1/*-
2 * Copyright (c) 2008 David Schultz <das@FreeBSD.org>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in the
12 *    documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27/*
28 * Tests for fma{,f,l}().
29 */
30
31#include <sys/param.h>
32#include <fenv.h>
33#include <float.h>
34#include <math.h>
35#include <stdio.h>
36#include <stdlib.h>
37
38#include "test-utils.h"
39
40#pragma STDC FENV_ACCESS ON
41
42/*
43 * Test that a function returns the correct value and sets the
44 * exception flags correctly. The exceptmask specifies which
45 * exceptions we should check. We need to be lenient for several
46 * reasons, but mainly because on some architectures it's impossible
47 * to raise FE_OVERFLOW without raising FE_INEXACT.
48 *
49 * These are macros instead of functions so that assert provides more
50 * meaningful error messages.
51 */
52#define	test(func, x, y, z, result, exceptmask, excepts) do {		\
53	volatile long double _vx = (x), _vy = (y), _vz = (z);		\
54	ATF_CHECK(feclearexcept(FE_ALL_EXCEPT) == 0);			\
55	CHECK_FPEQUAL((func)(_vx, _vy, _vz), (result));		\
56	CHECK_FP_EXCEPTIONS_MSG(excepts, exceptmask, "for %s(%s)",	\
57	    #func, #x);							\
58} while (0)
59
60#define	testall(x, y, z, result, exceptmask, excepts)	do {		\
61	test(fma, (double)(x), (double)(y), (double)(z),		\
62		(double)(result), (exceptmask), (excepts));		\
63	test(fmaf, (float)(x), (float)(y), (float)(z),			\
64		(float)(result), (exceptmask), (excepts));		\
65	test(fmal, (x), (y), (z), (result), (exceptmask), (excepts));	\
66} while (0)
67
68/* Test in all rounding modes. */
69#define	testrnd(func, x, y, z, rn, ru, rd, rz, exceptmask, excepts)	do { \
70	fesetround(FE_TONEAREST);					\
71	test((func), (x), (y), (z), (rn), (exceptmask), (excepts));	\
72	fesetround(FE_UPWARD);						\
73	test((func), (x), (y), (z), (ru), (exceptmask), (excepts));	\
74	fesetround(FE_DOWNWARD);					\
75	test((func), (x), (y), (z), (rd), (exceptmask), (excepts));	\
76	fesetround(FE_TOWARDZERO);					\
77	test((func), (x), (y), (z), (rz), (exceptmask), (excepts));	\
78} while (0)
79
80/*
81 * This is needed because clang constant-folds fma in ways that are incorrect
82 * in rounding modes other than FE_TONEAREST.
83 */
84static volatile double one = 1.0;
85
86static void
87test_zeroes(void)
88{
89	const int rd = (fegetround() == FE_DOWNWARD);
90
91	testall(0.0, 0.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
92	testall(1.0, 0.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
93	testall(0.0, 1.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
94	testall(0.0, 0.0, 1.0, 1.0, ALL_STD_EXCEPT, 0);
95
96	testall(-0.0, 0.0, 0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
97	testall(0.0, -0.0, 0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
98	testall(-0.0, -0.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
99	testall(0.0, 0.0, -0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
100	testall(-0.0, -0.0, -0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
101
102	testall(-0.0, 0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
103	testall(0.0, -0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
104
105	testall(-one, one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
