Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
1 /*
2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
24 #include "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/divnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/memnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/superword.hpp"
36 #include "opto/vectornode.hpp"
38 //
39 // S U P E R W O R D T R A N S F O R M
40 //=============================================================================
42 //------------------------------SuperWord---------------------------
43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
44 _phase(phase),
45 _igvn(phase->_igvn),
46 _arena(phase->C->comp_arena()),
47 _packset(arena(), 8, 0, NULL), // packs for the current block
48 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
49 _block(arena(), 8, 0, NULL), // nodes in current block
50 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
51 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
52 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
53 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
54 _align_to_ref(NULL), // memory reference to align vectors to
55 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
56 _dg(_arena), // dependence graph
57 _visited(arena()), // visited node set
58 _post_visited(arena()), // post visited node set
59 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
60 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
61 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
62 _lpt(NULL), // loop tree node
63 _lp(NULL), // LoopNode
64 _bb(NULL), // basic block
65 _iv(NULL) // induction var
66 {}
68 //------------------------------transform_loop---------------------------
69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
70 assert(lpt->_head->is_CountedLoop(), "must be");
71 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
73 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
75 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
77 // Check for no control flow in body (other than exit)
78 Node *cl_exit = cl->loopexit();
79 if (cl_exit->in(0) != lpt->_head) return;
81 // Make sure the are no extra control users of the loop backedge
82 if (cl->back_control()->outcnt() != 1) {
83 return;
84 }
86 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
87 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
88 if (pre_end == NULL) return;
89 Node *pre_opaq1 = pre_end->limit();
90 if (pre_opaq1->Opcode() != Op_Opaque1) return;
92 // Do vectors exist on this architecture?
93 if (vector_width_in_bytes() == 0) return;
95 init(); // initialize data structures
97 set_lpt(lpt);
98 set_lp(cl);
100 // For now, define one block which is the entire loop body
101 set_bb(cl);
103 assert(_packset.length() == 0, "packset must be empty");
104 SLP_extract();
105 }
107 //------------------------------SLP_extract---------------------------
108 // Extract the superword level parallelism
109 //
110 // 1) A reverse post-order of nodes in the block is constructed. By scanning
111 // this list from first to last, all definitions are visited before their uses.
112 //
113 // 2) A point-to-point dependence graph is constructed between memory references.
114 // This simplies the upcoming "independence" checker.
115 //
116 // 3) The maximum depth in the node graph from the beginning of the block
117 // to each node is computed. This is used to prune the graph search
118 // in the independence checker.
119 //
120 // 4) For integer types, the necessary bit width is propagated backwards
121 // from stores to allow packed operations on byte, char, and short
122 // integers. This reverses the promotion to type "int" that javac
123 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
124 //
125 // 5) One of the memory references is picked to be an aligned vector reference.
126 // The pre-loop trip count is adjusted to align this reference in the
127 // unrolled body.
128 //
129 // 6) The initial set of pack pairs is seeded with memory references.
130 //
131 // 7) The set of pack pairs is extended by following use->def and def->use links.
132 //
133 // 8) The pairs are combined into vector sized packs.
134 //
135 // 9) Reorder the memory slices to co-locate members of the memory packs.
136 //
137 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
138 // inserting scalar promotion, vector creation from multiple scalars, and
139 // extraction of scalar values from vectors.
140 //
141 void SuperWord::SLP_extract() {
143 // Ready the block
145 construct_bb();
147 dependence_graph();
149 compute_max_depth();
151 compute_vector_element_type();
153 // Attempt vectorization
155 find_adjacent_refs();
157 extend_packlist();
159 combine_packs();
161 construct_my_pack_map();
163 filter_packs();
165 schedule();
167 output();
168 }
170 //------------------------------find_adjacent_refs---------------------------
171 // Find the adjacent memory references and create pack pairs for them.
172 // This is the initial set of packs that will then be extended by
173 // following use->def and def->use links. The align positions are
174 // assigned relative to the reference "align_to_ref"
175 void SuperWord::find_adjacent_refs() {
176 // Get list of memory operations
177 Node_List memops;
178 for (int i = 0; i < _block.length(); i++) {
179 Node* n = _block.at(i);
180 if (n->is_Mem() && in_bb(n) &&
181 is_java_primitive(n->as_Mem()->memory_type())) {
182 int align = memory_alignment(n->as_Mem(), 0);
183 if (align != bottom_align) {
184 memops.push(n);
185 }
186 }
187 }
188 if (memops.size() == 0) return;
190 // Find a memory reference to align to. The pre-loop trip count
191 // is modified to align this reference to a vector-aligned address
192 find_align_to_ref(memops);
193 if (align_to_ref() == NULL) return;
195 SWPointer align_to_ref_p(align_to_ref(), this);
196 int offset = align_to_ref_p.offset_in_bytes();
197 int scale = align_to_ref_p.scale_in_bytes();
198 int vw = vector_width_in_bytes();
199 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
200 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
202 #ifndef PRODUCT
203 if (TraceSuperWord)
204 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d",
205 offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
206 #endif
208 // Set alignment relative to "align_to_ref"
209 for (int i = memops.size() - 1; i >= 0; i--) {
210 MemNode* s = memops.at(i)->as_Mem();
211 SWPointer p2(s, this);
212 if (p2.comparable(align_to_ref_p)) {
213 int align = memory_alignment(s, iv_adjustment);
214 set_alignment(s, align);
215 } else {
216 memops.remove(i);
217 }
218 }
220 // Create initial pack pairs of memory operations
221 for (uint i = 0; i < memops.size(); i++) {
222 Node* s1 = memops.at(i);
223 for (uint j = 0; j < memops.size(); j++) {
224 Node* s2 = memops.at(j);
225 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
226 int align = alignment(s1);
227 if (stmts_can_pack(s1, s2, align)) {
228 Node_List* pair = new Node_List();
229 pair->push(s1);
230 pair->push(s2);
231 _packset.append(pair);
232 }
233 }
234 }
235 }
237 #ifndef PRODUCT
238 if (TraceSuperWord) {
239 tty->print_cr("\nAfter find_adjacent_refs");
240 print_packset();
241 }
242 #endif
243 }
245 //------------------------------find_align_to_ref---------------------------
246 // Find a memory reference to align the loop induction variable to.
247 // Looks first at stores then at loads, looking for a memory reference
248 // with the largest number of references similar to it.
249 void SuperWord::find_align_to_ref(Node_List &memops) {
250 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
252 // Count number of comparable memory ops
253 for (uint i = 0; i < memops.size(); i++) {
254 MemNode* s1 = memops.at(i)->as_Mem();
255 SWPointer p1(s1, this);
256 // Discard if pre loop can't align this reference
257 if (!ref_is_alignable(p1)) {
258 *cmp_ct.adr_at(i) = 0;
259 continue;
260 }
261 for (uint j = i+1; j < memops.size(); j++) {
262 MemNode* s2 = memops.at(j)->as_Mem();
263 if (isomorphic(s1, s2)) {
264 SWPointer p2(s2, this);
265 if (p1.comparable(p2)) {
266 (*cmp_ct.adr_at(i))++;
267 (*cmp_ct.adr_at(j))++;
268 }
269 }
270 }
271 }
273 // Find Store (or Load) with the greatest number of "comparable" references
274 int max_ct = 0;
275 int max_idx = -1;
276 int min_size = max_jint;
277 int min_iv_offset = max_jint;
278 for (uint j = 0; j < memops.size(); j++) {
279 MemNode* s = memops.at(j)->as_Mem();
280 if (s->is_Store()) {
281 SWPointer p(s, this);
282 if (cmp_ct.at(j) > max_ct ||
283 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
284 data_size(s) == min_size &&
285 p.offset_in_bytes() < min_iv_offset)) {
286 max_ct = cmp_ct.at(j);
287 max_idx = j;
288 min_size = data_size(s);
289 min_iv_offset = p.offset_in_bytes();
290 }
291 }
292 }
293 // If no stores, look at loads
294 if (max_ct == 0) {
295 for (uint j = 0; j < memops.size(); j++) {
296 MemNode* s = memops.at(j)->as_Mem();
297 if (s->is_Load()) {
298 SWPointer p(s, this);
299 if (cmp_ct.at(j) > max_ct ||
300 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
301 data_size(s) == min_size &&
302 p.offset_in_bytes() < min_iv_offset)) {
303 max_ct = cmp_ct.at(j);
304 max_idx = j;
305 min_size = data_size(s);
306 min_iv_offset = p.offset_in_bytes();
307 }
308 }
309 }
310 }
312 if (max_ct > 0)
313 set_align_to_ref(memops.at(max_idx)->as_Mem());
315 #ifndef PRODUCT
316 if (TraceSuperWord && Verbose) {
317 tty->print_cr("\nVector memops after find_align_to_refs");
318 for (uint i = 0; i < memops.size(); i++) {
319 MemNode* s = memops.at(i)->as_Mem();
320 s->dump();
321 }
322 }
323 #endif
324 }
326 //------------------------------ref_is_alignable---------------------------
327 // Can the preloop align the reference to position zero in the vector?
