2 * ====================================================
3 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
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
9 * ====================================================
13 * from: @(#)fdlibm.h 5.1 93/09/24
17 #ifndef _MATH_PRIVATE_H_
18 #define _MATH_PRIVATE_H_
20 #include <sys/types.h>
21 #include <machine/endian.h>
24 * The original fdlibm code used statements like:
25 * n0 = ((*(int*)&one)>>29)^1; * index of high word *
26 * ix0 = *(n0+(int*)&x); * high word of x *
27 * ix1 = *((1-n0)+(int*)&x); * low word of x *
28 * to dig two 32 bit words out of the 64 bit IEEE floating point
29 * value. That is non-ANSI, and, moreover, the gcc instruction
30 * scheduler gets it wrong. We instead use the following macros.
31 * Unlike the original code, we determine the endianness at compile
32 * time, not at run time; I don't see much benefit to selecting
33 * endianness at run time.
37 * A union which permits us to convert between a double and two 32 bit
42 #if defined(__VFP_FP__) || defined(__ARM_EABI__)
43 #define IEEE_WORD_ORDER BYTE_ORDER
45 #define IEEE_WORD_ORDER BIG_ENDIAN
48 #define IEEE_WORD_ORDER BYTE_ORDER
51 #if IEEE_WORD_ORDER == BIG_ENDIAN
65 } ieee_double_shape_type;
69 #if IEEE_WORD_ORDER == LITTLE_ENDIAN
83 } ieee_double_shape_type;
87 /* Get two 32 bit ints from a double. */
89 #define EXTRACT_WORDS(ix0,ix1,d) \
91 ieee_double_shape_type ew_u; \
93 (ix0) = ew_u.parts.msw; \
94 (ix1) = ew_u.parts.lsw; \
97 /* Get a 64-bit int from a double. */
98 #define EXTRACT_WORD64(ix,d) \
100 ieee_double_shape_type ew_u; \
102 (ix) = ew_u.xparts.w; \
105 /* Get the more significant 32 bit int from a double. */
107 #define GET_HIGH_WORD(i,d) \
109 ieee_double_shape_type gh_u; \
111 (i) = gh_u.parts.msw; \
114 /* Get the less significant 32 bit int from a double. */
116 #define GET_LOW_WORD(i,d) \
118 ieee_double_shape_type gl_u; \
120 (i) = gl_u.parts.lsw; \
123 /* Set a double from two 32 bit ints. */
125 #define INSERT_WORDS(d,ix0,ix1) \
127 ieee_double_shape_type iw_u; \
128 iw_u.parts.msw = (ix0); \
129 iw_u.parts.lsw = (ix1); \
133 /* Set a double from a 64-bit int. */
134 #define INSERT_WORD64(d,ix) \
136 ieee_double_shape_type iw_u; \
137 iw_u.xparts.w = (ix); \
141 /* Set the more significant 32 bits of a double from an int. */
143 #define SET_HIGH_WORD(d,v) \
145 ieee_double_shape_type sh_u; \
147 sh_u.parts.msw = (v); \
151 /* Set the less significant 32 bits of a double from an int. */
153 #define SET_LOW_WORD(d,v) \
155 ieee_double_shape_type sl_u; \
157 sl_u.parts.lsw = (v); \
162 * A union which permits us to convert between a float and a 32 bit
169 /* FIXME: Assumes 32 bit int. */
171 } ieee_float_shape_type;
173 /* Get a 32 bit int from a float. */
175 #define GET_FLOAT_WORD(i,d) \
177 ieee_float_shape_type gf_u; \
182 /* Set a float from a 32 bit int. */
184 #define SET_FLOAT_WORD(d,i) \
186 ieee_float_shape_type sf_u; \
192 * Get expsign and mantissa as 16 bit and 64 bit ints from an 80 bit long
196 #define EXTRACT_LDBL80_WORDS(ix0,ix1,d) \
198 union IEEEl2bits ew_u; \
200 (ix0) = ew_u.xbits.expsign; \
201 (ix1) = ew_u.xbits.man; \
205 * Get expsign and mantissa as one 16 bit and two 64 bit ints from a 128 bit
209 #define EXTRACT_LDBL128_WORDS(ix0,ix1,ix2,d) \
211 union IEEEl2bits ew_u; \
213 (ix0) = ew_u.xbits.expsign; \
214 (ix1) = ew_u.xbits.manh; \
215 (ix2) = ew_u.xbits.manl; \
218 /* Get expsign as a 16 bit int from a long double. */
220 #define GET_LDBL_EXPSIGN(i,d) \
222 union IEEEl2bits ge_u; \
224 (i) = ge_u.xbits.expsign; \
228 * Set an 80 bit long double from a 16 bit int expsign and a 64 bit int
232 #define INSERT_LDBL80_WORDS(d,ix0,ix1) \
234 union IEEEl2bits iw_u; \
235 iw_u.xbits.expsign = (ix0); \
236 iw_u.xbits.man = (ix1); \
241 * Set a 128 bit long double from a 16 bit int expsign and two 64 bit ints
242 * comprising the mantissa.
