1 | // Copyright (c) 2013-2014 Sandstorm Development Group, Inc. and contributors |
2 | // Licensed under the MIT License: |
3 | // |
4 | // Permission is hereby granted, free of charge, to any person obtaining a copy |
5 | // of this software and associated documentation files (the "Software"), to deal |
6 | // in the Software without restriction, including without limitation the rights |
7 | // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
8 | // copies of the Software, and to permit persons to whom the Software is |
9 | // furnished to do so, subject to the following conditions: |
10 | // |
11 | // The above copyright notice and this permission notice shall be included in |
12 | // all copies or substantial portions of the Software. |
13 | // |
14 | // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
15 | // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
16 | // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
17 | // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
18 | // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
19 | // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
20 | // THE SOFTWARE. |
21 | |
22 | #pragma once |
23 | |
24 | #if defined(__GNUC__) && !defined(CAPNP_HEADER_WARNINGS) |
25 | #pragma GCC system_header |
26 | #endif |
27 | |
28 | #include "common.h" |
29 | #include <inttypes.h> |
30 | #include <string.h> // memcpy |
31 | |
32 | namespace capnp { |
33 | namespace _ { // private |
34 | |
35 | // WireValue |
36 | // |
37 | // Wraps a primitive value as it appears on the wire. Namely, values are little-endian on the |
38 | // wire, because little-endian is the most common endianness in modern CPUs. |
39 | // |
40 | // Note: In general, code that depends cares about byte ordering is bad. See: |
41 | // http://commandcenter.blogspot.com/2012/04/byte-order-fallacy.html |
42 | // Cap'n Proto is special because it is essentially doing compiler-like things, fussing over |
43 | // allocation and layout of memory, in order to squeeze out every last drop of performance. |
44 | |
45 | #if _MSC_VER |
46 | // Assume Windows is little-endian. |
47 | // |
48 | // TODO(msvc): This is ugly. Maybe refactor later checks to be based on CAPNP_BYTE_ORDER or |
49 | // CAPNP_SWAP_BYTES or something, and define that in turn based on _MSC_VER or the GCC |
50 | // intrinsics. |
51 | |
52 | #ifndef __ORDER_BIG_ENDIAN__ |
53 | #define __ORDER_BIG_ENDIAN__ 4321 |
54 | #endif |
55 | #ifndef __ORDER_LITTLE_ENDIAN__ |
56 | #define __ORDER_LITTLE_ENDIAN__ 1234 |
57 | #endif |
58 | #ifndef __BYTE_ORDER__ |
59 | #define __BYTE_ORDER__ __ORDER_LITTLE_ENDIAN__ |
60 | #endif |
61 | #endif |
62 | |
63 | #if CAPNP_REVERSE_ENDIAN |
64 | #define CAPNP_WIRE_BYTE_ORDER __ORDER_BIG_ENDIAN__ |
65 | #define CAPNP_OPPOSITE_OF_WIRE_BYTE_ORDER __ORDER_LITTLE_ENDIAN__ |
66 | #else |
67 | #define CAPNP_WIRE_BYTE_ORDER __ORDER_LITTLE_ENDIAN__ |
68 | #define CAPNP_OPPOSITE_OF_WIRE_BYTE_ORDER __ORDER_BIG_ENDIAN__ |
69 | #endif |
70 | |
71 | #if defined(__BYTE_ORDER__) && \ |
72 | __BYTE_ORDER__ == CAPNP_WIRE_BYTE_ORDER && \ |
73 | !CAPNP_DISABLE_ENDIAN_DETECTION |
74 | // CPU is little-endian. We can just read/write the memory directly. |
75 | |
76 | template <typename T> |
77 | class DirectWireValue { |
78 | public: |
79 | KJ_ALWAYS_INLINE(T get() const) { return value; } |
80 | KJ_ALWAYS_INLINE(void set(T newValue)) { value = newValue; } |
81 | |
82 | private: |
83 | T value; |
84 | }; |
85 | |
86 | template <typename T> |
87 | using WireValue = DirectWireValue<T>; |
88 | // To prevent ODR problems when endian-test, endian-reverse-test, and endian-fallback-test are |