328 bool SuperWord::ref_is_alignable(SWPointer& p) {
329 if (!p.has_iv()) {
330 return true; // no induction variable
331 }
332 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
333 assert(pre_end->stride_is_con(), "pre loop stride is constant");
334 int preloop_stride = pre_end->stride_con();
336 int span = preloop_stride * p.scale_in_bytes();
338 // Stride one accesses are alignable.
339 if (ABS(span) == p.memory_size())
340 return true;
342 // If initial offset from start of object is computable,
343 // compute alignment within the vector.
344 int vw = vector_width_in_bytes();
345 if (vw % span == 0) {
346 Node* init_nd = pre_end->init_trip();
347 if (init_nd->is_Con() && p.invar() == NULL) {
348 int init = init_nd->bottom_type()->is_int()->get_con();
350 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
351 assert(init_offset >= 0, "positive offset from object start");
353 if (span > 0) {
354 return (vw - (init_offset % vw)) % span == 0;
355 } else {
356 assert(span < 0, "nonzero stride * scale");
357 return (init_offset % vw) % -span == 0;
358 }
359 }
360 }
361 return false;
362 }
364 //---------------------------dependence_graph---------------------------
365 // Construct dependency graph.
366 // Add dependence edges to load/store nodes for memory dependence
367 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
368 void SuperWord::dependence_graph() {
369 // First, assign a dependence node to each memory node
370 for (int i = 0; i < _block.length(); i++ ) {
371 Node *n = _block.at(i);
372 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
373 _dg.make_node(n);
374 }
375 }
377 // For each memory slice, create the dependences
378 for (int i = 0; i < _mem_slice_head.length(); i++) {
379 Node* n = _mem_slice_head.at(i);
380 Node* n_tail = _mem_slice_tail.at(i);
382 // Get slice in predecessor order (last is first)
383 mem_slice_preds(n_tail, n, _nlist);
385 // Make the slice dependent on the root
386 DepMem* slice = _dg.dep(n);
387 _dg.make_edge(_dg.root(), slice);
389 // Create a sink for the slice
390 DepMem* slice_sink = _dg.make_node(NULL);
391 _dg.make_edge(slice_sink, _dg.tail());
393 // Now visit each pair of memory ops, creating the edges
394 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
395 Node* s1 = _nlist.at(j);
397 // If no dependency yet, use slice
398 if (_dg.dep(s1)->in_cnt() == 0) {
399 _dg.make_edge(slice, s1);
400 }
401 SWPointer p1(s1->as_Mem(), this);
402 bool sink_dependent = true;
403 for (int k = j - 1; k >= 0; k--) {
404 Node* s2 = _nlist.at(k);
405 if (s1->is_Load() && s2->is_Load())
406 continue;
407 SWPointer p2(s2->as_Mem(), this);
409 int cmp = p1.cmp(p2);
410 if (SuperWordRTDepCheck &&
411 p1.base() != p2.base() && p1.valid() && p2.valid()) {
412 // Create a runtime check to disambiguate
413 OrderedPair pp(p1.base(), p2.base());
414 _disjoint_ptrs.append_if_missing(pp);
415 } else if (!SWPointer::not_equal(cmp)) {
416 // Possibly same address
417 _dg.make_edge(s1, s2);
418 sink_dependent = false;
419 }
420 }
421 if (sink_dependent) {
422 _dg.make_edge(s1, slice_sink);
423 }
424 }
425 #ifndef PRODUCT
426 if (TraceSuperWord) {
427 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
428 for (int q = 0; q < _nlist.length(); q++) {
429 _dg.print(_nlist.at(q));
430 }
431 tty->cr();
432 }
433 #endif
434 _nlist.clear();
435 }
437 #ifndef PRODUCT
438 if (TraceSuperWord) {
439 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
440 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
441 _disjoint_ptrs.at(r).print();
442 tty->cr();
443 }
444 tty->cr();
445 }
446 #endif
447 }
449 //---------------------------mem_slice_preds---------------------------
450 // Return a memory slice (node list) in predecessor order starting at "start"
451 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
452 assert(preds.length() == 0, "start empty");
453 Node* n = start;
454 Node* prev = NULL;
455 while (true) {
456 assert(in_bb(n), "must be in block");
457 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
458 Node* out = n->fast_out(i);
459 if (out->is_Load()) {
460 if (in_bb(out)) {
461 preds.push(out);
462 }
463 } else {
464 // FIXME
465 if (out->is_MergeMem() && !in_bb(out)) {
466 // Either unrolling is causing a memory edge not to disappear,
467 // or need to run igvn.optimize() again before SLP
468 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
469 // Ditto. Not sure what else to check further.
470 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
471 // StoreCM has an input edge used as a precedence edge.
472 // Maybe an issue when oop stores are vectorized.
473 } else {
474 assert(out == prev || prev == NULL, "no branches off of store slice");
475 }
476 }
477 }
478 if (n == stop) break;
479 preds.push(n);
480 prev = n;
481 n = n->in(MemNode::Memory);
482 }
483 }
485 //------------------------------stmts_can_pack---------------------------
486 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
487 // s1 aligned at "align"
488 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
490 // Do not use superword for non-primitives
491 if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) ||
492 (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type())))
493 return false;
495 if (isomorphic(s1, s2)) {
496 if (independent(s1, s2)) {
497 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
498 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
499 int s1_align = alignment(s1);
500 int s2_align = alignment(s2);
501 if (s1_align == top_align || s1_align == align) {
502 if (s2_align == top_align || s2_align == align + data_size(s1)) {
503 return true;
504 }
505 }
506 }
507 }
508 }
509 }
510 return false;
511 }
513 //------------------------------exists_at---------------------------
514 // Does s exist in a pack at position pos?
515 bool SuperWord::exists_at(Node* s, uint pos) {
516 for (int i = 0; i < _packset.length(); i++) {
517 Node_List* p = _packset.at(i);
518 if (p->at(pos) == s) {
519 return true;
520 }
521 }
522 return false;
523 }
525 //------------------------------are_adjacent_refs---------------------------
526 // Is s1 immediately before s2 in memory?
527 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
528 if (!s1->is_Mem() || !s2->is_Mem()) return false;
529 if (!in_bb(s1) || !in_bb(s2)) return false;
531 // Do not use superword for non-primitives
532 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
533 !is_java_primitive(s2->as_Mem()->memory_type())) {
534 return false;
535 }
537 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
538 // only pack memops that are in the same alias set until that's fixed.
539 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
540 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
541 return false;
542 SWPointer p1(s1->as_Mem(), this);
543 SWPointer p2(s2->as_Mem(), this);
544 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
545 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
546 return diff == data_size(s1);
547 }
549 //------------------------------isomorphic---------------------------
550 // Are s1 and s2 similar?
551 bool SuperWord::isomorphic(Node* s1, Node* s2) {
552 if (s1->Opcode() != s2->Opcode()) return false;
553 if (s1->req() != s2->req()) return false;
554 if (s1->in(0) != s2->in(0)) return false;
555 if (velt_type(s1) != velt_type(s2)) return false;
556 return true;
557 }
559 //------------------------------independent---------------------------
560 // Is there no data path from s1 to s2 or s2 to s1?
561 bool SuperWord::independent(Node* s1, Node* s2) {
562 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
563 int d1 = depth(s1);
564 int d2 = depth(s2);
565 if (d1 == d2) return s1 != s2;
566 Node* deep = d1 > d2 ? s1 : s2;
567 Node* shallow = d1 > d2 ? s2 : s1;
569 visited_clear();
571 return independent_path(shallow, deep);
572 }
574 //------------------------------independent_path------------------------------
575 // Helper for independent
576 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
577 if (dp >= 1000) return false; // stop deep recursion
578 visited_set(deep);
579 int shal_depth = depth(shallow);
580 assert(shal_depth <= depth(deep), "must be");
581 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
582 Node* pred = preds.current();
583 if (in_bb(pred) && !visited_test(pred)) {
584 if (shallow == pred) {
585 return false;
586 }
587 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
588 return false;
589 }
590 }
591 }
592 return true;
593 }
595 //------------------------------set_alignment---------------------------
596 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
597 set_alignment(s1, align);
598 set_alignment(s2, align + data_size(s1));
599 }
601 //------------------------------data_size---------------------------
602 int SuperWord::data_size(Node* s) {
603 const Type* t = velt_type(s);
604 BasicType bt = t->array_element_basic_type();
605 int bsize = type2aelembytes(bt);
606 assert(bsize != 0, "valid size");
607 return bsize;
608 }
610 //------------------------------extend_packlist---------------------------
611 // Extend packset by following use->def and def->use links from pack members.
