| Summary: | Alias can go funny with pointer addition | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Andrew Pinski <pinskia> |
| Component: | tree-optimization | Assignee: | Andrew Pinski <pinskia> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | CC: | gcc-bugs, rguenth |
| Priority: | P3 | Keywords: | alias, missed-optimization |
| Version: | 4.3.0 | ||
| Target Milestone: | 4.3.0 | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2006-11-04 17:21:20 | |
Confirmed. Mine, we always get the correct answer on the pointer plus branch. On the trunk: ANYTHING = &ANYTHING READONLY = &ANYTHING INTEGER = &ANYTHING a = &ANYTHING c_1 = &a b = &ANYTHING c_1 = &b c_6 = c_1 + 32 D.1639_11 = &ANYTHING D.1640_12 = D.1639_11 D.1640_12 = c_6 *D.1640_12 = 2; While on the branch: ANYTHING = &ANYTHING READONLY = &ANYTHING INTEGER = &ANYTHING a = &ANYTHING c_1 = &a b = &ANYTHING c_1 = &b c_6 = c_1 + 32 D.1639_11 = c_6 *D.1639_11 = 2; So there is no extra Constraint for the store to &ANYTHING. Fixed at revision 125755. |
With the following testcase: int a[100], b[100], d[100]; void link_error (void); int g(int i, int j) { int *c; int t; c = i?a:b; c++; d[j] = 1; c[j] = 2; if (d[j] != 1) link_error (); return 0; } ---- May-alias should figure out that c can only pointer to a or b but it gets confused as we get: j.2_11 = (unsigned int) j_7; D.1537_12 = j.2_11 * 4; D.1538_13 = (int *) D.1537_12; D.1539_14 = D.1538_13 + c_6; *D.1539_14 = 2; as may-alias only looks at the left hand side of a PLUS expression, we get the following result: D.1538_13 = &ANYTHING D.1539_14 = D.1538_13 Note 4.1 optimzed this testcase but I think that was by accident and is not actually getting it correct. If we add PTR_PLUS_EXPR we get the correct answer all the time unless someone does some real weird tricks by doing something like: size_t t = (size_t)a; int *c = (int*)(size_t)j; *(c+t) = 2;