1 /*
2 * Copyright (c) 2007, 2014, 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 */
23
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/castnode.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/divnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/memnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/opcodes.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/superword.hpp"
39 #include "opto/vectornode.hpp"
40
41 //
42 // S U P E R W O R D T R A N S F O R M
43 //=============================================================================
44
45 //------------------------------SuperWord---------------------------
46 SuperWord::SuperWord(PhaseIdealLoop* phase) :
47 _phase(phase),
48 _igvn(phase->_igvn),
49 _arena(phase->C->comp_arena()),
50 _packset(arena(), 8, 0, NULL), // packs for the current block
51 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
52 _block(arena(), 8, 0, NULL), // nodes in current block
53 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
54 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
55 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
56 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
57 _align_to_ref(NULL), // memory reference to align vectors to
58 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
59 _dg(_arena), // dependence graph
60 _visited(arena()), // visited node set
61 _post_visited(arena()), // post visited node set
62 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
63 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
64 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
65 _lpt(NULL), // loop tree node
66 _lp(NULL), // LoopNode
67 _bb(NULL), // basic block
68 _iv(NULL), // induction var
69 _race_possible(false) // cases where SDMU is true
70 {}
71
72 //------------------------------transform_loop---------------------------
73 void SuperWord::transform_loop(IdealLoopTree* lpt) {
74 assert(UseSuperWord, "should be");
75 // Do vectors exist on this architecture?
76 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
77
78 assert(lpt->_head->is_CountedLoop(), "must be");
79 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
80
81 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
82
83 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
84
85 // Check for no control flow in body (other than exit)
86 Node *cl_exit = cl->loopexit();
87 if (cl_exit->in(0) != lpt->_head) return;
88
89 // Make sure the are no extra control users of the loop backedge
90 if (cl->back_control()->outcnt() != 1) {
91 return;
92 }
93
94 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
95 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
96 if (pre_end == NULL) return;
97 Node *pre_opaq1 = pre_end->limit();
98 if (pre_opaq1->Opcode() != Op_Opaque1) return;
99
100 init(); // initialize data structures
101
102 set_lpt(lpt);
103 set_lp(cl);
104
105 // For now, define one block which is the entire loop body
106 set_bb(cl);
107
108 assert(_packset.length() == 0, "packset must be empty");
109 SLP_extract();
110 }
111
112 //------------------------------SLP_extract---------------------------
113 // Extract the superword level parallelism
114 //
115 // 1) A reverse post-order of nodes in the block is constructed. By scanning
116 // this list from first to last, all definitions are visited before their uses.
117 //
118 // 2) A point-to-point dependence graph is constructed between memory references.
119 // This simplies the upcoming "independence" checker.
120 //
121 // 3) The maximum depth in the node graph from the beginning of the block
122 // to each node is computed. This is used to prune the graph search
123 // in the independence checker.
124 //
125 // 4) For integer types, the necessary bit width is propagated backwards
126 // from stores to allow packed operations on byte, char, and short
127 // integers. This reverses the promotion to type "int" that javac
128 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
129 //
130 // 5) One of the memory references is picked to be an aligned vector reference.
131 // The pre-loop trip count is adjusted to align this reference in the
132 // unrolled body.
133 //
134 // 6) The initial set of pack pairs is seeded with memory references.
135 //
136 // 7) The set of pack pairs is extended by following use->def and def->use links.
137 //
138 // 8) The pairs are combined into vector sized packs.
139 //
140 // 9) Reorder the memory slices to co-locate members of the memory packs.
141 //
142 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
143 // inserting scalar promotion, vector creation from multiple scalars, and
144 // extraction of scalar values from vectors.
145 //
146 void SuperWord::SLP_extract() {
147
148 // Ready the block
149 if (!construct_bb())
150 return; // Exit if no interesting nodes or complex graph.
151
152 dependence_graph();
153
154 compute_max_depth();
155
156 compute_vector_element_type();
157
158 // Attempt vectorization
159
160 find_adjacent_refs();
161
162 extend_packlist();
163
164 combine_packs();
165
166 construct_my_pack_map();
167
168 filter_packs();
169
170 schedule();
171
172 output();
173 }
174
175 //------------------------------find_adjacent_refs---------------------------
176 // Find the adjacent memory references and create pack pairs for them.
177 // This is the initial set of packs that will then be extended by
178 // following use->def and def->use links. The align positions are
179 // assigned relative to the reference "align_to_ref"
180 void SuperWord::find_adjacent_refs() {
181 // Get list of memory operations
182 Node_List memops;
183 for (int i = 0; i < _block.length(); i++) {
184 Node* n = _block.at(i);
185 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
186 is_java_primitive(n->as_Mem()->memory_type())) {
187 int align = memory_alignment(n->as_Mem(), 0);
188 if (align != bottom_align) {
189 memops.push(n);
190 }
191 }
192 }
193
194 Node_List align_to_refs;
195 int best_iv_adjustment = 0;
196 MemNode* best_align_to_mem_ref = NULL;
197
198 while (memops.size() != 0) {
199 // Find a memory reference to align to.
200 MemNode* mem_ref = find_align_to_ref(memops);
201 if (mem_ref == NULL) break;
202 align_to_refs.push(mem_ref);
203 int iv_adjustment = get_iv_adjustment(mem_ref);
204
205 if (best_align_to_mem_ref == NULL) {
206 // Set memory reference which is the best from all memory operations
207 // to be used for alignment. The pre-loop trip count is modified to align
208 // this reference to a vector-aligned address.
209 best_align_to_mem_ref = mem_ref;
210 best_iv_adjustment = iv_adjustment;
211 }
212
213 SWPointer align_to_ref_p(mem_ref, this);
214 // Set alignment relative to "align_to_ref" for all related memory operations.
215 for (int i = memops.size() - 1; i >= 0; i--) {
216 MemNode* s = memops.at(i)->as_Mem();
217 if (isomorphic(s, mem_ref)) {
218 SWPointer p2(s, this);
219 if (p2.comparable(align_to_ref_p)) {
220 int align = memory_alignment(s, iv_adjustment);
221 set_alignment(s, align);
222 }
223 }
224 }
225
226 // Create initial pack pairs of memory operations for which
227 // alignment is set and vectors will be aligned.
228 bool create_pack = true;
229 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
230 if (!Matcher::misaligned_vectors_ok()) {
231 int vw = vector_width(mem_ref);
232 int vw_best = vector_width(best_align_to_mem_ref);
233 if (vw > vw_best) {
234 // Do not vectorize a memory access with more elements per vector
235 // if unaligned memory access is not allowed because number of
236 // iterations in pre-loop will be not enough to align it.
237 create_pack = false;
238 }
239 }
240 } else {
241 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
242 // Can't allow vectorization of unaligned memory accesses with the
243 // same type since it could be overlapped accesses to the same array.
244 create_pack = false;
245 } else {
246 // Allow independent (different type) unaligned memory operations
247 // if HW supports them.
248 if (!Matcher::misaligned_vectors_ok()) {
249 create_pack = false;
250 } else {
251 // Check if packs of the same memory type but
252 // with a different alignment were created before.
253 for (uint i = 0; i < align_to_refs.size(); i++) {
254 MemNode* mr = align_to_refs.at(i)->as_Mem();
255 if (same_velt_type(mr, mem_ref) &&
256 memory_alignment(mr, iv_adjustment) != 0)
257 create_pack = false;
258 }
259 }
260 }
261 }
262 if (create_pack) {
263 for (uint i = 0; i < memops.size(); i++) {
264 Node* s1 = memops.at(i);
265 int align = alignment(s1);
266 if (align == top_align) continue;
267 for (uint j = 0; j < memops.size(); j++) {
268 Node* s2 = memops.at(j);
269 if (alignment(s2) == top_align) continue;
270 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
271 if (stmts_can_pack(s1, s2, align)) {
272 Node_List* pair = new Node_List();
273 pair->push(s1);
274 pair->push(s2);
275 _packset.append(pair);
276 }
277 }
278 }
279 }
280 } else { // Don't create unaligned pack
281 // First, remove remaining memory ops of the same type from the list.
282 for (int i = memops.size() - 1; i >= 0; i--) {
283 MemNode* s = memops.at(i)->as_Mem();
284 if (same_velt_type(s, mem_ref)) {
285 memops.remove(i);
286 }
287 }
288
289 // Second, remove already constructed packs of the same type.
290 for (int i = _packset.length() - 1; i >= 0; i--) {
291 Node_List* p = _packset.at(i);
292 MemNode* s = p->at(0)->as_Mem();
293 if (same_velt_type(s, mem_ref)) {
294 remove_pack_at(i);
295 }
296 }
297
298 // If needed find the best memory reference for loop alignment again.
299 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
300 // Put memory ops from remaining packs back on memops list for
301 // the best alignment search.
302 uint orig_msize = memops.size();
303 for (int i = 0; i < _packset.length(); i++) {
304 Node_List* p = _packset.at(i);
305 MemNode* s = p->at(0)->as_Mem();
306 assert(!same_velt_type(s, mem_ref), "sanity");
307 memops.push(s);
308 }
309 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
310 if (best_align_to_mem_ref == NULL) break;
311 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
312 // Restore list.
313 while (memops.size() > orig_msize)
314 (void)memops.pop();
315 }
316 } // unaligned memory accesses
317
318 // Remove used mem nodes.
319 for (int i = memops.size() - 1; i >= 0; i--) {
320 MemNode* m = memops.at(i)->as_Mem();
321 if (alignment(m) != top_align) {
322 memops.remove(i);
323 }
324 }
325
326 } // while (memops.size() != 0
327 set_align_to_ref(best_align_to_mem_ref);
328
329 #ifndef PRODUCT
330 if (TraceSuperWord) {
331 tty->print_cr("\nAfter find_adjacent_refs");
332 print_packset();
333 }
334 #endif
335 }
336
337 //------------------------------find_align_to_ref---------------------------
338 // Find a memory reference to align the loop induction variable to.
339 // Looks first at stores then at loads, looking for a memory reference
340 // with the largest number of references similar to it.
341 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
342 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
343
344 // Count number of comparable memory ops
345 for (uint i = 0; i < memops.size(); i++) {
346 MemNode* s1 = memops.at(i)->as_Mem();
347 SWPointer p1(s1, this);
348 // Discard if pre loop can't align this reference
349 if (!ref_is_alignable(p1)) {
350 *cmp_ct.adr_at(i) = 0;
351 continue;
352 }
353 for (uint j = i+1; j < memops.size(); j++) {
354 MemNode* s2 = memops.at(j)->as_Mem();
355 if (isomorphic(s1, s2)) {
356 SWPointer p2(s2, this);
357 if (p1.comparable(p2)) {
358 (*cmp_ct.adr_at(i))++;
359 (*cmp_ct.adr_at(j))++;
360 }
361 }
362 }
363 }
364
365 // Find Store (or Load) with the greatest number of "comparable" references,
366 // biggest vector size, smallest data size and smallest iv offset.