612 void SuperWord::extend_packlist() {
613 bool changed;
614 do {
615 changed = false;
616 for (int i = 0; i < _packset.length(); i++) {
617 Node_List* p = _packset.at(i);
618 changed |= follow_use_defs(p);
619 changed |= follow_def_uses(p);
620 }
621 } while (changed);
623 #ifndef PRODUCT
624 if (TraceSuperWord) {
625 tty->print_cr("\nAfter extend_packlist");
626 print_packset();
627 }
628 #endif
629 }
631 //------------------------------follow_use_defs---------------------------
632 // Extend the packset by visiting operand definitions of nodes in pack p
633 bool SuperWord::follow_use_defs(Node_List* p) {
634 Node* s1 = p->at(0);
635 Node* s2 = p->at(1);
636 assert(p->size() == 2, "just checking");
637 assert(s1->req() == s2->req(), "just checking");
638 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
640 if (s1->is_Load()) return false;
642 int align = alignment(s1);
643 bool changed = false;
644 int start = s1->is_Store() ? MemNode::ValueIn : 1;
645 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
646 for (int j = start; j < end; j++) {
647 Node* t1 = s1->in(j);
648 Node* t2 = s2->in(j);
649 if (!in_bb(t1) || !in_bb(t2))
650 continue;
651 if (stmts_can_pack(t1, t2, align)) {
652 if (est_savings(t1, t2) >= 0) {
653 Node_List* pair = new Node_List();
654 pair->push(t1);
655 pair->push(t2);
656 _packset.append(pair);
657 set_alignment(t1, t2, align);
658 changed = true;
659 }
660 }
661 }
662 return changed;
663 }
665 //------------------------------follow_def_uses---------------------------
666 // Extend the packset by visiting uses of nodes in pack p
667 bool SuperWord::follow_def_uses(Node_List* p) {
668 bool changed = false;
669 Node* s1 = p->at(0);
670 Node* s2 = p->at(1);
671 assert(p->size() == 2, "just checking");
672 assert(s1->req() == s2->req(), "just checking");
673 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
675 if (s1->is_Store()) return false;
677 int align = alignment(s1);
678 int savings = -1;
679 Node* u1 = NULL;
680 Node* u2 = NULL;
681 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
682 Node* t1 = s1->fast_out(i);
683 if (!in_bb(t1)) continue;
684 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
685 Node* t2 = s2->fast_out(j);
686 if (!in_bb(t2)) continue;
687 if (!opnd_positions_match(s1, t1, s2, t2))
688 continue;
689 if (stmts_can_pack(t1, t2, align)) {
690 int my_savings = est_savings(t1, t2);
691 if (my_savings > savings) {
692 savings = my_savings;
693 u1 = t1;
694 u2 = t2;
695 }
696 }
697 }
698 }
699 if (savings >= 0) {
700 Node_List* pair = new Node_List();
701 pair->push(u1);
702 pair->push(u2);
703 _packset.append(pair);
704 set_alignment(u1, u2, align);
705 changed = true;
706 }
707 return changed;
708 }
710 //---------------------------opnd_positions_match-------------------------
711 // Is the use of d1 in u1 at the same operand position as d2 in u2?
712 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
713 uint ct = u1->req();
714 if (ct != u2->req()) return false;
715 uint i1 = 0;
716 uint i2 = 0;
717 do {
718 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
719 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
720 if (i1 != i2) {
721 return false;
722 }
723 } while (i1 < ct);
724 return true;
725 }
727 //------------------------------est_savings---------------------------
728 // Estimate the savings from executing s1 and s2 as a pack
729 int SuperWord::est_savings(Node* s1, Node* s2) {
730 int save = 2 - 1; // 2 operations per instruction in packed form
732 // inputs
733 for (uint i = 1; i < s1->req(); i++) {
734 Node* x1 = s1->in(i);
735 Node* x2 = s2->in(i);
736 if (x1 != x2) {
737 if (are_adjacent_refs(x1, x2)) {
738 save += adjacent_profit(x1, x2);
739 } else if (!in_packset(x1, x2)) {
740 save -= pack_cost(2);
741 } else {
742 save += unpack_cost(2);
743 }
744 }
745 }
747 // uses of result
748 uint ct = 0;
749 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
750 Node* s1_use = s1->fast_out(i);
751 for (int j = 0; j < _packset.length(); j++) {
752 Node_List* p = _packset.at(j);
753 if (p->at(0) == s1_use) {
754 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
755 Node* s2_use = s2->fast_out(k);
756 if (p->at(p->size()-1) == s2_use) {
757 ct++;
758 if (are_adjacent_refs(s1_use, s2_use)) {
759 save += adjacent_profit(s1_use, s2_use);
760 }
761 }
762 }
763 }
764 }
765 }
767 if (ct < s1->outcnt()) save += unpack_cost(1);
768 if (ct < s2->outcnt()) save += unpack_cost(1);
770 return save;
771 }
773 //------------------------------costs---------------------------
774 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
775 int SuperWord::pack_cost(int ct) { return ct; }
776 int SuperWord::unpack_cost(int ct) { return ct; }
778 //------------------------------combine_packs---------------------------
779 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
780 void SuperWord::combine_packs() {
781 bool changed;
782 do {
783 changed = false;
784 for (int i = 0; i < _packset.length(); i++) {
785 Node_List* p1 = _packset.at(i);
786 if (p1 == NULL) continue;
787 for (int j = 0; j < _packset.length(); j++) {
788 Node_List* p2 = _packset.at(j);
789 if (p2 == NULL) continue;
790 if (p1->at(p1->size()-1) == p2->at(0)) {
791 for (uint k = 1; k < p2->size(); k++) {
792 p1->push(p2->at(k));
793 }
794 _packset.at_put(j, NULL);
795 changed = true;
796 }
797 }
798 }
799 } while (changed);
801 for (int i = _packset.length() - 1; i >= 0; i--) {
802 Node_List* p1 = _packset.at(i);
803 if (p1 == NULL) {
804 _packset.remove_at(i);
805 }
806 }
808 #ifndef PRODUCT
809 if (TraceSuperWord) {
810 tty->print_cr("\nAfter combine_packs");
811 print_packset();
812 }
813 #endif
814 }
816 //-----------------------------construct_my_pack_map--------------------------
817 // Construct the map from nodes to packs. Only valid after the
818 // point where a node is only in one pack (after combine_packs).
819 void SuperWord::construct_my_pack_map() {
820 Node_List* rslt = NULL;
821 for (int i = 0; i < _packset.length(); i++) {
822 Node_List* p = _packset.at(i);
823 for (uint j = 0; j < p->size(); j++) {
824 Node* s = p->at(j);
825 assert(my_pack(s) == NULL, "only in one pack");
826 set_my_pack(s, p);
827 }
828 }
829 }
831 //------------------------------filter_packs---------------------------
832 // Remove packs that are not implemented or not profitable.
833 void SuperWord::filter_packs() {
835 // Remove packs that are not implemented
836 for (int i = _packset.length() - 1; i >= 0; i--) {
837 Node_List* pk = _packset.at(i);
838 bool impl = implemented(pk);
839 if (!impl) {
840 #ifndef PRODUCT
841 if (TraceSuperWord && Verbose) {
842 tty->print_cr("Unimplemented");
843 pk->at(0)->dump();
844 }
845 #endif
846 remove_pack_at(i);
847 }
848 }
850 // Remove packs that are not profitable
851 bool changed;
852 do {
853 changed = false;
854 for (int i = _packset.length() - 1; i >= 0; i--) {
855 Node_List* pk = _packset.at(i);
856 bool prof = profitable(pk);
857 if (!prof) {
858 #ifndef PRODUCT
859 if (TraceSuperWord && Verbose) {
860 tty->print_cr("Unprofitable");
861 pk->at(0)->dump();
862 }
863 #endif
864 remove_pack_at(i);
865 changed = true;
866 }
867 }
868 } while (changed);
870 #ifndef PRODUCT
871 if (TraceSuperWord) {
872 tty->print_cr("\nAfter filter_packs");
873 print_packset();
874 tty->cr();
875 }
876 #endif
877 }
879 //------------------------------implemented---------------------------
880 // Can code be generated for pack p?
881 bool SuperWord::implemented(Node_List* p) {
882 Node* p0 = p->at(0);
883 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
884 return vopc > 0 && Matcher::has_match_rule(vopc);
885 }
887 //------------------------------profitable---------------------------
888 // For pack p, are all operands and all uses (with in the block) vector?
889 bool SuperWord::profitable(Node_List* p) {
890 Node* p0 = p->at(0);
891 uint start, end;
892 vector_opd_range(p0, &start, &end);
894 // Return false if some input is not vector and inside block
895 for (uint i = start; i < end; i++) {
896 if (!is_vector_use(p0, i)) {
897 // For now, return false if not scalar promotion case (inputs are the same.)
898 // Later, implement PackNode and allow differing, non-vector inputs
899 // (maybe just the ones from outside the block.)
