Mon, 11 May 2015 07:44:46 +0200
8079343: Crash in PhaseIdealLoop with "assert(!had_error) failed: bad dominance"
Summary: C2 should not try to vectorize loops with loop variant vector base address.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2007, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
24 #include "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/divnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/memnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/superword.hpp"
36 #include "opto/vectornode.hpp"
38 //
39 // S U P E R W O R D T R A N S F O R M
40 //=============================================================================
42 //------------------------------SuperWord---------------------------
43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
44 _phase(phase),
45 _igvn(phase->_igvn),
46 _arena(phase->C->comp_arena()),
47 _packset(arena(), 8, 0, NULL), // packs for the current block
48 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
49 _block(arena(), 8, 0, NULL), // nodes in current block
50 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
51 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
52 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
53 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
54 _align_to_ref(NULL), // memory reference to align vectors to
55 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
56 _dg(_arena), // dependence graph
57 _visited(arena()), // visited node set
58 _post_visited(arena()), // post visited node set
59 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
60 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
61 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
62 _lpt(NULL), // loop tree node
63 _lp(NULL), // LoopNode
64 _bb(NULL), // basic block
65 _iv(NULL) // induction var
66 {}
68 //------------------------------transform_loop---------------------------
69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
70 assert(UseSuperWord, "should be");
71 // Do vectors exist on this architecture?
72 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
74 assert(lpt->_head->is_CountedLoop(), "must be");
75 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
77 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
79 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
81 // Check for no control flow in body (other than exit)
82 Node *cl_exit = cl->loopexit();
83 if (cl_exit->in(0) != lpt->_head) return;
85 // Make sure the are no extra control users of the loop backedge
86 if (cl->back_control()->outcnt() != 1) {
87 return;
88 }
90 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
91 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
92 if (pre_end == NULL) return;
93 Node *pre_opaq1 = pre_end->limit();
94 if (pre_opaq1->Opcode() != Op_Opaque1) return;
96 init(); // initialize data structures
98 set_lpt(lpt);
99 set_lp(cl);
101 // For now, define one block which is the entire loop body
102 set_bb(cl);
104 assert(_packset.length() == 0, "packset must be empty");
105 SLP_extract();
106 }
108 //------------------------------SLP_extract---------------------------
109 // Extract the superword level parallelism
110 //
111 // 1) A reverse post-order of nodes in the block is constructed. By scanning
112 // this list from first to last, all definitions are visited before their uses.
113 //
114 // 2) A point-to-point dependence graph is constructed between memory references.
115 // This simplies the upcoming "independence" checker.
116 //
117 // 3) The maximum depth in the node graph from the beginning of the block
118 // to each node is computed. This is used to prune the graph search
119 // in the independence checker.
120 //
121 // 4) For integer types, the necessary bit width is propagated backwards
122 // from stores to allow packed operations on byte, char, and short
123 // integers. This reverses the promotion to type "int" that javac
124 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
125 //
126 // 5) One of the memory references is picked to be an aligned vector reference.
127 // The pre-loop trip count is adjusted to align this reference in the
128 // unrolled body.
129 //
130 // 6) The initial set of pack pairs is seeded with memory references.
131 //
132 // 7) The set of pack pairs is extended by following use->def and def->use links.
133 //
134 // 8) The pairs are combined into vector sized packs.
135 //
136 // 9) Reorder the memory slices to co-locate members of the memory packs.
137 //
138 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
139 // inserting scalar promotion, vector creation from multiple scalars, and
140 // extraction of scalar values from vectors.
141 //
142 void SuperWord::SLP_extract() {
144 // Ready the block
146 if (!construct_bb())
147 return; // Exit if no interesting nodes or complex graph.
149 dependence_graph();
151 compute_max_depth();
153 compute_vector_element_type();
155 // Attempt vectorization
157 find_adjacent_refs();
159 extend_packlist();
161 combine_packs();
163 construct_my_pack_map();
165 filter_packs();
167 schedule();
169 output();
170 }
172 //------------------------------find_adjacent_refs---------------------------
173 // Find the adjacent memory references and create pack pairs for them.
174 // This is the initial set of packs that will then be extended by
175 // following use->def and def->use links. The align positions are
176 // assigned relative to the reference "align_to_ref"
177 void SuperWord::find_adjacent_refs() {
178 // Get list of memory operations
179 Node_List memops;
180 for (int i = 0; i < _block.length(); i++) {
181 Node* n = _block.at(i);
182 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
183 is_java_primitive(n->as_Mem()->memory_type())) {
184 int align = memory_alignment(n->as_Mem(), 0);
185 if (align != bottom_align) {
186 memops.push(n);
187 }
188 }
189 }
191 Node_List align_to_refs;
192 int best_iv_adjustment = 0;
193 MemNode* best_align_to_mem_ref = NULL;
195 while (memops.size() != 0) {
196 // Find a memory reference to align to.
197 MemNode* mem_ref = find_align_to_ref(memops);
198 if (mem_ref == NULL) break;
199 align_to_refs.push(mem_ref);
200 int iv_adjustment = get_iv_adjustment(mem_ref);
202 if (best_align_to_mem_ref == NULL) {
203 // Set memory reference which is the best from all memory operations
204 // to be used for alignment. The pre-loop trip count is modified to align
205 // this reference to a vector-aligned address.
206 best_align_to_mem_ref = mem_ref;
207 best_iv_adjustment = iv_adjustment;
208 }
210 SWPointer align_to_ref_p(mem_ref, this);
211 // Set alignment relative to "align_to_ref" for all related memory operations.
212 for (int i = memops.size() - 1; i >= 0; i--) {
213 MemNode* s = memops.at(i)->as_Mem();
214 if (isomorphic(s, mem_ref)) {
215 SWPointer p2(s, this);
216 if (p2.comparable(align_to_ref_p)) {
217 int align = memory_alignment(s, iv_adjustment);
218 set_alignment(s, align);
219 }
220 }
221 }
223 // Create initial pack pairs of memory operations for which
224 // alignment is set and vectors will be aligned.
225 bool create_pack = true;
226 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
227 if (!Matcher::misaligned_vectors_ok()) {
228 int vw = vector_width(mem_ref);
229 int vw_best = vector_width(best_align_to_mem_ref);
230 if (vw > vw_best) {
231 // Do not vectorize a memory access with more elements per vector
232 // if unaligned memory access is not allowed because number of
233 // iterations in pre-loop will be not enough to align it.
234 create_pack = false;
235 }
236 }
237 } else {
238 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
239 // Can't allow vectorization of unaligned memory accesses with the
240 // same type since it could be overlapped accesses to the same array.
241 create_pack = false;
242 } else {
243 // Allow independent (different type) unaligned memory operations
244 // if HW supports them.
245 if (!Matcher::misaligned_vectors_ok()) {
246 create_pack = false;
247 } else {
248 // Check if packs of the same memory type but
249 // with a different alignment were created before.
250 for (uint i = 0; i < align_to_refs.size(); i++) {
251 MemNode* mr = align_to_refs.at(i)->as_Mem();
252 if (same_velt_type(mr, mem_ref) &&
253 memory_alignment(mr, iv_adjustment) != 0)
254 create_pack = false;
255 }
256 }
257 }
258 }
259 if (create_pack) {
260 for (uint i = 0; i < memops.size(); i++) {
261 Node* s1 = memops.at(i);
262 int align = alignment(s1);
263 if (align == top_align) continue;
264 for (uint j = 0; j < memops.size(); j++) {
265 Node* s2 = memops.at(j);
266 if (alignment(s2) == top_align) continue;
267 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
268 if (stmts_can_pack(s1, s2, align)) {
269 Node_List* pair = new Node_List();
270 pair->push(s1);
271 pair->push(s2);
272 _packset.append(pair);
273 }
274 }
275 }
276 }
277 } else { // Don't create unaligned pack
278 // First, remove remaining memory ops of the same type from the list.
279 for (int i = memops.size() - 1; i >= 0; i--) {
280 MemNode* s = memops.at(i)->as_Mem();
281 if (same_velt_type(s, mem_ref)) {
282 memops.remove(i);
283 }
284 }
286 // Second, remove already constructed packs of the same type.
287 for (int i = _packset.length() - 1; i >= 0; i--) {
288 Node_List* p = _packset.at(i);
289 MemNode* s = p->at(0)->as_Mem();
290 if (same_velt_type(s, mem_ref)) {
291 remove_pack_at(i);
292 }
293 }
295 // If needed find the best memory reference for loop alignment again.
296 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
297 // Put memory ops from remaining packs back on memops list for
298 // the best alignment search.
299 uint orig_msize = memops.size();
300 for (int i = 0; i < _packset.length(); i++) {
301 Node_List* p = _packset.at(i);
302 MemNode* s = p->at(0)->as_Mem();
303 assert(!same_velt_type(s, mem_ref), "sanity");
304 memops.push(s);
305 }
306 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
307 if (best_align_to_mem_ref == NULL) break;
308 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
309 // Restore list.
310 while (memops.size() > orig_msize)
311 (void)memops.pop();
312 }
313 } // unaligned memory accesses
315 // Remove used mem nodes.
316 for (int i = memops.size() - 1; i >= 0; i--) {
317 MemNode* m = memops.at(i)->as_Mem();
318 if (alignment(m) != top_align) {
319 memops.remove(i);
320 }
321 }
323 } // while (memops.size() != 0
324 set_align_to_ref(best_align_to_mem_ref);
326 #ifndef PRODUCT
327 if (TraceSuperWord) {
328 tty->print_cr("\nAfter find_adjacent_refs");
329 print_packset();
330 }
331 #endif
332 }
334 //------------------------------find_align_to_ref---------------------------
335 // Find a memory reference to align the loop induction variable to.
336 // Looks first at stores then at loads, looking for a memory reference
337 // with the largest number of references similar to it.
338 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
339 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
341 // Count number of comparable memory ops
342 for (uint i = 0; i < memops.size(); i++) {
343 MemNode* s1 = memops.at(i)->as_Mem();
344 SWPointer p1(s1, this);
345 // Discard if pre loop can't align this reference
346 if (!ref_is_alignable(p1)) {
347 *cmp_ct.adr_at(i) = 0;
348 continue;
349 }
350 for (uint j = i+1; j < memops.size(); j++) {
351 MemNode* s2 = memops.at(j)->as_Mem();
352 if (isomorphic(s1, s2)) {
353 SWPointer p2(s2, this);
354 if (p1.comparable(p2)) {
355 (*cmp_ct.adr_at(i))++;
356 (*cmp_ct.adr_at(j))++;
357 }
358 }
359 }
360 }
362 // Find Store (or Load) with the greatest number of "comparable" references,
363 // biggest vector size, smallest data size and smallest iv offset.
