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 _early_return(true)
71 {}
72
73 //------------------------------transform_loop---------------------------
74 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
75 assert(UseSuperWord, "should be");
76 // Do vectors exist on this architecture?
77 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
78
79 assert(lpt->_head->is_CountedLoop(), "must be");
80 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
81
82 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
83
84 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
85
86 // Check for no control flow in body (other than exit)
87 Node *cl_exit = cl->loopexit();
88 if (cl_exit->in(0) != lpt->_head) return;
89
90 // Make sure the are no extra control users of the loop backedge
91 if (cl->back_control()->outcnt() != 1) {
92 return;
93 }
94
95 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
96 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
97 if (pre_end == NULL) return;
98 Node *pre_opaq1 = pre_end->limit();
99 if (pre_opaq1->Opcode() != Op_Opaque1) return;
100
101 init(); // initialize data structures
102
103 set_lpt(lpt);
104 set_lp(cl);
105
106 // For now, define one block which is the entire loop body
107 set_bb(cl);
108
109 if (do_optimization) {
110 assert(_packset.length() == 0, "packset must be empty");
111 SLP_extract();
112 }
113 }
114
115 //------------------------------SLP_extract---------------------------
116 // Extract the superword level parallelism
117 //
118 // 1) A reverse post-order of nodes in the block is constructed. By scanning
119 // this list from first to last, all definitions are visited before their uses.
120 //
121 // 2) A point-to-point dependence graph is constructed between memory references.
122 // This simplies the upcoming "independence" checker.
123 //
124 // 3) The maximum depth in the node graph from the beginning of the block
125 // to each node is computed. This is used to prune the graph search
126 // in the independence checker.
127 //
128 // 4) For integer types, the necessary bit width is propagated backwards
129 // from stores to allow packed operations on byte, char, and short
130 // integers. This reverses the promotion to type "int" that javac
131 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
132 //
133 // 5) One of the memory references is picked to be an aligned vector reference.
134 // The pre-loop trip count is adjusted to align this reference in the
135 // unrolled body.
136 //
137 // 6) The initial set of pack pairs is seeded with memory references.
138 //
139 // 7) The set of pack pairs is extended by following use->def and def->use links.
140 //
141 // 8) The pairs are combined into vector sized packs.
142 //
143 // 9) Reorder the memory slices to co-locate members of the memory packs.
144 //
145 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
146 // inserting scalar promotion, vector creation from multiple scalars, and
147 // extraction of scalar values from vectors.
148 //
149 void SuperWord::SLP_extract() {
150
151 // Ready the block
152 if (!construct_bb())
153 return; // Exit if no interesting nodes or complex graph.
154
155 dependence_graph();
156
157 compute_max_depth();
158
159 compute_vector_element_type();
160
161 // Attempt vectorization
162
163 find_adjacent_refs();
164
165 extend_packlist();
166
167 combine_packs();
168
169 construct_my_pack_map();
170
171 filter_packs();
172
173 schedule();
174
175 output();
176 }
177
178 //------------------------------find_adjacent_refs---------------------------
179 // Find the adjacent memory references and create pack pairs for them.
180 // This is the initial set of packs that will then be extended by
181 // following use->def and def->use links. The align positions are
182 // assigned relative to the reference "align_to_ref"
183 void SuperWord::find_adjacent_refs() {
184 // Get list of memory operations
185 Node_List memops;
186 for (int i = 0; i < _block.length(); i++) {
187 Node* n = _block.at(i);
188 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
189 is_java_primitive(n->as_Mem()->memory_type())) {
190 int align = memory_alignment(n->as_Mem(), 0);
191 if (align != bottom_align) {
192 memops.push(n);
193 }
194 }
195 }
196
197 Node_List align_to_refs;
198 int best_iv_adjustment = 0;
199 MemNode* best_align_to_mem_ref = NULL;
200
201 while (memops.size() != 0) {
202 // Find a memory reference to align to.
203 MemNode* mem_ref = find_align_to_ref(memops);
204 if (mem_ref == NULL) break;
205 align_to_refs.push(mem_ref);
206 int iv_adjustment = get_iv_adjustment(mem_ref);
207
208 if (best_align_to_mem_ref == NULL) {
209 // Set memory reference which is the best from all memory operations
210 // to be used for alignment. The pre-loop trip count is modified to align
211 // this reference to a vector-aligned address.
212 best_align_to_mem_ref = mem_ref;
213 best_iv_adjustment = iv_adjustment;
214 }
215
216 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
217 // Set alignment relative to "align_to_ref" for all related memory operations.
218 for (int i = memops.size() - 1; i >= 0; i--) {
219 MemNode* s = memops.at(i)->as_Mem();
220 if (isomorphic(s, mem_ref)) {
221 SWPointer p2(s, this, NULL, false);
222 if (p2.comparable(align_to_ref_p)) {
223 int align = memory_alignment(s, iv_adjustment);
224 set_alignment(s, align);
225 }
226 }
227 }
228
229 // Create initial pack pairs of memory operations for which
230 // alignment is set and vectors will be aligned.
231 bool create_pack = true;
232 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
233 if (!Matcher::misaligned_vectors_ok()) {
234 int vw = vector_width(mem_ref);
235 int vw_best = vector_width(best_align_to_mem_ref);
236 if (vw > vw_best) {
237 // Do not vectorize a memory access with more elements per vector
238 // if unaligned memory access is not allowed because number of
239 // iterations in pre-loop will be not enough to align it.
240 create_pack = false;
241 }
242 }
243 } else {
244 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
245 // Can't allow vectorization of unaligned memory accesses with the
246 // same type since it could be overlapped accesses to the same array.
247 create_pack = false;
248 } else {
249 // Allow independent (different type) unaligned memory operations
250 // if HW supports them.
251 if (!Matcher::misaligned_vectors_ok()) {
252 create_pack = false;
253 } else {
254 // Check if packs of the same memory type but
255 // with a different alignment were created before.
256 for (uint i = 0; i < align_to_refs.size(); i++) {
257 MemNode* mr = align_to_refs.at(i)->as_Mem();
258 if (same_velt_type(mr, mem_ref) &&
259 memory_alignment(mr, iv_adjustment) != 0)
260 create_pack = false;
261 }
262 }
263 }
264 }
265 if (create_pack) {
266 for (uint i = 0; i < memops.size(); i++) {
267 Node* s1 = memops.at(i);
268 int align = alignment(s1);
269 if (align == top_align) continue;
270 for (uint j = 0; j < memops.size(); j++) {
271 Node* s2 = memops.at(j);
272 if (alignment(s2) == top_align) continue;
273 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
274 if (stmts_can_pack(s1, s2, align)) {
275 Node_List* pair = new Node_List();
276 pair->push(s1);
277 pair->push(s2);
278 _packset.append(pair);
279 }
280 }
281 }
282 }
283 } else { // Don't create unaligned pack
284 // First, remove remaining memory ops of the same type from the list.
285 for (int i = memops.size() - 1; i >= 0; i--) {
286 MemNode* s = memops.at(i)->as_Mem();
287 if (same_velt_type(s, mem_ref)) {
288 memops.remove(i);
289 }
290 }
291
292 // Second, remove already constructed packs of the same type.
293 for (int i = _packset.length() - 1; i >= 0; i--) {
294 Node_List* p = _packset.at(i);
295 MemNode* s = p->at(0)->as_Mem();
296 if (same_velt_type(s, mem_ref)) {
297 remove_pack_at(i);
298 }
299 }
300
301 // If needed find the best memory reference for loop alignment again.
302 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
303 // Put memory ops from remaining packs back on memops list for
304 // the best alignment search.
305 uint orig_msize = memops.size();
306 for (int i = 0; i < _packset.length(); i++) {
307 Node_List* p = _packset.at(i);
308 MemNode* s = p->at(0)->as_Mem();
309 assert(!same_velt_type(s, mem_ref), "sanity");
310 memops.push(s);
311 }
312 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
313 if (best_align_to_mem_ref == NULL) break;
314 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
315 // Restore list.
316 while (memops.size() > orig_msize)
317 (void)memops.pop();
318 }
319 } // unaligned memory accesses
320
321 // Remove used mem nodes.
322 for (int i = memops.size() - 1; i >= 0; i--) {
323 MemNode* m = memops.at(i)->as_Mem();
324 if (alignment(m) != top_align) {
325 memops.remove(i);
326 }
327 }
328
329 } // while (memops.size() != 0
330 set_align_to_ref(best_align_to_mem_ref);
331
332 #ifndef PRODUCT
333 if (TraceSuperWord) {
334 tty->print_cr("\nAfter find_adjacent_refs");
335 print_packset();
336 }
337 #endif
338 }
339
340 //------------------------------find_align_to_ref---------------------------
341 // Find a memory reference to align the loop induction variable to.
342 // Looks first at stores then at loads, looking for a memory reference
343 // with the largest number of references similar to it.
344 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
345 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
346
347 // Count number of comparable memory ops
348 for (uint i = 0; i < memops.size(); i++) {
349 MemNode* s1 = memops.at(i)->as_Mem();
350 SWPointer p1(s1, this, NULL, false);
351 // Discard if pre loop can't align this reference
352 if (!ref_is_alignable(p1)) {
353 *cmp_ct.adr_at(i) = 0;
354 continue;
355 }
356 for (uint j = i+1; j < memops.size(); j++) {
357 MemNode* s2 = memops.at(j)->as_Mem();
358 if (isomorphic(s1, s2)) {
359 SWPointer p2(s2, this, NULL, false);
360 if (p1.comparable(p2)) {
361 (*cmp_ct.adr_at(i))++;
362 (*cmp_ct.adr_at(j))++;
363 }
364 }
365 }
366 }
367
368 // Find Store (or Load) with the greatest number of "comparable" references,
369 // biggest vector size, smallest data size and smallest iv offset.
