| 1 | //===-- strtofloatingpoint comparison test --------------------------------===// |
| 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 | #include "src/__support/CPP/bit.h" |
| 10 | #include "src/stdio/fclose.h" |
| 11 | #include "src/stdio/fgets.h" |
| 12 | #include "src/stdio/fopen.h" |
| 13 | #include "src/stdio/printf.h" |
| 14 | #include "src/stdlib/getenv.h" |
| 15 | #include "src/stdlib/strtod.h" |
| 16 | #include "src/stdlib/strtof.h" |
| 17 | #include "src/string/strdup.h" |
| 18 | #include "src/string/strtok.h" |
| 19 | #include "test/UnitTest/Test.h" |
| 20 | #include <stdint.h> |
| 21 | |
| 22 | // The intent of this test is to read in files in the format used in this test |
| 23 | // dataset: https://github.com/nigeltao/parse-number-fxx-test-data |
| 24 | // The format is as follows: |
| 25 | // Hexadecimal representations of IEEE754 floats in 16 bits, 32 bits, and 64 |
| 26 | // bits, then the string that matches to them. |
| 27 | |
| 28 | // 3C00 3F800000 3FF0000000000000 1.0 |
| 29 | |
| 30 | // By default, float_comp_in.txt is used as the test set, but once built this |
| 31 | // file can be run against the larger test set. To do that, clone the repository |
| 32 | // with the dataset, then navigate to the compiled binary of this file (it |
| 33 | // should be in llvm_project/build/bin). Run the following command: |
| 34 | // ./libc_str_to_float_comparison_test <path/to/dataset/repo>/data/* |
| 35 | // It will take a few seconds to run. |
| 36 | |
| 37 | struct ParseResult { |
| 38 | uint32_t totalFails; |
| 39 | uint32_t totalBitDiffs; |
| 40 | uint32_t detailedBitDiffs[4]; |
| 41 | uint32_t total; |
| 42 | }; |
| 43 | |
| 44 | enum class ParseStatus : uint8_t { |
| 45 | SUCCESS, |
| 46 | FILE_ERROR, |
| 47 | PARSE_ERROR, |
| 48 | }; |
| 49 | |
| 50 | static inline uint32_t hexCharToU32(char in) { |
| 51 | return in > '9' ? in + 10 - 'A' : in - '0'; |
| 52 | } |
| 53 | |
| 54 | // Fast because it assumes inStr points to exactly 8 uppercase hex chars |
| 55 | static inline uint32_t fastHexToU32(const char *inStr) { |
| 56 | uint32_t result = 0; |
| 57 | result = (hexCharToU32(in: inStr[0]) << 28) + (hexCharToU32(in: inStr[1]) << 24) + |
| 58 | (hexCharToU32(in: inStr[2]) << 20) + (hexCharToU32(in: inStr[3]) << 16) + |
| 59 | (hexCharToU32(in: inStr[4]) << 12) + (hexCharToU32(in: inStr[5]) << 8) + |
| 60 | (hexCharToU32(in: inStr[6]) << 4) + hexCharToU32(in: inStr[7]); |
| 61 | return result; |
| 62 | } |
| 63 | |
| 64 | // Fast because it assumes inStr points to exactly 8 uppercase hex chars |
| 65 | static inline uint64_t fastHexToU64(const char *inStr) { |
| 66 | uint64_t result = 0; |
| 67 | result = (static_cast<uint64_t>(fastHexToU32(inStr)) << 32) + |
| 68 | fastHexToU32(inStr: inStr + 8); |
| 69 | return result; |
| 70 | } |
| 71 | |
| 72 | static void parseLine(const char *line, ParseResult &parseResult, |
| 73 | int32_t &curFails, int32_t &curBitDiffs) { |
| 74 | |
| 75 | if (line[0] == '#') |
| 76 | return; |
| 77 | |
| 78 | parseResult.total += 1; |
| 79 | uint32_t expectedFloatRaw; |
| 80 | uint64_t expectedDoubleRaw; |
| 81 | |
| 82 | expectedFloatRaw = fastHexToU32(inStr: line + 5); |
