &rem1 );
3047     while ( (sbits64) rem0 < 0 ) {
3048         --zSig0;
3049         add128( rem0, rem1, 0, bSig, &rem0, &rem1 );
3050     }
3051     zSig1 = estimateDiv128To64( rem1, 0, bSig );
3052     if ( (bits64) ( zSig1<<1 ) <= 8 ) {
3053         mul64To128( bSig, zSig1, &term1, &term2 );
3054         sub128( rem1, 0, term1, term2, &rem1, &rem2 );
3055         while ( (sbits64) rem1 < 0 ) {
3056             --zSig1;
3057             add128( rem1, rem2, 0, bSig, &rem1, &rem2 );
3058         }
3059         zSig1 |= ( ( rem1 | rem2 ) != 0 );
3060     }
3061     return
3062         roundAndPackFloatx80(
3063             floatx80_rounding_precision, zSign, zExp, zSig0, zSig1 );
3064 
3065 }
3066 
3067 /*
3068 -------------------------------------------------------------------------------
3069 Returns the remainder of the extended double-precision floating-point value
3070 `a' with respect to the corresponding value `b'.  The operation is performed
3071 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
3072 -------------------------------------------------------------------------------
3073 */
floatx80_rem(floatx80 a,floatx80 b)3074 floatx80 floatx80_rem( floatx80 a, floatx80 b )
3075 {
3076     flag aSign, bSign, zSign;
3077     int32 aExp, bExp, expDiff;
3078     bits64 aSig0, aSig1, bSig;
3079     bits64 q, term0, term1, alternateASig0, alternateASig1;
3080     floatx80 z;
3081 
3082     aSig0 = extractFloatx80Frac( a );
3083     aExp = extractFloatx80Exp( a );
3084     aSign = extractFloatx80Sign( a );
3085     bSig = extractFloatx80Frac( b );
3086     bExp = extractFloatx80Exp( b );
3087     bSign = extractFloatx80Sign( b );
3088     if ( aExp == 0x7FFF ) {
3089         if (    (bits64) ( aSig0<<1 )
3090              || ( ( bExp == 0x7FFF ) && (bits64) ( bSig<<1 ) ) ) {
3091             return propagateFloatx80NaN( a, b );
3092         }
3093         goto invalid;
3094     }
3095     if ( bExp == 0x7FFF ) {
3096         if ( (bits64) ( bSig<<1 ) ) return propagateFloatx80NaN( a, b );
3097         return a;
3098     }
3099     if ( bExp == 0 ) {
3100         if ( bSig == 0 ) {
3101  invalid:
3102             float_raise( float_flag_invalid );
3103             z.low = floatx80_default_nan_low;
3104             z.high = floatx80_default_nan_high;
3105             return z;
3106         }
3107         normalizeFloatx80Subnormal( bSig, &bExp, &bSig );
3108     }
3109     if ( aExp == 0 ) {
3110         if ( (bits64) ( aSig0<<1 ) == 0 ) return a;
3111         normalizeFloatx80Subnormal( aSig0, &aExp, &aSig0 );
3112     }
3113     bSig |= LIT64( 0x8000000000000000 );
3114     zSign = aSign;
3115     expDiff = aExp - bExp;
3116     aSig1 = 0;
3117     if ( expDiff < 0 ) {
3118         if ( expDiff < -1 ) return a;
3119         shift128Right( aSig0, 0, 1, &aSig0, &aSig1 );
3120         expDiff = 0;
3121     }
3122     q = ( bSig <= aSig0 );
3123     if ( q ) aSig0 -= bSig;
3124     expDiff -= 64;
3125     while ( 0 < expDiff ) {
3126         q = estimateDiv128To64( aSig0, aSig1, bSig );
3127         q = ( 2 < q ) ? q - 2 : 0;
3128         mul64To128( bSig, q, &term0, &term1 );
3129         sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 );
3130         shortShift128Left( aSig0, aSig1, 62, &aSig0, &aSig1 );
3131         expDiff -= 62;
3132     }
3133     expDiff += 64;
3134     if ( 0 < expDiff ) {
3135         q = estimateDiv128To64( aSig0, aSig1, bSig );
3136         q = ( 2 < q ) ? q - 2 : 0;
3137         q >>= 64 - expDiff;
