1 | /* Inline math functions for i387 and SSE. |
2 | Copyright (C) 1995-2018 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
4 | |
5 | The GNU C Library is free software; you can redistribute it and/or |
6 | modify it under the terms of the GNU Lesser General Public |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the License, or (at your option) any later version. |
9 | |
10 | The GNU C Library is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | Lesser General Public License for more details. |
14 | |
15 | You should have received a copy of the GNU Lesser General Public |
16 | License along with the GNU C Library; if not, see |
17 | <http://www.gnu.org/licenses/>. */ |
18 | |
19 | #ifndef _MATH_H |
20 | # error "Never use <bits/mathinline.h> directly; include <math.h> instead." |
21 | #endif |
22 | |
23 | #ifndef __extern_always_inline |
24 | # define __MATH_INLINE __inline |
25 | #else |
26 | # define __MATH_INLINE __extern_always_inline |
27 | #endif |
28 | |
29 | /* The gcc, version 2.7 or below, has problems with all this inlining |
30 | code. So disable it for this version of the compiler. */ |
31 | #if __GNUC_PREREQ (2, 8) |
32 | # if !__GNUC_PREREQ (3, 4) && !defined __NO_MATH_INLINES \ |
33 | && defined __OPTIMIZE__ |
34 | /* GCC 3.4 introduced builtins for all functions below, so |
35 | there's no need to define any of these inline functions. */ |
36 | |
37 | # ifdef __USE_ISOC99 |
38 | |
39 | /* Round to nearest integer. */ |
40 | # ifdef __SSE_MATH__ |
41 | __MATH_INLINE long int |
42 | __NTH (lrintf (float __x)) |
43 | { |
44 | long int __res; |
45 | /* Mark as volatile since the result is dependent on the state of |
46 | the SSE control register (the rounding mode). Otherwise GCC might |
47 | remove these assembler instructions since it does not know about |
48 | the rounding mode change and cannot currently be told. */ |
49 | __asm __volatile__ ("cvtss2si %1, %0" : "=r" (__res) : "xm" (__x)); |
50 | return __res; |
51 | } |
52 | # endif |
53 | # ifdef __SSE2_MATH__ |
54 | __MATH_INLINE long int |
55 | __NTH (lrint (double __x)) |
56 | { |
57 | long int __res; |
58 | /* Mark as volatile since the result is dependent on the state of |
59 | the SSE control register (the rounding mode). Otherwise GCC might |
60 | remove these assembler instructions since it does not know about |
61 | the rounding mode change and cannot currently be told. */ |
62 | __asm __volatile__ ("cvtsd2si %1, %0" : "=r" (__res) : "xm" (__x)); |
63 | return __res; |
64 | } |
65 | # endif |
66 | # ifdef __x86_64__ |
67 | __extension__ |
68 | __MATH_INLINE long long int |
69 | __NTH (llrintf (float __x)) |
70 | { |
71 | long long int __res; |
72 | /* Mark as volatile since the result is dependent on the state of |
73 | the SSE control register (the rounding mode). Otherwise GCC might |
74 | remove these assembler instructions since it does not know about |
75 | the rounding mode change and cannot currently be told. */ |
76 | __asm __volatile__ ("cvtss2si %1, %0" : "=r" (__res) : "xm" (__x)); |
77 | return __res; |
78 | } |
79 | __extension__ |
80 | __MATH_INLINE long long int |
81 | __NTH (llrint (double __x)) |
82 | { |
83 | long long int __res; |
84 | /* Mark as volatile since the result is dependent on the state of |
85 | the SSE control register (the rounding mode). Otherwise GCC might |
86 | remove these assembler instructions since it does not know about |
87 | the rounding mode change and cannot currently be told. */ |
88 | __asm __volatile__ ("cvtsd2si %1, %0" : "=r" (__res) : "xm" (__x)); |
89 | return __res; |
90 | } |
91 | # endif |
92 | |
93 | # if defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ > 0 \ |
94 | && defined __SSE2_MATH__ |
95 | /* Determine maximum of two values. */ |
