Sat, 09 Nov 2019 20:29:45 +0800
Merge
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 } else {
236 SWPointer p2(best_align_to_mem_ref, this);
237 if (align_to_ref_p.invar() != p2.invar()) {
238 // Do not vectorize memory accesses with different invariants
239 // if unaligned memory accesses are not allowed.
240 create_pack = false;
241 }
242 }
243 }
244 } else {
245 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
246 // Can't allow vectorization of unaligned memory accesses with the
247 // same type since it could be overlapped accesses to the same array.
248 create_pack = false;
249 } else {
250 // Allow independent (different type) unaligned memory operations
251 // if HW supports them.
252 if (!Matcher::misaligned_vectors_ok()) {
253 create_pack = false;
254 } else {
255 // Check if packs of the same memory type but
256 // with a different alignment were created before.
257 for (uint i = 0; i < align_to_refs.size(); i++) {
258 MemNode* mr = align_to_refs.at(i)->as_Mem();
259 if (same_velt_type(mr, mem_ref) &&
260 memory_alignment(mr, iv_adjustment) != 0)
261 create_pack = false;
262 }
263 }
264 }
265 }
266 if (create_pack) {
267 for (uint i = 0; i < memops.size(); i++) {
268 Node* s1 = memops.at(i);
269 int align = alignment(s1);
270 if (align == top_align) continue;
271 for (uint j = 0; j < memops.size(); j++) {
272 Node* s2 = memops.at(j);
273 if (alignment(s2) == top_align) continue;
274 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
275 if (stmts_can_pack(s1, s2, align)) {
276 Node_List* pair = new Node_List();
277 pair->push(s1);
278 pair->push(s2);
279 _packset.append(pair);
280 }
281 }
282 }
283 }
284 } else { // Don't create unaligned pack
285 // First, remove remaining memory ops of the same type from the list.
286 for (int i = memops.size() - 1; i >= 0; i--) {
287 MemNode* s = memops.at(i)->as_Mem();
288 if (same_velt_type(s, mem_ref)) {
289 memops.remove(i);
290 }
291 }
293 // Second, remove already constructed packs of the same type.
294 for (int i = _packset.length() - 1; i >= 0; i--) {
295 Node_List* p = _packset.at(i);
296 MemNode* s = p->at(0)->as_Mem();
297 if (same_velt_type(s, mem_ref)) {
298 remove_pack_at(i);
299 }
300 }
302 // If needed find the best memory reference for loop alignment again.
303 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
304 // Put memory ops from remaining packs back on memops list for
305 // the best alignment search.
306 uint orig_msize = memops.size();
307 for (int i = 0; i < _packset.length(); i++) {
308 Node_List* p = _packset.at(i);
309 MemNode* s = p->at(0)->as_Mem();
310 assert(!same_velt_type(s, mem_ref), "sanity");
311 memops.push(s);
312 }
313 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
314 if (best_align_to_mem_ref == NULL) break;
315 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
316 // Restore list.
317 while (memops.size() > orig_msize)
318 (void)memops.pop();
319 }
320 } // unaligned memory accesses
322 // Remove used mem nodes.
323 for (int i = memops.size() - 1; i >= 0; i--) {
324 MemNode* m = memops.at(i)->as_Mem();
325 if (alignment(m) != top_align) {
326 memops.remove(i);
327 }
328 }
330 } // while (memops.size() != 0
331 set_align_to_ref(best_align_to_mem_ref);
333 #ifndef PRODUCT
334 if (TraceSuperWord) {
335 tty->print_cr("\nAfter find_adjacent_refs");
336 print_packset();
337 }
338 #endif
339 }
341 //------------------------------find_align_to_ref---------------------------
342 // Find a memory reference to align the loop induction variable to.
343 // Looks first at stores then at loads, looking for a memory reference
344 // with the largest number of references similar to it.
345 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
346 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
348 // Count number of comparable memory ops
349 for (uint i = 0; i < memops.size(); i++) {
350 MemNode* s1 = memops.at(i)->as_Mem();
351 SWPointer p1(s1, this);
352 // Discard if pre loop can't align this reference
353 if (!ref_is_alignable(p1)) {
354 *cmp_ct.adr_at(i) = 0;
355 continue;
356 }
357 for (uint j = i+1; j < memops.size(); j++) {
358 MemNode* s2 = memops.at(j)->as_Mem();
359 if (isomorphic(s1, s2)) {
360 SWPointer p2(s2, this);
361 if (p1.comparable(p2)) {
362 (*cmp_ct.adr_at(i))++;
363 (*cmp_ct.adr_at(j))++;
364 }
365 }
366 }
367 }
369 // Find Store (or Load) with the greatest number of "comparable" references,
370 // biggest vector size, smallest data size and smallest iv offset.
