xref: /freebsd/lib/msun/src/math_private.h (revision 3085fc9d97bd83785ba3ba43e0378d7d67987d1f)
1 /*
2  * ====================================================
3  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
4  *
5  * Developed at SunPro, a Sun Microsystems, Inc. business.
6  * Permission to use, copy, modify, and distribute this
7  * software is freely granted, provided that this notice
8  * is preserved.
9  * ====================================================
10  */
11 
12 /*
13  */
14 
15 #ifndef _MATH_PRIVATE_H_
16 #define	_MATH_PRIVATE_H_
17 
18 #include <sys/types.h>
19 #include <machine/endian.h>
20 
21 /*
22  * The original fdlibm code used statements like:
23  *	n0 = ((*(int*)&one)>>29)^1;		* index of high word *
24  *	ix0 = *(n0+(int*)&x);			* high word of x *
25  *	ix1 = *((1-n0)+(int*)&x);		* low word of x *
26  * to dig two 32 bit words out of the 64 bit IEEE floating point
27  * value.  That is non-ANSI, and, moreover, the gcc instruction
28  * scheduler gets it wrong.  We instead use the following macros.
29  * Unlike the original code, we determine the endianness at compile
30  * time, not at run time; I don't see much benefit to selecting
31  * endianness at run time.
32  */
33 
34 /*
35  * A union which permits us to convert between a double and two 32 bit
36  * ints.
37  */
38 
39 #ifdef __arm__
40 #if defined(__VFP_FP__) || defined(__ARM_EABI__)
41 #define	IEEE_WORD_ORDER	BYTE_ORDER
42 #else
43 #define	IEEE_WORD_ORDER	BIG_ENDIAN
44 #endif
45 #else /* __arm__ */
46 #define	IEEE_WORD_ORDER	BYTE_ORDER
47 #endif
48 
49 /* A union which permits us to convert between a long double and
50    four 32 bit ints.  */
51 
52 #if IEEE_WORD_ORDER == BIG_ENDIAN
53 
54 typedef union
55 {
56   long double value;
57   struct {
58     u_int32_t mswhi;
59     u_int32_t mswlo;
60     u_int32_t lswhi;
61     u_int32_t lswlo;
62   } parts32;
63   struct {
64     u_int64_t msw;
65     u_int64_t lsw;
66   } parts64;
67 } ieee_quad_shape_type;
68 
69 #endif
70 
71 #if IEEE_WORD_ORDER == LITTLE_ENDIAN
72 
73 typedef union
74 {
75   long double value;
76   struct {
77     u_int32_t lswlo;
78     u_int32_t lswhi;
79     u_int32_t mswlo;
80     u_int32_t mswhi;
81   } parts32;
82   struct {
83     u_int64_t lsw;
84     u_int64_t msw;
85   } parts64;
86 } ieee_quad_shape_type;
87 
88 #endif
89 
90 #if IEEE_WORD_ORDER == BIG_ENDIAN
91 
92 typedef union
93 {
94   double value;
95   struct
96   {
97     u_int32_t msw;
98     u_int32_t lsw;
99   } parts;
100   struct
101   {
102     u_int64_t w;
103   } xparts;
104 } ieee_double_shape_type;
105 
106 #endif
107 
108 #if IEEE_WORD_ORDER == LITTLE_ENDIAN
109 
110 typedef union
111 {
112   double value;
113   struct
114   {
115     u_int32_t lsw;
116     u_int32_t msw;
117   } parts;
118   struct
119   {
120     u_int64_t w;
121   } xparts;
122 } ieee_double_shape_type;
123 
124 #endif
125 
126 /* Get two 32 bit ints from a double.  */
127 
128 #define EXTRACT_WORDS(ix0,ix1,d)				\
129 do {								\
130   ieee_double_shape_type ew_u;					\
131   ew_u.value = (d);						\
132   (ix0) = ew_u.parts.msw;					\
133   (ix1) = ew_u.parts.lsw;					\
134 } while (0)
135 
136 /* Get a 64-bit int from a double. */
137 #define EXTRACT_WORD64(ix,d)					\
138 do {								\
139   ieee_double_shape_type ew_u;					\
140   ew_u.value = (d);						\
141   (ix) = ew_u.xparts.w;						\
