fma_test.c revision 292492
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/cdefs.h>
32__FBSDID("$FreeBSD: head/tools/regression/lib/msun/test-fma.c 292492 2015-12-20 04:28:37Z ngie $");
33
34#include <sys/param.h>
35#include <assert.h>
36#include <fenv.h>
37#include <float.h>
38#include <math.h>
39#include <stdio.h>
40
41#include "test-utils.h"
42
43#pragma STDC FENV_ACCESS ON
44
45/*
46 * Test that a function returns the correct value and sets the
47 * exception flags correctly. The exceptmask specifies which
48 * exceptions we should check. We need to be lenient for several
49 * reasons, but mainly because on some architectures it's impossible
50 * to raise FE_OVERFLOW without raising FE_INEXACT.
51 *
52 * These are macros instead of functions so that assert provides more
53 * meaningful error messages.
54 */
55#define	test(func, x, y, z, result, exceptmask, excepts) do {		\
56	volatile long double _vx = (x), _vy = (y), _vz = (z);		\
57	assert(feclearexcept(FE_ALL_EXCEPT) == 0);			\
58	assert(fpequal((func)(_vx, _vy, _vz), (result)));		\
59	assert(((void)(func), fetestexcept(exceptmask) == (excepts)));	\
60} while (0)
61
62#define	testall(x, y, z, result, exceptmask, excepts)	do {		\
63	test(fma, (double)(x), (double)(y), (double)(z),		\
64		(double)(result), (exceptmask), (excepts));		\
65	test(fmaf, (float)(x), (float)(y), (float)(z),			\
66		(float)(result), (exceptmask), (excepts));		\
67	test(fmal, (x), (y), (z), (result), (exceptmask), (excepts));	\
68} while (0)
69
70/* Test in all rounding modes. */
71#define	testrnd(func, x, y, z, rn, ru, rd, rz, exceptmask, excepts)	do { \
72	fesetround(FE_TONEAREST);					\
73	test((func), (x), (y), (z), (rn), (exceptmask), (excepts));	\
74	fesetround(FE_UPWARD);						\
75	test((func), (x), (y), (z), (ru), (exceptmask), (excepts));	\
76	fesetround(FE_DOWNWARD);					\
77	test((func), (x), (y), (z), (rd), (exceptmask), (excepts));	\
78	fesetround(FE_TOWARDZERO);					\
79	test((func), (x), (y), (z), (rz), (exceptmask), (excepts));	\
80} while (0)
81
82/*
83 * This is needed because clang constant-folds fma in ways that are incorrect
84 * in rounding modes other than FE_TONEAREST.
85 */
86volatile double one = 1.0;
87
88static void
89test_zeroes(void)
90{
91	const int rd = (fegetround() == FE_DOWNWARD);
92
93	testall(0.0, 0.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
94	testall(1.0, 0.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
95	testall(0.0, 1.0, 0.0, 0.0, ALL_STD_EXCEPT, 0);
96	testall(0.0, 0.0, 1.0, 1.0, ALL_STD_EXCEPT, 0);
97
98	testall(-0.0, 0.0, 0.0, rd ? -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, 0.0, ALL_STD_EXCEPT, 0);
101	testall(0.0, 0.0, -0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
102	testall(-0.0, -0.0, -0.0, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
103
104	testall(-0.0, 0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
105	testall(0.0, -0.0, -0.0, -0.0, ALL_STD_EXCEPT, 0);
106
107	testall(-one, one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
108	testall(one, -one, one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