89 | // linked together, we define each implementation with a different name and define an alias to the |
90 | // one we want to use. |
91 | |
92 | #elif defined(__BYTE_ORDER__) && \ |
93 | __BYTE_ORDER__ == CAPNP_OPPOSITE_OF_WIRE_BYTE_ORDER && \ |
94 | defined(__GNUC__) && !CAPNP_DISABLE_ENDIAN_DETECTION |
95 | // Big-endian, but GCC's __builtin_bswap() is available. |
96 | |
97 | // TODO(perf): Use dedicated instructions to read little-endian data on big-endian CPUs that have |
98 | // them. |
99 | |
100 | // TODO(perf): Verify that this code optimizes reasonably. In particular, ensure that the |
101 | // compiler optimizes away the memcpy()s and keeps everything in registers. |
102 | |
103 | template <typename T, size_t size = sizeof(T)> |
104 | class SwappingWireValue; |
105 | |
106 | template <typename T> |
107 | class SwappingWireValue<T, 1> { |
108 | public: |
109 | KJ_ALWAYS_INLINE(T get() const) { return value; } |
110 | KJ_ALWAYS_INLINE(void set(T newValue)) { value = newValue; } |
111 | |
112 | private: |
113 | T value; |
114 | }; |
115 | |
116 | template <typename T> |
117 | class SwappingWireValue<T, 2> { |
118 | public: |
119 | KJ_ALWAYS_INLINE(T get() const) { |
120 | // Not all platforms have __builtin_bswap16() for some reason. In particular, it is missing |
121 | // on gcc-4.7.3-cygwin32 (but present on gcc-4.8.1-cygwin64). |
122 | uint16_t swapped = (value << 8) | (value >> 8); |
123 | T result; |
124 | memcpy(&result, &swapped, sizeof(T)); |
125 | return result; |
126 | } |
127 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
128 | uint16_t raw; |
129 | memcpy(&raw, &newValue, sizeof(T)); |
130 | // Not all platforms have __builtin_bswap16() for some reason. In particular, it is missing |
131 | // on gcc-4.7.3-cygwin32 (but present on gcc-4.8.1-cygwin64). |
132 | value = (raw << 8) | (raw >> 8); |
133 | } |
134 | |
135 | private: |
136 | uint16_t value; |
137 | }; |
138 | |
139 | template <typename T> |
140 | class SwappingWireValue<T, 4> { |
141 | public: |
142 | KJ_ALWAYS_INLINE(T get() const) { |
143 | uint32_t swapped = __builtin_bswap32(value); |
144 | T result; |
145 | memcpy(&result, &swapped, sizeof(T)); |
146 | return result; |
147 | } |
148 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
149 | uint32_t raw; |
150 | memcpy(&raw, &newValue, sizeof(T)); |
151 | value = __builtin_bswap32(raw); |
152 | } |
153 | |
154 | private: |
155 | uint32_t value; |
156 | }; |
157 | |
158 | template <typename T> |
159 | class SwappingWireValue<T, 8> { |
160 | public: |
161 | KJ_ALWAYS_INLINE(T get() const) { |
162 | uint64_t swapped = __builtin_bswap64(value); |
163 | T result; |
164 | memcpy(&result, &swapped, sizeof(T)); |
165 | return result; |
166 | } |
167 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
168 | uint64_t raw; |
169 | memcpy(&raw, &newValue, sizeof(T)); |
170 | value = __builtin_bswap64(raw); |
171 | } |
172 | |
173 | private: |
174 | uint64_t value; |
175 | }; |
176 | |
177 | template <typename T> |
178 | using WireValue = SwappingWireValue<T>; |
179 | // To prevent ODR problems when endian-test, endian-reverse-test, and endian-fallback-test are |
180 | // linked together, we define each implementation with a different name and define an alias to the |
181 | // one we want to use. |
182 | |
183 | #else |
184 | // Unknown endianness. Fall back to bit shifts. |
185 | |
186 | #if !CAPNP_DISABLE_ENDIAN_DETECTION |
187 | #if _MSC_VER |
188 | #pragma message("Couldn't detect endianness of your platform. Using unoptimized fallback implementation.") |
189 | #pragma message("Consider changing this code to detect your platform and send us a patch!") |