364 int max_ct = 0;
365 int max_vw = 0;
366 int max_idx = -1;
367 int min_size = max_jint;
368 int min_iv_offset = max_jint;
369 for (uint j = 0; j < memops.size(); j++) {
370 MemNode* s = memops.at(j)->as_Mem();
371 if (s->is_Store()) {
372 int vw = vector_width_in_bytes(s);
373 assert(vw > 1, "sanity");
374 SWPointer p(s, this);
375 if (cmp_ct.at(j) > max_ct ||
376 cmp_ct.at(j) == max_ct &&
377 (vw > max_vw ||
378 vw == max_vw &&
379 (data_size(s) < min_size ||
380 data_size(s) == min_size &&
381 (p.offset_in_bytes() < min_iv_offset)))) {
382 max_ct = cmp_ct.at(j);
383 max_vw = vw;
384 max_idx = j;
385 min_size = data_size(s);
386 min_iv_offset = p.offset_in_bytes();
387 }
388 }
389 }
390 // If no stores, look at loads
391 if (max_ct == 0) {
392 for (uint j = 0; j < memops.size(); j++) {
393 MemNode* s = memops.at(j)->as_Mem();
394 if (s->is_Load()) {
395 int vw = vector_width_in_bytes(s);
396 assert(vw > 1, "sanity");
397 SWPointer p(s, this);
398 if (cmp_ct.at(j) > max_ct ||
399 cmp_ct.at(j) == max_ct &&
400 (vw > max_vw ||
401 vw == max_vw &&
402 (data_size(s) < min_size ||
403 data_size(s) == min_size &&
404 (p.offset_in_bytes() < min_iv_offset)))) {
405 max_ct = cmp_ct.at(j);
406 max_vw = vw;
407 max_idx = j;
408 min_size = data_size(s);
409 min_iv_offset = p.offset_in_bytes();
410 }
411 }
412 }
413 }
415 #ifdef ASSERT
416 if (TraceSuperWord && Verbose) {
417 tty->print_cr("\nVector memops after find_align_to_refs");
418 for (uint i = 0; i < memops.size(); i++) {
419 MemNode* s = memops.at(i)->as_Mem();
420 s->dump();
421 }
422 }
423 #endif
425 if (max_ct > 0) {
426 #ifdef ASSERT
427 if (TraceSuperWord) {
428 tty->print("\nVector align to node: ");
429 memops.at(max_idx)->as_Mem()->dump();
430 }
431 #endif
432 return memops.at(max_idx)->as_Mem();
433 }
434 return NULL;
435 }
437 //------------------------------ref_is_alignable---------------------------
438 // Can the preloop align the reference to position zero in the vector?
439 bool SuperWord::ref_is_alignable(SWPointer& p) {
440 if (!p.has_iv()) {
441 return true; // no induction variable
442 }
443 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
444 assert(pre_end->stride_is_con(), "pre loop stride is constant");
445 int preloop_stride = pre_end->stride_con();
447 int span = preloop_stride * p.scale_in_bytes();
448 int mem_size = p.memory_size();
449 int offset = p.offset_in_bytes();
450 // Stride one accesses are alignable if offset is aligned to memory operation size.
451 // Offset can be unaligned when UseUnalignedAccesses is used.
452 if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
453 return true;
454 }
455 // If initial offset from start of object is computable,
456 // compute alignment within the vector.
457 int vw = vector_width_in_bytes(p.mem());
458 assert(vw > 1, "sanity");
459 if (vw % span == 0) {
460 Node* init_nd = pre_end->init_trip();
461 if (init_nd->is_Con() && p.invar() == NULL) {
462 int init = init_nd->bottom_type()->is_int()->get_con();
464 int init_offset = init * p.scale_in_bytes() + offset;
465 assert(init_offset >= 0, "positive offset from object start");
467 if (span > 0) {
468 return (vw - (init_offset % vw)) % span == 0;
469 } else {
470 assert(span < 0, "nonzero stride * scale");
471 return (init_offset % vw) % -span == 0;
472 }
473 }
474 }
475 return false;
476 }
478 //---------------------------get_iv_adjustment---------------------------
479 // Calculate loop's iv adjustment for this memory ops.
480 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
481 SWPointer align_to_ref_p(mem_ref, this);
482 int offset = align_to_ref_p.offset_in_bytes();
483 int scale = align_to_ref_p.scale_in_bytes();
484 int vw = vector_width_in_bytes(mem_ref);
485 assert(vw > 1, "sanity");
486 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
487 // At least one iteration is executed in pre-loop by default. As result
488 // several iterations are needed to align memory operations in main-loop even
489 // if offset is 0.
490 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
491 int elt_size = align_to_ref_p.memory_size();
492 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
493 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
494 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
496 #ifndef PRODUCT
497 if (TraceSuperWord)
498 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
499 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
500 #endif
501 return iv_adjustment;
502 }
504 //---------------------------dependence_graph---------------------------
505 // Construct dependency graph.
506 // Add dependence edges to load/store nodes for memory dependence
507 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
508 void SuperWord::dependence_graph() {
509 // First, assign a dependence node to each memory node
510 for (int i = 0; i < _block.length(); i++ ) {
511 Node *n = _block.at(i);
512 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
513 _dg.make_node(n);
514 }
515 }
517 // For each memory slice, create the dependences
518 for (int i = 0; i < _mem_slice_head.length(); i++) {
519 Node* n = _mem_slice_head.at(i);
520 Node* n_tail = _mem_slice_tail.at(i);
522 // Get slice in predecessor order (last is first)
523 mem_slice_preds(n_tail, n, _nlist);
525 // Make the slice dependent on the root
526 DepMem* slice = _dg.dep(n);
527 _dg.make_edge(_dg.root(), slice);
529 // Create a sink for the slice
530 DepMem* slice_sink = _dg.make_node(NULL);
531 _dg.make_edge(slice_sink, _dg.tail());
533 // Now visit each pair of memory ops, creating the edges
534 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
535 Node* s1 = _nlist.at(j);
537 // If no dependency yet, use slice
538 if (_dg.dep(s1)->in_cnt() == 0) {
539 _dg.make_edge(slice, s1);
540 }
541 SWPointer p1(s1->as_Mem(), this);
542 bool sink_dependent = true;
543 for (int k = j - 1; k >= 0; k--) {
544 Node* s2 = _nlist.at(k);
545 if (s1->is_Load() && s2->is_Load())
546 continue;
547 SWPointer p2(s2->as_Mem(), this);
549 int cmp = p1.cmp(p2);
550 if (SuperWordRTDepCheck &&
551 p1.base() != p2.base() && p1.valid() && p2.valid()) {
552 // Create a runtime check to disambiguate
553 OrderedPair pp(p1.base(), p2.base());
554 _disjoint_ptrs.append_if_missing(pp);
555 } else if (!SWPointer::not_equal(cmp)) {
556 // Possibly same address
557 _dg.make_edge(s1, s2);
558 sink_dependent = false;
559 }
560 }
561 if (sink_dependent) {
562 _dg.make_edge(s1, slice_sink);
563 }
564 }
565 #ifndef PRODUCT
566 if (TraceSuperWord) {
567 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
568 for (int q = 0; q < _nlist.length(); q++) {
569 _dg.print(_nlist.at(q));
570 }
571 tty->cr();
572 }
573 #endif
574 _nlist.clear();
575 }
577 #ifndef PRODUCT
578 if (TraceSuperWord) {
579 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
580 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
581 _disjoint_ptrs.at(r).print();
582 tty->cr();
583 }
584 tty->cr();
585 }
586 #endif
587 }
589 //---------------------------mem_slice_preds---------------------------
590 // Return a memory slice (node list) in predecessor order starting at "start"
591 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
592 assert(preds.length() == 0, "start empty");
593 Node* n = start;
594 Node* prev = NULL;
595 while (true) {
596 assert(in_bb(n), "must be in block");
597 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
598 Node* out = n->fast_out(i);
599 if (out->is_Load()) {
600 if (in_bb(out)) {
601 preds.push(out);
602 }
603 } else {
604 // FIXME
605 if (out->is_MergeMem() && !in_bb(out)) {
606 // Either unrolling is causing a memory edge not to disappear,
607 // or need to run igvn.optimize() again before SLP
608 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
609 // Ditto. Not sure what else to check further.
610 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
611 // StoreCM has an input edge used as a precedence edge.
612 // Maybe an issue when oop stores are vectorized.
613 } else {
614 assert(out == prev || prev == NULL, "no branches off of store slice");
615 }
616 }
617 }
618 if (n == stop) break;
619 preds.push(n);
620 prev = n;
621 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
622 n = n->in(MemNode::Memory);
623 }
624 }
626 //------------------------------stmts_can_pack---------------------------
627 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
628 // s1 aligned at "align"
629 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
631 // Do not use superword for non-primitives
632 BasicType bt1 = velt_basic_type(s1);
633 BasicType bt2 = velt_basic_type(s2);
634 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
635 return false;
636 if (Matcher::max_vector_size(bt1) < 2) {
637 return false; // No vectors for this type
638 }
640 if (isomorphic(s1, s2)) {
641 if (independent(s1, s2)) {
642 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
643 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
644 int s1_align = alignment(s1);
645 int s2_align = alignment(s2);
646 if (s1_align == top_align || s1_align == align) {
647 if (s2_align == top_align || s2_align == align + data_size(s1)) {
648 return true;
649 }
650 }
651 }
652 }
653 }
654 }
655 return false;
656 }
658 //------------------------------exists_at---------------------------
659 // Does s exist in a pack at position pos?
660 bool SuperWord::exists_at(Node* s, uint pos) {
661 for (int i = 0; i < _packset.length(); i++) {
662 Node_List* p = _packset.at(i);
663 if (p->at(pos) == s) {
664 return true;
665 }
666 }
667 return false;
668 }
670 //------------------------------are_adjacent_refs---------------------------
671 // Is s1 immediately before s2 in memory?