367 int max_ct = 0;
368 int max_vw = 0;
369 int max_idx = -1;
370 int min_size = max_jint;
371 int min_iv_offset = max_jint;
372 for (uint j = 0; j < memops.size(); j++) {
373 MemNode* s = memops.at(j)->as_Mem();
374 if (s->is_Store()) {
375 int vw = vector_width_in_bytes(s);
376 assert(vw > 1, "sanity");
377 SWPointer p(s, this);
378 if (cmp_ct.at(j) > max_ct ||
379 cmp_ct.at(j) == max_ct &&
380 (vw > max_vw ||
381 vw == max_vw &&
382 (data_size(s) < min_size ||
383 data_size(s) == min_size &&
384 (p.offset_in_bytes() < min_iv_offset)))) {
385 max_ct = cmp_ct.at(j);
386 max_vw = vw;
387 max_idx = j;
388 min_size = data_size(s);
389 min_iv_offset = p.offset_in_bytes();
390 }
391 }
392 }
393 // If no stores, look at loads
394 if (max_ct == 0) {
395 for (uint j = 0; j < memops.size(); j++) {
396 MemNode* s = memops.at(j)->as_Mem();
397 if (s->is_Load()) {
398 int vw = vector_width_in_bytes(s);
399 assert(vw > 1, "sanity");
400 SWPointer p(s, this);
401 if (cmp_ct.at(j) > max_ct ||
402 cmp_ct.at(j) == max_ct &&
403 (vw > max_vw ||
404 vw == max_vw &&
405 (data_size(s) < min_size ||
406 data_size(s) == min_size &&
407 (p.offset_in_bytes() < min_iv_offset)))) {
408 max_ct = cmp_ct.at(j);
409 max_vw = vw;
410 max_idx = j;
411 min_size = data_size(s);
412 min_iv_offset = p.offset_in_bytes();
413 }
414 }
415 }
416 }
417
418 #ifdef ASSERT
419 if (TraceSuperWord && Verbose) {
420 tty->print_cr("\nVector memops after find_align_to_refs");
421 for (uint i = 0; i < memops.size(); i++) {
422 MemNode* s = memops.at(i)->as_Mem();
423 s->dump();
424 }
425 }
426 #endif
427
428 if (max_ct > 0) {
429 #ifdef ASSERT
430 if (TraceSuperWord) {
431 tty->print("\nVector align to node: ");
432 memops.at(max_idx)->as_Mem()->dump();
433 }
434 #endif
435 return memops.at(max_idx)->as_Mem();
436 }
437 return NULL;
438 }
439
440 //------------------------------ref_is_alignable---------------------------
441 // Can the preloop align the reference to position zero in the vector?
442 bool SuperWord::ref_is_alignable(SWPointer& p) {
443 if (!p.has_iv()) {
444 return true; // no induction variable
445 }
446 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
447 assert(pre_end != NULL, "we must have a correct pre-loop");
448 assert(pre_end->stride_is_con(), "pre loop stride is constant");
449 int preloop_stride = pre_end->stride_con();
450
451 int span = preloop_stride * p.scale_in_bytes();
452
453 // Stride one accesses are alignable.
454 if (ABS(span) == p.memory_size())
455 return true;
456
457 // If initial offset from start of object is computable,
458 // compute alignment within the vector.
459 int vw = vector_width_in_bytes(p.mem());
460 assert(vw > 1, "sanity");
461 if (vw % span == 0) {
462 Node* init_nd = pre_end->init_trip();
463 if (init_nd->is_Con() && p.invar() == NULL) {
464 int init = init_nd->bottom_type()->is_int()->get_con();
465
466 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
467 assert(init_offset >= 0, "positive offset from object start");
468
469 if (span > 0) {
470 return (vw - (init_offset % vw)) % span == 0;
471 } else {
472 assert(span < 0, "nonzero stride * scale");
473 return (init_offset % vw) % -span == 0;
474 }
475 }
476 }
477 return false;
478 }
479
480 //---------------------------get_iv_adjustment---------------------------
481 // Calculate loop's iv adjustment for this memory ops.
482 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
483 SWPointer align_to_ref_p(mem_ref, this);
484 int offset = align_to_ref_p.offset_in_bytes();
485 int scale = align_to_ref_p.scale_in_bytes();
486 int vw = vector_width_in_bytes(mem_ref);
487 assert(vw > 1, "sanity");
488 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
489 // At least one iteration is executed in pre-loop by default. As result
490 // several iterations are needed to align memory operations in main-loop even
491 // if offset is 0.
492 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
493 int elt_size = align_to_ref_p.memory_size();
494 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
495 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
496 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
497
498 #ifndef PRODUCT
499 if (TraceSuperWord)
500 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
501 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
502 #endif
503 return iv_adjustment;
504 }
505
506 //---------------------------dependence_graph---------------------------
507 // Construct dependency graph.
508 // Add dependence edges to load/store nodes for memory dependence
509 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
510 void SuperWord::dependence_graph() {
511 // First, assign a dependence node to each memory node
512 for (int i = 0; i < _block.length(); i++ ) {
513 Node *n = _block.at(i);
514 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
515 _dg.make_node(n);
516 }
517 }
518
519 // For each memory slice, create the dependences
520 for (int i = 0; i < _mem_slice_head.length(); i++) {
521 Node* n = _mem_slice_head.at(i);
522 Node* n_tail = _mem_slice_tail.at(i);
523
524 // Get slice in predecessor order (last is first)
525 mem_slice_preds(n_tail, n, _nlist);
526
527 // Make the slice dependent on the root
528 DepMem* slice = _dg.dep(n);
529 _dg.make_edge(_dg.root(), slice);
530
531 // Create a sink for the slice
532 DepMem* slice_sink = _dg.make_node(NULL);
533 _dg.make_edge(slice_sink, _dg.tail());
534
535 // Now visit each pair of memory ops, creating the edges
536 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
537 Node* s1 = _nlist.at(j);
538
539 // If no dependency yet, use slice
540 if (_dg.dep(s1)->in_cnt() == 0) {
541 _dg.make_edge(slice, s1);
542 }
543 SWPointer p1(s1->as_Mem(), this);
544 bool sink_dependent = true;
545 for (int k = j - 1; k >= 0; k--) {
546 Node* s2 = _nlist.at(k);
547 if (s1->is_Load() && s2->is_Load())
548 continue;
549 SWPointer p2(s2->as_Mem(), this);
550
551 int cmp = p1.cmp(p2);
552 if (SuperWordRTDepCheck &&
553 p1.base() != p2.base() && p1.valid() && p2.valid()) {
554 // Create a runtime check to disambiguate
555 OrderedPair pp(p1.base(), p2.base());
556 _disjoint_ptrs.append_if_missing(pp);
557 } else if (!SWPointer::not_equal(cmp)) {
558 // Possibly same address
559 _dg.make_edge(s1, s2);
560 sink_dependent = false;
561 }
562 }
563 if (sink_dependent) {
564 _dg.make_edge(s1, slice_sink);
565 }
566 }
567 #ifndef PRODUCT
568 if (TraceSuperWord) {
569 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
570 for (int q = 0; q < _nlist.length(); q++) {
571 _dg.print(_nlist.at(q));
572 }
573 tty->cr();
574 }
575 #endif
576 _nlist.clear();
577 }
578
579 #ifndef PRODUCT
580 if (TraceSuperWord) {
581 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
582 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
583 _disjoint_ptrs.at(r).print();
584 tty->cr();
585 }
586 tty->cr();
587 }
588 #endif
589 }
590
591 //---------------------------mem_slice_preds---------------------------
592 // Return a memory slice (node list) in predecessor order starting at "start"
593 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
594 assert(preds.length() == 0, "start empty");
595 Node* n = start;
596 Node* prev = NULL;
597 while (true) {
598 assert(in_bb(n), "must be in block");
599 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
600 Node* out = n->fast_out(i);
601 if (out->is_Load()) {
602 if (in_bb(out)) {
603 preds.push(out);
604 }
605 } else {
606 // FIXME
607 if (out->is_MergeMem() && !in_bb(out)) {
608 // Either unrolling is causing a memory edge not to disappear,
609 // or need to run igvn.optimize() again before SLP
610 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
611 // Ditto. Not sure what else to check further.
612 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
613 // StoreCM has an input edge used as a precedence edge.
614 // Maybe an issue when oop stores are vectorized.
615 } else {
616 assert(out == prev || prev == NULL, "no branches off of store slice");
617 }
618 }
619 }
620 if (n == stop) break;
621 preds.push(n);
622 prev = n;
623 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
624 n = n->in(MemNode::Memory);
625 }
626 }
627
628 //------------------------------stmts_can_pack---------------------------
629 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
630 // s1 aligned at "align"
631 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
632
633 // Do not use superword for non-primitives
634 BasicType bt1 = velt_basic_type(s1);
635 BasicType bt2 = velt_basic_type(s2);
636 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
637 return false;
638 if (Matcher::max_vector_size(bt1) < 2) {
639 return false; // No vectors for this type
640 }
641
642 if (isomorphic(s1, s2)) {
643 if (independent(s1, s2) || reduction(s1, s2)) {
644 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
645 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
646 int s1_align = alignment(s1);
647 int s2_align = alignment(s2);
648 if (s1_align == top_align || s1_align == align) {
649 if (s2_align == top_align || s2_align == align + data_size(s1)) {
650 return true;
651 }
652 }
653 }
654 }
655 }
656 }
657 return false;
658 }
659
660 //------------------------------exists_at---------------------------
661 // Does s exist in a pack at position pos?
662 bool SuperWord::exists_at(Node* s, uint pos) {
663 for (int i = 0; i < _packset.length(); i++) {
664 Node_List* p = _packset.at(i);
665 if (p->at(pos) == s) {
666 return true;
667 }
668 }
669 return false;
670 }
671
672 //------------------------------are_adjacent_refs---------------------------
673 // Is s1 immediately before s2 in memory?
674 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
675 if (!s1->is_Mem() || !s2->is_Mem()) return false;
676 if (!in_bb(s1) || !in_bb(s2)) return false;
677
678 // Do not use superword for non-primitives
679 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
680 !is_java_primitive(s2->as_Mem()->memory_type())) {
681 return false;
682 }
683
684 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
685 // only pack memops that are in the same alias set until that's fixed.
686 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
687 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
688 return false;
689 SWPointer p1(s1->as_Mem(), this);
690 SWPointer p2(s2->as_Mem(), this);
691 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
692 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
693 return diff == data_size(s1);
694 }
695
696 //------------------------------isomorphic---------------------------
697 // Are s1 and s2 similar?
698 bool SuperWord::isomorphic(Node* s1, Node* s2) {
699 if (s1->Opcode() != s2->Opcode()) return false;
700 if (s1->req() != s2->req()) return false;
701 if (s1->in(0) != s2->in(0)) return false;
702 if (!same_velt_type(s1, s2)) return false;
703 return true;
704 }
705
706 //------------------------------independent---------------------------
707 // Is there no data path from s1 to s2 or s2 to s1?
708 bool SuperWord::independent(Node* s1, Node* s2) {
709 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
710 int d1 = depth(s1);
711 int d2 = depth(s2);
712 if (d1 == d2) return s1 != s2;
713 Node* deep = d1 > d2 ? s1 : s2;
714 Node* shallow = d1 > d2 ? s2 : s1;
715
716 visited_clear();
717
718 return independent_path(shallow, deep);
719 }
720
721 //------------------------------reduction---------------------------
722 // Is there a data path between s1 and s2 and the nodes reductions?
