public inbox for gcc-bugs@sourceware.org help / color / mirror / Atom feed
From: "rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org> To: gcc-bugs@gcc.gnu.org Subject: [Bug fortran/62283] basic-block vectorization fails Date: Wed, 27 Aug 2014 12:43:00 -0000 [thread overview] Message-ID: <bug-62283-4-fx3eiCoBKV@http.gcc.gnu.org/bugzilla/> (raw) In-Reply-To: <bug-62283-4@http.gcc.gnu.org/bugzilla/> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62283 --- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- Like Index: gcc/tree-vect-slp.c =================================================================== --- gcc/tree-vect-slp.c (revision 214572) +++ gcc/tree-vect-slp.c (working copy) @@ -329,6 +329,13 @@ vect_get_and_check_slp_defs (loop_vec_in } } + /* If this is an internal def but the definition is not vectorizable + treat it as external def for basic-block SLP. */ + if (!loop + && dt == vect_internal_def + && !STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (def_stmt))) + dt = vect_external_def; + if (first) { oprnd_info->first_dt = dt; if one can sort out placing of the initializer (currently we place it at the start of the basic-block which is wrong). Like with Index: gcc/tree-vect-slp.c =================================================================== --- gcc/tree-vect-slp.c (revision 214572) +++ gcc/tree-vect-slp.c (working copy) @@ -2447,6 +2454,8 @@ vect_get_constant_vectors (tree op, slp_ number_of_places_left_in_vector = nunits; elts = XALLOCAVEC (tree, nunits); + gimple_stmt_iterator where; + bool where_p = false; for (j = 0; j < number_of_copies; j++) { for (i = group_size - 1; stmts.iterate (i, &stmt); i--) @@ -2517,6 +2526,20 @@ vect_get_constant_vectors (tree op, slp_ /* Create 'vect_ = {op0,op1,...,opn}'. */ number_of_places_left_in_vector--; + /* For "externals" defined in BB compute an insert location. */ + if (TREE_CODE (op) == SSA_NAME + && (STMT_VINFO_BB_VINFO (stmt_vinfo)->bb + == gimple_bb (SSA_NAME_DEF_STMT (op)))) + { + gimple stmt = NULL; + if (where_p) + stmt = get_earlier_stmt (SSA_NAME_DEF_STMT (op), + gsi_stmt (where)); + if (!stmt || stmt == gsi_stmt (where)) + stmt = SSA_NAME_DEF_STMT (op); + where = gsi_for_stmt (stmt); + where_p = true; + } if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op))) { if (CONSTANT_CLASS_P (op)) @@ -2558,8 +2581,12 @@ vect_get_constant_vectors (tree op, slp_ CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, elts[k]); vec_cst = build_constructor (vector_type, v); } + if (where_p) + gsi_next (&where); voprnds.quick_push (vect_init_vector (stmt, vec_cst, - vector_type, NULL)); + vector_type, + where_p ? &where : NULL)); + where_p = false; if (ctor_seq != NULL) { gimple init_stmt = SSA_NAME_DEF_STMT (voprnds.last ()); a bit too hackish for my taste. Also handling unvectorizable stmts as external should instead turned into sth more general like forcing 'external' handling as soon as the SLP build fails for children (and letting cost handling reject the awful cases).
next prev parent reply other threads:[~2014-08-27 12:43 UTC|newest] Thread overview: 35+ messages / expand[flat|nested] mbox.gz Atom feed top 2014-08-27 11:49 [Bug fortran/62283] New: SSE optimization fails holst at matmech dot com 2014-08-27 11:52 ` [Bug fortran/62283] " holst at matmech dot com 2014-08-27 12:00 ` [Bug fortran/62283] basic-block vectorization fails rguenth at gcc dot gnu.org 2014-08-27 12:43 ` rguenth at gcc dot gnu.org [this message] 2014-08-27 14:26 ` holst at matmech dot com 2014-08-27 14:52 ` holst at matmech dot com 2014-08-28 9:17 ` rguenth at gcc dot gnu.org 2014-08-28 9:42 ` rguenth at gcc dot gnu.org 2014-08-28 13:14 ` rguenth at gcc dot gnu.org 2014-08-28 13:14 ` rguenth at gcc dot gnu.org 2014-09-01 13:33 ` ro at gcc dot gnu.org 2014-09-01 13:35 ` rguenther at suse dot de 2014-09-01 13:37 ` ro at gcc dot gnu.org 2014-09-01 14:11 ` rguenth at gcc dot gnu.org 2014-09-01 14:29 ` ro at CeBiTec dot Uni-Bielefeld.DE 2014-09-01 14:41 ` rguenther at suse dot de 2014-09-02 11:08 ` ro at gcc dot gnu.org 2014-09-02 11:08 ` ro at gcc dot gnu.org 2014-09-02 11:12 ` ro at CeBiTec dot Uni-Bielefeld.DE 2014-09-02 12:11 ` rguenth at gcc dot gnu.org 2014-09-02 12:17 ` rguenth at gcc dot gnu.org 2015-04-28 8:31 ` rguenth at gcc dot gnu.org 2015-04-28 8:32 ` rguenth at gcc dot gnu.org 2015-05-04 14:43 ` ro at gcc dot gnu.org 2015-05-04 14:44 ` ro at gcc dot gnu.org 2015-05-06 6:47 ` rguenth at gcc dot gnu.org 2015-05-06 6:48 ` rguenth at gcc dot gnu.org 2015-05-06 6:48 ` rguenth at gcc dot gnu.org 2015-05-06 11:31 ` ro at gcc dot gnu.org 2015-05-06 12:18 ` rguenther at suse dot de 2015-05-06 12:21 ` rguenth at gcc dot gnu.org 2024-02-28 14:59 ` ro at gcc dot gnu.org 2024-02-28 15:01 ` ro at gcc dot gnu.org 2024-03-15 21:27 ` anlauf at gcc dot gnu.org 2024-03-15 21:41 ` [Bug tree-optimization/62283] " pinskia at gcc dot gnu.org
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=bug-62283-4-fx3eiCoBKV@http.gcc.gnu.org/bugzilla/ \ --to=gcc-bugzilla@gcc.gnu.org \ --cc=gcc-bugs@gcc.gnu.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).