diff options
author | Thiago Macieira <thiago.macieira@nokia.com> | 2010-08-17 18:32:27 (GMT) |
---|---|---|
committer | Thiago Macieira <thiago.macieira@nokia.com> | 2010-08-24 10:36:35 (GMT) |
commit | c58293407519e7413cb6493f831ef69cf2323030 (patch) | |
tree | 9321f5d1b14315c6b3a705abe16c4705e71aaa41 /tests/benchmarks | |
parent | 64a6ce5fb031f151994f77b9d9abea52bb9e00d3 (diff) | |
download | Qt-c58293407519e7413cb6493f831ef69cf2323030.zip Qt-c58293407519e7413cb6493f831ef69cf2323030.tar.gz Qt-c58293407519e7413cb6493f831ef69cf2323030.tar.bz2 |
Major improvements to the comparison functions.
Also use the real-world data that I collected. The resulting files are
quite large, so I've added to Git only the smallest dump (apparently
Teambuilder scanning the environment for "TEAMBUILDER=").
Diffstat (limited to 'tests/benchmarks')
-rw-r--r-- | tests/benchmarks/corelib/tools/qstring/data.cpp | 1283 | ||||
-rw-r--r-- | tests/benchmarks/corelib/tools/qstring/main.cpp | 675 |
2 files changed, 1770 insertions, 188 deletions
diff --git a/tests/benchmarks/corelib/tools/qstring/data.cpp b/tests/benchmarks/corelib/tools/qstring/data.cpp new file mode 100644 index 0000000..89f50d0 --- /dev/null +++ b/tests/benchmarks/corelib/tools/qstring/data.cpp @@ -0,0 +1,1283 @@ +static const ushort stringCollectionData[] __attribute__((aligned(16))) = { + // #0 + 65535, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 118, 101, 114, 115, 105, 111, 110, 115, 47, + 65535,65534,65533,65532,65531, // 24 + 65535,65534,65533,65532,65531, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 118, 101, 114, 115, 105, 111, 110, 115, 47, + 65535, // 48 + + // #1 + 65535,65534,65533,65532,65531, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 118, 101, 114, 115, 105, 111, 110, 115, 47, + 65535, // 72 + 65535,65534,65533,65532,65531, + 67, 111, 109, 112, 105, 108, 101, 114, 32, 86, 101, 114, 115, 105, 111, 110, 115, 47, + 65535, // 96 + + // #2 + 65535, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 116, 105, 109, 101, 115, 116, 97, 109, 112, 115, 47, + 65535,65534,65533, // 120 + 65535,65534,65533,65532,65531, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 116, 105, 109, 101, 115, 116, 97, 109, 112, 115, 47, + 65535,65534,65533,65532,65531,65530,65529, // 152 + + // #3 + 65535,65534,65533,65532,65531, + 99, 111, 109, 112, 105, 108, 101, 114, 32, 116, 105, 109, 101, 115, 116, 97, 109, 112, 115, 47, + 65535,65534,65533,65532,65531,65530,65529, // 184 + 65535, + 67, 111, 109, 112, 105, 108, 101, 114, 32, 84, 105, 109, 101, 115, 116, 97, 109, 112, 115, 47, + 65535,65534,65533, // 208 + + // #4 + 65535,65534,65533,65532,65531,65530,65529,65528,65527,65526,65525,65524,65523,65522,65521,65520,65519, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532, // 280+ + + + // #5 + 65535,65534,65533,65532,65531,65530,65529,65528,65527,65526,65525,65524,65523, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532,65531,65530,65529,65528, // 352+ + 65535, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532, // 408+ + + // #6 + 65535,65534,65533,65532,65531,65530,65529,65528,65527, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532, // 472+ + + + // #7 + 65535, + 97, 114, 99, 104, 105, 118, 101, 100, 32, 99, 111, 109, 112, 105, 108, 101, 114, 115, 47, + 65535,65534,65533,65532, // 496 + 65535,65534,65533,65532,65531, + 97, 114, 99, 104, 105, 118, 101, 100, 32, 99, 111, 109, 112, 105, 108, 101, 114, 115, 47, + 65535,65534,65533,65532,65531,65530,65529,65528, // 528 + + // #8 + 65535,65534,65533,65532,65531, + 97, 114, 99, 104, 105, 118, 101, 100, 32, 99, 111, 109, 112, 105, 108, 101, 114, 115, 47, + 65535,65534,65533,65532,65531,65530,65529,65528, // 560 + 65535,65534,65533,65532,65531, + 65, 114, 99, 104, 105, 118, 101, 100, 32, 67, 111, 109, 112, 105, 108, 101, 114, 115, 47, + 65535,65534,65533,65532,65531,65530,65529,65528, // 592 + + // #9 + 65535,65534,65533,65532,65531,65530,65529,65528,65527,65526,65525,65524,65523,65522,65521,65520,65519, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532, // 664+ + 65535,65534,65533,65532,65531,65530,65529,65528,65527,65526,65525,65524,65523, + 47, 118, 97, 114, 47, 116, 109, 112, 47, 116, 101, 97, 109, 98, 117, 105, 108, 100, 101, 114, 45, 116, 109, 97, 99, 105, 101, 105, 114, 47, 99, 108, 105, 101, 110, 116, 47, 99, 111, 109, 112, 105, 108, 101, 114, 115, 46, 99, 111, 110, 102, + 65535,65534,65533,65532,65531,65530,65529,65528, // 736+ + + // #10 + 65535,65534,65533,65532,65531, + 76, 105, 110, 117, 120, + 65535,65534,65533,65532,65531,65530, // 752 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 760 + + // #11 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 776 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 792 + + // #12 + 65535, + 105, 99, 99, + 65535,65534,65533,65532, // 800 + 65535,65534,65533,65532,65531, + 103, 43, 43, + 65535,65534,65533,65532,65531,65530,65529,65528, // 816 + + // #13 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 824 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 832 + + // #14 + 65535, + 105, 51, 56, 54, + 65535,65534,65533, // 840 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 856 + + // #15 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 864 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 872 + + // #16 + 65535, + 105, 51, 56, 54, + 65535,65534,65533, // 880 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 896 + + // #17 + 65535, + 103, 99, 99, + 65535,65534,65533,65532, // 904 + 65535,65534,65533,65532,65531, + 103, 43, 43, + 65535,65534,65533,65532,65531,65530,65529,65528, // 920 + + // #18 + 65535,65534,65533,65532,65531, + 76, 105, 110, 117, 120, + 65535,65534,65533,65532,65531,65530, // 936 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 944 + + // #19 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 960 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 976 + + // #20 + 65535, + 103, 43, 43, + 65535,65534,65533,65532, // 984 + 65535,65534,65533,65532,65531, + 103, 43, 43, + 65535,65534,65533,65532,65531,65530,65529,65528, // 1000 + + // #21 + 65535,65534,65533,65532,65531, + 52, 46, 52, 46, 51, + 65535,65534,65533,65532,65531,65530, // 1016 + 65535,65534,65533,65532,65531, + 52, 46, 52, 46, 51, + 65535,65534,65533,65532,65531,65530, // 1032 + + // #22 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1040 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1048 + + // #23 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1064 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1080 + + // #24 + 65535, + 47, 117, 115, 114, 47, 98, 105, 110, 47, 103, 43, 43, + 65535,65534,65533, // 1096 + 65535, + 47, 117, 115, 114, 47, 98, 105, 110, 47, 103, 43, 43, + 65535,65534,65533, // 1112 + + // #25 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1120 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1128 + + // #26 + 65535, + 105, 51, 56, 54, + 65535,65534,65533, // 1136 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1152 + + // #27 + 65535, + 105, 99, 99, + 65535,65534,65533,65532, // 1160 + 