672 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
673 if (!s1->is_Mem() || !s2->is_Mem()) return false;
674 if (!in_bb(s1) || !in_bb(s2)) return false;
676 // Do not use superword for non-primitives
677 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
678 !is_java_primitive(s2->as_Mem()->memory_type())) {
679 return false;
680 }
682 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
683 // only pack memops that are in the same alias set until that's fixed.
684 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
685 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
686 return false;
687 SWPointer p1(s1->as_Mem(), this);
688 SWPointer p2(s2->as_Mem(), this);
689 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
690 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
691 return diff == data_size(s1);
692 }
694 //------------------------------isomorphic---------------------------
695 // Are s1 and s2 similar?
696 bool SuperWord::isomorphic(Node* s1, Node* s2) {
697 if (s1->Opcode() != s2->Opcode()) return false;
698 if (s1->req() != s2->req()) return false;
699 if (s1->in(0) != s2->in(0)) return false;
700 if (!same_velt_type(s1, s2)) return false;
701 return true;
702 }
704 //------------------------------independent---------------------------
705 // Is there no data path from s1 to s2 or s2 to s1?
706 bool SuperWord::independent(Node* s1, Node* s2) {
707 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
708 int d1 = depth(s1);
709 int d2 = depth(s2);
710 if (d1 == d2) return s1 != s2;
711 Node* deep = d1 > d2 ? s1 : s2;
712 Node* shallow = d1 > d2 ? s2 : s1;
714 visited_clear();
716 return independent_path(shallow, deep);
717 }
719 //------------------------------independent_path------------------------------
720 // Helper for independent
721 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
722 if (dp >= 1000) return false; // stop deep recursion
723 visited_set(deep);
724 int shal_depth = depth(shallow);
725 assert(shal_depth <= depth(deep), "must be");
726 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
727 Node* pred = preds.current();
728 if (in_bb(pred) && !visited_test(pred)) {
729 if (shallow == pred) {
730 return false;
731 }
732 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
733 return false;
734 }
735 }
736 }
737 return true;
738 }
740 //------------------------------set_alignment---------------------------
741 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
742 set_alignment(s1, align);
743 if (align == top_align || align == bottom_align) {
744 set_alignment(s2, align);
745 } else {
746 set_alignment(s2, align + data_size(s1));
747 }
748 }
750 //------------------------------data_size---------------------------
751 int SuperWord::data_size(Node* s) {
752 int bsize = type2aelembytes(velt_basic_type(s));
753 assert(bsize != 0, "valid size");
754 return bsize;
755 }
757 //------------------------------extend_packlist---------------------------
758 // Extend packset by following use->def and def->use links from pack members.
759 void SuperWord::extend_packlist() {
760 bool changed;
761 do {
762 changed = false;
763 for (int i = 0; i < _packset.length(); i++) {
764 Node_List* p = _packset.at(i);
765 changed |= follow_use_defs(p);
766 changed |= follow_def_uses(p);
767 }
768 } while (changed);
770 #ifndef PRODUCT
771 if (TraceSuperWord) {
772 tty->print_cr("\nAfter extend_packlist");
773 print_packset();
774 }
775 #endif
776 }
778 //------------------------------follow_use_defs---------------------------
779 // Extend the packset by visiting operand definitions of nodes in pack p
780 bool SuperWord::follow_use_defs(Node_List* p) {
781 assert(p->size() == 2, "just checking");
782 Node* s1 = p->at(0);
783 Node* s2 = p->at(1);
784 assert(s1->req() == s2->req(), "just checking");
785 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
787 if (s1->is_Load()) return false;
789 int align = alignment(s1);
790 bool changed = false;
791 int start = s1->is_Store() ? MemNode::ValueIn : 1;
792 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
793 for (int j = start; j < end; j++) {
794 Node* t1 = s1->in(j);
795 Node* t2 = s2->in(j);
796 if (!in_bb(t1) || !in_bb(t2))
797 continue;
798 if (stmts_can_pack(t1, t2, align)) {
799 if (est_savings(t1, t2) >= 0) {
800 Node_List* pair = new Node_List();
801 pair->push(t1);
802 pair->push(t2);
803 _packset.append(pair);
804 set_alignment(t1, t2, align);
805 changed = true;
806 }
807 }
808 }
809 return changed;
810 }
812 //------------------------------follow_def_uses---------------------------
813 // Extend the packset by visiting uses of nodes in pack p
814 bool SuperWord::follow_def_uses(Node_List* p) {
815 bool changed = false;
816 Node* s1 = p->at(0);
817 Node* s2 = p->at(1);
818 assert(p->size() == 2, "just checking");
819 assert(s1->req() == s2->req(), "just checking");
820 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
822 if (s1->is_Store()) return false;
824 int align = alignment(s1);
825 int savings = -1;
826 Node* u1 = NULL;
827 Node* u2 = NULL;
828 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
829 Node* t1 = s1->fast_out(i);
830 if (!in_bb(t1)) continue;
831 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
832 Node* t2 = s2->fast_out(j);
833 if (!in_bb(t2)) continue;
834 if (!opnd_positions_match(s1, t1, s2, t2))
835 continue;
836 if (stmts_can_pack(t1, t2, align)) {
837 int my_savings = est_savings(t1, t2);
838 if (my_savings > savings) {
839 savings = my_savings;
840 u1 = t1;
841 u2 = t2;
842 }
843 }
844 }
845 }
846 if (savings >= 0) {
847 Node_List* pair = new Node_List();
848 pair->push(u1);
849 pair->push(u2);
850 _packset.append(pair);
851 set_alignment(u1, u2, align);
852 changed = true;
853 }
854 return changed;
855 }
857 //---------------------------opnd_positions_match-------------------------
858 // Is the use of d1 in u1 at the same operand position as d2 in u2?
859 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
860 uint ct = u1->req();
861 if (ct != u2->req()) return false;
862 uint i1 = 0;
863 uint i2 = 0;
864 do {
865 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
866 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
867 if (i1 != i2) {
868 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
869 // Further analysis relies on operands position matching.
870 u2->swap_edges(i1, i2);
871 } else {
872 return false;
873 }
874 }
875 } while (i1 < ct);
876 return true;
877 }
879 //------------------------------est_savings---------------------------
880 // Estimate the savings from executing s1 and s2 as a pack
881 int SuperWord::est_savings(Node* s1, Node* s2) {
882 int save_in = 2 - 1; // 2 operations per instruction in packed form
884 // inputs
885 for (uint i = 1; i < s1->req(); i++) {
886 Node* x1 = s1->in(i);
887 Node* x2 = s2->in(i);
888 if (x1 != x2) {
889 if (are_adjacent_refs(x1, x2)) {
890 save_in += adjacent_profit(x1, x2);
891 } else if (!in_packset(x1, x2)) {
892 save_in -= pack_cost(2);
893 } else {
894 save_in += unpack_cost(2);
895 }
896 }
897 }
899 // uses of result
900 uint ct = 0;
901 int save_use = 0;
902 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
903 Node* s1_use = s1->fast_out(i);
904 for (int j = 0; j < _packset.length(); j++) {
905 Node_List* p = _packset.at(j);
906 if (p->at(0) == s1_use) {
907 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
908 Node* s2_use = s2->fast_out(k);
909 if (p->at(p->size()-1) == s2_use) {
910 ct++;
911 if (are_adjacent_refs(s1_use, s2_use)) {
912 save_use += adjacent_profit(s1_use, s2_use);
913 }
914 }
915 }
916 }
917 }
918 }
920 if (ct < s1->outcnt()) save_use += unpack_cost(1);
921 if (ct < s2->outcnt()) save_use += unpack_cost(1);
923 return MAX2(save_in, save_use);
924 }
926 //------------------------------costs---------------------------
927 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
928 int SuperWord::pack_cost(int ct) { return ct; }
929 int SuperWord::unpack_cost(int ct) { return ct; }
931 //------------------------------combine_packs---------------------------
932 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
933 void SuperWord::combine_packs() {
934 bool changed = true;
935 // Combine packs regardless max vector size.
936 while (changed) {
937 changed = false;
938 for (int i = 0; i < _packset.length(); i++) {
939 Node_List* p1 = _packset.at(i);
940 if (p1 == NULL) continue;
941 for (int j = 0; j < _packset.length(); j++) {
942 Node_List* p2 = _packset.at(j);
943 if (p2 == NULL) continue;
944 if (i == j) continue;
945 if (p1->at(p1->size()-1) == p2->at(0)) {
946 for (uint k = 1; k < p2->size(); k++) {
947 p1->push(p2->at(k));
948 }
949 _packset.at_put(j, NULL);
950 changed = true;
951 }
952 }
953 }
954 }
956 // Split packs which have size greater then max vector size.
957 for (int i = 0; i < _packset.length(); i++) {
958 Node_List* p1 = _packset.at(i);
959 if (p1 != NULL) {
960 BasicType bt = velt_basic_type(p1->at(0));
961 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
962 assert(is_power_of_2(max_vlen), "sanity");
963 uint psize = p1->size();
964 if (!is_power_of_2(psize)) {
965 // Skip pack which can't be vector.
966 // case1: for(...) { a[i] = i; } elements values are different (i+x)
967 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
968 _packset.at_put(i, NULL);
969 continue;
970 }
971 if (psize > max_vlen) {
972 Node_List* pack = new Node_List();
973 for (uint j = 0; j < psize; j++) {
974 pack->push(p1->at(j));
975 if (pack->size() >= max_vlen) {
976 assert(is_power_of_2(pack->size()), "sanity");
977 _packset.append(pack);
978 pack = new Node_List();
979 }
980 }
981 _packset.at_put(i, NULL);
982 }
983 }
984 }
986 // Compress list.
987 for (int i = _packset.length() - 1; i >= 0; i--) {
988 Node_List* p1 = _packset.at(i);
989 if (p1 == NULL) {
990 _packset.remove_at(i);
991 }
992 }
994 #ifndef PRODUCT
995 if (TraceSuperWord) {
996 tty->print_cr("\nAfter combine_packs");
997 print_packset();
998 }
999 #endif
1000 }
1002 //-----------------------------construct_my_pack_map--------------------------
1003 // Construct the map from nodes to packs. Only valid after the
1004 // point where a node is only in one pack (after combine_packs).