96 | __MATH_INLINE float |
97 | __NTH (fmaxf (float __x, float __y)) |
98 | { |
99 | # ifdef __AVX__ |
100 | float __res; |
101 | __asm ("vmaxss %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); |
102 | return __res; |
103 | # else |
104 | __asm ("maxss %1, %0" : "+x" (__x) : "xm" (__y)); |
105 | return __x; |
106 | # endif |
107 | } |
108 | __MATH_INLINE double |
109 | __NTH (fmax (double __x, double __y)) |
110 | { |
111 | # ifdef __AVX__ |
112 | float __res; |
113 | __asm ("vmaxsd %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); |
114 | return __res; |
115 | # else |
116 | __asm ("maxsd %1, %0" : "+x" (__x) : "xm" (__y)); |
117 | return __x; |
118 | # endif |
119 | } |
120 | |
121 | /* Determine minimum of two values. */ |
122 | __MATH_INLINE float |
123 | __NTH (fminf (float __x, float __y)) |
124 | { |
125 | # ifdef __AVX__ |
126 | float __res; |
127 | __asm ("vminss %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); |
128 | return __res; |
129 | # else |
130 | __asm ("minss %1, %0" : "+x" (__x) : "xm" (__y)); |
131 | return __x; |
132 | # endif |
133 | } |
134 | __MATH_INLINE double |
135 | __NTH (fmin (double __x, double __y)) |
136 | { |
137 | # ifdef __AVX__ |
138 | float __res; |
139 | __asm ("vminsd %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); |
140 | return __res; |
141 | # else |
142 | __asm ("minsd %1, %0" : "+x" (__x) : "xm" (__y)); |
143 | return __x; |
144 | # endif |
145 | } |
146 | # endif |
147 | |
148 | # endif |
149 | |
150 | # if defined __SSE4_1__ && defined __SSE2_MATH__ |
151 | # if defined __USE_XOPEN_EXTENDED || defined __USE_ISOC99 |
152 | |
153 | /* Round to nearest integer. */ |
154 | __MATH_INLINE double |
155 | __NTH (rint (double __x)) |
156 | { |
157 | double __res; |
158 | /* Mark as volatile since the result is dependent on the state of |
159 | the SSE control register (the rounding mode). Otherwise GCC might |
160 | remove these assembler instructions since it does not know about |
161 | the rounding mode change and cannot currently be told. */ |
162 | __asm __volatile__ ("roundsd $4, %1, %0" : "=x" (__res) : "xm" (__x)); |
163 | return __res; |
164 | } |
165 | __MATH_INLINE float |
166 | __NTH (rintf (float __x)) |
167 | { |
168 | float __res; |
169 | /* Mark as volatile since the result is dependent on the state of |
170 | the SSE control register (the rounding mode). Otherwise GCC might |
171 | remove these assembler instructions since it does not know about |
172 | the rounding mode change and cannot currently be told. */ |
173 | __asm __volatile__ ("roundss $4, %1, %0" : "=x" (__res) : "xm" (__x)); |
174 | return __res; |
175 | } |
176 | |
177 | # ifdef __USE_ISOC99 |
178 | /* Round to nearest integer without raising inexact exception. */ |
179 | __MATH_INLINE double |
180 | __NTH (nearbyint (double __x)) |
181 | { |
182 | double __res; |
183 | /* Mark as volatile since the result is dependent on the state of |
184 | the SSE control register (the rounding mode). Otherwise GCC might |
185 | remove these assembler instructions since it does not know about |
186 | the rounding mode change and cannot currently be told. */ |
187 | __asm __volatile__ ("roundsd $0xc, %1, %0" : "=x" (__res) : "xm" (__x)); |
188 | return __res; |
189 | } |
190 | __MATH_INLINE float |
191 | __NTH (nearbyintf (float __x)) |
192 | { |
193 | float __res; |
194 | /* Mark as volatile since the result is dependent on the state of |
195 | the SSE control register (the rounding mode). Otherwise GCC might |
196 | remove these assembler instructions since it does not know about |
197 | the rounding mode change and cannot currently be told. */ |
198 | __asm __volatile__ ("roundss $0xc, %1, %0" : "=x" (__res) : "xm" (__x)); |
199 | return __res; |
200 | } |
201 | # endif |