370 int max_ct = 0;
371 int max_vw = 0;
372 int max_idx = -1;
373 int min_size = max_jint;
374 int min_iv_offset = max_jint;
375 for (uint j = 0; j < memops.size(); j++) {
376 MemNode* s = memops.at(j)->as_Mem();
377 if (s->is_Store()) {
378 int vw = vector_width_in_bytes(s);
379 assert(vw > 1, "sanity");
380 SWPointer p(s, this, NULL, false);
381 if (cmp_ct.at(j) > max_ct ||
382 cmp_ct.at(j) == max_ct &&
383 (vw > max_vw ||
384 vw == max_vw &&
385 (data_size(s) < min_size ||
386 data_size(s) == min_size &&
387 (p.offset_in_bytes() < min_iv_offset)))) {
388 max_ct = cmp_ct.at(j);
389 max_vw = vw;
390 max_idx = j;
391 min_size = data_size(s);
392 min_iv_offset = p.offset_in_bytes();
393 }
394 }
395 }
396 // If no stores, look at loads
397 if (max_ct == 0) {
398 for (uint j = 0; j < memops.size(); j++) {
399 MemNode* s = memops.at(j)->as_Mem();
400 if (s->is_Load()) {
401 int vw = vector_width_in_bytes(s);
402 assert(vw > 1, "sanity");
403 SWPointer p(s, this, NULL, false);
404 if (cmp_ct.at(j) > max_ct ||
405 cmp_ct.at(j) == max_ct &&
406 (vw > max_vw ||
407 vw == max_vw &&
408 (data_size(s) < min_size ||
409 data_size(s) == min_size &&
410 (p.offset_in_bytes() < min_iv_offset)))) {
411 max_ct = cmp_ct.at(j);
412 max_vw = vw;
413 max_idx = j;
414 min_size = data_size(s);
415 min_iv_offset = p.offset_in_bytes();
416 }
417 }
418 }
419 }
420
421 #ifdef ASSERT
422 if (TraceSuperWord && Verbose) {
423 tty->print_cr("\nVector memops after find_align_to_refs");
424 for (uint i = 0; i < memops.size(); i++) {
425 MemNode* s = memops.at(i)->as_Mem();
426 s->dump();
427 }
428 }
429 #endif
430
431 if (max_ct > 0) {
432 #ifdef ASSERT
433 if (TraceSuperWord) {
434 tty->print("\nVector align to node: ");
435 memops.at(max_idx)->as_Mem()->dump();
436 }
437 #endif
438 return memops.at(max_idx)->as_Mem();
439 }
440 return NULL;
441 }
442
443 //------------------------------ref_is_alignable---------------------------
444 // Can the preloop align the reference to position zero in the vector?
445 bool SuperWord::ref_is_alignable(SWPointer& p) {
446 if (!p.has_iv()) {
447 return true; // no induction variable
448 }
449 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
450 assert(pre_end != NULL, "we must have a correct pre-loop");
451 assert(pre_end->stride_is_con(), "pre loop stride is constant");
452 int preloop_stride = pre_end->stride_con();
453
454 int span = preloop_stride * p.scale_in_bytes();
455 int mem_size = p.memory_size();
456 int offset = p.offset_in_bytes();
457 // Stride one accesses are alignable if offset is aligned to memory operation size.
458 // Offset can be unaligned when UseUnalignedAccesses is used.
459 if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
460 return true;
461 }
462 // If initial offset from start of object is computable,
463 // compute alignment within the vector.
464 int vw = vector_width_in_bytes(p.mem());
465 assert(vw > 1, "sanity");
466 if (vw % span == 0) {
467 Node* init_nd = pre_end->init_trip();
468 if (init_nd->is_Con() && p.invar() == NULL) {
469 int init = init_nd->bottom_type()->is_int()->get_con();
470
471 int init_offset = init * p.scale_in_bytes() + offset;
472 assert(init_offset >= 0, "positive offset from object start");
473
474 if (span > 0) {
475 return (vw - (init_offset % vw)) % span == 0;
476 } else {
477 assert(span < 0, "nonzero stride * scale");
478 return (init_offset % vw) % -span == 0;
479 }
480 }
481 }
482 return false;
483 }
484
485 //---------------------------get_iv_adjustment---------------------------
486 // Calculate loop's iv adjustment for this memory ops.
487 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
488 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
489 int offset = align_to_ref_p.offset_in_bytes();
490 int scale = align_to_ref_p.scale_in_bytes();
491 int vw = vector_width_in_bytes(mem_ref);
492 assert(vw > 1, "sanity");
493 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
494 // At least one iteration is executed in pre-loop by default. As result
495 // several iterations are needed to align memory operations in main-loop even
496 // if offset is 0.
497 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
498 int elt_size = align_to_ref_p.memory_size();
499 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
500 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
501 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
502
503 #ifndef PRODUCT
504 if (TraceSuperWord)
505 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
506 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
507 #endif
508 return iv_adjustment;
509 }
510
511 //---------------------------dependence_graph---------------------------
512 // Construct dependency graph.
513 // Add dependence edges to load/store nodes for memory dependence
514 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
515 void SuperWord::dependence_graph() {
516 // First, assign a dependence node to each memory node
517 for (int i = 0; i < _block.length(); i++ ) {
518 Node *n = _block.at(i);
519 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
520 _dg.make_node(n);
521 }
522 }
523
524 // For each memory slice, create the dependences
525 for (int i = 0; i < _mem_slice_head.length(); i++) {
526 Node* n = _mem_slice_head.at(i);
527 Node* n_tail = _mem_slice_tail.at(i);
528
529 // Get slice in predecessor order (last is first)
530 mem_slice_preds(n_tail, n, _nlist);
531
532 // Make the slice dependent on the root
533 DepMem* slice = _dg.dep(n);
534 _dg.make_edge(_dg.root(), slice);
535
536 // Create a sink for the slice
537 DepMem* slice_sink = _dg.make_node(NULL);
538 _dg.make_edge(slice_sink, _dg.tail());
539
540 // Now visit each pair of memory ops, creating the edges
541 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
542 Node* s1 = _nlist.at(j);
543
544 // If no dependency yet, use slice
545 if (_dg.dep(s1)->in_cnt() == 0) {
546 _dg.make_edge(slice, s1);
547 }
548 SWPointer p1(s1->as_Mem(), this, NULL, false);
549 bool sink_dependent = true;
550 for (int k = j - 1; k >= 0; k--) {
551 Node* s2 = _nlist.at(k);
552 if (s1->is_Load() && s2->is_Load())
553 continue;
554 SWPointer p2(s2->as_Mem(), this, NULL, false);
555
556 int cmp = p1.cmp(p2);
557 if (SuperWordRTDepCheck &&
558 p1.base() != p2.base() && p1.valid() && p2.valid()) {
559 // Create a runtime check to disambiguate
560 OrderedPair pp(p1.base(), p2.base());
561 _disjoint_ptrs.append_if_missing(pp);
562 } else if (!SWPointer::not_equal(cmp)) {
563 // Possibly same address
564 _dg.make_edge(s1, s2);
565 sink_dependent = false;
566 }
567 }
568 if (sink_dependent) {
569 _dg.make_edge(s1, slice_sink);
570 }
571 }
572 #ifndef PRODUCT
573 if (TraceSuperWord) {
574 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
575 for (int q = 0; q < _nlist.length(); q++) {
576 _dg.print(_nlist.at(q));
577 }
578 tty->cr();
579 }
580 #endif
581 _nlist.clear();
582 }
583
584 #ifndef PRODUCT
585 if (TraceSuperWord) {
586 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
587 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
588 _disjoint_ptrs.at(r).print();
589 tty->cr();
590 }
591 tty->cr();
592 }
593 #endif
594 }
595
596 //---------------------------mem_slice_preds---------------------------
597 // Return a memory slice (node list) in predecessor order starting at "start"
598 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
599 assert(preds.length() == 0, "start empty");
600 Node* n = start;
601 Node* prev = NULL;
602 while (true) {
603 assert(in_bb(n), "must be in block");
604 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
605 Node* out = n->fast_out(i);
606 if (out->is_Load()) {
607 if (in_bb(out)) {
608 preds.push(out);
609 }
610 } else {
611 // FIXME
612 if (out->is_MergeMem() && !in_bb(out)) {
613 // Either unrolling is causing a memory edge not to disappear,
614 // or need to run igvn.optimize() again before SLP
615 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
616 // Ditto. Not sure what else to check further.
617 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
618 // StoreCM has an input edge used as a precedence edge.
619 // Maybe an issue when oop stores are vectorized.
620 } else {
621 assert(out == prev || prev == NULL, "no branches off of store slice");
622 }
623 }
624 }
625 if (n == stop) break;
626 preds.push(n);
627 prev = n;
628 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
629 n = n->in(MemNode::Memory);
630 }
631 }
632
633 //------------------------------stmts_can_pack---------------------------
634 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
635 // s1 aligned at "align"
636 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
637
638 // Do not use superword for non-primitives
639 BasicType bt1 = velt_basic_type(s1);
640 BasicType bt2 = velt_basic_type(s2);
641 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
642 return false;
643 if (Matcher::max_vector_size(bt1) < 2) {
644 return false; // No vectors for this type
645 }
646
647 if (isomorphic(s1, s2)) {
648 if (independent(s1, s2) || reduction(s1, s2)) {
649 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
650 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
651 int s1_align = alignment(s1);
652 int s2_align = alignment(s2);
653 if (s1_align == top_align || s1_align == align) {
654 if (s2_align == top_align || s2_align == align + data_size(s1)) {
655 return true;
656 }
657 }
658 }
659 }
660 }
661 }
662 return false;
663 }
664
665 //------------------------------exists_at---------------------------
666 // Does s exist in a pack at position pos?
667 bool SuperWord::exists_at(Node* s, uint pos) {
668 for (int i = 0; i < _packset.length(); i++) {
669 Node_List* p = _packset.at(i);
670 if (p->at(pos) == s) {
671 return true;
672 }
673 }
674 return false;
675 }
676
677 //------------------------------are_adjacent_refs---------------------------
678 // Is s1 immediately before s2 in memory?
679 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
680 if (!s1->is_Mem() || !s2->is_Mem()) return false;
681 if (!in_bb(s1) || !in_bb(s2)) return false;
682
683 // Do not use superword for non-primitives
684 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
685 !is_java_primitive(s2->as_Mem()->memory_type())) {
686 return false;
687 }
688
689 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
690 // only pack memops that are in the same alias set until that's fixed.
691 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
692 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
693 return false;
694 SWPointer p1(s1->as_Mem(), this, NULL, false);
695 SWPointer p2(s2->as_Mem(), this, NULL, false);
696 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
697 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
698 return diff == data_size(s1);
699 }
700
701 //------------------------------isomorphic---------------------------
702 // Are s1 and s2 similar?
703 bool SuperWord::isomorphic(Node* s1, Node* s2) {
704 if (s1->Opcode() != s2->Opcode()) return false;
705 if (s1->req() != s2->req()) return false;
706 if (s1->in(0) != s2->in(0)) return false;
707 if (!same_velt_type(s1, s2)) return false;
708 return true;
709 }
710
711 //------------------------------independent---------------------------
712 // Is there no data path from s1 to s2 or s2 to s1?
713 bool SuperWord::independent(Node* s1, Node* s2) {
714 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
715 int d1 = depth(s1);
716 int d2 = depth(s2);
717 if (d1 == d2) return s1 != s2;
718 Node* deep = d1 > d2 ? s1 : s2;
719 Node* shallow = d1 > d2 ? s2 : s1;
720
721 visited_clear();
722
723 return independent_path(shallow, deep);
724 }
725
726 //------------------------------reduction---------------------------
727 // Is there a data path between s1 and s2 and the nodes reductions?
