The following tries to improve PR92645 in a minimal invasive way.
Currently as heuristic the BB vectorizer throws away vector stmts
when all of the vector stmts need to be built via a vector CTOR.
That makes sense unless the stmt only needs a single such vector CTOR
which would still mean eliding N scalar stmts.
With this foo() in the testcase generates optimal code using
unpacks (with some help of forwprop after BB SLP).
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
Richard.
2019-11-25 Richard Biener <rguent...@suse.de>
PR tree-optimization/92645
* tree-vect-slp.c (vect_build_slp_tree_2): For unary ops
do not build the operation from scalars if the operand is.
* gcc.target/i386/pr92645.c: New testcase.
Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c (revision 278681)
+++ gcc/tree-vect-slp.c (working copy)
@@ -1410,10 +1411,11 @@ vect_build_slp_tree_2 (vec_info *vinfo,
matches, npermutes,
&this_tree_size, bst_map)) != NULL)
{
- /* If we have all children of child built up from scalars then just
- throw that away and build it up this node from scalars. */
+ /* If we have all children of a non-unary child built up from
+ scalars then just throw that away and build it up this node
+ from scalars. */
if (is_a <bb_vec_info> (vinfo)
- && !SLP_TREE_CHILDREN (child).is_empty ()
+ && SLP_TREE_CHILDREN (child).length () > 1
/* ??? Rejecting patterns this way doesn't work. We'd have to
do extra work to cancel the pattern so the uses see the
scalar version. */
@@ -1549,10 +1551,11 @@ vect_build_slp_tree_2 (vec_info *vinfo,
tem, npermutes,
&this_tree_size, bst_map)) != NULL)
{
- /* If we have all children of child built up from scalars then
- just throw that away and build it up this node from scalars.
*/
+ /* If we have all children of a non-unary child built up from
+ scalars then just throw that away and build it up this node
+ from scalars. */
if (is_a <bb_vec_info> (vinfo)
- && !SLP_TREE_CHILDREN (child).is_empty ()
+ && SLP_TREE_CHILDREN (child).length () > 1
/* ??? Rejecting patterns this way doesn't work. We'd have
to do extra work to cancel the pattern so the uses see the
scalar version. */
Index: gcc/testsuite/gcc.target/i386/pr92645.c
===================================================================
--- gcc/testsuite/gcc.target/i386/pr92645.c (nonexistent)
+++ gcc/testsuite/gcc.target/i386/pr92645.c (working copy)
@@ -0,0 +1,36 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-tree-optimized -msse2 -Wno-psabi" } */
+
+typedef unsigned short v8hi __attribute__((vector_size(16)));
+typedef unsigned int v4si __attribute__((vector_size(16)));
+
+void bar (v4si *dst, v8hi * __restrict src)
+{
+ unsigned int tem[8];
+ tem[0] = (*src)[0];
+ tem[1] = (*src)[1];
+ tem[2] = (*src)[2];
+ tem[3] = (*src)[3];
+ tem[4] = (*src)[4];
+ tem[5] = (*src)[5];
+ tem[6] = (*src)[6];
+ tem[7] = (*src)[7];
+ dst[0] = *(v4si *)tem;
+ dst[1] = *(v4si *)&tem[4];
+}
+void foo (v4si *dst, v8hi src)
+{
+ unsigned int tem[8];
+ tem[0] = src[0];
+ tem[1] = src[1];
+ tem[2] = src[2];
+ tem[3] = src[3];
+ tem[4] = src[4];
+ tem[5] = src[5];
+ tem[6] = src[6];
+ tem[7] = src[7];
+ dst[0] = *(v4si *)tem;
+ dst[1] = *(v4si *)&tem[4];
+}
+
+/* { dg-final { scan-tree-dump-times "vec_unpack_" 4 "optimized" } } */