1005 void SuperWord::construct_my_pack_map() {
1006 Node_List* rslt = NULL;
1007 for (int i = 0; i < _packset.length(); i++) {
1008 Node_List* p = _packset.at(i);
1009 for (uint j = 0; j < p->size(); j++) {
1010 Node* s = p->at(j);
1011 assert(my_pack(s) == NULL, "only in one pack");
1012 set_my_pack(s, p);
1013 }
1014 }
1015 }
1017 //------------------------------filter_packs---------------------------
1018 // Remove packs that are not implemented or not profitable.
1019 void SuperWord::filter_packs() {
1021 // Remove packs that are not implemented
1022 for (int i = _packset.length() - 1; i >= 0; i--) {
1023 Node_List* pk = _packset.at(i);
1024 bool impl = implemented(pk);
1025 if (!impl) {
1026 #ifndef PRODUCT
1027 if (TraceSuperWord && Verbose) {
1028 tty->print_cr("Unimplemented");
1029 pk->at(0)->dump();
1030 }
1031 #endif
1032 remove_pack_at(i);
1033 }
1034 }
1036 // Remove packs that are not profitable
1037 bool changed;
1038 do {
1039 changed = false;
1040 for (int i = _packset.length() - 1; i >= 0; i--) {
1041 Node_List* pk = _packset.at(i);
1042 bool prof = profitable(pk);
1043 if (!prof) {
1044 #ifndef PRODUCT
1045 if (TraceSuperWord && Verbose) {
1046 tty->print_cr("Unprofitable");
1047 pk->at(0)->dump();
1048 }
1049 #endif
1050 remove_pack_at(i);
1051 changed = true;
1052 }
1053 }
1054 } while (changed);
1056 #ifndef PRODUCT
1057 if (TraceSuperWord) {
1058 tty->print_cr("\nAfter filter_packs");
1059 print_packset();
1060 tty->cr();
1061 }
1062 #endif
1063 }
1065 //------------------------------implemented---------------------------
1066 // Can code be generated for pack p?
1067 bool SuperWord::implemented(Node_List* p) {
1068 Node* p0 = p->at(0);
1069 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1070 }
1072 //------------------------------same_inputs--------------------------
1073 // For pack p, are all idx operands the same?
1074 static bool same_inputs(Node_List* p, int idx) {
1075 Node* p0 = p->at(0);
1076 uint vlen = p->size();
1077 Node* p0_def = p0->in(idx);
1078 for (uint i = 1; i < vlen; i++) {
1079 Node* pi = p->at(i);
1080 Node* pi_def = pi->in(idx);
1081 if (p0_def != pi_def)
1082 return false;
1083 }
1084 return true;
1085 }
1087 //------------------------------profitable---------------------------
1088 // For pack p, are all operands and all uses (with in the block) vector?
1089 bool SuperWord::profitable(Node_List* p) {
1090 Node* p0 = p->at(0);
1091 uint start, end;
1092 VectorNode::vector_operands(p0, &start, &end);
1094 // Return false if some inputs are not vectors or vectors with different
1095 // size or alignment.
1096 // Also, for now, return false if not scalar promotion case when inputs are
1097 // the same. Later, implement PackNode and allow differing, non-vector inputs
1098 // (maybe just the ones from outside the block.)
1099 for (uint i = start; i < end; i++) {
1100 if (!is_vector_use(p0, i))
1101 return false;
1102 }
1103 if (VectorNode::is_shift(p0)) {
1104 // For now, return false if shift count is vector or not scalar promotion
1105 // case (different shift counts) because it is not supported yet.
1106 Node* cnt = p0->in(2);
1107 Node_List* cnt_pk = my_pack(cnt);
1108 if (cnt_pk != NULL)
1109 return false;
1110 if (!same_inputs(p, 2))
1111 return false;
1112 }
1113 if (!p0->is_Store()) {
1114 // For now, return false if not all uses are vector.
1115 // Later, implement ExtractNode and allow non-vector uses (maybe
1116 // just the ones outside the block.)
1117 for (uint i = 0; i < p->size(); i++) {
1118 Node* def = p->at(i);
1119 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1120 Node* use = def->fast_out(j);
1121 for (uint k = 0; k < use->req(); k++) {
1122 Node* n = use->in(k);
1123 if (def == n) {
1124 if (!is_vector_use(use, k)) {
1125 return false;
1126 }
1127 }
1128 }
1129 }
1130 }
1131 }
1132 return true;
1133 }
1135 //------------------------------schedule---------------------------
1136 // Adjust the memory graph for the packed operations
1137 void SuperWord::schedule() {
1139 // Co-locate in the memory graph the members of each memory pack
1140 for (int i = 0; i < _packset.length(); i++) {
1141 co_locate_pack(_packset.at(i));
1142 }
1143 }
1145 //-------------------------------remove_and_insert-------------------
1146 // Remove "current" from its current position in the memory graph and insert
1147 // it after the appropriate insertion point (lip or uip).
1148 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1149 Node *uip, Unique_Node_List &sched_before) {
1150 Node* my_mem = current->in(MemNode::Memory);
1151 bool sched_up = sched_before.member(current);
1153 // remove current_store from its current position in the memmory graph
1154 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1155 Node* use = current->out(i);
1156 if (use->is_Mem()) {
1157 assert(use->in(MemNode::Memory) == current, "must be");
1158 if (use == prev) { // connect prev to my_mem
1159 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1160 --i; //deleted this edge; rescan position
1161 } else if (sched_before.member(use)) {
1162 if (!sched_up) { // Will be moved together with current
1163 _igvn.replace_input_of(use, MemNode::Memory, uip);
1164 --i; //deleted this edge; rescan position
1165 }
1166 } else {
1167 if (sched_up) { // Will be moved together with current
1168 _igvn.replace_input_of(use, MemNode::Memory, lip);
1169 --i; //deleted this edge; rescan position
1170 }
1171 }
1172 }
1173 }
1175 Node *insert_pt = sched_up ? uip : lip;
1177 // all uses of insert_pt's memory state should use current's instead
1178 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1179 Node* use = insert_pt->out(i);
1180 if (use->is_Mem()) {
1181 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1182 _igvn.replace_input_of(use, MemNode::Memory, current);
1183 --i; //deleted this edge; rescan position
1184 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1185 uint pos; //lip (lower insert point) must be the last one in the memory slice
1186 for (pos=1; pos < use->req(); pos++) {
1187 if (use->in(pos) == insert_pt) break;
1188 }
1189 _igvn.replace_input_of(use, pos, current);
1190 --i;
1191 }
1192 }
1194 //connect current to insert_pt
1195 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1196 }
1198 //------------------------------co_locate_pack----------------------------------
1199 // To schedule a store pack, we need to move any sandwiched memory ops either before
1200 // or after the pack, based upon dependence information:
1201 // (1) If any store in the pack depends on the sandwiched memory op, the
1202 // sandwiched memory op must be scheduled BEFORE the pack;
1203 // (2) If a sandwiched memory op depends on any store in the pack, the
1204 // sandwiched memory op must be scheduled AFTER the pack;
1205 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1206 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1207 // scheduled before the pack, memB must also be scheduled before the pack;
1208 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1209 // schedule this store AFTER the pack
1210 // (5) We know there is no dependence cycle, so there in no other case;
1211 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1212 //
1213 // To schedule a load pack, we use the memory state of either the first or the last load in
1214 // the pack, based on the dependence constraint.
1215 void SuperWord::co_locate_pack(Node_List* pk) {
1216 if (pk->at(0)->is_Store()) {
1217 MemNode* first = executed_first(pk)->as_Mem();
1218 MemNode* last = executed_last(pk)->as_Mem();
1219 Unique_Node_List schedule_before_pack;
1220 Unique_Node_List memops;
1222 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1223 MemNode* previous = last;
1224 while (true) {
1225 assert(in_bb(current), "stay in block");
1226 memops.push(previous);
1227 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1228 Node* use = current->out(i);
1229 if (use->is_Mem() && use != previous)
1230 memops.push(use);
1231 }
1232 if (current == first) break;
1233 previous = current;
1234 current = current->in(MemNode::Memory)->as_Mem();
1235 }
1237 // determine which memory operations should be scheduled before the pack
1238 for (uint i = 1; i < memops.size(); i++) {
1239 Node *s1 = memops.at(i);
1240 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1241 for (uint j = 0; j< i; j++) {
1242 Node *s2 = memops.at(j);
1243 if (!independent(s1, s2)) {
1244 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1245 schedule_before_pack.push(s1); // s1 must be scheduled before
1246 Node_List* mem_pk = my_pack(s1);
1247 if (mem_pk != NULL) {
1248 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1249 Node* s = mem_pk->at(ii); // follow partner
1250 if (memops.member(s) && !schedule_before_pack.member(s))
1251 schedule_before_pack.push(s);
1252 }
1253 }
1254 break;
1255 }
1256 }
1257 }
1258 }
1259 }
1261 Node* upper_insert_pt = first->in(MemNode::Memory);
1262 // Following code moves loads connected to upper_insert_pt below aliased stores.
1263 // Collect such loads here and reconnect them back to upper_insert_pt later.
1264 memops.clear();
1265 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1266 Node* use = upper_insert_pt->out(i);
1267 if (use->is_Mem() && !use->is_Store()) {
1268 memops.push(use);
1269 }
1270 }
1272 MemNode* lower_insert_pt = last;
1273 previous = last; //previous store in pk
1274 current = last->in(MemNode::Memory)->as_Mem();
1276 // start scheduling from "last" to "first"
1277 while (true) {
1278 assert(in_bb(current), "stay in block");
1279 assert(in_pack(previous, pk), "previous stays in pack");
1280 Node* my_mem = current->in(MemNode::Memory);
1282 if (in_pack(current, pk)) {
1283 // Forward users of my memory state (except "previous) to my input memory state
1284 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1285 Node* use = current->out(i);
1286 if (use->is_Mem() && use != previous) {
1287 assert(use->in(MemNode::Memory) == current, "must be");
1288 if (schedule_before_pack.member(use)) {
1289 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1290 } else {
1291 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1292 }
1293 --i; // deleted this edge; rescan position
1294 }
1295 }
1296 previous = current;
1297 } else { // !in_pack(current, pk) ==> a sandwiched store
1298 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1299 }
1301 if (current == first) break;
1302 current = my_mem->as_Mem();
1303 } // end while
1305 // Reconnect loads back to upper_insert_pt.