728 bool SuperWord::reduction(Node* s1, Node* s2) {
729 bool retValue = false;
730 int d1 = depth(s1);
731 int d2 = depth(s2);
732 if (d1 + 1 == d2) {
733 if (s1->is_reduction() && s2->is_reduction()) {
734 // This is an ordered set, so s1 should define s2
735 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
736 Node* t1 = s1->fast_out(i);
737 if (t1 == s2) {
738 // both nodes are reductions and connected
739 retValue = true;
740 }
741 }
742 }
743 }
744
745 return retValue;
746 }
747
748 //------------------------------independent_path------------------------------
749 // Helper for independent
750 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
751 if (dp >= 1000) return false; // stop deep recursion
752 visited_set(deep);
753 int shal_depth = depth(shallow);
754 assert(shal_depth <= depth(deep), "must be");
755 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
756 Node* pred = preds.current();
757 if (in_bb(pred) && !visited_test(pred)) {
758 if (shallow == pred) {
759 return false;
760 }
761 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
762 return false;
763 }
764 }
765 }
766 return true;
767 }
768
769 //------------------------------set_alignment---------------------------
770 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
771 set_alignment(s1, align);
772 if (align == top_align || align == bottom_align) {
773 set_alignment(s2, align);
774 } else {
775 set_alignment(s2, align + data_size(s1));
776 }
777 }
778
779 //------------------------------data_size---------------------------
780 int SuperWord::data_size(Node* s) {
781 int bsize = type2aelembytes(velt_basic_type(s));
782 assert(bsize != 0, "valid size");
783 return bsize;
784 }
785
786 //------------------------------extend_packlist---------------------------
787 // Extend packset by following use->def and def->use links from pack members.
788 void SuperWord::extend_packlist() {
789 bool changed;
790 do {
791 packset_sort(_packset.length());
792 changed = false;
793 for (int i = 0; i < _packset.length(); i++) {
794 Node_List* p = _packset.at(i);
795 changed |= follow_use_defs(p);
796 changed |= follow_def_uses(p);
797 }
798 } while (changed);
799
800 if (_race_possible) {
801 for (int i = 0; i < _packset.length(); i++) {
802 Node_List* p = _packset.at(i);
803 order_def_uses(p);
804 }
805 }
806
807 #ifndef PRODUCT
808 if (TraceSuperWord) {
809 tty->print_cr("\nAfter extend_packlist");
810 print_packset();
811 }
812 #endif
813 }
814
815 //------------------------------follow_use_defs---------------------------
816 // Extend the packset by visiting operand definitions of nodes in pack p
817 bool SuperWord::follow_use_defs(Node_List* p) {
818 assert(p->size() == 2, "just checking");
819 Node* s1 = p->at(0);
820 Node* s2 = p->at(1);
821 assert(s1->req() == s2->req(), "just checking");
822 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
823
824 if (s1->is_Load()) return false;
825
826 int align = alignment(s1);
827 bool changed = false;
828 int start = s1->is_Store() ? MemNode::ValueIn : 1;
829 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
830 for (int j = start; j < end; j++) {
831 Node* t1 = s1->in(j);
832 Node* t2 = s2->in(j);
833 if (!in_bb(t1) || !in_bb(t2))
834 continue;
835 if (stmts_can_pack(t1, t2, align)) {
836 if (est_savings(t1, t2) >= 0) {
837 Node_List* pair = new Node_List();
838 pair->push(t1);
839 pair->push(t2);
840 _packset.append(pair);
841 set_alignment(t1, t2, align);
842 changed = true;
843 }
844 }
845 }
846 return changed;
847 }
848
849 //------------------------------follow_def_uses---------------------------
850 // Extend the packset by visiting uses of nodes in pack p
851 bool SuperWord::follow_def_uses(Node_List* p) {
852 bool changed = false;
853 Node* s1 = p->at(0);
854 Node* s2 = p->at(1);
855 assert(p->size() == 2, "just checking");
856 assert(s1->req() == s2->req(), "just checking");
857 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
858
859 if (s1->is_Store()) return false;
860
861 int align = alignment(s1);
862 int savings = -1;
863 int num_s1_uses = 0;
864 Node* u1 = NULL;
865 Node* u2 = NULL;
866 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
867 Node* t1 = s1->fast_out(i);
868 num_s1_uses++;
869 if (!in_bb(t1)) continue;
870 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
871 Node* t2 = s2->fast_out(j);
872 if (!in_bb(t2)) continue;
873 if (!opnd_positions_match(s1, t1, s2, t2))
874 continue;
875 if (stmts_can_pack(t1, t2, align)) {
876 int my_savings = est_savings(t1, t2);
877 if (my_savings > savings) {
878 savings = my_savings;
879 u1 = t1;
880 u2 = t2;
881 }
882 }
883 }
884 }
885 if (num_s1_uses > 1) {
886 _race_possible = true;
887 }
888 if (savings >= 0) {
889 Node_List* pair = new Node_List();
890 pair->push(u1);
891 pair->push(u2);
892 _packset.append(pair);
893 set_alignment(u1, u2, align);
894 changed = true;
895 }
896 return changed;
897 }
898
899 //------------------------------order_def_uses---------------------------
900 // For extended packsets, ordinally arrange uses packset by major component
901 void SuperWord::order_def_uses(Node_List* p) {
902 Node* s1 = p->at(0);
903
904 if (s1->is_Store()) return;
905
906 // reductions are always managed beforehand
907 if (s1->is_reduction()) return;
908
909 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
910 Node* t1 = s1->fast_out(i);
911
912 // Only allow operand swap on commuting operations
913 if (!t1->is_Add() && !t1->is_Mul()) {
914 break;
915 }
916
917 // Now find t1's packset
918 Node_List* p2 = NULL;
919 for (int j = 0; j < _packset.length(); j++) {
920 p2 = _packset.at(j);
921 Node* first = p2->at(0);
922 if (t1 == first) {
923 break;
924 }
925 p2 = NULL;
926 }
927 // Arrange all sub components by the major component
928 if (p2 != NULL) {
929 for (uint j = 1; j < p->size(); j++) {
930 Node* d1 = p->at(j);
931 Node* u1 = p2->at(j);
932 opnd_positions_match(s1, t1, d1, u1);
933 }
934 }
935 }
936 }
937
938 //---------------------------opnd_positions_match-------------------------
939 // Is the use of d1 in u1 at the same operand position as d2 in u2?
940 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
941 // check reductions to see if they are marshalled to represent the reduction
942 // operator in a specified opnd
943 if (u1->is_reduction() && u2->is_reduction()) {
944 // ensure reductions have phis and reduction definitions feeding the 1st operand
945 Node* first = u1->in(2);
946 if (first->is_Phi() || first->is_reduction()) {
947 u1->swap_edges(1, 2);
948 }
949 // ensure reductions have phis and reduction definitions feeding the 1st operand
950 first = u2->in(2);
951 if (first->is_Phi() || first->is_reduction()) {
952 u2->swap_edges(1, 2);
953 }
954 return true;
955 }
956
957 uint ct = u1->req();
958 if (ct != u2->req()) return false;
959 uint i1 = 0;
960 uint i2 = 0;
961 do {
962 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
963 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
964 if (i1 != i2) {
965 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
966 // Further analysis relies on operands position matching.
967 u2->swap_edges(i1, i2);
968 } else {
969 return false;
970 }
971 }
972 } while (i1 < ct);
973 return true;
974 }
975
976 //------------------------------est_savings---------------------------
977 // Estimate the savings from executing s1 and s2 as a pack
978 int SuperWord::est_savings(Node* s1, Node* s2) {
979 int save_in = 2 - 1; // 2 operations per instruction in packed form
980
981 // inputs
982 for (uint i = 1; i < s1->req(); i++) {
983 Node* x1 = s1->in(i);
984 Node* x2 = s2->in(i);
985 if (x1 != x2) {
986 if (are_adjacent_refs(x1, x2)) {
987 save_in += adjacent_profit(x1, x2);
988 } else if (!in_packset(x1, x2)) {
989 save_in -= pack_cost(2);
990 } else {
991 save_in += unpack_cost(2);
992 }
993 }
994 }
995
996 // uses of result
997 uint ct = 0;
998 int save_use = 0;
999 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1000 Node* s1_use = s1->fast_out(i);
1001 for (int j = 0; j < _packset.length(); j++) {
1002 Node_List* p = _packset.at(j);
1003 if (p->at(0) == s1_use) {
1004 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1005 Node* s2_use = s2->fast_out(k);
1006 if (p->at(p->size()-1) == s2_use) {
1007 ct++;
1008 if (are_adjacent_refs(s1_use, s2_use)) {
1009 save_use += adjacent_profit(s1_use, s2_use);
1010 }
1011 }
1012 }
1013 }
1014 }
1015 }
1016
1017 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1018 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1019
1020 return MAX2(save_in, save_use);
1021 }
1022
1023 //------------------------------costs---------------------------
1024 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1025 int SuperWord::pack_cost(int ct) { return ct; }
1026 int SuperWord::unpack_cost(int ct) { return ct; }
1027
1028 //------------------------------combine_packs---------------------------
1029 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1030 void SuperWord::combine_packs() {
1031 bool changed = true;
1032 // Combine packs regardless max vector size.
1033 while (changed) {
1034 changed = false;
1035 for (int i = 0; i < _packset.length(); i++) {
1036 Node_List* p1 = _packset.at(i);
1037 if (p1 == NULL) continue;
1038 // Because of sorting we can start at i + 1
1039 for (int j = i + 1; j < _packset.length(); j++) {
1040 Node_List* p2 = _packset.at(j);
1041 if (p2 == NULL) continue;
1042 if (i == j) continue;
1043 if (p1->at(p1->size()-1) == p2->at(0)) {
1044 for (uint k = 1; k < p2->size(); k++) {
1045 p1->push(p2->at(k));
1046 }
1047 _packset.at_put(j, NULL);
1048 changed = true;
1049 }
1050 }
1051 }
1052 }
1053
1054 // Split packs which have size greater then max vector size.
1055 for (int i = 0; i < _packset.length(); i++) {
1056 Node_List* p1 = _packset.at(i);
1057 if (p1 != NULL) {
1058 BasicType bt = velt_basic_type(p1->at(0));
1059 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1060 assert(is_power_of_2(max_vlen), "sanity");
1061 uint psize = p1->size();
1062 if (!is_power_of_2(psize)) {
1063 // Skip pack which can't be vector.
1064 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1065 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1066 _packset.at_put(i, NULL);
1067 continue;
1068 }
1069 if (psize > max_vlen) {
1070 Node_List* pack = new Node_List();
1071 for (uint j = 0; j < psize; j++) {
1072 pack->push(p1->at(j));
1073 if (pack->size() >= max_vlen) {
1074 assert(is_power_of_2(pack->size()), "sanity");
1075 _packset.append(pack);
1076 pack = new Node_List();
1077 }
1078 }
1079 _packset.at_put(i, NULL);
1080 }
1081 }
1082 }
1083
1084 // Compress list.
1085 for (int i = _packset.length() - 1; i >= 0; i--) {
1086 Node_List* p1 = _packset.at(i);
1087 if (p1 == NULL) {
1088 _packset.remove_at(i);
1089 }
1090 }
1091
1092 #ifndef PRODUCT
1093 if (TraceSuperWord) {
1094 tty->print_cr("\nAfter combine_packs");
1095 print_packset();
1096 }
1097 #endif
1098 }
1099
1100 //-----------------------------construct_my_pack_map--------------------------
1101 // Construct the map from nodes to packs. Only valid after the
1102 // point where a node is only in one pack (after combine_packs).
