From cf07eea8429c923b7eb884ffc1b267c80a0a839c Mon Sep 17 00:00:00 2001 From: Richard Biener Date: Mon, 22 Jun 2020 12:14:54 +0200 Subject: [PATCH] tree-optimization/95770 - fix SLP vectorized stmt placement compute This fixes the vectorized stmt placement compute for the case of external defs. 2020-06-22 Richard Biener PR tree-optimization/95770 * tree-vect-slp.c (vect_schedule_slp_instance): Also consider external defs. * gcc.dg/pr95770.c: New testcase. --- gcc/testsuite/gcc.dg/pr95770.c | 9 +++++++++ gcc/tree-vect-slp.c | 18 +++++++++++++++--- 2 files changed, 24 insertions(+), 3 deletions(-) create mode 100644 gcc/testsuite/gcc.dg/pr95770.c diff --git a/gcc/testsuite/gcc.dg/pr95770.c b/gcc/testsuite/gcc.dg/pr95770.c new file mode 100644 index 00000000000..06714ea1fa5 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr95770.c @@ -0,0 +1,9 @@ +/* { dg-do compile } */ +/* { dg-options "-O3" } */ + +float *a; +void b(float c, float d) +{ + a[0] = a[1] = 0.5f * (c - 2 + d); + a[2] = a[3] = 0.5f * (c + 2 + d); +} diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c index 5c169f37022..4031db4a9c6 100644 --- a/gcc/tree-vect-slp.c +++ b/gcc/tree-vect-slp.c @@ -4216,9 +4216,6 @@ vect_schedule_slp_instance (vec_info *vinfo, children vectorized defs. */ gimple *last_stmt = NULL; FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child) - /* ??? With only external defs the following breaks. Note - external defs do not get all vector init stmts generated - at the same place. */ if (SLP_TREE_DEF_TYPE (child) == vect_internal_def) { /* We are emitting all vectorized stmts in the same place and @@ -4232,6 +4229,21 @@ vect_schedule_slp_instance (vec_info *vinfo, || vect_stmt_dominates_stmt_p (last_stmt, vstmt)) last_stmt = vstmt; } + else + { + /* For externals we have to look at all defs since their + insertion place is decided per vector. */ + unsigned j; + tree vdef; + FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef) + if (TREE_CODE (vdef) == SSA_NAME) + { + gimple *vstmt = SSA_NAME_DEF_STMT (vdef); + if (!last_stmt + || vect_stmt_dominates_stmt_p (last_stmt, vstmt)) + last_stmt = vstmt; + } + } if (is_a (last_stmt)) si = gsi_after_labels (gimple_bb (last_stmt)); else -- 2.30.2