This adjusts the SLP build to allow a pattern root stmt to be
built from scalars. I've noticed this in PR98211 where we fail
to promote a SLP subtree to a simple splat operation and instead
emit a series of uniform vector operations. The bb-slp-div-1.c
testcase is now vectorized on x86_64 but only the store so I
adjusted it to expect the load to be vectorized.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2020-12-10 Richard Biener <rguent...@suse.de>
* tree-vect-slp.c (vect_get_and_check_slp_defs): Do
not mark the defs to occur in a pattern if it is the
pattern root and record the original stmt defs in that
case.
* gcc.dg/vect/bb-slp-div-1.c: Expect the load to be
vectorized.
---
gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c | 5 ++++-
gcc/tree-vect-slp.c | 15 ++++++++++++---
2 files changed, 16 insertions(+), 4 deletions(-)
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
b/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
index 87ffc9b897b..1eea9233b70 100644
--- a/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
+++ b/gcc/testsuite/gcc.dg/vect/bb-slp-div-1.c
@@ -16,4 +16,7 @@ f (void)
x[7] /= 9;
}
-/* { dg-final { scan-tree-dump "optimized: basic block" "slp2" { xfail *-*-* }
} } */
+/* We can vectorize the store from a CTOR built from scalar division
+ results but ideally we'd like to see vectorizing the load and the
+ division as well. */
+/* { dg-final { scan-tree-dump "transform load" "slp2" { xfail *-*-* } } } */
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index d248ce2c3f7..e93e9c7a2d3 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -544,12 +544,21 @@ vect_get_and_check_slp_defs (vec_info *vinfo, unsigned
char swap,
continue;
}
- if (def_stmt_info && is_pattern_stmt_p (def_stmt_info))
- oprnd_info->any_pattern = true;
-
oprnd_info->def_stmts.quick_push (def_stmt_info);
oprnd_info->ops.quick_push (oprnd);
+ if (def_stmt_info
+ && is_pattern_stmt_p (def_stmt_info))
+ {
+ if (STMT_VINFO_RELATED_STMT (vect_orig_stmt (def_stmt_info))
+ != def_stmt_info)
+ oprnd_info->any_pattern = true;
+ else
+ /* If we promote this to external use the original stmt def. */
+ oprnd_info->ops.last ()
+ = gimple_get_lhs (vect_orig_stmt (def_stmt_info)->stmt);
+ }
+
/* If there's a extern def on a backedge make sure we can
code-generate at the region start.
??? This is another case that could be fixed by adjusting
--
2.26.2