1306 for (uint i = 0; i < memops.size(); i++) {
1307 Node *ld = memops.at(i);
1308 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1309 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1310 }
1311 }
1312 } else if (pk->at(0)->is_Load()) { //load
1313 // all loads in the pack should have the same memory state. By default,
1314 // we use the memory state of the last load. However, if any load could
1315 // not be moved down due to the dependence constraint, we use the memory
1316 // state of the first load.
1317 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1318 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1319 bool schedule_last = true;
1320 for (uint i = 0; i < pk->size(); i++) {
1321 Node* ld = pk->at(i);
1322 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1323 current=current->in(MemNode::Memory)) {
1324 assert(current != first_mem, "corrupted memory graph");
1325 if(current->is_Mem() && !independent(current, ld)){
1326 schedule_last = false; // a later store depends on this load
1327 break;
1328 }
1329 }
1330 }
1332 Node* mem_input = schedule_last ? last_mem : first_mem;
1333 _igvn.hash_delete(mem_input);
1334 // Give each load the same memory state
1335 for (uint i = 0; i < pk->size(); i++) {
1336 LoadNode* ld = pk->at(i)->as_Load();
1337 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1338 }
1339 }
1340 }
1342 //------------------------------output---------------------------
1343 // Convert packs into vector node operations
1344 void SuperWord::output() {
1345 if (_packset.length() == 0) return;
1347 #ifndef PRODUCT
1348 if (TraceLoopOpts) {
1349 tty->print("SuperWord ");
1350 lpt()->dump_head();
1351 }
1352 #endif
1354 // MUST ENSURE main loop's initial value is properly aligned:
1355 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1357 align_initial_loop_index(align_to_ref());
1359 // Insert extract (unpack) operations for scalar uses
1360 for (int i = 0; i < _packset.length(); i++) {
1361 insert_extracts(_packset.at(i));
1362 }
1364 Compile* C = _phase->C;
1365 uint max_vlen_in_bytes = 0;
1366 for (int i = 0; i < _block.length(); i++) {
1367 Node* n = _block.at(i);
1368 Node_List* p = my_pack(n);
1369 if (p && n == executed_last(p)) {
1370 uint vlen = p->size();
1371 uint vlen_in_bytes = 0;
1372 Node* vn = NULL;
1373 Node* low_adr = p->at(0);
1374 Node* first = executed_first(p);
1375 int opc = n->Opcode();
1376 if (n->is_Load()) {
1377 Node* ctl = n->in(MemNode::Control);
1378 Node* mem = first->in(MemNode::Memory);
1379 SWPointer p1(n->as_Mem(), this);
1380 // Identify the memory dependency for the new loadVector node by
1381 // walking up through memory chain.
1382 // This is done to give flexibility to the new loadVector node so that
1383 // it can move above independent storeVector nodes.
1384 while (mem->is_StoreVector()) {
1385 SWPointer p2(mem->as_Mem(), this);
1386 int cmp = p1.cmp(p2);
1387 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1388 mem = mem->in(MemNode::Memory);
1389 } else {
1390 break; // dependent memory
1391 }
1392 }
1393 Node* adr = low_adr->in(MemNode::Address);
1394 const TypePtr* atyp = n->adr_type();
1395 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1396 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1397 } else if (n->is_Store()) {
1398 // Promote value to be stored to vector
1399 Node* val = vector_opd(p, MemNode::ValueIn);
1400 Node* ctl = n->in(MemNode::Control);
1401 Node* mem = first->in(MemNode::Memory);
1402 Node* adr = low_adr->in(MemNode::Address);
1403 const TypePtr* atyp = n->adr_type();
1404 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1405 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1406 } else if (n->req() == 3) {
1407 // Promote operands to vector
1408 Node* in1 = vector_opd(p, 1);
1409 Node* in2 = vector_opd(p, 2);
1410 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1411 // Move invariant vector input into second position to avoid register spilling.
1412 Node* tmp = in1;
1413 in1 = in2;
1414 in2 = tmp;
1415 }
1416 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1417 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1418 } else {
1419 ShouldNotReachHere();
1420 }
1421 assert(vn != NULL, "sanity");
1422 _igvn.register_new_node_with_optimizer(vn);
1423 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1424 for (uint j = 0; j < p->size(); j++) {
1425 Node* pm = p->at(j);
1426 _igvn.replace_node(pm, vn);
1427 }
1428 _igvn._worklist.push(vn);
1430 if (vlen_in_bytes > max_vlen_in_bytes) {
1431 max_vlen_in_bytes = vlen_in_bytes;
1432 }
1433 #ifdef ASSERT
1434 if (TraceNewVectors) {
1435 tty->print("new Vector node: ");
1436 vn->dump();
1437 }
1438 #endif
1439 }
1440 }
1441 C->set_max_vector_size(max_vlen_in_bytes);
1442 }
1444 //------------------------------vector_opd---------------------------
1445 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1446 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1447 Node* p0 = p->at(0);
1448 uint vlen = p->size();
1449 Node* opd = p0->in(opd_idx);
1451 if (same_inputs(p, opd_idx)) {
1452 if (opd->is_Vector() || opd->is_LoadVector()) {
1453 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1454 return opd; // input is matching vector
1455 }
1456 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1457 Compile* C = _phase->C;
1458 Node* cnt = opd;
1459 // Vector instructions do not mask shift count, do it here.
1460 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1461 const TypeInt* t = opd->find_int_type();
1462 if (t != NULL && t->is_con()) {
1463 juint shift = t->get_con();
1464 if (shift > mask) { // Unsigned cmp
1465 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1466 }
1467 } else {
1468 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1469 cnt = ConNode::make(C, TypeInt::make(mask));
1470 _igvn.register_new_node_with_optimizer(cnt);
1471 cnt = new (C) AndINode(opd, cnt);
1472 _igvn.register_new_node_with_optimizer(cnt);
1473 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1474 }
1475 assert(opd->bottom_type()->isa_int(), "int type only");
1476 // Move non constant shift count into vector register.
1477 cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1478 }
1479 if (cnt != opd) {
1480 _igvn.register_new_node_with_optimizer(cnt);
1481 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1482 }
1483 return cnt;
1484 }
1485 assert(!opd->is_StoreVector(), "such vector is not expected here");
1486 // Convert scalar input to vector with the same number of elements as
1487 // p0's vector. Use p0's type because size of operand's container in
1488 // vector should match p0's size regardless operand's size.
1489 const Type* p0_t = velt_type(p0);
1490 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1492 _igvn.register_new_node_with_optimizer(vn);
1493 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1494 #ifdef ASSERT
1495 if (TraceNewVectors) {
1496 tty->print("new Vector node: ");
1497 vn->dump();
1498 }
1499 #endif
1500 return vn;
1501 }
1503 // Insert pack operation
1504 BasicType bt = velt_basic_type(p0);
1505 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1506 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1508 for (uint i = 1; i < vlen; i++) {
1509 Node* pi = p->at(i);
1510 Node* in = pi->in(opd_idx);
1511 assert(my_pack(in) == NULL, "Should already have been unpacked");
1512 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1513 pk->add_opd(in);
1514 }
1515 _igvn.register_new_node_with_optimizer(pk);
1516 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1517 #ifdef ASSERT
1518 if (TraceNewVectors) {
1519 tty->print("new Vector node: ");
1520 pk->dump();
1521 }
1522 #endif
1523 return pk;
1524 }
1526 //------------------------------insert_extracts---------------------------
1527 // If a use of pack p is not a vector use, then replace the
1528 // use with an extract operation.
1529 void SuperWord::insert_extracts(Node_List* p) {
1530 if (p->at(0)->is_Store()) return;
1531 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1533 // Inspect each use of each pack member. For each use that is
1534 // not a vector use, replace the use with an extract operation.
1536 for (uint i = 0; i < p->size(); i++) {
1537 Node* def = p->at(i);
1538 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1539 Node* use = def->fast_out(j);
1540 for (uint k = 0; k < use->req(); k++) {
1541 Node* n = use->in(k);
1542 if (def == n) {
1543 if (!is_vector_use(use, k)) {
1544 _n_idx_list.push(use, k);
1545 }
1546 }
1547 }
1548 }
1549 }
1551 while (_n_idx_list.is_nonempty()) {
1552 Node* use = _n_idx_list.node();
1553 int idx = _n_idx_list.index();
1554 _n_idx_list.pop();
1555 Node* def = use->in(idx);
1557 // Insert extract operation
1558 _igvn.hash_delete(def);
1559 int def_pos = alignment(def) / data_size(def);
1561 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1562 _igvn.register_new_node_with_optimizer(ex);
1563 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1564 _igvn.replace_input_of(use, idx, ex);
1565 _igvn._worklist.push(def);
1567 bb_insert_after(ex, bb_idx(def));
1568 set_velt_type(ex, velt_type(def));
1569 }
1570 }
1572 //------------------------------is_vector_use---------------------------
1573 // Is use->in(u_idx) a vector use?
1574 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1575 Node_List* u_pk = my_pack(use);
1576 if (u_pk == NULL) return false;
1577 Node* def = use->in(u_idx);
1578 Node_List* d_pk = my_pack(def);
1579 if (d_pk == NULL) {
1580 // check for scalar promotion
1581 Node* n = u_pk->at(0)->in(u_idx);
1582 for (uint i = 1; i < u_pk->size(); i++) {
1583 if (u_pk->at(i)->in(u_idx) != n) return false;
1584 }
1585 return true;
1586 }
1587 if (u_pk->size() != d_pk->size())
1588 return false;
1589 for (uint i = 0; i < u_pk->size(); i++) {
1590 Node* ui = u_pk->at(i);
1591 Node* di = d_pk->at(i);
1592 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1593 return false;
1594 }
1595 return true;
1596 }
1598 //------------------------------construct_bb---------------------------
1599 // Construct reverse postorder list of block members
1600 bool SuperWord::construct_bb() {
1601 Node* entry = bb();
1603 assert(_stk.length() == 0, "stk is empty");
1604 assert(_block.length() == 0, "block is empty");
1605 assert(_data_entry.length() == 0, "data_entry is empty");
1606 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1607 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1609 // Find non-control nodes with no inputs from within block,
1610 // create a temporary map from node _idx to bb_idx for use
1611 // by the visited and post_visited sets,
1612 // and count number of nodes in block.
