The following fixes a long-standing issue with SLP vectorization where the dependence checking didn't really reflect reality... (oops).
I have sofar prepared trunk and GCC 8 variants. Bootstrap and regtest running on x86_64-unknown-linux-gnu. Richard. 2018-10-23 Richard Biener <rguent...@suse.de> PR tree-optimization/87665 * tree-vect-data-refs.c (vect_preserves_scalar_order_p): Adjust to reflect reality. * gcc.dg/torture/pr87665.c: New testcase. Index: gcc/tree-vect-data-refs.c =================================================================== --- gcc/tree-vect-data-refs.c (revision 265422) +++ gcc/tree-vect-data-refs.c (working copy) @@ -210,16 +210,26 @@ vect_preserves_scalar_order_p (dr_vec_in return true; /* STMT_A and STMT_B belong to overlapping groups. All loads in a - group are emitted at the position of the first scalar load and all + group are emitted at the position of the last scalar load and all stores in a group are emitted at the position of the last scalar store. - Thus writes will happen no earlier than their current position - (but could happen later) while reads will happen no later than their - current position (but could happen earlier). Reordering is therefore - only possible if the first access is a write. */ - stmtinfo_a = vect_orig_stmt (stmtinfo_a); - stmtinfo_b = vect_orig_stmt (stmtinfo_b); - stmt_vec_info earlier_stmt_info = get_earlier_stmt (stmtinfo_a, stmtinfo_b); - return !DR_IS_WRITE (STMT_VINFO_DATA_REF (earlier_stmt_info)); + Compute that position and check whether the resulting order matches + the current one. */ + stmt_vec_info last_a = DR_GROUP_FIRST_ELEMENT (stmtinfo_a); + if (last_a) + for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (last_a); s; + s = DR_GROUP_NEXT_ELEMENT (s)) + last_a = get_later_stmt (last_a, s); + else + last_a = stmtinfo_a; + stmt_vec_info last_b = DR_GROUP_FIRST_ELEMENT (stmtinfo_b); + if (last_b) + for (stmt_vec_info s = DR_GROUP_NEXT_ELEMENT (last_b); s; + s = DR_GROUP_NEXT_ELEMENT (s)) + last_b = get_later_stmt (last_b, s); + else + last_b = stmtinfo_b; + return ((get_later_stmt (last_a, last_b) == last_a) + == (get_later_stmt (stmtinfo_a, stmtinfo_b) == stmtinfo_a)); } /* A subroutine of vect_analyze_data_ref_dependence. Handle Index: gcc/testsuite/gcc.dg/torture/pr87665.c =================================================================== --- gcc/testsuite/gcc.dg/torture/pr87665.c (nonexistent) +++ gcc/testsuite/gcc.dg/torture/pr87665.c (working copy) @@ -0,0 +1,27 @@ +/* { dg-do run } */ + +struct X { long x; long y; }; + +struct X a[1024], b[1024]; + +void foo () +{ + for (int i = 0; i < 1024; ++i) + { + long tem = a[i].x; + a[i].x = 0; + b[i].x = tem; + b[i].y = a[i].y; + } +} + +int main() +{ + for (int i = 0; i < 1024; ++i) + a[i].x = i; + foo (); + for (int i = 0; i < 1024; ++i) + if (b[i].x != i) + __builtin_abort(); + return 0; +}