1 | //===-- strtofloat_fuzz.cpp -----------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// |
9 | /// Fuzzing test for llvm-libc atof implementation. |
10 | /// |
11 | //===----------------------------------------------------------------------===// |
12 | #include "src/stdlib/atof.h" |
13 | #include "src/stdlib/strtod.h" |
14 | #include "src/stdlib/strtof.h" |
15 | #include "src/stdlib/strtold.h" |
16 | |
17 | #include "src/__support/FPUtil/FPBits.h" |
18 | |
19 | #include "hdr/math_macros.h" |
20 | #include <stddef.h> |
21 | #include <stdint.h> |
22 | |
23 | #include "utils/MPFRWrapper/mpfr_inc.h" |
24 | |
25 | using LIBC_NAMESPACE::fputil::FPBits; |
26 | |
27 | // This function calculates the effective precision for a given float type and |
28 | // exponent. Subnormals have a lower effective precision since they don't |
29 | // necessarily use all of the bits of the mantissa. |
30 | template <typename F> inline constexpr int effective_precision(int exponent) { |
31 | const int full_precision = FPBits<F>::FRACTION_LEN + 1; |
32 | |
33 | // This is intended to be 0 when the exponent is the lowest normal and |
34 | // increase as the exponent's magnitude increases. |
35 | const int bits_below_normal = (-exponent) - (FPBits<F>::EXP_BIAS - 1); |
36 | |
37 | // The precision should be the normal, full precision, minus the bits lost |
38 | // by this being a subnormal, minus one for the implicit leading one. |
39 | const int bits_if_subnormal = full_precision - bits_below_normal - 1; |
40 | |
41 | if (bits_below_normal >= 0) { |
42 | return bits_if_subnormal; |
43 | } |
44 | return full_precision; |
45 | } |
46 | |
47 | extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) { |
48 | // const char newstr[] = "123"; |
49 | // data = reinterpret_cast<const uint8_t *>(newstr); |
50 | // size = sizeof(newstr); |
51 | uint8_t *container = new uint8_t[size + 1]; |
52 | if (!container) |
53 | __builtin_trap(); |
54 | size_t i; |
55 | |
56 | for (i = 0; i < size; ++i) { |
57 | // MPFR's strtofr uses "@" as a base-independent exponent symbol |
58 | if (data[i] != '@') |
59 | container[i] = data[i]; |
60 | else { |
61 | container[i] = '#'; |
62 | } |
63 | } |
64 | container[size] = '\0'; // Add null terminator to container. |
65 | |
66 | const char *str_ptr = reinterpret_cast<const char *>(container); |
67 | |
68 | char *out_ptr = nullptr; |
69 | |
70 | size_t base = 0; |
71 | |
72 | // This is just used to determine the base and precision. |
73 | mpfr_t result; |
74 | mpfr_init2(result, 256); |
75 | mpfr_t bin_result; |
76 | mpfr_init2(bin_result, 256); |
77 | mpfr_strtofr(result, str_ptr, &out_ptr, 0 /* base */, MPFR_RNDN); |
78 | ptrdiff_t result_strlen = out_ptr - str_ptr; |
79 | mpfr_strtofr(bin_result, str_ptr, &out_ptr, 2 /* base */, MPFR_RNDN); |
80 | ptrdiff_t bin_result_strlen = out_ptr - str_ptr; |
81 | |
82 | long double bin_result_ld = mpfr_get_ld(bin_result, MPFR_RNDN); |
83 | long double result_ld = mpfr_get_ld(result, MPFR_RNDN); |
84 | |
85 | // This detects if mpfr's strtofr selected a base of 2, which libc does not |
86 | // support. If a base 2 decoding is detected, it is replaced by a base 10 |
87 | // decoding. |
88 | if ((bin_result_ld != 0.0 || bin_result_strlen == result_strlen) && |
89 | bin_result_ld == result_ld) { |
90 | mpfr_strtofr(result, str_ptr, &out_ptr, 10 /* base */, MPFR_RNDN); |
91 | result_strlen = out_ptr - str_ptr; |
92 | base = 10; |
93 | } |
94 | |
95 | auto result_exp = mpfr_get_exp(result); |
96 | |