106	testall(one, -one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
107	testall(-one, -one, -one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
108
109	switch (fegetround()) {
110	case FE_TONEAREST:
111	case FE_TOWARDZERO:
112		test(fmaf, -FLT_MIN, FLT_MIN, 0.0, -0.0,
113		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
114		test(fma, -DBL_MIN, DBL_MIN, 0.0, -0.0,
115		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
116		test(fmal, -LDBL_MIN, LDBL_MIN, 0.0, -0.0,
117		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
118	}
119}
120
121static void
122test_infinities(void)
123{
124	testall(INFINITY, 1.0, -1.0, INFINITY, ALL_STD_EXCEPT, 0);
125	testall(-1.0, INFINITY, 0.0, -INFINITY, ALL_STD_EXCEPT, 0);
126	testall(0.0, 0.0, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
127	testall(1.0, 1.0, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
128	testall(1.0, 1.0, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
129
130	testall(INFINITY, -INFINITY, 1.0, -INFINITY, ALL_STD_EXCEPT, 0);
131	testall(INFINITY, INFINITY, 1.0, INFINITY, ALL_STD_EXCEPT, 0);
132	testall(-INFINITY, -INFINITY, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
133
134	testall(0.0, INFINITY, 1.0, NAN, ALL_STD_EXCEPT, FE_INVALID);
135	testall(INFINITY, 0.0, -0.0, NAN, ALL_STD_EXCEPT, FE_INVALID);
136
137	/* The invalid exception is optional in this case. */
138	testall(INFINITY, 0.0, NAN, NAN, ALL_STD_EXCEPT & ~FE_INVALID, 0);
139
140	testall(INFINITY, INFINITY, -INFINITY, NAN,
141		ALL_STD_EXCEPT, FE_INVALID);
142	testall(-INFINITY, INFINITY, INFINITY, NAN,
143		ALL_STD_EXCEPT, FE_INVALID);
144	testall(INFINITY, -1.0, INFINITY, NAN,
145		ALL_STD_EXCEPT, FE_INVALID);
146
147	test(fmaf, FLT_MAX, FLT_MAX, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
148	test(fma, DBL_MAX, DBL_MAX, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
149	test(fmal, LDBL_MAX, LDBL_MAX, -INFINITY, -INFINITY,
150	     ALL_STD_EXCEPT, 0);
151	test(fmaf, FLT_MAX, -FLT_MAX, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
152	test(fma, DBL_MAX, -DBL_MAX, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
153	test(fmal, LDBL_MAX, -LDBL_MAX, INFINITY, INFINITY,
154	     ALL_STD_EXCEPT, 0);
155}
156
157static void
158test_nans(void)
159{
160	testall(NAN, 0.0, 0.0, NAN, ALL_STD_EXCEPT, 0);
161	testall(1.0, NAN, 1.0, NAN, ALL_STD_EXCEPT, 0);
162	testall(1.0, -1.0, NAN, NAN, ALL_STD_EXCEPT, 0);
163	testall(0.0, 0.0, NAN, NAN, ALL_STD_EXCEPT, 0);
164	testall(NAN, NAN, NAN, NAN, ALL_STD_EXCEPT, 0);
165
166	/* x*y should not raise an inexact/overflow/underflow if z is NaN. */
167	testall(M_PI, M_PI, NAN, NAN, ALL_STD_EXCEPT, 0);
168	test(fmaf, FLT_MIN, FLT_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
169	test(fma, DBL_MIN, DBL_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
170	test(fmal, LDBL_MIN, LDBL_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
171	test(fmaf, FLT_MAX, FLT_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
172	test(fma, DBL_MAX, DBL_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
173	test(fmal, LDBL_MAX, LDBL_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
174}
175
176/*
177 * Tests for cases where z is very small compared to x*y.