245 #define INSERT_LDBL128_WORDS(d,ix0,ix1,ix2) \
247 union IEEEl2bits iw_u; \
248 iw_u.xbits.expsign = (ix0); \
249 iw_u.xbits.manh = (ix1); \
250 iw_u.xbits.manl = (ix2); \
254 /* Set expsign of a long double from a 16 bit int. */
256 #define SET_LDBL_EXPSIGN(d,v) \
258 union IEEEl2bits se_u; \
260 se_u.xbits.expsign = (v); \
265 /* Long double constants are broken on i386. */
266 #define LD80C(m, ex, v) { \
267 .xbits.man = __CONCAT(m, ULL), \
268 .xbits.expsign = (0x3fff + (ex)) | ((v) < 0 ? 0x8000 : 0), \
271 /* The above works on non-i386 too, but we use this to check v. */
272 #define LD80C(m, ex, v) { .e = (v), }
275 #ifdef FLT_EVAL_METHOD
277 * Attempt to get strict C99 semantics for assignment with non-C99 compilers.
279 #if FLT_EVAL_METHOD == 0 || __GNUC__ == 0
280 #define STRICT_ASSIGN(type, lval, rval) ((lval) = (rval))
282 #define STRICT_ASSIGN(type, lval, rval) do { \
283 volatile type __lval; \
285 if (sizeof(type) >= sizeof(long double)) \
293 #endif /* FLT_EVAL_METHOD */
295 /* Support switching the mode to FP_PE if necessary. */
296 #if defined(__i386__) && !defined(NO_FPSETPREC)
297 #define ENTERI() ENTERIT(long double)
298 #define ENTERIT(returntype) \
299 returntype __retval; \
302 if ((__oprec = fpgetprec()) != FP_PE) \
304 #define RETURNI(x) do { \
306 if (__oprec != FP_PE) \
307 fpsetprec(__oprec); \
313 if ((__oprec = fpgetprec()) != FP_PE) \
315 #define RETURNV() do { \
316 if (__oprec != FP_PE) \
317 fpsetprec(__oprec); \
323 #define RETURNI(x) RETURNF(x)
325 #define RETURNV() return
328 /* Default return statement if hack*_t() is not used. */
329 #define RETURNF(v) return (v)
332 * 2sum gives the same result as 2sumF without requiring |a| >= |b| or
333 * a == 0, but is slower.
335 #define _2sum(a, b) do { \
336 __typeof(a) __s, __w; \
340 (b) = ((a) - (__w - __s)) + ((b) - __s); \
347 * "Normalize" the terms in the infinite-precision expression a + b for
348 * the sum of 2 floating point values so that b is as small as possible
349 * relative to 'a'. (The resulting 'a' is the value of the expression in
350 * the same precision as 'a' and the resulting b is the rounding error.)
351 * |a| must be >= |b| or 0, b's type must be no larger than 'a's type, and
352 * exponent overflow or underflow must not occur. This uses a Theorem of
353 * Dekker (1971). See Knuth (1981) 4.2.2 Theorem C. The name "TwoSum"
354 * is apparently due to Skewchuk (1997).