97 | mpfr_clear(result); |
98 | mpfr_clear(bin_result); |
99 | |
100 | // These must be calculated with the correct precision, and not any more, to |
101 | // prevent numbers like 66336650.00...01 (many zeroes) from causing an issue. |
102 | // 66336650 is exactly between two float values (66336652 and 66336648) so the |
103 | // correct float result for 66336650.00...01 is rounding up to 66336652. The |
104 | // correct double is instead 66336650, which when converted to float is |
105 | // rounded down to 66336648. This means we have to compare against the correct |
106 | // precision to get the correct result. |
107 | |
108 | // TODO: Add support for other rounding modes. |
109 | int float_precision = effective_precision<float>(exponent: result_exp); |
110 | if (float_precision >= 2) { |
111 | mpfr_t mpfr_float; |
112 | mpfr_init2(mpfr_float, float_precision); |
113 | mpfr_strtofr(mpfr_float, str_ptr, &out_ptr, base, MPFR_RNDN); |
114 | float volatile float_result = mpfr_get_flt(mpfr_float, MPFR_RNDN); |
115 | auto volatile strtof_result = LIBC_NAMESPACE::strtof(str: str_ptr, str_end: &out_ptr); |
116 | ptrdiff_t strtof_strlen = out_ptr - str_ptr; |
117 | if (result_strlen != strtof_strlen) |
118 | __builtin_trap(); |
119 | // If any result is NaN, all of them should be NaN. We can't use the usual |
120 | // comparisons because NaN != NaN. |
121 | if (isnan(x: float_result) ^ isnan(x: strtof_result)) |
122 | __builtin_trap(); |
123 | if (!isnan(x: float_result) && float_result != strtof_result) |
124 | __builtin_trap(); |
125 | mpfr_clear(mpfr_float); |
126 | } |
127 | |
128 | int double_precision = effective_precision<double>(exponent: result_exp); |
129 | if (double_precision >= 2) { |
130 | mpfr_t mpfr_double; |
131 | mpfr_init2(mpfr_double, double_precision); |
132 | mpfr_strtofr(mpfr_double, str_ptr, &out_ptr, base, MPFR_RNDN); |
133 | double volatile double_result = mpfr_get_d(mpfr_double, MPFR_RNDN); |
134 | auto volatile strtod_result = LIBC_NAMESPACE::strtod(str: str_ptr, str_end: &out_ptr); |
135 | auto volatile atof_result = LIBC_NAMESPACE::atof(str: str_ptr); |
136 | ptrdiff_t strtod_strlen = out_ptr - str_ptr; |
137 | if (result_strlen != strtod_strlen) |
138 | __builtin_trap(); |
139 | if (isnan(x: double_result) ^ isnan(x: strtod_result) || |
140 | isnan(x: double_result) ^ isnan(x: atof_result)) |
141 | __builtin_trap(); |
142 | if (!isnan(x: double_result) && |
143 | (double_result != strtod_result || double_result != atof_result)) |
144 | __builtin_trap(); |
145 | mpfr_clear(mpfr_double); |
146 | } |
147 | |
148 | int long_double_precision = effective_precision<long double>(exponent: result_exp); |
149 | if (long_double_precision >= 2) { |
150 | mpfr_t mpfr_long_double; |
151 | mpfr_init2(mpfr_long_double, long_double_precision); |
152 | mpfr_strtofr(mpfr_long_double, str_ptr, &out_ptr, base, MPFR_RNDN); |
153 | long double volatile long_double_result = |
154 | mpfr_get_ld(mpfr_long_double, MPFR_RNDN); |
155 | auto volatile strtold_result = LIBC_NAMESPACE::strtold(str: str_ptr, str_end: &out_ptr); |
156 | ptrdiff_t strtold_strlen = out_ptr - str_ptr; |
157 | if (result_strlen != strtold_strlen) |
158 | __builtin_trap(); |
159 | if (isnan(x: long_double_result) ^ isnan(x: strtold_result)) |
160 | __builtin_trap(); |
161 | if (!isnan(x: long_double_result) && long_double_result != strtold_result) |
162 | __builtin_trap(); |
163 | mpfr_clear(mpfr_long_double); |
164 | } |
165 | |
166 | delete[] container; |
167 | return 0; |
168 | } |
169 | |