202 | |
203 | # endif |
204 | |
205 | /* Smallest integral value not less than X. */ |
206 | __MATH_INLINE double |
207 | __NTH (ceil (double __x)) |
208 | { |
209 | double __res; |
210 | __asm ("roundsd $2, %1, %0" : "=x" (__res) : "xm" (__x)); |
211 | return __res; |
212 | } |
213 | |
214 | __MATH_INLINE float |
215 | __NTH (ceilf (float __x)) |
216 | { |
217 | float __res; |
218 | __asm ("roundss $2, %1, %0" : "=x" (__res) : "xm" (__x)); |
219 | return __res; |
220 | } |
221 | |
222 | /* Largest integer not greater than X. */ |
223 | __MATH_INLINE double |
224 | __NTH (floor (double __x)) |
225 | { |
226 | double __res; |
227 | __asm ("roundsd $1, %1, %0" : "=x" (__res) : "xm" (__x)); |
228 | return __res; |
229 | } |
230 | |
231 | __MATH_INLINE float |
232 | __NTH (floorf (float __x)) |
233 | { |
234 | float __res; |
235 | __asm ("roundss $1, %1, %0" : "=x" (__res) : "xm" (__x)); |
236 | return __res; |
237 | } |
238 | # endif |
239 | # endif |
240 | #endif |
241 | |
242 | /* Disable x87 inlines when -fpmath=sse is passed and also when we're building |
243 | on x86_64. Older gcc (gcc-3.2 for example) does not define __SSE2_MATH__ |
244 | for x86_64. */ |
245 | #if !defined __SSE2_MATH__ && !defined __x86_64__ |
246 | # if ((!defined __NO_MATH_INLINES || defined __LIBC_INTERNAL_MATH_INLINES) \ |
247 | && defined __OPTIMIZE__) |
248 | |
249 | /* The inline functions do not set errno or raise necessarily the |
250 | correct exceptions. */ |
251 | # undef math_errhandling |
252 | |
253 | /* A macro to define float, double, and long double versions of various |
254 | math functions for the ix87 FPU. FUNC is the function name (which will |
255 | be suffixed with f and l for the float and long double version, |
256 | respectively). OP is the name of the FPU operation. |
257 | We define two sets of macros. The set with the additional NP |
258 | doesn't add a prototype declaration. */ |
259 | |
260 | # ifdef __USE_ISOC99 |
261 | # define __inline_mathop(func, op) \ |
262 | __inline_mathop_ (double, func, op) \ |
263 | __inline_mathop_ (float, __CONCAT(func,f), op) \ |
264 | __inline_mathop_ (long double, __CONCAT(func,l), op) |
265 | # define __inline_mathopNP(func, op) \ |
266 | __inline_mathopNP_ (double, func, op) \ |
267 | __inline_mathopNP_ (float, __CONCAT(func,f), op) \ |
268 | __inline_mathopNP_ (long double, __CONCAT(func,l), op) |
269 | # else |
270 | # define __inline_mathop(func, op) \ |
271 | __inline_mathop_ (double, func, op) |
272 | # define __inline_mathopNP(func, op) \ |
273 | __inline_mathopNP_ (double, func, op) |
274 | # endif |
275 | |
276 | # define __inline_mathop_(float_type, func, op) \ |
277 | __inline_mathop_decl_ (float_type, func, op, "0" (__x)) |
278 | # define __inline_mathopNP_(float_type, func, op) \ |
279 | __inline_mathop_declNP_ (float_type, func, op, "0" (__x)) |
280 | |
281 | |
282 | # ifdef __USE_ISOC99 |
283 | # define __inline_mathop_decl(func, op, params...) \ |
284 | __inline_mathop_decl_ (double, func, op, params) \ |
285 | __inline_mathop_decl_ (float, __CONCAT(func,f), op, params) \ |
286 | __inline_mathop_decl_ (long double, __CONCAT(func,l), op, params) |
287 | # define __inline_mathop_declNP(func, op, params...) \ |
288 | __inline_mathop_declNP_ (double, func, op, params) \ |
289 | __inline_mathop_declNP_ (float, __CONCAT(func,f), op, params) \ |
290 | __inline_mathop_declNP_ (long double, __CONCAT(func,l), op, params) |
291 | # else |
292 | # define __inline_mathop_decl(func, op, params...) \ |
293 | __inline_mathop_decl_ (double, func, op, params) |
294 | # define __inline_mathop_declNP(func, op, params...) \ |
295 | __inline_mathop_declNP_ (double, func, op, params) |
296 | # endif |
297 | |
298 | # define __inline_mathop_decl_(float_type, func, op, params...) \ |