109	testall(-one, -one, -one, rd ? -0.0 : 0.0, ALL_STD_EXCEPT, 0);
110
111	switch (fegetround()) {
112	case FE_TONEAREST:
113	case FE_TOWARDZERO:
114		test(fmaf, -FLT_MIN, FLT_MIN, 0.0, -0.0,
115		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
116		test(fma, -DBL_MIN, DBL_MIN, 0.0, -0.0,
117		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
118		test(fmal, -LDBL_MIN, LDBL_MIN, 0.0, -0.0,
119		     ALL_STD_EXCEPT, FE_INEXACT | FE_UNDERFLOW);
120	}
121}
122
123static void
124test_infinities(void)
125{
126
127	testall(INFINITY, 1.0, -1.0, INFINITY, ALL_STD_EXCEPT, 0);
128	testall(-1.0, INFINITY, 0.0, -INFINITY, ALL_STD_EXCEPT, 0);
129	testall(0.0, 0.0, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
130	testall(1.0, 1.0, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
131	testall(1.0, 1.0, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
132
133	testall(INFINITY, -INFINITY, 1.0, -INFINITY, ALL_STD_EXCEPT, 0);
134	testall(INFINITY, INFINITY, 1.0, INFINITY, ALL_STD_EXCEPT, 0);
135	testall(-INFINITY, -INFINITY, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
136
137	testall(0.0, INFINITY, 1.0, NAN, ALL_STD_EXCEPT, FE_INVALID);
138	testall(INFINITY, 0.0, -0.0, NAN, ALL_STD_EXCEPT, FE_INVALID);
139
140	/* The invalid exception is optional in this case. */
141	testall(INFINITY, 0.0, NAN, NAN, ALL_STD_EXCEPT & ~FE_INVALID, 0);
142
143	testall(INFINITY, INFINITY, -INFINITY, NAN,
144		ALL_STD_EXCEPT, FE_INVALID);
145	testall(-INFINITY, INFINITY, INFINITY, NAN,
146		ALL_STD_EXCEPT, FE_INVALID);
147	testall(INFINITY, -1.0, INFINITY, NAN,
148		ALL_STD_EXCEPT, FE_INVALID);
149
150	test(fmaf, FLT_MAX, FLT_MAX, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
151	test(fma, DBL_MAX, DBL_MAX, -INFINITY, -INFINITY, ALL_STD_EXCEPT, 0);
152	test(fmal, LDBL_MAX, LDBL_MAX, -INFINITY, -INFINITY,
153	     ALL_STD_EXCEPT, 0);
154	test(fmaf, FLT_MAX, -FLT_MAX, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
155	test(fma, DBL_MAX, -DBL_MAX, INFINITY, INFINITY, ALL_STD_EXCEPT, 0);
156	test(fmal, LDBL_MAX, -LDBL_MAX, INFINITY, INFINITY,
157	     ALL_STD_EXCEPT, 0);
158}
159
160static void
161test_nans(void)
162{
163
164	testall(NAN, 0.0, 0.0, NAN, ALL_STD_EXCEPT, 0);
165	testall(1.0, NAN, 1.0, NAN, ALL_STD_EXCEPT, 0);
166	testall(1.0, -1.0, NAN, NAN, ALL_STD_EXCEPT, 0);
167	testall(0.0, 0.0, NAN, NAN, ALL_STD_EXCEPT, 0);
168	testall(NAN, NAN, NAN, NAN, ALL_STD_EXCEPT, 0);
169
170	/* x*y should not raise an inexact/overflow/underflow if z is NaN. */
171	testall(M_PI, M_PI, NAN, NAN, ALL_STD_EXCEPT, 0);
172	test(fmaf, FLT_MIN, FLT_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
173	test(fma, DBL_MIN, DBL_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
174	test(fmal, LDBL_MIN, LDBL_MIN, NAN, NAN, ALL_STD_EXCEPT, 0);
175	test(fmaf, FLT_MAX, FLT_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
176	test(fma, DBL_MAX, DBL_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
177	test(fmal, LDBL_MAX, LDBL_MAX, NAN, NAN, ALL_STD_EXCEPT, 0);
178}
179
180/*
181 * Tests for cases where z is very small compared to x*y.