723 bool SuperWord::reduction(Node* s1, Node* s2) {
724 bool retValue = false;
725 int d1 = depth(s1);
726 int d2 = depth(s2);
727 if (d1 + 1 == d2) {
728 if (s1->is_reduction() && s2->is_reduction()) {
729 // This is an ordered set, so s1 should define s2
730 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
731 Node* t1 = s1->fast_out(i);
732 if (t1 == s2) {
733 // both nodes are reductions and connected
734 retValue = true;
735 }
736 }
737 }
738 }
739
740 return retValue;
741 }
742
743 //------------------------------independent_path------------------------------
744 // Helper for independent
745 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
746 if (dp >= 1000) return false; // stop deep recursion
747 visited_set(deep);
748 int shal_depth = depth(shallow);
749 assert(shal_depth <= depth(deep), "must be");
750 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
751 Node* pred = preds.current();
752 if (in_bb(pred) && !visited_test(pred)) {
753 if (shallow == pred) {
754 return false;
755 }
756 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
757 return false;
758 }
759 }
760 }
761 return true;
762 }
763
764 //------------------------------set_alignment---------------------------
765 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
766 set_alignment(s1, align);
767 if (align == top_align || align == bottom_align) {
768 set_alignment(s2, align);
769 } else {
770 set_alignment(s2, align + data_size(s1));
771 }
772 }
773
774 //------------------------------data_size---------------------------
775 int SuperWord::data_size(Node* s) {
776 int bsize = type2aelembytes(velt_basic_type(s));
777 assert(bsize != 0, "valid size");
778 return bsize;
779 }
780
781 //------------------------------extend_packlist---------------------------
782 // Extend packset by following use->def and def->use links from pack members.
783 void SuperWord::extend_packlist() {
784 bool changed;
785 do {
786 packset_sort(_packset.length());
787 changed = false;
788 for (int i = 0; i < _packset.length(); i++) {
789 Node_List* p = _packset.at(i);
790 changed |= follow_use_defs(p);
791 changed |= follow_def_uses(p);
792 }
793 } while (changed);
794
795 if (_race_possible) {
796 for (int i = 0; i < _packset.length(); i++) {
797 Node_List* p = _packset.at(i);
798 order_def_uses(p);
799 }
800 }
801
802 #ifndef PRODUCT
803 if (TraceSuperWord) {
804 tty->print_cr("\nAfter extend_packlist");
805 print_packset();
806 }
807 #endif
808 }
809
810 //------------------------------follow_use_defs---------------------------
811 // Extend the packset by visiting operand definitions of nodes in pack p
812 bool SuperWord::follow_use_defs(Node_List* p) {
813 assert(p->size() == 2, "just checking");
814 Node* s1 = p->at(0);
815 Node* s2 = p->at(1);
816 assert(s1->req() == s2->req(), "just checking");
817 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
818
819 if (s1->is_Load()) return false;
820
821 int align = alignment(s1);
822 bool changed = false;
823 int start = s1->is_Store() ? MemNode::ValueIn : 1;
824 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
825 for (int j = start; j < end; j++) {
826 Node* t1 = s1->in(j);
827 Node* t2 = s2->in(j);
828 if (!in_bb(t1) || !in_bb(t2))
829 continue;
830 if (stmts_can_pack(t1, t2, align)) {
831 if (est_savings(t1, t2) >= 0) {
832 Node_List* pair = new Node_List();
833 pair->push(t1);
834 pair->push(t2);
835 _packset.append(pair);
836 set_alignment(t1, t2, align);
837 changed = true;
838 }
839 }
840 }
841 return changed;
842 }
843
844 //------------------------------follow_def_uses---------------------------
845 // Extend the packset by visiting uses of nodes in pack p
846 bool SuperWord::follow_def_uses(Node_List* p) {
847 bool changed = false;
848 Node* s1 = p->at(0);
849 Node* s2 = p->at(1);
850 assert(p->size() == 2, "just checking");
851 assert(s1->req() == s2->req(), "just checking");
852 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
853
854 if (s1->is_Store()) return false;
855
856 int align = alignment(s1);
857 int savings = -1;
858 int num_s1_uses = 0;
859 Node* u1 = NULL;
860 Node* u2 = NULL;
861 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
862 Node* t1 = s1->fast_out(i);
863 num_s1_uses++;
864 if (!in_bb(t1)) continue;
865 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
866 Node* t2 = s2->fast_out(j);
867 if (!in_bb(t2)) continue;
868 if (!opnd_positions_match(s1, t1, s2, t2))
869 continue;
870 if (stmts_can_pack(t1, t2, align)) {
871 int my_savings = est_savings(t1, t2);
872 if (my_savings > savings) {
873 savings = my_savings;
874 u1 = t1;
875 u2 = t2;
876 }
877 }
878 }
879 }
880 if (num_s1_uses > 1) {
881 _race_possible = true;
882 }
883 if (savings >= 0) {
884 Node_List* pair = new Node_List();
885 pair->push(u1);
886 pair->push(u2);
887 _packset.append(pair);
888 set_alignment(u1, u2, align);
889 changed = true;
890 }
891 return changed;
892 }
893
894 //------------------------------order_def_uses---------------------------
895 // For extended packsets, ordinally arrange uses packset by major component
896 void SuperWord::order_def_uses(Node_List* p) {
897 Node* s1 = p->at(0);
898
899 if (s1->is_Store()) return;
900
901 // reductions are always managed beforehand
902 if (s1->is_reduction()) return;
903
904 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
905 Node* t1 = s1->fast_out(i);
906
907 // Only allow operand swap on commuting operations
908 if (!t1->is_Add() && !t1->is_Mul()) {
909 break;
910 }
911
912 // Now find t1's packset
913 Node_List* p2 = NULL;
914 for (int j = 0; j < _packset.length(); j++) {
915 p2 = _packset.at(j);
916 Node* first = p2->at(0);
917 if (t1 == first) {
918 break;
919 }
920 p2 = NULL;
921 }
922 // Arrange all sub components by the major component
923 if (p2 != NULL) {
924 for (uint j = 1; j < p->size(); j++) {
925 Node* d1 = p->at(j);
926 Node* u1 = p2->at(j);
927 opnd_positions_match(s1, t1, d1, u1);
928 }
929 }
930 }
931 }
932
933 //---------------------------opnd_positions_match-------------------------
934 // Is the use of d1 in u1 at the same operand position as d2 in u2?
935 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
936 // check reductions to see if they are marshalled to represent the reduction
937 // operator in a specified opnd
938 if (u1->is_reduction() && u2->is_reduction()) {
939 // ensure reductions have phis and reduction definitions feeding the 1st operand
940 Node* first = u1->in(2);
941 if (first->is_Phi() || first->is_reduction()) {
942 u1->swap_edges(1, 2);
943 }
944 // ensure reductions have phis and reduction definitions feeding the 1st operand
945 first = u2->in(2);
946 if (first->is_Phi() || first->is_reduction()) {
947 u2->swap_edges(1, 2);
948 }
949 return true;
950 }
951
952 uint ct = u1->req();
953 if (ct != u2->req()) return false;
954 uint i1 = 0;
955 uint i2 = 0;
956 do {
957 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
958 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
959 if (i1 != i2) {
960 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
961 // Further analysis relies on operands position matching.
962 u2->swap_edges(i1, i2);
963 } else {
964 return false;
965 }
966 }
967 } while (i1 < ct);
968 return true;
969 }
970
971 //------------------------------est_savings---------------------------
972 // Estimate the savings from executing s1 and s2 as a pack
973 int SuperWord::est_savings(Node* s1, Node* s2) {
974 int save_in = 2 - 1; // 2 operations per instruction in packed form
975
976 // inputs
977 for (uint i = 1; i < s1->req(); i++) {
978 Node* x1 = s1->in(i);
979 Node* x2 = s2->in(i);
980 if (x1 != x2) {
981 if (are_adjacent_refs(x1, x2)) {
982 save_in += adjacent_profit(x1, x2);
983 } else if (!in_packset(x1, x2)) {
984 save_in -= pack_cost(2);
985 } else {
986 save_in += unpack_cost(2);
987 }
988 }
989 }
990
991 // uses of result
992 uint ct = 0;
993 int save_use = 0;
994 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
995 Node* s1_use = s1->fast_out(i);
996 for (int j = 0; j < _packset.length(); j++) {
997 Node_List* p = _packset.at(j);
998 if (p->at(0) == s1_use) {
999 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1000 Node* s2_use = s2->fast_out(k);
1001 if (p->at(p->size()-1) == s2_use) {
1002 ct++;
1003 if (are_adjacent_refs(s1_use, s2_use)) {
1004 save_use += adjacent_profit(s1_use, s2_use);
1005 }
1006 }
1007 }
1008 }
1009 }
1010 }
1011
1012 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1013 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1014
1015 return MAX2(save_in, save_use);
1016 }
1017
1018 //------------------------------costs---------------------------
1019 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1020 int SuperWord::pack_cost(int ct) { return ct; }
1021 int SuperWord::unpack_cost(int ct) { return ct; }
1022
1023 //------------------------------combine_packs---------------------------
1024 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1025 void SuperWord::combine_packs() {
1026 bool changed = true;
1027 // Combine packs regardless max vector size.
1028 while (changed) {
1029 changed = false;
1030 for (int i = 0; i < _packset.length(); i++) {
1031 Node_List* p1 = _packset.at(i);
1032 if (p1 == NULL) continue;
1033 // Because of sorting we can start at i + 1
1034 for (int j = i + 1; j < _packset.length(); j++) {
1035 Node_List* p2 = _packset.at(j);
1036 if (p2 == NULL) continue;
1037 if (i == j) continue;
1038 if (p1->at(p1->size()-1) == p2->at(0)) {
1039 for (uint k = 1; k < p2->size(); k++) {
1040 p1->push(p2->at(k));
1041 }
1042 _packset.at_put(j, NULL);
1043 changed = true;
1044 }
1045 }
1046 }
1047 }
1048
1049 // Split packs which have size greater then max vector size.
1050 for (int i = 0; i < _packset.length(); i++) {
1051 Node_List* p1 = _packset.at(i);
1052 if (p1 != NULL) {
1053 BasicType bt = velt_basic_type(p1->at(0));
1054 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1055 assert(is_power_of_2(max_vlen), "sanity");
1056 uint psize = p1->size();
1057 if (!is_power_of_2(psize)) {
1058 // Skip pack which can't be vector.
1059 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1060 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1061 _packset.at_put(i, NULL);
1062 continue;
1063 }
1064 if (psize > max_vlen) {
1065 Node_List* pack = new Node_List();
1066 for (uint j = 0; j < psize; j++) {
1067 pack->push(p1->at(j));
1068 if (pack->size() >= max_vlen) {
1069 assert(is_power_of_2(pack->size()), "sanity");
1070 _packset.append(pack);
1071 pack = new Node_List();
1072 }
1073 }
1074 _packset.at_put(i, NULL);
1075 }
1076 }
1077 }
1078
1079 // Compress list.
