Fossil SCM

Use intrinsic byte-swapping functions to boost the performance of delta checksums.

drh 2015-12-31 15:45 trunk merge
Commit 7338b3a9510281e7d49643f55a5c27c24107f083
Parent ec8394e166117a5…
1 file changed +60 -27
~ src/delta.c
M src/delta.c
+60 -27
--- src/delta.c
+++ src/delta.c
@@ -209,43 +209,76 @@
209209
 unsigned int i, x;
210210
 for(i=1, x=64; v>=x; i++, x <<= 6){}
211211
 return i;
212212
 }
213213
214
+#ifdef __GNUC__
215
+# define GCC_VERSION (__GNUC__*1000000+__GNUC_MINOR__*1000+__GNUC_PATCHLEVEL__)
216
+#else
217
+# define GCC_VERSION 0
218
+#endif
219
+
214220
 /*
215
-** Compute a 32-bit checksum on the N-byte buffer. Return the result.
221
+** Compute a 32-bit big-endian checksum on the N-byte buffer. If the
222
+** buffer is not a multiple of 4 bytes length, compute the sum that would
223
+** have occurred if the buffer was padded with zeros to the next multiple
224
+** of four bytes.
216225
 */
217226
 static unsigned int checksum(const char *zIn, size_t N){
227
+ static const int byteOrderTest = 1;
218228
 const unsigned char *z = (const unsigned char *)zIn;
219
- unsigned sum0 = 0;
220
- unsigned sum1 = 0;
221
- unsigned sum2 = 0;
222
- unsigned sum3 = 0;
223
- while(N >= 16){
224
- sum0 += ((unsigned)z[0] + z[4] + z[8] + z[12]);
225
- sum1 += ((unsigned)z[1] + z[5] + z[9] + z[13]);
226
- sum2 += ((unsigned)z[2] + z[6] + z[10]+ z[14]);
227
- sum3 += ((unsigned)z[3] + z[7] + z[11]+ z[15]);
228
- z += 16;
229
- N -= 16;
230
- }
231
- while(N >= 4){
232
- sum0 += z[0];
233
- sum1 += z[1];
234
- sum2 += z[2];
235
- sum3 += z[3];
236
- z += 4;
237
- N -= 4;
238
- }
239
- sum3 += (sum2 << 8) + (sum1 << 16) + (sum0 << 24);
240
- switch(N){
241
- case 3: sum3 += (z[2] << 8);
242
- case 2: sum3 += (z[1] << 16);
243
- case 1: sum3 += (z[0] << 24);
229
+ const unsigned char *zEnd = (const unsigned char*)&zIn[N&~3];
230
+ unsigned sum = 0;
231
+ assert( (3&(sqlite3_uint64)z)==0 ); /* Four-byte alignment */
232
+ if( 0==*(char*)&byteOrderTest ){
233
+ /* This is a big-endian machine */
234
+ while( z<zEnd ){
235
+ sum += *(unsigned*)z;
236
+ z += 4;
237
+ }
238
+ }else{
239
+ /* A little-endian machine */
240
+#if GCC_VERSION>=4003000
241
+ while( z<zEnd ){
242
+ sum += __builtin_bswap32(*(unsigned*)z);
243
+ z += 4;
244
+ }
245
+#elif defined(_MSC_VER) && _MSC_VER>=1300
246
+ while( z<zEnd ){
247
+ sum += _byteswap_ulong(*(unsigned*)z);
248
+ z += 4;
249
+ }
250
+#else
251
+ unsigned sum0 = 0;
252
+ unsigned sum1 = 0;
253
+ unsigned sum2 = 0;
254
+ while(N >= 16){
255
+ sum0 += ((unsigned)z[0] + z[4] + z[8] + z[12]);
256
+ sum1 += ((unsigned)z[1] + z[5] + z[9] + z[13]);
257
+ sum2 += ((unsigned)z[2] + z[6] + z[10]+ z[14]);
258
+ sum += ((unsigned)z[3] + z[7] + z[11]+ z[15]);
259
+ z += 16;
260
+ N -= 16;
261
+ }
262
+ while(N >= 4){
263
+ sum0 += z[0];
264
+ sum1 += z[1];
265
+ sum2 += z[2];
266
+ sum += z[3];
267
+ z += 4;
268
+ N -= 4;
269
+ }
270
+ sum += (sum2 << 8) + (sum1 << 16) + (sum0 << 24);
271
+#endif
272
+ }
273
+ switch(N&3){
274
+ case 3: sum += (z[2] << 8);
275
+ case 2: sum += (z[1] << 16);
276
+ case 1: sum += (z[0] << 24);
244277
 default: ;
245278
 }
246
- return sum3;
279
+ return sum;
247280
 }
248281
249282
 /*
250283
 ** Create a new delta.
251284
 **
252285
--- src/delta.c
+++ src/delta.c
@@ -209,43 +209,76 @@
209 unsigned int i, x;
210 for(i=1, x=64; v>=x; i++, x <<= 6){}
211 return i;
212 }
213
 
 
 
 
 
 
214 /*
215 ** Compute a 32-bit checksum on the N-byte buffer. Return the result.
 
