xref: /qemu/include/fpu/softfloat-types.h (revision 09951f5a27a7f8633118c1808cf17e66b30c3c62)
1 /*
2  * QEMU float support
3  *
4  * The code in this source file is derived from release 2a of the SoftFloat
5  * IEC/IEEE Floating-point Arithmetic Package. Those parts of the code (and
6  * some later contributions) are provided under that license, as detailed below.
7  * It has subsequently been modified by contributors to the QEMU Project,
8  * so some portions are provided under:
9  *  the SoftFloat-2a license
10  *  the BSD license
11  *  GPL-v2-or-later
12  *
13  * This header holds definitions for code that might be dealing with
14  * softfloat types but not need access to the actual library functions.
15  */
16 /*
17 ===============================================================================
18 This C header file is part of the SoftFloat IEC/IEEE Floating-point
19 Arithmetic Package, Release 2a.
20 
21 Written by John R. Hauser.  This work was made possible in part by the
22 International Computer Science Institute, located at Suite 600, 1947 Center
23 Street, Berkeley, California 94704.  Funding was partially provided by the
24 National Science Foundation under grant MIP-9311980.  The original version
25 of this code was written as part of a project to build a fixed-point vector
26 processor in collaboration with the University of California at Berkeley,
27 overseen by Profs. Nelson Morgan and John Wawrzynek.  More information
28 is available through the Web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
29 arithmetic/SoftFloat.html'.
30 
31 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE.  Although reasonable effort
32 has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
33 TIMES RESULT IN INCORRECT BEHAVIOR.  USE OF THIS SOFTWARE IS RESTRICTED TO
34 PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
35 AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
36 
37 Derivative works are acceptable, even for commercial purposes, so long as
38 (1) they include prominent notice that the work is derivative, and (2) they
39 include prominent notice akin to these four paragraphs for those parts of
40 this code that are retained.
41 
42 ===============================================================================
43 */
44 
45 /* BSD licensing:
46  * Copyright (c) 2006, Fabrice Bellard
47  * All rights reserved.
48  *
49  * Redistribution and use in source and binary forms, with or without
50  * modification, are permitted provided that the following conditions are met:
51  *
52  * 1. Redistributions of source code must retain the above copyright notice,
53  * this list of conditions and the following disclaimer.
54  *
55  * 2. Redistributions in binary form must reproduce the above copyright notice,
56  * this list of conditions and the following disclaimer in the documentation
57  * and/or other materials provided with the distribution.
58  *
59  * 3. Neither the name of the copyright holder nor the names of its contributors
60  * may be used to endorse or promote products derived from this software without
61  * specific prior written permission.
62  *
63  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
64  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
65  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
66  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
67  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
68  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
69  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
70  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
71  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
72  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
73  * THE POSSIBILITY OF SUCH DAMAGE.
74  */
75 
76 /* Portions of this work are licensed under the terms of the GNU GPL,
77  * version 2 or later. See the COPYING file in the top-level directory.
78  */
79 
80 #ifndef SOFTFLOAT_TYPES_H
81 #define SOFTFLOAT_TYPES_H
82 
83 #include "hw/registerfields.h"
84 
85 /*
86  * Software IEC/IEEE floating-point types.
87  */
88 
89 typedef uint16_t float16;
90 typedef uint32_t float32;
91 typedef uint64_t float64;
92 #define float16_val(x) (x)
93 #define float32_val(x) (x)
94 #define float64_val(x) (x)
95 #define make_float16(x) (x)
96 #define make_float32(x) (x)
97 #define make_float64(x) (x)
98 #define const_float16(x) (x)
99 #define const_float32(x) (x)
100 #define const_float64(x) (x)
101 typedef struct {
102     uint64_t low;
103     uint16_t high;
104 } floatx80;
105 #define make_floatx80(exp, mant) ((floatx80) { mant, exp })
106 #define make_floatx80_init(exp, mant) { .low = mant, .high = exp }
107 typedef struct {
108 #if HOST_BIG_ENDIAN
109     uint64_t high, low;
110 #else
111     uint64_t low, high;
112 #endif
113 } float128;
114 #define make_float128(high_, low_) ((float128) { .high = high_, .low = low_ })
115 #define make_float128_init(high_, low_) { .high = high_, .low = low_ }
116 
117 /*
118  * Software neural-network floating-point types.
119  */
120 typedef uint16_t bfloat16;
121 
122 /*
123  * Software IEC/IEEE floating-point underflow tininess-detection mode.
124  */
125 
126 #define float_tininess_after_rounding  false
127 #define float_tininess_before_rounding true
128 
129 /*
130  *Software IEC/IEEE floating-point rounding mode.