182 */
183static void
184test_small_z(void)
185{
186
187	/* x*y positive, z positive */
188	if (fegetround() == FE_UPWARD) {
189		test(fmaf, one, one, 0x1.0p-100, 1.0 + FLT_EPSILON,
190		     ALL_STD_EXCEPT, FE_INEXACT);
191		test(fma, one, one, 0x1.0p-200, 1.0 + DBL_EPSILON,
192		     ALL_STD_EXCEPT, FE_INEXACT);
193		test(fmal, one, one, 0x1.0p-200, 1.0 + LDBL_EPSILON,
194		     ALL_STD_EXCEPT, FE_INEXACT);
195	} else {
196		testall(0x1.0p100, one, 0x1.0p-100, 0x1.0p100,
197			ALL_STD_EXCEPT, FE_INEXACT);
198	}
199
200	/* x*y negative, z negative */
201	if (fegetround() == FE_DOWNWARD) {
202		test(fmaf, -one, one, -0x1.0p-100, -(1.0 + FLT_EPSILON),
203		     ALL_STD_EXCEPT, FE_INEXACT);
204		test(fma, -one, one, -0x1.0p-200, -(1.0 + DBL_EPSILON),
205		     ALL_STD_EXCEPT, FE_INEXACT);
206		test(fmal, -one, one, -0x1.0p-200, -(1.0 + LDBL_EPSILON),
207		     ALL_STD_EXCEPT, FE_INEXACT);
208	} else {
209		testall(0x1.0p100, -one, -0x1.0p-100, -0x1.0p100,
210			ALL_STD_EXCEPT, FE_INEXACT);
211	}
212
213	/* x*y positive, z negative */
214	if (fegetround() == FE_DOWNWARD || fegetround() == FE_TOWARDZERO) {
215		test(fmaf, one, one, -0x1.0p-100, 1.0 - FLT_EPSILON / 2,
216		     ALL_STD_EXCEPT, FE_INEXACT);
217		test(fma, one, one, -0x1.0p-200, 1.0 - DBL_EPSILON / 2,
218		     ALL_STD_EXCEPT, FE_INEXACT);
219		test(fmal, one, one, -0x1.0p-200, 1.0 - LDBL_EPSILON / 2,
220		     ALL_STD_EXCEPT, FE_INEXACT);
221	} else {
222		testall(0x1.0p100, one, -0x1.0p-100, 0x1.0p100,
223			ALL_STD_EXCEPT, FE_INEXACT);
224	}
225
226	/* x*y negative, z positive */
227	if (fegetround() == FE_UPWARD || fegetround() == FE_TOWARDZERO) {
228		test(fmaf, -one, one, 0x1.0p-100, -1.0 + FLT_EPSILON / 2,
229		     ALL_STD_EXCEPT, FE_INEXACT);
230		test(fma, -one, one, 0x1.0p-200, -1.0 + DBL_EPSILON / 2,
231		     ALL_STD_EXCEPT, FE_INEXACT);
232		test(fmal, -one, one, 0x1.0p-200, -1.0 + LDBL_EPSILON / 2,
233		     ALL_STD_EXCEPT, FE_INEXACT);
234	} else {
235		testall(-0x1.0p100, one, 0x1.0p-100, -0x1.0p100,
236			ALL_STD_EXCEPT, FE_INEXACT);
237	}
238}
239
240/*
241 * Tests for cases where z is very large compared to x*y.
242 */
243static void
244test_big_z(void)
245{
246
247	/* z positive, x*y positive */
248	if (fegetround() == FE_UPWARD) {
249		test(fmaf, 0x1.0p-50, 0x1.0p-50, 1.0, 1.0 + FLT_EPSILON,
250		     ALL_STD_EXCEPT, FE_INEXACT);
251		test(fma, 0x1.0p-100, 0x1.0p-100, 1.0, 1.0 + DBL_EPSILON,
252		     ALL_STD_EXCEPT, FE_INEXACT);
253		test(fmal, 0x1.0p-100, 0x1.0p-100, 1.0, 1.0 + LDBL_EPSILON,
254		     ALL_STD_EXCEPT, FE_INEXACT);
255	} else {
256		testall(-0x1.0p-50, -0x1.0p-50, 0x1.0p100, 0x1.0p100,