 
 
216 */
217 static unsigned int checksum(const char *zIn, size_t N){
 
218 const unsigned char *z = (const unsigned char *)zIn;
219 unsigned sum0 = 0;
220 unsigned sum1 = 0;
221 unsigned sum2 = 0;
222 unsigned sum3 = 0;
223 while(N >= 16){
224 sum0 += ((unsigned)z[0] + z[4] + z[8] + z[12]);
225 sum1 += ((unsigned)z[1] + z[5] + z[9] + z[13]);
226 sum2 += ((unsigned)z[2] + z[6] + z[10]+ z[14]);
227 sum3 += ((unsigned)z[3] + z[7] + z[11]+ z[15]);
228 z += 16;
229 N -= 16;
230 }
231 while(N >= 4){
232 sum0 += z[0];
233 sum1 += z[1];
234 sum2 += z[2];
235 sum3 += z[3];
236 z += 4;
237 N -= 4;
238 }
239 sum3 += (sum2 << 8) + (sum1 << 16) + (sum0 << 24);
240 switch(N){
241 case 3: sum3 += (z[2] << 8);
242 case 2: sum3 += (z[1] << 16);
243 case 1: sum3 += (z[0] << 24);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
244 default: ;
245 }
246 return sum3;
247 }
248
249 /*
250 ** Create a new delta.
251 **
252
--- src/delta.c
+++ src/delta.c
@@ -209,43 +209,76 @@
209 unsigned int i, x;
210 for(i=1, x=64; v>=x; i++, x <<= 6){}
211 return i;
212 }
213
214 #ifdef __GNUC__
215 # define GCC_VERSION (__GNUC__*1000000+__GNUC_MINOR__*1000+__GNUC_PATCHLEVEL__)
216 #else
217 # define GCC_VERSION 0
218 #endif
219
220 /*
221 ** Compute a 32-bit big-endian checksum on the N-byte buffer. If the
222 ** buffer is not a multiple of 4 bytes length, compute the sum that would
223 ** have occurred if the buffer was padded with zeros to the next multiple
224 ** of four bytes.
225 */
226 static unsigned int checksum(const char *zIn, size_t N){
227 static const int byteOrderTest = 1;
228 const unsigned char *z = (const unsigned char *)zIn;
229 const unsigned char *zEnd = (const unsigned char*)&zIn[N&~3];
230 unsigned sum = 0;
231 assert( (3&(sqlite3_uint64)z)==0 ); /* Four-byte alignment */
232 if( 0==*(char*)&byteOrderTest ){
233 /* This is a big-endian machine */
234 while( z<zEnd ){
235 sum += *(unsigned*)z;
236 z += 4;
237 }
238 }else{
239 /* A little-endian machine */
240 #if GCC_VERSION>=4003000
241 while( z<zEnd ){
242 sum += __builtin_bswap32(*(unsigned*)z);
243 z += 4;
244 }
245 #elif defined(_MSC_VER) && _MSC_VER>=1300
246 while( z<zEnd ){
247 sum += _byteswap_ulong(*(unsigned*)z);
248 z += 4;
249 }
250 #else
251 unsigned sum0 = 0;
252 unsigned sum1 = 0;
253 unsigned sum2 = 0;
254 while(N >= 16){
255 sum0 += ((unsigned)z[0] + z[4] + z[8] + z[12]);
256 sum1 += ((unsigned)z[1] + z[5] + z[9] + z[13]);
257 sum2 += ((unsigned)z[2] + z[6] + z[10]+ z[14]);
258 sum += ((unsigned)z[3] + z[7] + z[11]+ z[15]);
259 z += 16;
260 N -= 16;
261 }
262 while(N >= 4){
263 sum0 += z[0];
264 sum1 += z[1];
265 sum2 += z[2];
266 sum += z[3];
267 z += 4;
268 N -= 4;
269 }
270 sum += (sum2 << 8) + (sum1 << 16) + (sum0 << 24);
271 #endif
272 }
273 switch(N&3){
274 case 3: sum += (z[2] << 8);
275 case 2: sum += (z[1] << 16);
276 case 1: sum += (z[0] << 24);
277 default: ;
278 }
279 return sum;
280 }
281
282 /*
283 ** Create a new delta.
284 **
285

Keyboard Shortcuts

Open search /
Next entry (timeline) j
Previous entry (timeline) k
Open focused entry Enter
Show this help ?
Toggle theme Top nav button