1080 for (int i = _packset.length() - 1; i >= 0; i--) {
1081 Node_List* p1 = _packset.at(i);
1082 if (p1 == NULL) {
1083 _packset.remove_at(i);
1084 }
1085 }
1086
1087 #ifndef PRODUCT
1088 if (TraceSuperWord) {
1089 tty->print_cr("\nAfter combine_packs");
1090 print_packset();
1091 }
1092 #endif
1093 }
1094
1095 //-----------------------------construct_my_pack_map--------------------------
1096 // Construct the map from nodes to packs. Only valid after the
1097 // point where a node is only in one pack (after combine_packs).
1098 void SuperWord::construct_my_pack_map() {
1099 Node_List* rslt = NULL;
1100 for (int i = 0; i < _packset.length(); i++) {
1101 Node_List* p = _packset.at(i);
1102 for (uint j = 0; j < p->size(); j++) {
1103 Node* s = p->at(j);
1104 assert(my_pack(s) == NULL, "only in one pack");
1105 set_my_pack(s, p);
1106 }
1107 }
1108 }
1109
1110 //------------------------------filter_packs---------------------------
1111 // Remove packs that are not implemented or not profitable.
1112 void SuperWord::filter_packs() {
1113
1114 // Remove packs that are not implemented
1115 for (int i = _packset.length() - 1; i >= 0; i--) {
1116 Node_List* pk = _packset.at(i);
1117 bool impl = implemented(pk);
1118 if (!impl) {
1119 #ifndef PRODUCT
1120 if (TraceSuperWord && Verbose) {
1121 tty->print_cr("Unimplemented");
1122 pk->at(0)->dump();
1123 }
1124 #endif
1125 remove_pack_at(i);
1126 }
1127 }
1128
1129 // Remove packs that are not profitable
1130 bool changed;
1131 do {
1132 changed = false;
1133 for (int i = _packset.length() - 1; i >= 0; i--) {
1134 Node_List* pk = _packset.at(i);
1135 bool prof = profitable(pk);
1136 if (!prof) {
1137 #ifndef PRODUCT
1138 if (TraceSuperWord && Verbose) {
1139 tty->print_cr("Unprofitable");
1140 pk->at(0)->dump();
1141 }
1142 #endif
1143 remove_pack_at(i);
1144 changed = true;
1145 }
1146 }
1147 } while (changed);
1148
1149 #ifndef PRODUCT
1150 if (TraceSuperWord) {
1151 tty->print_cr("\nAfter filter_packs");
1152 print_packset();
1153 tty->cr();
1154 }
1155 #endif
1156 }
1157
1158 //------------------------------implemented---------------------------
1159 // Can code be generated for pack p?
1160 bool SuperWord::implemented(Node_List* p) {
1161 bool retValue = false;
1162 Node* p0 = p->at(0);
1163 if (p0 != NULL) {
1164 int opc = p0->Opcode();
1165 uint size = p->size();
1166 if (p0->is_reduction()) {
1167 const Type *arith_type = p0->bottom_type();
1168 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1169 } else {
1170 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1171 }
1172 }
1173 return retValue;
1174 }
1175
1176 //------------------------------same_inputs--------------------------
1177 // For pack p, are all idx operands the same?
1178 static bool same_inputs(Node_List* p, int idx) {
1179 Node* p0 = p->at(0);
1180 uint vlen = p->size();
1181 Node* p0_def = p0->in(idx);
1182 for (uint i = 1; i < vlen; i++) {
1183 Node* pi = p->at(i);
1184 Node* pi_def = pi->in(idx);
1185 if (p0_def != pi_def)
1186 return false;
1187 }
1188 return true;
1189 }
1190
1191 //------------------------------profitable---------------------------
1192 // For pack p, are all operands and all uses (with in the block) vector?
1193 bool SuperWord::profitable(Node_List* p) {
1194 Node* p0 = p->at(0);
1195 uint start, end;
1196 VectorNode::vector_operands(p0, &start, &end);
1197
1198 // Return false if some inputs are not vectors or vectors with different
1199 // size or alignment.
1200 // Also, for now, return false if not scalar promotion case when inputs are
1201 // the same. Later, implement PackNode and allow differing, non-vector inputs
1202 // (maybe just the ones from outside the block.)
1203 for (uint i = start; i < end; i++) {
1204 if (!is_vector_use(p0, i))
1205 return false;
1206 }
1207 // Check if reductions are connected
1208 if (p0->is_reduction()) {
1209 Node* second_in = p0->in(2);
1210 Node_List* second_pk = my_pack(second_in);
1211 if (second_pk == NULL) {
1212 // Remove reduction flag if no parent pack, it is not profitable
1213 p0->remove_flag(Node::Flag_is_reduction);
1214 return false;
1215 } else if (second_pk->size() != p->size()) {
1216 return false;
1217 }
1218 }
1219 if (VectorNode::is_shift(p0)) {
1220 // For now, return false if shift count is vector or not scalar promotion
1221 // case (different shift counts) because it is not supported yet.
1222 Node* cnt = p0->in(2);
1223 Node_List* cnt_pk = my_pack(cnt);
1224 if (cnt_pk != NULL)
1225 return false;
1226 if (!same_inputs(p, 2))
1227 return false;
1228 }
1229 if (!p0->is_Store()) {
1230 // For now, return false if not all uses are vector.
1231 // Later, implement ExtractNode and allow non-vector uses (maybe
1232 // just the ones outside the block.)
1233 for (uint i = 0; i < p->size(); i++) {
1234 Node* def = p->at(i);
1235 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1236 Node* use = def->fast_out(j);
1237 for (uint k = 0; k < use->req(); k++) {
1238 Node* n = use->in(k);
1239 if (def == n) {
1240 // reductions can be loop carried dependences
1241 if (def->is_reduction() && use->is_Phi())
1242 continue;
1243 if (!is_vector_use(use, k)) {
1244 return false;
1245 }
1246 }
1247 }
1248 }
1249 }
1250 }
1251 return true;
1252 }
1253
1254 //------------------------------schedule---------------------------
1255 // Adjust the memory graph for the packed operations
1256 void SuperWord::schedule() {
1257
1258 // Co-locate in the memory graph the members of each memory pack
1259 for (int i = 0; i < _packset.length(); i++) {
1260 co_locate_pack(_packset.at(i));
1261 }
1262 }
1263
1264 //-------------------------------remove_and_insert-------------------
1265 // Remove "current" from its current position in the memory graph and insert
1266 // it after the appropriate insertion point (lip or uip).
1267 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1268 Node *uip, Unique_Node_List &sched_before) {
1269 Node* my_mem = current->in(MemNode::Memory);
1270 bool sched_up = sched_before.member(current);
1271
1272 // remove current_store from its current position in the memmory graph
1273 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1274 Node* use = current->out(i);
1275 if (use->is_Mem()) {
1276 assert(use->in(MemNode::Memory) == current, "must be");
1277 if (use == prev) { // connect prev to my_mem
1278 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1279 --i; //deleted this edge; rescan position
1280 } else if (sched_before.member(use)) {
1281 if (!sched_up) { // Will be moved together with current
1282 _igvn.replace_input_of(use, MemNode::Memory, uip);
1283 --i; //deleted this edge; rescan position
1284 }
1285 } else {
1286 if (sched_up) { // Will be moved together with current
1287 _igvn.replace_input_of(use, MemNode::Memory, lip);
1288 --i; //deleted this edge; rescan position
1289 }
1290 }
1291 }
1292 }
1293
1294 Node *insert_pt = sched_up ? uip : lip;
1295
1296 // all uses of insert_pt's memory state should use current's instead
1297 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1298 Node* use = insert_pt->out(i);
1299 if (use->is_Mem()) {
1300 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1301 _igvn.replace_input_of(use, MemNode::Memory, current);
1302 --i; //deleted this edge; rescan position
1303 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1304 uint pos; //lip (lower insert point) must be the last one in the memory slice
1305 for (pos=1; pos < use->req(); pos++) {
1306 if (use->in(pos) == insert_pt) break;
1307 }
1308 _igvn.replace_input_of(use, pos, current);
1309 --i;
1310 }
1311 }
1312
1313 //connect current to insert_pt
1314 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1315 }
1316
1317 //------------------------------co_locate_pack----------------------------------
1318 // To schedule a store pack, we need to move any sandwiched memory ops either before
1319 // or after the pack, based upon dependence information:
1320 // (1) If any store in the pack depends on the sandwiched memory op, the
1321 // sandwiched memory op must be scheduled BEFORE the pack;
1322 // (2) If a sandwiched memory op depends on any store in the pack, the
1323 // sandwiched memory op must be scheduled AFTER the pack;
1324 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1325 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1326 // scheduled before the pack, memB must also be scheduled before the pack;
1327 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1328 // schedule this store AFTER the pack
1329 // (5) We know there is no dependence cycle, so there in no other case;
1330 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1331 //
1332 // To schedule a load pack, we use the memory state of either the first or the last load in
1333 // the pack, based on the dependence constraint.
1334 void SuperWord::co_locate_pack(Node_List* pk) {
1335 if (pk->at(0)->is_Store()) {
1336 MemNode* first = executed_first(pk)->as_Mem();
1337 MemNode* last = executed_last(pk)->as_Mem();
1338 Unique_Node_List schedule_before_pack;
1339 Unique_Node_List memops;
1340
1341 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1342 MemNode* previous = last;
1343 while (true) {
1344 assert(in_bb(current), "stay in block");
1345 memops.push(previous);
1346 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1347 Node* use = current->out(i);
1348 if (use->is_Mem() && use != previous)
1349 memops.push(use);
1350 }
1351 if (current == first) break;
1352 previous = current;
1353 current = current->in(MemNode::Memory)->as_Mem();
1354 }
1355
1356 // determine which memory operations should be scheduled before the pack
1357 for (uint i = 1; i < memops.size(); i++) {
1358 Node *s1 = memops.at(i);
1359 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1360 for (uint j = 0; j< i; j++) {
1361 Node *s2 = memops.at(j);
1362 if (!independent(s1, s2)) {
1363 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1364 schedule_before_pack.push(s1); // s1 must be scheduled before
1365 Node_List* mem_pk = my_pack(s1);
1366 if (mem_pk != NULL) {
1367 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1368 Node* s = mem_pk->at(ii); // follow partner
1369 if (memops.member(s) && !schedule_before_pack.member(s))
1370 schedule_before_pack.push(s);
1371 }
1372 }
1373 break;
1374 }
1375 }
1376 }
1377 }
1378 }
1379
1380 Node* upper_insert_pt = first->in(MemNode::Memory);
1381 // Following code moves loads connected to upper_insert_pt below aliased stores.
1382 // Collect such loads here and reconnect them back to upper_insert_pt later.