299 | __MATH_INLINE float_type func (float_type) __THROW; \ |
300 | __inline_mathop_declNP_ (float_type, func, op, params) |
301 | |
302 | # define __inline_mathop_declNP_(float_type, func, op, params...) \ |
303 | __MATH_INLINE float_type __NTH (func (float_type __x)) \ |
304 | { \ |
305 | register float_type __result; \ |
306 | __asm __volatile__ (op : "=t" (__result) : params); \ |
307 | return __result; \ |
308 | } |
309 | |
310 | |
311 | # ifdef __USE_ISOC99 |
312 | # define __inline_mathcode(func, arg, code) \ |
313 | __inline_mathcode_ (double, func, arg, code) \ |
314 | __inline_mathcode_ (float, __CONCAT(func,f), arg, code) \ |
315 | __inline_mathcode_ (long double, __CONCAT(func,l), arg, code) |
316 | # define __inline_mathcodeNP(func, arg, code) \ |
317 | __inline_mathcodeNP_ (double, func, arg, code) \ |
318 | __inline_mathcodeNP_ (float, __CONCAT(func,f), arg, code) \ |
319 | __inline_mathcodeNP_ (long double, __CONCAT(func,l), arg, code) |
320 | # define __inline_mathcode2(func, arg1, arg2, code) \ |
321 | __inline_mathcode2_ (double, func, arg1, arg2, code) \ |
322 | __inline_mathcode2_ (float, __CONCAT(func,f), arg1, arg2, code) \ |
323 | __inline_mathcode2_ (long double, __CONCAT(func,l), arg1, arg2, code) |
324 | # define __inline_mathcodeNP2(func, arg1, arg2, code) \ |
325 | __inline_mathcodeNP2_ (double, func, arg1, arg2, code) \ |
326 | __inline_mathcodeNP2_ (float, __CONCAT(func,f), arg1, arg2, code) \ |
327 | __inline_mathcodeNP2_ (long double, __CONCAT(func,l), arg1, arg2, code) |
328 | # define __inline_mathcode3(func, arg1, arg2, arg3, code) \ |
329 | __inline_mathcode3_ (double, func, arg1, arg2, arg3, code) \ |
330 | __inline_mathcode3_ (float, __CONCAT(func,f), arg1, arg2, arg3, code) \ |
331 | __inline_mathcode3_ (long double, __CONCAT(func,l), arg1, arg2, arg3, code) |
332 | # define __inline_mathcodeNP3(func, arg1, arg2, arg3, code) \ |
333 | __inline_mathcodeNP3_ (double, func, arg1, arg2, arg3, code) \ |
334 | __inline_mathcodeNP3_ (float, __CONCAT(func,f), arg1, arg2, arg3, code) \ |
335 | __inline_mathcodeNP3_ (long double, __CONCAT(func,l), arg1, arg2, arg3, code) |
336 | # else |
337 | # define __inline_mathcode(func, arg, code) \ |
338 | __inline_mathcode_ (double, func, (arg), code) |
339 | # define __inline_mathcodeNP(func, arg, code) \ |
340 | __inline_mathcodeNP_ (double, func, (arg), code) |
341 | # define __inline_mathcode2(func, arg1, arg2, code) \ |
342 | __inline_mathcode2_ (double, func, arg1, arg2, code) |
343 | # define __inline_mathcodeNP2(func, arg1, arg2, code) \ |
344 | __inline_mathcodeNP2_ (double, func, arg1, arg2, code) |
345 | # define __inline_mathcode3(func, arg1, arg2, arg3, code) \ |
346 | __inline_mathcode3_ (double, func, arg1, arg2, arg3, code) |
347 | # define __inline_mathcodeNP3(func, arg1, arg2, arg3, code) \ |
348 | __inline_mathcodeNP3_ (double, func, arg1, arg2, arg3, code) |
349 | # endif |
350 | |
351 | # define __inline_mathcode_(float_type, func, arg, code) \ |
352 | __MATH_INLINE float_type func (float_type) __THROW; \ |
353 | __inline_mathcodeNP_(float_type, func, arg, code) |
354 | |
355 | # define __inline_mathcodeNP_(float_type, func, arg, code) \ |
356 | __MATH_INLINE float_type __NTH (func (float_type arg)) \ |
357 | { \ |
358 | code; \ |
359 | } |
360 | |
361 | |
362 | # define __inline_mathcode2_(float_type, func, arg1, arg2, code) \ |
363 | __MATH_INLINE float_type func (float_type, float_type) __THROW; \ |
364 | __inline_mathcodeNP2_ (float_type, func, arg1, arg2, code) |
365 | |
366 | # define __inline_mathcodeNP2_(float_type, func, arg1, arg2, code) \ |
367 | __MATH_INLINE float_type __NTH (func (float_type arg1, float_type arg2)) \ |
368 | { \ |
369 | code; \ |
370 | } |
371 | |