178 */
179static void
180test_small_z(void)
181{
182	/* x*y positive, z positive */
183	if (fegetround() == FE_UPWARD) {
184		test(fmaf, one, one, 0x1.0p-100, 1.0 + FLT_EPSILON,
185		     ALL_STD_EXCEPT, FE_INEXACT);
186		test(fma, one, one, 0x1.0p-200, 1.0 + DBL_EPSILON,
187		     ALL_STD_EXCEPT, FE_INEXACT);
188		test(fmal, one, one, 0x1.0p-200, 1.0 + LDBL_EPSILON,
189		     ALL_STD_EXCEPT, FE_INEXACT);
190	} else {
191		testall(0x1.0p100, one, 0x1.0p-100, 0x1.0p100,
192			ALL_STD_EXCEPT, FE_INEXACT);
193	}
194
195	/* x*y negative, z negative */
196	if (fegetround() == FE_DOWNWARD) {
197		test(fmaf, -one, one, -0x1.0p-100, -(1.0 + FLT_EPSILON),
198		     ALL_STD_EXCEPT, FE_INEXACT);
199		test(fma, -one, one, -0x1.0p-200, -(1.0 + DBL_EPSILON),
200		     ALL_STD_EXCEPT, FE_INEXACT);
201		test(fmal, -one, one, -0x1.0p-200, -(1.0 + LDBL_EPSILON),
202		     ALL_STD_EXCEPT, FE_INEXACT);
203	} else {
204		testall(0x1.0p100, -one, -0x1.0p-100, -0x1.0p100,
205			ALL_STD_EXCEPT, FE_INEXACT);
206	}
207
208	/* x*y positive, z negative */
209	if (fegetround() == FE_DOWNWARD || fegetround() == FE_TOWARDZERO) {
210		test(fmaf, one, one, -0x1.0p-100, 1.0 - FLT_EPSILON / 2,
211		     ALL_STD_EXCEPT, FE_INEXACT);
212		test(fma, one, one, -0x1.0p-200, 1.0 - DBL_EPSILON / 2,
213		     ALL_STD_EXCEPT, FE_INEXACT);
214		test(fmal, one, one, -0x1.0p-200, 1.0 - LDBL_EPSILON / 2,
215		     ALL_STD_EXCEPT, FE_INEXACT);
216	} else {
217		testall(0x1.0p100, one, -0x1.0p-100, 0x1.0p100,
218			ALL_STD_EXCEPT, FE_INEXACT);
219	}
220
221	/* x*y negative, z positive */
222	if (fegetround() == FE_UPWARD || fegetround() == FE_TOWARDZERO) {
223		test(fmaf, -one, one, 0x1.0p-100, -1.0 + FLT_EPSILON / 2,
224		     ALL_STD_EXCEPT, FE_INEXACT);
225		test(fma, -one, one, 0x1.0p-200, -1.0 + DBL_EPSILON / 2,
226		     ALL_STD_EXCEPT, FE_INEXACT);
227		test(fmal, -one, one, 0x1.0p-200, -1.0 + LDBL_EPSILON / 2,
228		     ALL_STD_EXCEPT, FE_INEXACT);
229	} else {
230		testall(-0x1.0p100, one, 0x1.0p-100, -0x1.0p100,
231			ALL_STD_EXCEPT, FE_INEXACT);
232	}
233}
234
235/*
236 * Tests for cases where z is very large compared to x*y.
237 */
238static void
239test_big_z(void)
240{
241	/* z positive, x*y positive */
242	if (fegetround() == FE_UPWARD) {
243		test(fmaf, 0x1.0p-50, 0x1.0p-50, 1.0, 1.0 + FLT_EPSILON,
244		     ALL_STD_EXCEPT, FE_INEXACT);
245		test(fma, 0x1.0p-100, 0x1.0p-100, 1.0, 1.0 + DBL_EPSILON,
246		     ALL_STD_EXCEPT, FE_INEXACT);
247		test(fmal, 0x1.0p-100, 0x1.0p-100, 1.0, 1.0 + LDBL_EPSILON,
248		     ALL_STD_EXCEPT, FE_INEXACT);
249	} else {
250		testall(-0x1.0p-50, -0x1.0p-50, 0x1.0p100, 0x1.0p100,