1103 void SuperWord::construct_my_pack_map() {
1104 Node_List* rslt = NULL;
1105 for (int i = 0; i < _packset.length(); i++) {
1106 Node_List* p = _packset.at(i);
1107 for (uint j = 0; j < p->size(); j++) {
1108 Node* s = p->at(j);
1109 assert(my_pack(s) == NULL, "only in one pack");
1110 set_my_pack(s, p);
1111 }
1112 }
1113 }
1114
1115 //------------------------------filter_packs---------------------------
1116 // Remove packs that are not implemented or not profitable.
1117 void SuperWord::filter_packs() {
1118
1119 // Remove packs that are not implemented
1120 for (int i = _packset.length() - 1; i >= 0; i--) {
1121 Node_List* pk = _packset.at(i);
1122 bool impl = implemented(pk);
1123 if (!impl) {
1124 #ifndef PRODUCT
1125 if (TraceSuperWord && Verbose) {
1126 tty->print_cr("Unimplemented");
1127 pk->at(0)->dump();
1128 }
1129 #endif
1130 remove_pack_at(i);
1131 }
1132 }
1133
1134 // Remove packs that are not profitable
1135 bool changed;
1136 do {
1137 changed = false;
1138 for (int i = _packset.length() - 1; i >= 0; i--) {
1139 Node_List* pk = _packset.at(i);
1140 bool prof = profitable(pk);
1141 if (!prof) {
1142 #ifndef PRODUCT
1143 if (TraceSuperWord && Verbose) {
1144 tty->print_cr("Unprofitable");
1145 pk->at(0)->dump();
1146 }
1147 #endif
1148 remove_pack_at(i);
1149 changed = true;
1150 }
1151 }
1152 } while (changed);
1153
1154 #ifndef PRODUCT
1155 if (TraceSuperWord) {
1156 tty->print_cr("\nAfter filter_packs");
1157 print_packset();
1158 tty->cr();
1159 }
1160 #endif
1161 }
1162
1163 //------------------------------implemented---------------------------
1164 // Can code be generated for pack p?
1165 bool SuperWord::implemented(Node_List* p) {
1166 bool retValue = false;
1167 Node* p0 = p->at(0);
1168 if (p0 != NULL) {
1169 int opc = p0->Opcode();
1170 uint size = p->size();
1171 if (p0->is_reduction()) {
1172 const Type *arith_type = p0->bottom_type();
1173 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1174 } else {
1175 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1176 }
1177 }
1178 return retValue;
1179 }
1180
1181 //------------------------------same_inputs--------------------------
1182 // For pack p, are all idx operands the same?
1183 static bool same_inputs(Node_List* p, int idx) {
1184 Node* p0 = p->at(0);
1185 uint vlen = p->size();
1186 Node* p0_def = p0->in(idx);
1187 for (uint i = 1; i < vlen; i++) {
1188 Node* pi = p->at(i);
1189 Node* pi_def = pi->in(idx);
1190 if (p0_def != pi_def)
1191 return false;
1192 }
1193 return true;
1194 }
1195
1196 //------------------------------profitable---------------------------
1197 // For pack p, are all operands and all uses (with in the block) vector?
1198 bool SuperWord::profitable(Node_List* p) {
1199 Node* p0 = p->at(0);
1200 uint start, end;
1201 VectorNode::vector_operands(p0, &start, &end);
1202
1203 // Return false if some inputs are not vectors or vectors with different
1204 // size or alignment.
1205 // Also, for now, return false if not scalar promotion case when inputs are
1206 // the same. Later, implement PackNode and allow differing, non-vector inputs
1207 // (maybe just the ones from outside the block.)
1208 for (uint i = start; i < end; i++) {
1209 if (!is_vector_use(p0, i))
1210 return false;
1211 }
1212 // Check if reductions are connected
1213 if (p0->is_reduction()) {
1214 Node* second_in = p0->in(2);
1215 Node_List* second_pk = my_pack(second_in);
1216 if (second_pk == NULL) {
1217 // Remove reduction flag if no parent pack, it is not profitable
1218 p0->remove_flag(Node::Flag_is_reduction);
1219 return false;
1220 } else if (second_pk->size() != p->size()) {
1221 return false;
1222 }
1223 }
1224 if (VectorNode::is_shift(p0)) {
1225 // For now, return false if shift count is vector or not scalar promotion
1226 // case (different shift counts) because it is not supported yet.
1227 Node* cnt = p0->in(2);
1228 Node_List* cnt_pk = my_pack(cnt);
1229 if (cnt_pk != NULL)
1230 return false;
1231 if (!same_inputs(p, 2))
1232 return false;
1233 }
1234 if (!p0->is_Store()) {
1235 // For now, return false if not all uses are vector.
1236 // Later, implement ExtractNode and allow non-vector uses (maybe
1237 // just the ones outside the block.)
1238 for (uint i = 0; i < p->size(); i++) {
1239 Node* def = p->at(i);
1240 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1241 Node* use = def->fast_out(j);
1242 for (uint k = 0; k < use->req(); k++) {
1243 Node* n = use->in(k);
1244 if (def == n) {
1245 // reductions can be loop carried dependences
1246 if (def->is_reduction() && use->is_Phi())
1247 continue;
1248 if (!is_vector_use(use, k)) {
1249 return false;
1250 }
1251 }
1252 }
1253 }
1254 }
1255 }
1256 return true;
1257 }
1258
1259 //------------------------------schedule---------------------------
1260 // Adjust the memory graph for the packed operations
1261 void SuperWord::schedule() {
1262
1263 // Co-locate in the memory graph the members of each memory pack
1264 for (int i = 0; i < _packset.length(); i++) {
1265 co_locate_pack(_packset.at(i));
1266 }
1267 }
1268
1269 //-------------------------------remove_and_insert-------------------
1270 // Remove "current" from its current position in the memory graph and insert
1271 // it after the appropriate insertion point (lip or uip).
1272 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1273 Node *uip, Unique_Node_List &sched_before) {
1274 Node* my_mem = current->in(MemNode::Memory);
1275 bool sched_up = sched_before.member(current);
1276
1277 // remove current_store from its current position in the memmory graph
1278 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1279 Node* use = current->out(i);
1280 if (use->is_Mem()) {
1281 assert(use->in(MemNode::Memory) == current, "must be");
1282 if (use == prev) { // connect prev to my_mem
1283 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1284 --i; //deleted this edge; rescan position
1285 } else if (sched_before.member(use)) {
1286 if (!sched_up) { // Will be moved together with current
1287 _igvn.replace_input_of(use, MemNode::Memory, uip);
1288 --i; //deleted this edge; rescan position
1289 }
1290 } else {
1291 if (sched_up) { // Will be moved together with current
1292 _igvn.replace_input_of(use, MemNode::Memory, lip);
1293 --i; //deleted this edge; rescan position
1294 }
1295 }
1296 }
1297 }
1298
1299 Node *insert_pt = sched_up ? uip : lip;
1300
1301 // all uses of insert_pt's memory state should use current's instead
1302 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1303 Node* use = insert_pt->out(i);
1304 if (use->is_Mem()) {
1305 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1306 _igvn.replace_input_of(use, MemNode::Memory, current);
1307 --i; //deleted this edge; rescan position
1308 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1309 uint pos; //lip (lower insert point) must be the last one in the memory slice
1310 for (pos=1; pos < use->req(); pos++) {
1311 if (use->in(pos) == insert_pt) break;
1312 }
1313 _igvn.replace_input_of(use, pos, current);
1314 --i;
1315 }
1316 }
1317
1318 //connect current to insert_pt
1319 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1320 }
1321
1322 //------------------------------co_locate_pack----------------------------------
1323 // To schedule a store pack, we need to move any sandwiched memory ops either before
1324 // or after the pack, based upon dependence information:
1325 // (1) If any store in the pack depends on the sandwiched memory op, the
1326 // sandwiched memory op must be scheduled BEFORE the pack;
1327 // (2) If a sandwiched memory op depends on any store in the pack, the
1328 // sandwiched memory op must be scheduled AFTER the pack;
1329 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1330 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1331 // scheduled before the pack, memB must also be scheduled before the pack;
1332 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1333 // schedule this store AFTER the pack
1334 // (5) We know there is no dependence cycle, so there in no other case;
1335 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1336 //
1337 // To schedule a load pack, we use the memory state of either the first or the last load in
1338 // the pack, based on the dependence constraint.
1339 void SuperWord::co_locate_pack(Node_List* pk) {
1340 if (pk->at(0)->is_Store()) {
1341 MemNode* first = executed_first(pk)->as_Mem();
1342 MemNode* last = executed_last(pk)->as_Mem();
1343 Unique_Node_List schedule_before_pack;
1344 Unique_Node_List memops;
1345
1346 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1347 MemNode* previous = last;
1348 while (true) {
1349 assert(in_bb(current), "stay in block");
1350 memops.push(previous);
1351 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1352 Node* use = current->out(i);
1353 if (use->is_Mem() && use != previous)
1354 memops.push(use);
1355 }
1356 if (current == first) break;
1357 previous = current;
1358 current = current->in(MemNode::Memory)->as_Mem();
1359 }
1360
1361 // determine which memory operations should be scheduled before the pack
1362 for (uint i = 1; i < memops.size(); i++) {
1363 Node *s1 = memops.at(i);
1364 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1365 for (uint j = 0; j< i; j++) {
1366 Node *s2 = memops.at(j);
1367 if (!independent(s1, s2)) {
1368 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1369 schedule_before_pack.push(s1); // s1 must be scheduled before
1370 Node_List* mem_pk = my_pack(s1);
1371 if (mem_pk != NULL) {
1372 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1373 Node* s = mem_pk->at(ii); // follow partner
1374 if (memops.member(s) && !schedule_before_pack.member(s))
1375 schedule_before_pack.push(s);
1376 }
1377 }
1378 break;
1379 }
1380 }
1381 }
1382 }
1383 }
1384
1385 Node* upper_insert_pt = first->in(MemNode::Memory);
1386 // Following code moves loads connected to upper_insert_pt below aliased stores.
1387 // Collect such loads here and reconnect them back to upper_insert_pt later.
1388 memops.clear();
1389 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1390 Node* use = upper_insert_pt->out(i);
1391 if (use->is_Mem() && !use->is_Store()) {
1392 memops.push(use);
1393 }
1394 }
1395
1396 MemNode* lower_insert_pt = last;
1397 previous = last; //previous store in pk
1398 current = last->in(MemNode::Memory)->as_Mem();
1399
1400 // start scheduling from "last" to "first"
1401 while (true) {
1402 assert(in_bb(current), "stay in block");
1403 assert(in_pack(previous, pk), "previous stays in pack");
1404 Node* my_mem = current->in(MemNode::Memory);
1405
1406 if (in_pack(current, pk)) {
1407 // Forward users of my memory state (except "previous) to my input memory state
1408 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1409 Node* use = current->out(i);
1410 if (use->is_Mem() && use != previous) {
1411 assert(use->in(MemNode::Memory) == current, "must be");
1412 if (schedule_before_pack.member(use)) {
1413 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1414 } else {
1415 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1416 }
1417 --i; // deleted this edge; rescan position
1418 }
1419 }
1420 previous = current;
1421 } else { // !in_pack(current, pk) ==> a sandwiched store
1422 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1423 }
1424
1425 if (current == first) break;
1426 current = my_mem->as_Mem();
1427 } // end while
1428
1429 // Reconnect loads back to upper_insert_pt.