1613 int bb_ct = 0;
1614 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1615 Node *n = lpt()->_body.at(i);
1616 set_bb_idx(n, i); // Create a temporary map
1617 if (in_bb(n)) {
1618 if (n->is_LoadStore() || n->is_MergeMem() ||
1619 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1620 // Bailout if the loop has LoadStore, MergeMem or data Proj
1621 // nodes. Superword optimization does not work with them.
1622 return false;
1623 }
1624 bb_ct++;
1625 if (!n->is_CFG()) {
1626 bool found = false;
1627 for (uint j = 0; j < n->req(); j++) {
1628 Node* def = n->in(j);
1629 if (def && in_bb(def)) {
1630 found = true;
1631 break;
1632 }
1633 }
1634 if (!found) {
1635 assert(n != entry, "can't be entry");
1636 _data_entry.push(n);
1637 }
1638 }
1639 }
1640 }
1642 // Find memory slices (head and tail)
1643 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1644 Node *n = lp()->fast_out(i);
1645 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1646 Node* n_tail = n->in(LoopNode::LoopBackControl);
1647 if (n_tail != n->in(LoopNode::EntryControl)) {
1648 if (!n_tail->is_Mem()) {
1649 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1650 return false; // Bailout
1651 }
1652 _mem_slice_head.push(n);
1653 _mem_slice_tail.push(n_tail);
1654 }
1655 }
1656 }
1658 // Create an RPO list of nodes in block
1660 visited_clear();
1661 post_visited_clear();
1663 // Push all non-control nodes with no inputs from within block, then control entry
1664 for (int j = 0; j < _data_entry.length(); j++) {
1665 Node* n = _data_entry.at(j);
1666 visited_set(n);
1667 _stk.push(n);
1668 }
1669 visited_set(entry);
1670 _stk.push(entry);
1672 // Do a depth first walk over out edges
1673 int rpo_idx = bb_ct - 1;
1674 int size;
1675 while ((size = _stk.length()) > 0) {
1676 Node* n = _stk.top(); // Leave node on stack
1677 if (!visited_test_set(n)) {
1678 // forward arc in graph
1679 } else if (!post_visited_test(n)) {
1680 // cross or back arc
1681 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1682 Node *use = n->fast_out(i);
1683 if (in_bb(use) && !visited_test(use) &&
1684 // Don't go around backedge
1685 (!use->is_Phi() || n == entry)) {
1686 _stk.push(use);
1687 }
1688 }
1689 if (_stk.length() == size) {
1690 // There were no additional uses, post visit node now
1691 _stk.pop(); // Remove node from stack
1692 assert(rpo_idx >= 0, "");
1693 _block.at_put_grow(rpo_idx, n);
1694 rpo_idx--;
1695 post_visited_set(n);
1696 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1697 }
1698 } else {
1699 _stk.pop(); // Remove post-visited node from stack
1700 }
1701 }
1703 // Create real map of block indices for nodes
1704 for (int j = 0; j < _block.length(); j++) {
1705 Node* n = _block.at(j);
1706 set_bb_idx(n, j);
1707 }
1709 initialize_bb(); // Ensure extra info is allocated.
1711 #ifndef PRODUCT
1712 if (TraceSuperWord) {
1713 print_bb();
1714 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1715 for (int m = 0; m < _data_entry.length(); m++) {
1716 tty->print("%3d ", m);
1717 _data_entry.at(m)->dump();
1718 }
1719 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1720 for (int m = 0; m < _mem_slice_head.length(); m++) {
1721 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1722 tty->print(" "); _mem_slice_tail.at(m)->dump();
1723 }
1724 }
1725 #endif
1726 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1727 return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1728 }
1730 //------------------------------initialize_bb---------------------------
1731 // Initialize per node info
1732 void SuperWord::initialize_bb() {
1733 Node* last = _block.at(_block.length() - 1);
1734 grow_node_info(bb_idx(last));
1735 }
1737 //------------------------------bb_insert_after---------------------------
1738 // Insert n into block after pos
1739 void SuperWord::bb_insert_after(Node* n, int pos) {
1740 int n_pos = pos + 1;
1741 // Make room
1742 for (int i = _block.length() - 1; i >= n_pos; i--) {
1743 _block.at_put_grow(i+1, _block.at(i));
1744 }
1745 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1746 _node_info.at_put_grow(j+1, _node_info.at(j));
1747 }
1748 // Set value
1749 _block.at_put_grow(n_pos, n);
1750 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1751 // Adjust map from node->_idx to _block index
1752 for (int i = n_pos; i < _block.length(); i++) {
1753 set_bb_idx(_block.at(i), i);
1754 }
1755 }
1757 //------------------------------compute_max_depth---------------------------
1758 // Compute max depth for expressions from beginning of block
1759 // Use to prune search paths during test for independence.
1760 void SuperWord::compute_max_depth() {
1761 int ct = 0;
1762 bool again;
1763 do {
1764 again = false;
1765 for (int i = 0; i < _block.length(); i++) {
1766 Node* n = _block.at(i);
1767 if (!n->is_Phi()) {
1768 int d_orig = depth(n);
1769 int d_in = 0;
1770 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1771 Node* pred = preds.current();
1772 if (in_bb(pred)) {
1773 d_in = MAX2(d_in, depth(pred));
1774 }
1775 }
1776 if (d_in + 1 != d_orig) {
1777 set_depth(n, d_in + 1);
1778 again = true;
1779 }
1780 }
1781 }
1782 ct++;
1783 } while (again);
1784 #ifndef PRODUCT
1785 if (TraceSuperWord && Verbose)
1786 tty->print_cr("compute_max_depth iterated: %d times", ct);
1787 #endif
1788 }
1790 //-------------------------compute_vector_element_type-----------------------
1791 // Compute necessary vector element type for expressions
1792 // This propagates backwards a narrower integer type when the
1793 // upper bits of the value are not needed.
1794 // Example: char a,b,c; a = b + c;
1795 // Normally the type of the add is integer, but for packed character
1796 // operations the type of the add needs to be char.
1797 void SuperWord::compute_vector_element_type() {
1798 #ifndef PRODUCT
1799 if (TraceSuperWord && Verbose)
1800 tty->print_cr("\ncompute_velt_type:");
1801 #endif
1803 // Initial type
1804 for (int i = 0; i < _block.length(); i++) {
1805 Node* n = _block.at(i);
1806 set_velt_type(n, container_type(n));
1807 }
1809 // Propagate integer narrowed type backwards through operations
1810 // that don't depend on higher order bits
1811 for (int i = _block.length() - 1; i >= 0; i--) {
1812 Node* n = _block.at(i);
1813 // Only integer types need be examined
1814 const Type* vtn = velt_type(n);
1815 if (vtn->basic_type() == T_INT) {
1816 uint start, end;
1817 VectorNode::vector_operands(n, &start, &end);
1819 for (uint j = start; j < end; j++) {
1820 Node* in = n->in(j);
1821 // Don't propagate through a memory
1822 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1823 data_size(n) < data_size(in)) {
1824 bool same_type = true;
1825 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1826 Node *use = in->fast_out(k);
1827 if (!in_bb(use) || !same_velt_type(use, n)) {
1828 same_type = false;
1829 break;
1830 }
1831 }
1832 if (same_type) {
1833 // For right shifts of small integer types (bool, byte, char, short)
1834 // we need precise information about sign-ness. Only Load nodes have
1835 // this information because Store nodes are the same for signed and
1836 // unsigned values. And any arithmetic operation after a load may
1837 // expand a value to signed Int so such right shifts can't be used
1838 // because vector elements do not have upper bits of Int.
1839 const Type* vt = vtn;
1840 if (VectorNode::is_shift(in)) {
1841 Node* load = in->in(1);
1842 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1843 vt = velt_type(load);
1844 } else if (in->Opcode() != Op_LShiftI) {
1845 // Widen type to Int to avoid creation of right shift vector
1846 // (align + data_size(s1) check in stmts_can_pack() will fail).
1847 // Note, left shifts work regardless type.
1848 vt = TypeInt::INT;
1849 }
1850 }
1851 set_velt_type(in, vt);
1852 }
1853 }
1854 }
1855 }
1856 }
1857 #ifndef PRODUCT
1858 if (TraceSuperWord && Verbose) {
1859 for (int i = 0; i < _block.length(); i++) {
1860 Node* n = _block.at(i);
1861 velt_type(n)->dump();
1862 tty->print("\t");
1863 n->dump();
1864 }
1865 }
1866 #endif
1867 }
1869 //------------------------------memory_alignment---------------------------
1870 // Alignment within a vector memory reference
1871 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1872 SWPointer p(s, this);
1873 if (!p.valid()) {
1874 return bottom_align;
1875 }
1876 int vw = vector_width_in_bytes(s);
1877 if (vw < 2) {
1878 return bottom_align; // No vectors for this type
1879 }
1880 int offset = p.offset_in_bytes();
1881 offset += iv_adjust*p.memory_size();
1882 int off_rem = offset % vw;
1883 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1884 return off_mod;
1885 }
1887 //---------------------------container_type---------------------------
1888 // Smallest type containing range of values
1889 const Type* SuperWord::container_type(Node* n) {
1890 if (n->is_Mem()) {
1891 BasicType bt = n->as_Mem()->memory_type();
1892 if (n->is_Store() && (bt == T_CHAR)) {
1893 // Use T_SHORT type instead of T_CHAR for stored values because any
1894 // preceding arithmetic operation extends values to signed Int.
1895 bt = T_SHORT;
1896 }
1897 if (n->Opcode() == Op_LoadUB) {
1898 // Adjust type for unsigned byte loads, it is important for right shifts.
1899 // T_BOOLEAN is used because there is no basic type representing type
1900 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1901 // size (one byte) and sign is important.
1902 bt = T_BOOLEAN;
1903 }
1904 return Type::get_const_basic_type(bt);
1905 }
1906 const Type* t = _igvn.type(n);
1907 if (t->basic_type() == T_INT) {
1908 // A narrow type of arithmetic operations will be determined by
1909 // propagating the type of memory operations.
1910 return TypeInt::INT;
1911 }
1912 return t;
1913 }
1915 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1916 const Type* vt1 = velt_type(n1);
1917 const Type* vt2 = velt_type(n2);
1918 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1919 // Compare vectors element sizes for integer types.