257			ALL_STD_EXCEPT, FE_INEXACT);
258	}
259
260	/* z negative, x*y negative */
261	if (fegetround() == FE_DOWNWARD) {
262		test(fmaf, -0x1.0p-50, 0x1.0p-50, -1.0, -(1.0 + FLT_EPSILON),
263		     ALL_STD_EXCEPT, FE_INEXACT);
264		test(fma, -0x1.0p-100, 0x1.0p-100, -1.0, -(1.0 + DBL_EPSILON),
265		     ALL_STD_EXCEPT, FE_INEXACT);
266		test(fmal, -0x1.0p-100, 0x1.0p-100, -1.0, -(1.0 + LDBL_EPSILON),
267		     ALL_STD_EXCEPT, FE_INEXACT);
268	} else {
269		testall(0x1.0p-50, -0x1.0p-50, -0x1.0p100, -0x1.0p100,
270			ALL_STD_EXCEPT, FE_INEXACT);
271	}
272
273	/* z negative, x*y positive */
274	if (fegetround() == FE_UPWARD || fegetround() == FE_TOWARDZERO) {
275		test(fmaf, -0x1.0p-50, -0x1.0p-50, -1.0,
276		     -1.0 + FLT_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
277		test(fma, -0x1.0p-100, -0x1.0p-100, -1.0,
278		     -1.0 + DBL_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
279		test(fmal, -0x1.0p-100, -0x1.0p-100, -1.0,
280		     -1.0 + LDBL_EPSILON / 2, ALL_STD_EXCEPT, FE_INEXACT);
281	} else {
282		testall(0x1.0p-50, 0x1.0p-50, -0x1.0p100, -0x1.0p100,
283			ALL_STD_EXCEPT, FE_INEXACT);
284	}
285
286	/* z positive, x*y negative */
287	if (fegetround() == FE_DOWNWARD || fegetround() == FE_TOWARDZERO) {
288		test(fmaf, 0x1.0p-50, -0x1.0p-50, 1.0, 1.0 - FLT_EPSILON / 2,
289		     ALL_STD_EXCEPT, FE_INEXACT);
290		test(fma, 0x1.0p-100, -0x1.0p-100, 1.0, 1.0 - DBL_EPSILON / 2,
291		     ALL_STD_EXCEPT, FE_INEXACT);
292		test(fmal, 0x1.0p-100, -0x1.0p-100, 1.0, 1.0 - LDBL_EPSILON / 2,
293		     ALL_STD_EXCEPT, FE_INEXACT);
294	} else {
295		testall(-0x1.0p-50, 0x1.0p-50, 0x1.0p100, 0x1.0p100,
296			ALL_STD_EXCEPT, FE_INEXACT);
297	}
298}
299
300static void
301test_accuracy(void)
302{
303
304	/* ilogb(x*y) - ilogb(z) = 20 */
305	testrnd(fmaf, -0x1.c139d8p-51, -0x1.600e7ap32, 0x1.26558cp-38,
306		0x1.34e48ap-18, 0x1.34e48cp-18, 0x1.34e48ap-18, 0x1.34e48ap-18,
307		ALL_STD_EXCEPT, FE_INEXACT);
308	testrnd(fma, -0x1.c139d7b84f1a3p-51, -0x1.600e7a2a16484p32,
309		0x1.26558cac31580p-38, 0x1.34e48a78aae97p-18,
310		0x1.34e48a78aae97p-18, 0x1.34e48a78aae96p-18,
311		0x1.34e48a78aae96p-18, ALL_STD_EXCEPT, FE_INEXACT);
312#if LDBL_MANT_DIG == 113
313	testrnd(fmal, -0x1.c139d7b84f1a3079263afcc5bae3p-51L,
314		-0x1.600e7a2a164840edbe2e7d301a72p32L,
315		0x1.26558cac315807eb07e448042101p-38L,
316		0x1.34e48a78aae96c76ed36077dd387p-18L,
317		0x1.34e48a78aae96c76ed36077dd388p-18L,
318		0x1.34e48a78aae96c76ed36077dd387p-18L,
319		0x1.34e48a78aae96c76ed36077dd387p-18L,
320		ALL_STD_EXCEPT, FE_INEXACT);
321#elif LDBL_MANT_DIG == 64