372 | # define __inline_mathcode3_(float_type, func, arg1, arg2, arg3, code) \ |
373 | __MATH_INLINE float_type func (float_type, float_type, float_type) __THROW; \ |
374 | __inline_mathcodeNP3_(float_type, func, arg1, arg2, arg3, code) |
375 | |
376 | # define __inline_mathcodeNP3_(float_type, func, arg1, arg2, arg3, code) \ |
377 | __MATH_INLINE float_type __NTH (func (float_type arg1, float_type arg2, \ |
378 | float_type arg3)) \ |
379 | { \ |
380 | code; \ |
381 | } |
382 | # endif |
383 | |
384 | |
385 | # if !defined __NO_MATH_INLINES && defined __OPTIMIZE__ |
386 | /* Miscellaneous functions */ |
387 | |
388 | /* __FAST_MATH__ is defined by gcc -ffast-math. */ |
389 | # ifdef __FAST_MATH__ |
390 | # ifdef __USE_GNU |
391 | # define __sincos_code \ |
392 | register long double __cosr; \ |
393 | register long double __sinr; \ |
394 | register unsigned int __swtmp; \ |
395 | __asm __volatile__ \ |
396 | ("fsincos\n\t" \ |
397 | "fnstsw %w2\n\t" \ |
398 | "testl $0x400, %2\n\t" \ |
399 | "jz 1f\n\t" \ |
400 | "fldpi\n\t" \ |
401 | "fadd %%st(0)\n\t" \ |
402 | "fxch %%st(1)\n\t" \ |
403 | "2: fprem1\n\t" \ |
404 | "fnstsw %w2\n\t" \ |
405 | "testl $0x400, %2\n\t" \ |
406 | "jnz 2b\n\t" \ |
407 | "fstp %%st(1)\n\t" \ |
408 | "fsincos\n\t" \ |
409 | "1:" \ |
410 | : "=t" (__cosr), "=u" (__sinr), "=a" (__swtmp) : "0" (__x)); \ |
411 | *__sinx = __sinr; \ |
412 | *__cosx = __cosr |
413 | |
414 | __MATH_INLINE void |
415 | __NTH (__sincos (double __x, double *__sinx, double *__cosx)) |
416 | { |
417 | __sincos_code; |
418 | } |
419 | |
420 | __MATH_INLINE void |
421 | __NTH (__sincosf (float __x, float *__sinx, float *__cosx)) |
422 | { |
423 | __sincos_code; |
424 | } |
425 | |
426 | __MATH_INLINE void |
427 | __NTH (__sincosl (long double __x, long double *__sinx, long double *__cosx)) |
428 | { |
429 | __sincos_code; |
430 | } |
431 | # endif |
432 | |
433 | |
434 | /* Optimized inline implementation, sometimes with reduced precision |
435 | and/or argument range. */ |
436 | |
437 | # if __GNUC_PREREQ (3, 5) |
438 | # define __expm1_code \ |
439 | register long double __temp; \ |
440 | __temp = __builtin_expm1l (__x); \ |
441 | return __temp ? __temp : __x |
442 | # else |
443 | # define __expm1_code \ |
444 | register long double __value; \ |
445 | register long double __exponent; \ |
446 | register long double __temp; \ |
447 | __asm __volatile__ \ |
448 | ("fldl2e # e^x - 1 = 2^(x * log2(e)) - 1\n\t" \ |
449 | "fmul %%st(1) # x * log2(e)\n\t" \ |
450 | "fst %%st(1)\n\t" \ |
451 | "frndint # int(x * log2(e))\n\t" \ |
452 | "fxch\n\t" \ |
453 | "fsub %%st(1) # fract(x * log2(e))\n\t" \ |
454 | "f2xm1 # 2^(fract(x * log2(e))) - 1\n\t" \ |
455 | "fscale # 2^(x * log2(e)) - 2^(int(x * log2(e)))\n\t" \ |
456 | : "=t" (__value), "=u" (__exponent) : "0" (__x)); \ |
457 | __asm __volatile__ \ |
458 | ("fscale # 2^int(x * log2(e))\n\t" \ |
459 | : "=t" (__temp) : "0" (1.0), "u" (__exponent)); \ |
460 | __temp -= 1.0; \ |
461 | __temp += __value; \ |
462 | return __temp ? __temp : __x |
463 | # endif |
464 | __inline_mathcodeNP_ (long double, __expm1l, __x, __expm1_code) |
465 | |
466 | # if __GNUC_PREREQ (3, 4) |
467 | __inline_mathcodeNP_ (long double, __expl, __x, return __builtin_expl (__x)) |
468 | # else |
469 | # define __exp_code \ |
470 | register long double __value; \ |
471 | register long double __exponent; \ |
472 | __asm __volatile__ \ |
473 | ("fldl2e # e^x = 2^(x * log2(e))\n\t" \ |
474 | "fmul %%st(1) # x * log2(e)\n\t" \ |
475 | "fst %%st(1)\n\t" \ |
476 | "frndint # int(x * log2(e))\n\t" \ |
477 | "fxch\n\t" \ |
478 | "fsub %%st(1) # fract(x * log2(e))\n\t" \ |
479 | "f2xm1 # 2^(fract(x * log2(e))) - 1\n\t" \ |
480 | : "=t" (__value), "=u" (__exponent) : "0" (__x)); \ |
481 | __value += 1.0; \ |