1430 for (uint i = 0; i < memops.size(); i++) {
1431 Node *ld = memops.at(i);
1432 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1433 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1434 }
1435 }
1436 } else if (pk->at(0)->is_Load()) { //load
1437 // all loads in the pack should have the same memory state. By default,
1438 // we use the memory state of the last load. However, if any load could
1439 // not be moved down due to the dependence constraint, we use the memory
1440 // state of the first load.
1441 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1442 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1443 bool schedule_last = true;
1444 for (uint i = 0; i < pk->size(); i++) {
1445 Node* ld = pk->at(i);
1446 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1447 current=current->in(MemNode::Memory)) {
1448 assert(current != first_mem, "corrupted memory graph");
1449 if(current->is_Mem() && !independent(current, ld)){
1450 schedule_last = false; // a later store depends on this load
1451 break;
1452 }
1453 }
1454 }
1455
1456 Node* mem_input = schedule_last ? last_mem : first_mem;
1457 _igvn.hash_delete(mem_input);
1458 // Give each load the same memory state
1459 for (uint i = 0; i < pk->size(); i++) {
1460 LoadNode* ld = pk->at(i)->as_Load();
1461 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1462 }
1463 }
1464 }
1465
1466 //------------------------------output---------------------------
1467 // Convert packs into vector node operations
1468 void SuperWord::output() {
1469 if (_packset.length() == 0) return;
1470
1471 #ifndef PRODUCT
1472 if (TraceLoopOpts) {
1473 tty->print("SuperWord ");
1474 lpt()->dump_head();
1475 }
1476 #endif
1477
1478 // MUST ENSURE main loop's initial value is properly aligned:
1479 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1480
1481 align_initial_loop_index(align_to_ref());
1482
1483 // Insert extract (unpack) operations for scalar uses
1484 for (int i = 0; i < _packset.length(); i++) {
1485 insert_extracts(_packset.at(i));
1486 }
1487
1488 Compile* C = _phase->C;
1489 uint max_vlen_in_bytes = 0;
1490 for (int i = 0; i < _block.length(); i++) {
1491 Node* n = _block.at(i);
1492 Node_List* p = my_pack(n);
1493 if (p && n == executed_last(p)) {
1494 uint vlen = p->size();
1495 uint vlen_in_bytes = 0;
1496 Node* vn = NULL;
1497 Node* low_adr = p->at(0);
1498 Node* first = executed_first(p);
1499 int opc = n->Opcode();
1500 if (n->is_Load()) {
1501 Node* ctl = n->in(MemNode::Control);
1502 Node* mem = first->in(MemNode::Memory);
1503 SWPointer p1(n->as_Mem(), this, NULL, false);
1504 // Identify the memory dependency for the new loadVector node by
1505 // walking up through memory chain.
1506 // This is done to give flexibility to the new loadVector node so that
1507 // it can move above independent storeVector nodes.
1508 while (mem->is_StoreVector()) {
1509 SWPointer p2(mem->as_Mem(), this, NULL, false);
1510 int cmp = p1.cmp(p2);
1511 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1512 mem = mem->in(MemNode::Memory);
1513 } else {
1514 break; // dependent memory
1515 }
1516 }
1517 Node* adr = low_adr->in(MemNode::Address);
1518 const TypePtr* atyp = n->adr_type();
1519 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1520 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1521 } else if (n->is_Store()) {
1522 // Promote value to be stored to vector
1523 Node* val = vector_opd(p, MemNode::ValueIn);
1524 Node* ctl = n->in(MemNode::Control);
1525 Node* mem = first->in(MemNode::Memory);
1526 Node* adr = low_adr->in(MemNode::Address);
1527 const TypePtr* atyp = n->adr_type();
1528 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1529 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1530 } else if (n->req() == 3) {
1531 // Promote operands to vector
1532 Node* in1 = NULL;
1533 bool node_isa_reduction = n->is_reduction();
1534 if (node_isa_reduction) {
1535 // the input to the first reduction operation is retained
1536 in1 = low_adr->in(1);
1537 } else {
1538 in1 = vector_opd(p, 1);
1539 }
1540 Node* in2 = vector_opd(p, 2);
1541 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1542 // Move invariant vector input into second position to avoid register spilling.
1543 Node* tmp = in1;
1544 in1 = in2;
1545 in2 = tmp;
1546 }
1547 if (node_isa_reduction) {
1548 const Type *arith_type = n->bottom_type();
1549 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1550 if (in2->is_Load()) {
1551 vlen_in_bytes = in2->as_LoadVector()->memory_size();
1552 } else {
1553 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1554 }
1555 } else {
1556 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1557 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1558 }
1559 } else {
1560 ShouldNotReachHere();
1561 }
1562 assert(vn != NULL, "sanity");
1563 _igvn.register_new_node_with_optimizer(vn);
1564 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1565 for (uint j = 0; j < p->size(); j++) {
1566 Node* pm = p->at(j);
1567 _igvn.replace_node(pm, vn);
1568 }
1569 _igvn._worklist.push(vn);
1570
1571 if (vlen_in_bytes > max_vlen_in_bytes) {
1572 max_vlen_in_bytes = vlen_in_bytes;
1573 }
1574 #ifdef ASSERT
1575 if (TraceNewVectors) {
1576 tty->print("new Vector node: ");
1577 vn->dump();
1578 }
1579 #endif
1580 }
1581 }
1582 C->set_max_vector_size(max_vlen_in_bytes);
1583 }
1584
1585 //------------------------------vector_opd---------------------------
1586 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1587 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1588 Node* p0 = p->at(0);
1589 uint vlen = p->size();
1590 Node* opd = p0->in(opd_idx);
1591
1592 if (same_inputs(p, opd_idx)) {
1593 if (opd->is_Vector() || opd->is_LoadVector()) {
1594 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1595 return opd; // input is matching vector
1596 }
1597 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1598 Compile* C = _phase->C;
1599 Node* cnt = opd;
1600 // Vector instructions do not mask shift count, do it here.
1601 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1602 const TypeInt* t = opd->find_int_type();
1603 if (t != NULL && t->is_con()) {
1604 juint shift = t->get_con();
1605 if (shift > mask) { // Unsigned cmp
1606 cnt = ConNode::make(TypeInt::make(shift & mask));
1607 }
1608 } else {
1609 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1610 cnt = ConNode::make(TypeInt::make(mask));
1611 _igvn.register_new_node_with_optimizer(cnt);
1612 cnt = new AndINode(opd, cnt);
1613 _igvn.register_new_node_with_optimizer(cnt);
1614 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1615 }
1616 assert(opd->bottom_type()->isa_int(), "int type only");
1617 // Move non constant shift count into vector register.
1618 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1619 }
1620 if (cnt != opd) {
1621 _igvn.register_new_node_with_optimizer(cnt);
1622 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1623 }
1624 return cnt;
1625 }
1626 assert(!opd->is_StoreVector(), "such vector is not expected here");
1627 // Convert scalar input to vector with the same number of elements as
1628 // p0's vector. Use p0's type because size of operand's container in
1629 // vector should match p0's size regardless operand's size.
1630 const Type* p0_t = velt_type(p0);
1631 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1632
1633 _igvn.register_new_node_with_optimizer(vn);
1634 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1635 #ifdef ASSERT
1636 if (TraceNewVectors) {
1637 tty->print("new Vector node: ");
1638 vn->dump();
1639 }
1640 #endif
1641 return vn;
1642 }
1643
1644 // Insert pack operation
1645 BasicType bt = velt_basic_type(p0);
1646 PackNode* pk = PackNode::make(opd, vlen, bt);
1647 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1648
1649 for (uint i = 1; i < vlen; i++) {
1650 Node* pi = p->at(i);
1651 Node* in = pi->in(opd_idx);
1652 assert(my_pack(in) == NULL, "Should already have been unpacked");
1653 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1654 pk->add_opd(in);
1655 }
1656 _igvn.register_new_node_with_optimizer(pk);
1657 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1658 #ifdef ASSERT
1659 if (TraceNewVectors) {
1660 tty->print("new Vector node: ");
1661 pk->dump();
1662 }
1663 #endif
1664 return pk;
1665 }
1666
1667 //------------------------------insert_extracts---------------------------
1668 // If a use of pack p is not a vector use, then replace the
1669 // use with an extract operation.
1670 void SuperWord::insert_extracts(Node_List* p) {
1671 if (p->at(0)->is_Store()) return;
1672 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1673
1674 // Inspect each use of each pack member. For each use that is
1675 // not a vector use, replace the use with an extract operation.
1676
1677 for (uint i = 0; i < p->size(); i++) {
1678 Node* def = p->at(i);
1679 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1680 Node* use = def->fast_out(j);
1681 for (uint k = 0; k < use->req(); k++) {
1682 Node* n = use->in(k);
1683 if (def == n) {
1684 if (!is_vector_use(use, k)) {
1685 _n_idx_list.push(use, k);
1686 }
1687 }
1688 }
1689 }
1690 }
1691
1692 while (_n_idx_list.is_nonempty()) {
1693 Node* use = _n_idx_list.node();
1694 int idx = _n_idx_list.index();
1695 _n_idx_list.pop();
1696 Node* def = use->in(idx);
1697
1698 if (def->is_reduction()) continue;
1699
1700 // Insert extract operation
1701 _igvn.hash_delete(def);
1702 int def_pos = alignment(def) / data_size(def);
1703
1704 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1705 _igvn.register_new_node_with_optimizer(ex);
1706 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1707 _igvn.replace_input_of(use, idx, ex);
1708 _igvn._worklist.push(def);
1709
1710 bb_insert_after(ex, bb_idx(def));
1711 set_velt_type(ex, velt_type(def));
1712 }
1713 }
1714
1715 //------------------------------is_vector_use---------------------------
1716 // Is use->in(u_idx) a vector use?
1717 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1718 Node_List* u_pk = my_pack(use);
1719 if (u_pk == NULL) return false;
1720 if (use->is_reduction()) return true;
1721 Node* def = use->in(u_idx);
1722 Node_List* d_pk = my_pack(def);
1723 if (d_pk == NULL) {
1724 // check for scalar promotion
1725 Node* n = u_pk->at(0)->in(u_idx);
1726 for (uint i = 1; i < u_pk->size(); i++) {
1727 if (u_pk->at(i)->in(u_idx) != n) return false;
1728 }
1729 return true;
1730 }
1731 if (u_pk->size() != d_pk->size())
1732 return false;
1733 for (uint i = 0; i < u_pk->size(); i++) {
1734 Node* ui = u_pk->at(i);
1735 Node* di = d_pk->at(i);
1736 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1737 return false;
1738 }
1739 return true;
1740 }
1741
1742 //------------------------------construct_bb---------------------------
1743 // Construct reverse postorder list of block members
1744 bool SuperWord::construct_bb() {
1745 Node* entry = bb();
1746
1747 assert(_stk.length() == 0, "stk is empty");
1748 assert(_block.length() == 0, "block is empty");
1749 assert(_data_entry.length() == 0, "data_entry is empty");
1750 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1751 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1752
1753 // Find non-control nodes with no inputs from within block,
1754 // create a temporary map from node _idx to bb_idx for use
1755 // by the visited and post_visited sets,
1756 // and count number of nodes in block.