65535,65534,65533,65532,65531, + 103, 43, 43, + 65535,65534,65533,65532,65531,65530,65529,65528, // 1176 + + // #28 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1184 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1192 + + // #29 + 65535, + 105, 51, 56, 54, + 65535,65534,65533, // 1200 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1216 + + // #30 + 65535,65534,65533,65532,65531, + 76, 105, 110, 117, 120, + 65535,65534,65533,65532,65531,65530, // 1232 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1240 + + // #31 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1256 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1272 + + // #32 + 65535,65534,65533,65532,65531, + 76, 105, 110, 117, 120, + 65535,65534,65533,65532,65531,65530, // 1288 + 65535, + 76, 105, 110, 117, 120, + 65535,65534, // 1296 + + // #33 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1312 + 65535,65534,65533,65532,65531, + 105, 51, 56, 54, + 65535,65534,65533,65532,65531,65530,65529, // 1328 + + // #34 + 65535, + 45, 109, 97, 114, 99, 104, 61, 99, 111, 114, 101, 50, + 65535,65534,65533, // 1344 + 65535, + 116, 98, 51, 54, 57, 54, 56, 95, 50, 46, 105, 105, + 65535,65534,65533, // 1360 + + // #35 + 65535,65534,65533,65532,65531, + 45, 102, 108, 111, 111, 112, 45, 98, 108, 111, 99, 107, + 65535,65534,65533,65532,65531,65530,65529, // 1384 + 65535, + 116, 98, 51, 54, 57, 54, 56, 95, 50, 46, 105, 105, + 65535,65534,65533, // 1400 + + // #36 + 65535,65534,65533,65532,65531, + 116, 98, 51, 54, 57, 54, 56, 95, 50, 46, 105, 105, + 65535,65534,65533,65532,65531,65530,65529, // 1424 + 65535, + 116, 98, 51, 54, 57, 54, 56, 95, 50, 46, 105, 105, + 65535,65534,65533, // 1440 + + // #37 + 65535,65534,65533,65532,65531, + 45, 109, 115, 115, 101, 52, + 65535,65534,65533,65532,65531, // 1456 + 65535,65534,65533,65532,65531, + 108, 101, 110, 103, 116, 104, + 65535,65534,65533,65532,65531, // 1472 + + // #38 + 65535,65534,65533,65532,65531, + 116, 98, 51, 54, 57, 54, 56, 95, 49, 46, 111, + 65535,65534,65533,65532,65531,65530,65529,65528, // 1496 + 65535,65534,65533,65532,65531, + 116, 98, 51, 54, 57, 54, 56, 95, 49, 46, 111, + 65535,65534,65533,65532,65531,65530,65529,65528, // 1520 + + // #39 + 65535,65534,65533,65532,65531, + 68, 69, 83, 75, 84, 79, 80, 95, 83, 69, 83, 83, + 65535,65534,65533,65532,65531,65530,65529, // 1544 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1560 + + // #40 + 65535,65534,65533,65532,65531, + 76, 67, 95, 83, 79, 85, 82, 67, 69, 68, 61, 49, + 65535,65534,65533,65532,65531,65530,65529, // 1584 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1600 + + // #41 + 65535,65534,65533,65532,65531, + 81, 84, 68, 73, 82, 61, 47, 104, 111, 109, 101, 47, + 65535,65534,65533,65532,65531,65530,65529, // 1624 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1640 + + // #42 + 65535, + 76, 67, 95, 67, 84, 89, 80, 69, 61, 112, 116, 95, + 65535,65534,65533, // 1656 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1672 + + // #43 + 65535, + 71, 84, 75, 95, 82, 67, 95, 70, 73, 76, 69, 83, + 65535,65534,65533, // 1688 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1704 + + // #44 + 65535, + 88, 77, 79, 68, 73, 70, 73, 69, 82, 83, 61, 64, + 65535,65534,65533, // 1720 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1736 + + // #45 + 65535, + 83, 72, 69, 76, 76, 61, 47, 98, 105, 110, 47, 122, + 65535,65534,65533, // 1752 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1768 + + // #46 + 65535,65534,65533,65532,65531, + 85, 61, 64, 123, 117, 112, 115, 116, 114, 101, 97, 109, + 65535,65534,65533,65532,65531,65530,65529, // 1792 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1808 + + // #47 + 65535, + 95, 61, 47, 117, 115, 114, 47, 98, 105, 110, 47, 105, + 65535,65534,65533, // 1824 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1840 + + // #48 + 65535, + 88, 68, 71, 95, 67, 79, 78, 70, 73, 71, 95, 68, + 65535,65534,65533, // 1856 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1872 + + // #49 + 65535,65534,65533,65532,65531, + 83, 65, 86, 69, 72, 73, 83, 84, 61, 49, 48, 48, + 65535,65534,65533,65532,65531,65530,65529, // 1896 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1912 + + // #50 + 65535, + 75, 68, 69, 95, 77, 85, 76, 84, 73, 72, 69, 65, + 65535,65534,65533, // 1928 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1944 + + // #51 + 65535,65534,65533,65532,65531, + 77, 65, 76, 76, 79, 67, 95, 67, 72, 69, 67, 75, + 65535,65534,65533,65532,65531,65530,65529, // 1968 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 1984 + + // #52 + 65535,65534,65533,65532,65531, + 72, 73, 83, 84, 67, 79, 78, 84, 82, 79, 76, 61, + 65535,65534,65533,65532,65531,65530,65529, // 2008 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2024 + + // #53 + 65535,65534,65533,65532,65531, + 88, 68, 71, 95, 68, 65, 84, 65, 95, 68, 73, 82, + 65535,65534,65533,65532,65531,65530,65529, // 2048 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2064 + + // #54 + 65535,65534,65533,65532,65531, + 88, 68, 77, 95, 77, 65, 78, 65, 71, 69, 68, 61, + 65535,65534,65533,65532,65531,65530,65529, // 2088 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2104 + + // #55 + 65535, + 76, 67, 95, 67, 79, 76, 76, 65, 84, 69, 61, 112, + 65535,65534,65533, // 2120 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2136 + + // #56 + 65535, + 81, 84, 95, 80, 76, 85, 71, 73, 78, 95, 80, 65, + 65535,65534,65533, // 2152 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2168 + + // #57 + 65535, + 83, 67, 82, 69, 69, 78, 68, 73, 82, 61, 47, 104, + 65535,65534,65533, // 2184 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2200 + + // #58 + 65535,65534,65533,65532,65531, + 76, 69, 83, 83, 79, 80, 69, 78, 61, 124, 47, 117, + 65535,65534,65533,65532,65531,65530,65529, // 2224 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2240 + + // #59 + 65535, + 76, 67, 95, 78, 65, 77, 69, 61, 110, 98, 95, 78, + 65535,65534,65533, // 2256 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2272 + + // #60 + 65535,65534,65533,65532,65531, + 80, 52, 67, 76, 73, 69, 78, 84, 61, 116, 109, 97, + 65535,65534,65533,65532,65531,65530,65529, // 2296 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2312 + + // #61 + 65535, + 80, 65, 84, 72, 61, 47, 104, 111, 109, 101, 47, 116, + 65535,65534,65533, // 2328 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2344 + + // #62 + 65535,65534,65533,65532,65531, + 71, 80, 71, 95, 65, 71, 69, 78, 84, 95, 73, 78, + 65535,65534,65533,65532,65531,65530,65529, // 2368 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2384 + + // #63 + 65535,65534,65533,65532,65531, + 88, 67, 85, 82, 83, 79, 82, 95, 84, 72, 69, 77, + 65535,65534,65533,65532,65531,65530,65529, // 2408 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2424 + + // #64 + 65535,65534,65533,65532,65531, + 83, 69, 83, 83, 73, 79, 78, 95, 77, 65, 78, 65, + 65535,65534,65533,65532,65531,65530,65529, // 2448 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2464 + + // #65 + 65535, + 81, 84, 83, 82, 67, 68, 73, 82, 61, 47, 104, 111, + 65535,65534,65533, // 2480 