356 * For this to always work, assignment of a + b to 'a' must not retain any
357 * extra precision in a + b. This is required by C standards but broken
358 * in many compilers. The brokenness cannot be worked around using
359 * STRICT_ASSIGN() like we do elsewhere, since the efficiency of this
360 * algorithm would be destroyed by non-null strict assignments. (The
361 * compilers are correct to be broken -- the efficiency of all floating
362 * point code calculations would be destroyed similarly if they forced the
365 * Fortunately, a case that works well can usually be arranged by building
366 * any extra precision into the type of 'a' -- 'a' should have type float_t,
367 * double_t or long double. b's type should be no larger than 'a's type.
368 * Callers should use these types with scopes as large as possible, to
369 * reduce their own extra-precision and efficiciency problems. In
370 * particular, they shouldn't convert back and forth just to call here.
373 #define _2sumF(a, b) do { \
375 volatile __typeof(a) __ia, __ib, __r, __vw; \
379 assert(__ia == 0 || fabsl(__ia) >= fabsl(__ib)); \
382 (b) = ((a) - __w) + (b); \
385 /* The next 2 assertions are weak if (a) is already long double. */ \
386 assert((long double)__ia + __ib == (long double)(a) + (b)); \
387 __vw = __ia + __ib; \
390 assert(__vw == (a) && __r == (b)); \
393 #define _2sumF(a, b) do { \
397 (b) = ((a) - __w) + (b); \
403 * Set x += c, where x is represented in extra precision as a + b.
404 * x must be sufficiently normalized and sufficiently larger than c,
405 * and the result is then sufficiently normalized.
407 * The details of ordering are that |a| must be >= |c| (so that (a, c)
408 * can be normalized without extra work to swap 'a' with c). The details of
409 * the normalization are that b must be small relative to the normalized 'a'.
410 * Normalization of (a, c) makes the normalized c tiny relative to the
411 * normalized a, so b remains small relative to 'a' in the result. However,
412 * b need not ever be tiny relative to 'a'. For example, b might be about
413 * 2**20 times smaller than 'a' to give about 20 extra bits of precision.
414 * That is usually enough, and adding c (which by normalization is about
415 * 2**53 times smaller than a) cannot change b significantly. However,
416 * cancellation of 'a' with c in normalization of (a, c) may reduce 'a'
417 * significantly relative to b. The caller must ensure that significant
418 * cancellation doesn't occur, either by having c of the same sign as 'a',
419 * or by having |c| a few percent smaller than |a|. Pre-normalization of
422 * This is is a variant of an algorithm of Kahan (see Knuth (1981) 4.2.2
423 * exercise 19). We gain considerable efficiency by requiring the terms to
424 * be sufficiently normalized and sufficiently increasing.
426 #define _3sumF(a, b, c) do { \
430 _2sumF(__tmp, (a)); \
436 * Common routine to process the arguments to nan(), nanf(), and nanl().
438 void _scan_nan(uint32_t *__words, int __num_words, const char *__s);
443 * C99 specifies that complex numbers have the same representation as
444 * an array of two elements, where the first element is the real part
445 * and the second element is the imaginary part.
456 long double complex f;
458 } long_double_complex;
459 #define REALPART(z) ((z).a[0])
460 #define IMAGPART(z) ((z).a[1])
463 * Inline functions that can be used to construct complex values.
465 * The C99 standard intends x+I*y to be used for this, but x+I*y is
466 * currently unusable in general since gcc introduces many overflow,
467 * underflow, sign and efficiency bugs by rewriting I*y as
468 * (0.0+I)*(y+0.0*I) and laboriously computing the full complex product.
469 * In particular, I*Inf is corrupted to NaN+I*Inf, and I*-0 is corrupted
472 * The C11 standard introduced the macros CMPLX(), CMPLXF() and CMPLXL()
473 * to construct complex values. Compilers that conform to the C99
474 * standard require the following functions to avoid the above issues.