1383 memops.clear();
1384 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1385 Node* use = upper_insert_pt->out(i);
1386 if (use->is_Mem() && !use->is_Store()) {
1387 memops.push(use);
1388 }
1389 }
1390
1391 MemNode* lower_insert_pt = last;
1392 previous = last; //previous store in pk
1393 current = last->in(MemNode::Memory)->as_Mem();
1394
1395 // start scheduling from "last" to "first"
1396 while (true) {
1397 assert(in_bb(current), "stay in block");
1398 assert(in_pack(previous, pk), "previous stays in pack");
1399 Node* my_mem = current->in(MemNode::Memory);
1400
1401 if (in_pack(current, pk)) {
1402 // Forward users of my memory state (except "previous) to my input memory state
1403 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1404 Node* use = current->out(i);
1405 if (use->is_Mem() && use != previous) {
1406 assert(use->in(MemNode::Memory) == current, "must be");
1407 if (schedule_before_pack.member(use)) {
1408 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1409 } else {
1410 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1411 }
1412 --i; // deleted this edge; rescan position
1413 }
1414 }
1415 previous = current;
1416 } else { // !in_pack(current, pk) ==> a sandwiched store
1417 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1418 }
1419
1420 if (current == first) break;
1421 current = my_mem->as_Mem();
1422 } // end while
1423
1424 // Reconnect loads back to upper_insert_pt.
1425 for (uint i = 0; i < memops.size(); i++) {
1426 Node *ld = memops.at(i);
1427 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1428 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1429 }
1430 }
1431 } else if (pk->at(0)->is_Load()) { //load
1432 // all loads in the pack should have the same memory state. By default,
1433 // we use the memory state of the last load. However, if any load could
1434 // not be moved down due to the dependence constraint, we use the memory
1435 // state of the first load.
1436 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1437 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1438 bool schedule_last = true;
1439 for (uint i = 0; i < pk->size(); i++) {
1440 Node* ld = pk->at(i);
1441 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1442 current=current->in(MemNode::Memory)) {
1443 assert(current != first_mem, "corrupted memory graph");
1444 if(current->is_Mem() && !independent(current, ld)){
1445 schedule_last = false; // a later store depends on this load
1446 break;
1447 }
1448 }
1449 }
1450
1451 Node* mem_input = schedule_last ? last_mem : first_mem;
1452 _igvn.hash_delete(mem_input);
1453 // Give each load the same memory state
1454 for (uint i = 0; i < pk->size(); i++) {
1455 LoadNode* ld = pk->at(i)->as_Load();
1456 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1457 }
1458 }
1459 }
1460
1461 //------------------------------output---------------------------
1462 // Convert packs into vector node operations
1463 void SuperWord::output() {
1464 if (_packset.length() == 0) return;
1465
1466 #ifndef PRODUCT
1467 if (TraceLoopOpts) {
1468 tty->print("SuperWord ");
1469 lpt()->dump_head();
1470 }
1471 #endif
1472
1473 // MUST ENSURE main loop's initial value is properly aligned:
1474 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1475
1476 align_initial_loop_index(align_to_ref());
1477
1478 // Insert extract (unpack) operations for scalar uses
1479 for (int i = 0; i < _packset.length(); i++) {
1480 insert_extracts(_packset.at(i));
1481 }
1482
1483 Compile* C = _phase->C;
1484 uint max_vlen_in_bytes = 0;
1485 for (int i = 0; i < _block.length(); i++) {
1486 Node* n = _block.at(i);
1487 Node_List* p = my_pack(n);
1488 if (p && n == executed_last(p)) {
1489 uint vlen = p->size();
1490 uint vlen_in_bytes = 0;
1491 Node* vn = NULL;
1492 Node* low_adr = p->at(0);
1493 Node* first = executed_first(p);
1494 int opc = n->Opcode();
1495 if (n->is_Load()) {
1496 Node* ctl = n->in(MemNode::Control);
1497 Node* mem = first->in(MemNode::Memory);
1498 SWPointer p1(n->as_Mem(), this);
1499 // Identify the memory dependency for the new loadVector node by
1500 // walking up through memory chain.
1501 // This is done to give flexibility to the new loadVector node so that
1502 // it can move above independent storeVector nodes.
1503 while (mem->is_StoreVector()) {
1504 SWPointer p2(mem->as_Mem(), this);
1505 int cmp = p1.cmp(p2);
1506 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1507 mem = mem->in(MemNode::Memory);
1508 } else {
1509 break; // dependent memory
1510 }
1511 }
1512 Node* adr = low_adr->in(MemNode::Address);
1513 const TypePtr* atyp = n->adr_type();
1514 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1515 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1516 } else if (n->is_Store()) {
1517 // Promote value to be stored to vector
1518 Node* val = vector_opd(p, MemNode::ValueIn);
1519 Node* ctl = n->in(MemNode::Control);
1520 Node* mem = first->in(MemNode::Memory);
1521 Node* adr = low_adr->in(MemNode::Address);
1522 const TypePtr* atyp = n->adr_type();
1523 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1524 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1525 } else if (n->req() == 3) {
1526 // Promote operands to vector
1527 Node* in1 = NULL;
1528 bool node_isa_reduction = n->is_reduction();
1529 if (node_isa_reduction) {
1530 // the input to the first reduction operation is retained
1531 in1 = low_adr->in(1);
1532 } else {
1533 in1 = vector_opd(p, 1);
1534 }
1535 Node* in2 = vector_opd(p, 2);
1536 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1537 // Move invariant vector input into second position to avoid register spilling.
1538 Node* tmp = in1;
1539 in1 = in2;
1540 in2 = tmp;
1541 }
1542 if (node_isa_reduction) {
1543 const Type *arith_type = n->bottom_type();
1544 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1545 if (in2->is_Load()) {
1546 vlen_in_bytes = in2->as_LoadVector()->memory_size();
1547 } else {
1548 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1549 }
1550 } else {
1551 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1552 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1553 }
1554 } else {
1555 ShouldNotReachHere();
1556 }
1557 assert(vn != NULL, "sanity");
1558 _igvn.register_new_node_with_optimizer(vn);
1559 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1560 for (uint j = 0; j < p->size(); j++) {
1561 Node* pm = p->at(j);
1562 _igvn.replace_node(pm, vn);
1563 }
1564 _igvn._worklist.push(vn);
1565
1566 if (vlen_in_bytes > max_vlen_in_bytes) {
1567 max_vlen_in_bytes = vlen_in_bytes;
1568 }
1569 #ifdef ASSERT
1570 if (TraceNewVectors) {
1571 tty->print("new Vector node: ");
1572 vn->dump();
1573 }
1574 #endif
1575 }
1576 }
1577 C->set_max_vector_size(max_vlen_in_bytes);
1578 }
1579
1580 //------------------------------vector_opd---------------------------
1581 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1582 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1583 Node* p0 = p->at(0);
1584 uint vlen = p->size();
1585 Node* opd = p0->in(opd_idx);
1586
1587 if (same_inputs(p, opd_idx)) {
1588 if (opd->is_Vector() || opd->is_LoadVector()) {
1589 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1590 return opd; // input is matching vector
1591 }
1592 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1593 Compile* C = _phase->C;
1594 Node* cnt = opd;
1595 // Vector instructions do not mask shift count, do it here.
1596 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1597 const TypeInt* t = opd->find_int_type();
1598 if (t != NULL && t->is_con()) {
1599 juint shift = t->get_con();
1600 if (shift > mask) { // Unsigned cmp
1601 cnt = ConNode::make(TypeInt::make(shift & mask));
1602 }
1603 } else {
1604 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1605 cnt = ConNode::make(TypeInt::make(mask));
1606 _igvn.register_new_node_with_optimizer(cnt);
1607 cnt = new AndINode(opd, cnt);
1608 _igvn.register_new_node_with_optimizer(cnt);
1609 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1610 }
1611 assert(opd->bottom_type()->isa_int(), "int type only");
1612 // Move non constant shift count into vector register.
1613 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1614 }
1615 if (cnt != opd) {
1616 _igvn.register_new_node_with_optimizer(cnt);
1617 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1618 }
1619 return cnt;
1620 }
1621 assert(!opd->is_StoreVector(), "such vector is not expected here");
1622 // Convert scalar input to vector with the same number of elements as
1623 // p0's vector. Use p0's type because size of operand's container in
1624 // vector should match p0's size regardless operand's size.
1625 const Type* p0_t = velt_type(p0);
1626 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1627
1628 _igvn.register_new_node_with_optimizer(vn);
1629 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1630 #ifdef ASSERT
1631 if (TraceNewVectors) {
1632 tty->print("new Vector node: ");
1633 vn->dump();
1634 }
1635 #endif
1636 return vn;
1637 }
1638
1639 // Insert pack operation
1640 BasicType bt = velt_basic_type(p0);
1641 PackNode* pk = PackNode::make(opd, vlen, bt);
1642 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1643
1644 for (uint i = 1; i < vlen; i++) {
1645 Node* pi = p->at(i);
1646 Node* in = pi->in(opd_idx);
1647 assert(my_pack(in) == NULL, "Should already have been unpacked");
1648 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1649 pk->add_opd(in);
1650 }
1651 _igvn.register_new_node_with_optimizer(pk);
1652 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1653 #ifdef ASSERT
1654 if (TraceNewVectors) {
1655 tty->print("new Vector node: ");
1656 pk->dump();
1657 }
1658 #endif
1659 return pk;
1660 }
1661
1662 //------------------------------insert_extracts---------------------------
1663 // If a use of pack p is not a vector use, then replace the
1664 // use with an extract operation.
1665 void SuperWord::insert_extracts(Node_List* p) {
1666 if (p->at(0)->is_Store()) return;
1667 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1668
1669 // Inspect each use of each pack member. For each use that is
1670 // not a vector use, replace the use with an extract operation.
1671
1672 for (uint i = 0; i < p->size(); i++) {
1673 Node* def = p->at(i);
1674 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1675 Node* use = def->fast_out(j);
1676 for (uint k = 0; k < use->req(); k++) {
1677 Node* n = use->in(k);
1678 if (def == n) {
1679 if (!is_vector_use(use, k)) {
1680 _n_idx_list.push(use, k);
1681 }
1682 }
1683 }
1684 }
1685 }
1686
1687 while (_n_idx_list.is_nonempty()) {
1688 Node* use = _n_idx_list.node();
1689 int idx = _n_idx_list.index();
1690 _n_idx_list.pop();
1691 Node* def = use->in(idx);
1692
1693 if (def->is_reduction()) continue;
1694
1695 // Insert extract operation
1696 _igvn.hash_delete(def);
1697 int def_pos = alignment(def) / data_size(def);
1698
1699 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1700 _igvn.register_new_node_with_optimizer(ex);
1701 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1702 _igvn.replace_input_of(use, idx, ex);
1703 _igvn._worklist.push(def);
1704
1705 bb_insert_after(ex, bb_idx(def));
1706 set_velt_type(ex, velt_type(def));
1707 }
1708 }
1709
1710 //------------------------------is_vector_use---------------------------
1711 // Is use->in(u_idx) a vector use?