322	testrnd(fmal, -0x1.c139d7b84f1a307ap-51L, -0x1.600e7a2a164840eep32L,
323		0x1.26558cac315807ecp-38L, 0x1.34e48a78aae96c78p-18L,
324		0x1.34e48a78aae96c78p-18L, 0x1.34e48a78aae96c76p-18L,
325		0x1.34e48a78aae96c76p-18L, ALL_STD_EXCEPT, FE_INEXACT);
326#elif LDBL_MANT_DIG == 53
327	testrnd(fmal, -0x1.c139d7b84f1a3p-51L, -0x1.600e7a2a16484p32L,
328		0x1.26558cac31580p-38L, 0x1.34e48a78aae97p-18L,
329		0x1.34e48a78aae97p-18L, 0x1.34e48a78aae96p-18L,
330		0x1.34e48a78aae96p-18L, ALL_STD_EXCEPT, FE_INEXACT);
331#endif
332
333	/* ilogb(x*y) - ilogb(z) = -40 */
334	testrnd(fmaf, 0x1.98210ap53, 0x1.9556acp-24, 0x1.d87da4p70,
335		0x1.d87da4p70, 0x1.d87da6p70, 0x1.d87da4p70, 0x1.d87da4p70,
336		ALL_STD_EXCEPT, FE_INEXACT);
337	testrnd(fma, 0x1.98210ac83fe2bp53, 0x1.9556ac1475f0fp-24,
338		0x1.d87da3aafc60ep70, 0x1.d87da3aafda40p70,
339		0x1.d87da3aafda40p70, 0x1.d87da3aafda3fp70,
340		0x1.d87da3aafda3fp70, ALL_STD_EXCEPT, FE_INEXACT);
341#if LDBL_MANT_DIG == 113
342	testrnd(fmal, 0x1.98210ac83fe2a8f65b6278b74cebp53L,
343		0x1.9556ac1475f0f28968b61d0de65ap-24L,
344		0x1.d87da3aafc60d830aa4c6d73b749p70L,
345		0x1.d87da3aafda3f36a69eb86488224p70L,
346		0x1.d87da3aafda3f36a69eb86488225p70L,
347		0x1.d87da3aafda3f36a69eb86488224p70L,
348		0x1.d87da3aafda3f36a69eb86488224p70L,
349		ALL_STD_EXCEPT, FE_INEXACT);
350#elif LDBL_MANT_DIG == 64
351	testrnd(fmal, 0x1.98210ac83fe2a8f6p53L, 0x1.9556ac1475f0f28ap-24L,
352		0x1.d87da3aafc60d83p70L, 0x1.d87da3aafda3f36ap70L,
353		0x1.d87da3aafda3f36ap70L, 0x1.d87da3aafda3f368p70L,
354		0x1.d87da3aafda3f368p70L, ALL_STD_EXCEPT, FE_INEXACT);
355#elif LDBL_MANT_DIG == 53
356	testrnd(fmal, 0x1.98210ac83fe2bp53L, 0x1.9556ac1475f0fp-24L,
357		0x1.d87da3aafc60ep70L, 0x1.d87da3aafda40p70L,
358		0x1.d87da3aafda40p70L, 0x1.d87da3aafda3fp70L,
359		0x1.d87da3aafda3fp70L, ALL_STD_EXCEPT, FE_INEXACT);
360#endif
361
362	/* ilogb(x*y) - ilogb(z) = 0 */
363	testrnd(fmaf, 0x1.31ad02p+100, 0x1.2fbf7ap-42, -0x1.c3e106p+58,
364		-0x1.64c27cp+56, -0x1.64c27ap+56, -0x1.64c27cp+56,
365		-0x1.64c27ap+56, ALL_STD_EXCEPT, FE_INEXACT);
366	testrnd(fma, 0x1.31ad012ede8aap+100, 0x1.2fbf79c839067p-42,
367		-0x1.c3e106929056ep+58, -0x1.64c282b970a5fp+56,
368		-0x1.64c282b970a5ep+56, -0x1.64c282b970a5fp+56,
369		-0x1.64c282b970a5ep+56, ALL_STD_EXCEPT, FE_INEXACT);
370#if LDBL_MANT_DIG == 113
371	testrnd(fmal, 0x1.31ad012ede8aa282fa1c19376d16p+100L,
372		 0x1.2fbf79c839066f0f5c68f6d2e814p-42L,
373		-0x1.c3e106929056ec19de72bfe64215p+58L,
374		-0x1.64c282b970a612598fc025ca8cddp+56L,
375		-0x1.64c282b970a612598fc025ca8cddp+56L,
376		-0x1.64c282b970a612598fc025ca8cdep+56L,
377		-0x1.64c282b970a612598fc025ca8cddp+56L,