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2496 + + // #66 + 65535,65534,65533,65532,65531, + 87, 73, 78, 68, 79, 87, 73, 68, 61, 52, 54, 49, + 65535,65534,65533,65532,65531,65530,65529, // 2520 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2536 + + // #67 + 65535,65534,65533,65532,65531, + 76, 67, 95, 77, 69, 83, 83, 65, 71, 69, 83, 61, + 65535,65534,65533,65532,65531,65530,65529, // 2560 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2576 + + // #68 + 65535, + 76, 67, 95, 78, 85, 77, 69, 82, 73, 67, 61, 110, + 65535,65534,65533, // 2592 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2608 + + // #69 + 65535, + 71, 84, 75, 50, 95, 82, 67, 95, 70, 73, 76, 69, + 65535,65534,65533, // 2624 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2640 + + // #70 + 65535, + 80, 82, 79, 70, 73, 76, 69, 72, 79, 77, 69, 61, + 65535,65534,65533, // 2656 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2672 + + // #71 + 65535,65534,65533,65532,65531, + 68, 77, 95, 67, 79, 78, 84, 82, 79, 76, 61, 47, + 65535,65534,65533,65532,65531,65530,65529, // 2696 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2712 + + // #72 + 65535, + 76, 83, 95, 67, 79, 76, 79, 82, 83, 61, 114, 115, + 65535,65534,65533, // 2728 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2744 + + // #73 + 65535,65534,65533,65532,65531, + 83, 83, 72, 95, 65, 85, 84, 72, 95, 83, 79, 67, + 65535,65534,65533,65532,65531,65530,65529, // 2768 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2784 + + // #74 + 65535,65534,65533,65532,65531, + 75, 68, 69, 68, 73, 82, 83, 61, 47, 104, 111, 109, + 65535,65534,65533,65532,65531,65530,65529, // 2808 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2824 + + // #75 + 65535,65534,65533,65532,65531, + 76, 68, 95, 80, 82, 69, 76, 79, 65, 68, 61, 47, + 65535,65534,65533,65532,65531,65530,65529, // 2848 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2864 + + // #76 + 65535,65534,65533,65532,65531, + 88, 67, 85, 82, 83, 79, 82, 95, 80, 65, 84, 72, + 65535,65534,65533,65532,65531,65530,65529, // 2888 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2904 + + // #77 + 65535, + 115, 114, 99, 100, 105, 114, 61, 47, 104, 111, 109, 101, + 65535,65534,65533, // 2920 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2936 + + // #78 + 65535,65534,65533,65532,65531, + 72, 79, 77, 69, 61, 47, 104, 111, 109, 101, 47, 116, + 65535,65534,65533,65532,65531,65530,65529, // 2960 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 2976 + + // #79 + 65535, + 81, 84, 52, 68, 79, 67, 68, 73, 82, 61, 47, 117, + 65535,65534,65533, // 2992 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3008 + + // #80 + 65535,65534,65533,65532,65531, + 80, 87, 68, 61, 47, 104, 111, 109, 101, 47, 116, 109, + 65535,65534,65533,65532,65531,65530,65529, // 3032 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3048 + + // #81 + 65535,65534,65533,65532,65531, + 75, 68, 69, 95, 83, 69, 83, 83, 73, 79, 78, 95, + 65535,65534,65533,65532,65531,65530,65529, // 3072 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3088 + + // #82 + 65535, + 73, 78, 83, 73, 68, 69, 95, 83, 80, 69, 67, 73, + 65535,65534,65533, // 3104 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3120 + + // #83 + 65535, + 83, 83, 72, 95, 65, 71, 69, 78, 84, 95, 80, 73, + 65535,65534,65533, // 3136 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3152 + + // #84 + 65535, + 80, 75, 71, 95, 67, 79, 78, 70, 73, 71, 95, 80, + 65535,65534,65533, // 3168 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3184 + + // #85 + 65535,65534,65533,65532,65531, + 68, 66, 85, 83, 95, 83, 69, 83, 83, 73, 79, 78, + 65535,65534,65533,65532,65531,65530,65529, // 3208 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3224 + + // #86 + 65535, + 76, 68, 95, 76, 73, 66, 82, 65, 82, 89, 95, 80, + 65535,65534,65533, // 3240 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3256 + + // #87 + 65535,65534,65533,65532,65531, + 80, 52, 85, 83, 69, 82, 61, 116, 106, 109, 97, 99, + 65535,65534,65533,65532,65531,65530,65529, // 3280 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3296 + + // #88 + 65535,65534,65533,65532,65531, + 81, 84, 69, 83, 84, 95, 67, 79, 76, 79, 82, 69, + 65535,65534,65533,65532,65531,65530,65529, // 3320 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3336 + + // #89 + 65535,65534,65533,65532,65531, + 88, 68, 71, 95, 83, 69, 83, 83, 73, 79, 78, 95, + 65535,65534,65533,65532,65531,65530,65529, // 3360 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3376 + + // #90 + 65535,65534,65533,65532,65531, + 76, 69, 83, 83, 75, 69, 89, 61, 47, 101, 116, 99, + 65535,65534,65533,65532,65531,65530,65529, // 3400 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3416 + + // #91 + 65535, + 76, 79, 71, 78, 65, 77, 69, 61, 116, 109, 97, 99, + 65535,65534,65533, // 3432 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3448 + + // #92 + 65535, + 71, 95, 70, 73, 76, 69, 78, 65, 77, 69, 95, 69, + 65535,65534,65533, // 3464 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3480 + + // #93 + 65535, + 75, 68, 69, 95, 70, 85, 76, 76, 95, 83, 69, 83, + 65535,65534,65533, // 3496 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3512 + + // #94 + 65535,65534,65533,65532,65531, + 72, 79, 83, 84, 78, 65, 77, 69, 61, 108, 111, 116, + 65535,65534,65533,65532,65531,65530,65529, // 3536 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3552 + + // #95 + 65535,65534,65533,65532,65531, + 76, 67, 95, 84, 73, 77, 69, 61, 112, 116, 95, 66, + 65535,65534,65533,65532,65531,65530,65529, // 3576 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3592 + + // #96 + 65535, + 83, 83, 72, 95, 65, 83, 75, 80, 65, 83, 83, 61, + 65535,65534,65533, // 3608 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3624 + + // #97 + 65535,65534,65533,65532,65531, + 72, 73, 83, 84, 70, 73, 76, 69, 61, 47, 104, 111, + 65535,65534,65533,65532,65531,65530,65529, // 3648 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3664 + + // #98 + 65535, + 75, 79, 78, 83, 79, 76, 69, 95, 68, 66, 85, 83, + 65535,65534,65533, // 3680 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3696 + + // #99 + 65535, + 77, 65, 75, 69, 61, 47, 117, 115, 114, 47, 98, 105, + 65535,65534,65533, // 3712 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3728 + + // #100 + 65535, + 67, 65, 78, 66, 69, 82, 82, 65, 95, 68, 82, 73, + 65535,65534,65533, // 3744 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3760 + + // #101 + 65535, + 71, 67, 79, 78, 70, 95, 84, 77, 80, 68, 73, 82, + 65535,65534,65533, // 3776 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3792 + + // #102 + 65535,65534,65533,65532,65531, + 85, 83, 69, 82, 61, 116, 109, 97, 99, 105, 101, 105, + 65535,65534,65533,65532,65531,65530,65529, // 3816 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3832 + + // #103 + 65535, + 111, 98, 106, 100, 105, 114, 61, 47, 104, 111, 109, 101, + 65535,65534,65533, // 3848 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3864 + + // #104 + 65535,65534,65533,65532,65531, + 76, 67, 95, 77, 79, 78, 69, 84, 65, 82, 