--- a/hotspot/src/share/vm/opto/c2_globals.hpp Wed Jul 19 09:00:13 2017 +0200
+++ b/hotspot/src/share/vm/opto/c2_globals.hpp Wed Jul 19 08:29:44 2017 -0700
@@ -192,6 +192,9 @@
"of rounds of unroll,optimize,..") \
range(0, max_jint) \
\
+ product(bool, UseSubwordForMaxVector, true, \
+ "Use Subword Analysis to set maximum vector size") \
+ \
develop(intx, UnrollLimitForProfileCheck, 1, \
"Don't use profile_trip_cnt() to restrict unrolling until " \
"unrolling would push the number of unrolled iterations above " \
--- a/hotspot/src/share/vm/opto/loopTransform.cpp Wed Jul 19 09:00:13 2017 +0200
+++ b/hotspot/src/share/vm/opto/loopTransform.cpp Wed Jul 19 08:29:44 2017 -0700
@@ -784,6 +784,9 @@
}
int slp_max_unroll_factor = cl->slp_max_unroll();
+ if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
+ LoopMaxUnroll = slp_max_unroll_factor;
+ }
if (cl->has_passed_slp()) {
if (slp_max_unroll_factor >= future_unroll_ct) return true;
// Normal case: loop too big
@@ -792,7 +795,7 @@
// Check for being too big
if (body_size > (uint)_local_loop_unroll_limit) {
- if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
+ if ((UseSubwordForMaxVector || xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) return true;
// Normal case: loop too big
return false;
}
--- a/hotspot/src/share/vm/opto/superword.cpp Wed Jul 19 09:00:13 2017 +0200
+++ b/hotspot/src/share/vm/opto/superword.cpp Wed Jul 19 08:29:44 2017 -0700
@@ -299,6 +299,7 @@
// Now we try to find the maximum supported consistent vector which the machine
// description can use
bool small_basic_type = false;
+ bool flag_small_bt = false;
for (uint i = 0; i < lpt()->_body.size(); i++) {
if (ignored_loop_nodes[i] != -1) continue;
@@ -334,7 +335,7 @@
if (is_java_primitive(bt) == false) continue;
- int cur_max_vector = Matcher::max_vector_size(bt);
+ int cur_max_vector = Matcher::max_vector_size(bt);
// If a max vector exists which is not larger than _local_loop_unroll_factor
// stop looking, we already have the max vector to map to.
@@ -348,10 +349,36 @@
// Map the maximal common vector
if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
- if (cur_max_vector < max_vector) {
+ if (cur_max_vector < max_vector && !flag_small_bt) {
max_vector = cur_max_vector;
+ } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) {
+ // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types.
+ // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively.
+ // We check for possibility of narrowing by looking through chain operations using subword types.
+ if (is_subword_type(bt)) {
+ uint start, end;
+ VectorNode::vector_operands(n, &start, &end);
+
+ for (uint j = start; j < end; j++) {
+ Node* in = n->in(j);
+ // Don't propagate through a memory
+ if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) {
+ bool same_type = true;
+ for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
+ Node *use = in->fast_out(k);
+ if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
+ same_type = false;
+ break;
+ }
+ }
+ if (same_type) {
+ max_vector = cur_max_vector;
+ flag_small_bt = true;
+ }
+ }
+ }
+ }
}
-
// We only process post loops on predicated targets where we want to
// mask map the loop to a single iteration
if (post_loop_allowed) {
@@ -2368,7 +2395,7 @@
}
}
- if (vlen_in_bytes > max_vlen_in_bytes) {
+ if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
max_vlen = vlen;
max_vlen_in_bytes = vlen_in_bytes;
}