190 | #else |
191 | #warning "Couldn't detect endianness of your platform. Using unoptimized fallback implementation." |
192 | #warning "Consider changing this code to detect your platform and send us a patch!" |
193 | #endif |
194 | #endif // !CAPNP_DISABLE_ENDIAN_DETECTION |
195 | |
196 | template <typename T, size_t size = sizeof(T)> |
197 | class ShiftingWireValue; |
198 | |
199 | template <typename T> |
200 | class ShiftingWireValue<T, 1> { |
201 | public: |
202 | KJ_ALWAYS_INLINE(T get() const) { return value; } |
203 | KJ_ALWAYS_INLINE(void set(T newValue)) { value = newValue; } |
204 | |
205 | private: |
206 | T value; |
207 | }; |
208 | |
209 | template <typename T> |
210 | class ShiftingWireValue<T, 2> { |
211 | public: |
212 | KJ_ALWAYS_INLINE(T get() const) { |
213 | uint16_t raw = (static_cast<uint16_t>(bytes[0]) ) | |
214 | (static_cast<uint16_t>(bytes[1]) << 8); |
215 | T result; |
216 | memcpy(&result, &raw, sizeof(T)); |
217 | return result; |
218 | } |
219 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
220 | uint16_t raw; |
221 | memcpy(&raw, &newValue, sizeof(T)); |
222 | bytes[0] = raw; |
223 | bytes[1] = raw >> 8; |
224 | } |
225 | |
226 | private: |
227 | union { |
228 | byte bytes[2]; |
229 | uint16_t align; |
230 | }; |
231 | }; |
232 | |
233 | template <typename T> |
234 | class ShiftingWireValue<T, 4> { |
235 | public: |
236 | KJ_ALWAYS_INLINE(T get() const) { |
237 | uint32_t raw = (static_cast<uint32_t>(bytes[0]) ) | |
238 | (static_cast<uint32_t>(bytes[1]) << 8) | |
239 | (static_cast<uint32_t>(bytes[2]) << 16) | |
240 | (static_cast<uint32_t>(bytes[3]) << 24); |
241 | T result; |
242 | memcpy(&result, &raw, sizeof(T)); |
243 | return result; |
244 | } |
245 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
246 | uint32_t raw; |
247 | memcpy(&raw, &newValue, sizeof(T)); |
248 | bytes[0] = raw; |
249 | bytes[1] = raw >> 8; |
250 | bytes[2] = raw >> 16; |
251 | bytes[3] = raw >> 24; |
252 | } |
253 | |
254 | private: |
255 | union { |
256 | byte bytes[4]; |
257 | uint32_t align; |
258 | }; |
259 | }; |
260 | |
261 | template <typename T> |
262 | class ShiftingWireValue<T, 8> { |
263 | public: |
264 | KJ_ALWAYS_INLINE(T get() const) { |
265 | uint64_t raw = (static_cast<uint64_t>(bytes[0]) ) | |
266 | (static_cast<uint64_t>(bytes[1]) << 8) | |
267 | (static_cast<uint64_t>(bytes[2]) << 16) | |
268 | (static_cast<uint64_t>(bytes[3]) << 24) | |
269 | (static_cast<uint64_t>(bytes[4]) << 32) | |
270 | (static_cast<uint64_t>(bytes[5]) << 40) | |
271 | (static_cast<uint64_t>(bytes[6]) << 48) | |
272 | (static_cast<uint64_t>(bytes[7]) << 56); |
273 | T result; |
274 | memcpy(&result, &raw, sizeof(T)); |
275 | return result; |
276 | } |
277 | KJ_ALWAYS_INLINE(void set(T newValue)) { |
278 | uint64_t raw; |
279 | memcpy(&raw, &newValue, sizeof(T)); |
280 | bytes[0] = raw; |
281 | bytes[1] = raw >> 8; |
282 | bytes[2] = raw >> 16; |
283 | bytes[3] = raw >> 24; |
284 | bytes[4] = raw >> 32; |
285 | bytes[5] = raw >> 40; |
286 | bytes[6] = raw >> 48; |
287 | bytes[7] = raw >> 56; |
288 | } |
289 | |
290 | private: |
291 | union { |
292 | byte bytes[8]; |
293 | uint64_t align; |
294 | }; |
295 | }; |
296 | |
297 | template <typename T> |
298 | using WireValue = ShiftingWireValue<T>; |
299 | // To prevent ODR problems when endian-test, endian-reverse-test, and endian-fallback-test are |
300 | // linked together, we define each implementation with a different name and define an alias to the |
301 | // one we want to use. |
302 | |
303 | #endif |
304 | |
305 | } // namespace _ (private) |
306 | } // namespace capnp |
307 | |