3138         mul64To128( bSig, q<<( 64 - expDiff ), &term0, &term1 );
3139         sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 );
3140         shortShift128Left( 0, bSig, 64 - expDiff, &term0, &term1 );
3141         while ( le128( term0, term1, aSig0, aSig1 ) ) {
3142             ++q;
3143             sub128( aSig0, aSig1, term0, term1, &aSig0, &aSig1 );
3144         }
3145     }
3146     else {
3147         term1 = 0;
3148         term0 = bSig;
3149     }
3150     sub128( term0, term1, aSig0, aSig1, &alternateASig0, &alternateASig1 );
3151     if (    lt128( alternateASig0, alternateASig1, aSig0, aSig1 )
3152          || (    eq128( alternateASig0, alternateASig1, aSig0, aSig1 )
3153               && ( q & 1 ) )
3154        ) {
3155         aSig0 = alternateASig0;
3156         aSig1 = alternateASig1;
3157         zSign = ! zSign;
3158     }
3159     return
3160         normalizeRoundAndPackFloatx80(
3161             80, zSign, bExp + expDiff, aSig0, aSig1 );
3162 
3163 }
3164 
3165 /*
3166 -------------------------------------------------------------------------------
3167 Returns the square root of the extended double-precision floating-point
3168 value `a'.  The operation is performed according to the IEC/IEEE Standard
3169 for Binary Floating-point Arithmetic.
3170 -------------------------------------------------------------------------------
3171 */
floatx80_sqrt(floatx80 a)3172 floatx80 floatx80_sqrt( floatx80 a )
3173 {
3174     flag aSign;
3175     int32 aExp, zExp;
3176     bits64 aSig0, aSig1, zSig0, zSig1;
3177     bits64 rem0, rem1, rem2, rem3, term0, term1, term2, term3;
3178     bits64 shiftedRem0, shiftedRem1;
3179     floatx80 z;
3180 
3181     aSig0 = extractFloatx80Frac( a );
3182     aExp = extractFloatx80Exp( a );
3183     aSign = extractFloatx80Sign( a );
3184     if ( aExp == 0x7FFF ) {
3185         if ( (bits64) ( aSig0<<1 ) ) return propagateFloatx80NaN( a, a );
3186         if ( ! aSign ) return a;
3187         goto invalid;
3188     }
3189     if ( aSign ) {
3190         if ( ( aExp | aSig0 ) == 0 ) return a;
3191  invalid:
3192         float_raise( float_flag_invalid );
3193         z.low = floatx80_default_nan_low;
3194         z.high = floatx80_default_nan_high;
3195         return z;
3196     }
3197     if ( aExp == 0 ) {
3198         if ( aSig0 == 0 ) return packFloatx80( 0, 0, 0 );
3199         normalizeFloatx80Subnormal( aSig0, &aExp, &aSig0 );
3200     }
3201     zExp = ( ( aExp - 0x3FFF )>>1 ) + 0x3FFF;
3202     zSig0 = estimateSqrt32( aExp, aSig0>>32 );
3203     zSig0 <<= 31;
3204     aSig1 = 0;
3205     shift128Right( aSig0, 0, ( aExp & 1 ) + 2, &aSig0, &aSig1 );
3206     zSig0 = estimateDiv128To64( aSig0, aSig1, zSig0 ) + zSig0 + 4;
3207     if ( 0 <= (sbits64) zSig0 ) zSig0 = LIT64( 0xFFFFFFFFFFFFFFFF );
3208     shortShift128Left( aSig0, aSig1, 2, &aSig0, &aSig1 );
3209     mul64To128( zSig0, zSig0, &term0, &term1 );
3210     sub128( aSig0, aSig1, term0, term1, &rem0, &rem1 );
3211     while ( (sbits64) rem0 < 0 ) {
3212         --zSig0;
3213         shortShift128Left( 0, zSig0, 1, &term0, &term1 );
3214         term1 |= 1;
3215         add128( rem0, rem1, term0, term1, &rem0, &rem1 );
3216     }
3217     shortShift128Left( rem0, rem1, 63, &shiftedRem0, &shiftedRem1 );
3218     zSig1 = estimateDiv128To64( shiftedRem0, shiftedRem1, zSig0 );
3219     if ( (bits64) ( zSig1<<1 ) <= 10 ) {
3220         if ( zSig1 == 0 ) zSig1 = 1;
3221         mul64To128( zSig0, zSig1, &term1, &term2 );