89, 61, + 65535,65534,65533,65532,65531,65530,65529, // 3888 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3904 + + // #105 + 65535,65534,65533,65532,65531, + 81, 84, 76, 73, 66, 61, 47, 117, 115, 114, 47, 108, + 65535,65534,65533,65532,65531,65530,65529, // 3928 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3944 + + // #106 + 65535,65534,65533,65532,65531, + 76, 67, 95, 84, 69, 76, 69, 80, 72, 79, 78, 69, + 65535,65534,65533,65532,65531,65530,65529, // 3968 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 3984 + + // #107 + 65535, + 80, 89, 84, 72, 79, 78, 68, 79, 78, 84, 87, 82, + 65535,65534,65533, // 4000 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4016 + + // #108 + 65535,65534,65533,65532,65531, + 84, 77, 80, 68, 73, 82, 61, 47, 116, 109, 112, 47, + 65535,65534,65533,65532,65531,65530,65529, // 4040 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4056 + + // #109 + 65535,65534,65533,65532,65531, + 65, 82, 77, 76, 77, 68, 95, 76, 73, 67, 69, 78, + 65535,65534,65533,65532,65531,65530,65529, // 4080 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4096 + + // #110 + 65535, + 80, 89, 84, 72, 79, 78, 80, 65, 84, 72, 61, 47, + 65535,65534,65533, // 4112 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4128 + + // #111 + 65535,65534,65533,65532,65531, + 77, 65, 75, 69, 70, 76, 65, 71, 83, 61, 119, 32, + 65535,65534,65533,65532,65531,65530,65529, // 4152 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4168 + + // #112 + 65535, + 77, 70, 76, 65, 71, 83, 61, 45, 119, 32, 45, 45, + 65535,65534,65533, // 4184 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4200 + + // #113 + 65535, + 77, 65, 73, 76, 61, 47, 118, 97, 114, 47, 115, 112, + 65535,65534,65533, // 4216 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4232 + + // #114 + 65535,65534,65533,65532,65531, + 83, 72, 69, 76, 76, 95, 83, 69, 83, 83, 73, 79, + 65535,65534,65533,65532,65531,65530,65529, // 4256 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4272 + + // #115 + 65535, + 75, 68, 69, 68, 73, 82, 61, 47, 104, 111, 109, 101, + 65535,65534,65533, // 4288 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4304 + + // #116 + 65535,65534,65533,65532,65531, + 76, 69, 83, 83, 67, 72, 65, 82, 83, 69, 84, 61, + 65535,65534,65533,65532,65531,65530,65529, // 4328 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4344 + + // #117 + 65535,65534,65533,65532,65531, + 76, 67, 95, 80, 65, 80, 69, 82, 61, 110, 98, 95, + 65535,65534,65533,65532,65531,65530,65529, // 4368 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4384 + + // #118 + 65535, + 66, 82, 79, 87, 83, 69, 82, 61, 47, 117, 115, 114, + 65535,65534,65533, // 4400 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4416 + + // #119 + 65535, + 77, 69, 84, 65, 95, 67, 76, 65, 83, 83, 61, 100, + 65535,65534,65533, // 4432 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4448 + + // #120 + 65535,65534,65533,65532,65531, + 77, 68, 86, 95, 77, 69, 78, 85, 95, 83, 84, 89, + 65535,65534,65533,65532,65531,65530,65529, // 4472 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4488 + + // #121 + 65535,65534,65533,65532,65531, + 67, 79, 76, 79, 82, 70, 71, 66, 71, 61, 49, 53, + 65535,65534,65533,65532,65531,65530,65529, // 4512 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4528 + + // #122 + 65535,65534,65533,65532,65531, + 80, 89, 84, 72, 79, 78, 83, 84, 65, 82, 84, 85, + 65535,65534,65533,65532,65531,65530,65529, // 4552 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4568 + + // #123 + 65535, + 76, 67, 95, 77, 69, 65, 83, 85, 82, 69, 77, 69, + 65535,65534,65533, // 4584 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4600 + + // #124 + 65535,65534,65533,65532,65531, + 69, 68, 73, 84, 79, 82, 61, 47, 117, 115, 114, 47, + 65535,65534,65533,65532,65531,65530,65529, // 4624 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4640 + + // #125 + 65535,65534,65533,65532,65531, + 69, 78, 95, 84, 66, 61, 109, 111, 99, 58, 117, 105, + 65535,65534,65533,65532,65531,65530,65529, // 4664 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4680 + + // #126 + 65535, + 72, 73, 83, 84, 83, 73, 90, 69, 61, 49, 48, 48, + 65535,65534,65533, // 4696 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4712 + + // #127 + 65535,65534,65533,65532,65531, + 71, 83, 95, 76, 73, 66, 61, 47, 104, 111, 109, 101, + 65535,65534,65533,65532,65531,65530,65529, // 4736 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4752 + + // #128 + 65535,65534,65533,65532,65531, + 78, 76, 83, 80, 65, 84, 72, 61, 47, 117, 115, 114, + 65535,65534,65533,65532,65531,65530,65529, // 4776 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4792 + + // #129 + 65535,65534,65533,65532,65531, + 87, 73, 78, 68, 79, 87, 80, 65, 84, 72, 61, 55, + 65535,65534,65533,65532,65531,65530,65529, // 4816 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4832 + + // #130 + 65535,65534,65533,65532,65531, + 75, 79, 78, 83, 79, 76, 69, 95, 68, 66, 85, 83, + 65535,65534,65533,65532,65531,65530,65529, // 4856 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4872 + + // #131 + 65535, + 76, 67, 95, 73, 68, 69, 78, 84, 73, 70, 73, 67, + 65535,65534,65533, // 4888 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4904 + + // #132 + 65535, + 73, 78, 80, 85, 84, 82, 67, 61, 47, 101, 116, 99, + 65535,65534,65533, // 4920 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4936 + + // #133 + 65535,65534,65533,65532,65531, + 81, 84, 73, 78, 67, 61, 47, 117, 115, 114, 47, 108, + 65535,65534,65533,65532,65531,65530,65529, // 4960 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 4976 + + // #134 + 65535, + 76, 67, 95, 65, 68, 68, 82, 69, 83, 83, 61, 110, + 65535,65534,65533, // 4992 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5008 + + // #135 + 65535,65534,65533,65532,65531, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 95, + 65535,65534,65533,65532,65531,65530,65529, // 5032 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5048 + + // #136 + 65535, + 76, 65, 78, 71, 61, 112, 116, 95, 66, 82, 46, 85, + 65535,65534,65533, // 5064 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5080 + + // #137 + 65535, + 80, 52, 80, 79, 82, 84, 61, 112, 52, 46, 116, 114, + 65535,65534,65533, // 5096 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5112 + + // #138 + 65535,65534,65533,65532,65531, + 80, 73, 76, 79, 84, 80, 79, 82, 84, 61, 117, 115, + 65535,65534,65533,65532,65531,65530,65529, // 5136 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5152 + + // #139 + 65535, + 75, 68, 69, 95, 83, 69, 83, 83, 73, 79, 78, 95, + 65535,65534,65533, // 5168 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5184 + + // #140 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5200 + 65535, + 84, 69, 65, 77, 66, 85, 73, 76, 68, 69, 82, 61, + 65535,65534,65533, // 5216 + + +}; +static struct StringCollection +{ + int len; + int offset1, offset2; + ushort align1, align2; +} stringCollection[] = { + {18, 1, 29, 3666, 106}, // #0 + {18, 53, 77, 106, 1978}, // #1 + {20, 97, 125, 2850, 3210}, // #2 + {20, 157, 185, 3210, 3138}, // #3 + {51, 225, 225, 3362, 3362}, // #4 + {51, 293, 353, 1434, 3362}, // #5 + {51, 417, 417, 3362, 3362}, // #6 + {19, 473, 501, 2850, 10}, // #7 + {19, 533, 565, 10, 442}, // #8 + {51, 609, 677, 3362, 1434}, // #9 + {5, 741, 753, 2666, 2066}, // #10 + {4, 765, 781, 2362, 3930}, // #11 + {3, 793, 805, 3330, 2138}, // #12 + {5, 817, 825, 738, 2066}, // #13 + {4, 833, 845, 434, 3930}, // #14 + {5, 857, 865, 3842, 2066}, // #15 + {4, 873, 885, 3538, 3930}, // #16 + {3, 897, 909, 3330, 2138}, // #17 + {5, 925, 937, 1898, 2066}, // #18 + {4, 949, 965, 1594, 3930}, // #19 + {3, 977, 989, 3330, 2138}, // #20 + {5, 1005, 1021, 2218, 762}, // #21 + {5, 1033, 1041, 3346, 2066}, // #22 + {4, 1053, 1069, 3082, 3930}, // #23 + {12, 1081, 1097, 2082, 962}, // #24 + {5, 1113, 1121, 3362, 2066}, // #25 + {4, 1129, 1141, 322, 3930}, // #26 + {3, 1153, 1165, 2050, 2138}, // #27 + {5, 1177, 1185, 1538, 2066}, // #28 + {4, 1193, 1205, 1234, 3930}, // #29 + {5, 1221, 1233, 554, 2066}, // #30 + {4, 1245, 1261, 250, 3930}, // #31 + {5, 1277, 1289, 2858, 2066}, // #32 + {4, 1301, 1317, 2554, 3930}, // #33 + {12, 1329, 1345, 2194, 1762}, // #34 + {12, 1365, 1385, 2170, 1762}, // #35 + {12, 1405, 1425, 2314, 1762}, // #36 + {6, 1445, 1461, 3626, 666}, // #37 + {11, 1477, 1501, 3882, 842}, // #38 + {12, 1525, 1545, 1722, 2930}, // #39 + {12, 1565, 1585, 1914, 2930}, // #40 + {12, 1605, 1625, 442, 2930}, // #41 + {12, 1641, 1657, 626, 2930}, // #42 + {12, 1673, 1689, 946, 2930}, // #43 + {12, 1705, 1721, 738, 2930}, // #44 + {12, 1737, 1753, 2066, 2930}, // #45 + {12, 1773, 1793, 1210, 2930}, // #46 + {12, 1809, 1825, 1426, 2930}, // #47 + {12, 1841, 1857, 1650, 2930}, // #48 + {12, 1877, 1897, 1530, 2930}, // #49 + {12, 1913, 1929, 1858, 2930}, // #50 + {12, 1949, 1969, 2106, 2930}, // #51 + {12, 1989, 2009, 2202, 2930}, // #52 + {12, 2029, 2049, 2490, 2930}, // #53 + {12, 2069, 2089, 2794, 2930}, // #54 + {12, 2105, 2121, 2322, 2930}, // #55 + {12, 2137, 2153, 2834, 2930}, // #56 + {12, 2169, 2185, 1266, 2930}, // #57 + {12, 2205, 2225, 2538, 2930}, // #58 + {12, 2241, 2257, 2706, 2930}, // #59 + {12, 2277, 2297, 3402, 2930}, // #60 + {12, 2313, 2329, 146, 2930}, // #61 + {12, 2349, 2369, 3690, 2930}, // #62 + {12, 2389, 2409, 810, 2930}, // #63 + {12, 2429, 2449, 1178, 2930}, // #64 + {12, 2465, 2481, 1442, 2930}, // #65 + {12, 2501, 2521, 3546, 2930}, // #66 + {12, 2541, 2561, 1930, 2930}, // #67 + {12, 2577, 2593, 1634, 2930}, // #68 + {12, 2609, 2625, 1986, 2930}, // #69 + {12, 2641, 2657, 1970, 2930}, // #70 + {12, 2677, 2697, 1834, 2930}, // #71 + {12, 2713, 2729, 1474, 2930}, // #72 + {12, 2749, 2769, 2250, 2930}, // #73 + {12, 2789, 2809, 2458, 2930}, // #74 + {12, 2829, 2849, 2618, 2930}, // #75 + {12, 2869, 2889, 3066, 2930}, // #76 + {12, 2905, 2921, 3330, 2930}, // #77 + {12, 2941, 2961, 1706, 2930}, // #78 + {12, 2977, 2993, 2802, 2930}, // #79 + {12, 3013, 3033, 3770, 2930}, // #80 + {12, 3053, 3073, 3594, 2930}, // #81 + {12, 3089, 3105, 2, 2930}, // #82 + {12, 3121, 3137, 2962, 2930}, // #83 + {12, 3153, 3169, 290, 2930}, // #84 + {12, 3189, 3209, 794, 2930}, // #85 + {12, 3225, 3241, 1058, 2930}, // #86 + {12, 3261, 3281, 2394, 2930}, // #87 + {12, 3301, 3321, 138, 2930}, // #88 + {12, 3341, 3361, 1482, 2930}, // #89 + {12, 3381, 3401, 570, 2930}, // #90 + {12, 3417, 3433, 674, 2930}, // #91 + {12, 3449, 3465, 1282, 2930}, // #92 + {12, 3481, 3497, 1746, 2930}, // #93 + {12, 3517, 3537, 1866, 2930}, // #94 + {12, 3557, 3577, 1978, 2930}, // #95 + {12, 3593, 3609, 3954, 2930}, // #96 + {12, 3629, 3649, 2570, 2930}, // #97 + {12, 3665, 3681, 2754, 2930}, // #98 + {12, 3697, 3713, 3666, 2930}, // #99 + {12, 3729, 3745, 34, 2930}, // #100 + {12, 3761, 3777, 2914, 2930}, // #101 + {12, 3797, 3817, 1194, 2930}, // #102 + {12, 3833, 3849, 3202, 2930}, // #103 + {12, 3869, 3889, 3018, 2930}, // #104 + {12, 3909, 3929, 202, 2930}, // #105 + {12, 3949, 3969, 3546, 2930}, // #106 + {12, 3985, 4001, 3682, 2930}, // #107 + {12, 4021, 4041, 3466, 2930}, // #108 + {12, 4061, 4081, 4074, 2930}, // #109 + {12, 4097, 4113, 306, 2930}, // #110 + {12, 4133, 4153, 634, 2930}, // #111 + {12, 4169, 4185, 802, 2930}, // #112 + {12, 4201, 4217, 962, 2930}, // #113 + {12, 4237, 4257, 1114, 2930}, // #114 + {12, 4273, 4289, 1250, 2930}, // #115 + {12, 4309, 4329, 3898, 2930}, // #116 + {12, 4349, 4369, 1386, 2930}, // #117 + {12, 4385, 4401, 1586, 2930}, // #118 + {12, 4417, 4433, 1730, 2930}, // #119 + {12, 4453, 4473, 1914, 2930}, // #120 + {12, 4493, 4513, 1498, 2930}, // #121 + {12, 4533, 4553, 2138, 2930}, // #122 + {12, 4569, 4585, 2290, 2930}, // #123 + {12, 4605, 4625, 2426, 2930}, // #124 + {12, 4645, 4665, 2666, 2930}, // #125 + {12, 4681, 4697, 2050, 2930}, // #126 + {12, 4717, 4737, 2874, 2930}, // #127 + {12, 4757, 4777, 3018, 2930}, // #128 + {12, 4797, 4817, 1834, 2930}, // #129 + {12, 4837, 4857, 3178, 2930}, // #130 + {12, 4873, 4889, 3314, 2930}, // #131 + {12, 4905, 4921, 2546, 2930}, // #132 + {12, 4941, 4961, 3546, 2930}, // #133 + {12, 4977, 4993, 3682, 2930}, // #134 + {12, 5013, 5033, 3802, 2930}, // #135 + {12, 5049, 5065, 3922, 2930}, // #136 + {12, 5081, 5097, 4018, 2930}, // #137 + {12, 5117, 5137, 42, 2930}, // #138 + {12, 5153, 5169, 130, 2930}, // #139 + {12, 5185, 5201, 242, 2930}, // #140 +}; +static const int stringCollectionCount = 141; +static const int stringCollectionMaxLen = 51; +// average comparison length: 12.0922 +// cache-line crosses: 6 (2.1%) +// alignment histogram: +// 0xXXX2 = 188 (66.7%) strings, 57 (30.3%) of which same-aligned +// 0xXXXa = 94 (33.3%) strings, 10 (10.6%) of which same-aligned +// total = 282 (100%) strings, 67 (23.8%) of which same-aligned diff --git a/tests/benchmarks/corelib/tools/qstring/main.cpp b/tests/benchmarks/corelib/tools/qstring/main.cpp index fb65f65..88dc40b 100644 --- a/tests/benchmarks/corelib/tools/qstring/main.cpp +++ b/tests/benchmarks/corelib/tools/qstring/main.cpp @@ -48,8 +48,15 @@ #define SRCDIR "" #endif +#ifdef Q_OS_UNIX +#include <sys/mman.h> +#include <unistd.h> +#endif + #include <private/qsimd_p.h> +#include "data.cpp" + class tst_QString: public QObject { Q_OBJECT @@ -73,20 +80,8 @@ void tst_QString::equals() const } } -static ushort databuffer[4096]; - tst_QString::tst_QString() { - // populate databuffer with our seed, each byte 3 times in a row - // include the NUL! - static const char seed[] = "AAAAAAAAAEhlbGxvIFdvcmxkIAAAAAA="; - static const int repeat = 3; - int pos = 0; - for (ushort *p = databuffer; p < databuffer + (sizeof(databuffer) / sizeof(databuffer[0])); p += repeat) { - for (int j = 0; j < repeat; ++p, ++j) - *p = seed[pos]; - pos = (pos + 1) % sizeof(seed); - } } void tst_QString::equals_data() const @@ -148,13 +143,15 @@ void tst_QString::equals_data() const << QString::fromRawData(ptr + 1, 58) << QString::fromRawData(ptr + 3, 58); } -static bool equals2_memcmp_call(ushort *p1, ushort *p2, int len) +static bool equals2_memcmp_call(const ushort *p1, const ushort *p2, int len) { return memcmp(p1, p2, len * 2) == 0; } -static bool equals2_bytewise(ushort *p1, ushort *p2, int