142 } while (0)
143 
144 /* Get the more significant 32 bit int from a double.  */
145 
146 #define GET_HIGH_WORD(i,d)					\
147 do {								\
148   ieee_double_shape_type gh_u;					\
149   gh_u.value = (d);						\
150   (i) = gh_u.parts.msw;						\
151 } while (0)
152 
153 /* Get the less significant 32 bit int from a double.  */
154 
155 #define GET_LOW_WORD(i,d)					\
156 do {								\
157   ieee_double_shape_type gl_u;					\
158   gl_u.value = (d);						\
159   (i) = gl_u.parts.lsw;						\
160 } while (0)
161 
162 /* Set a double from two 32 bit ints.  */
163 
164 #define INSERT_WORDS(d,ix0,ix1)					\
165 do {								\
166   ieee_double_shape_type iw_u;					\
167   iw_u.parts.msw = (ix0);					\
168   iw_u.parts.lsw = (ix1);					\
169   (d) = iw_u.value;						\
170 } while (0)
171 
172 /* Set a double from a 64-bit int. */
173 #define INSERT_WORD64(d,ix)					\
174 do {								\
175   ieee_double_shape_type iw_u;					\
176   iw_u.xparts.w = (ix);						\
177   (d) = iw_u.value;						\
178 } while (0)
179 
180 /* Set the more significant 32 bits of a double from an int.  */
181 
182 #define SET_HIGH_WORD(d,v)					\
183 do {								\
184   ieee_double_shape_type sh_u;					\
185   sh_u.value = (d);						\
186   sh_u.parts.msw = (v);						\
187   (d) = sh_u.value;						\
188 } while (0)
189 
190 /* Set the less significant 32 bits of a double from an int.  */
191 
192 #define SET_LOW_WORD(d,v)					\
193 do {								\
194   ieee_double_shape_type sl_u;					\
195   sl_u.value = (d);						\
196   sl_u.parts.lsw = (v);						\
197   (d) = sl_u.value;						\
198 } while (0)
199 
200 /*
201  * A union which permits us to convert between a float and a 32 bit
202  * int.
203  */
204 
205 typedef union
206 {
207   float value;
208   /* FIXME: Assumes 32 bit int.  */
209   unsigned int word;
210 } ieee_float_shape_type;
211 
212 /* Get a 32 bit int from a float.  */
213 
214 #define GET_FLOAT_WORD(i,d)					\
215 do {								\
216   ieee_float_shape_type gf_u;					\
217   gf_u.value = (d);						\
218   (i) = gf_u.word;						\
219 } while (0)
220 
221 /* Set a float from a 32 bit int.  */
222 
223 #define SET_FLOAT_WORD(d,i)					\
224 do {								\
225   ieee_float_shape_type sf_u;					\
226   sf_u.word = (i);						\
227   (d) = sf_u.value;						\
228 } while (0)
229 
230 /*
231  * Get expsign and mantissa as 16 bit and 64 bit ints from an 80 bit long
232  * double.
233  */
234 
235 #define	EXTRACT_LDBL80_WORDS(ix0,ix1,d)				\
236 do {								\
237   union IEEEl2bits ew_u;					\
238   ew_u.e = (d);							\
239   (ix0) = ew_u.xbits.expsign;					\
240   (ix1) = ew_u.xbits.man;					\
241 } while (0)
242 
243 /*
244  * Get expsign and mantissa as one 16 bit and two 64 bit ints from a 128 bit
245  * long double.
246  */
247 
248 #define	EXTRACT_LDBL128_WORDS(ix0,ix1,ix2,d)			\
249 do {								\
250   union IEEEl2bits ew_u;					\
251   ew_u.e = (d);							\
252   (ix0) = ew_u.xbits.expsign;					\
253   (ix1) = ew_u.xbits.manh;					\
254   (ix2) = ew_u.xbits.manl;					\
255 } while (0)
256 
257 /* Get expsign as a 16 bit int from a long double.  */
258 
259 #define	GET_LDBL_EXPSIGN(i,d)					\
260 do {								\
261   union IEEEl2bits ge_u;					\
262   ge_u.e = (d);							\
263   (i) = ge_u.xbits.expsign;					\
264 } while (0)
265 
266 /*
267  * Set an 80 bit long double from a 16 bit int expsign and a 64 bit int
268  * mantissa.