900 Node* p0_def = p0->in(i);
901 for (uint j = 1; j < p->size(); j++) {
902 Node* use = p->at(j);
903 Node* def = use->in(i);
904 if (p0_def != def)
905 return false;
906 }
907 }
908 }
909 if (!p0->is_Store()) {
910 // For now, return false if not all uses are vector.
911 // Later, implement ExtractNode and allow non-vector uses (maybe
912 // just the ones outside the block.)
913 for (uint i = 0; i < p->size(); i++) {
914 Node* def = p->at(i);
915 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
916 Node* use = def->fast_out(j);
917 for (uint k = 0; k < use->req(); k++) {
918 Node* n = use->in(k);
919 if (def == n) {
920 if (!is_vector_use(use, k)) {
921 return false;
922 }
923 }
924 }
925 }
926 }
927 }
928 return true;
929 }
931 //------------------------------schedule---------------------------
932 // Adjust the memory graph for the packed operations
933 void SuperWord::schedule() {
935 // Co-locate in the memory graph the members of each memory pack
936 for (int i = 0; i < _packset.length(); i++) {
937 co_locate_pack(_packset.at(i));
938 }
939 }
941 //-------------------------------remove_and_insert-------------------
942 //remove "current" from its current position in the memory graph and insert
943 //it after the appropriate insertion point (lip or uip)
944 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
945 Node *uip, Unique_Node_List &sched_before) {
946 Node* my_mem = current->in(MemNode::Memory);
947 _igvn.hash_delete(current);
948 _igvn.hash_delete(my_mem);
950 //remove current_store from its current position in the memmory graph
951 for (DUIterator i = current->outs(); current->has_out(i); i++) {
952 Node* use = current->out(i);
953 if (use->is_Mem()) {
954 assert(use->in(MemNode::Memory) == current, "must be");
955 _igvn.hash_delete(use);
956 if (use == prev) { // connect prev to my_mem
957 use->set_req(MemNode::Memory, my_mem);
958 } else if (sched_before.member(use)) {
959 _igvn.hash_delete(uip);
960 use->set_req(MemNode::Memory, uip);
961 } else {
962 _igvn.hash_delete(lip);
963 use->set_req(MemNode::Memory, lip);
964 }
965 _igvn._worklist.push(use);
966 --i; //deleted this edge; rescan position
967 }
968 }
970 bool sched_up = sched_before.member(current);
971 Node *insert_pt = sched_up ? uip : lip;
972 _igvn.hash_delete(insert_pt);
974 // all uses of insert_pt's memory state should use current's instead
975 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
976 Node* use = insert_pt->out(i);
977 if (use->is_Mem()) {
978 assert(use->in(MemNode::Memory) == insert_pt, "must be");
979 _igvn.hash_delete(use);
980 use->set_req(MemNode::Memory, current);
981 _igvn._worklist.push(use);
982 --i; //deleted this edge; rescan position
983 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
984 uint pos; //lip (lower insert point) must be the last one in the memory slice
985 _igvn.hash_delete(use);
986 for (pos=1; pos < use->req(); pos++) {
987 if (use->in(pos) == insert_pt) break;
988 }
989 use->set_req(pos, current);
990 _igvn._worklist.push(use);
991 --i;
992 }
993 }
995 //connect current to insert_pt
996 current->set_req(MemNode::Memory, insert_pt);
997 _igvn._worklist.push(current);
998 }
1000 //------------------------------co_locate_pack----------------------------------
1001 // To schedule a store pack, we need to move any sandwiched memory ops either before
1002 // or after the pack, based upon dependence information:
1003 // (1) If any store in the pack depends on the sandwiched memory op, the
1004 // sandwiched memory op must be scheduled BEFORE the pack;
1005 // (2) If a sandwiched memory op depends on any store in the pack, the
1006 // sandwiched memory op must be scheduled AFTER the pack;
1007 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1008 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1009 // scheduled before the pack, memB must also be scheduled before the pack;
1010 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1011 // schedule this store AFTER the pack
1012 // (5) We know there is no dependence cycle, so there in no other case;
1013 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1014 //
1015 // To schedule a load pack, we use the memory state of either the first or the last load in
1016 // the pack, based on the dependence constraint.
1017 void SuperWord::co_locate_pack(Node_List* pk) {
1018 if (pk->at(0)->is_Store()) {
1019 MemNode* first = executed_first(pk)->as_Mem();
1020 MemNode* last = executed_last(pk)->as_Mem();
1021 Unique_Node_List schedule_before_pack;
1022 Unique_Node_List memops;
1024 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1025 MemNode* previous = last;
1026 while (true) {
1027 assert(in_bb(current), "stay in block");
1028 memops.push(previous);
1029 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1030 Node* use = current->out(i);
1031 if (use->is_Mem() && use != previous)
1032 memops.push(use);
1033 }
1034 if(current == first) break;
1035 previous = current;
1036 current = current->in(MemNode::Memory)->as_Mem();
1037 }
1039 // determine which memory operations should be scheduled before the pack
1040 for (uint i = 1; i < memops.size(); i++) {
1041 Node *s1 = memops.at(i);
1042 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1043 for (uint j = 0; j< i; j++) {
1044 Node *s2 = memops.at(j);
1045 if (!independent(s1, s2)) {
1046 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1047 schedule_before_pack.push(s1); //s1 must be scheduled before
1048 Node_List* mem_pk = my_pack(s1);
1049 if (mem_pk != NULL) {
1050 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1051 Node* s = mem_pk->at(ii); // follow partner
1052 if (memops.member(s) && !schedule_before_pack.member(s))
1053 schedule_before_pack.push(s);
1054 }
1055 }
1056 }
1057 }
1058 }
1059 }
1060 }
1062 MemNode* lower_insert_pt = last;
1063 Node* upper_insert_pt = first->in(MemNode::Memory);
1064 previous = last; //previous store in pk
1065 current = last->in(MemNode::Memory)->as_Mem();
1067 //start scheduling from "last" to "first"
1068 while (true) {
1069 assert(in_bb(current), "stay in block");
1070 assert(in_pack(previous, pk), "previous stays in pack");
1071 Node* my_mem = current->in(MemNode::Memory);
1073 if (in_pack(current, pk)) {
1074 // Forward users of my memory state (except "previous) to my input memory state
1075 _igvn.hash_delete(current);
1076 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1077 Node* use = current->out(i);
1078 if (use->is_Mem() && use != previous) {
1079 assert(use->in(MemNode::Memory) == current, "must be");
1080 _igvn.hash_delete(use);
1081 if (schedule_before_pack.member(use)) {
1082 _igvn.hash_delete(upper_insert_pt);
1083 use->set_req(MemNode::Memory, upper_insert_pt);
1084 } else {
1085 _igvn.hash_delete(lower_insert_pt);
1086 use->set_req(MemNode::Memory, lower_insert_pt);
1087 }
1088 _igvn._worklist.push(use);
1089 --i; // deleted this edge; rescan position
1090 }
1091 }
1092 previous = current;
1093 } else { // !in_pack(current, pk) ==> a sandwiched store
1094 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1095 }
1097 if (current == first) break;
1098 current = my_mem->as_Mem();
1099 } // end while
1100 } else if (pk->at(0)->is_Load()) { //load
1101 // all loads in the pack should have the same memory state. By default,
1102 // we use the memory state of the last load. However, if any load could
1103 // not be moved down due to the dependence constraint, we use the memory
1104 // state of the first load.