1712 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1713 Node_List* u_pk = my_pack(use);
1714 if (u_pk == NULL) return false;
1715 if (use->is_reduction()) return true;
1716 Node* def = use->in(u_idx);
1717 Node_List* d_pk = my_pack(def);
1718 if (d_pk == NULL) {
1719 // check for scalar promotion
1720 Node* n = u_pk->at(0)->in(u_idx);
1721 for (uint i = 1; i < u_pk->size(); i++) {
1722 if (u_pk->at(i)->in(u_idx) != n) return false;
1723 }
1724 return true;
1725 }
1726 if (u_pk->size() != d_pk->size())
1727 return false;
1728 for (uint i = 0; i < u_pk->size(); i++) {
1729 Node* ui = u_pk->at(i);
1730 Node* di = d_pk->at(i);
1731 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1732 return false;
1733 }
1734 return true;
1735 }
1736
1737 //------------------------------construct_bb---------------------------
1738 // Construct reverse postorder list of block members
1739 bool SuperWord::construct_bb() {
1740 Node* entry = bb();
1741
1742 assert(_stk.length() == 0, "stk is empty");
1743 assert(_block.length() == 0, "block is empty");
1744 assert(_data_entry.length() == 0, "data_entry is empty");
1745 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1746 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1747
1748 // Find non-control nodes with no inputs from within block,
1749 // create a temporary map from node _idx to bb_idx for use
1750 // by the visited and post_visited sets,
1751 // and count number of nodes in block.
1752 int bb_ct = 0;
1753 for (uint i = 0; i < lpt()->_body.size(); i++) {
1754 Node *n = lpt()->_body.at(i);
1755 set_bb_idx(n, i); // Create a temporary map
1756 if (in_bb(n)) {
1757 if (n->is_LoadStore() || n->is_MergeMem() ||
1758 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1759 // Bailout if the loop has LoadStore, MergeMem or data Proj
1760 // nodes. Superword optimization does not work with them.
1761 return false;
1762 }
1763 bb_ct++;
1764 if (!n->is_CFG()) {
1765 bool found = false;
1766 for (uint j = 0; j < n->req(); j++) {
1767 Node* def = n->in(j);
1768 if (def && in_bb(def)) {
1769 found = true;
1770 break;
1771 }
1772 }
1773 if (!found) {
1774 assert(n != entry, "can't be entry");
1775 _data_entry.push(n);
1776 }
1777 }
1778 }
1779 }
1780
1781 // Find memory slices (head and tail)
1782 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1783 Node *n = lp()->fast_out(i);
1784 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1785 Node* n_tail = n->in(LoopNode::LoopBackControl);
1786 if (n_tail != n->in(LoopNode::EntryControl)) {
1787 if (!n_tail->is_Mem()) {
1788 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1789 return false; // Bailout
1790 }
1791 _mem_slice_head.push(n);
1792 _mem_slice_tail.push(n_tail);
1793 }
1794 }
1795 }
1796
1797 // Create an RPO list of nodes in block
1798
1799 visited_clear();
1800 post_visited_clear();
1801
1802 // Push all non-control nodes with no inputs from within block, then control entry
1803 for (int j = 0; j < _data_entry.length(); j++) {
1804 Node* n = _data_entry.at(j);
1805 visited_set(n);
1806 _stk.push(n);
1807 }
1808 visited_set(entry);
1809 _stk.push(entry);
1810
1811 // Do a depth first walk over out edges
1812 int rpo_idx = bb_ct - 1;
1813 int size;
1814 int reduction_uses = 0;
1815 while ((size = _stk.length()) > 0) {
1816 Node* n = _stk.top(); // Leave node on stack
1817 if (!visited_test_set(n)) {
1818 // forward arc in graph
1819 } else if (!post_visited_test(n)) {
1820 // cross or back arc
1821 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1822 Node *use = n->fast_out(i);
1823 if (in_bb(use) && !visited_test(use) &&
1824 // Don't go around backedge
1825 (!use->is_Phi() || n == entry)) {
1826 if (use->is_reduction()) {
1827 // First see if we can map the reduction on the given system we are on, then
1828 // make a data entry operation for each reduction we see.
1829 const Type *arith_type = use->bottom_type();
1830 int vopc = ReductionNode::opcode(use->Opcode(), arith_type->basic_type());
1831 if (vopc != use->Opcode()) {
1832 if (Matcher::match_rule_supported(vopc)) {
1833 _stk.push(use);
1834 reduction_uses++;
1835 } else {
1836 // failed a support issue
1837 return false;
1838 }
1839 }
1840 } else {
1841 _stk.push(use);
1842 }
1843 }
1844 }
1845 if (_stk.length() == size) {
1846 // There were no additional uses, post visit node now
1847 _stk.pop(); // Remove node from stack
1848 assert(rpo_idx >= 0, "");
1849 _block.at_put_grow(rpo_idx, n);
1850 rpo_idx--;
1851 post_visited_set(n);
1852 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1853 }
1854 } else {
1855 _stk.pop(); // Remove post-visited node from stack
1856 }
1857 }
1858
1859 // Create real map of block indices for nodes
1860 for (int j = 0; j < _block.length(); j++) {
1861 Node* n = _block.at(j);
1862 set_bb_idx(n, j);
1863 }
1864
1865 // Ensure extra info is allocated.
1866 initialize_bb();
1867
1868 #ifndef PRODUCT
1869 if (TraceSuperWord) {
1870 print_bb();
1871 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1872 for (int m = 0; m < _data_entry.length(); m++) {
1873 tty->print("%3d ", m);
1874 _data_entry.at(m)->dump();
1875 }
1876 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1877 for (int m = 0; m < _mem_slice_head.length(); m++) {
1878 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1879 tty->print(" "); _mem_slice_tail.at(m)->dump();
1880 }
1881 }
1882 #endif
1883 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1884 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
1885 }
1886
1887 //------------------------------initialize_bb---------------------------
1888 // Initialize per node info
1889 void SuperWord::initialize_bb() {
1890 Node* last = _block.at(_block.length() - 1);
1891 grow_node_info(bb_idx(last));
1892 }
1893
1894 //------------------------------bb_insert_after---------------------------
1895 // Insert n into block after pos
1896 void SuperWord::bb_insert_after(Node* n, int pos) {
1897 int n_pos = pos + 1;
1898 // Make room
1899 for (int i = _block.length() - 1; i >= n_pos; i--) {
1900 _block.at_put_grow(i+1, _block.at(i));
1901 }
1902 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1903 _node_info.at_put_grow(j+1, _node_info.at(j));
1904 }
1905 // Set value
1906 _block.at_put_grow(n_pos, n);
1907 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1908 // Adjust map from node->_idx to _block index
1909 for (int i = n_pos; i < _block.length(); i++) {
1910 set_bb_idx(_block.at(i), i);
1911 }
1912 }
1913
1914 //------------------------------compute_max_depth---------------------------
1915 // Compute max depth for expressions from beginning of block
1916 // Use to prune search paths during test for independence.
1917 void SuperWord::compute_max_depth() {
1918 int ct = 0;
1919 bool again;
1920 do {
1921 again = false;
1922 for (int i = 0; i < _block.length(); i++) {
1923 Node* n = _block.at(i);
1924 if (!n->is_Phi()) {
1925 int d_orig = depth(n);
1926 int d_in = 0;
1927 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1928 Node* pred = preds.current();
1929 if (in_bb(pred)) {
1930 d_in = MAX2(d_in, depth(pred));
1931 }
1932 }
1933 if (d_in + 1 != d_orig) {
1934 set_depth(n, d_in + 1);
1935 again = true;
1936 }
1937 }
1938 }
1939 ct++;
1940 } while (again);
1941 #ifndef PRODUCT
1942 if (TraceSuperWord && Verbose)
1943 tty->print_cr("compute_max_depth iterated: %d times", ct);
1944 #endif
1945 }
1946
1947 //-------------------------compute_vector_element_type-----------------------
1948 // Compute necessary vector element type for expressions
1949 // This propagates backwards a narrower integer type when the
1950 // upper bits of the value are not needed.
1951 // Example: char a,b,c; a = b + c;
1952 // Normally the type of the add is integer, but for packed character
1953 // operations the type of the add needs to be char.
1954 void SuperWord::compute_vector_element_type() {
1955 #ifndef PRODUCT
1956 if (TraceSuperWord && Verbose)
1957 tty->print_cr("\ncompute_velt_type:");
1958 #endif
1959
1960 // Initial type
1961 for (int i = 0; i < _block.length(); i++) {
1962 Node* n = _block.at(i);
1963 set_velt_type(n, container_type(n));
1964 }
1965
1966 // Propagate integer narrowed type backwards through operations
1967 // that don't depend on higher order bits
1968 for (int i = _block.length() - 1; i >= 0; i--) {
1969 Node* n = _block.at(i);
1970 // Only integer types need be examined
1971 const Type* vtn = velt_type(n);
1972 if (vtn->basic_type() == T_INT) {
1973 uint start, end;
1974 VectorNode::vector_operands(n, &start, &end);
1975
1976 for (uint j = start; j < end; j++) {
1977 Node* in = n->in(j);
1978 // Don't propagate through a memory
1979 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1980 data_size(n) < data_size(in)) {
1981 bool same_type = true;
1982 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1983 Node *use = in->fast_out(k);
1984 if (!in_bb(use) || !same_velt_type(use, n)) {
1985 same_type = false;
1986 break;
1987 }
1988 }
1989 if (same_type) {
1990 // For right shifts of small integer types (bool, byte, char, short)
1991 // we need precise information about sign-ness. Only Load nodes have
1992 // this information because Store nodes are the same for signed and
1993 // unsigned values. And any arithmetic operation after a load may
1994 // expand a value to signed Int so such right shifts can't be used
1995 // because vector elements do not have upper bits of Int.
1996 const Type* vt = vtn;
1997 if (VectorNode::is_shift(in)) {
1998 Node* load = in->in(1);
1999 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2000 vt = velt_type(load);
2001 } else if (in->Opcode() != Op_LShiftI) {
2002 // Widen type to Int to avoid creation of right shift vector
2003 // (align + data_size(s1) check in stmts_can_pack() will fail).
2004 // Note, left shifts work regardless type.
2005 vt = TypeInt::INT;
2006 }
2007 }
2008 set_velt_type(in, vt);
2009 }
2010 }
2011 }
2012 }
2013 }
2014 #ifndef PRODUCT
2015 if (TraceSuperWord && Verbose) {
2016 for (int i = 0; i < _block.length(); i++) {
2017 Node* n = _block.at(i);
2018 velt_type(n)->dump();
2019 tty->print("\t");
2020 n->dump();
2021 }
2022 }
2023 #endif
2024 }
2025
2026 //------------------------------memory_alignment---------------------------
2027 // Alignment within a vector memory reference
2028 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2029 SWPointer p(s, this);
2030 if (!p.valid()) {
2031 return bottom_align;
2032 }
2033 int vw = vector_width_in_bytes(s);
2034 if (vw < 2) {
2035 return bottom_align; // No vectors for this type
2036 }
2037 int offset = p.offset_in_bytes();
2038 offset += iv_adjust*p.memory_size();
2039 int off_rem = offset % vw;
2040 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2041 return off_mod;
2042 }
2043
2044 //---------------------------container_type---------------------------
2045 // Smallest type containing range of values
2046 const Type* SuperWord::container_type(Node* n) {
2047 if (n->is_Mem()) {
2048 BasicType bt = n->as_Mem()->memory_type();
2049 if (n->is_Store() && (bt == T_CHAR)) {
2050 // Use T_SHORT type instead of T_CHAR for stored values because any
2051 // preceding arithmetic operation extends values to signed Int.
