duke@435: /* xdono@631: * Copyright 2007-2008 Sun Microsystems, Inc. All Rights Reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: */ duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_superword.cpp.incl" duke@435: duke@435: // duke@435: // S U P E R W O R D T R A N S F O R M duke@435: //============================================================================= duke@435: duke@435: //------------------------------SuperWord--------------------------- duke@435: SuperWord::SuperWord(PhaseIdealLoop* phase) : duke@435: _phase(phase), duke@435: _igvn(phase->_igvn), duke@435: _arena(phase->C->comp_arena()), duke@435: _packset(arena(), 8, 0, NULL), // packs for the current block duke@435: _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb duke@435: _block(arena(), 8, 0, NULL), // nodes in current block duke@435: _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside duke@435: _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads duke@435: _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails duke@435: _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node duke@435: _align_to_ref(NULL), // memory reference to align vectors to duke@435: _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs duke@435: _dg(_arena), // dependence graph duke@435: _visited(arena()), // visited node set duke@435: _post_visited(arena()), // post visited node set duke@435: _n_idx_list(arena(), 8), // scratch list of (node,index) pairs duke@435: _stk(arena(), 8, 0, NULL), // scratch stack of nodes duke@435: _nlist(arena(), 8, 0, NULL), // scratch list of nodes duke@435: _lpt(NULL), // loop tree node duke@435: _lp(NULL), // LoopNode duke@435: _bb(NULL), // basic block duke@435: _iv(NULL) // induction var duke@435: {} duke@435: duke@435: //------------------------------transform_loop--------------------------- duke@435: void SuperWord::transform_loop(IdealLoopTree* lpt) { duke@435: assert(lpt->_head->is_CountedLoop(), "must be"); duke@435: CountedLoopNode *cl = lpt->_head->as_CountedLoop(); duke@435: duke@435: if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops duke@435: duke@435: // Check for no control flow in body (other than exit) duke@435: Node *cl_exit = cl->loopexit(); duke@435: if (cl_exit->in(0) != lpt->_head) return; duke@435: never@540: // Make sure the are no extra control users of the loop backedge never@540: if (cl->back_control()->outcnt() != 1) { never@540: return; never@540: } never@540: duke@435: // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit)))) duke@435: CountedLoopEndNode* pre_end = get_pre_loop_end(cl); duke@435: if (pre_end == NULL) return; duke@435: Node *pre_opaq1 = pre_end->limit(); duke@435: if (pre_opaq1->Opcode() != Op_Opaque1) return; duke@435: duke@435: // Do vectors exist on this architecture? duke@435: if (vector_width_in_bytes() == 0) return; duke@435: duke@435: init(); // initialize data structures duke@435: duke@435: set_lpt(lpt); duke@435: set_lp(cl); duke@435: duke@435: // For now, define one block which is the entire loop body duke@435: set_bb(cl); duke@435: duke@435: assert(_packset.length() == 0, "packset must be empty"); duke@435: SLP_extract(); duke@435: } duke@435: duke@435: //------------------------------SLP_extract--------------------------- duke@435: // Extract the superword level parallelism duke@435: // duke@435: // 1) A reverse post-order of nodes in the block is constructed. By scanning duke@435: // this list from first to last, all definitions are visited before their uses. duke@435: // duke@435: // 2) A point-to-point dependence graph is constructed between memory references. duke@435: // This simplies the upcoming "independence" checker. duke@435: // duke@435: // 3) The maximum depth in the node graph from the beginning of the block duke@435: // to each node is computed. This is used to prune the graph search duke@435: // in the independence checker. duke@435: // duke@435: // 4) For integer types, the necessary bit width is propagated backwards duke@435: // from stores to allow packed operations on byte, char, and short duke@435: // integers. This reverses the promotion to type "int" that javac duke@435: // did for operations like: char c1,c2,c3; c1 = c2 + c3. duke@435: // duke@435: // 5) One of the memory references is picked to be an aligned vector reference. duke@435: // The pre-loop trip count is adjusted to align this reference in the duke@435: // unrolled body. duke@435: // duke@435: // 6) The initial set of pack pairs is seeded with memory references. duke@435: // duke@435: // 7) The set of pack pairs is extended by following use->def and def->use links. duke@435: // duke@435: // 8) The pairs are combined into vector sized packs. duke@435: // duke@435: // 9) Reorder the memory slices to co-locate members of the memory packs. duke@435: // duke@435: // 10) Generate ideal vector nodes for the final set of packs and where necessary, duke@435: // inserting scalar promotion, vector creation from multiple scalars, and duke@435: // extraction of scalar values from vectors. duke@435: // duke@435: void SuperWord::SLP_extract() { duke@435: duke@435: // Ready the block duke@435: duke@435: construct_bb(); duke@435: duke@435: dependence_graph(); duke@435: duke@435: compute_max_depth(); duke@435: duke@435: compute_vector_element_type(); duke@435: duke@435: // Attempt vectorization duke@435: duke@435: find_adjacent_refs(); duke@435: duke@435: extend_packlist(); duke@435: duke@435: combine_packs(); duke@435: duke@435: construct_my_pack_map(); duke@435: duke@435: filter_packs(); duke@435: duke@435: schedule(); duke@435: duke@435: output(); duke@435: } duke@435: duke@435: //------------------------------find_adjacent_refs--------------------------- duke@435: // Find the adjacent memory references and create pack pairs for them. duke@435: // This is the initial set of packs that will then be extended by duke@435: // following use->def and def->use links. The align positions are duke@435: // assigned relative to the reference "align_to_ref" duke@435: void SuperWord::find_adjacent_refs() { duke@435: // Get list of memory operations duke@435: Node_List memops; duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); kvn@464: if (n->is_Mem() && in_bb(n) && kvn@464: is_java_primitive(n->as_Mem()->memory_type())) { duke@435: int align = memory_alignment(n->as_Mem(), 0); duke@435: if (align != bottom_align) { duke@435: memops.push(n); duke@435: } duke@435: } duke@435: } duke@435: if (memops.size() == 0) return; duke@435: duke@435: // Find a memory reference to align to. The pre-loop trip count duke@435: // is modified to align this reference to a vector-aligned address duke@435: find_align_to_ref(memops); duke@435: if (align_to_ref() == NULL) return; duke@435: duke@435: SWPointer align_to_ref_p(align_to_ref(), this); duke@435: int offset = align_to_ref_p.offset_in_bytes(); duke@435: int scale = align_to_ref_p.scale_in_bytes(); duke@435: int vw = vector_width_in_bytes(); duke@435: int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1; duke@435: int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw; duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) never@507: tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d", never@507: offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride()); duke@435: #endif duke@435: duke@435: // Set alignment relative to "align_to_ref" duke@435: for (int i = memops.size() - 1; i >= 0; i--) { duke@435: MemNode* s = memops.at(i)->as_Mem(); duke@435: SWPointer p2(s, this); duke@435: if (p2.comparable(align_to_ref_p)) { duke@435: int align = memory_alignment(s, iv_adjustment); duke@435: set_alignment(s, align); duke@435: } else { duke@435: memops.remove(i); duke@435: } duke@435: } duke@435: duke@435: // Create initial pack pairs of memory operations duke@435: for (uint i = 0; i < memops.size(); i++) { duke@435: Node* s1 = memops.at(i); duke@435: for (uint j = 0; j < memops.size(); j++) { duke@435: Node* s2 = memops.at(j); duke@435: if (s1 != s2 && are_adjacent_refs(s1, s2)) { duke@435: int align = alignment(s1); duke@435: if (stmts_can_pack(s1, s2, align)) { duke@435: Node_List* pair = new Node_List(); duke@435: pair->push(s1); duke@435: pair->push(s2); duke@435: _packset.append(pair); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\nAfter find_adjacent_refs"); duke@435: print_packset(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------find_align_to_ref--------------------------- duke@435: // Find a memory reference to align the loop induction variable to. duke@435: // Looks first at stores then at loads, looking for a memory reference duke@435: // with the largest number of references similar to it. duke@435: void SuperWord::find_align_to_ref(Node_List &memops) { duke@435: GrowableArray cmp_ct(arena(), memops.size(), memops.size(), 0); duke@435: duke@435: // Count number of comparable memory ops duke@435: for (uint i = 0; i < memops.size(); i++) { duke@435: MemNode* s1 = memops.at(i)->as_Mem(); duke@435: SWPointer p1(s1, this); duke@435: // Discard if pre loop can't align this reference duke@435: if (!ref_is_alignable(p1)) { duke@435: *cmp_ct.adr_at(i) = 0; duke@435: continue; duke@435: } duke@435: for (uint j = i+1; j < memops.size(); j++) { duke@435: MemNode* s2 = memops.at(j)->as_Mem(); duke@435: if (isomorphic(s1, s2)) { duke@435: SWPointer p2(s2, this); duke@435: if (p1.comparable(p2)) { duke@435: (*cmp_ct.adr_at(i))++; duke@435: (*cmp_ct.adr_at(j))++; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Find Store (or Load) with the greatest number of "comparable" references duke@435: int max_ct = 0; duke@435: int max_idx = -1; duke@435: int min_size = max_jint; duke@435: int min_iv_offset = max_jint; duke@435: for (uint j = 0; j < memops.size(); j++) { duke@435: MemNode* s = memops.at(j)->as_Mem(); duke@435: if (s->is_Store()) { duke@435: SWPointer p(s, this); duke@435: if (cmp_ct.at(j) > max_ct || duke@435: cmp_ct.at(j) == max_ct && (data_size(s) < min_size || duke@435: data_size(s) == min_size && duke@435: p.offset_in_bytes() < min_iv_offset)) { duke@435: max_ct = cmp_ct.at(j); duke@435: max_idx = j; duke@435: min_size = data_size(s); duke@435: min_iv_offset = p.offset_in_bytes(); duke@435: } duke@435: } duke@435: } duke@435: // If no stores, look at loads duke@435: if (max_ct == 0) { duke@435: for (uint j = 0; j < memops.size(); j++) { duke@435: MemNode* s = memops.at(j)->as_Mem(); duke@435: if (s->is_Load()) { duke@435: SWPointer p(s, this); duke@435: if (cmp_ct.at(j) > max_ct || duke@435: cmp_ct.at(j) == max_ct && (data_size(s) < min_size || duke@435: data_size(s) == min_size && duke@435: p.offset_in_bytes() < min_iv_offset)) { duke@435: max_ct = cmp_ct.at(j); duke@435: max_idx = j; duke@435: min_size = data_size(s); duke@435: min_iv_offset = p.offset_in_bytes(); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: if (max_ct > 0) duke@435: set_align_to_ref(memops.at(max_idx)->as_Mem()); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) { duke@435: tty->print_cr("\nVector memops after find_align_to_refs"); duke@435: for (uint i = 0; i < memops.size(); i++) { duke@435: MemNode* s = memops.at(i)->as_Mem(); duke@435: s->dump(); duke@435: } duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------ref_is_alignable--------------------------- duke@435: // Can the preloop align the reference to position zero in the vector? duke@435: bool SuperWord::ref_is_alignable(SWPointer& p) { duke@435: if (!p.has_iv()) { duke@435: return true; // no induction variable duke@435: } duke@435: CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop()); duke@435: assert(pre_end->stride_is_con(), "pre loop stride is constant"); duke@435: int preloop_stride = pre_end->stride_con(); duke@435: duke@435: int span = preloop_stride * p.scale_in_bytes(); duke@435: duke@435: // Stride one accesses are alignable. duke@435: if (ABS(span) == p.memory_size()) duke@435: return true; duke@435: duke@435: // If initial offset from start of object is computable, duke@435: // compute alignment within the vector. duke@435: int vw = vector_width_in_bytes(); duke@435: if (vw % span == 0) { duke@435: Node* init_nd = pre_end->init_trip(); duke@435: if (init_nd->is_Con() && p.invar() == NULL) { duke@435: int init = init_nd->bottom_type()->is_int()->get_con(); duke@435: duke@435: int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes(); duke@435: assert(init_offset >= 0, "positive offset from object start"); duke@435: duke@435: if (span > 0) { duke@435: return (vw - (init_offset % vw)) % span == 0; duke@435: } else { duke@435: assert(span < 0, "nonzero stride * scale"); duke@435: return (init_offset % vw) % -span == 0; duke@435: } duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //---------------------------dependence_graph--------------------------- duke@435: // Construct dependency graph. duke@435: // Add dependence edges to load/store nodes for memory dependence duke@435: // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x) duke@435: void SuperWord::dependence_graph() { duke@435: // First, assign a dependence node to each memory node duke@435: for (int i = 0; i < _block.length(); i++ ) { duke@435: Node *n = _block.at(i); duke@435: if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) { duke@435: _dg.make_node(n); duke@435: } duke@435: } duke@435: duke@435: // For each memory slice, create the dependences duke@435: for (int i = 0; i < _mem_slice_head.length(); i++) { duke@435: Node* n = _mem_slice_head.at(i); duke@435: Node* n_tail = _mem_slice_tail.at(i); duke@435: duke@435: // Get slice in predecessor order (last is first) duke@435: mem_slice_preds(n_tail, n, _nlist); duke@435: duke@435: // Make the slice dependent on the root duke@435: DepMem* slice = _dg.dep(n); duke@435: _dg.make_edge(_dg.root(), slice); duke@435: duke@435: // Create a sink for the slice duke@435: DepMem* slice_sink = _dg.make_node(NULL); duke@435: _dg.make_edge(slice_sink, _dg.tail()); duke@435: duke@435: // Now visit each pair of memory ops, creating the edges duke@435: for (int j = _nlist.length() - 1; j >= 0 ; j--) { duke@435: Node* s1 = _nlist.at(j); duke@435: duke@435: // If no dependency yet, use slice duke@435: if (_dg.dep(s1)->in_cnt() == 0) { duke@435: _dg.make_edge(slice, s1); duke@435: } duke@435: SWPointer p1(s1->as_Mem(), this); duke@435: bool sink_dependent = true; duke@435: for (int k = j - 1; k >= 0; k--) { duke@435: Node* s2 = _nlist.at(k); duke@435: if (s1->is_Load() && s2->is_Load()) duke@435: continue; duke@435: SWPointer p2(s2->as_Mem(), this); duke@435: duke@435: int cmp = p1.cmp(p2); duke@435: if (SuperWordRTDepCheck && duke@435: p1.base() != p2.base() && p1.valid() && p2.valid()) { duke@435: // Create a runtime check to disambiguate duke@435: OrderedPair pp(p1.base(), p2.base()); duke@435: _disjoint_ptrs.append_if_missing(pp); duke@435: } else if (!SWPointer::not_equal(cmp)) { duke@435: // Possibly same address duke@435: _dg.make_edge(s1, s2); duke@435: sink_dependent = false; duke@435: } duke@435: } duke@435: if (sink_dependent) { duke@435: _dg.make_edge(s1, slice_sink); duke@435: } duke@435: } duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\nDependence graph for slice: %d", n->_idx); duke@435: for (int q = 0; q < _nlist.length(); q++) { duke@435: _dg.print(_nlist.at(q)); duke@435: } duke@435: tty->cr(); duke@435: } duke@435: #endif duke@435: _nlist.clear(); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE"); duke@435: for (int r = 0; r < _disjoint_ptrs.length(); r++) { duke@435: _disjoint_ptrs.at(r).print(); duke@435: tty->cr(); duke@435: } duke@435: tty->cr(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //---------------------------mem_slice_preds--------------------------- duke@435: // Return a memory slice (node list) in predecessor order starting at "start" duke@435: void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray &preds) { duke@435: assert(preds.length() == 0, "start empty"); duke@435: Node* n = start; duke@435: Node* prev = NULL; duke@435: while (true) { duke@435: assert(in_bb(n), "must be in block"); duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { duke@435: Node* out = n->fast_out(i); duke@435: if (out->is_Load()) { duke@435: if (in_bb(out)) { duke@435: preds.push(out); duke@435: } duke@435: } else { duke@435: // FIXME duke@435: if (out->is_MergeMem() && !in_bb(out)) { duke@435: // Either unrolling is causing a memory edge not to disappear, duke@435: // or need to run igvn.optimize() again before SLP duke@435: } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) { duke@435: // Ditto. Not sure what else to check further. duke@435: } else if (out->Opcode() == Op_StoreCM && out->in(4) == n) { duke@435: // StoreCM has an input edge used as a precedence edge. duke@435: // Maybe an issue when oop stores are vectorized. duke@435: } else { duke@435: assert(out == prev || prev == NULL, "no branches off of store slice"); duke@435: } duke@435: } duke@435: } duke@435: if (n == stop) break; duke@435: preds.push(n); duke@435: prev = n; duke@435: n = n->in(MemNode::Memory); duke@435: } duke@435: } duke@435: duke@435: //------------------------------stmts_can_pack--------------------------- twisti@1040: // Can s1 and s2 be in a pack with s1 immediately preceding s2 and duke@435: // s1 aligned at "align" duke@435: bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) { duke@435: if (isomorphic(s1, s2)) { duke@435: if (independent(s1, s2)) { duke@435: if (!exists_at(s1, 0) && !exists_at(s2, 1)) { duke@435: if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) { duke@435: int s1_align = alignment(s1); duke@435: int s2_align = alignment(s2); duke@435: if (s1_align == top_align || s1_align == align) { duke@435: if (s2_align == top_align || s2_align == align + data_size(s1)) { duke@435: return true; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------exists_at--------------------------- duke@435: // Does s exist in a pack at position pos? duke@435: bool SuperWord::exists_at(Node* s, uint pos) { duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: Node_List* p = _packset.at(i); duke@435: if (p->at(pos) == s) { duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------are_adjacent_refs--------------------------- duke@435: // Is s1 immediately before s2 in memory? duke@435: bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) { duke@435: if (!s1->is_Mem() || !s2->is_Mem()) return false; duke@435: if (!in_bb(s1) || !in_bb(s2)) return false; duke@435: // FIXME - co_locate_pack fails on Stores in different mem-slices, so duke@435: // only pack memops that are in the same alias set until that's fixed. duke@435: if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) != duke@435: _phase->C->get_alias_index(s2->as_Mem()->adr_type())) duke@435: return false; duke@435: SWPointer p1(s1->as_Mem(), this); duke@435: SWPointer p2(s2->as_Mem(), this); duke@435: if (p1.base() != p2.base() || !p1.comparable(p2)) return