478 static __inline float complex
479 CMPLXF(float x, float y)
490 static __inline double complex
491 CMPLX(double x, double y)
502 static __inline long double complex
503 CMPLXL(long double x, long double y)
505 long_double_complex z;
513 #endif /* _COMPLEX_H */
515 #ifdef __GNUCLIKE_ASM
517 /* Asm versions of some functions. */
525 asm("cvtsd2si %1,%0" : "=r" (n) : "x" (x));
528 #define HAVE_EFFICIENT_IRINT
537 asm("fistl %0" : "=m" (n) : "t" (x));
540 #define HAVE_EFFICIENT_IRINT
543 #if defined(__amd64__) || defined(__i386__)
545 irintl(long double x)
549 asm("fistl %0" : "=m" (n) : "t" (x));
552 #define HAVE_EFFICIENT_IRINTL
555 #endif /* __GNUCLIKE_ASM */
558 #if defined(__amd64__) || defined(__i386__)
559 #define breakpoint() asm("int $3")
563 #define breakpoint() raise(SIGTRAP)
567 /* Write a pari script to test things externally. */
571 #ifndef DOPRINT_SWIZZLE
572 #define DOPRINT_SWIZZLE 0
577 #define DOPRINT_START(xp) do { \
581 /* Hack to give more-problematic args. */ \
582 EXTRACT_LDBL80_WORDS(__hx, __lx, *xp); \
583 __lx ^= DOPRINT_SWIZZLE; \
584 INSERT_LDBL80_WORDS(*xp, __hx, __lx); \
585 printf("x = %.21Lg; ", (long double)*xp); \
587 #define DOPRINT_END1(v) \
588 printf("y = %.21Lg; z = 0; show(x, y, z);\n", (long double)(v))
589 #define DOPRINT_END2(hi, lo) \
590 printf("y = %.21Lg; z = %.21Lg; show(x, y, z);\n", \
591 (long double)(hi), (long double)(lo))
593 #elif defined(DOPRINT_D64)
595 #define DOPRINT_START(xp) do { \
596 uint32_t __hx, __lx; \
598 EXTRACT_WORDS(__hx, __lx, *xp); \
599 __lx ^= DOPRINT_SWIZZLE; \
600 INSERT_WORDS(*xp, __hx, __lx); \
601 printf("x = %.21Lg; ", (long double)*xp); \
603 #define DOPRINT_END1(v) \
604 printf("y = %.21Lg; z = 0; show(x, y, z);\n", (long double)(v))
605 #define DOPRINT_END2(hi, lo) \
606 printf("y = %.21Lg; z = %.21Lg; show(x, y, z);\n", \
607 (long double)(hi), (long double)(lo))
609 #elif defined(DOPRINT_F32)
611 #define DOPRINT_START(xp) do { \
614 GET_FLOAT_WORD(__hx, *xp); \
615 __hx ^= DOPRINT_SWIZZLE; \
616 SET_FLOAT_WORD(*xp, __hx); \
617 printf("x = %.21Lg; ", (long double)*xp); \
619 #define DOPRINT_END1(v) \
620 printf("y = %.21Lg; z = 0; show(x, y, z);\n", (long double)(v))
621 #define DOPRINT_END2(hi, lo) \
622 printf("y = %.21Lg; z = %.21Lg; show(x, y, z);\n", \
623 (long double)(hi), (long double)(lo))
625 #else /* !DOPRINT_LD80 && !DOPRINT_D64 (LD128 only) */
627 #ifndef DOPRINT_SWIZZLE_HIGH
628 #define DOPRINT_SWIZZLE_HIGH 0
631 #define DOPRINT_START(xp) do { \
632 uint64_t __lx, __llx; \
635 EXTRACT_LDBL128_WORDS(__hx, __lx, __llx, *xp); \
636 __llx ^= DOPRINT_SWIZZLE; \
637 __lx ^= DOPRINT_SWIZZLE_HIGH; \
638 INSERT_LDBL128_WORDS(*xp, __hx, __lx, __llx); \
639 printf("x = %.36Lg; ", (long double)*xp); \
641 #define DOPRINT_END1(v) \
642 printf("y = %.36Lg; z = 0; show(x, y, z);\n", (long double)(v))
643 #define DOPRINT_END2(hi, lo) \
644 printf("y = %.36Lg; z = %.36Lg; show(x, y, z);\n", \
645 (long double)(hi), (long double)(lo))
647 #endif /* DOPRINT_LD80 */
650 #define DOPRINT_START(xp)
651 #define DOPRINT_END1(v)
652 #define DOPRINT_END2(hi, lo)
655 #define RETURNP(x) do { \
659 #define RETURNPI(x) do { \
663 #define RETURN2P(x, y) do { \
664 DOPRINT_END2((x), (y)); \
665 RETURNF((x) + (y)); \
667 #define RETURN2PI(x, y) do { \
668 DOPRINT_END2((x), (y)); \