1757 int bb_ct = 0;
1758 for (uint i = 0; i < lpt()->_body.size(); i++) {
1759 Node *n = lpt()->_body.at(i);
1760 set_bb_idx(n, i); // Create a temporary map
1761 if (in_bb(n)) {
1762 if (n->is_LoadStore() || n->is_MergeMem() ||
1763 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1764 // Bailout if the loop has LoadStore, MergeMem or data Proj
1765 // nodes. Superword optimization does not work with them.
1766 return false;
1767 }
1768 bb_ct++;
1769 if (!n->is_CFG()) {
1770 bool found = false;
1771 for (uint j = 0; j < n->req(); j++) {
1772 Node* def = n->in(j);
1773 if (def && in_bb(def)) {
1774 found = true;
1775 break;
1776 }
1777 }
1778 if (!found) {
1779 assert(n != entry, "can't be entry");
1780 _data_entry.push(n);
1781 }
1782 }
1783 }
1784 }
1785
1786 // Find memory slices (head and tail)
1787 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1788 Node *n = lp()->fast_out(i);
1789 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1790 Node* n_tail = n->in(LoopNode::LoopBackControl);
1791 if (n_tail != n->in(LoopNode::EntryControl)) {
1792 if (!n_tail->is_Mem()) {
1793 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1794 return false; // Bailout
1795 }
1796 _mem_slice_head.push(n);
1797 _mem_slice_tail.push(n_tail);
1798 }
1799 }
1800 }
1801
1802 // Create an RPO list of nodes in block
1803
1804 visited_clear();
1805 post_visited_clear();
1806
1807 // Push all non-control nodes with no inputs from within block, then control entry
1808 for (int j = 0; j < _data_entry.length(); j++) {
1809 Node* n = _data_entry.at(j);
1810 visited_set(n);
1811 _stk.push(n);
1812 }
1813 visited_set(entry);
1814 _stk.push(entry);
1815
1816 // Do a depth first walk over out edges
1817 int rpo_idx = bb_ct - 1;
1818 int size;
1819 int reduction_uses = 0;
1820 while ((size = _stk.length()) > 0) {
1821 Node* n = _stk.top(); // Leave node on stack
1822 if (!visited_test_set(n)) {
1823 // forward arc in graph
1824 } else if (!post_visited_test(n)) {
1825 // cross or back arc
1826 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1827 Node *use = n->fast_out(i);
1828 if (in_bb(use) && !visited_test(use) &&
1829 // Don't go around backedge
1830 (!use->is_Phi() || n == entry)) {
1831 if (use->is_reduction()) {
1832 // First see if we can map the reduction on the given system we are on, then
1833 // make a data entry operation for each reduction we see.
1834 BasicType bt = use->bottom_type()->basic_type();
1835 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
1836 reduction_uses++;
1837 }
1838 }
1839 _stk.push(use);
1840 }
1841 }
1842 if (_stk.length() == size) {
1843 // There were no additional uses, post visit node now
1844 _stk.pop(); // Remove node from stack
1845 assert(rpo_idx >= 0, "");
1846 _block.at_put_grow(rpo_idx, n);
1847 rpo_idx--;
1848 post_visited_set(n);
1849 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1850 }
1851 } else {
1852 _stk.pop(); // Remove post-visited node from stack
1853 }
1854 }
1855
1856 // Create real map of block indices for nodes
1857 for (int j = 0; j < _block.length(); j++) {
1858 Node* n = _block.at(j);
1859 set_bb_idx(n, j);
1860 }
1861
1862 // Ensure extra info is allocated.
1863 initialize_bb();
1864
1865 #ifndef PRODUCT
1866 if (TraceSuperWord) {
1867 print_bb();
1868 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1869 for (int m = 0; m < _data_entry.length(); m++) {
1870 tty->print("%3d ", m);
1871 _data_entry.at(m)->dump();
1872 }
1873 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1874 for (int m = 0; m < _mem_slice_head.length(); m++) {
1875 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1876 tty->print(" "); _mem_slice_tail.at(m)->dump();
1877 }
1878 }
1879 #endif
1880 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1881 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
1882 }
1883
1884 //------------------------------initialize_bb---------------------------
1885 // Initialize per node info
1886 void SuperWord::initialize_bb() {
1887 Node* last = _block.at(_block.length() - 1);
1888 grow_node_info(bb_idx(last));
1889 }
1890
1891 //------------------------------bb_insert_after---------------------------
1892 // Insert n into block after pos
1893 void SuperWord::bb_insert_after(Node* n, int pos) {
1894 int n_pos = pos + 1;
1895 // Make room
1896 for (int i = _block.length() - 1; i >= n_pos; i--) {
1897 _block.at_put_grow(i+1, _block.at(i));
1898 }
1899 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1900 _node_info.at_put_grow(j+1, _node_info.at(j));
1901 }
1902 // Set value
1903 _block.at_put_grow(n_pos, n);
1904 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1905 // Adjust map from node->_idx to _block index
1906 for (int i = n_pos; i < _block.length(); i++) {
1907 set_bb_idx(_block.at(i), i);
1908 }
1909 }
1910
1911 //------------------------------compute_max_depth---------------------------
1912 // Compute max depth for expressions from beginning of block
1913 // Use to prune search paths during test for independence.
1914 void SuperWord::compute_max_depth() {
1915 int ct = 0;
1916 bool again;
1917 do {
1918 again = false;
1919 for (int i = 0; i < _block.length(); i++) {
1920 Node* n = _block.at(i);
1921 if (!n->is_Phi()) {
1922 int d_orig = depth(n);
1923 int d_in = 0;
1924 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1925 Node* pred = preds.current();
1926 if (in_bb(pred)) {
1927 d_in = MAX2(d_in, depth(pred));
1928 }
1929 }
1930 if (d_in + 1 != d_orig) {
1931 set_depth(n, d_in + 1);
1932 again = true;
1933 }
1934 }
1935 }
1936 ct++;
1937 } while (again);
1938 #ifndef PRODUCT
1939 if (TraceSuperWord && Verbose)
1940 tty->print_cr("compute_max_depth iterated: %d times", ct);
1941 #endif
1942 }
1943
1944 //-------------------------compute_vector_element_type-----------------------
1945 // Compute necessary vector element type for expressions
1946 // This propagates backwards a narrower integer type when the
1947 // upper bits of the value are not needed.
1948 // Example: char a,b,c; a = b + c;
1949 // Normally the type of the add is integer, but for packed character
1950 // operations the type of the add needs to be char.
1951 void SuperWord::compute_vector_element_type() {
1952 #ifndef PRODUCT
1953 if (TraceSuperWord && Verbose)
1954 tty->print_cr("\ncompute_velt_type:");
1955 #endif
1956
1957 // Initial type
1958 for (int i = 0; i < _block.length(); i++) {
1959 Node* n = _block.at(i);
1960 set_velt_type(n, container_type(n));
1961 }
1962
1963 // Propagate integer narrowed type backwards through operations
1964 // that don't depend on higher order bits
1965 for (int i = _block.length() - 1; i >= 0; i--) {
1966 Node* n = _block.at(i);
1967 // Only integer types need be examined
1968 const Type* vtn = velt_type(n);
1969 if (vtn->basic_type() == T_INT) {
1970 uint start, end;
1971 VectorNode::vector_operands(n, &start, &end);
1972
1973 for (uint j = start; j < end; j++) {
1974 Node* in = n->in(j);
1975 // Don't propagate through a memory
1976 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1977 data_size(n) < data_size(in)) {
1978 bool same_type = true;
1979 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1980 Node *use = in->fast_out(k);
1981 if (!in_bb(use) || !same_velt_type(use, n)) {
1982 same_type = false;
1983 break;
1984 }
1985 }
1986 if (same_type) {
1987 // For right shifts of small integer types (bool, byte, char, short)
1988 // we need precise information about sign-ness. Only Load nodes have
1989 // this information because Store nodes are the same for signed and
1990 // unsigned values. And any arithmetic operation after a load may
1991 // expand a value to signed Int so such right shifts can't be used
1992 // because vector elements do not have upper bits of Int.
1993 const Type* vt = vtn;
1994 if (VectorNode::is_shift(in)) {
1995 Node* load = in->in(1);
1996 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1997 vt = velt_type(load);
1998 } else if (in->Opcode() != Op_LShiftI) {
1999 // Widen type to Int to avoid creation of right shift vector
2000 // (align + data_size(s1) check in stmts_can_pack() will fail).
2001 // Note, left shifts work regardless type.
2002 vt = TypeInt::INT;
2003 }
2004 }
2005 set_velt_type(in, vt);
2006 }
2007 }
2008 }
2009 }
2010 }
2011 #ifndef PRODUCT
2012 if (TraceSuperWord && Verbose) {
2013 for (int i = 0; i < _block.length(); i++) {
2014 Node* n = _block.at(i);
2015 velt_type(n)->dump();
2016 tty->print("\t");
2017 n->dump();
2018 }
2019 }
2020 #endif
2021 }
2022
2023 //------------------------------memory_alignment---------------------------
2024 // Alignment within a vector memory reference
2025 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2026 SWPointer p(s, this, NULL, false);
2027 if (!p.valid()) {
2028 return bottom_align;
2029 }
2030 int vw = vector_width_in_bytes(s);
2031 if (vw < 2) {
2032 return bottom_align; // No vectors for this type
2033 }
2034 int offset = p.offset_in_bytes();
2035 offset += iv_adjust*p.memory_size();
2036 int off_rem = offset % vw;
2037 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2038 return off_mod;
2039 }
2040
2041 //---------------------------container_type---------------------------
2042 // Smallest type containing range of values
2043 const Type* SuperWord::container_type(Node* n) {
2044 if (n->is_Mem()) {
2045 BasicType bt = n->as_Mem()->memory_type();
2046 if (n->is_Store() && (bt == T_CHAR)) {
2047 // Use T_SHORT type instead of T_CHAR for stored values because any
2048 // preceding arithmetic operation extends values to signed Int.
2049 bt = T_SHORT;
2050 }
2051 if (n->Opcode() == Op_LoadUB) {
2052 // Adjust type for unsigned byte loads, it is important for right shifts.
2053 // T_BOOLEAN is used because there is no basic type representing type
2054 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2055 // size (one byte) and sign is important.
2056 bt = T_BOOLEAN;
2057 }
2058 return Type::get_const_basic_type(bt);
2059 }
2060 const Type* t = _igvn.type(n);
2061 if (t->basic_type() == T_INT) {
2062 // A narrow type of arithmetic operations will be determined by
2063 // propagating the type of memory operations.
