On Mon, Jun 13, 2016 at 1:00 PM, Alan Hayward <alan.hayw...@arm.com> wrote:
> vectorizable_live_operation checks that there is exactly one use of each
> operation that has been marked live.
>
> However, it is possible for the operation is used more than once if the
> usage PHIs are identical.
>
> For example in 71416-1.c, _6 is used twice after the loop in bb 9.
>
> <bb 8>:
> # e.6_21 = PHI <e.6_20(7), _9(10)>
> _5 = g_16(D) >= 0;
> _6 = (int) _5;
> e.5_7 = (unsigned char) e.6_21;
> _8 = e.5_7 + 1;
> _9 = (char) _8;
> if (_9 != 0)
> goto <bb 10>;
> else
> goto <bb 9>;
>
> <bb 9>:
> # _12 = PHI <_6(8)>
> # _19 = PHI <_6(8)>
> goto <bb 12>;
Wonder where this inefficiency comes from ...
>
> This patch relaxes the restriction in vectorizable_live_operation, and now
> makes sure there is at least one use of each operation that has been marked
> live.
>
> Tested on x86 and aarch64.
> Ok to commit?
Ok.
Thanks,
Richard.
> gcc/
> PR tree-optimization/71416
> * tree-vect-loop.c (vectorizable_live_operation): Let worklist have
> multiple entries
>
>
> Alan.
>
>
> diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
> index
> 90ade75bcd212b542ad680877a79df717751ff4b..4c8678505df6ec572b69fd7d12ac55cf4
> 619ece6 100644
> --- a/gcc/tree-vect-loop.c
> +++ b/gcc/tree-vect-loop.c
> @@ -6355,7 +6355,7 @@ vectorizable_live_operation (gimple *stmt,
> FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, lhs)
> if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
> worklist.safe_push (use_stmt);
> - gcc_assert (worklist.length () == 1);
> + gcc_assert (worklist.length () >= 1);
>
> bitsize = TYPE_SIZE (TREE_TYPE (vectype));
> vec_bitsize = TYPE_SIZE (vectype);
> @@ -6413,9 +6413,12 @@ vectorizable_live_operation (gimple *stmt,
>
> /* Replace all uses of the USE_STMT in the worklist with the newly
> inserted
> statement. */
> - use_stmt = worklist.pop ();
> - replace_uses_by (gimple_phi_result (use_stmt), new_tree);
> - update_stmt (use_stmt);
> + while (!worklist.is_empty ())
> + {
> + use_stmt = worklist.pop ();
> + replace_uses_by (gimple_phi_result (use_stmt), new_tree);
> + update_stmt (use_stmt);
> + }
>
> return true;
> }
>
>
>
>
>