3222         shortShift128Left( term1, term2, 1, &term1, &term2 );
3223         sub128( rem1, 0, term1, term2, &rem1, &rem2 );
3224         mul64To128( zSig1, zSig1, &term2, &term3 );
3225         sub192( rem1, rem2, 0, 0, term2, term3, &rem1, &rem2, &rem3 );
3226         while ( (sbits64) rem1 < 0 ) {
3227             --zSig1;
3228             shortShift192Left( 0, zSig0, zSig1, 1, &term1, &term2, &term3 );
3229             term3 |= 1;
3230             add192(
3231                 rem1, rem2, rem3, term1, term2, term3, &rem1, &rem2, &rem3 );
3232         }
3233         zSig1 |= ( ( rem1 | rem2 | rem3 ) != 0 );
3234     }
3235     return
3236         roundAndPackFloatx80(
3237             floatx80_rounding_precision, 0, zExp, zSig0, zSig1 );
3238 
3239 }
3240 
3241 /*
3242 -------------------------------------------------------------------------------
3243 Returns 1 if the extended double-precision floating-point value `a' is
3244 equal to the corresponding value `b', and 0 otherwise.  The comparison is
3245 performed according to the IEC/IEEE Standard for Binary Floating-point
3246 Arithmetic.
3247 -------------------------------------------------------------------------------
3248 */
floatx80_eq(floatx80 a,floatx80 b)3249 flag floatx80_eq( floatx80 a, floatx80 b )
3250 {
3251 
3252     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3253               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3254          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3255               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3256        ) {
3257         if (    floatx80_is_signaling_nan( a )
3258              || floatx80_is_signaling_nan( b ) ) {
3259             float_raise( float_flag_invalid );
3260         }
3261         return 0;
3262     }
3263     return
3264            ( a.low == b.low )
3265         && (    ( a.high == b.high )
3266              || (    ( a.low == 0 )
3267                   && ( (bits16) ( ( a.high | b.high )<<1 ) == 0 ) )
3268            );
3269 
3270 }
3271 
3272 /*
3273 -------------------------------------------------------------------------------
3274 Returns 1 if the extended double-precision floating-point value `a' is
3275 less than or equal to the corresponding value `b', and 0 otherwise.  The
3276 comparison is performed according to the IEC/IEEE Standard for Binary
3277 Floating-point Arithmetic.
3278 -------------------------------------------------------------------------------
3279 */
floatx80_le(floatx80 a,floatx80 b)3280 flag floatx80_le( floatx80 a, floatx80 b )
3281 {
3282     flag aSign, bSign;
3283 
3284     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3285               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3286          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3287               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3288        ) {
3289         float_raise( float_flag_invalid );
3290         return 0;
3291     }
3292     aSign = extractFloatx80Sign( a );
3293     bSign = extractFloatx80Sign( b );
3294     if ( aSign != bSign ) {
3295         return
3296                aSign
3297             || (    ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low )
3298                  == 0 );
3299     }
3300     return
3301           aSign ? le128( b.high, b.low, a.high, a.low )
3302         : le128( a.high, a.low, b.high, b.low );
3303 
3304 }
3305 
3306 /*
3307 -------------------------------------------------------------------------------