269  */
270 
271 #define	INSERT_LDBL80_WORDS(d,ix0,ix1)				\
272 do {								\
273   union IEEEl2bits iw_u;					\
274   iw_u.xbits.expsign = (ix0);					\
275   iw_u.xbits.man = (ix1);					\
276   (d) = iw_u.e;							\
277 } while (0)
278 
279 /*
280  * Set a 128 bit long double from a 16 bit int expsign and two 64 bit ints
281  * comprising the mantissa.
282  */
283 
284 #define	INSERT_LDBL128_WORDS(d,ix0,ix1,ix2)			\
285 do {								\
286   union IEEEl2bits iw_u;					\
287   iw_u.xbits.expsign = (ix0);					\
288   iw_u.xbits.manh = (ix1);					\
289   iw_u.xbits.manl = (ix2);					\
290   (d) = iw_u.e;							\
291 } while (0)
292 
293 /* Set expsign of a long double from a 16 bit int.  */
294 
295 #define	SET_LDBL_EXPSIGN(d,v)					\
296 do {								\
297   union IEEEl2bits se_u;					\
298   se_u.e = (d);							\
299   se_u.xbits.expsign = (v);					\
300   (d) = se_u.e;							\
301 } while (0)
302 
303 #ifdef __i386__
304 /* Long double constants are broken on i386. */
305 #define	LD80C(m, ex, v) {						\
306 	.xbits.man = __CONCAT(m, ULL),					\
307 	.xbits.expsign = (0x3fff + (ex)) | ((v) < 0 ? 0x8000 : 0),	\
308 }
309 #else
310 /* The above works on non-i386 too, but we use this to check v. */
311 #define	LD80C(m, ex, v)	{ .e = (v), }
312 #endif
313 
314 #ifdef FLT_EVAL_METHOD
315 /*
316  * Attempt to get strict C99 semantics for assignment with non-C99 compilers.
317  */
318 #if FLT_EVAL_METHOD == 0 || __GNUC__ == 0
319 #define	STRICT_ASSIGN(type, lval, rval)	((lval) = (rval))
320 #else
321 #define	STRICT_ASSIGN(type, lval, rval) do {	\
322 	volatile type __lval;			\
323 						\
324 	if (sizeof(type) >= sizeof(long double))	\
325 		(lval) = (rval);		\
326 	else {					\
327 		__lval = (rval);		\
328 		(lval) = __lval;		\
329 	}					\
330 } while (0)
331 #endif
332 #endif /* FLT_EVAL_METHOD */
333 
334 /* Support switching the mode to FP_PE if necessary. */
335 #if defined(__i386__) && !defined(NO_FPSETPREC)
336 #define	ENTERI() ENTERIT(long double)
337 #define	ENTERIT(returntype)			\
338 	returntype __retval;			\
339 	fp_prec_t __oprec;			\
340 						\
341 	if ((__oprec = fpgetprec()) != FP_PE)	\
342 		fpsetprec(FP_PE)
343 #define	RETURNI(x) do {				\
344 	__retval = (x);				\
345 	if (__oprec != FP_PE)			\
346 		fpsetprec(__oprec);		\
347 	RETURNF(__retval);			\
348 } while (0)
349 #define	ENTERV()				\
350 	fp_prec_t __oprec;			\
351 						\
352 	if ((__oprec = fpgetprec()) != FP_PE)	\
353 		fpsetprec(FP_PE)
354 #define	RETURNV() do {				\
355 	if (__oprec != FP_PE)			\
356 		fpsetprec(__oprec);		\
357 	return;			\
358 } while (0)
359 #else
360 #define	ENTERI()
361 #define	ENTERIT(x)
362 #define	RETURNI(x)	RETURNF(x)
363 #define	ENTERV()
364 #define	RETURNV()	return
365 #endif
366 
367 /* Default return statement if hack*_t() is not used. */
368 #define      RETURNF(v)      return (v)
369 
370 /*
371  * 2sum gives the same result as 2sumF without requiring |a| >= |b| or
372  * a == 0, but is slower.
373  */
374 #define	_2sum(a, b) do {	\
375 	__typeof(a) __s, __w;	\
376 				\
377 	__w = (a) + (b);	\
378 	__s = __w - (a);	\
379 	(b) = ((a) - (__w - __s)) + ((b) - __s); \
380 	(a) = __w;		\
381 } while (0)
382 
383 /*
384  * 2sumF algorithm.
385  *
386  * "Normalize" the terms in the infinite-precision expression a + b for
387  * the sum of 2 floating point values so that b is as small as possible
388  * relative to 'a'.  (The resulting 'a' is the value of the expression in
389  * the same precision as 'a' and the resulting b is the rounding error.)
