1.1 --- a/src/share/vm/opto/superword.cpp Tue Sep 25 10:41:15 2012 -0700
1.2 +++ b/src/share/vm/opto/superword.cpp Tue Sep 25 15:48:17 2012 -0700
1.3 @@ -1088,23 +1088,22 @@
1.4 uint start, end;
1.5 VectorNode::vector_operands(p0, &start, &end);
1.6
1.7 - // Return false if some input is not vector and inside block
1.8 + // Return false if some inputs are not vectors or vectors with different
1.9 + // size or alignment.
1.10 + // Also, for now, return false if not scalar promotion case when inputs are
1.11 + // the same. Later, implement PackNode and allow differing, non-vector inputs
1.12 + // (maybe just the ones from outside the block.)
1.13 for (uint i = start; i < end; i++) {
1.14 - if (!is_vector_use(p0, i)) {
1.15 - // For now, return false if not scalar promotion case (inputs are the same.)
1.16 - // Later, implement PackNode and allow differing, non-vector inputs
1.17 - // (maybe just the ones from outside the block.)
1.18 - if (!same_inputs(p, i)) {
1.19 - return false;
1.20 - }
1.21 - }
1.22 + if (!is_vector_use(p0, i))
1.23 + return false;
1.24 }
1.25 if (VectorNode::is_shift(p0)) {
1.26 - // For now, return false if shift count is vector because
1.27 - // hw does not support it.
1.28 - if (is_vector_use(p0, 2))
1.29 + // For now, return false if shift count is vector or not scalar promotion
1.30 + // case (different shift counts) because it is not supported yet.
1.31 + Node* cnt = p0->in(2);
1.32 + Node_List* cnt_pk = my_pack(cnt);
1.33 + if (cnt_pk != NULL)
1.34 return false;
1.35 - // For the same reason return false if different shift counts.
1.36 if (!same_inputs(p, 2))
1.37 return false;
1.38 }
1.39 @@ -1402,7 +1401,7 @@
1.40 ShouldNotReachHere();
1.41 }
1.42 assert(vn != NULL, "sanity");
1.43 - _phase->_igvn.register_new_node_with_optimizer(vn);
1.44 + _igvn.register_new_node_with_optimizer(vn);
1.45 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1.46 for (uint j = 0; j < p->size(); j++) {
1.47 Node* pm = p->at(j);
1.48 @@ -1451,9 +1450,9 @@
1.49 } else {
1.50 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1.51 cnt = ConNode::make(C, TypeInt::make(mask));
1.52 - _phase->_igvn.register_new_node_with_optimizer(cnt);
1.53 + _igvn.register_new_node_with_optimizer(cnt);
1.54 cnt = new (C, 3) AndINode(opd, cnt);
1.55 - _phase->_igvn.register_new_node_with_optimizer(cnt);
1.56 + _igvn.register_new_node_with_optimizer(cnt);
1.57 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1.58 }
1.59 assert(opd->bottom_type()->isa_int(), "int type only");
1.60 @@ -1461,7 +1460,7 @@
1.61 cnt = new (C, 2) MoveI2FNode(cnt);
1.62 }
1.63 if (cnt != opd) {
1.64 - _phase->_igvn.register_new_node_with_optimizer(cnt);
1.65 + _igvn.register_new_node_with_optimizer(cnt);
1.66 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1.67 }
1.68 return cnt;
1.69 @@ -1473,7 +1472,7 @@
1.70 const Type* p0_t = velt_type(p0);
1.71 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1.72
1.73 - _phase->_igvn.register_new_node_with_optimizer(vn);
1.74 + _igvn.register_new_node_with_optimizer(vn);
1.75 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1.76 #ifdef ASSERT
1.77 if (TraceNewVectors) {
1.78 @@ -1496,7 +1495,7 @@
1.79 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1.80 pk->add_opd(in);
1.81 }
1.82 - _phase->_igvn.register_new_node_with_optimizer(pk);
1.83 + _igvn.register_new_node_with_optimizer(pk);
1.84 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1.85 #ifdef ASSERT
1.86 if (TraceNewVectors) {
1.87 @@ -1543,7 +1542,7 @@
1.88 int def_pos = alignment(def) / data_size(def);
1.89
1.90 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1.91 - _phase->_igvn.register_new_node_with_optimizer(ex);
