Hi, the following fixes a thinko in propagate_subaccesses_across_link. When we cannot copy over a tree of accesses of RHS to LHS because of some conflict, we must mark the whole LHS access (sub-)tree as potentially written, even when the RHS access describing this level is not, because some of its sub-accesses might be, which is exactly what happens in the PR.
Bootstrapped and tested on x86_64-linux. OK for trunk? Thanks, Martin 2017-10-02 Martin Jambor <mjam...@suse.cz> PR tree-optimization/82363 * tree-sra.c (propagate_subaccesses_across_link): In unrecoverable mismatch, mark lacc written regardless of racc. testsuite/ * gcc.dg/tree-ssa/pr82363.c: New test. --- gcc/testsuite/gcc.dg/tree-ssa/pr82363.c | 50 +++++++++++++++++++++++++++++++++ gcc/tree-sra.c | 2 +- 2 files changed, 51 insertions(+), 1 deletion(-) create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr82363.c diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr82363.c b/gcc/testsuite/gcc.dg/tree-ssa/pr82363.c new file mode 100644 index 00000000000..6652468976b --- /dev/null +++ b/gcc/testsuite/gcc.dg/tree-ssa/pr82363.c @@ -0,0 +1,50 @@ +/* { dg-do run } */ +/* { dg-options "-O" } */ + +struct A +{ + int b; + int c; + int d; +}; + +struct E +{ + int f; + int g:18; + struct A h; +}; + +struct I +{ + int b; + int j; + struct E k; +}; + +int l, *m = &l; + +struct A n; +struct I o; + +void __attribute__ ((noipa)) +test_l (void) +{ + if (l != 1) + __builtin_abort (); +} + +int main () +{ + while (1) + { + struct I q = { 0, 0, {0, 0, {1, 1, 1}}}, p = q, r = p, *s = &q; + if (p.k.h.c) + o = p; + *m = r.k.h.d; + n = (*s).k.h; + break; + } + test_l (); + return 0; +} diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c index f5675edc7f1..bac593951e7 100644 --- a/gcc/tree-sra.c +++ b/gcc/tree-sra.c @@ -2691,7 +2691,7 @@ propagate_subaccesses_across_link (struct access *lacc, struct access *racc) } else { - if (rchild->grp_write && !lacc->grp_write) + if (!lacc->grp_write) { ret = true; subtree_mark_written_and_enqueue (lacc); -- 2.14.1