482 | __asm __volatile__ \ |
483 | ("fscale" \ |
484 | : "=t" (__value) : "0" (__value), "u" (__exponent)); \ |
485 | return __value |
486 | __inline_mathcodeNP (exp, __x, __exp_code) |
487 | __inline_mathcodeNP_ (long double, __expl, __x, __exp_code) |
488 | # endif |
489 | |
490 | |
491 | # if !__GNUC_PREREQ (3, 5) |
492 | __inline_mathcodeNP (tan, __x, \ |
493 | register long double __value; \ |
494 | register long double __value2 __attribute__ ((__unused__)); \ |
495 | __asm __volatile__ \ |
496 | ("fptan" \ |
497 | : "=t" (__value2), "=u" (__value) : "0" (__x)); \ |
498 | return __value) |
499 | # endif |
500 | # endif /* __FAST_MATH__ */ |
501 | |
502 | |
503 | # if __GNUC_PREREQ (3, 4) |
504 | __inline_mathcodeNP2_ (long double, __atan2l, __y, __x, |
505 | return __builtin_atan2l (__y, __x)) |
506 | # else |
507 | # define __atan2_code \ |
508 | register long double __value; \ |
509 | __asm __volatile__ \ |
510 | ("fpatan" \ |
511 | : "=t" (__value) : "0" (__x), "u" (__y) : "st(1)"); \ |
512 | return __value |
513 | # ifdef __FAST_MATH__ |
514 | __inline_mathcodeNP2 (atan2, __y, __x, __atan2_code) |
515 | # endif |
516 | __inline_mathcodeNP2_ (long double, __atan2l, __y, __x, __atan2_code) |
517 | # endif |
518 | |
519 | |
520 | # if defined __FAST_MATH__ && !__GNUC_PREREQ (3, 5) |
521 | __inline_mathcodeNP2 (fmod, __x, __y, \ |
522 | register long double __value; \ |
523 | __asm __volatile__ \ |
524 | ("1: fprem\n\t" \ |
525 | "fnstsw %%ax\n\t" \ |
526 | "sahf\n\t" \ |
527 | "jp 1b" \ |
528 | : "=t" (__value) : "0" (__x), "u" (__y) : "ax" , "cc" ); \ |
529 | return __value) |
530 | # endif |
531 | |
532 | |
533 | # ifdef __FAST_MATH__ |
534 | # if !__GNUC_PREREQ (3,3) |
535 | __inline_mathopNP (sqrt, "fsqrt" ) |
536 | __inline_mathopNP_ (long double, __sqrtl, "fsqrt" ) |
537 | # define __libc_sqrtl(n) __sqrtl (n) |
538 | # else |
539 | # define __libc_sqrtl(n) __builtin_sqrtl (n) |
540 | # endif |
541 | # endif |
542 | |
543 | # if __GNUC_PREREQ (2, 8) |
544 | __inline_mathcodeNP_ (double, fabs, __x, return __builtin_fabs (__x)) |
545 | # ifdef __USE_ISOC99 |
546 | __inline_mathcodeNP_ (float, fabsf, __x, return __builtin_fabsf (__x)) |
547 | __inline_mathcodeNP_ (long double, fabsl, __x, return __builtin_fabsl (__x)) |
548 | # endif |
549 | __inline_mathcodeNP_ (long double, __fabsl, __x, return __builtin_fabsl (__x)) |
550 | # else |
551 | __inline_mathop (fabs, "fabs" ) |
552 | __inline_mathop_ (long double, __fabsl, "fabs" ) |
553 | # endif |
554 | |
555 | # ifdef __FAST_MATH__ |
556 | # if !__GNUC_PREREQ (3, 4) |
557 | /* The argument range of this inline version is reduced. */ |
558 | __inline_mathopNP (sin, "fsin" ) |
559 | /* The argument range of this inline version is reduced. */ |
560 | __inline_mathopNP (cos, "fcos" ) |
561 | |
562 | __inline_mathop_declNP (log, "fldln2; fxch; fyl2x" , "0" (__x) : "st(1)" ) |
563 | # endif |
564 | |
565 | # if !__GNUC_PREREQ (3, 5) |
566 | __inline_mathop_declNP (log10, "fldlg2; fxch; fyl2x" , "0" (__x) : "st(1)" ) |
567 | |
568 | __inline_mathcodeNP (asin, __x, return __atan2l (__x, __libc_sqrtl (1.0 - __x * __x))) |
569 | __inline_mathcodeNP (acos, __x, return __atan2l (__libc_sqrtl (1.0 - __x * __x), __x)) |
570 | # endif |
571 | |
572 | # if !__GNUC_PREREQ (3, 4) |
573 | __inline_mathop_declNP (atan, "fld1; fpatan" , "0" (__x) : "st(1)" ) |
574 | # endif |
575 | # endif /* __FAST_MATH__ */ |
576 | |
577 | __inline_mathcode_ (long double, __sgn1l, __x, \ |
578 | __extension__ union { long double __xld; unsigned int __xi[3]; } __n = \ |
579 | { __xld: __x }; \ |
580 | __n.__xi[2] = (__n.__xi[2] & 0x8000) | 0x3fff; \ |
581 | __n.__xi[1] = 0x80000000; \ |
582 | __n.__xi[0] = 0; \ |
583 | return __n.__xld) |
584 | |
585 | |
586 | # ifdef __FAST_MATH__ |