1920 return data_size(n1) == data_size(n2);
1921 }
1922 return vt1 == vt2;
1923 }
1925 //------------------------------in_packset---------------------------
1926 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1927 bool SuperWord::in_packset(Node* s1, Node* s2) {
1928 for (int i = 0; i < _packset.length(); i++) {
1929 Node_List* p = _packset.at(i);
1930 assert(p->size() == 2, "must be");
1931 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1932 return true;
1933 }
1934 }
1935 return false;
1936 }
1938 //------------------------------in_pack---------------------------
1939 // Is s in pack p?
1940 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1941 for (uint i = 0; i < p->size(); i++) {
1942 if (p->at(i) == s) {
1943 return p;
1944 }
1945 }
1946 return NULL;
1947 }
1949 //------------------------------remove_pack_at---------------------------
1950 // Remove the pack at position pos in the packset
1951 void SuperWord::remove_pack_at(int pos) {
1952 Node_List* p = _packset.at(pos);
1953 for (uint i = 0; i < p->size(); i++) {
1954 Node* s = p->at(i);
1955 set_my_pack(s, NULL);
1956 }
1957 _packset.remove_at(pos);
1958 }
1960 //------------------------------executed_first---------------------------
1961 // Return the node executed first in pack p. Uses the RPO block list
1962 // to determine order.
1963 Node* SuperWord::executed_first(Node_List* p) {
1964 Node* n = p->at(0);
1965 int n_rpo = bb_idx(n);
1966 for (uint i = 1; i < p->size(); i++) {
1967 Node* s = p->at(i);
1968 int s_rpo = bb_idx(s);
1969 if (s_rpo < n_rpo) {
1970 n = s;
1971 n_rpo = s_rpo;
1972 }
1973 }
1974 return n;
1975 }
1977 //------------------------------executed_last---------------------------
1978 // Return the node executed last in pack p.
1979 Node* SuperWord::executed_last(Node_List* p) {
1980 Node* n = p->at(0);
1981 int n_rpo = bb_idx(n);
1982 for (uint i = 1; i < p->size(); i++) {
1983 Node* s = p->at(i);
1984 int s_rpo = bb_idx(s);
1985 if (s_rpo > n_rpo) {
1986 n = s;
1987 n_rpo = s_rpo;
1988 }
1989 }
1990 return n;
1991 }
1993 //----------------------------align_initial_loop_index---------------------------
1994 // Adjust pre-loop limit so that in main loop, a load/store reference
1995 // to align_to_ref will be a position zero in the vector.
1996 // (iv + k) mod vector_align == 0
1997 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1998 CountedLoopNode *main_head = lp()->as_CountedLoop();
1999 assert(main_head->is_main_loop(), "");
2000 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2001 assert(pre_end != NULL, "");
2002 Node *pre_opaq1 = pre_end->limit();
2003 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2004 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2005 Node *lim0 = pre_opaq->in(1);
2007 // Where we put new limit calculations
2008 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2010 // Ensure the original loop limit is available from the
2011 // pre-loop Opaque1 node.
2012 Node *orig_limit = pre_opaq->original_loop_limit();
2013 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2015 SWPointer align_to_ref_p(align_to_ref, this);
2016 assert(align_to_ref_p.valid(), "sanity");
2018 // Given:
2019 // lim0 == original pre loop limit
2020 // V == v_align (power of 2)
2021 // invar == extra invariant piece of the address expression
2022 // e == offset [ +/- invar ]
2023 //
2024 // When reassociating expressions involving '%' the basic rules are:
2025 // (a - b) % k == 0 => a % k == b % k
2026 // and:
2027 // (a + b) % k == 0 => a % k == (k - b) % k
2028 //
2029 // For stride > 0 && scale > 0,
2030 // Derive the new pre-loop limit "lim" such that the two constraints:
2031 // (1) lim = lim0 + N (where N is some positive integer < V)
2032 // (2) (e + lim) % V == 0
2033 // are true.
2034 //
2035 // Substituting (1) into (2),
2036 // (e + lim0 + N) % V == 0
2037 // solve for N:
2038 // N = (V - (e + lim0)) % V
2039 // substitute back into (1), so that new limit
2040 // lim = lim0 + (V - (e + lim0)) % V
2041 //
2042 // For stride > 0 && scale < 0
2043 // Constraints:
2044 // lim = lim0 + N
2045 // (e - lim) % V == 0
2046 // Solving for lim:
2047 // (e - lim0 - N) % V == 0
2048 // N = (e - lim0) % V
2049 // lim = lim0 + (e - lim0) % V
2050 //
2051 // For stride < 0 && scale > 0
2052 // Constraints:
2053 // lim = lim0 - N
2054 // (e + lim) % V == 0
2055 // Solving for lim:
2056 // (e + lim0 - N) % V == 0
2057 // N = (e + lim0) % V
2058 // lim = lim0 - (e + lim0) % V
2059 //
2060 // For stride < 0 && scale < 0
2061 // Constraints:
2062 // lim = lim0 - N
2063 // (e - lim) % V == 0
2064 // Solving for lim:
2065 // (e - lim0 + N) % V == 0
2066 // N = (V - (e - lim0)) % V
2067 // lim = lim0 - (V - (e - lim0)) % V
2069 int vw = vector_width_in_bytes(align_to_ref);
2070 int stride = iv_stride();
2071 int scale = align_to_ref_p.scale_in_bytes();
2072 int elt_size = align_to_ref_p.memory_size();
2073 int v_align = vw / elt_size;
2074 assert(v_align > 1, "sanity");
2075 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2076 Node *offsn = _igvn.intcon(offset);
2078 Node *e = offsn;
2079 if (align_to_ref_p.invar() != NULL) {
2080 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2081 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2082 Node* aref = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2083 _igvn.register_new_node_with_optimizer(aref);
2084 _phase->set_ctrl(aref, pre_ctrl);
2085 if (align_to_ref_p.negate_invar()) {
2086 e = new (_phase->C) SubINode(e, aref);
2087 } else {
2088 e = new (_phase->C) AddINode(e, aref);
2089 }
2090 _igvn.register_new_node_with_optimizer(e);
2091 _phase->set_ctrl(e, pre_ctrl);
2092 }
2093 if (vw > ObjectAlignmentInBytes) {
2094 // incorporate base e +/- base && Mask >>> log2(elt)
2095 Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2096 _igvn.register_new_node_with_optimizer(xbase);
2097 #ifdef _LP64
2098 xbase = new (_phase->C) ConvL2INode(xbase);
2099 _igvn.register_new_node_with_optimizer(xbase);
2100 #endif
2101 Node* mask = _igvn.intcon(vw-1);
2102 Node* masked_xbase = new (_phase->C) AndINode(xbase, mask);
2103 _igvn.register_new_node_with_optimizer(masked_xbase);
2104 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2105 Node* bref = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2106 _igvn.register_new_node_with_optimizer(bref);
2107 _phase->set_ctrl(bref, pre_ctrl);
2108 e = new (_phase->C) AddINode(e, bref);
2109 _igvn.register_new_node_with_optimizer(e);
2110 _phase->set_ctrl(e, pre_ctrl);
2111 }
2113 // compute e +/- lim0
2114 if (scale < 0) {
2115 e = new (_phase->C) SubINode(e, lim0);
2116 } else {
2117 e = new (_phase->C) AddINode(e, lim0);
2118 }
2119 _igvn.register_new_node_with_optimizer(e);
2120 _phase->set_ctrl(e, pre_ctrl);
2122 if (stride * scale > 0) {
2123 // compute V - (e +/- lim0)
2124 Node* va = _igvn.intcon(v_align);
2125 e = new (_phase->C) SubINode(va, e);
2126 _igvn.register_new_node_with_optimizer(e);
2127 _phase->set_ctrl(e, pre_ctrl);
2128 }
2129 // compute N = (exp) % V
2130 Node* va_msk = _igvn.intcon(v_align - 1);
2131 Node* N = new (_phase->C) AndINode(e, va_msk);
2132 _igvn.register_new_node_with_optimizer(N);
2133 _phase->set_ctrl(N, pre_ctrl);
2135 // substitute back into (1), so that new limit
2136 // lim = lim0 + N
2137 Node* lim;
2138 if (stride < 0) {
2139 lim = new (_phase->C) SubINode(lim0, N);
2140 } else {
2141 lim = new (_phase->C) AddINode(lim0, N);
2142 }
2143 _igvn.register_new_node_with_optimizer(lim);
2144 _phase->set_ctrl(lim, pre_ctrl);
2145 Node* constrained =
2146 (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2147 : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2148 _igvn.register_new_node_with_optimizer(constrained);
2149 _phase->set_ctrl(constrained, pre_ctrl);
2150 _igvn.hash_delete(pre_opaq);
2151 pre_opaq->set_req(1, constrained);
2152 }
2154 //----------------------------get_pre_loop_end---------------------------
2155 // Find pre loop end from main loop. Returns null if none.