251			ALL_STD_EXCEPT, FE_INEXACT);
252	}
253
254	/* z negative, x*y negative */
255	if (fegetround() == FE_DOWNWARD) {
256		test(fmaf, -0x1.0p-50, 0x1.0p-50, -1.0, -(1.0 + FLT_EPSILON),
257		     ALL_STD_EXCEPT, FE_INEXACT);
258		test(fma, -0x1.0p-100, 0x1.0p-100, -1.0, -(1.0 + DBL_EPSILON),
259		     ALL_STD_EXCEPT, FE_INEXACT);
260		test(fmal, -0x1.0p-100, 0x1.0p-100, -1.0, -(1.0 + LDBL_EPSILON),
261		     ALL_STD_EXCEPT, FE_INEXACT);
262	} else {
263		testall(0x1.0p-50, -0x1.0p-50, -0x1.0p100, -0x1.0p100,
264			ALL_STD_EXCEPT, FE_INEXACT);
265	}
266
267	/* z negative, x*y positive */
268	if (fegetround() == FE_UPWARD || fegetround() == FE_TOWARDZERO) {
269		test(fmaf, -0x1.0p-50, -0x1.0p-50, -1.0,
270		     -1.0 + FLT_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
271		test(fma, -0x1.0p-100, -0x1.0p-100, -1.0,
272		     -1.0 + DBL_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
273		test(fmal, -0x1.0p-100, -0x1.0p-100, -1.0,
274		     -1.0 + LDBL_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
275	} else {
276		testall(0x1.0p-50, 0x1.0p-50, -0x1.0p100, -0x1.0p100,
277			ALL_STD_EXCEPT, FE_INEXACT);
278	}
279
280	/* z positive, x*y negative */
281	if (fegetround() == FE_DOWNWARD || fegetround() == FE_TOWARDZERO) {
282		test(fmaf, 0x1.0p-50, -0x1.0p-50, 1.0, 1.0 - FLT_EPSILON / 2,
283		     ALL_STD_EXCEPT, FE_INEXACT);
284		test(fma, 0x1.0p-100, -0x1.0p-100, 1.0, 1.0 - DBL_EPSILON / 2,
285		     ALL_STD_EXCEPT, FE_INEXACT);
286		test(fmal, 0x1.0p-100, -0x1.0p-100, 1.0, 1.0 - LDBL_EPSILON / 2,
287		     ALL_STD_EXCEPT, FE_INEXACT);
288	} else {
289		testall(-0x1.0p-50, 0x1.0p-50, 0x1.0p100, 0x1.0p100,
290			ALL_STD_EXCEPT, FE_INEXACT);
291	}
292}
293
294static void
295test_accuracy(void)
296{
297
298	/* ilogb(x*y) - ilogb(z) = 20 */
299	testrnd(fmaf, -0x1.c139d8p-51, -0x1.600e7ap32, 0x1.26558cp-38,
300		0x1.34e48ap-18, 0x1.34e48cp-18, 0x1.34e48ap-18, 0x1.34e48ap-18,
301		ALL_STD_EXCEPT, FE_INEXACT);
302	testrnd(fma, -0x1.c139d7b84f1a3p-51, -0x1.600e7a2a16484p32,
303		0x1.26558cac31580p-38, 0x1.34e48a78aae97p-18,
304		0x1.34e48a78aae97p-18, 0x1.34e48a78aae96p-18,
305		0x1.34e48a78aae96p-18, ALL_STD_EXCEPT, FE_INEXACT);
306#if LDBL_MANT_DIG == 113
307	testrnd(fmal, -0x1.c139d7b84f1a3079263afcc5bae3p-51L,
308		-0x1.600e7a2a164840edbe2e7d301a72p32L,
309		0x1.26558cac315807eb07e448042101p-38L,
310		0x1.34e48a78aae96c76ed36077dd387p-18L,
311		0x1.34e48a78aae96c76ed36077dd388p-18L,
312		0x1.34e48a78aae96c76ed36077dd387p-18L,
313		0x1.34e48a78aae96c76ed36077dd387p-18L,
314		ALL_STD_EXCEPT, FE_INEXACT);
315#elif LDBL_MANT_DIG == 64