2064 return TypeInt::INT;
2065 }
2066 return t;
2067 }
2068
2069 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2070 const Type* vt1 = velt_type(n1);
2071 const Type* vt2 = velt_type(n2);
2072 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2073 // Compare vectors element sizes for integer types.
2074 return data_size(n1) == data_size(n2);
2075 }
2076 return vt1 == vt2;
2077 }
2078
2079 //------------------------------in_packset---------------------------
2080 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2081 bool SuperWord::in_packset(Node* s1, Node* s2) {
2082 for (int i = 0; i < _packset.length(); i++) {
2083 Node_List* p = _packset.at(i);
2084 assert(p->size() == 2, "must be");
2085 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2086 return true;
2087 }
2088 }
2089 return false;
2090 }
2091
2092 //------------------------------in_pack---------------------------
2093 // Is s in pack p?
2094 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2095 for (uint i = 0; i < p->size(); i++) {
2096 if (p->at(i) == s) {
2097 return p;
2098 }
2099 }
2100 return NULL;
2101 }
2102
2103 //------------------------------remove_pack_at---------------------------
2104 // Remove the pack at position pos in the packset
2105 void SuperWord::remove_pack_at(int pos) {
2106 Node_List* p = _packset.at(pos);
2107 for (uint i = 0; i < p->size(); i++) {
2108 Node* s = p->at(i);
2109 set_my_pack(s, NULL);
2110 }
2111 _packset.remove_at(pos);
2112 }
2113
2114 void SuperWord::packset_sort(int n) {
2115 // simple bubble sort so that we capitalize with O(n) when its already sorted
2116 while (n != 0) {
2117 bool swapped = false;
2118 for (int i = 1; i < n; i++) {
2119 Node_List* q_low = _packset.at(i-1);
2120 Node_List* q_i = _packset.at(i);
2121
2122 // only swap when we find something to swap
2123 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2124 Node_List* t = q_i;
2125 *(_packset.adr_at(i)) = q_low;
2126 *(_packset.adr_at(i-1)) = q_i;
2127 swapped = true;
2128 }
2129 }
2130 if (swapped == false) break;
2131 n--;
2132 }
2133 }
2134
2135 //------------------------------executed_first---------------------------
2136 // Return the node executed first in pack p. Uses the RPO block list
2137 // to determine order.
2138 Node* SuperWord::executed_first(Node_List* p) {
2139 Node* n = p->at(0);
2140 int n_rpo = bb_idx(n);
2141 for (uint i = 1; i < p->size(); i++) {
2142 Node* s = p->at(i);
2143 int s_rpo = bb_idx(s);
2144 if (s_rpo < n_rpo) {
2145 n = s;
2146 n_rpo = s_rpo;
2147 }
2148 }
2149 return n;
2150 }
2151
2152 //------------------------------executed_last---------------------------
2153 // Return the node executed last in pack p.
2154 Node* SuperWord::executed_last(Node_List* p) {
2155 Node* n = p->at(0);
2156 int n_rpo = bb_idx(n);
2157 for (uint i = 1; i < p->size(); i++) {
2158 Node* s = p->at(i);
2159 int s_rpo = bb_idx(s);
2160 if (s_rpo > n_rpo) {
2161 n = s;
2162 n_rpo = s_rpo;
2163 }
2164 }
2165 return n;
2166 }
2167
2168 //----------------------------align_initial_loop_index---------------------------
2169 // Adjust pre-loop limit so that in main loop, a load/store reference
2170 // to align_to_ref will be a position zero in the vector.
2171 // (iv + k) mod vector_align == 0
2172 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2173 CountedLoopNode *main_head = lp()->as_CountedLoop();
2174 assert(main_head->is_main_loop(), "");
2175 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2176 assert(pre_end != NULL, "we must have a correct pre-loop");
2177 Node *pre_opaq1 = pre_end->limit();
2178 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2179 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2180 Node *lim0 = pre_opaq->in(1);
2181
2182 // Where we put new limit calculations
2183 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2184
2185 // Ensure the original loop limit is available from the
2186 // pre-loop Opaque1 node.
2187 Node *orig_limit = pre_opaq->original_loop_limit();
2188 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2189
2190 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2191 assert(align_to_ref_p.valid(), "sanity");
2192
2193 // Given:
2194 // lim0 == original pre loop limit
2195 // V == v_align (power of 2)
2196 // invar == extra invariant piece of the address expression
2197 // e == offset [ +/- invar ]
2198 //
2199 // When reassociating expressions involving '%' the basic rules are:
2200 // (a - b) % k == 0 => a % k == b % k
2201 // and:
2202 // (a + b) % k == 0 => a % k == (k - b) % k
2203 //
2204 // For stride > 0 && scale > 0,
2205 // Derive the new pre-loop limit "lim" such that the two constraints:
2206 // (1) lim = lim0 + N (where N is some positive integer < V)
2207 // (2) (e + lim) % V == 0
2208 // are true.
2209 //
2210 // Substituting (1) into (2),
2211 // (e + lim0 + N) % V == 0
2212 // solve for N:
2213 // N = (V - (e + lim0)) % V
2214 // substitute back into (1), so that new limit
2215 // lim = lim0 + (V - (e + lim0)) % V
2216 //
2217 // For stride > 0 && scale < 0
2218 // Constraints:
2219 // lim = lim0 + N
2220 // (e - lim) % V == 0
2221 // Solving for lim:
2222 // (e - lim0 - N) % V == 0
2223 // N = (e - lim0) % V
2224 // lim = lim0 + (e - lim0) % V
2225 //
2226 // For stride < 0 && scale > 0
2227 // Constraints:
2228 // lim = lim0 - N
2229 // (e + lim) % V == 0
2230 // Solving for lim:
2231 // (e + lim0 - N) % V == 0
2232 // N = (e + lim0) % V
2233 // lim = lim0 - (e + lim0) % V
2234 //
2235 // For stride < 0 && scale < 0
2236 // Constraints:
2237 // lim = lim0 - N
2238 // (e - lim) % V == 0
2239 // Solving for lim:
2240 // (e - lim0 + N) % V == 0
2241 // N = (V - (e - lim0)) % V
2242 // lim = lim0 - (V - (e - lim0)) % V
2243
2244 int vw = vector_width_in_bytes(align_to_ref);
2245 int stride = iv_stride();
2246 int scale = align_to_ref_p.scale_in_bytes();
2247 int elt_size = align_to_ref_p.memory_size();
2248 int v_align = vw / elt_size;
2249 assert(v_align > 1, "sanity");
2250 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2251 Node *offsn = _igvn.intcon(offset);
2252
2253 Node *e = offsn;
2254 if (align_to_ref_p.invar() != NULL) {
2255 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2256 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2257 Node* aref = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2258 _igvn.register_new_node_with_optimizer(aref);
2259 _phase->set_ctrl(aref, pre_ctrl);
2260 if (align_to_ref_p.negate_invar()) {
2261 e = new SubINode(e, aref);
2262 } else {
2263 e = new AddINode(e, aref);
2264 }
2265 _igvn.register_new_node_with_optimizer(e);
2266 _phase->set_ctrl(e, pre_ctrl);
2267 }
2268 if (vw > ObjectAlignmentInBytes) {
2269 // incorporate base e +/- base && Mask >>> log2(elt)
2270 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2271 _igvn.register_new_node_with_optimizer(xbase);
2272 #ifdef _LP64
2273 xbase = new ConvL2INode(xbase);
2274 _igvn.register_new_node_with_optimizer(xbase);
2275 #endif
2276 Node* mask = _igvn.intcon(vw-1);
2277 Node* masked_xbase = new AndINode(xbase, mask);
2278 _igvn.register_new_node_with_optimizer(masked_xbase);
2279 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2280 Node* bref = new URShiftINode(masked_xbase, log2_elt);
2281 _igvn.register_new_node_with_optimizer(bref);
2282 _phase->set_ctrl(bref, pre_ctrl);
2283 e = new AddINode(e, bref);
2284 _igvn.register_new_node_with_optimizer(e);
2285 _phase->set_ctrl(e, pre_ctrl);
2286 }
2287
2288 // compute e +/- lim0
2289 if (scale < 0) {
2290 e = new SubINode(e, lim0);
2291 } else {
2292 e = new AddINode(e, lim0);
2293 }
2294 _igvn.register_new_node_with_optimizer(e);
2295 _phase->set_ctrl(e, pre_ctrl);
2296
2297 if (stride * scale > 0) {
2298 // compute V - (e +/- lim0)
2299 Node* va = _igvn.intcon(v_align);
2300 e = new SubINode(va, e);
2301 _igvn.register_new_node_with_optimizer(e);
2302 _phase->set_ctrl(e, pre_ctrl);
2303 }
2304 // compute N = (exp) % V
2305 Node* va_msk = _igvn.intcon(v_align - 1);
2306 Node* N = new AndINode(e, va_msk);
2307 _igvn.register_new_node_with_optimizer(N);
2308 _phase->set_ctrl(N, pre_ctrl);
2309
2310 // substitute back into (1), so that new limit
2311 // lim = lim0 + N
2312 Node* lim;
2313 if (stride < 0) {
2314 lim = new SubINode(lim0, N);
2315 } else {
2316 lim = new AddINode(lim0, N);
2317 }
2318 _igvn.register_new_node_with_optimizer(lim);
2319 _phase->set_ctrl(lim, pre_ctrl);
2320 Node* constrained =
2321 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2322 : (Node*) new MaxINode(lim, orig_limit);
2323 _igvn.register_new_node_with_optimizer(constrained);
2324 _phase->set_ctrl(constrained, pre_ctrl);
2325 _igvn.hash_delete(pre_opaq);
2326 pre_opaq->set_req(1, constrained);
2327 }
2328
2329 //----------------------------get_pre_loop_end---------------------------
2330 // Find pre loop end from main loop. Returns null if none.