371 int max_ct = 0;
372 int max_vw = 0;
373 int max_idx = -1;
374 int min_size = max_jint;
375 int min_iv_offset = max_jint;
376 for (uint j = 0; j < memops.size(); j++) {
377 MemNode* s = memops.at(j)->as_Mem();
378 if (s->is_Store()) {
379 int vw = vector_width_in_bytes(s);
380 assert(vw > 1, "sanity");
381 SWPointer p(s, this);
382 if (cmp_ct.at(j) > max_ct ||
383 cmp_ct.at(j) == max_ct &&
384 (vw > max_vw ||
385 vw == max_vw &&
386 (data_size(s) < min_size ||
387 data_size(s) == min_size &&
388 (p.offset_in_bytes() < min_iv_offset)))) {
389 max_ct = cmp_ct.at(j);
390 max_vw = vw;
391 max_idx = j;
392 min_size = data_size(s);
393 min_iv_offset = p.offset_in_bytes();
394 }
395 }
396 }
397 // If no stores, look at loads
398 if (max_ct == 0) {
399 for (uint j = 0; j < memops.size(); j++) {
400 MemNode* s = memops.at(j)->as_Mem();
401 if (s->is_Load()) {
402 int vw = vector_width_in_bytes(s);
403 assert(vw > 1, "sanity");
404 SWPointer p(s, this);
405 if (cmp_ct.at(j) > max_ct ||
406 cmp_ct.at(j) == max_ct &&
407 (vw > max_vw ||
408 vw == max_vw &&
409 (data_size(s) < min_size ||
410 data_size(s) == min_size &&
411 (p.offset_in_bytes() < min_iv_offset)))) {
412 max_ct = cmp_ct.at(j);
413 max_vw = vw;
414 max_idx = j;
415 min_size = data_size(s);
416 min_iv_offset = p.offset_in_bytes();
417 }
418 }
419 }
420 }
422 #ifdef ASSERT
423 if (TraceSuperWord && Verbose) {
424 tty->print_cr("\nVector memops after find_align_to_refs");
425 for (uint i = 0; i < memops.size(); i++) {
426 MemNode* s = memops.at(i)->as_Mem();
427 s->dump();
428 }
429 }
430 #endif
432 if (max_ct > 0) {
433 #ifdef ASSERT
434 if (TraceSuperWord) {
435 tty->print("\nVector align to node: ");
436 memops.at(max_idx)->as_Mem()->dump();
437 }
438 #endif
439 return memops.at(max_idx)->as_Mem();
440 }
441 return NULL;
442 }
444 //------------------------------ref_is_alignable---------------------------
445 // Can the preloop align the reference to position zero in the vector?
446 bool SuperWord::ref_is_alignable(SWPointer& p) {
447 if (!p.has_iv()) {
448 return true; // no induction variable
449 }
450 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
451 assert(pre_end->stride_is_con(), "pre loop stride is constant");
452 int preloop_stride = pre_end->stride_con();
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 the initial offset from start of the object is computable,
463 // check if the pre-loop can align the final offset accordingly.
464 //
465 // In other words: Can we find an i such that the offset
466 // after i pre-loop iterations is aligned to vw?
467 // (init_offset + pre_loop) % vw == 0 (1)
468 // where
469 // pre_loop = i * span
470 // is the number of bytes added to the offset by i pre-loop iterations.
471 //
472 // For this to hold we need pre_loop to increase init_offset by
473 // pre_loop = vw - (init_offset % vw)
474 //
475 // This is only possible if pre_loop is divisible by span because each
476 // pre-loop iteration increases the initial offset by 'span' bytes:
477 // (vw - (init_offset % vw)) % span == 0
478 //
479 int vw = vector_width_in_bytes(p.mem());
480 assert(vw > 1, "sanity");
481 Node* init_nd = pre_end->init_trip();
482 if (init_nd->is_Con() && p.invar() == NULL) {
483 int init = init_nd->bottom_type()->is_int()->get_con();
484 int init_offset = init * p.scale_in_bytes() + offset;
485 if (init_offset < 0) { // negative offset from object start?
486 return false; // may happen in dead loop
487 }
488 if (vw % span == 0) {
489 // If vm is a multiple of span, we use formula (1).
490 if (span > 0) {
491 return (vw - (init_offset % vw)) % span == 0;
492 } else {
493 assert(span < 0, "nonzero stride * scale");
494 return (init_offset % vw) % -span == 0;
495 }
496 } else if (span % vw == 0) {
497 // If span is a multiple of vw, we can simplify formula (1) to:
498 // (init_offset + i * span) % vw == 0
499 // =>
500 // (init_offset % vw) + ((i * span) % vw) == 0
501 // =>
502 // init_offset % vw == 0
503 //
504 // Because we add a multiple of vw to the initial offset, the final
505 // offset is a multiple of vw if and only if init_offset is a multiple.
506 //
507 return (init_offset % vw) == 0;
508 }
509 }
510 return false;
511 }
513 //---------------------------get_iv_adjustment---------------------------
514 // Calculate loop's iv adjustment for this memory ops.
515 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
516 SWPointer align_to_ref_p(mem_ref, this);
517 int offset = align_to_ref_p.offset_in_bytes();
518 int scale = align_to_ref_p.scale_in_bytes();
519 int elt_size = align_to_ref_p.memory_size();
520 int vw = vector_width_in_bytes(mem_ref);
521 assert(vw > 1, "sanity");
522 int iv_adjustment;
523 if (scale != 0) {
524 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
525 // At least one iteration is executed in pre-loop by default. As result
526 // several iterations are needed to align memory operations in main-loop even
527 // if offset is 0.