316	testrnd(fmal, -0x1.c139d7b84f1a307ap-51L, -0x1.600e7a2a164840eep32L,
317		0x1.26558cac315807ecp-38L, 0x1.34e48a78aae96c78p-18L,
318		0x1.34e48a78aae96c78p-18L, 0x1.34e48a78aae96c76p-18L,
319		0x1.34e48a78aae96c76p-18L, ALL_STD_EXCEPT, FE_INEXACT);
320#elif LDBL_MANT_DIG == 53
321	testrnd(fmal, -0x1.c139d7b84f1a3p-51L, -0x1.600e7a2a16484p32L,
322		0x1.26558cac31580p-38L, 0x1.34e48a78aae97p-18L,
323		0x1.34e48a78aae97p-18L, 0x1.34e48a78aae96p-18L,
324		0x1.34e48a78aae96p-18L, ALL_STD_EXCEPT, FE_INEXACT);
325#endif
326
327	/* ilogb(x*y) - ilogb(z) = -40 */
328	testrnd(fmaf, 0x1.98210ap53, 0x1.9556acp-24, 0x1.d87da4p70,
329		0x1.d87da4p70, 0x1.d87da6p70, 0x1.d87da4p70, 0x1.d87da4p70,
330		ALL_STD_EXCEPT, FE_INEXACT);
331	testrnd(fma, 0x1.98210ac83fe2bp53, 0x1.9556ac1475f0fp-24,
332		0x1.d87da3aafc60ep70, 0x1.d87da3aafda40p70,
333		0x1.d87da3aafda40p70, 0x1.d87da3aafda3fp70,
334		0x1.d87da3aafda3fp70, ALL_STD_EXCEPT, FE_INEXACT);
335#if LDBL_MANT_DIG == 113
336	testrnd(fmal, 0x1.98210ac83fe2a8f65b6278b74cebp53L,
337		0x1.9556ac1475f0f28968b61d0de65ap-24L,
338		0x1.d87da3aafc60d830aa4c6d73b749p70L,
339		0x1.d87da3aafda3f36a69eb86488224p70L,
340		0x1.d87da3aafda3f36a69eb86488225p70L,
341		0x1.d87da3aafda3f36a69eb86488224p70L,
342		0x1.d87da3aafda3f36a69eb86488224p70L,
343		ALL_STD_EXCEPT, FE_INEXACT);
344#elif LDBL_MANT_DIG == 64
345	testrnd(fmal, 0x1.98210ac83fe2a8f6p53L, 0x1.9556ac1475f0f28ap-24L,
346		0x1.d87da3aafc60d83p70L, 0x1.d87da3aafda3f36ap70L,
347		0x1.d87da3aafda3f36ap70L, 0x1.d87da3aafda3f368p70L,
348		0x1.d87da3aafda3f368p70L, ALL_STD_EXCEPT, FE_INEXACT);
349#elif LDBL_MANT_DIG == 53
350	testrnd(fmal, 0x1.98210ac83fe2bp53L, 0x1.9556ac1475f0fp-24L,
351		0x1.d87da3aafc60ep70L, 0x1.d87da3aafda40p70L,
352		0x1.d87da3aafda40p70L, 0x1.d87da3aafda3fp70L,
353		0x1.d87da3aafda3fp70L, ALL_STD_EXCEPT, FE_INEXACT);
354#endif
355
356	/* ilogb(x*y) - ilogb(z) = 0 */
357	testrnd(fmaf, 0x1.31ad02p+100, 0x1.2fbf7ap-42, -0x1.c3e106p+58,
358		-0x1.64c27cp+56, -0x1.64c27ap+56, -0x1.64c27cp+56,
359		-0x1.64c27ap+56, ALL_STD_EXCEPT, FE_INEXACT);
360	testrnd(fma, 0x1.31ad012ede8aap+100, 0x1.2fbf79c839067p-42,
361		-0x1.c3e106929056ep+58, -0x1.64c282b970a5fp+56,
362		-0x1.64c282b970a5ep+56, -0x1.64c282b970a5fp+56,
363		-0x1.64c282b970a5ep+56, ALL_STD_EXCEPT, FE_INEXACT);
364#if LDBL_MANT_DIG == 113
365	testrnd(fmal, 0x1.31ad012ede8aa282fa1c19376d16p+100L,
366		 0x1.2fbf79c839066f0f5c68f6d2e814p-42L,
367		-0x1.c3e106929056ec19de72bfe64215p+58L,
368		-0x1.64c282b970a612598fc025ca8cddp+56L,
369		-0x1.64c282b970a612598fc025ca8cddp+56L,
370		-0x1.64c282b970a612598fc025ca8cdep+56L,
371		-0x1.64c282b970a612598fc025ca8cddp+56L,