131  */
132 
133 typedef enum __attribute__((__packed__)) {
134     float_round_nearest_even = 0,
135     float_round_down         = 1,
136     float_round_up           = 2,
137     float_round_to_zero      = 3,
138     float_round_ties_away    = 4,
139     /* Not an IEEE rounding mode: round to closest odd, overflow to max */
140     float_round_to_odd       = 5,
141     /* Not an IEEE rounding mode: round to closest odd, overflow to inf */
142     float_round_to_odd_inf   = 6,
143     /* Not an IEEE rounding mode: round to nearest even, overflow to max */
144     float_round_nearest_even_max = 7,
145 } FloatRoundMode;
146 
147 /*
148  * Software IEC/IEEE floating-point exception flags.
149  */
150 
151 enum {
152     float_flag_invalid         = 0x0001,
153     float_flag_divbyzero       = 0x0002,
154     float_flag_overflow        = 0x0004,
155     float_flag_underflow       = 0x0008,
156     float_flag_inexact         = 0x0010,
157     /* We flushed an input denormal to 0 (because of flush_inputs_to_zero) */
158     float_flag_input_denormal_flushed = 0x0020,
159     /* We flushed an output denormal to 0 (because of flush_to_zero) */
160     float_flag_output_denormal_flushed = 0x0040,
161     float_flag_invalid_isi     = 0x0080,  /* inf - inf */
162     float_flag_invalid_imz     = 0x0100,  /* inf * 0 */
163     float_flag_invalid_idi     = 0x0200,  /* inf / inf */
164     float_flag_invalid_zdz     = 0x0400,  /* 0 / 0 */
165     float_flag_invalid_sqrt    = 0x0800,  /* sqrt(-x) */
166     float_flag_invalid_cvti    = 0x1000,  /* non-nan to integer */
167     float_flag_invalid_snan    = 0x2000,  /* any operand was snan */
168     /*
169      * An input was denormal and we used it (without flushing it to zero).
170      * Not set if we do not actually use the denormal input (e.g.
171      * because some other input was a NaN, or because the operation
172      * wasn't actually carried out (divide-by-zero; invalid))
173      */
174     float_flag_input_denormal_used = 0x4000,
175 };
176 
177 /*
178  * Rounding precision for floatx80.
179  */
180 typedef enum __attribute__((__packed__)) {
181     floatx80_precision_x,
182     floatx80_precision_d,
183     floatx80_precision_s,
184 } FloatX80RoundPrec;
185 
186 /*
187  * 2-input NaN propagation rule. Individual architectures have
188  * different rules for which input NaN is propagated to the output
189  * when there is more than one NaN on the input.
190  *
191  * If default_nan_mode is enabled then it is valid not to set a
192  * NaN propagation rule, because the softfloat code guarantees
193  * not to try to pick a NaN to propagate in default NaN mode.
194  * When not in default-NaN mode, it is an error for the target
195  * not to set the rule in float_status, and we will assert if
196  * we need to handle an input NaN and no rule was selected.
197  */
198 typedef enum __attribute__((__packed__)) {
199     /* No propagation rule specified */
200     float_2nan_prop_none = 0,
201     /* Prefer SNaN over QNaN, then operand A over B */
202     float_2nan_prop_s_ab,
203     /* Prefer SNaN over QNaN, then operand B over A */
204     float_2nan_prop_s_ba,
205     /* Prefer A over B regardless of SNaN vs QNaN */
206     float_2nan_prop_ab,
207     /* Prefer B over A regardless of SNaN vs QNaN */
208     float_2nan_prop_ba,
209     /*
210      * This implements x87 NaN propagation rules:
211      * SNaN + QNaN => return the QNaN
212      * two SNaNs => return the one with the larger significand, silenced
213      * two QNaNs => return the one with the larger significand
214      * SNaN and a non-NaN => return the SNaN, silenced
215      * QNaN and a non-NaN => return the QNaN
216      *
217      * If we get down to comparing significands and they are the same,
218      * return the NaN with the positive sign bit (if any).
219      */
220     float_2nan_prop_x87,
221 } Float2NaNPropRule;
222 
223 /*
224  * 3-input NaN propagation rule, for fused multiply-add. Individual
225  * architectures have different rules for which input NaN is
226  * propagated to the output when there is more than one NaN on the
227  * input.
228  *
229  * If default_nan_mode is enabled then it is valid not to set a NaN
230  * propagation rule, because the softfloat code guarantees not to try
231  * to pick a NaN to propagate in default NaN mode.  When not in
232  * default-NaN mode, it is an error for the target not to set the rule
233  * in float_status if it uses a muladd, and we will assert if we need
234  * to handle an input NaN and no rule was selected.
