https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59354
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|[4.8/4.9/5/Regression] |[4.8/4.9/5/Regression]
|Element swizzling produces |Element swizzling produces
|invalid result with -O3 |invalid result with -O3
|(tree-cunrolli pass |
|problem?) |
--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
This is tree vectorizer problem, involving VECT_PACK_TRUNC expressions.
Initializing the source array with:
for (i = 0; i < 256; i++)
a[i] = i;
we can see the difference in the destination array:
-O2 (correct):
0, 1, 2, 3, 16, 17, 18, 19, 32, 33, 34, 35, 48, 49, 50, 51, 64, 65, 66, 67, 80,
81, 82, 83, 96, 97, 98, 99, 112, 113, 114, 115, 128, 129, 130, 131, 144, 145,
146, 147, 160, 161, 162, 163, 176, 177, 178, 179, 192, 193, 194, 195, 208, 209,
210, 211, 224, 225, 226, 227, 240, 241, 242, 243, 4, 5, 6, 7, 20, 21, 22, 23,
36, 37, 38, 39, 52, 53, 54, 55, 68, 69, 70, 71, 84, 85, 86, 87, 100, 101, 102,
103, 116, 117, 118, 119, 132, 133, 134, 135, 148, 149, 150, 151, 164, 165, 166,
167, 180, 181, 182, 183, 196, 197, 198, 199, 212, 213, 214, 215, 228, 229, 230,
231, 244, 245, 246, 247, 8, 9, 10, 11, 24, 25, 26, 27, 40, 41, 42, 43, 56, 57,
58, 59, 72, 73, 74, 75, 88, 89, 90, 91, 104, 105, 106, 107, 120, 121, 122, 123,
136, 137, 138, 139, 152, 153, 154, 155, 168, 169, 170, 171, 184, 185, 186, 187,
200, 201, 202, 203, 216, 217, 218, 219, 232, 233, 234, 235, 248, 249, 250, 251,
12, 13, 14, 15, 28, 29, 30, 31, 44, 45, 46, 47, 60, 61, 62, 63, 76, 77, 78, 79,
92, 93, 94, 95, 108, 109, 110, 111, 124, 125, 126, 127, 140, 141, 142, 143,
156, 157, 158, 159, 172, 173, 174, 175, 188, 189, 190, 191, 204, 205, 206, 207,
220, 221, 222, 223, 236, 237, 238, 239, 252, 253, 254, 255
-O3 (incorrect):
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41,
42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61,
62, 63, 4, 5, 6, 7, 20, 21, 22, 23, 36, 37, 38, 39, 52, 53, 54, 55, 68, 69, 70,
71, 84, 85, 86, 87, 100, 101, 102, 103, 116, 117, 118, 119, 132, 133, 134, 135,
148, 149, 150, 151, 164, 165, 166, 167, 180, 181, 182, 183, 196, 197, 198, 199,
212, 213, 214, 215, 228, 229, 230, 231, 244, 245, 246, 247, 8, 9, 10, 11, 24,
25, 26, 27, 40, 41, 42, 43, 56, 57, 58, 59, 72, 73, 74, 75, 88, 89, 90, 91,
104, 105, 106, 107, 120, 121, 122, 123, 136, 137, 138, 139, 152, 153, 154, 155,
168, 169, 170, 171, 184, 185, 186, 187, 200, 201, 202, 203, 216, 217, 218, 219,
232, 233, 234, 235, 248, 249, 250, 251, 12, 13, 14, 15, 28, 29, 30, 31, 44, 45,
46, 47, 60, 61, 62, 63, 76, 77, 78, 79, 92, 93, 94, 95, 108, 109, 110, 111,
124, 125, 126, 127, 140, 141, 142, 143, 156, 157, 158, 159, 172, 173, 174, 175,
188, 189, 190, 191, 204, 205, 206, 207, 220, 221, 222, 223, 236, 237, 238, 239,
252, 253, 254, 255
b[4] is already wrong.
This is the beginning of the optimized dump:
vect__24.6_185 = MEM[(unsigned int *)&a];
vect__24.7_182 = MEM[(unsigned int *)&a + 16B];
vect__24.8_176 = MEM[(unsigned int *)&a + 32B];
vect__24.9_173 = MEM[(unsigned int *)&a + 48B];
vect__24.10_160 = MEM[(unsigned int *)&a + 64B];
vect__24.11_158 = MEM[(unsigned int *)&a + 80B];
vect__24.12_155 = MEM[(unsigned int *)&a + 96B];
vect__24.13_148 = MEM[(unsigned int *)&a + 112B];
vect__24.14_145 = MEM[(unsigned int *)&a + 128B];
vect__24.15_138 = MEM[(unsigned int *)&a + 144B];
vect__24.16_136 = MEM[(unsigned int *)&a + 160B];
vect__24.17_133 = MEM[(unsigned int *)&a + 176B];
vect__24.18_126 = MEM[(unsigned int *)&a + 192B];
vect__24.19_124 = MEM[(unsigned int *)&a + 208B];
vect__24.20_111 = MEM[(unsigned int *)&a + 224B];
vect__24.21_108 = MEM[(unsigned int *)&a + 240B];
vect__25.23_107 = VEC_PACK_TRUNC_EXPR <vect__24.6_185, vect__24.7_182>;
vect__25.23_101 = VEC_PACK_TRUNC_EXPR <vect__24.8_176, vect__24.9_173>;
vect__25.23_100 = VEC_PACK_TRUNC_EXPR <vect__24.10_160, vect__24.11_158>;
vect__25.23_99 = VEC_PACK_TRUNC_EXPR <vect__24.12_155, vect__24.13_148>;
vect__25.23_87 = VEC_PACK_TRUNC_EXPR <vect__24.14_145, vect__24.15_138>;
vect__25.23_86 = VEC_PACK_TRUNC_EXPR <vect__24.16_136, vect__24.17_133>;
vect__25.23_85 = VEC_PACK_TRUNC_EXPR <vect__24.18_126, vect__24.19_124>;
vect__25.23_81 = VEC_PACK_TRUNC_EXPR <vect__24.20_111, vect__24.21_108>;
vect__25.22_80 = VEC_PACK_TRUNC_EXPR <vect__25.23_107, vect__25.23_101>;
vect__25.22_79 = VEC_PACK_TRUNC_EXPR <vect__25.23_100, vect__25.23_99>;
vect__25.22_78 = VEC_PACK_TRUNC_EXPR <vect__25.23_87, vect__25.23_86>;
vect__25.22_77 = VEC_PACK_TRUNC_EXPR <vect__25.23_85, vect__25.23_81>;
MEM[(unsigned char[256] *)&b] = vect__25.22_80;
MEM[(unsigned char[256] *)&b + 16B] = vect__25.22_79;
MEM[(unsigned char[256] *)&b + 32B] = vect__25.22_78;
MEM[(unsigned char[256] *)&b + 48B] = vect__25.22_77;
As can be seen, the first 64 elements of the destination array are simply
truncated first 64 elements of source array.