372		ALL_STD_EXCEPT, FE_INEXACT);
373#elif LDBL_MANT_DIG == 64
374	testrnd(fmal, 0x1.31ad012ede8aa4eap+100L, 0x1.2fbf79c839066aeap-42L,
375		-0x1.c3e106929056e61p+58L, -0x1.64c282b970a60298p+56L,
376		-0x1.64c282b970a60298p+56L, -0x1.64c282b970a6029ap+56L,
377		-0x1.64c282b970a60298p+56L, ALL_STD_EXCEPT, FE_INEXACT);
378#elif LDBL_MANT_DIG == 53
379	testrnd(fmal, 0x1.31ad012ede8aap+100L, 0x1.2fbf79c839067p-42L,
380		-0x1.c3e106929056ep+58L, -0x1.64c282b970a5fp+56L,
381		-0x1.64c282b970a5ep+56L, -0x1.64c282b970a5fp+56L,
382		-0x1.64c282b970a5ep+56L, ALL_STD_EXCEPT, FE_INEXACT);
383#endif
384
385	/* x*y (rounded) ~= -z */
386	/* XXX spurious inexact exceptions */
387	testrnd(fmaf, 0x1.bbffeep-30, -0x1.1d164cp-74, 0x1.ee7296p-104,
388		-0x1.c46ea8p-128, -0x1.c46ea8p-128, -0x1.c46ea8p-128,
389		-0x1.c46ea8p-128, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
390	testrnd(fma, 0x1.bbffeea6fc7d6p-30, 0x1.1d164c6cbf078p-74,
391		-0x1.ee72993aff948p-104, -0x1.71f72ac7d9d8p-159,
392		-0x1.71f72ac7d9d8p-159, -0x1.71f72ac7d9d8p-159,
393		-0x1.71f72ac7d9d8p-159, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
394#if LDBL_MANT_DIG == 113
395	testrnd(fmal, 0x1.bbffeea6fc7d65927d147f437675p-30L,
396		0x1.1d164c6cbf078b7a22607d1cd6a2p-74L,
397		-0x1.ee72993aff94973876031bec0944p-104L,
398		0x1.64e086175b3a2adc36e607058814p-217L,
399		0x1.64e086175b3a2adc36e607058814p-217L,
400		0x1.64e086175b3a2adc36e607058814p-217L,
401		0x1.64e086175b3a2adc36e607058814p-217L,
402		ALL_STD_EXCEPT & ~FE_INEXACT, 0);
403#elif LDBL_MANT_DIG == 64
404	testrnd(fmal, 0x1.bbffeea6fc7d6592p-30L, 0x1.1d164c6cbf078b7ap-74L,
405		-0x1.ee72993aff949736p-104L, 0x1.af190e7a1ee6ad94p-168L,
406		0x1.af190e7a1ee6ad94p-168L, 0x1.af190e7a1ee6ad94p-168L,
407		0x1.af190e7a1ee6ad94p-168L, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
408#elif LDBL_MANT_DIG == 53
409	testrnd(fmal, 0x1.bbffeea6fc7d6p-30L, 0x1.1d164c6cbf078p-74L,
410		-0x1.ee72993aff948p-104L, -0x1.71f72ac7d9d8p-159L,
411		-0x1.71f72ac7d9d8p-159L, -0x1.71f72ac7d9d8p-159L,
412		-0x1.71f72ac7d9d8p-159L, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
413#endif
414}
415
416static void
417test_double_rounding(void)
418{
419
420	/*
421	 *     a =  0x1.8000000000001p0
422	 *     b =  0x1.8000000000001p0
423	 *     c = -0x0.0000000000000000000000000080...1p+1
424	 * a * b =  0x1.2000000000001800000000000080p+1
425	 *
426	 * The correct behavior is to round DOWN to 0x1.2000000000001p+1 in
427	 * round-to-nearest mode.  An implementation that computes a*b+c in
428	 * double+double precision, however, will get 0x1.20000000000018p+1,
429	 * and then round UP.
