1 | /* Round to integer type. Common helper functions. |
2 | Copyright (C) 2016-2024 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 | <https://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <errno.h> |
20 | #include <fenv.h> |
21 | #include <float.h> |
22 | #include <math.h> |
23 | #include <math-barriers.h> |
24 | #include <stdbool.h> |
25 | #include <stdint.h> |
26 | |
27 | /* The including file should have defined UNSIGNED to 0 (signed return |
28 | type) or 1 (unsigned return type), INEXACT to 0 (no inexact |
29 | exceptions) or 1 (raise inexact exceptions) and RET_TYPE to the |
30 | return type (intmax_t or uintmax_t). */ |
31 | |
32 | /* Return the maximum unbiased exponent for an argument (negative if |
33 | NEGATIVE is set) that might be in range for a call to a fromfp |
34 | function with width WIDTH (greater than 0, and not exceeding that |
35 | of intmax_t). The truncated argument may still be out of range in |
36 | the case of negative arguments, and if not out of range it may |
37 | become out of range as a result of rounding. */ |
38 | |
39 | static int |
40 | fromfp_max_exponent (bool negative, int width) |
41 | { |
42 | if (UNSIGNED) |
43 | return negative ? -1 : width - 1; |
44 | else |
45 | return negative ? width - 1 : width - 2; |
46 | } |
47 | |
48 | /* Return the result of rounding an integer value X (passed as the |
49 | absolute value; NEGATIVE is true if the value is negative), where |
50 | HALF_BIT is true if the bit with value 0.5 is set and MORE_BITS is |
51 | true if any lower bits are set, in the rounding direction |
52 | ROUND. */ |
53 | |
54 | static uintmax_t |
55 | fromfp_round (bool negative, uintmax_t x, bool half_bit, bool more_bits, |
56 | int round) |
57 | { |
58 | switch (round) |
59 | { |
60 | case FP_INT_UPWARD: |
61 | return x + (!negative && (half_bit || more_bits)); |
62 | |
63 | case FP_INT_DOWNWARD: |
64 | return x + (negative && (half_bit || more_bits)); |
65 | |
66 | case FP_INT_TOWARDZERO: |
67 | default: |
68 | /* Unknown rounding directions are defined to mean unspecified |
69 | rounding; treat this as truncation. */ |
70 | return x; |
71 | |
72 | case FP_INT_TONEARESTFROMZERO: |
73 | return x + half_bit; |
74 | |
75 | case FP_INT_TONEAREST: |
76 | return x + (half_bit && ((x & 1) || more_bits)); |
77 | } |
78 | } |
79 | |
80 | /* Integer rounding, of a value whose exponent EXPONENT did not exceed |
81 | the maximum exponent MAX_EXPONENT and so did not necessarily |
82 | overflow, has produced X (possibly wrapping to 0); the sign is |
83 | negative if NEGATIVE is true. Return whether this overflowed the |
84 | allowed width. */ |
85 | |
86 | static bool |
87 | fromfp_overflowed (bool negative, uintmax_t x, int exponent, int max_exponent) |
88 | { |
89 | if (UNSIGNED) |
90 | { |
91 | if (negative) |
92 | return x != 0; |
93 | else if (max_exponent == INTMAX_WIDTH - 1) |
94 | return exponent == INTMAX_WIDTH - 1 && x == 0; |
95 | else |
96 | return x == (1ULL << (max_exponent + 1)); |
97 | } |
98 | else |
99 | { |
100 | if (negative) |
101 | return exponent == max_exponent && x != (1ULL << max_exponent); |
102 | else |
103 | return x == (1ULL << (max_exponent + 1)); |
104 | } |
105 | } |
106 | |
107 | /* Handle a domain error for a call to a fromfp function with an |
108 | argument which is negative if NEGATIVE is set, and specified width |
109 | (not exceeding that of intmax_t) WIDTH. The return value is |
110 | unspecified (with it being unclear if the result needs to fit |
111 | within WIDTH bits in this case); we choose to saturate to the given |
112 | number of bits (treating NaNs like any other value). */ |
113 | |
114 | static RET_TYPE |
115 | fromfp_domain_error (bool negative, unsigned int width) |
116 | { |
117 | feraiseexcept (FE_INVALID); |
118 | __set_errno (EDOM); |
119 | /* The return value is unspecified; we choose to saturate to the |
120 | given number of bits (treating NaNs like any other value). */ |
121 | if (UNSIGNED) |
122 | { |
123 | if (negative) |
124 | return 0; |
125 | else if (width == INTMAX_WIDTH) |
126 | return -1; |
127 | else |
128 | return (1ULL << width) - 1; |
129 | } |
130 | else |
131 | { |
132 | if (width == 0) |
133 | return 0; |
134 | else if (negative) |
135 | return -(1ULL << (width - 1)); |
136 | else |
137 | return (1ULL << (width - 1)) - 1; |
138 | } |
139 | } |
140 | |
141 | /* Given X, the absolute value of a floating-point number (negative if |
142 | NEGATIVE is set) truncated towards zero, where HALF_BIT is true if |
143 | the bit with value 0.5 is set and MORE_BITS is true if any lower |
144 | bits are set, round it in the rounding direction ROUND, handle |
145 | errors and exceptions and return the appropriate return value for a |
146 | fromfp function. X originally had floating-point exponent |
147 | EXPONENT, which does not exceed MAX_EXPONENT, the return value from |
148 | fromfp_max_exponent with width WIDTH. */ |
149 | |
150 | static RET_TYPE |
151 | fromfp_round_and_return (bool negative, uintmax_t x, bool half_bit, |
152 | bool more_bits, int round, int exponent, |
153 | int max_exponent, unsigned int width) |
154 | { |
155 | uintmax_t uret = fromfp_round (negative, x, half_bit, more_bits, round); |
156 | if (fromfp_overflowed (negative, x: uret, exponent, max_exponent)) |
157 | return fromfp_domain_error (negative, width); |
158 | |
159 | if (INEXACT && (half_bit || more_bits)) |
160 | { |
161 | /* There is no need for this to use the specific floating-point |
162 | type for which this header is included, and there is no need |
163 | for this header to know that type at all, so just use float |
164 | here. */ |
165 | float force_inexact = 1.0f + FLT_MIN; |
166 | math_force_eval (force_inexact); |
167 | } |
168 | if (UNSIGNED) |
169 | /* A negative argument not rounding to zero will already have |
170 | produced a domain error. */ |
171 | return uret; |
172 | else |
173 | return negative ? -uret : uret; |
174 | } |
175 | |