2052 bt = T_SHORT;
2053 }
2054 if (n->Opcode() == Op_LoadUB) {
2055 // Adjust type for unsigned byte loads, it is important for right shifts.
2056 // T_BOOLEAN is used because there is no basic type representing type
2057 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2058 // size (one byte) and sign is important.
2059 bt = T_BOOLEAN;
2060 }
2061 return Type::get_const_basic_type(bt);
2062 }
2063 const Type* t = _igvn.type(n);
2064 if (t->basic_type() == T_INT) {
2065 // A narrow type of arithmetic operations will be determined by
2066 // propagating the type of memory operations.
2067 return TypeInt::INT;
2068 }
2069 return t;
2070 }
2071
2072 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2073 const Type* vt1 = velt_type(n1);
2074 const Type* vt2 = velt_type(n2);
2075 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2076 // Compare vectors element sizes for integer types.
2077 return data_size(n1) == data_size(n2);
2078 }
2079 return vt1 == vt2;
2080 }
2081
2082 //------------------------------in_packset---------------------------
2083 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2084 bool SuperWord::in_packset(Node* s1, Node* s2) {
2085 for (int i = 0; i < _packset.length(); i++) {
2086 Node_List* p = _packset.at(i);
2087 assert(p->size() == 2, "must be");
2088 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2089 return true;
2090 }
2091 }
2092 return false;
2093 }
2094
2095 //------------------------------in_pack---------------------------
2096 // Is s in pack p?
2097 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2098 for (uint i = 0; i < p->size(); i++) {
2099 if (p->at(i) == s) {
2100 return p;
2101 }
2102 }
2103 return NULL;
2104 }
2105
2106 //------------------------------remove_pack_at---------------------------
2107 // Remove the pack at position pos in the packset
2108 void SuperWord::remove_pack_at(int pos) {
2109 Node_List* p = _packset.at(pos);
2110 for (uint i = 0; i < p->size(); i++) {
2111 Node* s = p->at(i);
2112 set_my_pack(s, NULL);
2113 }
2114 _packset.remove_at(pos);
2115 }
2116
2117 void SuperWord::packset_sort(int n)
2118 {
2119 // simple bubble sort so that we capitalize with O(n) when its already sorted
2120 while (n != 0) {
2121 bool swapped = false;
2122 for (int i = 1; i < n; i++) {
2123 Node_List* q_low = _packset.at(i-1);
2124 Node_List* q_i = _packset.at(i);
2125
2126 // only swap when we find something to swap
2127 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2128 Node_List* t = q_i;
2129 *(_packset.adr_at(i)) = q_low;
2130 *(_packset.adr_at(i-1)) = q_i;
2131 swapped = true;
2132 }
2133 }
2134 if (swapped == false) break;
2135 n--;
2136 }
2137 }
2138
2139 //------------------------------executed_first---------------------------
2140 // Return the node executed first in pack p. Uses the RPO block list
2141 // to determine order.
2142 Node* SuperWord::executed_first(Node_List* p) {
2143 Node* n = p->at(0);
2144 int n_rpo = bb_idx(n);
2145 for (uint i = 1; i < p->size(); i++) {
2146 Node* s = p->at(i);
2147 int s_rpo = bb_idx(s);
2148 if (s_rpo < n_rpo) {
2149 n = s;
2150 n_rpo = s_rpo;
2151 }
2152 }
2153 return n;
2154 }
2155
2156 //------------------------------executed_last---------------------------
2157 // Return the node executed last in pack p.
2158 Node* SuperWord::executed_last(Node_List* p) {
2159 Node* n = p->at(0);
2160 int n_rpo = bb_idx(n);
2161 for (uint i = 1; i < p->size(); i++) {
2162 Node* s = p->at(i);
2163 int s_rpo = bb_idx(s);
2164 if (s_rpo > n_rpo) {
2165 n = s;
2166 n_rpo = s_rpo;
2167 }
2168 }
2169 return n;
2170 }
2171
2172 //----------------------------align_initial_loop_index---------------------------
2173 // Adjust pre-loop limit so that in main loop, a load/store reference
2174 // to align_to_ref will be a position zero in the vector.
2175 // (iv + k) mod vector_align == 0
2176 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2177 CountedLoopNode *main_head = lp()->as_CountedLoop();
2178 assert(main_head->is_main_loop(), "");
2179 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2180 assert(pre_end != NULL, "we must have a correct pre-loop");
2181 Node *pre_opaq1 = pre_end->limit();
2182 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2183 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2184 Node *lim0 = pre_opaq->in(1);
2185
2186 // Where we put new limit calculations
2187 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2188
2189 // Ensure the original loop limit is available from the
2190 // pre-loop Opaque1 node.
2191 Node *orig_limit = pre_opaq->original_loop_limit();
2192 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2193
2194 SWPointer align_to_ref_p(align_to_ref, this);
2195 assert(align_to_ref_p.valid(), "sanity");
2196
2197 // Given:
2198 // lim0 == original pre loop limit
2199 // V == v_align (power of 2)
2200 // invar == extra invariant piece of the address expression
2201 // e == offset [ +/- invar ]
2202 //
2203 // When reassociating expressions involving '%' the basic rules are:
2204 // (a - b) % k == 0 => a % k == b % k
2205 // and:
2206 // (a + b) % k == 0 => a % k == (k - b) % k
2207 //
2208 // For stride > 0 && scale > 0,
2209 // Derive the new pre-loop limit "lim" such that the two constraints:
2210 // (1) lim = lim0 + N (where N is some positive integer < V)
2211 // (2) (e + lim) % V == 0
2212 // are true.
2213 //
2214 // Substituting (1) into (2),
2215 // (e + lim0 + N) % V == 0
2216 // solve for N:
2217 // N = (V - (e + lim0)) % V
2218 // substitute back into (1), so that new limit
2219 // lim = lim0 + (V - (e + lim0)) % V
2220 //
2221 // For stride > 0 && scale < 0
2222 // Constraints:
2223 // lim = lim0 + N
2224 // (e - lim) % V == 0
2225 // Solving for lim:
2226 // (e - lim0 - N) % V == 0
2227 // N = (e - lim0) % V
2228 // lim = lim0 + (e - lim0) % V
2229 //
2230 // For stride < 0 && scale > 0
2231 // Constraints:
2232 // lim = lim0 - N
2233 // (e + lim) % V == 0
2234 // Solving for lim:
2235 // (e + lim0 - N) % V == 0
2236 // N = (e + lim0) % V
2237 // lim = lim0 - (e + lim0) % V
2238 //
2239 // For stride < 0 && scale < 0
2240 // Constraints:
2241 // lim = lim0 - N
2242 // (e - lim) % V == 0
2243 // Solving for lim:
2244 // (e - lim0 + N) % V == 0
2245 // N = (V - (e - lim0)) % V
2246 // lim = lim0 - (V - (e - lim0)) % V
2247
2248 int vw = vector_width_in_bytes(align_to_ref);
2249 int stride = iv_stride();
2250 int scale = align_to_ref_p.scale_in_bytes();
2251 int elt_size = align_to_ref_p.memory_size();
2252 int v_align = vw / elt_size;
2253 assert(v_align > 1, "sanity");
2254 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2255 Node *offsn = _igvn.intcon(offset);
2256
2257 Node *e = offsn;
2258 if (align_to_ref_p.invar() != NULL) {
2259 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2260 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2261 Node* aref = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2262 _igvn.register_new_node_with_optimizer(aref);
2263 _phase->set_ctrl(aref, pre_ctrl);
2264 if (align_to_ref_p.negate_invar()) {
2265 e = new SubINode(e, aref);
2266 } else {
2267 e = new AddINode(e, aref);
2268 }
2269 _igvn.register_new_node_with_optimizer(e);
2270 _phase->set_ctrl(e, pre_ctrl);
2271 }
2272 if (vw > ObjectAlignmentInBytes) {
2273 // incorporate base e +/- base && Mask >>> log2(elt)
2274 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2275 _igvn.register_new_node_with_optimizer(xbase);
2276 #ifdef _LP64
2277 xbase = new ConvL2INode(xbase);
2278 _igvn.register_new_node_with_optimizer(xbase);
2279 #endif
2280 Node* mask = _igvn.intcon(vw-1);
2281 Node* masked_xbase = new AndINode(xbase, mask);
2282 _igvn.register_new_node_with_optimizer(masked_xbase);
2283 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2284 Node* bref = new URShiftINode(masked_xbase, log2_elt);
2285 _igvn.register_new_node_with_optimizer(bref);
2286 _phase->set_ctrl(bref, pre_ctrl);
2287 e = new AddINode(e, bref);
2288 _igvn.register_new_node_with_optimizer(e);
2289 _phase->set_ctrl(e, pre_ctrl);
2290 }
2291
2292 // compute e +/- lim0
2293 if (scale < 0) {
2294 e = new SubINode(e, lim0);
2295 } else {
2296 e = new AddINode(e, lim0);
2297 }
2298 _igvn.register_new_node_with_optimizer(e);
2299 _phase->set_ctrl(e, pre_ctrl);
2300
2301 if (stride * scale > 0) {
2302 // compute V - (e +/- lim0)
2303 Node* va = _igvn.intcon(v_align);
2304 e = new SubINode(va, e);
2305 _igvn.register_new_node_with_optimizer(e);
2306 _phase->set_ctrl(e, pre_ctrl);
2307 }
2308 // compute N = (exp) % V
2309 Node* va_msk = _igvn.intcon(v_align - 1);
2310 Node* N = new AndINode(e, va_msk);
2311 _igvn.register_new_node_with_optimizer(N);
2312 _phase->set_ctrl(N, pre_ctrl);
2313
2314 // substitute back into (1), so that new limit
2315 // lim = lim0 + N
2316 Node* lim;
2317 if (stride < 0) {
2318 lim = new SubINode(lim0, N);
2319 } else {
2320 lim = new AddINode(lim0, N);
2321 }
2322 _igvn.register_new_node_with_optimizer(lim);
2323 _phase->set_ctrl(lim, pre_ctrl);
2324 Node* constrained =
2325 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2326 : (Node*) new MaxINode(lim, orig_limit);
2327 _igvn.register_new_node_with_optimizer(constrained);
2328 _phase->set_ctrl(constrained, pre_ctrl);
2329 _igvn.hash_delete(pre_opaq);
2330 pre_opaq->set_req(1, constrained);
2331 }
2332
2333 //----------------------------get_pre_loop_end---------------------------
2334 // Find pre loop end from main loop. Returns null if none.