390  * |a| must be >= |b| or 0, b's type must be no larger than 'a's type, and
391  * exponent overflow or underflow must not occur.  This uses a Theorem of
392  * Dekker (1971).  See Knuth (1981) 4.2.2 Theorem C.  The name "TwoSum"
393  * is apparently due to Skewchuk (1997).
394  *
395  * For this to always work, assignment of a + b to 'a' must not retain any
396  * extra precision in a + b.  This is required by C standards but broken
397  * in many compilers.  The brokenness cannot be worked around using
398  * STRICT_ASSIGN() like we do elsewhere, since the efficiency of this
399  * algorithm would be destroyed by non-null strict assignments.  (The
400  * compilers are correct to be broken -- the efficiency of all floating
401  * point code calculations would be destroyed similarly if they forced the
402  * conversions.)
403  *
404  * Fortunately, a case that works well can usually be arranged by building
405  * any extra precision into the type of 'a' -- 'a' should have type float_t,
406  * double_t or long double.  b's type should be no larger than 'a's type.
407  * Callers should use these types with scopes as large as possible, to
408  * reduce their own extra-precision and efficiency problems.  In
409  * particular, they shouldn't convert back and forth just to call here.
410  */
411 #ifdef DEBUG
412 #define	_2sumF(a, b) do {				\
413 	__typeof(a) __w;				\
414 	volatile __typeof(a) __ia, __ib, __r, __vw;	\
415 							\
416 	__ia = (a);					\
417 	__ib = (b);					\
418 	assert(__ia == 0 || fabsl(__ia) >= fabsl(__ib));	\
419 							\
420 	__w = (a) + (b);				\
421 	(b) = ((a) - __w) + (b);			\
422 	(a) = __w;					\
423 							\
424 	/* The next 2 assertions are weak if (a) is already long double. */ \
425 	assert((long double)__ia + __ib == (long double)(a) + (b));	\
426 	__vw = __ia + __ib;				\
427 	__r = __ia - __vw;				\
428 	__r += __ib;					\
429 	assert(__vw == (a) && __r == (b));		\
430 } while (0)
431 #else /* !DEBUG */
432 #define	_2sumF(a, b) do {	\
433 	__typeof(a) __w;	\
434 				\
435 	__w = (a) + (b);	\
436 	(b) = ((a) - __w) + (b); \
437 	(a) = __w;		\
438 } while (0)
439 #endif /* DEBUG */
440 
441 /*
442  * Set x += c, where x is represented in extra precision as a + b.
443  * x must be sufficiently normalized and sufficiently larger than c,
444  * and the result is then sufficiently normalized.
445  *
446  * The details of ordering are that |a| must be >= |c| (so that (a, c)
447  * can be normalized without extra work to swap 'a' with c).  The details of
448  * the normalization are that b must be small relative to the normalized 'a'.
449  * Normalization of (a, c) makes the normalized c tiny relative to the
450  * normalized a, so b remains small relative to 'a' in the result.  However,
451  * b need not ever be tiny relative to 'a'.  For example, b might be about
452  * 2**20 times smaller than 'a' to give about 20 extra bits of precision.
453  * That is usually enough, and adding c (which by normalization is about
454  * 2**53 times smaller than a) cannot change b significantly.  However,
455  * cancellation of 'a' with c in normalization of (a, c) may reduce 'a'
456  * significantly relative to b.  The caller must ensure that significant
457  * cancellation doesn't occur, either by having c of the same sign as 'a',
458  * or by having |c| a few percent smaller than |a|.  Pre-normalization of
459  * (a, b) may help.
460  *
461  * This is a variant of an algorithm of Kahan (see Knuth (1981) 4.2.2
462  * exercise 19).  We gain considerable efficiency by requiring the terms to
463  * be sufficiently normalized and sufficiently increasing.
464  */
465 #define	_3sumF(a, b, c) do {	\
466 	__typeof(a) __tmp;	\
467 				\
468 	__tmp = (c);		\
469 	_2sumF(__tmp, (a));	\
470 	(b) += (a);		\
471 	(a) = __tmp;		\
472 } while (0)
473 
474 /*
475  * Split x into high and low bits where CC is 0x1p(N/2) + 1 where
476  * N is rounded up for types with odd precisions.