430	 */
431	fesetround(FE_TONEAREST);
432	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
433	     -0x1.0000000000001p-104, 0x1.2000000000001p+1,
434	     ALL_STD_EXCEPT, FE_INEXACT);
435	fesetround(FE_DOWNWARD);
436	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
437	     -0x1.0000000000001p-104, 0x1.2000000000001p+1,
438	     ALL_STD_EXCEPT, FE_INEXACT);
439	fesetround(FE_UPWARD);
440	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
441	     -0x1.0000000000001p-104, 0x1.2000000000002p+1,
442	     ALL_STD_EXCEPT, FE_INEXACT);
443
444	fesetround(FE_TONEAREST);
445	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200002p+1,
446	     ALL_STD_EXCEPT, FE_INEXACT);
447	fesetround(FE_DOWNWARD);
448	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200002p+1,
449	     ALL_STD_EXCEPT, FE_INEXACT);
450	fesetround(FE_UPWARD);
451	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200004p+1,
452	     ALL_STD_EXCEPT, FE_INEXACT);
453
454	fesetround(FE_TONEAREST);
455#if LDBL_MANT_DIG == 64
456	test(fmal, 0x1.4p+0L, 0x1.0000000000000004p+0L, 0x1p-128L,
457	     0x1.4000000000000006p+0L, ALL_STD_EXCEPT, FE_INEXACT);
458#elif LDBL_MANT_DIG == 113
459	test(fmal, 0x1.8000000000000000000000000001p+0L,
460	     0x1.8000000000000000000000000001p+0L,
461	     -0x1.0000000000000000000000000001p-224L,
462	     0x1.2000000000000000000000000001p+1L, ALL_STD_EXCEPT, FE_INEXACT);
463#endif
464
465}
466
467static const int rmodes[] = {
468	FE_TONEAREST, FE_UPWARD, FE_DOWNWARD, FE_TOWARDZERO
469};
470
471ATF_TC_WITHOUT_HEAD(zeroes);
472ATF_TC_BODY(zeroes, tc)
473{
474	for (size_t i = 0; i < nitems(rmodes); i++) {
475		printf("rmode = %d\n", rmodes[i]);
476		fesetround(rmodes[i]);
477		test_zeroes();
478	}
479}
480
481ATF_TC_WITHOUT_HEAD(infinities);
482ATF_TC_BODY(infinities, tc)
483{
484	for (size_t i = 0; i < nitems(rmodes); i++) {
485		printf("rmode = %d\n", rmodes[i]);
486		fesetround(rmodes[i]);
487		test_infinities();
488	}
489}
490
491ATF_TC_WITHOUT_HEAD(nans);
492ATF_TC_BODY(nans, tc)
493{
494	fesetround(FE_TONEAREST);
495	test_nans();
496}
497
498
499ATF_TC_WITHOUT_HEAD(small_z);
500ATF_TC_BODY(small_z, tc)
501{
502	for (size_t i = 0; i < nitems(rmodes); i++) {
503		printf("rmode = %d\n", rmodes[i]);
504		fesetround(rmodes[i]);
505		test_small_z();
506	}
507}
508
509
510ATF_TC_WITHOUT_HEAD(big_z);
511ATF_TC_BODY(big_z, tc)
512{
513	for (size_t i = 0; i < nitems(rmodes); i++) {
514		printf("rmode = %d\n", rmodes[i]);
515		fesetround(rmodes[i]);
516		test_big_z();
517	}
518}
519
520ATF_TC_WITHOUT_HEAD(accuracy);
521ATF_TC_BODY(accuracy, tc)
522{
523	fesetround(FE_TONEAREST);
524	test_accuracy();
525}
526
527ATF_TC_WITHOUT_HEAD(double_rounding);
528ATF_TC_BODY(double_rounding, tc) {
529	test_double_rounding();
530}
531
532ATF_TP_ADD_TCS(tp)
533{
534	ATF_TP_ADD_TC(tp, zeroes);
535	ATF_TP_ADD_TC(tp, infinities);
536	ATF_TP_ADD_TC(tp, nans);
537	ATF_TP_ADD_TC(tp, small_z);
538	ATF_TP_ADD_TC(tp, big_z);
539	ATF_TP_ADD_TC(tp, accuracy);
540	ATF_TP_ADD_TC(tp, double_rounding);
541	/*
542	 * TODO:
543	 * - Tests for subnormals
544	 * - Cancellation tests (e.g., z = (double)x*y, but x*y is inexact)
545	 */
546	return (atf_no_error());
547}
548