1105 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1106 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1107 bool schedule_last = true;
1108 for (uint i = 0; i < pk->size(); i++) {
1109 Node* ld = pk->at(i);
1110 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1111 current=current->in(MemNode::Memory)) {
1112 assert(current != first_mem, "corrupted memory graph");
1113 if(current->is_Mem() && !independent(current, ld)){
1114 schedule_last = false; // a later store depends on this load
1115 break;
1116 }
1117 }
1118 }
1120 Node* mem_input = schedule_last ? last_mem : first_mem;
1121 _igvn.hash_delete(mem_input);
1122 // Give each load the same memory state
1123 for (uint i = 0; i < pk->size(); i++) {
1124 LoadNode* ld = pk->at(i)->as_Load();
1125 _igvn.hash_delete(ld);
1126 ld->set_req(MemNode::Memory, mem_input);
1127 _igvn._worklist.push(ld);
1128 }
1129 }
1130 }
1132 //------------------------------output---------------------------
1133 // Convert packs into vector node operations
1134 void SuperWord::output() {
1135 if (_packset.length() == 0) return;
1137 #ifndef PRODUCT
1138 if (TraceLoopOpts) {
1139 tty->print("SuperWord ");
1140 lpt()->dump_head();
1141 }
1142 #endif
1144 // MUST ENSURE main loop's initial value is properly aligned:
1145 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1147 align_initial_loop_index(align_to_ref());
1149 // Insert extract (unpack) operations for scalar uses
1150 for (int i = 0; i < _packset.length(); i++) {
1151 insert_extracts(_packset.at(i));
1152 }
1154 for (int i = 0; i < _block.length(); i++) {
1155 Node* n = _block.at(i);
1156 Node_List* p = my_pack(n);
1157 if (p && n == executed_last(p)) {
1158 uint vlen = p->size();
1159 Node* vn = NULL;
1160 Node* low_adr = p->at(0);
1161 Node* first = executed_first(p);
1162 if (n->is_Load()) {
1163 int opc = n->Opcode();
1164 Node* ctl = n->in(MemNode::Control);
1165 Node* mem = first->in(MemNode::Memory);
1166 Node* adr = low_adr->in(MemNode::Address);
1167 const TypePtr* atyp = n->adr_type();
1168 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1170 } else if (n->is_Store()) {
1171 // Promote value to be stored to vector
1172 Node* val = vector_opd(p, MemNode::ValueIn);
1174 int opc = n->Opcode();
1175 Node* ctl = n->in(MemNode::Control);
1176 Node* mem = first->in(MemNode::Memory);
1177 Node* adr = low_adr->in(MemNode::Address);
1178 const TypePtr* atyp = n->adr_type();
1179 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1181 } else if (n->req() == 3) {
1182 // Promote operands to vector
1183 Node* in1 = vector_opd(p, 1);
1184 Node* in2 = vector_opd(p, 2);
1185 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1187 } else {
1188 ShouldNotReachHere();
1189 }
1191 _phase->_igvn.register_new_node_with_optimizer(vn);
1192 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1193 for (uint j = 0; j < p->size(); j++) {
1194 Node* pm = p->at(j);
1195 _igvn.replace_node(pm, vn);
1196 }
1197 _igvn._worklist.push(vn);
1198 }
1199 }
1200 }
1202 //------------------------------vector_opd---------------------------
1203 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1204 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1205 Node* p0 = p->at(0);
1206 uint vlen = p->size();
1207 Node* opd = p0->in(opd_idx);
1209 bool same_opd = true;
1210 for (uint i = 1; i < vlen; i++) {
1211 Node* pi = p->at(i);
1212 Node* in = pi->in(opd_idx);
1213 if (opd != in) {
1214 same_opd = false;
1215 break;
1216 }
1217 }
1219 if (same_opd) {
1220 if (opd->is_Vector() || opd->is_VectorLoad()) {
1221 return opd; // input is matching vector
1222 }
1223 assert(!opd->is_VectorStore(), "such vector is not expected here");
1224 // Convert scalar input to vector. Use p0's type because it's container
1225 // maybe smaller than the operand's container.
1226 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1227 const Type* p0_t = velt_type(p0);
1228 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1229 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1231 _phase->_igvn.register_new_node_with_optimizer(vn);
1232 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1233 return vn;
1234 }
1236 // Insert pack operation
1237 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1238 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1240 for (uint i = 1; i < vlen; i++) {
1241 Node* pi = p->at(i);
1242 Node* in = pi->in(opd_idx);
1243 assert(my_pack(in) == NULL, "Should already have been unpacked");
1244 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1245 pk->add_opd(in);
1246 }
1247 _phase->_igvn.register_new_node_with_optimizer(pk);
1248 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1249 return pk;
1250 }
1252 //------------------------------insert_extracts---------------------------
1253 // If a use of pack p is not a vector use, then replace the
1254 // use with an extract operation.
1255 void SuperWord::insert_extracts(Node_List* p) {
1256 if (p->at(0)->is_Store()) return;
1257 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1259 // Inspect each use of each pack member. For each use that is
1260 // not a vector use, replace the use with an extract operation.
1262 for (uint i = 0; i < p->size(); i++) {
1263 Node* def = p->at(i);
1264 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1265 Node* use = def->fast_out(j);
1266 for (uint k = 0; k < use->req(); k++) {
1267 Node* n = use->in(k);
1268 if (def == n) {
1269 if (!is_vector_use(use, k)) {
1270 _n_idx_list.push(use, k);
1271 }
1272 }
1273 }
1274 }
1275 }
1277 while (_n_idx_list.is_nonempty()) {
1278 Node* use = _n_idx_list.node();
1279 int idx = _n_idx_list.index();
1280 _n_idx_list.pop();
1281 Node* def = use->in(idx);
1283 // Insert extract operation
1284 _igvn.hash_delete(def);
1285 _igvn.hash_delete(use);
1286 int def_pos = alignment(def) / data_size(def);
1287 const Type* def_t = velt_type(def);
1289 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1290 _phase->_igvn.register_new_node_with_optimizer(ex);
1291 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1292 use->set_req(idx, ex);
1293 _igvn._worklist.push(def);
1294 _igvn._worklist.push(use);
1296 bb_insert_after(ex, bb_idx(def));
1297 set_velt_type(ex, def_t);
1298 }
1299 }
1301 //------------------------------is_vector_use---------------------------
1302 // Is use->in(u_idx) a vector use?
1303 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1304 Node_List* u_pk = my_pack(use);
1305 if (u_pk == NULL) return false;
1306 Node* def = use->in(u_idx);
1307 Node_List* d_pk = my_pack(def);
1308 if (d_pk == NULL) {
1309 // check for scalar promotion
1310 Node* n = u_pk->at(0)->in(u_idx);
1311 for (uint i = 1; i < u_pk->size(); i++) {
1312 if (u_pk->at(i)->in(u_idx) != n) return false;
1313 }
1314 return true;
1315 }
1316 if (u_pk->size() != d_pk->size())
1317 return false;
1318 for (uint i = 0; i < u_pk->size(); i++) {
1319 Node* ui = u_pk->at(i);
1320 Node* di = d_pk->at(i);
1321 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1322 return false;
1323 }
1324 return true;
1325 }
1327 //------------------------------construct_bb---------------------------
1328 // Construct reverse postorder list of block members
1329 void SuperWord::construct_bb() {
1330 Node* entry = bb();
1332 assert(_stk.length() == 0, "stk is empty");
1333 assert(_block.length() == 0, "block is empty");
1334 assert(_data_entry.length() == 0, "data_entry is empty");
1335 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1336 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1338 // Find non-control nodes with no inputs from within block,
1339 // create a temporary map from node _idx to bb_idx for use
1340 // by the visited and post_visited sets,
1341 // and count number of nodes in block.
1342 int bb_ct = 0;
1343 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1344 Node *n = lpt()->_body.at(i);
1345 set_bb_idx(n, i); // Create a temporary map
1346 if (in_bb(n)) {
1347 bb_ct++;
1348 if (!n->is_CFG()) {
1349 bool found = false;
1350 for (uint j = 0; j < n->req(); j++) {
1351 Node* def = n->in(j);
1352 if (def && in_bb(def)) {
1353 found = true;
1354 break;
1355 }
1356 }
1357 if (!found) {
1358 assert(n != entry, "can't be entry");
1359 _data_entry.push(n);
1360 }
1361 }
1362 }
1363 }
1365 // Find memory slices (head and tail)
1366 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1367 Node *n = lp()->fast_out(i);
1368 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1369 Node* n_tail = n->in(LoopNode::LoopBackControl);
1370 if (n_tail != n->in(LoopNode::EntryControl)) {
1371 _mem_slice_head.push(n);
1372 _mem_slice_tail.push(n_tail);
1373 }
1374 }
1375 }
1377 // Create an RPO list of nodes in block
1379 visited_clear();
1380 post_visited_clear();
1382 // Push all non-control nodes with no inputs from within block, then control entry
1383 for (int j = 0; j < _data_entry.length(); j++) {
1384 Node* n = _data_entry.at(j);
1385 visited_set(n);
1386 _stk.push(n);
1387 }
1388 visited_set(entry);
1389 _stk.push(entry);
1391 // Do a depth first walk over out edges
1392 int rpo_idx = bb_ct - 1;
1393 int size;
1394 while ((size = _stk.length()) > 0) {
1395 Node* n = _stk.top(); // Leave node on stack
1396 if (!visited_test_set(n)) {
1397 // forward arc in graph
1398 } else if (!post_visited_test(n)) {
1399 // cross or back arc
1400 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1401 Node *use = n->fast_out(i);
1402 if (in_bb(use) && !visited_test(use) &&
1403 // Don't go around backedge
1404 (!use->is_Phi() || n == entry)) {
1405 _stk.push(use);
1406 }
1407 }
1408 if (_stk.length() == size) {
1409 // There were no additional uses, post visit node now
1410 _stk.pop(); // Remove node from stack
1411 assert(rpo_idx >= 0, "");
1412 _block.at_put_grow(rpo_idx, n);
1413 rpo_idx--;
1414 post_visited_set(n);
1415 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1416 }
1417 } else {
1418 _stk.pop(); // Remove post-visited node from stack
1419 }
1420 }
1422 // Create real map of block indices for nodes
1423 for (int j = 0; j < _block.length(); j++) {
1424 Node* n = _block.at(j);
1425 set_bb_idx(n, j);
1426 }
1428 initialize_bb(); // Ensure extra info is allocated.