477  *
478  * #define _CC (0x1p12F + 1)	// float
479  * #define _CC (0x1p27  + 1)	// double
480  * #define _CC (0x1p32L + 1)	// long double (LD80)
481  * #define _CC (0x1p57L + 1)	// long double (LD128)
482  */
483 #define _SPLIT(x, xh, xl)		\
484 do {					\
485 	typeof(x) __t1;			\
486 	__t1 = (x) * _CC;		\
487 	xh = __t1 + ((x) - __t1);	\
488 	xl = (x) - xh;			\
489 } while(0)
490 
491 /*
492  * FAST2SUM requires |x| >= |y|.  x and y are full precision.
493  * Note, _2SUMF(x,y) above destroys x and y.
494  */
495 #define _FAST2SUM(x, y, hi, lo)	\
496 do {				\
497 	hi = (x) + (y);		\
498 	lo = (y) - (hi - (x));	\
499 } while(0)
500 
501 /*
502  * SLOW2SUM does not require |x| >= |y|.  Here, x and y are full precision.
503  * The t1 temporary variable is volatile to prevent compiler optimizations.
504  * Note, _2SUM(x,y) above destroys x and y.
505  */
506 #define _SLOW2SUM(x, y, hi, lo)			\
507 do {						\
508 	volatile typeof(x) __t1;		\
509 	typeof(x) __t2;				\
510 	hi = (x) + (y);				\
511 	__t1 = hi - (y);			\
512 	__t2 = hi - __t1;			\
513 	lo = ((x) - __t1) + ((y) - __t2);	\
514 } while(0)
515 
516 /*
517  * x and y are full precision quantities that have been split into high
518  * and low parts via the _SPLIT macro.  x and y are added to give z as
519  * high and low parts.
520  */
521 #define _XADD(xh, xl, yh, yl, zh, zl)			\
522 do {							\
523 	typeof(xh) __s1, __s2, __s3, __s4, __s5, __s6;	\
524 	_SLOW2SUM(xh, yh, __s1, __s2);			\
525 	_SLOW2SUM(xl, yl, __s3, __s4);			\
526 	_FAST2SUM(__s1, __s2 + __s3, __s5, __s6);	\
527 	_FAST2SUM(__s5, __s6 + __s4, zh, zl);		\
528 } while(0)
529 
530 /*
531  * x and y are full precision quantities.  r1 and r2 are full precision
532  * high and low parts of the multiplication x * y.
533  */
534 #define _MUL(x, y, r1 ,r2)				\
535 do  {							\
536 	typeof(x) __xh, __xl, __yh, __yl;		\
537 	typeof(x) __t1;				\
538 	_SPLIT(x, __xh, __xl);				\
539 	_SPLIT(y, __yh, __yl);				\
540 	r1 = (x) * (y);					\
541 	__t1 = __xh * __yl + (__xh * __yh - r1);	\
542 	r2 = __xl * __yl + (__xl * __yh + __t1);	\
543 } while(0)
544 
545 /*
546  * x and y are full precision quantities that have been split into high
547  * and low parts via the _SPLIT macro.  x and y are multiplied to give z
548  * as high and low parts.
549  */
550 #define _XMUL(xh, xl, yh, yl, ph, pl)	\
551 do {					\
552     _MUL(xh, yh, ph, pl);		\
553     pl += xl * yl + xl * yh + xh * yl;	\
554 } while(0)
555 
556 
557 /*
558  * Common routine to process the arguments to nan(), nanf(), and nanl().
559  */
560 void _scan_nan(uint32_t *__words, int __num_words, const char *__s);
561 
562 /*
563  * Mix 0, 1 or 2 NaNs.  First add 0 to each arg.  This normally just turns
564  * signaling NaNs into quiet NaNs by setting a quiet bit.  We do this
565  * because we want to never return a signaling NaN, and also because we
566  * don't want the quiet bit to affect the result.  Then mix the converted
567  * args using the specified operation.
568  *
569  * When one arg is NaN, the result is typically that arg quieted.  When both
570  * args are NaNs, the result is typically the quietening of the arg whose
571  * mantissa is largest after quietening.  When neither arg is NaN, the
572  * result may be NaN because it is indeterminate, or finite for subsequent
573  * construction of a NaN as the indeterminate 0.0L/0.0L.