len) +static bool equals2_bytewise(const ushort *p1, const ushort *p2, int len) { + if (p1 == p2 || !len) + return true; uchar *b1 = (uchar *)p1; uchar *b2 = (uchar *)p2; len *= 2; @@ -164,17 +161,24 @@ static bool equals2_bytewise(ushort *p1, ushort *p2, int len) return true; } -static bool equals2_shortwise(ushort *p1, ushort *p2, int len) +static bool __attribute__((optimize("unroll-loops"))) equals2_shortwise(const ushort *p1, const ushort *p2, int len) { - register ushort * const end = p1 + len; - for ( ; p1 != end; ++p1, ++p2) - if (*p1 != *p2) + if (p1 == p2 || !len) + return true; +// for (register int counter; counter < len; ++counter) +// if (p1[counter] != p2[counter]) +// return false; + while (len--) { + if (p1[len] != p2[len]) return false; + } return true; } -static bool equals2_intwise(ushort *p1, ushort *p2, int length) +static bool equals2_intwise(const ushort *p1, const ushort *p2, int length) { + if (p1 == p2 || !length) + return true; register union { const quint16 *w; const quint32 *d; @@ -219,71 +223,204 @@ static bool equals2_intwise(ushort *p1, ushort *p2, int length) return true; } +static inline bool equals2_short_tail(const ushort *p1, const ushort *p2, int len) +{ + if (len) { + if (*p1 != *p2) + return false; + if (--len) { + if (p1[1] != p2[1]) + return false; + if (--len) { + if (p1[2] != p2[2]) + return false; + if (--len) { + if (p1[3] != p2[3]) + return false; + if (--len) { + if (p1[4] != p2[4]) + return false; + if (--len) { + if (p1[5] != p2[5]) + return false; + if (--len) { + if (p1[6] != p2[6]) + return false; + return p1[7] == p2[7]; + } + } + } + } + } + } + } + return true; +} + +#pragma GCC optimize("no-unroll-loops") #ifdef __SSE2__ -static bool equals2_sse2(ushort *p1, ushort *p2, int len) +static bool equals2_sse2_aligned(const ushort *p1, const ushort *p2, int len) { - if (len > 8) { + if (len >= 8) { + qptrdiff counter = 0; while (len > 8) { - __m128i q1 = _mm_loadu_si128((__m128i *)p1); - __m128i q2 = _mm_loadu_si128((__m128i *)p2); + __m128i q1 = _mm_load_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_load_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != ushort(0xffff)) + return false; + + len -= 8; + counter += 8; + } + p1 += counter; + p2 += counter; + } + + return equals2_short_tail(p1, p2, len); +} + +static bool __attribute__((optimize("no-unroll-loops"))) equals2_sse2(const ushort *p1, const ushort *p2, int len) +{ + if (p1 == p2 || !len) + return true; + + if (len >= 8) { + qptrdiff counter = 0; + while (len >= 8) { + __m128i q1 = _mm_loadu_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 + counter)); __m128i cmp = _mm_cmpeq_epi16(q1, q2); if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) return false; len -= 8; - p1 += 8; - p2 += 8; + counter += 8; } + p1 += counter; + p2 += counter; } - return equals2_shortwise(p1, p2, len); + return equals2_short_tail(p1, p2, len); } -static inline -#ifdef Q_CC_GNU -__attribute__((always_inline)) -#endif -bool prolog_align(ushort *&p1, ushort *&p2, int &len) +//static bool equals2_sse2(const ushort *p1, const ushort *p2, int len) +//{ +// register int val1 = quintptr(p1) & 0xf; +// register int val2 = quintptr(p2) & 0xf; +// if (false && val1 + val2 == 0) +// return equals2_sse2_aligned(p1, p2, len); +// else +// return equals2_sse2_unaligned(p1, p2, len); +//} + +static bool equals2_sse2_aligning(const ushort *p1, const ushort *p2, int len) { - const ushort *end = (ushort*) ((quintptr(p1) + 15) & ~15); - if (end > p1 + len) - end = p1 + len; - for ( ; p1 != end; ++p1, ++p2, --len) - if (*p1 != *p2) + if (len < 8) + return equals2_short_tail(p1, p2, len); + + qptrdiff counter = 0; + + // which one is easier to align, p1 or p2 ? + register int val1 = quintptr(p1) & 0xf; + register int val2 = quintptr(p2) & 0xf; + if (val1 && val2) { +#if 0 + // we'll align the one which requires the least number of steps + if (val1 > val2) { + qSwap(p1, p2); + val1 = val2; + } + + // val1 contains the number of bytes past the 16-aligned mark + // we must read 16-val1 bytes to align + val1 = 16 - val1; + if (val1 & 0x2) { + if (*p1 != *p2) + return false; + --len; + ++counter; + } + while (val1 & 12) { + if (*(uint*)p1 != *(uint*)p2) + return false; + --len; + counter += 2; + val1 -= 4; + } +#else + // we'll align the one closest to the 16-byte mark + if (val1 > val2) { + qSwap(p1, p2); + val1 = val2; + } + + // we're reading val1 bytes too many + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 - val1/2)); + __m128i cmp = _mm_cmpeq_epi16(*(__m128i *)(p1 - val1/2), q2); + if (short(_mm_movemask_epi8(cmp)) >> val1 != short(-1)) return false; - return true; + + counter = 8 - val1/2; + len -= 8 - val1/2; +#endif + } else if (!val2) { + // p2 is already aligned + qSwap(p1, p2); + } + + // p1 is aligned + + while (len >= 8) { + __m128i q1 = _mm_load_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_loadu_si128((__m128i *)(p2 + counter)); + __m128i cmp = _mm_cmpeq_epi16(q1, q2); + if (ushort(_mm_movemask_epi8(cmp)) != ushort(0xffff)) + return false; + + len -= 8; + counter += 8; + } + + // tail + return equals2_short_tail(p1 + counter, p2 + counter, len); } -static bool equals2_sse2_aligning(ushort *p1, ushort *p2, int len) +#ifdef __SSE3__ +static bool __attribute__((optimize("no-unroll-loops"))) equals2_sse3(const ushort *p1, const ushort *p2, int len) { - if (len > 8) { - if (!prolog_align(p1, p2, len)) - return false; - while (len > 8) { - __m128i q1 = _mm_load_si128((__m128i *)p1); - __m128i q2 = _mm_loadu_si128((__m128i *)p2); + if (p1 == p2 || !len) + return true; + + if (len >= 8) { + qptrdiff counter = 0; + while (len >= 8) { + __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 + counter)); + __m128i q2 = _mm_lddqu_si128((__m128i *)(p2 + counter)); __m128i cmp = _mm_cmpeq_epi16(q1, q2); if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) return false; len -= 8; - p1 += 8; - p2 += 8; + counter += 8; } + p1 += counter; + p2 += counter; } - return equals2_shortwise(p1, p2, len); + return equals2_short_tail(p1, p2, len); } -template<int N> static inline bool equals2_ssse3_alignr(__m128i *m1, __m128i *m2, int len) +#ifdef __SSSE3__ +template<int N> static __attribute__((optimize("unroll-loops"))) inline bool equals2_ssse3_alignr(__m128i *m1, __m128i *m2, int len) { __m128i lower = _mm_load_si128(m1); - while (len > 8) { + while (len >= 8) { __m128i upper = _mm_load_si128(m1 + 1); __m128i correct; correct = _mm_alignr_epi8(upper, lower, N); - __m128i q2 = _mm_loadu_si128(m2); + __m128i q2 = _mm_lddqu_si128(m2); __m128i cmp = _mm_cmpeq_epi16(correct, q2); if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) return false; @@ -295,13 +432,13 @@ template<int N> static inline bool equals2_ssse3_alignr(__m128i *m1, __m128i *m2 } // tail - return len == 0 || equals2_shortwise((ushort *)m1 + N / 2, (ushort*)m2, len); + return len == 0 || equals2_short_tail((const ushort *)m1 + N / 2, (const ushort*)m2, len); } -static inline bool equals2_ssse3_aligned(__m128i *m1, __m128i *m2, int len) +static inline __attribute__((optimize("unroll-loops"))) bool