diff options
author | Richard Biener <rguenther@suse.de> | 2020-10-06 15:47:15 +0200 |
---|---|---|
committer | Richard Biener <rguenther@suse.de> | 2020-10-08 16:07:15 +0200 |
commit | 181702ef8ab76afbf5d2cd4d7bc0cef613397d6e (patch) | |
tree | 47224d10f36c8f065c9fbe74346cb5f7cd3e838c /gcc/tree-vectorizer.h | |
parent | f997b67550144c6c0562f94c9b9cb932125d0444 (diff) | |
download | gcc-181702ef8ab76afbf5d2cd4d7bc0cef613397d6e.zip gcc-181702ef8ab76afbf5d2cd4d7bc0cef613397d6e.tar.gz gcc-181702ef8ab76afbf5d2cd4d7bc0cef613397d6e.tar.bz2 |
SLP vectorize multiple BBs at once
This work from Martin Liska was motivated by gcc.dg/vect/bb-slp-22.c
which shows how poorly we currently BB vectorize code like
a0 = in[0] + 23;
a1 = in[1] + 142;
a2 = in[2] + 2;
a3 = in[3] + 31;
if (x > y)
{
b[0] = a0;
b[1] = a1;
b[2] = a2;
b[3] = a3;
}
else
{
out[0] = a0 * (x + 1);
out[1] = a1 * (y + 1);
out[2] = a2 * (x + 1);
out[3] = a3 * (y + 1);
}
namely by vectorizing the stores but not the common load (and add)
they are feeded with.
Thus with the following patch we change the BB vectorizer from
operating on a single basic-block at a time to consider somewhat
larger regions (but not the whole function yet because of issues
with vector size iteration).
I took the opportunity to remove the fancy region iterations again
now that we operate on BB granularity and in the end need to visit
PHI nodes as well.
2020-10-08 Martin Liska <mliska@suse.cz>
Richard Biener <rguenther@suse.de>
* tree-vectorizer.h (_bb_vec_info::const_iterator): Remove.
(_bb_vec_info::const_reverse_iterator): Likewise.
(_bb_vec_info::region_stmts): Likewise.
(_bb_vec_info::reverse_region_stmts): Likewise.
(_bb_vec_info::_bb_vec_info): Adjust.
(_bb_vec_info::bb): Remove.
(_bb_vec_info::region_begin): Remove.
(_bb_vec_info::region_end): Remove.
(_bb_vec_info::bbs): New vector of BBs.
(vect_slp_function): Declare.
* tree-vect-patterns.c (vect_determine_precisions): Use
regular stmt iteration.
(vect_pattern_recog): Likewise.
* tree-vect-slp.c: Include cfganal.h, tree-eh.h and tree-cfg.h.
(vect_build_slp_tree_1): Properly refuse to vectorize
volatile and throwing stmts.
(vect_build_slp_tree_2): Pass group-size down to
get_vectype_for_scalar_type.
(_bb_vec_info::_bb_vec_info): Use regular stmt iteration,
adjust for changed region specification.
(_bb_vec_info::~_bb_vec_info): Likewise.
(vect_slp_check_for_constructors): Likewise.
(vect_slp_region): Likewise.
(vect_slp_bbs): New worker operating on a vector of BBs.
(vect_slp_bb): Wrap it.
(vect_slp_function): New function splitting the function
into multi-BB regions.
(vect_create_constant_vectors): Handle the case of inserting
after a throwing def.
(vect_schedule_slp_instance): Adjust.
* tree-vectorizer.c (vec_info::remove_stmt): Simplify again.
(vec_info::insert_seq_on_entry): Adjust.
(pass_slp_vectorize::execute): Also init PHIs. Call
vect_slp_function.
* gcc.dg/vect/bb-slp-22.c: Adjust.
* gfortran.dg/pr68627.f: Likewise.
Diffstat (limited to 'gcc/tree-vectorizer.h')
-rw-r--r-- | gcc/tree-vectorizer.h | 93 |
1 files changed, 7 insertions, 86 deletions
diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h index 37b0915..38daa05 100644 --- a/gcc/tree-vectorizer.h +++ b/gcc/tree-vectorizer.h @@ -827,94 +827,14 @@ loop_vec_info_for_loop (class loop *loop) typedef class _bb_vec_info : public vec_info { public: - - /* GIMPLE statement iterator going from region_begin to region_end. */ - - struct const_iterator - { - const_iterator (gimple_stmt_iterator _gsi) : gsi (_gsi) {} - - const const_iterator & - operator++ () - { - gsi_next (&gsi); return *this; - } - - gimple *operator* () const { return gsi_stmt (gsi); } - - bool - operator== (const const_iterator &other) const - { - return gsi_stmt (gsi) == gsi_stmt (other.gsi); - } - - bool - operator!= (const const_iterator &other) const - { - return !(*this == other); - } - - gimple_stmt_iterator gsi; - }; - - /* GIMPLE statement iterator going from region_end to region_begin. */ - - struct const_reverse_iterator - { - const_reverse_iterator (gimple_stmt_iterator _gsi) : gsi (_gsi) {} - - const const_reverse_iterator & - operator++ () - { - gsi_prev (&gsi); return *this; - } - - gimple *operator* () const { return gsi_stmt (gsi); } - - bool - operator== (const const_reverse_iterator &other) const - { - return gsi_stmt (gsi) == gsi_stmt (other.gsi); - } - - bool - operator!= (const const_reverse_iterator &other) const - { - return !(*this == other); - } - - gimple_stmt_iterator gsi; - }; - - _bb_vec_info (gimple_stmt_iterator, gimple_stmt_iterator, vec_info_shared *); + _bb_vec_info (vec<basic_block> bbs, vec_info_shared *); ~_bb_vec_info (); - /* Returns iterator_range for range-based loop. */ - - iterator_range<const_iterator> - region_stmts () - { - return iterator_range<const_iterator> (region_begin, region_end); - } - - /* Returns iterator_range for range-based loop in a reverse order. */ - - iterator_range<const_reverse_iterator> - reverse_region_stmts () - { - const_reverse_iterator begin = region_end; - if (*begin == NULL) - begin = const_reverse_iterator (gsi_last_bb (gsi_bb (region_end))); - else - ++begin; - - const_reverse_iterator end = region_begin; - return iterator_range<const_reverse_iterator> (begin, ++end); - } - - basic_block bb; - gimple_stmt_iterator region_begin; - gimple_stmt_iterator region_end; + /* The region we are operating on. bbs[0] is the entry, excluding + its PHI nodes. In the future we might want to track an explicit + entry edge to cover bbs[0] PHI nodes and have a region entry + insert location. */ + vec<basic_block> bbs; } *bb_vec_info; #define BB_VINFO_BB(B) (B)->bb @@ -2035,6 +1955,7 @@ extern void vect_get_slp_defs (slp_tree, vec<tree> *); extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *, unsigned n = -1U); extern bool vect_slp_bb (basic_block); +extern bool vect_slp_function (function *); extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree); extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree); extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info); |