574  *
575  * Technical complications: the result in bits after rounding to the final
576  * precision might depend on the runtime precision and/or on compiler
577  * optimizations, especially when different register sets are used for
578  * different precisions.  Try to make the result not depend on at least the
579  * runtime precision by always doing the main mixing step in long double
580  * precision.  Try to reduce dependencies on optimizations by adding the
581  * the 0's in different precisions (unless everything is in long double
582  * precision).
583  */
584 #define	nan_mix(x, y)		(nan_mix_op((x), (y), +))
585 #define	nan_mix_op(x, y, op)	(((x) + 0.0L) op ((y) + 0))
586 
587 #ifdef _COMPLEX_H
588 
589 /*
590  * C99 specifies that complex numbers have the same representation as
591  * an array of two elements, where the first element is the real part
592  * and the second element is the imaginary part.
593  */
594 typedef union {
595 	float complex f;
596 	float a[2];
597 } float_complex;
598 typedef union {
599 	double complex f;
600 	double a[2];
601 } double_complex;
602 typedef union {
603 	long double complex f;
604 	long double a[2];
605 } long_double_complex;
606 #define	REALPART(z)	((z).a[0])
607 #define	IMAGPART(z)	((z).a[1])
608 
609 /*
610  * Inline functions that can be used to construct complex values.
611  *
612  * The C99 standard intends x+I*y to be used for this, but x+I*y is
613  * currently unusable in general since gcc introduces many overflow,
614  * underflow, sign and efficiency bugs by rewriting I*y as
615  * (0.0+I)*(y+0.0*I) and laboriously computing the full complex product.
616  * In particular, I*Inf is corrupted to NaN+I*Inf, and I*-0 is corrupted
617  * to -0.0+I*0.0.
618  *
619  * The C11 standard introduced the macros CMPLX(), CMPLXF() and CMPLXL()
620  * to construct complex values.  Compilers that conform to the C99
621  * standard require the following functions to avoid the above issues.
622  */
623 
624 #ifndef CMPLXF
625 static __inline float complex
CMPLXF(float x,float y)626 CMPLXF(float x, float y)
627 {
628 	float_complex z;
629 
630 	REALPART(z) = x;
631 	IMAGPART(z) = y;
632 	return (z.f);
633 }
634 #endif
635 
636 #ifndef CMPLX
637 static __inline double complex
CMPLX(double x,double y)638 CMPLX(double x, double y)
639 {
640 	double_complex z;
641 
642 	REALPART(z) = x;
643 	IMAGPART(z) = y;
644 	return (z.f);
645 }
646 #endif
647 
648 #ifndef CMPLXL
649 static __inline long double complex
CMPLXL(long double x,long double y)650 CMPLXL(long double x, long double y)
651 {
652 	long_double_complex z;
653 
654 	REALPART(z) = x;
655 	IMAGPART(z) = y;
656 	return (z.f);
657 }
658 #endif
659 
660 #endif /* _COMPLEX_H */
661 
662 /*
663  * The rnint() family rounds to the nearest integer for a restricted range
664  * range of args (up to about 2**MANT_DIG).  We assume that the current
665  * rounding mode is FE_TONEAREST so that this can be done efficiently.
666  * Extra precision causes more problems in practice, and we only centralize
667  * this here to reduce those problems, and have not solved the efficiency
668  * problems.  The exp2() family uses a more delicate version of this that
669  * requires extracting bits from the intermediate value, so it is not
670  * centralized here and should copy any solution of the efficiency problems.
671  */
672 
673 static inline double
rnint(__double_t x)674 rnint(__double_t x)
675 {
676 	/*
677 	 * This casts to double to kill any extra precision.  This depends
678 	 * on the cast being applied to a double_t to avoid compiler bugs
679 	 * (this is a cleaner version of STRICT_ASSIGN()).  This is
680 	 * inefficient if there actually is extra precision, but is hard
681 	 * to improve on.  We use double_t in the API to minimise conversions
682 	 * for just calling here.  Note that we cannot easily change the
683 	 * magic number to the one that works directly with double_t, since
684 	 * the rounding precision is variable at runtime on x86 so the
685 	 * magic number would need to be variable.  Assuming that the
686 	 * rounding precision is always the default is too fragile.  This
687 	 * and many other complications will move when the default is
688 	 * changed to FP_PE.
689 	 */
690 	return ((double)(x + 0x1.8p52) - 0x1.8p52);
691 }
692 
693 static inline float
rnintf(__float_t x)694 rnintf(__float_t x)
695 {
696 	/*
697 	 * As for rnint(), except we could just call that to handle the
698 	 * extra precision case, usually without losing efficiency.