| 83 | expectedDoubleRaw = fastHexToU64(inStr: line + 14); |
| 84 | |
| 85 | const char *num = line + 31; |
| 86 | |
| 87 | float floatResult = LIBC_NAMESPACE::strtof(num, nullptr); |
| 88 | |
| 89 | double doubleResult = LIBC_NAMESPACE::strtod(num, nullptr); |
| 90 | |
| 91 | uint32_t floatRaw = LIBC_NAMESPACE::cpp::bit_cast<uint32_t>(floatResult); |
| 92 | |
| 93 | uint64_t doubleRaw = LIBC_NAMESPACE::cpp::bit_cast<uint64_t>(doubleResult); |
| 94 | |
| 95 | if (!(expectedFloatRaw == floatRaw)) { |
| 96 | if (expectedFloatRaw == floatRaw + 1 || expectedFloatRaw == floatRaw - 1) { |
| 97 | curBitDiffs++; |
| 98 | if (expectedFloatRaw == floatRaw + 1) { |
| 99 | parseResult.detailedBitDiffs[0] = |
| 100 | parseResult.detailedBitDiffs[0] + 1; // float low |
| 101 | } else { |
| 102 | parseResult.detailedBitDiffs[1] = |
| 103 | parseResult.detailedBitDiffs[1] + 1; // float high |
| 104 | } |
| 105 | } else { |
| 106 | curFails++; |
| 107 | } |
| 108 | if (curFails + curBitDiffs < 10) { |
| 109 | LIBC_NAMESPACE::printf("Float fail for '%s'. Expected %x but got %x\n" , |
| 110 | num, expectedFloatRaw, floatRaw); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | if (!(expectedDoubleRaw == doubleRaw)) { |
| 115 | if (expectedDoubleRaw == doubleRaw + 1 || |
| 116 | expectedDoubleRaw == doubleRaw - 1) { |
| 117 | curBitDiffs++; |
| 118 | if (expectedDoubleRaw == doubleRaw + 1) { |
| 119 | parseResult.detailedBitDiffs[2] = |
| 120 | parseResult.detailedBitDiffs[2] + 1; // double low |
| 121 | } else { |
| 122 | parseResult.detailedBitDiffs[3] = |
| 123 | parseResult.detailedBitDiffs[3] + 1; // double high |
| 124 | } |
| 125 | } else { |
| 126 | curFails++; |
| 127 | } |
| 128 | if (curFails + curBitDiffs < 10) { |
| 129 | LIBC_NAMESPACE::printf("Double fail for '%s'. Expected %lx but got %lx\n" , |
| 130 | num, expectedDoubleRaw, doubleRaw); |
| 131 | } |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | ParseStatus checkBuffer(ParseResult &parseResult) { |
| 136 | constexpr const char *LINES[] = { |
| 137 | "3C00 3F800000 3FF0000000000000 1" , |
| 138 | "3D00 3FA00000 3FF4000000000000 1.25" , |
| 139 | "3D9A 3FB33333 3FF6666666666666 1.4" , |
| 140 | "57B7 42F6E979 405EDD2F1A9FBE77 123.456" , |
| 141 | "622A 44454000 4088A80000000000 789" , |
| 142 | "7C00 7F800000 7FF0000000000000 123.456e789" }; |
| 143 | |
| 144 | int32_t curFails = 0; // Only counts actual failures, not bitdiffs. |
| 145 | int32_t curBitDiffs = 0; // A bitdiff is when the expected result and actual |
| 146 | // result are off by +/- 1 bit. |
| 147 | |
| 148 | for (uint8_t i = 0; i < sizeof(LINES) / sizeof(LINES[0]); i++) { |
| 149 | parseLine(line: LINES[i], parseResult, curFails, curBitDiffs); |
| 150 | } |
| 151 | |
| 152 | parseResult.totalBitDiffs += curBitDiffs; |
| 153 | parseResult.totalFails += curFails; |
| 154 | |
| 155 | if (curFails > 1 || curBitDiffs > 1) { |
| 156 | return ParseStatus::PARSE_ERROR; |
| 157 | } |
| 158 | return ParseStatus::SUCCESS; |
| 159 | } |
| 160 | |
| 161 | ParseStatus checkFile(char *inputFileName, ParseResult &parseResult) { |
| 162 | int32_t curFails = 0; // Only counts actual failures, not bitdiffs. |