587 | /* The argument range of the inline version of sinhl is slightly reduced. */ |
588 | __inline_mathcodeNP (sinh, __x, \ |
589 | register long double __exm1 = __expm1l (__fabsl (__x)); \ |
590 | return 0.5 * (__exm1 / (__exm1 + 1.0) + __exm1) * __sgn1l (__x)) |
591 | |
592 | __inline_mathcodeNP (cosh, __x, \ |
593 | register long double __ex = __expl (__x); \ |
594 | return 0.5 * (__ex + 1.0 / __ex)) |
595 | |
596 | __inline_mathcodeNP (tanh, __x, \ |
597 | register long double __exm1 = __expm1l (-__fabsl (__x + __x)); \ |
598 | return __exm1 / (__exm1 + 2.0) * __sgn1l (-__x)) |
599 | # endif |
600 | |
601 | __inline_mathcodeNP (floor, __x, \ |
602 | register long double __value; \ |
603 | register int __ignore; \ |
604 | unsigned short int __cw; \ |
605 | unsigned short int __cwtmp; \ |
606 | __asm __volatile ("fnstcw %3\n\t" \ |
607 | "movzwl %3, %1\n\t" \ |
608 | "andl $0xf3ff, %1\n\t" \ |
609 | "orl $0x0400, %1\n\t" /* rounding down */ \ |
610 | "movw %w1, %2\n\t" \ |
611 | "fldcw %2\n\t" \ |
612 | "frndint\n\t" \ |
613 | "fldcw %3" \ |
614 | : "=t" (__value), "=&q" (__ignore), "=m" (__cwtmp), \ |
615 | "=m" (__cw) \ |
616 | : "0" (__x)); \ |
617 | return __value) |
618 | |
619 | __inline_mathcodeNP (ceil, __x, \ |
620 | register long double __value; \ |
621 | register int __ignore; \ |
622 | unsigned short int __cw; \ |
623 | unsigned short int __cwtmp; \ |
624 | __asm __volatile ("fnstcw %3\n\t" \ |
625 | "movzwl %3, %1\n\t" \ |
626 | "andl $0xf3ff, %1\n\t" \ |
627 | "orl $0x0800, %1\n\t" /* rounding up */ \ |
628 | "movw %w1, %2\n\t" \ |
629 | "fldcw %2\n\t" \ |
630 | "frndint\n\t" \ |
631 | "fldcw %3" \ |
632 | : "=t" (__value), "=&q" (__ignore), "=m" (__cwtmp), \ |
633 | "=m" (__cw) \ |
634 | : "0" (__x)); \ |
635 | return __value) |
636 | |
637 | # ifdef __FAST_MATH__ |
638 | # define __ldexp_code \ |
639 | register long double __value; \ |
640 | __asm __volatile__ \ |
641 | ("fscale" \ |
642 | : "=t" (__value) : "0" (__x), "u" ((long double) __y)); \ |
643 | return __value |
644 | |
645 | __MATH_INLINE double |
646 | __NTH (ldexp (double __x, int __y)) |
647 | { |
648 | __ldexp_code; |
649 | } |
650 | # endif |
651 | |
652 | |
653 | /* Optimized versions for some non-standardized functions. */ |
654 | # ifdef __USE_ISOC99 |
655 | |
656 | # ifdef __FAST_MATH__ |
657 | __inline_mathcodeNP (expm1, __x, __expm1_code) |
658 | |
659 | /* We cannot rely on M_SQRT being defined. So we do it for ourself |
660 | here. */ |
661 | # define __M_SQRT2 1.41421356237309504880L /* sqrt(2) */ |
662 | |
663 | # if !__GNUC_PREREQ (3, 5) |
664 | __inline_mathcodeNP (log1p, __x, \ |
665 | register long double __value; \ |
666 | if (__fabsl (__x) >= 1.0 - 0.5 * __M_SQRT2) \ |
667 | __value = logl (1.0 + __x); \ |
668 | else \ |
669 | __asm __volatile__ \ |
670 | ("fldln2\n\t" \ |
671 | "fxch\n\t" \ |
672 | "fyl2xp1" \ |
673 | : "=t" (__value) : "0" (__x) : "st(1)" ); \ |
674 | return __value) |
675 | # endif |
676 | |
677 | |
678 | /* The argument range of the inline version of asinhl is slightly reduced. */ |
679 | __inline_mathcodeNP (asinh, __x, \ |
680 | register long double __y = __fabsl (__x); \ |
681 | return (log1pl (__y * __y / (__libc_sqrtl (__y * __y + 1.0) + 1.0) + __y) \ |
682 | * __sgn1l (__x))) |
683 | |
684 | __inline_mathcodeNP (acosh, __x, \ |
685 | return logl (__x + __libc_sqrtl (__x - 1.0) * __libc_sqrtl (__x + 1.0))) |
686 | |
687 | __inline_mathcodeNP (atanh, __x, \ |
688 | register long double __y = __fabsl (__x); \ |
689 | return -0.5 * log1pl (-(__y + __y) / (1.0 + __y)) * __sgn1l (__x)) |
690 | |
691 | /* The argument range of the inline version of hypotl is slightly reduced. */ |
692 | __inline_mathcodeNP2 (hypot, __x, __y, |
693 | return __libc_sqrtl (__x * __x + __y * __y)) |