equals2_ssse3_aligned(__m128i *m1, __m128i *m2, int len) { - while (len > 8) { - __m128i q2 = _mm_loadu_si128(m2); + while (len >= 8) { + __m128i q2 = _mm_lddqu_si128(m2); __m128i cmp = _mm_cmpeq_epi16(*m1, q2); if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) return false; @@ -310,30 +447,30 @@ static inline bool equals2_ssse3_aligned(__m128i *m1, __m128i *m2, int len) ++m1; ++m2; } - return len == 0 || equals2_shortwise((ushort *)m1, (ushort *)m2, len); + return len == 0 || equals2_short_tail((const ushort *)m1, (const ushort *)m2, len); } -//#ifdef __SSSE3__ -static bool equals2_ssse3(ushort *p1, ushort *p2, int len) +static bool equals2_ssse3(const ushort *p1, const ushort *p2, int len) { // p1 & 0xf can be: // 0, 2, 4, 6, 8, 10, 12, 14 // If it's 0, we're aligned // If it's not, then we're interested in the 16 - (p1 & 0xf) bytes only - if (len > 8) { + if (len >= 8) { // find the last aligned position below the p1 memory __m128i *m1 = (__m128i *)(quintptr(p1) & ~0xf); __m128i *m2 = (__m128i *)p2; - uchar diff = quintptr(p1) - quintptr(m1); + qptrdiff diff = quintptr(p1) - quintptr(m1); // diff contains the number of extra bytes + if (diff == 10) + return equals2_ssse3_alignr<10>(m1, m2, len); + else if (diff == 2) + return equals2_ssse3_alignr<2>(m1, m2, len); if (diff < 8) { if (diff < 4) { - if (diff == 0) - return equals2_ssse3_aligned(m1, m2, len); - else // diff == 2 - return equals2_ssse3_alignr<2>(m1, m2, len); + return equals2_ssse3_aligned(m1, m2, len); } else { if (diff == 4) return equals2_ssse3_alignr<4>(m1, m2, len); @@ -342,10 +479,7 @@ static bool equals2_ssse3(ushort *p1, ushort *p2, int len) } } else { if (diff < 12) { - if (diff == 8) - return equals2_ssse3_alignr<8>(m1, m2, len); - else // diff == 10 - return equals2_ssse3_alignr<10>(m1, m2, len); + return equals2_ssse3_alignr<8>(m1, m2, len); } else { if (diff == 12) return equals2_ssse3_alignr<12>(m1, m2, len); @@ -353,34 +487,99 @@ static bool equals2_ssse3(ushort *p1, ushort *p2, int len) return equals2_ssse3_alignr<14>(m1, m2, len); } } + } + + // tail + return equals2_short_tail(p1, p2, len); +} + +template<int N> static inline bool equals2_ssse3_aligning_alignr(__m128i *m1, __m128i *m2, int len) +{ + __m128i lower = _mm_load_si128(m1); + while (len >= 8) { + __m128i upper = _mm_load_si128(m1 + 1); + __m128i correct; + correct = _mm_alignr_epi8(upper, lower, N); + + __m128i cmp = _mm_cmpeq_epi16(correct, *m2); + if (ushort(_mm_movemask_epi8(cmp)) != 0xffff) + return false; -// switch (diff) { -// case 0: -// return equals2_ssse3_aligned(m1, m2, len); -// case 2: -// return equals2_ssse3_alignr<2>(m1, m2, len); -// case 4: -// return equals2_ssse3_alignr<4>(m1, m2, len); -// case 6: -// return equals2_ssse3_alignr<6>(m1, m2, len); -// case 8: -// return equals2_ssse3_alignr<8>(m1, m2, len); -// case 10: -// return equals2_ssse3_alignr<10>(m1, m2, len); -// case 12: -// return equals2_ssse3_alignr<12>(m1, m2, len); -// case 14: -// return equals2_ssse3_alignr<14>(m1, m2, len); -// } + len -= 8; + ++m2; + ++m1; + lower = upper; } // tail - return equals2_shortwise(p1, p2, len); + return len == 0 || equals2_short_tail((const ushort *)m1 + N / 2, (const ushort*)m2, len); } -//#endif -//#ifdef __SSE4_1__ -static bool equals2_sse4(ushort *p1, ushort *p2, int len) +static bool equals2_ssse3_aligning(const ushort *p1, const ushort *p2, int len) +{ + if (len < 8) + return equals2_short_tail(p1, p2, len); + qptrdiff counter = 0; + + // which one is easier to align, p1 or p2 ? + { + register int val1 = quintptr(p1) & 0xf; + register int val2 = quintptr(p2) & 0xf; + if (val1 && val2) { + // we'll align the one closest to the 16-byte mark + if (val1 < val2) { + qSwap(p1, p2); + val2 = val1; + } + + // we're reading val1 bytes too many + __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 - val2/2)); + __m128i cmp = _mm_cmpeq_epi16(q1, *(__m128i *)(p2 - val2/2)); + if (short(_mm_movemask_epi8(cmp)) >> val1 != short(-1)) + return false; + + counter = 8 - val2/2; + len -= 8 - val2/2; + } else if (!val1) { + // p1 is already aligned + qSwap(p1, p2); + } + } + + // p2 is aligned now + // we want to use palignr in the mis-alignment of p1 + __m128i *m1 = (__m128i *)(quintptr(p1 + counter) & ~0xf); + __m128i *m2 = (__m128i *)(p2 + counter); + register int val1 = quintptr(p1 + counter) - quintptr(m1); + + // val1 contains the number of extra bytes + if (val1 == 8) + return equals2_ssse3_aligning_alignr<8>(m1, m2, len); + if (val1 == 0) + return equals2_sse2_aligned(p1 + counter, p2 + counter, len); + if (val1 < 8) { + if (val1 < 4) { + return equals2_ssse3_aligning_alignr<2>(m1, m2, len); + } else { + if (val1 == 4) + return equals2_ssse3_aligning_alignr<4>(m1, m2, len); + else // diff == 6 + return equals2_ssse3_aligning_alignr<6>(m1, m2, len); + } + } else { + if (val1 < 12) { + return equals2_ssse3_aligning_alignr<10>(m1, m2, len); + } else { + if (val1 == 12) + return equals2_ssse3_aligning_alignr<12>(m1, m2, len); + else // diff == 14 + return equals2_ssse3_aligning_alignr<14>(m1, m2, len); + } + } +} + +#ifdef __SSE4_1__ +static bool equals2_sse4(const ushort *p1, const ushort *p2, int len) { // We use the pcmpestrm instruction searching for differences (negative polarity) // it will reset CF if it's all equal @@ -390,36 +589,150 @@ static bool equals2_sse4(ushort *p1, ushort *p2, int len) // difference found: CF = 1 // all equal, not finished: CF = ZF = SF = 0 // all equal, finished: CF = 0, ZF = SF = 1 + // We use the JA instruction that jumps if ZF = 0 and CF = 0 + if (p1 == p2 || !len) + return true; + + // This function may read some bytes past the end of p1 or p2 + // It is safe to do that, as long as those extra bytes (beyond p1+len and p2+len) + // are on the same page as the last valid byte. + // If len is a multiple of 8, we'll never load invalid bytes. + if (len & 7) { + // The last load would load (len & ~7) valid bytes and (8 - (len & ~7)) invalid bytes. + // So we can't do the last load if any of those bytes is in a different + // page. That is, if: + // pX + len is on a different page from pX + (len & ~7) + 8 + // + // that is, if second-to-last load ended up less than 16 bytes from the page end: + // pX + (len & ~7) is the last ushort read in the second-to-last load + if (len < 8) + return equals2_short_tail(p1, p2, len); + if ((quintptr(p1 + (len & ~7)) & 0xfff) > 0xff0 || + (quintptr(p2 + (len & ~7)) & 0xfff) > 0xff0) { + + // yes, so we mustn't do the final 128-bit load + bool result; + asm ( + "sub %[p1], %[p2]\n\t" + "sub $16, %[p1]\n\t" + "add $8, %[len]\n\t" + + // main loop: + "0:\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p2],%[p1]), %%xmm0\n\t" + + "jna 1f\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p2],%[p1]), %%xmm0\n\t" + + "ja 0b\n\t" + "1:\n\t" + "setnc %[result]\n\t" + : [result] "=a" (result), + [p1] "+r" (p1), + [p2] "+r" (p2) + : [len] "0" (len & ~7), + [mode] "i" (_SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY) + : "%edx", "%ecx", "%xmm0" + ); + return result && equals2_short_tail(p1, (const ushort *)(quintptr(p1) + quintptr(p2)), len & 7); + } + } + +// const qptrdiff disp = p2 - p1; +// p1 -= 8; +// len += 8; +// while (true) { +// enum { Mode = _SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY }; + +// p1 += 8; +// len -= 8; +// if (!len) +// return true; + +// __m128i q1 = _mm_lddqu_si128((__m128i *)(p1 + disp)); +// __m128i *m2 = (__m128i *)p1; + +// bool cmp_a = _mm_cmpestra(q1, len, *m2, len, Mode); +// if (cmp_a) +// continue; +// return !