1.92 + _igvn.register_new_node_with_optimizer(ex);
1.93 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1.94 _igvn.replace_input_of(use, idx, ex);
1.95 _igvn._worklist.push(def);
1.96 @@ -2023,33 +2022,33 @@
1.97 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
1.98 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1.99 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1.100 - _phase->_igvn.register_new_node_with_optimizer(aref);
1.101 + _igvn.register_new_node_with_optimizer(aref);
1.102 _phase->set_ctrl(aref, pre_ctrl);
1.103 if (align_to_ref_p.negate_invar()) {
1.104 e = new (_phase->C, 3) SubINode(e, aref);
1.105 } else {
1.106 e = new (_phase->C, 3) AddINode(e, aref);
1.107 }
1.108 - _phase->_igvn.register_new_node_with_optimizer(e);
1.109 + _igvn.register_new_node_with_optimizer(e);
1.110 _phase->set_ctrl(e, pre_ctrl);
1.111 }
1.112 if (vw > ObjectAlignmentInBytes) {
1.113 // incorporate base e +/- base && Mask >>> log2(elt)
1.114 Node* xbase = new(_phase->C, 2) CastP2XNode(NULL, align_to_ref_p.base());
1.115 - _phase->_igvn.register_new_node_with_optimizer(xbase);
1.116 + _igvn.register_new_node_with_optimizer(xbase);
1.117 #ifdef _LP64
1.118 xbase = new (_phase->C, 2) ConvL2INode(xbase);
1.119 - _phase->_igvn.register_new_node_with_optimizer(xbase);
1.120 + _igvn.register_new_node_with_optimizer(xbase);
1.121 #endif
1.122 Node* mask = _igvn.intcon(vw-1);
1.123 Node* masked_xbase = new (_phase->C, 3) AndINode(xbase, mask);
1.124 - _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
1.125 + _igvn.register_new_node_with_optimizer(masked_xbase);
1.126 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1.127 Node* bref = new (_phase->C, 3) URShiftINode(masked_xbase, log2_elt);
1.128 - _phase->_igvn.register_new_node_with_optimizer(bref);
1.129 + _igvn.register_new_node_with_optimizer(bref);
1.130 _phase->set_ctrl(bref, pre_ctrl);
1.131 e = new (_phase->C, 3) AddINode(e, bref);
1.132 - _phase->_igvn.register_new_node_with_optimizer(e);
1.133 + _igvn.register_new_node_with_optimizer(e);
1.134 _phase->set_ctrl(e, pre_ctrl);
1.135 }
1.136
1.137 @@ -2059,20 +2058,20 @@
1.138 } else {
1.139 e = new (_phase->C, 3) AddINode(e, lim0);
1.140 }
1.141 - _phase->_igvn.register_new_node_with_optimizer(e);
1.142 + _igvn.register_new_node_with_optimizer(e);
1.143 _phase->set_ctrl(e, pre_ctrl);
1.144
1.145 if (stride * scale > 0) {
1.146 // compute V - (e +/- lim0)
1.147 Node* va = _igvn.intcon(v_align);
1.148 e = new (_phase->C, 3) SubINode(va, e);
1.149 - _phase->_igvn.register_new_node_with_optimizer(e);
1.150 + _igvn.register_new_node_with_optimizer(e);
1.151 _phase->set_ctrl(e, pre_ctrl);
1.152 }
1.153 // compute N = (exp) % V
1.154 Node* va_msk = _igvn.intcon(v_align - 1);
1.155 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1.156 - _phase->_igvn.register_new_node_with_optimizer(N);
1.157 + _igvn.register_new_node_with_optimizer(N);
1.158 _phase->set_ctrl(N, pre_ctrl);
1.159
1.160 // substitute back into (1), so that new limit
1.161 @@ -2083,12 +2082,12 @@
1.162 } else {
1.163 lim = new (_phase->C, 3) AddINode(lim0, N);
1.164 }
1.165 - _phase->_igvn.register_new_node_with_optimizer(lim);
1.166 + _igvn.register_new_node_with_optimizer(lim);
1.167 _phase->set_ctrl(lim, pre_ctrl);
1.168 Node* constrained =
1.169 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1.170 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1.171 - _phase->_igvn.register_new_node_with_optimizer(constrained);
1.172 + _igvn.register_new_node_with_optimizer(constrained);
1.173 _phase->set_ctrl(constrained, pre_ctrl);
1.174 _igvn.hash_delete(pre_opaq);
1.175 pre_opaq->set_req(1, constrained);