1430 #ifndef PRODUCT
1431 if (TraceSuperWord) {
1432 print_bb();
1433 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1434 for (int m = 0; m < _data_entry.length(); m++) {
1435 tty->print("%3d ", m);
1436 _data_entry.at(m)->dump();
1437 }
1438 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1439 for (int m = 0; m < _mem_slice_head.length(); m++) {
1440 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1441 tty->print(" "); _mem_slice_tail.at(m)->dump();
1442 }
1443 }
1444 #endif
1445 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1446 }
1448 //------------------------------initialize_bb---------------------------
1449 // Initialize per node info
1450 void SuperWord::initialize_bb() {
1451 Node* last = _block.at(_block.length() - 1);
1452 grow_node_info(bb_idx(last));
1453 }
1455 //------------------------------bb_insert_after---------------------------
1456 // Insert n into block after pos
1457 void SuperWord::bb_insert_after(Node* n, int pos) {
1458 int n_pos = pos + 1;
1459 // Make room
1460 for (int i = _block.length() - 1; i >= n_pos; i--) {
1461 _block.at_put_grow(i+1, _block.at(i));
1462 }
1463 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1464 _node_info.at_put_grow(j+1, _node_info.at(j));
1465 }
1466 // Set value
1467 _block.at_put_grow(n_pos, n);
1468 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1469 // Adjust map from node->_idx to _block index
1470 for (int i = n_pos; i < _block.length(); i++) {
1471 set_bb_idx(_block.at(i), i);
1472 }
1473 }
1475 //------------------------------compute_max_depth---------------------------
1476 // Compute max depth for expressions from beginning of block
1477 // Use to prune search paths during test for independence.
1478 void SuperWord::compute_max_depth() {
1479 int ct = 0;
1480 bool again;
1481 do {
1482 again = false;
1483 for (int i = 0; i < _block.length(); i++) {
1484 Node* n = _block.at(i);
1485 if (!n->is_Phi()) {
1486 int d_orig = depth(n);
1487 int d_in = 0;
1488 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1489 Node* pred = preds.current();
1490 if (in_bb(pred)) {
1491 d_in = MAX2(d_in, depth(pred));
1492 }
1493 }
1494 if (d_in + 1 != d_orig) {
1495 set_depth(n, d_in + 1);
1496 again = true;
1497 }
1498 }
1499 }
1500 ct++;
1501 } while (again);
1502 #ifndef PRODUCT
1503 if (TraceSuperWord && Verbose)
1504 tty->print_cr("compute_max_depth iterated: %d times", ct);
1505 #endif
1506 }
1508 //-------------------------compute_vector_element_type-----------------------
1509 // Compute necessary vector element type for expressions
1510 // This propagates backwards a narrower integer type when the
1511 // upper bits of the value are not needed.
1512 // Example: char a,b,c; a = b + c;
1513 // Normally the type of the add is integer, but for packed character
1514 // operations the type of the add needs to be char.
1515 void SuperWord::compute_vector_element_type() {
1516 #ifndef PRODUCT
1517 if (TraceSuperWord && Verbose)
1518 tty->print_cr("\ncompute_velt_type:");
1519 #endif
1521 // Initial type
1522 for (int i = 0; i < _block.length(); i++) {
1523 Node* n = _block.at(i);
1524 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1525 : _igvn.type(n);
1526 const Type* vt = container_type(t);
1527 set_velt_type(n, vt);
1528 }
1530 // Propagate narrowed type backwards through operations
1531 // that don't depend on higher order bits
1532 for (int i = _block.length() - 1; i >= 0; i--) {
1533 Node* n = _block.at(i);
1534 // Only integer types need be examined
1535 if (n->bottom_type()->isa_int()) {
1536 uint start, end;
1537 vector_opd_range(n, &start, &end);
1538 const Type* vt = velt_type(n);
1540 for (uint j = start; j < end; j++) {
1541 Node* in = n->in(j);
1542 // Don't propagate through a type conversion
1543 if (n->bottom_type() != in->bottom_type())
1544 continue;
1545 switch(in->Opcode()) {
1546 case Op_AddI: case Op_AddL:
1547 case Op_SubI: case Op_SubL:
1548 case Op_MulI: case Op_MulL:
1549 case Op_AndI: case Op_AndL:
1550 case Op_OrI: case Op_OrL:
1551 case Op_XorI: case Op_XorL:
1552 case Op_LShiftI: case Op_LShiftL:
1553 case Op_CMoveI: case Op_CMoveL:
1554 if (in_bb(in)) {
1555 bool same_type = true;
1556 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1557 Node *use = in->fast_out(k);
1558 if (!in_bb(use) || velt_type(use) != vt) {
1559 same_type = false;
1560 break;
1561 }
1562 }
1563 if (same_type) {
1564 set_velt_type(in, vt);
1565 }
1566 }
1567 }
1568 }
1569 }
1570 }
1571 #ifndef PRODUCT
1572 if (TraceSuperWord && Verbose) {
1573 for (int i = 0; i < _block.length(); i++) {
1574 Node* n = _block.at(i);
1575 velt_type(n)->dump();
1576 tty->print("\t");
1577 n->dump();
1578 }
1579 }
1580 #endif
1581 }
1583 //------------------------------memory_alignment---------------------------
1584 // Alignment within a vector memory reference
1585 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1586 SWPointer p(s, this);
1587 if (!p.valid()) {
1588 return bottom_align;
1589 }
1590 int offset = p.offset_in_bytes();
1591 offset += iv_adjust_in_bytes;
1592 int off_rem = offset % vector_width_in_bytes();
1593 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1594 return off_mod;
1595 }
1597 //---------------------------container_type---------------------------
1598 // Smallest type containing range of values
1599 const Type* SuperWord::container_type(const Type* t) {
1600 const Type* tp = t->make_ptr();
1601 if (tp && tp->isa_aryptr()) {
1602 t = tp->is_aryptr()->elem();
1603 }
1604 if (t->basic_type() == T_INT) {
1605 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
1606 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
1607 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
1608 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1609 return TypeInt::INT;
1610 }
1611 return t;
1612 }
1614 //-------------------------vector_opd_range-----------------------
1615 // (Start, end] half-open range defining which operands are vector
1616 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1617 switch (n->Opcode()) {
1618 case Op_LoadB: case Op_LoadUS:
1619 case Op_LoadI: case Op_LoadL:
1620 case Op_LoadF: case Op_LoadD:
1621 case Op_LoadP:
1622 *start = 0;
1623 *end = 0;
1624 return;
1625 case Op_StoreB: case Op_StoreC:
1626 case Op_StoreI: case Op_StoreL:
1627 case Op_StoreF: case Op_StoreD:
1628 case Op_StoreP:
1629 *start = MemNode::ValueIn;
1630 *end = *start + 1;
1631 return;
1632 case Op_LShiftI: case Op_LShiftL:
1633 *start = 1;
1634 *end = 2;
1635 return;
1636 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1637 *start = 2;
1638 *end = n->req();
1639 return;
1640 }
1641 *start = 1;
1642 *end = n->req(); // default is all operands
1643 }
1645 //------------------------------in_packset---------------------------
1646 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1647 bool SuperWord::in_packset(Node* s1, Node* s2) {
1648 for (int i = 0; i < _packset.length(); i++) {
1649 Node_List* p = _packset.at(i);
1650 assert(p->size() == 2, "must be");
1651 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1652 return true;
1653 }
1654 }
1655 return false;
1656 }
1658 //------------------------------in_pack---------------------------
1659 // Is s in pack p?
1660 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1661 for (uint i = 0; i < p->size(); i++) {
1662 if (p->at(i) == s) {
1663 return p;
1664 }
1665 }
1666 return NULL;
1667 }
1669 //------------------------------remove_pack_at---------------------------
1670 // Remove the pack at position pos in the packset
1671 void SuperWord::remove_pack_at(int pos) {
1672 Node_List* p = _packset.at(pos);
1673 for (uint i = 0; i < p->size(); i++) {
1674 Node* s = p->at(i);
1675 set_my_pack(s, NULL);
1676 }
1677 _packset.remove_at(pos);
1678 }
1680 //------------------------------executed_first---------------------------
1681 // Return the node executed first in pack p. Uses the RPO block list
1682 // to determine order.
1683 Node* SuperWord::executed_first(Node_List* p) {
1684 Node* n = p->at(0);
1685 int n_rpo = bb_idx(n);
1686 for (uint i = 1; i < p->size(); i++) {
1687 Node* s = p->at(i);
1688 int s_rpo = bb_idx(s);
1689 if (s_rpo < n_rpo) {
1690 n = s;
1691 n_rpo = s_rpo;
1692 }
1693 }
1694 return n;
1695 }
1697 //------------------------------executed_last---------------------------
1698 // Return the node executed last in pack p.