_mm_cmpestrc(q1, len, *m2, len, Mode); +// } +// return true; bool result; asm ( - "movd %%ebx, %%xmm1\n\t" "sub %[p1], %[p2]\n\t" - "mov %[p1], %%ebx\n\t" - "sub $16, %%ebx\n\t" + "sub $16, %[p1]\n\t" "add $8, %[len]\n\t" - "0:\n\t" - "add $16, %%ebx\n\t" + + "0:\n\t" + "add $16, %[p1]\n\t" "sub $8, %[len]\n\t" - "movdqu (%%ebx), %%xmm0\n\t" + "jz 1f\n\t" + "lddqu (%[p2],%[p1]), %%xmm0\n\t" "mov %[len], %%edx\n\t" - "pcmpestrm %[mode], (%[p2],%%ebx), %%xmm0\n\t" + "pcmpestri %[mode], (%[p1]), %%xmm0\n\t" + + "jna 1f\n\t" + "add $16, %[p1]\n\t" + "sub $8, %[len]\n\t" + "jz 1f\n\t" + "lddqu (%[p2],%[p1]), %%xmm0\n\t" + "mov %[len], %%edx\n\t" + "pcmpestri %[mode], (%[p1]), %%xmm0\n\t" + "ja 0b\n\t" - "1:\n\t" - "mov $0, %%eax\n\t" - "setnc %%al\n\t" - "movd %%xmm1, %%ebx\n\t" + + "1:\n\t" + "setnc %[result]\n\t" : [result] "=a" (result) : [len] "0" (len), - [p1] "d" (p1), + [p1] "r" (p1), [p2] "r" (p2), - [mode] "K" (_SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY) - : "%xmm0", "%xmm1" + [mode] "i" (_SIDD_UWORD_OPS | _SIDD_CMP_EQUAL_EACH | _SIDD_NEGATIVE_POLARITY) + : "%edx", "%ecx", "%xmm0" ); return result; } -//#endif + +#endif +#endif +#endif #endif +typedef bool (* FuncPtr)(const ushort *, const ushort *, int); +static const FuncPtr func[] = { + equals2_memcmp_call, // 0 + equals2_bytewise, // 1 + equals2_shortwise, // 1 + equals2_intwise, // 3 +#ifdef __SSE2__ + equals2_sse2, // 4 + equals2_sse2_aligning, // 5 +#ifdef __SSE3__ + equals2_sse3, // 6 +#ifdef __SSSE3__ + equals2_ssse3, // 7 + equals2_ssse3, // 8 +#ifdef __SSE4_1__ + equals2_sse4, // 9 +#endif +#endif +#endif +#endif + 0 +}; +static const int functionCount = sizeof(func)/sizeof(func[0]) - 1; + void tst_QString::equals2_data() const { QTest::addColumn<int>("algorithm"); @@ -431,108 +744,94 @@ void tst_QString::equals2_data() const #ifdef __SSE2__ QTest::newRow("sse2") << 4; QTest::newRow("sse2_aligning") << 5; +#ifdef __SSE3__ + QTest::newRow("sse3") << 6; #ifdef __SSSE3__ - QTest::newRow("ssse3") << 6; + QTest::newRow("ssse3") << 7; + QTest::newRow("ssse3_aligning") << 8; #ifdef __SSE4_1__ - QTest::newRow("sse4.2") << 7; + QTest::newRow("sse4.2") << 9; +#endif #endif #endif #endif } -void tst_QString::equals2() const +static void __attribute__((noinline)) equals2_selftest() { - static const short positions[] = { - 190, 1719, 2149, 1752, - 158, 244, 365, 1117, - 254, 265, 1047, 1785, - 1435, 552, 1476, 2030, - // 16 - 421, 1840, 2209, 232, - 1389, 907, 1500, 1479, - 1152, 541, 655, 1960, - 1642, 299, 740, 1995, - // 32 - 1946, 1407, 1272, 1946, - 1459, 1851, 1717, 1484, - 1761, 1630, 1377, 1675, - 629, 341, 661, 244 - // 48 - }; - // the length list must not contain 0 - static const int lens[] = { - 11, // 0 - 40, - 28, - 38, - 9, - 52, // 5 - 48, - 38, - 29, - 7, - 2, // 10 - 49, - 41, - 5, - 20, - 62 // 15 +#ifdef Q_OS_UNIX + const long pagesize = sysconf(_SC_PAGESIZE); + void *page1, *page3; + ushort *page2; + page1 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + page2 = (ushort *)mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE, -1, 0); + page3 = mmap(0, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + + Q_ASSERT(quintptr(page2) == quintptr(page1) + pagesize || quintptr(page2) == quintptr(page1) - pagesize); + Q_ASSERT(quintptr(page3) == quintptr(page2) + pagesize || quintptr(page3) == quintptr(page2) - pagesize); + munmap(page1, pagesize); + munmap(page3, pagesize); + + // populate our page + for (uint i = 0; i < pagesize / sizeof(long long); ++i) + ((long long *)page2)[i] = Q_INT64_C(0x0041004100410041); + + // the following should crash: + //page2[-1] = 0xdead; + //page2[pagesize / sizeof(ushort) + 1] = 0xbeef; + + static const ushort needle[] = { + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, 0x41, + 0x41 }; - typedef bool (* FuncPtr)(ushort *, ushort *, int); - static const FuncPtr func[] = { - equals2_memcmp_call, // 0 - equals2_bytewise, // 1 - equals2_shortwise, // 1 - equals2_intwise, // 3 -#ifdef __SSE2__ - equals2_sse2, // 4 - equals2_sse2_aligning, // 5 -#ifdef __SSSE3__ - equals2_ssse3, // 6 -#ifdef __SSE4_1__ - equals2_sse4, // 7 -#endif -#endif + for (int algo = 0; algo < functionCount; ++algo) { + // boundary condition test: + for (int i = 0; i < 8; ++i) { + (func[algo])(page2 + i, needle, sizeof needle / 2); + (func[algo])(page2 - i - 1 - sizeof(needle)/2 + pagesize/2, needle, sizeof needle/2); + } + } + + munmap(page2, pagesize); #endif - 0 - }; + for (int algo = 0; algo < functionCount; ++algo) { + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + bool expected = memcmp(p1, p2, stringCollection[i].len * 2) == 0; + + bool result = (func[algo])(p1, p2, stringCollection[i].len); + if (expected != result) + qWarning().nospace() + << "algo=" << algo + << " i=" << i + << " failed (" << result << "!=" << expected + << "); strings were " + << QByteArray((char*)p1, stringCollection[i].len).toHex() + << " and " + << QByteArray((char*)p2, stringCollection[i].len).toHex(); + } + } +} + +void tst_QString::equals2() const +{ QFETCH(int, algorithm); if (algorithm == -1) { - for (uint pos1 = 0; pos1 < sizeof positions / sizeof positions[0]; ++pos1) - for (uint pos2 = 0; pos2 < (sizeof positions / sizeof positions[0]) - 32; ++pos2) - for (uint len = 0; len < sizeof lens / sizeof lens[0]; ++len) { - ushort *p1 = databuffer + positions[pos1]; - ushort *p2 = databuffer + positions[pos2]; - bool expected = memcmp(p1, p2, lens[len] * 2) == 0; - - for (uint algo = 0; algo < -1 + (sizeof func / sizeof func[0]); ++algo) { - bool result = (func[algo])(p1, p2, lens[len]); - if (expected != result) - qWarning().nospace() - << "algo=" << algo - << " pos1=" << positions[pos1] - << " pos2=" << positions[pos2] - << " len=" << lens[len] - << " failed (" << result << "!=" << expected - << "); strings were " - << QByteArray((char*)p1, lens[len]).toHex() - << " and " - << QByteArray((char*)p2, lens[len]).toHex(); - } - - } + equals2_selftest(); return; } QBENCHMARK { - for (uint pos1 = 0; pos1 < sizeof positions / sizeof positions[0]; ++pos1) - for (uint pos2 = 0; pos2 < (sizeof positions / sizeof positions[0]) - 32; ++pos2) - for (uint len = 0; len < sizeof lens / sizeof lens[0]; ++len) { - bool result = (func[algorithm])(databuffer + positions[pos1], databuffer + positions[pos2], lens[len]); - Q_UNUSED(result); - } + for (int i = 0; i < stringCollectionCount; ++i) { + const ushort *p1 = stringCollectionData + stringCollection[i].offset1; + const ushort *p2 = stringCollectionData + stringCollection[i].offset2; + bool result = (func[algorithm])(p1, p2, stringCollection[i].len); + Q_UNUSED(result); + } } } |