2335 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2336 Node *ctrl = cl->in(LoopNode::EntryControl);
2337 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2338 Node *iffm = ctrl->in(0);
2339 if (!iffm->is_If()) return NULL;
2340 Node *p_f = iffm->in(0);
2341 if (!p_f->is_IfFalse()) return NULL;
2342 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2343 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2344 CountedLoopNode* loop_node = pre_end->loopnode();
2345 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2346 return pre_end;
2347 }
2348
2349
2350 //------------------------------init---------------------------
2351 void SuperWord::init() {
2352 _dg.init();
2353 _packset.clear();
2354 _disjoint_ptrs.clear();
2355 _block.clear();
2356 _data_entry.clear();
2357 _mem_slice_head.clear();
2358 _mem_slice_tail.clear();
2359 _node_info.clear();
2360 _align_to_ref = NULL;
2361 _lpt = NULL;
2362 _lp = NULL;
2363 _bb = NULL;
2364 _iv = NULL;
2365 }
2366
2367 //------------------------------print_packset---------------------------
2368 void SuperWord::print_packset() {
2369 #ifndef PRODUCT
2370 tty->print_cr("packset");
2371 for (int i = 0; i < _packset.length(); i++) {
2372 tty->print_cr("Pack: %d", i);
2373 Node_List* p = _packset.at(i);
2374 print_pack(p);
2375 }
2376 #endif
2377 }
2378
2379 //------------------------------print_pack---------------------------
2380 void SuperWord::print_pack(Node_List* p) {
2381 for (uint i = 0; i < p->size(); i++) {
2382 print_stmt(p->at(i));
2383 }
2384 }
2385
2386 //------------------------------print_bb---------------------------
2387 void SuperWord::print_bb() {
2388 #ifndef PRODUCT
2389 tty->print_cr("\nBlock");
2390 for (int i = 0; i < _block.length(); i++) {
2391 Node* n = _block.at(i);
2392 tty->print("%d ", i);
2393 if (n) {
2394 n->dump();
2395 }
2396 }
2397 #endif
2398 }
2399
2400 //------------------------------print_stmt---------------------------
2401 void SuperWord::print_stmt(Node* s) {
2402 #ifndef PRODUCT
2403 tty->print(" align: %d \t", alignment(s));
2404 s->dump();
2405 #endif
2406 }
2407
2408 //------------------------------blank---------------------------
2409 char* SuperWord::blank(uint depth) {
2410 static char blanks[101];
2411 assert(depth < 101, "too deep");
2412 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2413 blanks[depth] = '\0';
2414 return blanks;
2415 }
2416
2417
2418 //==============================SWPointer===========================
2419
2420 //----------------------------SWPointer------------------------
2421 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2422 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2423 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2424
2425 Node* adr = mem->in(MemNode::Address);
2426 if (!adr->is_AddP()) {
2427 assert(!valid(), "too complex");
2428 return;
2429 }
2430 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2431 Node* base = adr->in(AddPNode::Base);
2432 //unsafe reference could not be aligned appropriately without runtime checking
2433 if (base == NULL || base->bottom_type() == Type::TOP) {
2434 assert(!valid(), "unsafe access");
2435 return;
2436 }
2437 for (int i = 0; i < 3; i++) {
2438 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2439 assert(!valid(), "too complex");
2440 return;
2441 }
2442 adr = adr->in(AddPNode::Address);
2443 if (base == adr || !adr->is_AddP()) {
2444 break; // stop looking at addp's
2445 }
2446 }
2447 _base = base;
2448 _adr = adr;
2449 assert(valid(), "Usable");
2450 }
2451
2452 // Following is used to create a temporary object during
2453 // the pattern match of an address expression.
2454 SWPointer::SWPointer(SWPointer* p) :
2455 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2456 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2457
2458 //------------------------scaled_iv_plus_offset--------------------
2459 // Match: k*iv + offset
2460 // where: k is a constant that maybe zero, and
2461 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2462 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2463 if (scaled_iv(n)) {
2464 return true;
2465 }
2466 if (offset_plus_k(n)) {
2467 return true;
2468 }
2469 int opc = n->Opcode();
2470 if (opc == Op_AddI) {
2471 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2472 return true;
2473 }
2474 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2475 return true;
2476 }
2477 } else if (opc == Op_SubI) {
2478 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2479 return true;
2480 }
2481 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2482 _scale *= -1;
2483 return true;
2484 }
2485 }
2486 return false;
2487 }
2488
2489 //----------------------------scaled_iv------------------------
2490 // Match: k*iv where k is a constant that's not zero
2491 bool SWPointer::scaled_iv(Node* n) {
2492 if (_scale != 0) {
2493 return false; // already found a scale
2494 }
2495 if (n == iv()) {
2496 _scale = 1;
2497 return true;
2498 }
2499 int opc = n->Opcode();
2500 if (opc == Op_MulI) {
2501 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2502 _scale = n->in(2)->get_int();
2503 return true;
2504 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2505 _scale = n->in(1)->get_int();
2506 return true;
2507 }
2508 } else if (opc == Op_LShiftI) {
2509 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2510 _scale = 1 << n->in(2)->get_int();
2511 return true;
2512 }
2513 } else if (opc == Op_ConvI2L) {
2514 if (scaled_iv_plus_offset(n->in(1))) {
2515 return true;
2516 }
2517 } else if (opc == Op_LShiftL) {
2518 if (!has_iv() && _invar == NULL) {
2519 // Need to preserve the current _offset value, so
2520 // create a temporary object for this expression subtree.
2521 // Hacky, so should re-engineer the address pattern match.
2522 SWPointer tmp(this);
2523 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2524 if (tmp._invar == NULL) {
2525 int mult = 1 << n->in(2)->get_int();
2526 _scale = tmp._scale * mult;
2527 _offset += tmp._offset * mult;
2528 return true;
2529 }
2530 }
2531 }
2532 }
2533 return false;
2534 }
2535
2536 //----------------------------offset_plus_k------------------------
2537 // Match: offset is (k [+/- invariant])
2538 // where k maybe zero and invariant is optional, but not both.
2539 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2540 int opc = n->Opcode();
2541 if (opc == Op_ConI) {
2542 _offset += negate ? -(n->get_int()) : n->get_int();
2543 return true;
2544 } else if (opc == Op_ConL) {
2545 // Okay if value fits into an int
2546 const TypeLong* t = n->find_long_type();
2547 if (t->higher_equal(TypeLong::INT)) {
2548 jlong loff = n->get_long();
2549 jint off = (jint)loff;
2550 _offset += negate ? -off : loff;
2551 return true;
2552 }
2553 return false;
2554 }
2555 if (_invar != NULL) return false; // already have an invariant
2556 if (opc == Op_AddI) {
2557 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2558 _negate_invar = negate;
2559 _invar = n->in(1);
2560 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2561 return true;
2562 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2563 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2564 _negate_invar = negate;
2565 _invar = n->in(2);
2566 return true;
2567 }
2568 }
2569 if (opc == Op_SubI) {
2570 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2571 _negate_invar = negate;
2572 _invar = n->in(1);
2573 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2574 return true;
2575 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2576 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2577 _negate_invar = !negate;
2578 _invar = n->in(2);
2579 return true;
2580 }
2581 }
2582 if (invariant(n)) {
2583 _negate_invar = negate;
2584 _invar = n;
2585 return true;
2586 }
2587 return false;
2588 }
2589
2590 //----------------------------print------------------------
2591 void SWPointer::print() {
2592 #ifndef PRODUCT
2593 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2594 _base != NULL ? _base->_idx : 0,
2595 _adr != NULL ? _adr->_idx : 0,
2596 _scale, _offset,
2597 _negate_invar?'-':'+',
2598 _invar != NULL ? _invar->_idx : 0);
2599 #endif
2600 }
2601
2602 // ========================= OrderedPair =====================
2603
2604 const OrderedPair OrderedPair::initial;
2605
2606 // ========================= SWNodeInfo =====================
2607
2608 const SWNodeInfo SWNodeInfo::initial;
2609
2610
2611 // ============================ DepGraph ===========================
2612
2613 //------------------------------make_node---------------------------
2614 // Make a new dependence graph node for an ideal node.
2615 DepMem* DepGraph::make_node(Node* node) {
2616 DepMem* m = new (_arena) DepMem(node);
2617 if (node != NULL) {
2618 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2619 _map.at_put_grow(node->_idx, m);
2620 }
2621 return m;
2622 }
2623
2624 //------------------------------make_edge---------------------------
2625 // Make a new dependence graph edge from dpred -> dsucc
2626 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2627 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2628 dpred->set_out_head(e);
2629 dsucc->set_in_head(e);
2630 return e;
2631 }
2632
2633 // ========================== DepMem ========================
2634
2635 //------------------------------in_cnt---------------------------
2636 int DepMem::in_cnt() {
2637 int ct = 0;
2638 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2639 return ct;
2640 }
2641
2642 //------------------------------out_cnt---------------------------
2643 int DepMem::out_cnt() {
2644 int ct = 0;
2645 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2646 return ct;
2647 }
2648
2649 //------------------------------print-----------------------------
2650 void DepMem::print() {
2651 #ifndef PRODUCT
2652 tty->print(" DepNode %d (", _node->_idx);
2653 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2654 Node* pred = p->pred()->node();
2655 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2656 }
2657 tty->print(") [");
2658 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2659 Node* succ = s->succ()->node();
2660 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2661 }
2662 tty->print_cr(" ]");
2663 #endif
2664 }
2665
2666 // =========================== DepEdge =========================
2667
2668 //------------------------------DepPreds---------------------------
2669 void DepEdge::print() {
2670 #ifndef PRODUCT
2671 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2672 #endif
2673 }
2674
2675 // =========================== DepPreds =========================
2676 // Iterator over predecessor edges in the dependence graph.
2677
2678 //------------------------------DepPreds---------------------------
2679 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2680 _n = n;
2681 _done = false;
2682 if (_n->is_Store() || _n->is_Load()) {
2683 _next_idx = MemNode::Address;
2684 _end_idx = n->req();
2685 _dep_next = dg.dep(_n)->in_head();
2686 } else if (_n->is_Mem()) {
2687 _next_idx = 0;
2688 _end_idx = 0;
2689 _dep_next = dg.dep(_n)->in_head();
2690 } else {
2691 _next_idx = 1;
2692 _end_idx = _n->req();
2693 _dep_next = NULL;
2694 }
2695 next();
2696 }
2697
2698 //------------------------------next---------------------------
2699 void DepPreds::next() {
2700 if (_dep_next != NULL) {
2701 _current = _dep_next->pred()->node();
2702 _dep_next = _dep_next->next_in();
2703 } else if (_next_idx < _end_idx) {
2704 _current = _n->in(_next_idx++);
2705 } else {
2706 _done = true;
2707 }
2708 }
2709
2710 // =========================== DepSuccs =========================
2711 // Iterator over successor edges in the dependence graph.
2712
2713 //------------------------------DepSuccs---------------------------
2714 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2715 _n = n;
2716 _done = false;
2717 if (_n->is_Load()) {
2718 _next_idx = 0;
2719 _end_idx = _n->outcnt();
2720 _dep_next = dg.dep(_n)->out_head();
2721 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2722 _next_idx = 0;
2723 _end_idx = 0;
2724 _dep_next = dg.dep(_n)->out_head();
2725 } else {
2726 _next_idx = 0;
2727 _end_idx = _n->outcnt();
2728 _dep_next = NULL;
2729 }
2730 next();
2731 }
2732
2733 //-------------------------------next---------------------------
2734 void DepSuccs::next() {
2735 if (_dep_next != NULL) {
2736 _current = _dep_next->succ()->node();
2737 _dep_next = _dep_next->next_out();
2738 } else if (_next_idx < _end_idx) {
2739 _current = _n->raw_out(_next_idx++);
2740 } else {
2741 _done = true;
2742 }
2743 }