false; duke@435: int diff = p2.offset_in_bytes() - p1.offset_in_bytes(); duke@435: return diff == data_size(s1); duke@435: } duke@435: duke@435: //------------------------------isomorphic--------------------------- duke@435: // Are s1 and s2 similar? duke@435: bool SuperWord::isomorphic(Node* s1, Node* s2) { duke@435: if (s1->Opcode() != s2->Opcode()) return false; duke@435: if (s1->req() != s2->req()) return false; duke@435: if (s1->in(0) != s2->in(0)) return false; duke@435: if (velt_type(s1) != velt_type(s2)) return false; duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------independent--------------------------- duke@435: // Is there no data path from s1 to s2 or s2 to s1? duke@435: bool SuperWord::independent(Node* s1, Node* s2) { duke@435: // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first"); duke@435: int d1 = depth(s1); duke@435: int d2 = depth(s2); duke@435: if (d1 == d2) return s1 != s2; duke@435: Node* deep = d1 > d2 ? s1 : s2; duke@435: Node* shallow = d1 > d2 ? s2 : s1; duke@435: duke@435: visited_clear(); duke@435: duke@435: return independent_path(shallow, deep); duke@435: } duke@435: duke@435: //------------------------------independent_path------------------------------ duke@435: // Helper for independent duke@435: bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) { duke@435: if (dp >= 1000) return false; // stop deep recursion duke@435: visited_set(deep); duke@435: int shal_depth = depth(shallow); duke@435: assert(shal_depth <= depth(deep), "must be"); duke@435: for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) { duke@435: Node* pred = preds.current(); duke@435: if (in_bb(pred) && !visited_test(pred)) { duke@435: if (shallow == pred) { duke@435: return false; duke@435: } duke@435: if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) { duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------set_alignment--------------------------- duke@435: void SuperWord::set_alignment(Node* s1, Node* s2, int align) { duke@435: set_alignment(s1, align); duke@435: set_alignment(s2, align + data_size(s1)); duke@435: } duke@435: duke@435: //------------------------------data_size--------------------------- duke@435: int SuperWord::data_size(Node* s) { duke@435: const Type* t = velt_type(s); duke@435: BasicType bt = t->array_element_basic_type(); kvn@464: int bsize = type2aelembytes(bt); duke@435: assert(bsize != 0, "valid size"); duke@435: return bsize; duke@435: } duke@435: duke@435: //------------------------------extend_packlist--------------------------- duke@435: // Extend packset by following use->def and def->use links from pack members. duke@435: void SuperWord::extend_packlist() { duke@435: bool changed; duke@435: do { duke@435: changed = false; duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: Node_List* p = _packset.at(i); duke@435: changed |= follow_use_defs(p); duke@435: changed |= follow_def_uses(p); duke@435: } duke@435: } while (changed); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\nAfter extend_packlist"); duke@435: print_packset(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------follow_use_defs--------------------------- duke@435: // Extend the packset by visiting operand definitions of nodes in pack p duke@435: bool SuperWord::follow_use_defs(Node_List* p) { duke@435: Node* s1 = p->at(0); duke@435: Node* s2 = p->at(1); duke@435: assert(p->size() == 2, "just checking"); duke@435: assert(s1->req() == s2->req(), "just checking"); duke@435: assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); duke@435: duke@435: if (s1->is_Load()) return false; duke@435: duke@435: int align = alignment(s1); duke@435: bool changed = false; duke@435: int start = s1->is_Store() ? MemNode::ValueIn : 1; duke@435: int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req(); duke@435: for (int j = start; j < end; j++) { duke@435: Node* t1 = s1->in(j); duke@435: Node* t2 = s2->in(j); duke@435: if (!in_bb(t1) || !in_bb(t2)) duke@435: continue; duke@435: if (stmts_can_pack(t1, t2, align)) { duke@435: if (est_savings(t1, t2) >= 0) { duke@435: Node_List* pair = new Node_List(); duke@435: pair->push(t1); duke@435: pair->push(t2); duke@435: _packset.append(pair); duke@435: set_alignment(t1, t2, align); duke@435: changed = true; duke@435: } duke@435: } duke@435: } duke@435: return changed; duke@435: } duke@435: duke@435: //------------------------------follow_def_uses--------------------------- duke@435: // Extend the packset by visiting uses of nodes in pack p duke@435: bool SuperWord::follow_def_uses(Node_List* p) { duke@435: bool changed = false; duke@435: Node* s1 = p->at(0); duke@435: Node* s2 = p->at(1); duke@435: assert(p->size() == 2, "just checking"); duke@435: assert(s1->req() == s2->req(), "just checking"); duke@435: assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); duke@435: duke@435: if (s1->is_Store()) return false; duke@435: duke@435: int align = alignment(s1); duke@435: int savings = -1; duke@435: Node* u1 = NULL; duke@435: Node* u2 = NULL; duke@435: for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { duke@435: Node* t1 = s1->fast_out(i); duke@435: if (!in_bb(t1)) continue; duke@435: for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) { duke@435: Node* t2 = s2->fast_out(j); duke@435: if (!in_bb(t2)) continue; duke@435: if (!opnd_positions_match(s1, t1, s2, t2)) duke@435: continue; duke@435: if (stmts_can_pack(t1, t2, align)) { duke@435: int my_savings = est_savings(t1, t2); duke@435: if (my_savings > savings) { duke@435: savings = my_savings; duke@435: u1 = t1; duke@435: u2 = t2; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: if (savings >= 0) { duke@435: Node_List* pair = new Node_List(); duke@435: pair->push(u1); duke@435: pair->push(u2); duke@435: _packset.append(pair); duke@435: set_alignment(u1, u2, align); duke@435: changed = true; duke@435: } duke@435: return changed; duke@435: } duke@435: duke@435: //---------------------------opnd_positions_match------------------------- duke@435: // Is the use of d1 in u1 at the same operand position as d2 in u2? duke@435: bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) { duke@435: uint ct = u1->req(); duke@435: if (ct != u2->req()) return false; duke@435: uint i1 = 0; duke@435: uint i2 = 0; duke@435: do { duke@435: for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break; duke@435: for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break; duke@435: if (i1 != i2) { duke@435: return false; duke@435: } duke@435: } while (i1 < ct); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------est_savings--------------------------- duke@435: // Estimate the savings from executing s1 and s2 as a pack duke@435: int SuperWord::est_savings(Node* s1, Node* s2) { duke@435: int save = 2 - 1; // 2 operations per instruction in packed form duke@435: duke@435: // inputs duke@435: for (uint i = 1; i < s1->req(); i++) { duke@435: Node* x1 = s1->in(i); duke@435: Node* x2 = s2->in(i); duke@435: if (x1 != x2) { duke@435: if (are_adjacent_refs(x1, x2)) { duke@435: save += adjacent_profit(x1, x2); duke@435: } else if (!in_packset(x1, x2)) { duke@435: save -= pack_cost(2); duke@435: } else { duke@435: save += unpack_cost(2); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // uses of result duke@435: uint ct = 0; duke@435: for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { duke@435: Node* s1_use = s1->fast_out(i); duke@435: for (int j = 0; j < _packset.length(); j++) { duke@435: Node_List* p = _packset.at(j); duke@435: if (p->at(0) == s1_use) { duke@435: for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) { duke@435: Node* s2_use = s2->fast_out(k); duke@435: if (p->at(p->size()-1) == s2_use) { duke@435: ct++; duke@435: if (are_adjacent_refs(s1_use, s2_use)) { duke@435: save += adjacent_profit(s1_use, s2_use); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: if (ct < s1->outcnt()) save += unpack_cost(1); duke@435: if (ct < s2->outcnt()) save += unpack_cost(1); duke@435: duke@435: return save; duke@435: } duke@435: duke@435: //------------------------------costs--------------------------- duke@435: int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; } duke@435: int SuperWord::pack_cost(int ct) { return ct; } duke@435: int SuperWord::unpack_cost(int ct) { return ct; } duke@435: duke@435: //------------------------------combine_packs--------------------------- duke@435: // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last duke@435: void SuperWord::combine_packs() { duke@435: bool changed; duke@435: do { duke@435: changed = false; duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: Node_List* p1 = _packset.at(i); duke@435: if (p1 == NULL) continue; duke@435: for (int j = 0; j < _packset.length(); j++) { duke@435: Node_List* p2 = _packset.at(j); duke@435: if (p2 == NULL) continue; duke@435: if (p1->at(p1->size()-1) == p2->at(0)) { duke@435: for (uint k = 1; k < p2->size(); k++) { duke@435: p1->push(p2->at(k)); duke@435: } duke@435: _packset.at_put(j, NULL); duke@435: changed = true; duke@435: } duke@435: } duke@435: } duke@435: } while (changed); duke@435: duke@435: for (int i = _packset.length() - 1; i >= 0; i--) { duke@435: Node_List* p1 = _packset.at(i); duke@435: if (p1 == NULL) { duke@435: _packset.remove_at(i); duke@435: } duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\nAfter combine_packs"); duke@435: print_packset(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //-----------------------------construct_my_pack_map-------------------------- duke@435: // Construct the map from nodes to packs. Only valid after the duke@435: // point where a node is only in one pack (after combine_packs). duke@435: void SuperWord::construct_my_pack_map() { duke@435: Node_List* rslt = NULL; duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: Node_List* p = _packset.at(i); duke@435: for (uint j = 0; j < p->size(); j++) { duke@435: Node* s = p->at(j); duke@435: assert(my_pack(s) == NULL, "only in one pack"); duke@435: set_my_pack(s, p); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------filter_packs--------------------------- duke@435: // Remove packs that are not implemented or not profitable. duke@435: void SuperWord::filter_packs() { duke@435: duke@435: // Remove packs that are not implemented duke@435: for (int i = _packset.length() - 1; i >= 0; i--) { duke@435: Node_List* pk = _packset.at(i); duke@435: bool impl = implemented(pk); duke@435: if (!impl) { duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) { duke@435: tty->print_cr("Unimplemented"); duke@435: pk->at(0)->dump(); duke@435: } duke@435: #endif duke@435: remove_pack_at(i); duke@435: } duke@435: } duke@435: duke@435: // Remove packs that are not profitable duke@435: bool changed; duke@435: do { duke@435: changed = false; duke@435: for (int i = _packset.length() - 1; i >= 0; i--) { duke@435: Node_List* pk = _packset.at(i); duke@435: bool prof = profitable(pk); duke@435: if (!prof) { duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) { duke@435: tty->print_cr("Unprofitable"); duke@435: pk->at(0)->dump(); duke@435: } duke@435: #endif duke@435: remove_pack_at(i); duke@435: changed = true; duke@435: } duke@435: } duke@435: } while (changed); duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: tty->print_cr("\nAfter filter_packs"); duke@435: print_packset(); duke@435: tty->cr(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------implemented--------------------------- duke@435: // Can code be generated for pack p? duke@435: bool SuperWord::implemented(Node_List* p) { duke@435: Node* p0 = p->at(0); duke@435: int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0)); duke@435: return vopc > 0 && Matcher::has_match_rule(vopc); duke@435: } duke@435: duke@435: //------------------------------profitable--------------------------- duke@435: // For pack p, are all operands and all uses (with in the block) vector? duke@435: bool SuperWord::profitable(Node_List* p) { duke@435: Node* p0 = p->at(0); duke@435: uint start, end; duke@435: vector_opd_range(p0, &start, &end); duke@435: duke@435: // Return false if some input is not vector and inside block duke@435: for (uint i = start; i < end; i++) { duke@435: if (!is_vector_use(p0, i)) { duke@435: // For now, return false if not scalar promotion case (inputs are the same.) twisti@1040: // Later, implement PackNode and allow differing, non-vector inputs duke@435: // (maybe just the ones from outside the block.) duke@435: Node* p0_def = p0->in(i); duke@435: for (uint j = 1; j < p->size(); j++) { duke@435: Node* use = p->at(j); duke@435: Node* def = use->in(i); duke@435: if (p0_def != def) duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: if (!p0->is_Store()) { duke@435: // For now, return false if not all uses are vector. duke@435: // Later, implement ExtractNode and allow non-vector uses (maybe duke@435: // just the ones outside the block.) duke@435: for (uint i = 0; i < p->size(); i++) { duke@435: Node* def = p->at(i); duke@435: for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = def->fast_out(j); duke@435: for (uint k = 0; k < use->req(); k++) { duke@435: Node* n = use->in(k); duke@435: if (def == n) { duke@435: if (!is_vector_use(use, k)) { duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------schedule--------------------------- duke@435: // Adjust the memory graph for the packed operations duke@435: void SuperWord::schedule() { duke@435: duke@435: // Co-locate in the memory graph the members of each memory pack duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: co_locate_pack(_packset.at(i)); duke@435: } duke@435: } duke@435: duke@435: //------------------------------co_locate_pack--------------------------- duke@435: // Within a pack, move stores down to the last executed store, duke@435: // and move loads up to the first executed load. duke@435: void SuperWord::co_locate_pack(Node_List* pk) { duke@435: if (pk->at(0)->is_Store()) { duke@435: // Push Stores down towards last executed pack member duke@435: MemNode* first = executed_first(pk)->as_Mem(); duke@435: MemNode* last = executed_last(pk)->as_Mem(); duke@435: MemNode* insert_pt = last; duke@435: MemNode* current = last->in(MemNode::Memory)->as_Mem(); duke@435: while (true) { duke@435: assert(in_bb(current), "stay in block"); duke@435: Node* my_mem = current->in(MemNode::Memory); duke@435: if (in_pack(current, pk)) { duke@435: // Forward users of my memory state to my input memory state duke@435: _igvn.hash_delete(current); duke@435: _igvn.hash_delete(my_mem); duke@435: for (DUIterator i = current->outs(); current->has_out(i); i++) { duke@435: Node* use = current->out(i); duke@435: if (use->is_Mem()) { duke@435: assert(use->in(MemNode::Memory) == current, "must be"); duke@435: _igvn.hash_delete(use); duke@435: use->set_req(MemNode::Memory, my_mem); duke@435: _igvn._worklist.push(use); duke@435: --i; // deleted this edge; rescan position duke@435: } duke@435: } duke@435: // put current immediately before insert_pt duke@435: current->set_req(MemNode::Memory, insert_pt->in(MemNode::Memory)); duke@435: _igvn.hash_delete(insert_pt); duke@435: insert_pt->set_req(MemNode::Memory, current); duke@435: _igvn._worklist.push(insert_pt); duke@435: _igvn._worklist.push(current); duke@435: insert_pt = current; duke@435: } duke@435: if (current == first) break; duke@435: current = my_mem->as_Mem(); duke@435: } duke@435: } else if (pk->at(0)->is_Load()) { duke@435: // Pull Loads up towards first executed pack member duke@435: LoadNode* first = executed_first(pk)->as_Load(); duke@435: Node* first_mem = first->in(MemNode::Memory); duke@435: _igvn.hash_delete(first_mem); duke@435: // Give each load same memory state as first duke@435: for (uint i = 0; i < pk->size(); i++) { duke@435: LoadNode* ld = pk->at(i)->as_Load(); duke@435: _igvn.hash_delete(ld); duke@435: ld->set_req(MemNode::Memory, first_mem); duke@435: _igvn._worklist.push(ld); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------output--------------------------- duke@435: // Convert packs into vector node operations duke@435: void SuperWord::output() { duke@435: if (_packset.length() == 0) return; duke@435: duke@435: // MUST ENSURE main loop's initial value is properly aligned: duke@435: // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0 duke@435: duke@435: align_initial_loop_index(align_to_ref()); duke@435: duke@435: // Insert extract (unpack) operations for scalar uses duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: insert_extracts(_packset.at(i)); duke@435: } duke@435: duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); duke@435: Node_List* p = my_pack(n); duke@435: if (p && n == executed_last(p)) { duke@435: uint vlen = p->size(); duke@435: Node* vn = NULL; duke@435: Node* low_adr = p->at(0); duke@435: Node* first = executed_first(p); duke@435: if (n->is_Load()) { duke@435: int opc = n->Opcode(); duke@435: Node* ctl = n->in(MemNode::Control); duke@435: Node* mem = first->in(MemNode::Memory); duke@435: Node* adr = low_adr->in(MemNode::Address); duke@435: const TypePtr* atyp = n->adr_type(); duke@435: vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen); duke@435: duke@435: } else if (n->is_Store()) { duke@435: // Promote value to be stored to vector duke@435: VectorNode* val = vector_opd(p, MemNode::ValueIn); duke@435: duke@435: int opc = n->Opcode(); duke@435: Node* ctl = n->in(MemNode::Control); duke@435: Node* mem = first->in(MemNode::Memory); duke@435: Node* adr = low_adr->in(MemNode::Address); duke@435: const TypePtr* atyp = n->adr_type(); duke@435: vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen); duke@435: duke@435: } else if (n->req() == 3) { duke@435: // Promote operands to vector duke@435: Node* in1 = vector_opd(p, 1); duke@435: Node* in2 = vector_opd(p, 2); duke@435: vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n)); duke@435: duke@435: } else { duke@435: ShouldNotReachHere(); duke@435: } duke@435: duke@435: _phase->_igvn.register_new_node_with_optimizer(vn); duke@435: _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0))); duke@435: for (uint j = 0; j < p->size(); j++) { duke@435: Node* pm = p->at(j); duke@435: _igvn.hash_delete(pm); duke@435: _igvn.subsume_node(pm, vn); duke@435: } duke@435: _igvn._worklist.push(vn); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------vector_opd--------------------------- duke@435: // Create a vector operand for the nodes in pack p for operand: in(opd_idx) duke@435: VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) { duke@435: Node* p0 = p->at(0); duke@435: uint vlen = p->size(); duke@435: Node* opd = p0->in(opd_idx); duke@435: duke@435: bool same_opd = true; duke@435: for (uint i = 1; i < vlen; i++) { duke@435: Node* pi = p->at(i); duke@435: Node* in = pi->in(opd_idx); duke@435: if (opd != in) { duke@435: same_opd = false; duke@435: break; duke@435: } duke@435: } duke@435: duke@435: if (same_opd) { duke@435: if (opd->is_Vector()) { duke@435: return (VectorNode*)opd; // input is matching vector duke@435: } duke@435: // Convert scalar input to vector. Use p0's type because it's container duke@435: // maybe smaller than the operand's container. duke@435: const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd); duke@435: const Type* p0_t = velt_type(p0); duke@435: if (p0_t->higher_equal(opd_t)) opd_t = p0_t; duke@435: VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t); duke@435: duke@435: _phase->_igvn.register_new_node_with_optimizer(vn); duke@435: _phase->set_ctrl(vn, _phase->get_ctrl(opd)); duke@435: return vn; duke@435: } duke@435: duke@435: // Insert pack operation duke@435: const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd); duke@435: PackNode* pk = PackNode::make(_phase->C, opd, opd_t); duke@435: duke@435: for (uint i = 1; i < vlen; i++) { duke@435: Node* pi = p->at(i); duke@435: Node* in = pi->in(opd_idx); duke@435: assert(my_pack(in) == NULL, "Should already have been unpacked"); duke@435: assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type"); duke@435: pk->add_opd(in); duke@435: } duke@435: _phase->_igvn.register_new_node_with_optimizer(pk); duke@435: _phase->set_ctrl(pk, _phase->get_ctrl(opd)); duke@435: return pk; duke@435: } duke@435: duke@435: //------------------------------insert_extracts--------------------------- duke@435: // If a use of pack p is not a vector use, then replace the duke@435: // use with an extract operation. duke@435: void SuperWord::insert_extracts(Node_List* p) { duke@435: if (p->at(0)->is_Store()) return; duke@435: assert(_n_idx_list.is_empty(), "empty (node,index) list"); duke@435: duke@435: // Inspect each use of each pack member. For each use that is duke@435: // not a vector use, replace the use with an extract operation. duke@435: duke@435: for (uint i = 0; i < p->size(); i++) { duke@435: Node* def = p->at(i); duke@435: for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = def->fast_out(j); duke@435: for (uint k = 0; k < use->req(); k++) { duke@435: Node* n = use->in(k); duke@435: if (def == n) { duke@435: if (!is_vector_use(use, k)) { duke@435: _n_idx_list.push(use, k); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: while (_n_idx_list.is_nonempty()) { duke@435: Node* use = _n_idx_list.node(); duke@435: int idx = _n_idx_list.index(); duke@435: _n_idx_list.pop(); duke@435: Node* def = use->in(idx); duke@435: duke@435: // Insert extract operation duke@435: _igvn.hash_delete(def); duke@435: _igvn.hash_delete(use); duke@435: int def_pos = alignment(def) / data_size(def); duke@435: const Type* def_t = velt_type(def); duke@435: duke@435: Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t); duke@435: _phase->_igvn.register_new_node_with_optimizer(ex); duke@435: _phase->set_ctrl(ex, _phase->get_ctrl(def)); duke@435: use->set_req(idx, ex); duke@435: _igvn._worklist.push(def); duke@435: _igvn._worklist.push(use); duke@435: duke@435: bb_insert_after(ex, bb_idx(def)); duke@435: set_velt_type(ex, def_t); duke@435: } duke@435: } duke@435: duke@435: //------------------------------is_vector_use--------------------------- duke@435: // Is use->in(u_idx) a vector use? duke@435: bool SuperWord::is_vector_use(Node* use, int u_idx) { duke@435: Node_List* u_pk = my_pack(use); duke@435: if (u_pk == NULL) return false; duke@435: Node* def = use->in(u_idx); duke@435: Node_List* d_pk = my_pack(def); duke@435: if (d_pk == NULL) { duke@435: // check for scalar promotion duke@435: Node* n = u_pk->at(0)->in(u_idx); duke@435: for (uint i = 1; i < u_pk->size(); i++) { duke@435: if (u_pk->at(i)->in(u_idx) != n) return false; duke@435: } duke@435: return true; duke@435: } duke@435: if (u_pk->size() != d_pk->size()) duke@435: return false; duke@435: for (uint i = 0; i < u_pk->size(); i++) { duke@435: Node* ui = u_pk->at(i); duke@435: Node* di = d_pk->at(i); duke@435: if (ui->in(u_idx) != di || alignment(ui) != alignment(di)) duke@435: return false; duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------construct_bb--------------------------- duke@435: // Construct reverse postorder list of block members duke@435: void SuperWord::construct_bb() { duke@435: Node* entry = bb(); duke@435: duke@435: assert(_stk.length() == 0, "stk is empty"); duke@435: assert(_block.length() == 0, "block is empty"); duke@435: assert(_data_entry.length() == 0, "data_entry is empty"); duke@435: assert(_mem_slice_head.length() == 0, "mem_slice_head is empty"); duke@435: assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty"); duke@435: duke@435: // Find non-control nodes with no inputs from within block, duke@435: // create a temporary map from node _idx to bb_idx for use duke@435: // by the visited and post_visited sets, duke@435: // and count number of nodes in block. duke@435: int bb_ct = 0; duke@435: for (uint i = 0; i < lpt()->_body.size(); i++ ) { duke@435: Node *n = lpt()->_body.at(i); duke@435: set_bb_idx(n, i); // Create a temporary map duke@435: if (in_bb(n)) { duke@435: bb_ct++; duke@435: if (!n->is_CFG()) { duke@435: bool found = false; duke@435: for (uint j = 0; j < n->req(); j++) { duke@435: Node* def = n->in(j); duke@435: if (def && in_bb(def)) { duke@435: found = true; duke@435: break; duke@435: } duke@435: } duke@435: if (!found) { duke@435: assert(n != entry, "can't be entry"); duke@435: _data_entry.push(n); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Find memory slices (head and tail) duke@435: for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) { duke@435: Node *n = lp()->fast_out(i); duke@435: if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) { duke@435: Node* n_tail = n->in(LoopNode::LoopBackControl); kvn@688: if (n_tail != n->in(LoopNode::EntryControl)) { kvn@688: _mem_slice_head.push(n); kvn@688: _mem_slice_tail.push(n_tail); kvn@688: } duke@435: } duke@435: } duke@435: duke@435: // Create an RPO list of nodes in block duke@435: duke@435: visited_clear(); duke@435: post_visited_clear(); duke@435: duke@435: // Push all non-control nodes with no inputs from within block, then control entry duke@435: for (int j = 0; j < _data_entry.length(); j++) { duke@435: Node* n = _data_entry.at(j); duke@435: visited_set(n); duke@435: _stk.push(n); duke@435: } duke@435: visited_set(entry); duke@435: _stk.push(entry); duke@435: duke@435: // Do a depth first walk over out edges duke@435: int rpo_idx = bb_ct - 1; duke@435: int size; duke@435: while ((size = _stk.length()) > 0) { duke@435: Node* n = _stk.top(); // Leave node on stack duke@435: if (!visited_test_set(n)) { duke@435: // forward arc in graph duke@435: } else if (!post_visited_test(n)) { duke@435: // cross or back arc duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { duke@435: Node *use = n->fast_out(i); duke@435: if (in_bb(use) && !visited_test(use) && duke@435: // Don't go around backedge duke@435: (!use->is_Phi() || n == entry)) { duke@435: _stk.push(use); duke@435: } duke@435: } duke@435: if (_stk.length() == size) { duke@435: // There were no additional uses, post visit node now duke@435: _stk.pop(); // Remove node from stack duke@435: assert(rpo_idx >= 0, ""); duke@435: _block.at_put_grow(rpo_idx, n); duke@435: rpo_idx--; duke@435: post_visited_set(n); duke@435: assert(rpo_idx >= 0 || _stk.is_empty(), ""); duke@435: } duke@435: } else { duke@435: _stk.pop(); // Remove post-visited node from stack duke@435: } duke@435: } duke@435: duke@435: // Create real map of block indices for nodes duke@435: for (int j = 0; j < _block.length(); j++) { duke@435: Node* n = _block.at(j); duke@435: set_bb_idx(n, j); duke@435: } duke@435: duke@435: initialize_bb(); // Ensure extra info is allocated. duke@435: duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord) { duke@435: print_bb(); duke@435: tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE"); duke@435: for (int m = 0; m < _data_entry.length(); m++) { duke@435: tty->print("%3d ", m); duke@435: _data_entry.at(m)->dump(); duke@435: } duke@435: tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE"); duke@435: for (int m = 0; m < _mem_slice_head.length(); m++) { duke@435: tty->print("%3d ", m); _mem_slice_head.at(m)->dump(); duke@435: tty->print(" "); _mem_slice_tail.at(m)->dump(); duke@435: } duke@435: } duke@435: #endif duke@435: assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found"); duke@435: } duke@435: duke@435: //------------------------------initialize_bb--------------------------- duke@435: // Initialize per node info duke@435: void SuperWord::initialize_bb() { duke@435: Node* last = _block.at(_block.length() - 1); duke@435: grow_node_info(bb_idx(last)); duke@435: } duke@435: duke@435: //------------------------------bb_insert_after--------------------------- duke@435: // Insert n into block after pos duke@435: void SuperWord::bb_insert_after(Node* n, int pos) { duke@435: int n_pos = pos + 1; duke@435: // Make room duke@435: for (int i = _block.length() - 1; i >= n_pos; i--) { duke@435: _block.at_put_grow(i+1, _block.at(i)); duke@435: } duke@435: for (int j = _node_info.length() - 1; j >= n_pos; j--) { duke@435: _node_info.at_put_grow(j+1, _node_info.at(j)); duke@435: } duke@435: // Set value duke@435: _block.at_put_grow(n_pos, n); duke@435: _node_info.at_put_grow(n_pos, SWNodeInfo::initial); duke@435: // Adjust map from node->_idx to _block index duke@435: for (int i = n_pos; i < _block.length(); i++) { duke@435: set_bb_idx(_block.at(i), i); duke@435: } duke@435: } duke@435: duke@435: //------------------------------compute_max_depth--------------------------- duke@435: // Compute max depth for expressions from beginning of block duke@435: // Use to prune search paths during test for independence. duke@435: void SuperWord::compute_max_depth() { duke@435: int ct = 0; duke@435: bool again; duke@435: do { duke@435: again = false; duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); duke@435: if (!n->is_Phi()) { duke@435: int d_orig = depth(n); duke@435: int d_in = 0; duke@435: for (DepPreds preds(n, _dg); !preds.done(); preds.next()) { duke@435: Node* pred = preds.current(); duke@435: if (in_bb(pred)) { duke@435: d_in = MAX2(d_in, depth(pred)); duke@435: } duke@435: } duke@435: if (d_in + 1 != d_orig) { duke@435: set_depth(n, d_in + 1); duke@435: again = true; duke@435: } duke@435: } duke@435: } duke@435: ct++; duke@435: } while (again); duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) duke@435: tty->print_cr("compute_max_depth iterated: %d times", ct); duke@435: #endif duke@435: } duke@435: duke@435: //-------------------------compute_vector_element_type----------------------- duke@435: // Compute necessary vector element type for expressions duke@435: // This propagates backwards a narrower integer type when the duke@435: // upper bits of the value are not needed. duke@435: // Example: char a,b,c; a = b + c; duke@435: // Normally the type of the add is integer, but for packed character duke@435: // operations the type of the add needs to be char. duke@435: void SuperWord::compute_vector_element_type() { duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) duke@435: tty->print_cr("\ncompute_velt_type:"); duke@435: #endif duke@435: duke@435: // Initial type duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); duke@435: const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type()) duke@435: : _igvn.type(n); duke@435: const Type* vt = container_type(t); duke@435: set_velt_type(n, vt); duke@435: } duke@435: duke@435: // Propagate narrowed type backwards through operations duke@435: // that don't depend on higher order bits duke@435: for (int i = _block.length() - 1; i >= 0; i--) { duke@435: Node* n = _block.at(i); duke@435: // Only integer types need be examined duke@435: if (n->bottom_type()->isa_int()) { duke@435: uint start, end; duke@435: vector_opd_range(n, &start, &end); duke@435: const Type* vt = velt_type(n); duke@435: duke@435: for (uint j = start; j < end; j++) { duke@435: Node* in = n->in(j); duke@435: // Don't propagate through a type conversion duke@435: if (n->bottom_type() != in->bottom_type()) duke@435: continue; duke@435: switch(in->Opcode()) { duke@435: case Op_AddI: case Op_AddL: duke@435: case Op_SubI: case Op_SubL: duke@435: case Op_MulI: case Op_MulL: duke@435: case Op_AndI: case Op_AndL: duke@435: case Op_OrI: case Op_OrL: duke@435: case Op_XorI: case Op_XorL: duke@435: case Op_LShiftI: case Op_LShiftL: duke@435: case Op_CMoveI: case Op_CMoveL: duke@435: if (in_bb(in)) { duke@435: bool same_type = true; duke@435: for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) { duke@435: Node *use = in->fast_out(k); duke@435: if (!in_bb(use) || velt_type(use) != vt) { duke@435: same_type = false; duke@435: break; duke@435: } duke@435: } duke@435: if (same_type) { duke@435: set_velt_type(in, vt); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: #ifndef PRODUCT duke@435: if (TraceSuperWord && Verbose) { duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); duke@435: velt_type(n)->dump(); duke@435: tty->print("\t"); duke@435: n->dump(); duke@435: } duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------memory_alignment--------------------------- duke@435: // Alignment within a vector memory reference duke@435: int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) { duke@435: SWPointer p(s, this); duke@435: if (!p.valid()) { duke@435: return bottom_align; duke@435: } duke@435: int offset = p.offset_in_bytes(); duke@435: offset += iv_adjust_in_bytes; duke@435: int off_rem = offset % vector_width_in_bytes(); duke@435: int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes(); duke@435: return off_mod; duke@435: } duke@435: duke@435: //---------------------------container_type--------------------------- duke@435: // Smallest type containing range of values duke@435: const Type* SuperWord::container_type(const Type* t) { kvn@656: const Type* tp = t->make_ptr(); kvn@656: if (tp && tp->isa_aryptr()) { kvn@656: t = tp->is_aryptr()->elem(); duke@435: } duke@435: if (t->basic_type() == T_INT) { duke@435: if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL; duke@435: if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE; duke@435: if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR; duke@435: if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT; duke@435: return TypeInt::INT; duke@435: } duke@435: return t; duke@435: } duke@435: duke@435: //-------------------------vector_opd_range----------------------- duke@435: // (Start, end] half-open range defining which operands are vector duke@435: void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) { duke@435: switch (n->Opcode()) { twisti@993: case Op_LoadB: case Op_LoadUS: duke@435: case Op_LoadI: case Op_LoadL: duke@435: case Op_LoadF: case Op_LoadD: duke@435: case Op_LoadP: duke@435: *start = 0; duke@435: *end = 0; duke@435: return; duke@435: case Op_StoreB: case Op_StoreC: duke@435: case Op_StoreI: case Op_StoreL: duke@435: case Op_StoreF: case Op_StoreD: duke@435: case Op_StoreP: duke@435: *start = MemNode::ValueIn; duke@435: *end = *start + 1; duke@435: return; duke@435: case Op_LShiftI: case Op_LShiftL: duke@435: *start = 1; duke@435: *end = 2; duke@435: return; duke@435: case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD: duke@435: *start = 2; duke@435: *end = n->req(); duke@435: return; duke@435: } duke@435: *start = 1; duke@435: *end = n->req(); // default is all operands duke@435: } duke@435: duke@435: //------------------------------in_packset--------------------------- duke@435: // Are s1 and s2 in a pack pair and ordered as s1,s2? duke@435: bool SuperWord::in_packset(Node* s1, Node* s2) { duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: Node_List* p = _packset.at(i); duke@435: assert(p->size() == 2, "must be"); duke@435: if (p->at(0) == s1 && p->at(p->size()-1) == s2) { duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------in_pack--------------------------- duke@435: // Is s in pack p? duke@435: Node_List* SuperWord::in_pack(Node* s, Node_List* p) { duke@435: for (uint i = 0; i < p->size(); i++) { duke@435: if (p->at(i) == s) { duke@435: return p; duke@435: } duke@435: } duke@435: return NULL; duke@435: } duke@435: duke@435: //------------------------------remove_pack_at--------------------------- duke@435: // Remove the pack at position pos in the packset duke@435: void SuperWord::remove_pack_at(int pos) { duke@435: Node_List* p = _packset.at(pos); duke@435: for (uint i = 0; i < p->size(); i++) { duke@435: Node* s = p->at(i); duke@435: set_my_pack(s, NULL); duke@435: } duke@435: _packset.remove_at(pos); duke@435: } duke@435: duke@435: //------------------------------executed_first--------------------------- duke@435: // Return the node executed first in pack p. Uses the RPO block list duke@435: // to determine order. duke@435: Node* SuperWord::executed_first(Node_List* p) { duke@435: Node* n = p->at(0); duke@435: int n_rpo = bb_idx(n); duke@435: for (uint i = 1; i < p->size(); i++) { duke@435: Node* s = p->at(i); duke@435: int s_rpo = bb_idx(s); duke@435: if (s_rpo < n_rpo) { duke@435: n = s; duke@435: n_rpo = s_rpo; duke@435: } duke@435: } duke@435: return n; duke@435: } duke@435: duke@435: //------------------------------executed_last--------------------------- duke@435: // Return the node executed last in pack p. duke@435: Node* SuperWord::executed_last(Node_List* p) { duke@435: Node* n = p->at(0); duke@435: int n_rpo = bb_idx(n); duke@435: for (uint i = 1; i < p->size(); i++) { duke@435: Node* s = p->at(i); duke@435: int s_rpo = bb_idx(s); duke@435: if (s_rpo > n_rpo) { duke@435: n = s; duke@435: n_rpo = s_rpo; duke@435: } duke@435: } duke@435: return n; duke@435: } duke@435: duke@435: //----------------------------align_initial_loop_index--------------------------- duke@435: // Adjust pre-loop limit so that in main loop, a load/store reference duke@435: // to align_to_ref will be a position zero in the vector. duke@435: // (iv + k) mod vector_align == 0 duke@435: void SuperWord::align_initial_loop_index(MemNode* align_to_ref) { duke@435: CountedLoopNode *main_head = lp()->as_CountedLoop(); duke@435: assert(main_head->is_main_loop(), ""); duke@435: CountedLoopEndNode* pre_end = get_pre_loop_end(main_head); duke@435: assert(pre_end != NULL, ""); duke@435: Node *pre_opaq1 = pre_end->limit(); duke@435: assert(pre_opaq1->Opcode() == Op_Opaque1, ""); duke@435: Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; never@507: Node *lim0 = pre_opaq->in(1); duke@435: duke@435: // Where we put new limit calculations duke@435: Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); duke@435: duke@435: // Ensure the original loop limit is available from the duke@435: // pre-loop Opaque1 node. duke@435: Node *orig_limit = pre_opaq->original_loop_limit(); duke@435: assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, ""); duke@435: duke@435: SWPointer align_to_ref_p(align_to_ref, this); duke@435: never@507: // Given: never@507: // lim0 == original pre loop limit never@507: // V == v_align (power of 2) never@507: // invar == extra invariant piece of the address expression never@507: // e == k [ +/- invar ] duke@435: // never@507: // When reassociating expressions involving '%' the basic rules are: never@507: // (a - b) % k == 0 => a % k == b % k never@507: // and: never@507: // (a + b) % k == 0 => a % k == (k - b) % k never@507: // never@507: // For stride > 0 && scale > 0, never@507: // Derive the new pre-loop limit "lim" such that the two constraints: never@507: // (1) lim = lim0 + N (where N is some positive integer < V) never@507: // (2) (e + lim) % V == 0 never@507: // are true. never@507: // never@507: // Substituting (1) into (2), never@507: // (e + lim0 + N) % V == 0 never@507: // solve for N: never@507: // N = (V - (e + lim0)) % V never@507: // substitute back into (1), so that new limit never@507: // lim = lim0 + (V - (e + lim0)) % V never@507: // never@507: // For stride > 0 && scale < 0 never@507: // Constraints: never@507: // lim = lim0 + N never@507: // (e - lim) % V == 0 never@507: // Solving for lim: never@507: // (e - lim0 - N) % V == 0 never@507: // N = (e - lim0) % V never@507: // lim = lim0 + (e - lim0) % V never@507: // never@507: // For stride < 0 && scale > 0 never@507: // Constraints: never@507: // lim = lim0 - N never@507: // (e + lim) % V == 0 never@507: // Solving for lim: never@507: // (e + lim0 - N) % V == 0 never@507: // N = (e + lim0) % V never@507: // lim = lim0 - (e + lim0) % V never@507: // never@507: // For stride < 0 && scale < 0 never@507: // Constraints: never@507: // lim = lim0 - N never@507: // (e - lim) % V == 0 never@507: // Solving for lim: never@507: // (e - lim0 + N) % V == 0 never@507: // N = (V - (e - lim0)) % V never@507: // lim = lim0 - (V - (e - lim0)) % V duke@435: never@507: int stride = iv_stride(); never@507: int scale = align_to_ref_p.scale_in_bytes(); duke@435: int elt_size = align_to_ref_p.memory_size(); duke@435: int v_align = vector_width_in_bytes() / elt_size; duke@435: int k = align_to_ref_p.offset_in_bytes() / elt_size; duke@435: duke@435: Node *kn = _igvn.intcon(k); never@507: never@507: Node *e = kn; duke@435: if (align_to_ref_p.invar() != NULL) { never@507: // incorporate any extra invariant piece producing k +/- invar >>> log2(elt) duke@435: Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); duke@435: Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt); duke@435: _phase->_igvn.register_new_node_with_optimizer(aref); duke@435: _phase->set_ctrl(aref, pre_ctrl); never@507: if (align_to_ref_p.negate_invar()) { never@507: e = new (_phase->C, 3) SubINode(e, aref); duke@435: } else { never@507: e = new (_phase->C, 3) AddINode(e, aref); duke@435: } never@507: _phase->_igvn.register_new_node_with_optimizer(e); never@507: _phase->set_ctrl(e, pre_ctrl); duke@435: } never@507: never@507: // compute e +/- lim0 never@507: if (scale < 0) { never@507: e = new (_phase->C, 3) SubINode(e, lim0); never@507: } else { never@507: e = new (_phase->C, 3) AddINode(e, lim0); never@507: } never@507: _phase->_igvn.register_new_node_with_optimizer(e); never@507: _phase->set_ctrl(e, pre_ctrl); never@507: never@507: if (stride * scale > 0) { never@507: // compute V - (e +/- lim0) never@507: Node* va = _igvn.intcon(v_align); never@507: e = new (_phase->C, 3) SubINode(va, e); never@507: _phase->_igvn.register_new_node_with_optimizer(e); never@507: _phase->set_ctrl(e, pre_ctrl); never@507: } never@507: // compute N = (exp) % V duke@435: Node* va_msk = _igvn.intcon(v_align - 1); never@507: Node* N = new (_phase->C, 3) AndINode(e, va_msk); never@507: _phase->_igvn.register_new_node_with_optimizer(N); never@507: _phase->set_ctrl(N, pre_ctrl); never@507: never@507: // substitute back into (1), so that new limit never@507: // lim = lim0 + N never@507: Node* lim; never@507: if (stride < 0) { never@507: lim = new (_phase->C, 3) SubINode(lim0, N); duke@435: } else { never@507: lim = new (_phase->C, 3) AddINode(lim0, N); duke@435: } never@507: _phase->_igvn.register_new_node_with_optimizer(lim); never@507: _phase->set_ctrl(lim, pre_ctrl); duke@435: Node* constrained = never@507: (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit) never@507: : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit); duke@435: _phase->_igvn.register_new_node_with_optimizer(constrained); duke@435: _phase->set_ctrl(constrained, pre_ctrl); duke@435: _igvn.hash_delete(pre_opaq); duke@435: pre_opaq->set_req(1, constrained); duke@435: } duke@435: duke@435: //----------------------------get_pre_loop_end--------------------------- duke@435: // Find pre loop end from main loop. Returns null if none. duke@435: CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) { duke@435: Node *ctrl = cl->in(LoopNode::EntryControl); duke@435: if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL; duke@435: Node *iffm = ctrl->in(0); duke@435: if (!iffm->is_If()) return NULL; duke@435: Node *p_f = iffm->in(0); duke@435: if (!p_f->is_IfFalse()) return NULL; duke@435: if (!p_f->in(0)->is_CountedLoopEnd()) return NULL; duke@435: CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); duke@435: if (!pre_end->loopnode()->is_pre_loop()) return NULL; duke@435: return pre_end; duke@435: } duke@435: duke@435: duke@435: //------------------------------init--------------------------- duke@435: void SuperWord::init() { duke@435: _dg.init(); duke@435: _packset.clear(); duke@435: _disjoint_ptrs.clear(); duke@435: _block.clear(); duke@435: _data_entry.clear(); duke@435: _mem_slice_head.clear(); duke@435: _mem_slice_tail.clear(); duke@435: _node_info.clear(); duke@435: _align_to_ref = NULL; duke@435: _lpt = NULL; duke@435: _lp = NULL; duke@435: _bb = NULL; duke@435: _iv = NULL; duke@435: } duke@435: duke@435: //------------------------------print_packset--------------------------- duke@435: void SuperWord::print_packset() { duke@435: #ifndef PRODUCT duke@435: tty->print_cr("packset"); duke@435: for (int i = 0; i < _packset.length(); i++) { duke@435: tty->print_cr("Pack: %d", i); duke@435: Node_List* p = _packset.at(i); duke@435: print_pack(p); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------print_pack--------------------------- duke@435: void SuperWord::print_pack(Node_List* p) { duke@435: for (uint i = 0; i < p->size(); i++) { duke@435: print_stmt(p->at(i)); duke@435: } duke@435: } duke@435: duke@435: //------------------------------print_bb--------------------------- duke@435: void SuperWord::print_bb() { duke@435: #ifndef PRODUCT duke@435: tty->print_cr("\nBlock"); duke@435: for (int i = 0; i < _block.length(); i++) { duke@435: Node* n = _block.at(i); duke@435: tty->print("%d ", i); duke@435: if (n) { duke@435: n->dump(); duke@435: } duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------print_stmt--------------------------- duke@435: void SuperWord::print_stmt(Node* s) { duke@435: #ifndef PRODUCT duke@435: tty->print(" align: %d \t", alignment(s)); duke@435: s->dump(); duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------blank--------------------------- duke@435: char* SuperWord::blank(uint depth) { duke@435: static char blanks[101]; duke@435: assert(depth < 101, "too deep"); duke@435: for (uint i = 0; i < depth; i++) blanks[i] = ' '; duke@435: blanks[depth] = '\0'; duke@435: return blanks; duke@435: } duke@435: duke@435: duke@435: //==============================SWPointer=========================== duke@435: duke@435: //----------------------------SWPointer------------------------ duke@435: SWPointer::SWPointer(MemNode* mem, SuperWord* slp) : duke@435: _mem(mem), _slp(slp), _base(NULL), _adr(NULL), duke@435: _scale(0), _offset(0), _invar(NULL), _negate_invar(false) { duke@435: duke@435: Node* adr = mem->in(MemNode::Address); duke@435: if (!adr->is_AddP()) { duke@435: assert(!valid(), "too complex"); duke@435: return; duke@435: } duke@435: // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant) duke@435: Node* base = adr->in(AddPNode::Base); duke@435: for (int i = 0; i < 3; i++) { duke@435: if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) { duke@435: assert(!valid(), "too complex"); duke@435: return; duke@435: } duke@435: adr = adr->in(AddPNode::Address); duke@435: if (base == adr || !adr->is_AddP()) { duke@435: break; // stop looking at addp's duke@435: } duke@435: } duke@435: _base = base; duke@435: _adr = adr; duke@435: assert(valid(), "Usable"); duke@435: } duke@435: duke@435: // Following is used to create a temporary object during duke@435: // the pattern match of an address expression. duke@435: SWPointer::SWPointer(SWPointer* p) : duke@435: _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL), duke@435: _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {} duke@435: duke@435: //------------------------scaled_iv_plus_offset-------------------- duke@435: // Match: k*iv + offset duke@435: // where: k is a constant that maybe zero, and duke@435: // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional duke@435: bool SWPointer::scaled_iv_plus_offset(Node* n) { duke@435: if (scaled_iv(n)) { duke@435: return true; duke@435: } duke@435: if (offset_plus_k(n)) { duke@435: return true; duke@435: } duke@435: int opc = n->Opcode(); duke@435: if (opc == Op_AddI) { duke@435: if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) { duke@435: return true; duke@435: } duke@435: if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { duke@435: return true; duke@435: } duke@435: } else if (opc == Op_SubI) { duke@435: if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) { duke@435: return true; duke@435: } duke@435: if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { duke@435: _scale *= -1; duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //----------------------------scaled_iv------------------------ duke@435: // Match: k*iv where k is a constant that's not zero duke@435: bool SWPointer::scaled_iv(Node* n) { duke@435: if (_scale != 0) { duke@435: return false; // already found a scale duke@435: } duke@435: if (n == iv()) { duke@435: _scale = 1; duke@435: return true; duke@435: } duke@435: int opc = n->Opcode(); duke@435: if (opc == Op_MulI) { duke@435: if (n->in(1) == iv() && n->in(2)->is_Con()) { duke@435: _scale = n->in(2)->get_int(); duke@435: return true; duke@435: } else if (n->in(2) == iv() && n->in(1)->is_Con()) { duke@435: _scale = n->in(1)->get_int(); duke@435: return true; duke@435: } duke@435: } else if (opc == Op_LShiftI) { duke@435: if (n->in(1) == iv() && n->in(2)->is_Con()) { duke@435: _scale = 1 << n->in(2)->get_int(); duke@435: return true; duke@435: } duke@435: } else if (opc == Op_ConvI2L) { duke@435: if (scaled_iv_plus_offset(n->in(1))) { duke@435: return true; duke@435: } duke@435: } else if (opc == Op_LShiftL) { duke@435: if (!has_iv() && _invar == NULL) { duke@435: // Need to preserve the current _offset value, so duke@435: // create a temporary object for this expression subtree. duke@435: // Hacky, so should re-engineer the address pattern match. duke@435: SWPointer tmp(this); duke@435: if (tmp.scaled_iv_plus_offset(n->in(1))) { duke@435: if (tmp._invar == NULL) { duke@435: int mult = 1 << n->in(2)->get_int(); duke@435: _scale = tmp._scale * mult; duke@435: _offset += tmp._offset * mult; duke@435: return true; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //----------------------------offset_plus_k------------------------ duke@435: // Match: offset is (k [+/- invariant]) duke@435: // where k maybe zero and invariant is optional, but not both. duke@435: bool SWPointer::offset_plus_k(Node* n, bool negate) { duke@435: int opc = n->Opcode(); duke@435: if (opc == Op_ConI) { duke@435: _offset += negate ? -(n->get_int()) : n->get_int(); duke@435: return true; duke@435: } else if (opc == Op_ConL) { duke@435: // Okay if value fits into an int duke@435: const TypeLong* t = n->find_long_type(); duke@435: if (t->higher_equal(TypeLong::INT)) { duke@435: jlong loff = n->get_long(); duke@435: jint off = (jint)loff; duke@435: _offset += negate ? -off : loff; duke@435: return true; duke@435: } duke@435: return false; duke@435: } duke@435: if (_invar != NULL) return false; // already have an invariant duke@435: if (opc == Op_AddI) { duke@435: if (n->in(2)->is_Con() && invariant(n->in(1))) { duke@435: _negate_invar = negate; duke@435: _invar = n->in(1); duke@435: _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); duke@435: return true; duke@435: } else if (n->in(1)->is_Con() && invariant(n->in(2))) { duke@435: _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); duke@435: _negate_invar = negate; duke@435: _invar = n->in(2); duke@435: return true; duke@435: } duke@435: } duke@435: if (opc == Op_SubI) { duke@435: if (n->in(2)->is_Con() && invariant(n->in(1))) { duke@435: _negate_invar = negate; duke@435: _invar = n->in(1); duke@435: _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); duke@435: return true; duke@435: } else if (n->in(1)->is_Con() && invariant(n->in(2))) { duke@435: _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); duke@435: _negate_invar = !negate; duke@435: _invar = n->in(2); duke@435: return true; duke@435: } duke@435: } duke@435: if (invariant(n)) { duke@435: _negate_invar = negate; duke@435: _invar = n; duke@435: return true; duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //----------------------------print------------------------ duke@435: void SWPointer::print() { duke@435: #ifndef PRODUCT duke@435: tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n", duke@435: _base != NULL ? _base->_idx : 0, duke@435: _adr != NULL ? _adr->_idx : 0, duke@435: _scale, _offset, duke@435: _negate_invar?'-':'+', duke@435: _invar != NULL ? _invar->_idx : 0); duke@435: #endif duke@435: } duke@435: duke@435: // ========================= OrderedPair ===================== duke@435: duke@435: const OrderedPair OrderedPair::initial; duke@435: duke@435: // ========================= SWNodeInfo ===================== duke@435: duke@435: const SWNodeInfo SWNodeInfo::initial; duke@435: duke@435: duke@435: // ============================ DepGraph =========================== duke@435: duke@435: //------------------------------make_node--------------------------- duke@435: // Make a new dependence graph node for an ideal node. duke@435: DepMem* DepGraph::make_node(Node* node) { duke@435: DepMem* m = new (_arena) DepMem(node); duke@435: if (node != NULL) { duke@435: assert(_map.at_grow(node->_idx) == NULL, "one init only"); duke@435: _map.at_put_grow(node->_idx, m); duke@435: } duke@435: return m; duke@435: } duke@435: duke@435: //------------------------------make_edge--------------------------- duke@435: // Make a new dependence graph edge from dpred -> dsucc duke@435: DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) { duke@435: DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head()); duke@435: dpred->set_out_head(e); duke@435: dsucc->set_in_head(e); duke@435: return e; duke@435: } duke@435: duke@435: // ========================== DepMem ======================== duke@435: duke@435: //------------------------------in_cnt--------------------------- duke@435: int DepMem::in_cnt() { duke@435: int ct = 0; duke@435: for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++; duke@435: return ct; duke@435: } duke@435: duke@435: //------------------------------out_cnt--------------------------- duke@435: int DepMem::out_cnt() { duke@435: int ct = 0; duke@435: for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++; duke@435: return ct; duke@435: } duke@435: duke@435: //------------------------------print----------------------------- duke@435: void DepMem::print() { duke@435: #ifndef PRODUCT duke@435: tty->print(" DepNode %d (", _node->_idx); duke@435: for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) { duke@435: Node* pred = p->pred()->node(); duke@435: tty->print(" %d", pred != NULL ? pred->_idx : 0); duke@435: } duke@435: tty->print(") ["); duke@435: for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) { duke@435: Node* succ = s->succ()->node(); duke@435: tty->print(" %d", succ != NULL ? succ->_idx : 0); duke@435: } duke@435: tty->print_cr(" ]"); duke@435: #endif duke@435: } duke@435: duke@435: // =========================== DepEdge ========================= duke@435: duke@435: //------------------------------DepPreds--------------------------- duke@435: void DepEdge::print() { duke@435: #ifndef PRODUCT duke@435: tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx); duke@435: #endif duke@435: } duke@435: duke@435: // =========================== DepPreds ========================= duke@435: // Iterator over predecessor edges in the dependence graph. duke@435: duke@435: //------------------------------DepPreds--------------------------- duke@435: DepPreds::DepPreds(Node* n, DepGraph& dg) { duke@435: _n = n; duke@435: _done = false; duke@435: if (_n->is_Store() || _n->is_Load()) { duke@435: _next_idx = MemNode::Address; duke@435: _end_idx = n->req(); duke@435: _dep_next = dg.dep(_n)->in_head(); duke@435: } else if (_n->is_Mem()) { duke@435: _next_idx = 0; duke@435: _end_idx = 0; duke@435: _dep_next = dg.dep(_n)->in_head(); duke@435: } else { duke@435: _next_idx = 1; duke@435: _end_idx = _n->req(); duke@435: _dep_next = NULL; duke@435: } duke@435: next(); duke@435: } duke@435: duke@435: //------------------------------next--------------------------- duke@435: void DepPreds::next() { duke@435: if (_dep_next != NULL) { duke@435: _current = _dep_next->pred()->node(); duke@435: _dep_next = _dep_next->next_in(); duke@435: } else if (_next_idx < _end_idx) { duke@435: _current = _n->in(_next_idx++); duke@435: } else { duke@435: _done = true; duke@435: } duke@435: } duke@435: duke@435: // =========================== DepSuccs ========================= duke@435: // Iterator over successor edges in the dependence graph. duke@435: duke@435: //------------------------------DepSuccs--------------------------- duke@435: DepSuccs::DepSuccs(Node* n, DepGraph& dg) { duke@435: _n = n; duke@435: _done = false; duke@435: if (_n->is_Load()) { duke@435: _next_idx = 0; duke@435: _end_idx = _n->outcnt(); duke@435: _dep_next = dg.dep(_n)->out_head(); duke@435: } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) { duke@435: _next_idx = 0; duke@435: _end_idx = 0; duke@435: _dep_next = dg.dep(_n)->out_head(); duke@435: } else { duke@435: _next_idx = 0; duke@435: _end_idx = _n->outcnt(); duke@435: _dep_next = NULL; duke@435: } duke@435: next(); duke@435: } duke@435: duke@435: //-------------------------------next--------------------------- duke@435: void DepSuccs::next() { duke@435: if (_dep_next != NULL) { duke@435: _current = _dep_next->succ()->node(); duke@435: _dep_next = _dep_next->next_out(); duke@435: } else if (_next_idx < _end_idx) { duke@435: _current = _n->raw_out(_next_idx++); duke@435: } else { duke@435: _done = true; duke@435: } duke@435: }