378		ALL_STD_EXCEPT, FE_INEXACT);
379#elif LDBL_MANT_DIG == 64
380	testrnd(fmal, 0x1.31ad012ede8aa4eap+100L, 0x1.2fbf79c839066aeap-42L,
381		-0x1.c3e106929056e61p+58L, -0x1.64c282b970a60298p+56L,
382		-0x1.64c282b970a60298p+56L, -0x1.64c282b970a6029ap+56L,
383		-0x1.64c282b970a60298p+56L, ALL_STD_EXCEPT, FE_INEXACT);
384#elif LDBL_MANT_DIG == 53
385	testrnd(fmal, 0x1.31ad012ede8aap+100L, 0x1.2fbf79c839067p-42L,
386		-0x1.c3e106929056ep+58L, -0x1.64c282b970a5fp+56L,
387		-0x1.64c282b970a5ep+56L, -0x1.64c282b970a5fp+56L,
388		-0x1.64c282b970a5ep+56L, ALL_STD_EXCEPT, FE_INEXACT);
389#endif
390
391	/* x*y (rounded) ~= -z */
392	/* XXX spurious inexact exceptions */
393	testrnd(fmaf, 0x1.bbffeep-30, -0x1.1d164cp-74, 0x1.ee7296p-104,
394		-0x1.c46ea8p-128, -0x1.c46ea8p-128, -0x1.c46ea8p-128,
395		-0x1.c46ea8p-128, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
396	testrnd(fma, 0x1.bbffeea6fc7d6p-30, 0x1.1d164c6cbf078p-74,
397		-0x1.ee72993aff948p-104, -0x1.71f72ac7d9d8p-159,
398		-0x1.71f72ac7d9d8p-159, -0x1.71f72ac7d9d8p-159,
399		-0x1.71f72ac7d9d8p-159, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
400#if LDBL_MANT_DIG == 113
401	testrnd(fmal, 0x1.bbffeea6fc7d65927d147f437675p-30L,
402		0x1.1d164c6cbf078b7a22607d1cd6a2p-74L,
403		-0x1.ee72993aff94973876031bec0944p-104L,
404		0x1.64e086175b3a2adc36e607058814p-217L,
405		0x1.64e086175b3a2adc36e607058814p-217L,
406		0x1.64e086175b3a2adc36e607058814p-217L,
407		0x1.64e086175b3a2adc36e607058814p-217L,
408		ALL_STD_EXCEPT & ~FE_INEXACT, 0);
409#elif LDBL_MANT_DIG == 64
410	testrnd(fmal, 0x1.bbffeea6fc7d6592p-30L, 0x1.1d164c6cbf078b7ap-74L,
411		-0x1.ee72993aff949736p-104L, 0x1.af190e7a1ee6ad94p-168L,
412		0x1.af190e7a1ee6ad94p-168L, 0x1.af190e7a1ee6ad94p-168L,
413		0x1.af190e7a1ee6ad94p-168L, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
414#elif LDBL_MANT_DIG == 53
415	testrnd(fmal, 0x1.bbffeea6fc7d6p-30L, 0x1.1d164c6cbf078p-74L,
416		-0x1.ee72993aff948p-104L, -0x1.71f72ac7d9d8p-159L,
417		-0x1.71f72ac7d9d8p-159L, -0x1.71f72ac7d9d8p-159L,
418		-0x1.71f72ac7d9d8p-159L, ALL_STD_EXCEPT & ~FE_INEXACT, 0);
419#endif
420}
421
422static void
423test_double_rounding(void)
424{
425
426	/*
427	 *     a =  0x1.8000000000001p0
428	 *     b =  0x1.8000000000001p0
429	 *     c = -0x0.0000000000000000000000000080...1p+1
430	 * a * b =  0x1.2000000000001800000000000080p+1
431	 *
432	 * The correct behavior is to round DOWN to 0x1.2000000000001p+1 in
433	 * round-to-nearest mode.  An implementation that computes a*b+c in
434	 * double+double precision, however, will get 0x1.20000000000018p+1,
435	 * and then round UP.
