Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2011 The Android Open Source Project |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #define __STDC_LIMIT_MACROS |
| 18 | |
| 19 | #include <assert.h> |
| 20 | #include <stdint.h> |
| 21 | |
| 22 | #include <utils/LinearTransform.h> |
| 23 | |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 24 | // disable sanitize as these functions may intentionally overflow (see comments below). |
| 25 | // the ifdef can be removed when host builds use clang. |
| 26 | #if defined(__clang__) |
| 27 | #define ATTRIBUTE_NO_SANITIZE_INTEGER __attribute__((no_sanitize("integer"))) |
| 28 | #else |
| 29 | #define ATTRIBUTE_NO_SANITIZE_INTEGER |
| 30 | #endif |
| 31 | |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 32 | namespace android { |
| 33 | |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 34 | // sanitize failure with T = int32_t and x = 0x80000000 |
| 35 | template<class T> |
| 36 | ATTRIBUTE_NO_SANITIZE_INTEGER |
| 37 | static inline T ABS(T x) { return (x < 0) ? -x : x; } |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 38 | |
| 39 | // Static math methods involving linear transformations |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 40 | // remote sanitize failure on overflow case. |
| 41 | ATTRIBUTE_NO_SANITIZE_INTEGER |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 42 | static bool scale_u64_to_u64( |
| 43 | uint64_t val, |
| 44 | uint32_t N, |
| 45 | uint32_t D, |
| 46 | uint64_t* res, |
| 47 | bool round_up_not_down) { |
| 48 | uint64_t tmp1, tmp2; |
| 49 | uint32_t r; |
| 50 | |
| 51 | assert(res); |
| 52 | assert(D); |
| 53 | |
| 54 | // Let U32(X) denote a uint32_t containing the upper 32 bits of a 64 bit |
| 55 | // integer X. |
| 56 | // Let L32(X) denote a uint32_t containing the lower 32 bits of a 64 bit |
| 57 | // integer X. |
| 58 | // Let X[A, B] with A <= B denote bits A through B of the integer X. |
| 59 | // Let (A | B) denote the concatination of two 32 bit ints, A and B. |
| 60 | // IOW X = (A | B) => U32(X) == A && L32(X) == B |
| 61 | // |
| 62 | // compute M = val * N (a 96 bit int) |
| 63 | // --------------------------------- |
| 64 | // tmp2 = U32(val) * N (a 64 bit int) |
| 65 | // tmp1 = L32(val) * N (a 64 bit int) |
| 66 | // which means |
| 67 | // M = val * N = (tmp2 << 32) + tmp1 |
| 68 | tmp2 = (val >> 32) * N; |
| 69 | tmp1 = (val & UINT32_MAX) * N; |
| 70 | |
| 71 | // compute M[32, 95] |
| 72 | // tmp2 = tmp2 + U32(tmp1) |
| 73 | // = (U32(val) * N) + U32(L32(val) * N) |
| 74 | // = M[32, 95] |
| 75 | tmp2 += tmp1 >> 32; |
| 76 | |
| 77 | // if M[64, 95] >= D, then M/D has bits > 63 set and we have |
| 78 | // an overflow. |
| 79 | if ((tmp2 >> 32) >= D) { |
| 80 | *res = UINT64_MAX; |
| 81 | return false; |
| 82 | } |
| 83 | |
| 84 | // Divide. Going in we know |
| 85 | // tmp2 = M[32, 95] |
| 86 | // U32(tmp2) < D |
| 87 | r = tmp2 % D; |
| 88 | tmp2 /= D; |
| 89 | |
| 90 | // At this point |
| 91 | // tmp1 = L32(val) * N |
| 92 | // tmp2 = M[32, 95] / D |
| 93 | // = (M / D)[32, 95] |
| 94 | // r = M[32, 95] % D |
| 95 | // U32(tmp2) = 0 |
| 96 | // |
| 97 | // compute tmp1 = (r | M[0, 31]) |
| 98 | tmp1 = (tmp1 & UINT32_MAX) | ((uint64_t)r << 32); |