2156 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2157 Node *ctrl = cl->in(LoopNode::EntryControl);
2158 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2159 Node *iffm = ctrl->in(0);
2160 if (!iffm->is_If()) return NULL;
2161 Node *p_f = iffm->in(0);
2162 if (!p_f->is_IfFalse()) return NULL;
2163 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2164 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2165 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2166 return pre_end;
2167 }
2170 //------------------------------init---------------------------
2171 void SuperWord::init() {
2172 _dg.init();
2173 _packset.clear();
2174 _disjoint_ptrs.clear();
2175 _block.clear();
2176 _data_entry.clear();
2177 _mem_slice_head.clear();
2178 _mem_slice_tail.clear();
2179 _node_info.clear();
2180 _align_to_ref = NULL;
2181 _lpt = NULL;
2182 _lp = NULL;
2183 _bb = NULL;
2184 _iv = NULL;
2185 }
2187 //------------------------------print_packset---------------------------
2188 void SuperWord::print_packset() {
2189 #ifndef PRODUCT
2190 tty->print_cr("packset");
2191 for (int i = 0; i < _packset.length(); i++) {
2192 tty->print_cr("Pack: %d", i);
2193 Node_List* p = _packset.at(i);
2194 print_pack(p);
2195 }
2196 #endif
2197 }
2199 //------------------------------print_pack---------------------------
2200 void SuperWord::print_pack(Node_List* p) {
2201 for (uint i = 0; i < p->size(); i++) {
2202 print_stmt(p->at(i));
2203 }
2204 }
2206 //------------------------------print_bb---------------------------
2207 void SuperWord::print_bb() {
2208 #ifndef PRODUCT
2209 tty->print_cr("\nBlock");
2210 for (int i = 0; i < _block.length(); i++) {
2211 Node* n = _block.at(i);
2212 tty->print("%d ", i);
2213 if (n) {
2214 n->dump();
2215 }
2216 }
2217 #endif
2218 }
2220 //------------------------------print_stmt---------------------------
2221 void SuperWord::print_stmt(Node* s) {
2222 #ifndef PRODUCT
2223 tty->print(" align: %d \t", alignment(s));
2224 s->dump();
2225 #endif
2226 }
2228 //------------------------------blank---------------------------
2229 char* SuperWord::blank(uint depth) {
2230 static char blanks[101];
2231 assert(depth < 101, "too deep");
2232 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2233 blanks[depth] = '\0';
2234 return blanks;
2235 }
2238 //==============================SWPointer===========================
2240 //----------------------------SWPointer------------------------
2241 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2242 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2243 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2245 Node* adr = mem->in(MemNode::Address);
2246 if (!adr->is_AddP()) {
2247 assert(!valid(), "too complex");
2248 return;
2249 }
2250 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2251 Node* base = adr->in(AddPNode::Base);
2252 // The base address should be loop invariant
2253 if (!invariant(base)) {
2254 assert(!valid(), "base address is loop variant");
2255 return;
2256 }
2257 //unsafe reference could not be aligned appropriately without runtime checking
2258 if (base == NULL || base->bottom_type() == Type::TOP) {
2259 assert(!valid(), "unsafe access");
2260 return;
2261 }
2262 for (int i = 0; i < 3; i++) {
2263 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2264 assert(!valid(), "too complex");
2265 return;
2266 }
2267 adr = adr->in(AddPNode::Address);
2268 if (base == adr || !adr->is_AddP()) {
2269 break; // stop looking at addp's
2270 }
2271 }
2272 _base = base;
2273 _adr = adr;
2274 assert(valid(), "Usable");
2275 }
2277 // Following is used to create a temporary object during
2278 // the pattern match of an address expression.
2279 SWPointer::SWPointer(SWPointer* p) :
2280 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2281 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2283 //------------------------scaled_iv_plus_offset--------------------
2284 // Match: k*iv + offset
2285 // where: k is a constant that maybe zero, and
2286 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2287 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2288 if (scaled_iv(n)) {
2289 return true;
2290 }
2291 if (offset_plus_k(n)) {
2292 return true;
2293 }
2294 int opc = n->Opcode();
2295 if (opc == Op_AddI) {
2296 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2297 return true;
2298 }
2299 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2300 return true;
2301 }
2302 } else if (opc == Op_SubI) {
2303 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2304 return true;
2305 }
2306 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2307 _scale *= -1;
2308 return true;
2309 }
2310 }
2311 return false;
2312 }
2314 //----------------------------scaled_iv------------------------
2315 // Match: k*iv where k is a constant that's not zero
2316 bool SWPointer::scaled_iv(Node* n) {
2317 if (_scale != 0) {
2318 return false; // already found a scale
2319 }
2320 if (n == iv()) {
2321 _scale = 1;
2322 return true;
2323 }
2324 int opc = n->Opcode();
2325 if (opc == Op_MulI) {
2326 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2327 _scale = n->in(2)->get_int();
2328 return true;
2329 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2330 _scale = n->in(1)->get_int();
2331 return true;
2332 }
2333 } else if (opc == Op_LShiftI) {
2334 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2335 _scale = 1 << n->in(2)->get_int();
2336 return true;
2337 }
2338 } else if (opc == Op_ConvI2L) {
2339 if (scaled_iv_plus_offset(n->in(1))) {
2340 return true;
2341 }
2342 } else if (opc == Op_LShiftL) {
2343 if (!has_iv() && _invar == NULL) {
2344 // Need to preserve the current _offset value, so
2345 // create a temporary object for this expression subtree.
2346 // Hacky, so should re-engineer the address pattern match.
2347 SWPointer tmp(this);
2348 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2349 if (tmp._invar == NULL) {
2350 int mult = 1 << n->in(2)->get_int();
2351 _scale = tmp._scale * mult;
2352 _offset += tmp._offset * mult;
2353 return true;
2354 }
2355 }
2356 }
2357 }
2358 return false;
2359 }
2361 //----------------------------offset_plus_k------------------------
2362 // Match: offset is (k [+/- invariant])
2363 // where k maybe zero and invariant is optional, but not both.
2364 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2365 int opc = n->Opcode();
2366 if (opc == Op_ConI) {
2367 _offset += negate ? -(n->get_int()) : n->get_int();
2368 return true;
2369 } else if (opc == Op_ConL) {
2370 // Okay if value fits into an int
2371 const TypeLong* t = n->find_long_type();
2372 if (t->higher_equal(TypeLong::INT)) {
2373 jlong loff = n->get_long();
2374 jint off = (jint)loff;
2375 _offset += negate ? -off : loff;
2376 return true;
2377 }
2378 return false;
2379 }
2380 if (_invar != NULL) return false; // already have an invariant
2381 if (opc == Op_AddI) {
2382 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2383 _negate_invar = negate;
2384 _invar = n->in(1);
2385 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2386 return true;
2387 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2388 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2389 _negate_invar = negate;
2390 _invar = n->in(2);
2391 return true;
2392 }
2393 }
2394 if (opc == Op_SubI) {
2395 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2396 _negate_invar = negate;
2397 _invar = n->in(1);
2398 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2399 return true;
2400 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2401 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2402 _negate_invar = !negate;
2403 _invar = n->in(2);
2404 return true;
2405 }
2406 }
2407 if (invariant(n)) {
2408 _negate_invar = negate;
2409 _invar = n;
2410 return true;
2411 }
2412 return false;
2413 }
2415 //----------------------------print------------------------
2416 void SWPointer::print() {
2417 #ifndef PRODUCT
2418 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2419 _base != NULL ? _base->_idx : 0,
2420 _adr != NULL ? _adr->_idx : 0,
2421 _scale, _offset,
2422 _negate_invar?'-':'+',
2423 _invar != NULL ? _invar->_idx : 0);
2424 #endif
2425 }
2427 // ========================= OrderedPair =====================
2429 const OrderedPair OrderedPair::initial;
2431 // ========================= SWNodeInfo =====================
2433 const SWNodeInfo SWNodeInfo::initial;
2436 // ============================ DepGraph ===========================
2438 //------------------------------make_node---------------------------
2439 // Make a new dependence graph node for an ideal node.
2440 DepMem* DepGraph::make_node(Node* node) {
2441 DepMem* m = new (_arena) DepMem(node);
2442 if (node != NULL) {
2443 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2444 _map.at_put_grow(node->_idx, m);
2445 }
2446 return m;
2447 }
2449 //------------------------------make_edge---------------------------
2450 // Make a new dependence graph edge from dpred -> dsucc
2451 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2452 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2453 dpred->set_out_head(e);
2454 dsucc->set_in_head(e);
2455 return e;
2456 }
2458 // ========================== DepMem ========================
2460 //------------------------------in_cnt---------------------------
2461 int DepMem::in_cnt() {
2462 int ct = 0;
2463 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2464 return ct;
2465 }
2467 //------------------------------out_cnt---------------------------
2468 int DepMem::out_cnt() {
2469 int ct = 0;
2470 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2471 return ct;
2472 }
2474 //------------------------------print-----------------------------
2475 void DepMem::print() {
2476 #ifndef PRODUCT
2477 tty->print(" DepNode %d (", _node->_idx);
2478 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2479 Node* pred = p->pred()->node();
2480 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2481 }
2482 tty->print(") [");
2483 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2484 Node* succ = s->succ()->node();
2485 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2486 }
2487 tty->print_cr(" ]");
2488 #endif
2489 }
2491 // =========================== DepEdge =========================
2493 //------------------------------DepPreds---------------------------
2494 void DepEdge::print() {
2495 #ifndef PRODUCT
2496 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2497 #endif
2498 }
2500 // =========================== DepPreds =========================
2501 // Iterator over predecessor edges in the dependence graph.
2503 //------------------------------DepPreds---------------------------
2504 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2505 _n = n;
2506 _done = false;
2507 if (_n->is_Store() || _n->is_Load()) {
2508 _next_idx = MemNode::Address;
2509 _end_idx = n->req();
2510 _dep_next = dg.dep(_n)->in_head();
2511 } else if (_n->is_Mem()) {
2512 _next_idx = 0;
2513 _end_idx = 0;
2514 _dep_next = dg.dep(_n)->in_head();
2515 } else {
2516 _next_idx = 1;
2517 _end_idx = _n->req();
2518 _dep_next = NULL;
2519 }
2520 next();
2521 }
2523 //------------------------------next---------------------------
2524 void DepPreds::next() {
2525 if (_dep_next != NULL) {
2526 _current = _dep_next->pred()->node();
2527 _dep_next = _dep_next->next_in();
2528 } else if (_next_idx < _end_idx) {
2529 _current = _n->in(_next_idx++);
2530 } else {
2531 _done = true;
2532 }
2533 }
2535 // =========================== DepSuccs =========================
2536 // Iterator over successor edges in the dependence graph.
2538 //------------------------------DepSuccs---------------------------
2539 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2540 _n = n;
2541 _done = false;
2542 if (_n->is_Load()) {
2543 _next_idx = 0;
2544 _end_idx = _n->outcnt();
2545 _dep_next = dg.dep(_n)->out_head();
2546 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2547 _next_idx = 0;
2548 _end_idx = 0;
2549 _dep_next = dg.dep(_n)->out_head();
2550 } else {
2551 _next_idx = 0;
2552 _end_idx = _n->outcnt();
2553 _dep_next = NULL;
2554 }
2555 next();
2556 }
2558 //-------------------------------next---------------------------
2559 void DepSuccs::next() {
2560 if (_dep_next != NULL) {
2561 _current = _dep_next->succ()->node();
2562 _dep_next = _dep_next->next_out();
2563 } else if (_next_idx < _end_idx) {
2564 _current = _n->raw_out(_next_idx++);
2565 } else {
2566 _done = true;
2567 }
2568 }