3308 Returns 1 if the extended double-precision floating-point value `a' is
3309 less than the corresponding value `b', and 0 otherwise.  The comparison
3310 is performed according to the IEC/IEEE Standard for Binary Floating-point
3311 Arithmetic.
3312 -------------------------------------------------------------------------------
3313 */
floatx80_lt(floatx80 a,floatx80 b)3314 flag floatx80_lt( floatx80 a, floatx80 b )
3315 {
3316     flag aSign, bSign;
3317 
3318     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3319               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3320          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3321               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3322        ) {
3323         float_raise( float_flag_invalid );
3324         return 0;
3325     }
3326     aSign = extractFloatx80Sign( a );
3327     bSign = extractFloatx80Sign( b );
3328     if ( aSign != bSign ) {
3329         return
3330                aSign
3331             && (    ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low )
3332                  != 0 );
3333     }
3334     return
3335           aSign ? lt128( b.high, b.low, a.high, a.low )
3336         : lt128( a.high, a.low, b.high, b.low );
3337 
3338 }
3339 
3340 /*
3341 -------------------------------------------------------------------------------
3342 Returns 1 if the extended double-precision floating-point value `a' is equal
3343 to the corresponding value `b', and 0 otherwise.  The invalid exception is
3344 raised if either operand is a NaN.  Otherwise, the comparison is performed
3345 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
3346 -------------------------------------------------------------------------------
3347 */
floatx80_eq_signaling(floatx80 a,floatx80 b)3348 flag floatx80_eq_signaling( floatx80 a, floatx80 b )
3349 {
3350 
3351     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3352               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3353          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3354               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3355        ) {
3356         float_raise( float_flag_invalid );
3357         return 0;
3358     }
3359     return
3360            ( a.low == b.low )
3361         && (    ( a.high == b.high )
3362              || (    ( a.low == 0 )
3363                   && ( (bits16) ( ( a.high | b.high )<<1 ) == 0 ) )
3364            );
3365 
3366 }
3367 
3368 /*
3369 -------------------------------------------------------------------------------
3370 Returns 1 if the extended double-precision floating-point value `a' is less
3371 than or equal to the corresponding value `b', and 0 otherwise.  Quiet NaNs
3372 do not cause an exception.  Otherwise, the comparison is performed according
3373 to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
3374 -------------------------------------------------------------------------------
3375 */
floatx80_le_quiet(floatx80 a,floatx80 b)3376 flag floatx80_le_quiet( floatx80 a, floatx80 b )
3377 {
3378     flag aSign, bSign;
3379 
3380     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3381               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3382          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3383               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3384        ) {
3385         if (    floatx80_is_signaling_nan( a )
3386              || floatx80_is_signaling_nan( b ) ) {
3387             float_raise( float_flag_invalid );
3388         }
3389         return 0;
3390     }
3391     aSign = extractFloatx80Sign( a );
3392     bSign = extractFloatx80Sign( b );
3393     if ( aSign != bSign ) {
3394         return
3395                aSign
3396             || (    ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low )
3397                  == 0 );
3398     }
3399     return
3400           aSign ? le128( b.high, b.low, a.high, a.low )
3401         : le128( a.high, a.low, b.high, b.low );
3402 
3403 }
3404 
3405 /*
3406 -------------------------------------------------------------------------------
3407 Returns 1 if the extended double-precision floating-point value `a' is less
3408 than the corresponding value `b', and 0 otherwise.  Quiet NaNs do not cause
3409 an exception.  Otherwise, the comparison is performed according to the
3410 IEC/IEEE Standard for Binary Floating-point Arithmetic.
3411 -------------------------------------------------------------------------------
3412 */
floatx80_lt_quiet(floatx80 a,floatx80 b)3413 flag floatx80_lt_quiet( floatx80 a, floatx80 b )
3414 {
3415     flag aSign, bSign;
3416 
3417     if (    (    ( extractFloatx80Exp( a ) == 0x7FFF )
3418               && (bits64) ( extractFloatx80Frac( a )<<1 ) )
3419          || (    ( extractFloatx80Exp( b ) == 0x7FFF )
3420               && (bits64) ( extractFloatx80Frac( b )<<1 ) )
3421        ) {
3422         if (    floatx80_is_signaling_nan( a )
3423              || floatx80_is_signaling_nan( b ) ) {
3424             float_raise( float_flag_invalid );
3425         }
3426         return 0;
3427     }
3428     aSign = extractFloatx80Sign( a );
3429     bSign = extractFloatx80Sign( b );
3430     if ( aSign != bSign ) {
3431         return
3432                aSign
3433             && (    ( ( (bits16) ( ( a.high | b.high )<<1 ) ) | a.low | b.low )
3434                  != 0 );
3435     }
3436     return
3437           aSign ? lt128( b.high, b.low, a.high, a.low )
3438         : lt128( a.high, a.low, b.high, b.low );
3439 
3440 }
3441 
3442 #endif
3443 
3444