| 163 | int32_t curBitDiffs = 0; // A bitdiff is when the expected result and actual |
| 164 | // result are off by +/- 1 bit. |
| 165 | char line[1000]; |
| 166 | |
| 167 | auto *fileHandle = LIBC_NAMESPACE::fopen(inputFileName, "r" ); |
| 168 | |
| 169 | if (!fileHandle) { |
| 170 | LIBC_NAMESPACE::printf("file '%s' failed to open. Exiting.\n" , |
| 171 | inputFileName); |
| 172 | return ParseStatus::FILE_ERROR; |
| 173 | } |
| 174 | |
| 175 | while (LIBC_NAMESPACE::fgets(line, sizeof(line), fileHandle)) { |
| 176 | parseLine(line, parseResult, curFails, curBitDiffs); |
| 177 | } |
| 178 | |
| 179 | LIBC_NAMESPACE::fclose(fileHandle); |
| 180 | |
| 181 | parseResult.totalBitDiffs += curBitDiffs; |
| 182 | parseResult.totalFails += curFails; |
| 183 | |
| 184 | if (curFails > 1 || curBitDiffs > 1) { |
| 185 | return ParseStatus::PARSE_ERROR; |
| 186 | } |
| 187 | return ParseStatus::SUCCESS; |
| 188 | } |
| 189 | |
| 190 | ParseStatus updateStatus(ParseStatus parse_status, ParseStatus cur_status) { |
| 191 | if (cur_status == ParseStatus::FILE_ERROR) { |
| 192 | parse_status = ParseStatus::FILE_ERROR; |
| 193 | } else if (cur_status == ParseStatus::PARSE_ERROR) { |
| 194 | parse_status = ParseStatus::PARSE_ERROR; |
| 195 | } |
| 196 | return parse_status; |
| 197 | } |
| 198 | |
| 199 | TEST(LlvmLibcStrToFloatComparisonTest, CheckFloats) { |
| 200 | ParseStatus parseStatus = ParseStatus::SUCCESS; |
| 201 | |
| 202 | // Bitdiffs are cases where the expected result and actual result only differ |
| 203 | // by +/- the least significant bit. They are tracked separately from larger |
| 204 | // failures since a bitdiff is most likely the result of a rounding error, and |
| 205 | // splitting them off makes them easier to track down. |
| 206 | |
| 207 | ParseResult parseResult = { |
| 208 | .totalFails = 0, |
| 209 | .totalBitDiffs = 0, |
| 210 | .detailedBitDiffs = {0, 0, 0, 0}, |
| 211 | .total = 0, |
| 212 | }; |
| 213 | |
| 214 | char *files = LIBC_NAMESPACE::getenv("FILES" ); |
| 215 | |
| 216 | if (files == nullptr) { |
| 217 | ParseStatus cur_status = checkBuffer(parseResult); |
| 218 | parseStatus = updateStatus(parse_status: parseStatus, cur_status); |
| 219 | } else { |
| 220 | files = LIBC_NAMESPACE::strdup(files); |
| 221 | for (char *file = LIBC_NAMESPACE::strtok(files, "," ); file != nullptr; |
| 222 | file = LIBC_NAMESPACE::strtok(nullptr, "," )) { |
| 223 | ParseStatus cur_status = checkFile(inputFileName: file, parseResult); |
| 224 | parseStatus = updateStatus(parse_status: parseStatus, cur_status); |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | EXPECT_EQ(parseStatus, ParseStatus::SUCCESS); |
| 229 | EXPECT_EQ(parseResult.totalFails, 0u); |
| 230 | EXPECT_EQ(parseResult.totalBitDiffs, 0u); |
| 231 | EXPECT_EQ(parseResult.detailedBitDiffs[0], 0u); // float low |
| 232 | EXPECT_EQ(parseResult.detailedBitDiffs[1], 0u); // float high |
| 233 | EXPECT_EQ(parseResult.detailedBitDiffs[2], 0u); // double low |
| 234 | EXPECT_EQ(parseResult.detailedBitDiffs[3], 0u); // double high |
| 235 | LIBC_NAMESPACE::printf("Total lines: %d\n" , parseResult.total); |
| 236 | } |
| 237 | |