699 	 */
700 	return ((float)(x + 0x1.8p23F) - 0x1.8p23F);
701 }
702 
703 #ifdef LDBL_MANT_DIG
704 /*
705  * The complications for extra precision are smaller for rnintl() since it
706  * can safely assume that the rounding precision has been increased from
707  * its default to FP_PE on x86.  We don't exploit that here to get small
708  * optimizations from limiting the range to double.  We just need it for
709  * the magic number to work with long doubles.  ld128 callers should use
710  * rnint() instead of this if possible.  ld80 callers should prefer
711  * rnintl() since for amd64 this avoids swapping the register set, while
712  * for i386 it makes no difference (assuming FP_PE), and for other arches
713  * it makes little difference.
714  */
715 static inline long double
rnintl(long double x)716 rnintl(long double x)
717 {
718 	return (x + __CONCAT(0x1.8p, LDBL_MANT_DIG) / 2 -
719 	    __CONCAT(0x1.8p, LDBL_MANT_DIG) / 2);
720 }
721 #endif /* LDBL_MANT_DIG */
722 
723 /*
724  * irint() and i64rint() give the same result as casting to their integer
725  * return type provided their arg is a floating point integer.  They can
726  * sometimes be more efficient because no rounding is required.
727  */
728 #if defined(amd64) || defined(__i386__)
729 #define	irint(x)						\
730     (sizeof(x) == sizeof(float) &&				\
731     sizeof(__float_t) == sizeof(long double) ? irintf(x) :	\
732     sizeof(x) == sizeof(double) &&				\
733     sizeof(__double_t) == sizeof(long double) ? irintd(x) :	\
734     sizeof(x) == sizeof(long double) ? irintl(x) : (int)(x))
735 #else
736 #define	irint(x)	((int)(x))
737 #endif
738 
739 #define	i64rint(x)	((int64_t)(x))	/* only needed for ld128 so not opt. */
740 
741 #if defined(__i386__)
742 static __inline int
irintf(float x)743 irintf(float x)
744 {
745 	int n;
746 
747 	__asm("fistl %0" : "=m" (n) : "t" (x));
748 	return (n);
749 }
750 
751 static __inline int
irintd(double x)752 irintd(double x)
753 {
754 	int n;
755 
756 	__asm("fistl %0" : "=m" (n) : "t" (x));
757 	return (n);
758 }
759 #endif
760 
761 #if defined(__amd64__) || defined(__i386__)
762 static __inline int
irintl(long double x)763 irintl(long double x)
764 {
765 	int n;
766 
767 	__asm("fistl %0" : "=m" (n) : "t" (x));
768 	return (n);
769 }
770 #endif
771 
772 /*
773  * The following are fast floor macros for 0 <= |x| < 0x1p(N-1), where
774  * N is the precision of the type of x. These macros are used in the
775  * half-cycle trignometric functions (e.g., sinpi(x)).
776  */
777 #define	FFLOORF(x, j0, ix) do {			\
778 	(j0) = (((ix) >> 23) & 0xff) - 0x7f;	\
779 	(ix) &= ~(0x007fffff >> (j0));		\
780 	SET_FLOAT_WORD((x), (ix));		\
781 } while (0)
782 
783 #define	FFLOOR(x, j0, ix, lx) do {				\
784 	(j0) = (((ix) >> 20) & 0x7ff) - 0x3ff;			\
785 	if ((j0) < 20) {					\
786 		(ix) &= ~(0x000fffff >> (j0));			\
787 		(lx) = 0;					\
788 	} else {						\
789 		(lx) &= ~((uint32_t)0xffffffff >> ((j0) - 20));	\
790 	}							\
791 	INSERT_WORDS((x), (ix), (lx));				\
792 } while (0)
793 
794 #define	FFLOORL80(x, j0, ix, lx) do {			\
795 	j0 = ix - 0x3fff + 1;				\
796 	if ((j0) < 32) {				\
797 		(lx) = ((lx) >> 32) << 32;		\
798 		(lx) &= ~((((lx) << 32)-1) >> (j0));	\
799 	} else {					\
800 		uint64_t _m;				\
801 		_m = (uint64_t)-1 >> (j0);		\
802 		if ((lx) & _m) (lx) &= ~_m;		\
803 	}						\
804 	INSERT_LDBL80_WORDS((x), (ix), (lx));		\
805 } while (0)
806 
807 #define FFLOORL128(x, ai, ar) do {			\
808 	union IEEEl2bits u;				\
809 	uint64_t m;					\
810 	int e;						\
811 	u.e = (x);					\
812 	e = u.bits.exp - 16383;				\
813 	if (e < 48) {					\
814 		m = ((1llu << 49) - 1) >> (e + 1);	\
815 		u.bits.manh &= ~m;			\
816 		u.bits.manl = 0;			\
817 	} else {					\
818 		m = (uint64_t)-1 >> (e - 48);		\
819 		u.bits.manl &= ~m;			\
820 	}						\
821 	(ai) = u.e;					\
822 	(ar) = (x) - (ai);				\
823 } while (0)
824 
825 /*
826  * For a subnormal double entity split into high and low parts, compute ilogb.