235  *
236  * The naming scheme for Float3NaNPropRule values is:
237  *  float_3nan_prop_s_abc:
238  *    = "Prefer SNaN over QNaN, then operand A over B over C"
239  *  float_3nan_prop_abc:
240  *    = "Prefer A over B over C regardless of SNaN vs QNAN"
241  *
242  * For QEMU, the multiply-add operation is A * B + C.
243  */
244 
245 /*
246  * We set the Float3NaNPropRule enum values up so we can select the
247  * right value in pickNaNMulAdd in a data driven way.
248  */
249 FIELD(3NAN, 1ST, 0, 2)   /* which operand is most preferred ? */
250 FIELD(3NAN, 2ND, 2, 2)   /* which operand is next most preferred ? */
251 FIELD(3NAN, 3RD, 4, 2)   /* which operand is least preferred ? */
252 FIELD(3NAN, SNAN, 6, 1)  /* do we prefer SNaN over QNaN ? */
253 
254 #define PROPRULE(X, Y, Z) \
255     ((X << R_3NAN_1ST_SHIFT) | (Y << R_3NAN_2ND_SHIFT) | (Z << R_3NAN_3RD_SHIFT))
256 
257 typedef enum __attribute__((__packed__)) {
258     float_3nan_prop_none = 0,     /* No propagation rule specified */
259     float_3nan_prop_abc = PROPRULE(0, 1, 2),
260     float_3nan_prop_acb = PROPRULE(0, 2, 1),
261     float_3nan_prop_bac = PROPRULE(1, 0, 2),
262     float_3nan_prop_bca = PROPRULE(1, 2, 0),
263     float_3nan_prop_cab = PROPRULE(2, 0, 1),
264     float_3nan_prop_cba = PROPRULE(2, 1, 0),
265     float_3nan_prop_s_abc = float_3nan_prop_abc | R_3NAN_SNAN_MASK,
266     float_3nan_prop_s_acb = float_3nan_prop_acb | R_3NAN_SNAN_MASK,
267     float_3nan_prop_s_bac = float_3nan_prop_bac | R_3NAN_SNAN_MASK,
268     float_3nan_prop_s_bca = float_3nan_prop_bca | R_3NAN_SNAN_MASK,
269     float_3nan_prop_s_cab = float_3nan_prop_cab | R_3NAN_SNAN_MASK,
270     float_3nan_prop_s_cba = float_3nan_prop_cba | R_3NAN_SNAN_MASK,
271 } Float3NaNPropRule;
272 
273 #undef PROPRULE
274 
275 /*
276  * Rule for result of fused multiply-add 0 * Inf + NaN.
277  * This must be a NaN, but implementations differ on whether this
278  * is the input NaN or the default NaN.
279  *
280  * You don't need to set this if default_nan_mode is enabled.
281  * When not in default-NaN mode, it is an error for the target
282  * not to set the rule in float_status if it uses muladd, and we
283  * will assert if we need to handle an input NaN and no rule was
284  * selected.
285  */
286 typedef enum __attribute__((__packed__)) {
287     /* No propagation rule specified */
288     float_infzeronan_none = 0,
289     /* Result is never the default NaN (so always the input NaN) */
290     float_infzeronan_dnan_never = 1,
291     /* Result is always the default NaN */
292     float_infzeronan_dnan_always = 2,
293     /* Result is the default NaN if the input NaN is quiet */
294     float_infzeronan_dnan_if_qnan = 3,
295     /*
296      * Don't raise Invalid for 0 * Inf + NaN. Default is to raise.
297      * IEEE 754-2008 section 7.2 makes it implementation defined whether
298      * 0 * Inf + QNaN raises Invalid or not. Note that 0 * Inf + SNaN will
299      * raise the Invalid flag for the SNaN anyway.
300      *
301      * This is a flag which can be ORed in with any of the above
302      * DNaN behaviour options.
303      */
304     float_infzeronan_suppress_invalid = (1 << 7),
305 } FloatInfZeroNaNRule;
306 
307 /*
308  * When flush_to_zero is set, should we detect denormal results to
309  * be flushed before or after rounding? For most architectures this
310  * should be set to match the tininess_before_rounding setting,
311  * but a few architectures, e.g. MIPS MSA, detect FTZ before
312  * rounding but tininess after rounding.
313  *
314  * This enum is arranged so that the default if the target doesn't
315  * configure it matches the default for tininess_before_rounding
316  * (i.e. "after rounding").