669 RETURNI((x) + (y)); \
672 #define RETURNSP(rp) do { \
675 RETURN2P((rp)->hi, (rp)->lo); \
677 #define RETURNSPI(rp) do { \
679 RETURNPI((rp)->hi); \
680 RETURN2PI((rp)->hi, (rp)->lo); \
683 #define SUM2P(x, y) ({ \
684 const __typeof (x) __x = (x); \
685 const __typeof (y) __y = (y); \
687 DOPRINT_END2(__x, __y); \
692 * ieee style elementary functions
694 * We rename functions here to improve other sources' diffability
697 #define __ieee754_sqrt sqrt
698 #define __ieee754_acos acos
699 #define __ieee754_acosh acosh
700 #define __ieee754_log log
701 #define __ieee754_log2 log2
702 #define __ieee754_atanh atanh
703 #define __ieee754_asin asin
704 #define __ieee754_atan2 atan2
705 #define __ieee754_exp exp
706 #define __ieee754_cosh cosh
707 #define __ieee754_fmod fmod
708 #define __ieee754_pow pow
709 #define __ieee754_lgamma lgamma
710 #define __ieee754_gamma gamma
711 #define __ieee754_lgamma_r lgamma_r
712 #define __ieee754_gamma_r gamma_r
713 #define __ieee754_log10 log10
714 #define __ieee754_sinh sinh
715 #define __ieee754_hypot hypot
716 #define __ieee754_j0 j0
717 #define __ieee754_j1 j1
718 #define __ieee754_y0 y0
719 #define __ieee754_y1 y1
720 #define __ieee754_jn jn
721 #define __ieee754_yn yn
722 #define __ieee754_remainder remainder
723 #define __ieee754_scalb scalb
724 #define __ieee754_sqrtf sqrtf
725 #define __ieee754_acosf acosf
726 #define __ieee754_acoshf acoshf
727 #define __ieee754_logf logf
728 #define __ieee754_atanhf atanhf
729 #define __ieee754_asinf asinf
730 #define __ieee754_atan2f atan2f
731 #define __ieee754_expf expf
732 #define __ieee754_coshf coshf
733 #define __ieee754_fmodf fmodf
734 #define __ieee754_powf powf
735 #define __ieee754_lgammaf lgammaf
736 #define __ieee754_gammaf gammaf
737 #define __ieee754_lgammaf_r lgammaf_r
738 #define __ieee754_gammaf_r gammaf_r
739 #define __ieee754_log10f log10f
740 #define __ieee754_log2f log2f
741 #define __ieee754_sinhf sinhf
742 #define __ieee754_hypotf hypotf
743 #define __ieee754_j0f j0f
744 #define __ieee754_j1f j1f
745 #define __ieee754_y0f y0f
746 #define __ieee754_y1f y1f
747 #define __ieee754_jnf jnf
748 #define __ieee754_ynf ynf
749 #define __ieee754_remainderf remainderf
750 #define __ieee754_scalbf scalbf
752 /* fdlibm kernel function */
753 int __kernel_rem_pio2(double*,double*,int,int,int);
755 /* double precision kernel functions */
756 #ifndef INLINE_REM_PIO2
757 int __ieee754_rem_pio2(double,double*);
759 double __kernel_sin(double,double,int);
760 double __kernel_cos(double,double);
761 double __kernel_tan(double,double,int);
762 double __ldexp_exp(double,int);
764 double complex __ldexp_cexp(double complex,int);
767 /* float precision kernel functions */
768 #ifndef INLINE_REM_PIO2F
769 int __ieee754_rem_pio2f(float,double*);
771 #ifndef INLINE_KERNEL_SINDF
772 float __kernel_sindf(double);
774 #ifndef INLINE_KERNEL_COSDF
775 float __kernel_cosdf(double);
777 #ifndef INLINE_KERNEL_TANDF
778 float __kernel_tandf(double,int);
780 float __ldexp_expf(float,int);
782 float complex __ldexp_cexpf(float complex,int);
785 /* long double precision kernel functions */
786 long double __kernel_sinl(long double, long double, int);
787 long double __kernel_cosl(long double, long double);
788 long double __kernel_tanl(long double, long double, int);
790 #endif /* !_MATH_PRIVATE_H_ */