The following properly handles aggregate returns of the const marked STORE_LANES internal function to update virtual SSA form on-the-fly rather than relying on a costly virtual SSA rewrite.
Bootstrap and regtest running on x86_64-unknown-linux-gnu. PR tree-optimization/106196 * tree-vect-stmts.cc (vect_finish_stmt_generation): Properly handle aggregate returns of calls for VDEF updates. * gcc.dg/torture/pr106196.c: New testcase. --- gcc/testsuite/gcc.dg/torture/pr106196.c | 14 ++++++++++++++ gcc/tree-vect-stmts.cc | 6 ++++-- 2 files changed, 18 insertions(+), 2 deletions(-) create mode 100644 gcc/testsuite/gcc.dg/torture/pr106196.c diff --git a/gcc/testsuite/gcc.dg/torture/pr106196.c b/gcc/testsuite/gcc.dg/torture/pr106196.c new file mode 100644 index 00000000000..56723de42c2 --- /dev/null +++ b/gcc/testsuite/gcc.dg/torture/pr106196.c @@ -0,0 +1,14 @@ +/* { dg-do compile } */ +/* { dg-additional-options "-ftree-vectorize -fno-vect-cost-model" } */ + +extern char a[]; +char *b; +void e() { + char *d; + int c; + d = a; + for (; c; c++) { + d[2] = d[1] = d[0] = b[c]; + d += 3; + } +} diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc index 72107afc883..3db6620dd42 100644 --- a/gcc/tree-vect-stmts.cc +++ b/gcc/tree-vect-stmts.cc @@ -1638,8 +1638,10 @@ vect_finish_stmt_generation (vec_info *vinfo, && ((is_gimple_assign (vec_stmt) && !is_gimple_reg (gimple_assign_lhs (vec_stmt))) || (is_gimple_call (vec_stmt) - && !(gimple_call_flags (vec_stmt) - & (ECF_CONST|ECF_PURE|ECF_NOVOPS))))) + && (!(gimple_call_flags (vec_stmt) + & (ECF_CONST|ECF_PURE|ECF_NOVOPS)) + || (gimple_call_lhs (vec_stmt) + && !is_gimple_reg (gimple_call_lhs (vec_stmt))))))) { tree new_vdef = copy_ssa_name (vuse, vec_stmt); gimple_set_vdef (vec_stmt, new_vdef); -- 2.35.3