1699 Node* SuperWord::executed_last(Node_List* p) {
1700 Node* n = p->at(0);
1701 int n_rpo = bb_idx(n);
1702 for (uint i = 1; i < p->size(); i++) {
1703 Node* s = p->at(i);
1704 int s_rpo = bb_idx(s);
1705 if (s_rpo > n_rpo) {
1706 n = s;
1707 n_rpo = s_rpo;
1708 }
1709 }
1710 return n;
1711 }
1713 //----------------------------align_initial_loop_index---------------------------
1714 // Adjust pre-loop limit so that in main loop, a load/store reference
1715 // to align_to_ref will be a position zero in the vector.
1716 // (iv + k) mod vector_align == 0
1717 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1718 CountedLoopNode *main_head = lp()->as_CountedLoop();
1719 assert(main_head->is_main_loop(), "");
1720 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1721 assert(pre_end != NULL, "");
1722 Node *pre_opaq1 = pre_end->limit();
1723 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1724 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1725 Node *lim0 = pre_opaq->in(1);
1727 // Where we put new limit calculations
1728 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1730 // Ensure the original loop limit is available from the
1731 // pre-loop Opaque1 node.
1732 Node *orig_limit = pre_opaq->original_loop_limit();
1733 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1735 SWPointer align_to_ref_p(align_to_ref, this);
1737 // Given:
1738 // lim0 == original pre loop limit
1739 // V == v_align (power of 2)
1740 // invar == extra invariant piece of the address expression
1741 // e == k [ +/- invar ]
1742 //
1743 // When reassociating expressions involving '%' the basic rules are:
1744 // (a - b) % k == 0 => a % k == b % k
1745 // and:
1746 // (a + b) % k == 0 => a % k == (k - b) % k
1747 //
1748 // For stride > 0 && scale > 0,
1749 // Derive the new pre-loop limit "lim" such that the two constraints:
1750 // (1) lim = lim0 + N (where N is some positive integer < V)
1751 // (2) (e + lim) % V == 0
1752 // are true.
1753 //
1754 // Substituting (1) into (2),
1755 // (e + lim0 + N) % V == 0
1756 // solve for N:
1757 // N = (V - (e + lim0)) % V
1758 // substitute back into (1), so that new limit
1759 // lim = lim0 + (V - (e + lim0)) % V
1760 //
1761 // For stride > 0 && scale < 0
1762 // Constraints:
1763 // lim = lim0 + N
1764 // (e - lim) % V == 0
1765 // Solving for lim:
1766 // (e - lim0 - N) % V == 0
1767 // N = (e - lim0) % V
1768 // lim = lim0 + (e - lim0) % V
1769 //
1770 // For stride < 0 && scale > 0
1771 // Constraints:
1772 // lim = lim0 - N
1773 // (e + lim) % V == 0
1774 // Solving for lim:
1775 // (e + lim0 - N) % V == 0
1776 // N = (e + lim0) % V
1777 // lim = lim0 - (e + lim0) % V
1778 //
1779 // For stride < 0 && scale < 0
1780 // Constraints:
1781 // lim = lim0 - N
1782 // (e - lim) % V == 0
1783 // Solving for lim:
1784 // (e - lim0 + N) % V == 0
1785 // N = (V - (e - lim0)) % V
1786 // lim = lim0 - (V - (e - lim0)) % V
1788 int stride = iv_stride();
1789 int scale = align_to_ref_p.scale_in_bytes();
1790 int elt_size = align_to_ref_p.memory_size();
1791 int v_align = vector_width_in_bytes() / elt_size;
1792 int k = align_to_ref_p.offset_in_bytes() / elt_size;
1794 Node *kn = _igvn.intcon(k);
1796 Node *e = kn;
1797 if (align_to_ref_p.invar() != NULL) {
1798 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1799 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1800 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1801 _phase->_igvn.register_new_node_with_optimizer(aref);
1802 _phase->set_ctrl(aref, pre_ctrl);
1803 if (align_to_ref_p.negate_invar()) {
1804 e = new (_phase->C, 3) SubINode(e, aref);
1805 } else {
1806 e = new (_phase->C, 3) AddINode(e, aref);
1807 }
1808 _phase->_igvn.register_new_node_with_optimizer(e);
1809 _phase->set_ctrl(e, pre_ctrl);
1810 }
1812 // compute e +/- lim0
1813 if (scale < 0) {
1814 e = new (_phase->C, 3) SubINode(e, lim0);
1815 } else {
1816 e = new (_phase->C, 3) AddINode(e, lim0);
1817 }
1818 _phase->_igvn.register_new_node_with_optimizer(e);
1819 _phase->set_ctrl(e, pre_ctrl);
1821 if (stride * scale > 0) {
1822 // compute V - (e +/- lim0)
1823 Node* va = _igvn.intcon(v_align);
1824 e = new (_phase->C, 3) SubINode(va, e);
1825 _phase->_igvn.register_new_node_with_optimizer(e);
1826 _phase->set_ctrl(e, pre_ctrl);
1827 }
1828 // compute N = (exp) % V
1829 Node* va_msk = _igvn.intcon(v_align - 1);
1830 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1831 _phase->_igvn.register_new_node_with_optimizer(N);
1832 _phase->set_ctrl(N, pre_ctrl);
1834 // substitute back into (1), so that new limit
1835 // lim = lim0 + N
1836 Node* lim;
1837 if (stride < 0) {
1838 lim = new (_phase->C, 3) SubINode(lim0, N);
1839 } else {
1840 lim = new (_phase->C, 3) AddINode(lim0, N);
1841 }
1842 _phase->_igvn.register_new_node_with_optimizer(lim);
1843 _phase->set_ctrl(lim, pre_ctrl);
1844 Node* constrained =
1845 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1846 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1847 _phase->_igvn.register_new_node_with_optimizer(constrained);
1848 _phase->set_ctrl(constrained, pre_ctrl);
1849 _igvn.hash_delete(pre_opaq);
1850 pre_opaq->set_req(1, constrained);
1851 }
1853 //----------------------------get_pre_loop_end---------------------------
1854 // Find pre loop end from main loop. Returns null if none.
1855 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1856 Node *ctrl = cl->in(LoopNode::EntryControl);
1857 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1858 Node *iffm = ctrl->in(0);
1859 if (!iffm->is_If()) return NULL;
1860 Node *p_f = iffm->in(0);
1861 if (!p_f->is_IfFalse()) return NULL;
1862 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1863 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1864 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1865 return pre_end;
1866 }
1869 //------------------------------init---------------------------
1870 void SuperWord::init() {
1871 _dg.init();
1872 _packset.clear();
1873 _disjoint_ptrs.clear();
1874 _block.clear();
1875 _data_entry.clear();
1876 _mem_slice_head.clear();
1877 _mem_slice_tail.clear();
1878 _node_info.clear();
1879 _align_to_ref = NULL;
1880 _lpt = NULL;
1881 _lp = NULL;
1882 _bb = NULL;
1883 _iv = NULL;
1884 }
1886 //------------------------------print_packset---------------------------
1887 void SuperWord::print_packset() {
1888 #ifndef PRODUCT
1889 tty->print_cr("packset");
1890 for (int i = 0; i < _packset.length(); i++) {
1891 tty->print_cr("Pack: %d", i);
1892 Node_List* p = _packset.at(i);
1893 print_pack(p);
1894 }
1895 #endif
1896 }
1898 //------------------------------print_pack---------------------------
1899 void SuperWord::print_pack(Node_List* p) {
1900 for (uint i = 0; i < p->size(); i++) {
1901 print_stmt(p->at(i));
1902 }
1903 }
1905 //------------------------------print_bb---------------------------
1906 void SuperWord::print_bb() {
1907 #ifndef PRODUCT
1908 tty->print_cr("\nBlock");
1909 for (int i = 0; i < _block.length(); i++) {
1910 Node* n = _block.at(i);
1911 tty->print("%d ", i);
1912 if (n) {
1913 n->dump();
1914 }
1915 }
1916 #endif
1917 }
1919 //------------------------------print_stmt---------------------------
1920 void SuperWord::print_stmt(Node* s) {
1921 #ifndef PRODUCT
1922 tty->print(" align: %d \t", alignment(s));
1923 s->dump();
1924 #endif
1925 }
1927 //------------------------------blank---------------------------
1928 char* SuperWord::blank(uint depth) {
1929 static char blanks[101];
1930 assert(depth < 101, "too deep");
1931 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1932 blanks[depth] = '\0';
1933 return blanks;
1934 }
1937 //==============================SWPointer===========================
1939 //----------------------------SWPointer------------------------
1940 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1941 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
1942 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1944 Node* adr = mem->in(MemNode::Address);
1945 if (!adr->is_AddP()) {
1946 assert(!valid(), "too complex");
1947 return;
1948 }
1949 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1950 Node* base = adr->in(AddPNode::Base);
1951 //unsafe reference could not be aligned appropriately without runtime checking
1952 if (base == NULL || base->bottom_type() == Type::TOP) {
1953 assert(!valid(), "unsafe access");
1954 return;
1955 }
1956 for (int i = 0; i < 3; i++) {
1957 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1958 assert(!valid(), "too complex");
1959 return;
1960 }
1961 adr = adr->in(AddPNode::Address);
1962 if (base == adr || !adr->is_AddP()) {
1963 break; // stop looking at addp's
1964 }
1965 }
1966 _base = base;
1967 _adr = adr;
1968 assert(valid(), "Usable");
1969 }
1971 // Following is used to create a temporary object during
1972 // the pattern match of an address expression.