436	 */
437	fesetround(FE_TONEAREST);
438	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
439	     -0x1.0000000000001p-104, 0x1.2000000000001p+1,
440	     ALL_STD_EXCEPT, FE_INEXACT);
441	fesetround(FE_DOWNWARD);
442	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
443	     -0x1.0000000000001p-104, 0x1.2000000000001p+1,
444	     ALL_STD_EXCEPT, FE_INEXACT);
445	fesetround(FE_UPWARD);
446	test(fma, 0x1.8000000000001p0, 0x1.8000000000001p0,
447	     -0x1.0000000000001p-104, 0x1.2000000000002p+1,
448	     ALL_STD_EXCEPT, FE_INEXACT);
449
450	fesetround(FE_TONEAREST);
451	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200002p+1,
452	     ALL_STD_EXCEPT, FE_INEXACT);
453	fesetround(FE_DOWNWARD);
454	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200002p+1,
455	     ALL_STD_EXCEPT, FE_INEXACT);
456	fesetround(FE_UPWARD);
457	test(fmaf, 0x1.800002p+0, 0x1.800002p+0, -0x1.000002p-46, 0x1.200004p+1,
458	     ALL_STD_EXCEPT, FE_INEXACT);
459
460	fesetround(FE_TONEAREST);
461#if LDBL_MANT_DIG == 64
462	test(fmal, 0x1.4p+0L, 0x1.0000000000000004p+0L, 0x1p-128L,
463	     0x1.4000000000000006p+0L, ALL_STD_EXCEPT, FE_INEXACT);
464#elif LDBL_MANT_DIG == 113
465	test(fmal, 0x1.8000000000000000000000000001p+0L,
466	     0x1.8000000000000000000000000001p+0L,
467	     -0x1.0000000000000000000000000001p-224L,
468	     0x1.2000000000000000000000000001p+1L, ALL_STD_EXCEPT, FE_INEXACT);
469#endif
470
471}
472
473int
474main(int argc, char *argv[])
475{
476	int rmodes[] = { FE_TONEAREST, FE_UPWARD, FE_DOWNWARD, FE_TOWARDZERO };
477	int i, j;
478
479	printf("1..19\n");
480
481	for (i = 0; i < nitems(rmodes); i++, j++) {
482		printf("rmode = %d\n", rmodes[i]);
483		fesetround(rmodes[i]);
484		test_zeroes();
485		printf("ok %d - fma zeroes\n", i + 1);
486	}
487
488	for (i = 0; i < nitems(rmodes); i++, j++) {
489		printf("rmode = %d\n", rmodes[i]);
490		fesetround(rmodes[i]);
491		test_infinities();
492		printf("ok %d - fma infinities\n", j);
493	}
494
495	fesetround(FE_TONEAREST);
496	test_nans();
497	printf("ok 9 - fma NaNs\n");
498
499	for (i = 0; i < nitems(rmodes); i++, j++) {
500		printf("rmode = %d\n", rmodes[i]);
501		fesetround(rmodes[i]);
502		test_small_z();
503		printf("ok %d - fma small z\n", j);
504	}
505
506	for (i = 0; i < nitems(rmodes); i++, j++) {
507		printf("rmode = %d\n", rmodes[i]);
508		fesetround(rmodes[i]);
509		test_big_z();
510		printf("ok %d - fma big z\n", j);
511	}
512
513	fesetround(FE_TONEAREST);
514	test_accuracy();
515	printf("ok %d - fma accuracy\n", j);
516	j++;
517
518	test_double_rounding();
519	printf("ok %d - fma double rounding\n", j);
520	j++;
521
522	/*
523	 * TODO:
524	 * - Tests for subnormals
525	 * - Cancellation tests (e.g., z = (double)x*y, but x*y is inexact)
526	 */
527
528	return (0);
529}
530