| 99 | |
| 100 | // Divide again. Keep the remainder around in order to round properly. |
| 101 | r = tmp1 % D; |
| 102 | tmp1 /= D; |
| 103 | |
| 104 | // At this point |
| 105 | // tmp2 = (M / D)[32, 95] |
| 106 | // tmp1 = (M / D)[ 0, 31] |
| 107 | // r = M % D |
| 108 | // U32(tmp1) = 0 |
| 109 | // U32(tmp2) = 0 |
| 110 | |
| 111 | // Pack the result and deal with the round-up case (As well as the |
| 112 | // remote possiblility over overflow in such a case). |
| 113 | *res = (tmp2 << 32) | tmp1; |
| 114 | if (r && round_up_not_down) { |
| 115 | ++(*res); |
| 116 | if (!(*res)) { |
| 117 | *res = UINT64_MAX; |
| 118 | return false; |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | return true; |
| 123 | } |
| 124 | |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 125 | // at least one known sanitize failure (see comment below) |
| 126 | ATTRIBUTE_NO_SANITIZE_INTEGER |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 127 | static bool linear_transform_s64_to_s64( |
| 128 | int64_t val, |
| 129 | int64_t basis1, |
| 130 | int32_t N, |
| 131 | uint32_t D, |
John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 132 | bool invert_frac, |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 133 | int64_t basis2, |
| 134 | int64_t* out) { |
| 135 | uint64_t scaled, res; |
| 136 | uint64_t abs_val; |
| 137 | bool is_neg; |
| 138 | |
| 139 | if (!out) |
| 140 | return false; |
| 141 | |
| 142 | // Compute abs(val - basis_64). Keep track of whether or not this delta |
| 143 | // will be negative after the scale opertaion. |
| 144 | if (val < basis1) { |
| 145 | is_neg = true; |
| 146 | abs_val = basis1 - val; |
| 147 | } else { |
| 148 | is_neg = false; |
| 149 | abs_val = val - basis1; |
| 150 | } |
| 151 | |
| 152 | if (N < 0) |
| 153 | is_neg = !is_neg; |
| 154 | |
| 155 | if (!scale_u64_to_u64(abs_val, |
John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 156 | invert_frac ? D : ABS(N), |
| 157 | invert_frac ? ABS(N) : D, |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 158 | &scaled, |
| 159 | is_neg)) |
| 160 | return false; // overflow/undeflow |
| 161 | |
| 162 | // if scaled is >= 0x8000<etc>, then we are going to overflow or |
| 163 | // underflow unless ABS(basis2) is large enough to pull us back into the |
| 164 | // non-overflow/underflow region. |
| 165 | if (scaled & INT64_MIN) { |
| 166 | if (is_neg && (basis2 < 0)) |
| 167 | return false; // certain underflow |
| 168 | |
| 169 | if (!is_neg && (basis2 >= 0)) |
| 170 | return false; // certain overflow |
| 171 | |
| 172 | if (ABS(basis2) <= static_cast<int64_t>(scaled & INT64_MAX)) |
| 173 | return false; // not enough |
| 174 | |
| 175 | // Looks like we are OK |
| 176 | *out = (is_neg ? (-scaled) : scaled) + basis2; |
| 177 | } else { |
| 178 | // Scaled fits within signed bounds, so we just need to check for |
| 179 | // over/underflow for two signed integers. Basically, if both scaled |
| 180 | // and basis2 have the same sign bit, and the result has a different |
| 181 | // sign bit, then we have under/overflow. An easy way to compute this |
| 182 | // is |
| 183 | // (scaled_signbit XNOR basis_signbit) && |
| 184 | // (scaled_signbit XOR res_signbit) |