317  */
318 typedef enum __attribute__((__packed__)) {
319     float_ftz_after_rounding = 0,
320     float_ftz_before_rounding = 1,
321 } FloatFTZDetection;
322 
323 /*
324  * floatx80 is primarily used by x86 and m68k, and there are
325  * differences in the handling, largely related to the explicit
326  * Integer bit which floatx80 has and the other float formats do not.
327  * These flag values allow specification of the target's requirements
328  * and can be ORed together to set floatx80_behaviour.
329  */
330 typedef enum __attribute__((__packed__)) {
331     /* In the default Infinity value, is the Integer bit 0 ? */
332     floatx80_default_inf_int_bit_is_zero = 1,
333     /*
334      * Are Pseudo-infinities (Inf with the Integer bit zero) valid?
335      * If so, floatx80_is_infinity() will return true for them.
336      * If not, floatx80_invalid_encoding will return false for them,
337      * and using them as inputs to a float op will raise Invalid.
338      */
339     floatx80_pseudo_inf_valid = 2,
340     /*
341      * Are Pseudo-NaNs (NaNs where the Integer bit is zero) valid?
342      * If not, floatx80_invalid_encoding() will return false for them,
343      * and using them as inputs to a float op will raise Invalid.
344      */
345     floatx80_pseudo_nan_valid = 4,
346     /*
347      * Are Unnormals (0 < exp < 0x7fff, Integer bit zero) valid?
348      * If not, floatx80_invalid_encoding() will return false for them,
349      * and using them as inputs to a float op will raise Invalid.
350      */
351     floatx80_unnormal_valid = 8,
352 
353     /*
354      * If the exponent is 0 and the Integer bit is set, Intel call
355      * this a "pseudo-denormal"; x86 supports that only on input
356      * (treating them as denormals by ignoring the Integer bit).
357      * For m68k, the integer bit is considered validly part of the
358      * input value when the exponent is 0, and may be 0 or 1,
359      * giving extra range. They may also be generated as outputs.
360      * (The m68k manual actually calls these values part of the
361      * normalized number range, not the denormalized number range.)
362      *
363      * By default you get the Intel behaviour where the Integer
364      * bit is ignored; if this is set then the Integer bit value
365      * is honoured, m68k-style.
366      *
367      * Either way, floatx80_invalid_encoding() will always accept
368      * pseudo-denormals.
369      */
370     floatx80_pseudo_denormal_valid = 16,
371 } FloatX80Behaviour;
372 
373 /*
374  * Floating Point Status. Individual architectures may maintain
375  * several versions of float_status for different functions. The
376  * correct status for the operation is then passed by reference to
377  * most of the softfloat functions.
378  */
379 
380 typedef struct float_status {
381     uint16_t float_exception_flags;
382     FloatRoundMode float_rounding_mode;
383     FloatX80RoundPrec floatx80_rounding_precision;
384     FloatX80Behaviour floatx80_behaviour;
385     Float2NaNPropRule float_2nan_prop_rule;
386     Float3NaNPropRule float_3nan_prop_rule;
387     FloatInfZeroNaNRule float_infzeronan_rule;
388     bool tininess_before_rounding;
389     /* should denormalised results go to zero and set output_denormal_flushed? */
390     bool flush_to_zero;
391     /* do we detect and flush denormal results before or after rounding? */
392     FloatFTZDetection ftz_detection;
393     /* should denormalised inputs go to zero and set input_denormal_flushed? */
394     bool flush_inputs_to_zero;
395     bool default_nan_mode;
396     /*
397      * The pattern to use for the default NaN. Here the high bit specifies
398      * the default NaN's sign bit, and bits 6..0 specify the high bits of the
399      * fractional part. The low bits of the fractional part are copies of bit 0.
400      * The exponent of the default NaN is (as for any NaN) always all 1s.
401      * Note that a value of 0 here is not a valid NaN. The target must set
402      * this to the correct non-zero value, or we will assert when trying to
403      * create a default NaN.
404      */
405     uint8_t default_nan_pattern;
406     /*
407      * The flags below are not used on all specializations and may
408      * constant fold away (see snan_bit_is_one()/no_signalling_nans() in
409      * softfloat-specialize.inc.c)
410      */
411     bool snan_bit_is_one;
412     bool no_signaling_nans;
413     /* should overflowed results subtract re_bias to its exponent? */
414     bool rebias_overflow;
415     /* should underflowed results add re_bias to its exponent? */
416     bool rebias_underflow;
417 } float_status;
418 
419 #endif /* SOFTFLOAT_TYPES_H */
420