1973 SWPointer::SWPointer(SWPointer* p) :
1974 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
1975 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1977 //------------------------scaled_iv_plus_offset--------------------
1978 // Match: k*iv + offset
1979 // where: k is a constant that maybe zero, and
1980 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1981 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1982 if (scaled_iv(n)) {
1983 return true;
1984 }
1985 if (offset_plus_k(n)) {
1986 return true;
1987 }
1988 int opc = n->Opcode();
1989 if (opc == Op_AddI) {
1990 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1991 return true;
1992 }
1993 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1994 return true;
1995 }
1996 } else if (opc == Op_SubI) {
1997 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1998 return true;
1999 }
2000 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2001 _scale *= -1;
2002 return true;
2003 }
2004 }
2005 return false;
2006 }
2008 //----------------------------scaled_iv------------------------
2009 // Match: k*iv where k is a constant that's not zero
2010 bool SWPointer::scaled_iv(Node* n) {
2011 if (_scale != 0) {
2012 return false; // already found a scale
2013 }
2014 if (n == iv()) {
2015 _scale = 1;
2016 return true;
2017 }
2018 int opc = n->Opcode();
2019 if (opc == Op_MulI) {
2020 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2021 _scale = n->in(2)->get_int();
2022 return true;
2023 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2024 _scale = n->in(1)->get_int();
2025 return true;
2026 }
2027 } else if (opc == Op_LShiftI) {
2028 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2029 _scale = 1 << n->in(2)->get_int();
2030 return true;
2031 }
2032 } else if (opc == Op_ConvI2L) {
2033 if (scaled_iv_plus_offset(n->in(1))) {
2034 return true;
2035 }
2036 } else if (opc == Op_LShiftL) {
2037 if (!has_iv() && _invar == NULL) {
2038 // Need to preserve the current _offset value, so
2039 // create a temporary object for this expression subtree.
2040 // Hacky, so should re-engineer the address pattern match.
2041 SWPointer tmp(this);
2042 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2043 if (tmp._invar == NULL) {
2044 int mult = 1 << n->in(2)->get_int();
2045 _scale = tmp._scale * mult;
2046 _offset += tmp._offset * mult;
2047 return true;
2048 }
2049 }
2050 }
2051 }
2052 return false;
2053 }
2055 //----------------------------offset_plus_k------------------------
2056 // Match: offset is (k [+/- invariant])
2057 // where k maybe zero and invariant is optional, but not both.
2058 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2059 int opc = n->Opcode();
2060 if (opc == Op_ConI) {
2061 _offset += negate ? -(n->get_int()) : n->get_int();
2062 return true;
2063 } else if (opc == Op_ConL) {
2064 // Okay if value fits into an int
2065 const TypeLong* t = n->find_long_type();
2066 if (t->higher_equal(TypeLong::INT)) {
2067 jlong loff = n->get_long();
2068 jint off = (jint)loff;
2069 _offset += negate ? -off : loff;
2070 return true;
2071 }
2072 return false;
2073 }
2074 if (_invar != NULL) return false; // already have an invariant
2075 if (opc == Op_AddI) {
2076 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2077 _negate_invar = negate;
2078 _invar = n->in(1);
2079 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2080 return true;
2081 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2082 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2083 _negate_invar = negate;
2084 _invar = n->in(2);
2085 return true;
2086 }
2087 }
2088 if (opc == Op_SubI) {
2089 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2090 _negate_invar = negate;
2091 _invar = n->in(1);
2092 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2093 return true;
2094 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2095 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2096 _negate_invar = !negate;
2097 _invar = n->in(2);
2098 return true;
2099 }
2100 }
2101 if (invariant(n)) {
2102 _negate_invar = negate;
2103 _invar = n;
2104 return true;
2105 }
2106 return false;
2107 }
2109 //----------------------------print------------------------
2110 void SWPointer::print() {
2111 #ifndef PRODUCT
2112 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2113 _base != NULL ? _base->_idx : 0,
2114 _adr != NULL ? _adr->_idx : 0,
2115 _scale, _offset,
2116 _negate_invar?'-':'+',
2117 _invar != NULL ? _invar->_idx : 0);
2118 #endif
2119 }
2121 // ========================= OrderedPair =====================
2123 const OrderedPair OrderedPair::initial;
2125 // ========================= SWNodeInfo =====================
2127 const SWNodeInfo SWNodeInfo::initial;
2130 // ============================ DepGraph ===========================
2132 //------------------------------make_node---------------------------
2133 // Make a new dependence graph node for an ideal node.
2134 DepMem* DepGraph::make_node(Node* node) {
2135 DepMem* m = new (_arena) DepMem(node);
2136 if (node != NULL) {
2137 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2138 _map.at_put_grow(node->_idx, m);
2139 }
2140 return m;
2141 }
2143 //------------------------------make_edge---------------------------
2144 // Make a new dependence graph edge from dpred -> dsucc
2145 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2146 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2147 dpred->set_out_head(e);
2148 dsucc->set_in_head(e);
2149 return e;
2150 }
2152 // ========================== DepMem ========================
2154 //------------------------------in_cnt---------------------------
2155 int DepMem::in_cnt() {
2156 int ct = 0;
2157 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2158 return ct;
2159 }
2161 //------------------------------out_cnt---------------------------
2162 int DepMem::out_cnt() {
2163 int ct = 0;
2164 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2165 return ct;
2166 }
2168 //------------------------------print-----------------------------
2169 void DepMem::print() {
2170 #ifndef PRODUCT
2171 tty->print(" DepNode %d (", _node->_idx);
2172 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2173 Node* pred = p->pred()->node();
2174 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2175 }
2176 tty->print(") [");
2177 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2178 Node* succ = s->succ()->node();
2179 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2180 }
2181 tty->print_cr(" ]");
2182 #endif
2183 }
2185 // =========================== DepEdge =========================
2187 //------------------------------DepPreds---------------------------
2188 void DepEdge::print() {
2189 #ifndef PRODUCT
2190 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2191 #endif
2192 }
2194 // =========================== DepPreds =========================
2195 // Iterator over predecessor edges in the dependence graph.
2197 //------------------------------DepPreds---------------------------
2198 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2199 _n = n;
2200 _done = false;
2201 if (_n->is_Store() || _n->is_Load()) {
2202 _next_idx = MemNode::Address;
2203 _end_idx = n->req();
2204 _dep_next = dg.dep(_n)->in_head();
2205 } else if (_n->is_Mem()) {
2206 _next_idx = 0;
2207 _end_idx = 0;
2208 _dep_next = dg.dep(_n)->in_head();
2209 } else {
2210 _next_idx = 1;
2211 _end_idx = _n->req();
2212 _dep_next = NULL;
2213 }
2214 next();
2215 }
2217 //------------------------------next---------------------------
2218 void DepPreds::next() {
2219 if (_dep_next != NULL) {
2220 _current = _dep_next->pred()->node();
2221 _dep_next = _dep_next->next_in();
2222 } else if (_next_idx < _end_idx) {
2223 _current = _n->in(_next_idx++);
2224 } else {
2225 _done = true;
2226 }
2227 }
2229 // =========================== DepSuccs =========================
2230 // Iterator over successor edges in the dependence graph.
2232 //------------------------------DepSuccs---------------------------
2233 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2234 _n = n;
2235 _done = false;
2236 if (_n->is_Load()) {
2237 _next_idx = 0;
2238 _end_idx = _n->outcnt();
2239 _dep_next = dg.dep(_n)->out_head();
2240 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2241 _next_idx = 0;
2242 _end_idx = 0;
2243 _dep_next = dg.dep(_n)->out_head();
2244 } else {
2245 _next_idx = 0;
2246 _end_idx = _n->outcnt();
2247 _dep_next = NULL;
2248 }
2249 next();
2250 }
2252 //-------------------------------next---------------------------
2253 void DepSuccs::next() {
2254 if (_dep_next != NULL) {
2255 _current = _dep_next->succ()->node();
2256 _dep_next = _dep_next->next_out();
2257 } else if (_next_idx < _end_idx) {
2258 _current = _n->raw_out(_next_idx++);
2259 } else {
2260 _done = true;
2261 }
2262 }