827  */
828 static inline int32_t
subnormal_ilogb(int32_t hi,int32_t lo)829 subnormal_ilogb(int32_t hi, int32_t lo)
830 {
831 	int32_t j;
832 	uint32_t i;
833 
834 	j = -1022;
835 	if (hi == 0) {
836 	    j -= 21;
837 	    i = (uint32_t)lo;
838 	} else
839 	    i = (uint32_t)hi << 11;
840 
841 	for (; i < 0x7fffffff; i <<= 1) j -= 1;
842 
843 	return (j);
844 }
845 
846 /*
847  * For a subnormal float entity represented as an int32_t, compute ilogb.
848  */
849 static inline int32_t
subnormal_ilogbf(int32_t hx)850 subnormal_ilogbf(int32_t hx)
851 {
852 	int32_t j;
853 	uint32_t i;
854 	i = (uint32_t) hx << 8;
855 	for (j = -126; i < 0x7fffffff; i <<= 1) j -=1;
856 
857 	return (j);
858 }
859 
860 #ifdef DEBUG
861 #if defined(__amd64__) || defined(__i386__)
862 #define	breakpoint()	asm("int $3")
863 #else
864 #include <signal.h>
865 
866 #define	breakpoint()	raise(SIGTRAP)
867 #endif
868 #endif
869 
870 #ifdef STRUCT_RETURN
871 #define	RETURNSP(rp) do {		\
872 	if (!(rp)->lo_set)		\
873 		RETURNF((rp)->hi);	\
874 	RETURNF((rp)->hi + (rp)->lo);	\
875 } while (0)
876 #define	RETURNSPI(rp) do {		\
877 	if (!(rp)->lo_set)		\
878 		RETURNI((rp)->hi);	\
879 	RETURNI((rp)->hi + (rp)->lo);	\
880 } while (0)
881 #endif
882 
883 #define	SUM2P(x, y) ({			\
884 	const __typeof (x) __x = (x);	\
885 	const __typeof (y) __y = (y);	\
886 	__x + __y;			\
887 })
888 
889 /* fdlibm kernel function */
890 int	__kernel_rem_pio2(double*,double*,int,int,int);
891 
892 /* double precision kernel functions */
893 #ifndef INLINE_REM_PIO2
894 int	__ieee754_rem_pio2(double,double*);
895 #endif
896 double	__kernel_sin(double,double,int);
897 double	__kernel_cos(double,double);
898 double	__kernel_tan(double,double,int);
899 double	__ldexp_exp(double,int);
900 #ifdef _COMPLEX_H
901 double complex __ldexp_cexp(double complex,int);
902 #endif
903 
904 /* float precision kernel functions */
905 #ifndef INLINE_REM_PIO2F
906 int	__ieee754_rem_pio2f(float,double*);
907 #endif
908 #ifndef INLINE_KERNEL_SINDF
909 float	__kernel_sindf(double);
910 #endif
911 #ifndef INLINE_KERNEL_COSDF
912 float	__kernel_cosdf(double);
913 #endif
914 #ifndef INLINE_KERNEL_TANDF
915 float	__kernel_tandf(double,int);
916 #endif
917 float	__ldexp_expf(float,int);
918 #ifdef _COMPLEX_H
919 float complex __ldexp_cexpf(float complex,int);
920 #endif
921 
922 /* long double precision kernel functions */
923 long double __kernel_sinl(long double, long double, int);
924 long double __kernel_cosl(long double, long double);
925 long double __kernel_tanl(long double, long double, int);
926 
927 #endif /* !_MATH_PRIVATE_H_ */
928