2331 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2332 Node *ctrl = cl->in(LoopNode::EntryControl);
2333 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2334 Node *iffm = ctrl->in(0);
2335 if (!iffm->is_If()) return NULL;
2336 Node *p_f = iffm->in(0);
2337 if (!p_f->is_IfFalse()) return NULL;
2338 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2339 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2340 CountedLoopNode* loop_node = pre_end->loopnode();
2341 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2342 return pre_end;
2343 }
2344
2345
2346 //------------------------------init---------------------------
2347 void SuperWord::init() {
2348 _dg.init();
2349 _packset.clear();
2350 _disjoint_ptrs.clear();
2351 _block.clear();
2352 _data_entry.clear();
2353 _mem_slice_head.clear();
2354 _mem_slice_tail.clear();
2355 _node_info.clear();
2356 _align_to_ref = NULL;
2357 _lpt = NULL;
2358 _lp = NULL;
2359 _bb = NULL;
2360 _iv = NULL;
2361 _early_return = false;
2362 }
2363
2364 //------------------------------print_packset---------------------------
2365 void SuperWord::print_packset() {
2366 #ifndef PRODUCT
2367 tty->print_cr("packset");
2368 for (int i = 0; i < _packset.length(); i++) {
2369 tty->print_cr("Pack: %d", i);
2370 Node_List* p = _packset.at(i);
2371 print_pack(p);
2372 }
2373 #endif
2374 }
2375
2376 //------------------------------print_pack---------------------------
2377 void SuperWord::print_pack(Node_List* p) {
2378 for (uint i = 0; i < p->size(); i++) {
2379 print_stmt(p->at(i));
2380 }
2381 }
2382
2383 //------------------------------print_bb---------------------------
2384 void SuperWord::print_bb() {
2385 #ifndef PRODUCT
2386 tty->print_cr("\nBlock");
2387 for (int i = 0; i < _block.length(); i++) {
2388 Node* n = _block.at(i);
2389 tty->print("%d ", i);
2390 if (n) {
2391 n->dump();
2392 }
2393 }
2394 #endif
2395 }
2396
2397 //------------------------------print_stmt---------------------------
2398 void SuperWord::print_stmt(Node* s) {
2399 #ifndef PRODUCT
2400 tty->print(" align: %d \t", alignment(s));
2401 s->dump();
2402 #endif
2403 }
2404
2405 //------------------------------blank---------------------------
2406 char* SuperWord::blank(uint depth) {
2407 static char blanks[101];
2408 assert(depth < 101, "too deep");
2409 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2410 blanks[depth] = '\0';
2411 return blanks;
2412 }
2413
2414
2415 //==============================SWPointer===========================
2416
2417 //----------------------------SWPointer------------------------
2418 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
2419 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2420 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2421 _nstack(nstack), _analyze_only(analyze_only),
2422 _stack_idx(0) {
2423
2424 Node* adr = mem->in(MemNode::Address);
2425 if (!adr->is_AddP()) {
2426 assert(!valid(), "too complex");
2427 return;
2428 }
2429 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2430 Node* base = adr->in(AddPNode::Base);
2431 //unsafe reference could not be aligned appropriately without runtime checking
2432 if (base == NULL || base->bottom_type() == Type::TOP) {
2433 assert(!valid(), "unsafe access");
2434 return;
2435 }
2436 for (int i = 0; i < 3; i++) {
2437 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2438 assert(!valid(), "too complex");
2439 return;
2440 }
2441 adr = adr->in(AddPNode::Address);
2442 if (base == adr || !adr->is_AddP()) {
2443 break; // stop looking at addp's
2444 }
2445 }
2446 _base = base;
2447 _adr = adr;
2448 assert(valid(), "Usable");
2449 }
2450
2451 // Following is used to create a temporary object during
2452 // the pattern match of an address expression.
2453 SWPointer::SWPointer(SWPointer* p) :
2454 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2455 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2456 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
2457 _stack_idx(p->_stack_idx) {}
2458
2459 //------------------------scaled_iv_plus_offset--------------------
2460 // Match: k*iv + offset
2461 // where: k is a constant that maybe zero, and
2462 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2463 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2464 if (scaled_iv(n)) {
2465 return true;
2466 }
2467 if (offset_plus_k(n)) {
2468 return true;
2469 }
2470 int opc = n->Opcode();
2471 if (opc == Op_AddI) {
2472 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2473 return true;
2474 }
2475 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2476 return true;
2477 }
2478 } else if (opc == Op_SubI) {
2479 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2480 return true;
2481 }
2482 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2483 _scale *= -1;
2484 return true;
2485 }
2486 }
2487 return false;
2488 }
2489
2490 //----------------------------scaled_iv------------------------
2491 // Match: k*iv where k is a constant that's not zero
2492 bool SWPointer::scaled_iv(Node* n) {
2493 if (_scale != 0) {
2494 return false; // already found a scale
2495 }
2496 if (n == iv()) {
2497 _scale = 1;
2498 return true;
2499 }
2500 if (_analyze_only && (invariant(n) == false)) {
2501 _nstack->push(n, _stack_idx++);
2502 }
2503 int opc = n->Opcode();
2504 if (opc == Op_MulI) {
2505 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2506 _scale = n->in(2)->get_int();
2507 return true;
2508 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2509 _scale = n->in(1)->get_int();
2510 return true;
2511 }
2512 } else if (opc == Op_LShiftI) {
2513 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2514 _scale = 1 << n->in(2)->get_int();
2515 return true;
2516 }
2517 } else if (opc == Op_ConvI2L) {
2518 if (scaled_iv_plus_offset(n->in(1))) {
2519 return true;
2520 }
2521 } else if (opc == Op_LShiftL) {
2522 if (!has_iv() && _invar == NULL) {
2523 // Need to preserve the current _offset value, so
2524 // create a temporary object for this expression subtree.
2525 // Hacky, so should re-engineer the address pattern match.
2526 SWPointer tmp(this);
2527 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2528 if (tmp._invar == NULL) {
2529 int mult = 1 << n->in(2)->get_int();
2530 _scale = tmp._scale * mult;
2531 _offset += tmp._offset * mult;
2532 return true;
2533 }
2534 }
2535 }
2536 }
2537 return false;
2538 }
2539
2540 //----------------------------offset_plus_k------------------------
2541 // Match: offset is (k [+/- invariant])
2542 // where k maybe zero and invariant is optional, but not both.
2543 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2544 int opc = n->Opcode();
2545 if (opc == Op_ConI) {
2546 _offset += negate ? -(n->get_int()) : n->get_int();
2547 return true;
2548 } else if (opc == Op_ConL) {
2549 // Okay if value fits into an int
2550 const TypeLong* t = n->find_long_type();
2551 if (t->higher_equal(TypeLong::INT)) {
2552 jlong loff = n->get_long();
2553 jint off = (jint)loff;
2554 _offset += negate ? -off : loff;
2555 return true;
2556 }
2557 return false;
2558 }
2559 if (_invar != NULL) return false; // already have an invariant
2560 if (_analyze_only && (invariant(n) == false)) {
2561 _nstack->push(n, _stack_idx++);
2562 }
2563 if (opc == Op_AddI) {
2564 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2565 _negate_invar = negate;
2566 _invar = n->in(1);
2567 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2568 return true;
2569 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2570 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2571 _negate_invar = negate;
2572 _invar = n->in(2);
2573 return true;
2574 }
2575 }
2576 if (opc == Op_SubI) {
2577 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2578 _negate_invar = negate;
2579 _invar = n->in(1);
2580 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2581 return true;
2582 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2583 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2584 _negate_invar = !negate;
2585 _invar = n->in(2);
2586 return true;
2587 }
2588 }
2589 if (invariant(n)) {
2590 _negate_invar = negate;
2591 _invar = n;
2592 return true;
2593 }
2594 return false;
2595 }
2596
2597 //----------------------------print------------------------
2598 void SWPointer::print() {
2599 #ifndef PRODUCT
2600 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2601 _base != NULL ? _base->_idx : 0,
2602 _adr != NULL ? _adr->_idx : 0,
2603 _scale, _offset,
2604 _negate_invar?'-':'+',
2605 _invar != NULL ? _invar->_idx : 0);
2606 #endif
2607 }
2608
2609 // ========================= OrderedPair =====================
2610
2611 const OrderedPair OrderedPair::initial;
2612
2613 // ========================= SWNodeInfo =====================
2614
2615 const SWNodeInfo SWNodeInfo::initial;
2616
2617
2618 // ============================ DepGraph ===========================
2619
2620 //------------------------------make_node---------------------------
2621 // Make a new dependence graph node for an ideal node.
2622 DepMem* DepGraph::make_node(Node* node) {
2623 DepMem* m = new (_arena) DepMem(node);
2624 if (node != NULL) {
2625 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2626 _map.at_put_grow(node->_idx, m);
2627 }
2628 return m;
2629 }
2630
2631 //------------------------------make_edge---------------------------
2632 // Make a new dependence graph edge from dpred -> dsucc
2633 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2634 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2635 dpred->set_out_head(e);
2636 dsucc->set_in_head(e);
2637 return e;
2638 }
2639
2640 // ========================== DepMem ========================
2641
2642 //------------------------------in_cnt---------------------------
2643 int DepMem::in_cnt() {
2644 int ct = 0;
2645 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2646 return ct;
2647 }
2648
2649 //------------------------------out_cnt---------------------------
2650 int DepMem::out_cnt() {
2651 int ct = 0;
2652 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2653 return ct;
2654 }
2655
2656 //------------------------------print-----------------------------
2657 void DepMem::print() {
2658 #ifndef PRODUCT
2659 tty->print(" DepNode %d (", _node->_idx);
2660 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2661 Node* pred = p->pred()->node();
2662 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2663 }
2664 tty->print(") [");
2665 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2666 Node* succ = s->succ()->node();
2667 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2668 }
2669 tty->print_cr(" ]");
2670 #endif
2671 }
2672
2673 // =========================== DepEdge =========================
2674
2675 //------------------------------DepPreds---------------------------
2676 void DepEdge::print() {
2677 #ifndef PRODUCT
2678 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2679 #endif
2680 }
2681
2682 // =========================== DepPreds =========================
2683 // Iterator over predecessor edges in the dependence graph.
2684
2685 //------------------------------DepPreds---------------------------
2686 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2687 _n = n;
2688 _done = false;
2689 if (_n->is_Store() || _n->is_Load()) {
2690 _next_idx = MemNode::Address;
2691 _end_idx = n->req();
2692 _dep_next = dg.dep(_n)->in_head();
2693 } else if (_n->is_Mem()) {
2694 _next_idx = 0;
2695 _end_idx = 0;
2696 _dep_next = dg.dep(_n)->in_head();
2697 } else {
2698 _next_idx = 1;
2699 _end_idx = _n->req();
2700 _dep_next = NULL;
2701 }
2702 next();
2703 }
2704
2705 //------------------------------next---------------------------
2706 void DepPreds::next() {
2707 if (_dep_next != NULL) {
2708 _current = _dep_next->pred()->node();
2709 _dep_next = _dep_next->next_in();
2710 } else if (_next_idx < _end_idx) {
2711 _current = _n->in(_next_idx++);
2712 } else {
2713 _done = true;
2714 }
2715 }
2716
2717 // =========================== DepSuccs =========================
2718 // Iterator over successor edges in the dependence graph.
2719
2720 //------------------------------DepSuccs---------------------------
2721 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2722 _n = n;
2723 _done = false;
2724 if (_n->is_Load()) {
2725 _next_idx = 0;
2726 _end_idx = _n->outcnt();
2727 _dep_next = dg.dep(_n)->out_head();
2728 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2729 _next_idx = 0;
2730 _end_idx = 0;
2731 _dep_next = dg.dep(_n)->out_head();
2732 } else {
2733 _next_idx = 0;
2734 _end_idx = _n->outcnt();
2735 _dep_next = NULL;
2736 }
2737 next();
2738 }
2739
2740 //-------------------------------next---------------------------
2741 void DepSuccs::next() {
2742 if (_dep_next != NULL) {
2743 _current = _dep_next->succ()->node();
2744 _dep_next = _dep_next->next_out();
2745 } else if (_next_idx < _end_idx) {
2746 _current = _n->raw_out(_next_idx++);
2747 } else {
2748 _done = true;
2749 }
2750 }