| 185 | // == |
| 186 | // (scaled_signbit XOR basis_signbit XOR 1) && |
| 187 | // (scaled_signbit XOR res_signbit) |
| 188 | |
| 189 | if (is_neg) |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 190 | scaled = -scaled; // known sanitize failure |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 191 | res = scaled + basis2; |
| 192 | |
| 193 | if ((scaled ^ basis2 ^ INT64_MIN) & (scaled ^ res) & INT64_MIN) |
| 194 | return false; |
| 195 | |
| 196 | *out = res; |
| 197 | } |
| 198 | |
| 199 | return true; |
| 200 | } |
| 201 | |
| 202 | bool LinearTransform::doForwardTransform(int64_t a_in, int64_t* b_out) const { |
| 203 | if (0 == a_to_b_denom) |
| 204 | return false; |
| 205 | |
| 206 | return linear_transform_s64_to_s64(a_in, |
| 207 | a_zero, |
| 208 | a_to_b_numer, |
| 209 | a_to_b_denom, |
John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 210 | false, |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 211 | b_zero, |
| 212 | b_out); |
| 213 | } |
| 214 | |
| 215 | bool LinearTransform::doReverseTransform(int64_t b_in, int64_t* a_out) const { |
| 216 | if (0 == a_to_b_numer) |
| 217 | return false; |
| 218 | |
| 219 | return linear_transform_s64_to_s64(b_in, |
| 220 | b_zero, |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 221 | a_to_b_numer, |
John Grossman | 885a2fe | 2012-06-27 15:34:43 -0700 | [diff] [blame] | 222 | a_to_b_denom, |
| 223 | true, |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 224 | a_zero, |
| 225 | a_out); |
| 226 | } |
| 227 | |
| 228 | template <class T> void LinearTransform::reduce(T* N, T* D) { |
| 229 | T a, b; |
| 230 | if (!N || !D || !(*D)) { |
| 231 | assert(false); |
| 232 | return; |
| 233 | } |
| 234 | |
| 235 | a = *N; |
| 236 | b = *D; |
| 237 | |
| 238 | if (a == 0) { |
| 239 | *D = 1; |
| 240 | return; |
| 241 | } |
| 242 | |
| 243 | // This implements Euclid's method to find GCD. |
| 244 | if (a < b) { |
| 245 | T tmp = a; |
| 246 | a = b; |
| 247 | b = tmp; |
| 248 | } |
| 249 | |
| 250 | while (1) { |
| 251 | // a is now the greater of the two. |
| 252 | const T remainder = a % b; |
| 253 | if (remainder == 0) { |
| 254 | *N /= b; |
| 255 | *D /= b; |
| 256 | return; |
| 257 | } |
| 258 | // by swapping remainder and b, we are guaranteeing that a is |
| 259 | // still the greater of the two upon entrance to the loop. |
| 260 | a = b; |
| 261 | b = remainder; |
| 262 | } |
| 263 | }; |
| 264 | |
| 265 | template void LinearTransform::reduce<uint64_t>(uint64_t* N, uint64_t* D); |
| 266 | template void LinearTransform::reduce<uint32_t>(uint32_t* N, uint32_t* D); |
| 267 | |
Andy Hung | 13c19e7 | 2015-10-21 15:11:02 -0700 | [diff] [blame^] | 268 | // sanitize failure if *N = 0x80000000 |
| 269 | ATTRIBUTE_NO_SANITIZE_INTEGER |
Jason Simmons | b86c8e4 | 2011-06-28 17:43:30 -0700 | [diff] [blame] | 270 | void LinearTransform::reduce(int32_t* N, uint32_t* D) { |
| 271 | if (N && D && *D) { |
| 272 | if (*N < 0) { |
| 273 | *N = -(*N); |
| 274 | reduce(reinterpret_cast<uint32_t*>(N), D); |
| 275 | *N = -(*N); |
| 276 | } else { |
| 277 | reduce(reinterpret_cast<uint32_t*>(N), D); |
| 278 | } |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | } // namespace android |