694 | |
695 | # if !__GNUC_PREREQ (3, 5) |
696 | __inline_mathcodeNP(logb, __x, \ |
697 | register long double __value; \ |
698 | register long double __junk; \ |
699 | __asm __volatile__ \ |
700 | ("fxtract\n\t" \ |
701 | : "=t" (__junk), "=u" (__value) : "0" (__x)); \ |
702 | return __value) |
703 | # endif |
704 | |
705 | # endif |
706 | # endif |
707 | |
708 | # ifdef __USE_ISOC99 |
709 | # ifdef __FAST_MATH__ |
710 | |
711 | # if !__GNUC_PREREQ (3, 5) |
712 | __inline_mathop_declNP (log2, "fld1; fxch; fyl2x" , "0" (__x) : "st(1)" ) |
713 | # endif |
714 | |
715 | __MATH_INLINE float |
716 | __NTH (ldexpf (float __x, int __y)) |
717 | { |
718 | __ldexp_code; |
719 | } |
720 | |
721 | __MATH_INLINE long double |
722 | __NTH (ldexpl (long double __x, int __y)) |
723 | { |
724 | __ldexp_code; |
725 | } |
726 | |
727 | __inline_mathopNP (rint, "frndint" ) |
728 | # endif /* __FAST_MATH__ */ |
729 | |
730 | # define __lrint_code \ |
731 | long int __lrintres; \ |
732 | __asm__ __volatile__ \ |
733 | ("fistpl %0" \ |
734 | : "=m" (__lrintres) : "t" (__x) : "st"); \ |
735 | return __lrintres |
736 | __MATH_INLINE long int |
737 | __NTH (lrintf (float __x)) |
738 | { |
739 | __lrint_code; |
740 | } |
741 | __MATH_INLINE long int |
742 | __NTH (lrint (double __x)) |
743 | { |
744 | __lrint_code; |
745 | } |
746 | __MATH_INLINE long int |
747 | __NTH (lrintl (long double __x)) |
748 | { |
749 | __lrint_code; |
750 | } |
751 | # undef __lrint_code |
752 | |
753 | # define __llrint_code \ |
754 | long long int __llrintres; \ |
755 | __asm__ __volatile__ \ |
756 | ("fistpll %0" \ |
757 | : "=m" (__llrintres) : "t" (__x) : "st"); \ |
758 | return __llrintres |
759 | __extension__ |
760 | __MATH_INLINE long long int |
761 | __NTH (llrintf (float __x)) |
762 | { |
763 | __llrint_code; |
764 | } |
765 | __extension__ |
766 | __MATH_INLINE long long int |
767 | __NTH (llrint (double __x)) |
768 | { |
769 | __llrint_code; |
770 | } |
771 | __extension__ |
772 | __MATH_INLINE long long int |
773 | __NTH (llrintl (long double __x)) |
774 | { |
775 | __llrint_code; |
776 | } |
777 | # undef __llrint_code |
778 | |
779 | # endif |
780 | |
781 | |
782 | # ifdef __USE_MISC |
783 | |
784 | # if defined __FAST_MATH__ && !__GNUC_PREREQ (3, 5) |
785 | __inline_mathcodeNP2 (drem, __x, __y, \ |
786 | register double __value; \ |
787 | register int __clobbered; \ |
788 | __asm __volatile__ \ |
789 | ("1: fprem1\n\t" \ |
790 | "fstsw %%ax\n\t" \ |
791 | "sahf\n\t" \ |
792 | "jp 1b" \ |
793 | : "=t" (__value), "=&a" (__clobbered) : "0" (__x), "u" (__y) : "cc" ); \ |
794 | return __value) |
795 | # endif |
796 | |
797 | |
798 | /* This function is used in the `isfinite' macro. */ |
799 | __MATH_INLINE int |
800 | __NTH (__finite (double __x)) |
801 | { |
802 | return (__extension__ |
803 | (((((union { double __d; int __i[2]; }) {__d: __x}).__i[1] |
804 | | 0x800fffffu) + 1) >> 31)); |
805 | } |
806 | |
807 | # endif /* __USE_MISC */ |
808 | |
809 | /* Undefine some of the large macros which are not used anymore. */ |
810 | # undef __atan2_code |
811 | # ifdef __FAST_MATH__ |
812 | # undef __expm1_code |
813 | # undef __exp_code |
814 | # undef __sincos_code |
815 | # endif /* __FAST_MATH__ */ |
816 | |
817 | # endif /* __NO_MATH_INLINES */ |
818 | |
819 | |
820 | /* This code is used internally in the GNU libc. */ |
821 | # ifdef __LIBC_INTERNAL_MATH_INLINES |
822 | __inline_mathop (__ieee754_sqrt, "fsqrt" ) |
823 | __inline_mathcode2_ (long double, __ieee754_atan2l, __y, __x, |
824 | register long double __value; |
825 | __asm __volatile__ ("fpatan\n\t" |
826 | : "=t" (__value) |
827 | : "0" (__x), "u" (__y) : "st(1)" ); |
828 | return __value;) |
829 | # endif |
830 | |
831 | #endif /* !__SSE2_MATH__ && !__x86_64__ */ |
832 | |