528 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
529 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
530 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
531 iv_adjustment = iv_adjustment_in_bytes/elt_size;
532 } else {
533 // This memory op is not dependent on iv (scale == 0)
534 iv_adjustment = 0;
535 }
537 #ifndef PRODUCT
538 if (TraceSuperWord)
539 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
540 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
541 #endif
542 return iv_adjustment;
543 }
545 //---------------------------dependence_graph---------------------------
546 // Construct dependency graph.
547 // Add dependence edges to load/store nodes for memory dependence
548 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
549 void SuperWord::dependence_graph() {
550 // First, assign a dependence node to each memory node
551 for (int i = 0; i < _block.length(); i++ ) {
552 Node *n = _block.at(i);
553 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
554 _dg.make_node(n);
555 }
556 }
558 // For each memory slice, create the dependences
559 for (int i = 0; i < _mem_slice_head.length(); i++) {
560 Node* n = _mem_slice_head.at(i);
561 Node* n_tail = _mem_slice_tail.at(i);
563 // Get slice in predecessor order (last is first)
564 mem_slice_preds(n_tail, n, _nlist);
566 // Make the slice dependent on the root
567 DepMem* slice = _dg.dep(n);
568 _dg.make_edge(_dg.root(), slice);
570 // Create a sink for the slice
571 DepMem* slice_sink = _dg.make_node(NULL);
572 _dg.make_edge(slice_sink, _dg.tail());
574 // Now visit each pair of memory ops, creating the edges
575 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
576 Node* s1 = _nlist.at(j);
578 // If no dependency yet, use slice
579 if (_dg.dep(s1)->in_cnt() == 0) {
580 _dg.make_edge(slice, s1);
581 }
582 SWPointer p1(s1->as_Mem(), this);
583 bool sink_dependent = true;
584 for (int k = j - 1; k >= 0; k--) {
585 Node* s2 = _nlist.at(k);
586 if (s1->is_Load() && s2->is_Load())
587 continue;
588 SWPointer p2(s2->as_Mem(), this);
590 int cmp = p1.cmp(p2);
591 if (SuperWordRTDepCheck &&
592 p1.base() != p2.base() && p1.valid() && p2.valid()) {
593 // Create a runtime check to disambiguate
594 OrderedPair pp(p1.base(), p2.base());
595 _disjoint_ptrs.append_if_missing(pp);
596 } else if (!SWPointer::not_equal(cmp)) {
597 // Possibly same address
598 _dg.make_edge(s1, s2);
599 sink_dependent = false;
600 }
601 }
602 if (sink_dependent) {
603 _dg.make_edge(s1, slice_sink);
604 }
605 }
606 #ifndef PRODUCT
607 if (TraceSuperWord) {
608 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
609 for (int q = 0; q < _nlist.length(); q++) {
610 _dg.print(_nlist.at(q));
611 }
612 tty->cr();
613 }
614 #endif
615 _nlist.clear();
616 }
618 #ifndef PRODUCT
619 if (TraceSuperWord) {
620 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
621 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
622 _disjoint_ptrs.at(r).print();
623 tty->cr();
624 }
625 tty->cr();
626 }
627 #endif
628 }
630 //---------------------------mem_slice_preds---------------------------
631 // Return a memory slice (node list) in predecessor order starting at "start"
632 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
633 assert(preds.length() == 0, "start empty");
634 Node* n = start;
635 Node* prev = NULL;
636 while (true) {
637 assert(in_bb(n), "must be in block");
638 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
639 Node* out = n->fast_out(i);
640 if (out->is_Load()) {
641 if (in_bb(out)) {
642 preds.push(out);
643 }
644 } else {
645 // FIXME
646 if (out->is_MergeMem() && !in_bb(out)) {
647 // Either unrolling is causing a memory edge not to disappear,
648 // or need to run igvn.optimize() again before SLP
649 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
650 // Ditto. Not sure what else to check further.
651 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
652 // StoreCM has an input edge used as a precedence edge.
653 // Maybe an issue when oop stores are vectorized.
654 } else {
655 assert(out == prev || prev == NULL, "no branches off of store slice");
656 }
657 }
658 }
659 if (n == stop) break;
660 preds.push(n);
661 prev = n;
662 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
663 n = n->in(MemNode::Memory);
664 }
665 }
667 //------------------------------stmts_can_pack---------------------------
668 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
669 // s1 aligned at "align"
670 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
672 // Do not use superword for non-primitives
673 BasicType bt1 = velt_basic_type(s1);
674 BasicType bt2 = velt_basic_type(s2);
675 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
676 return false;
677 if (Matcher::max_vector_size(bt1) < 2) {
678 return false; // No vectors for this type
679 }
681 if (isomorphic(s1, s2)) {
682 if (independent(s1, s2)) {
683 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
684 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
685 int s1_align = alignment(s1);
686 int s2_align = alignment(s2);
687 if (s1_align == top_align || s1_align == align) {
688 if (s2_align == top_align || s2_align == align + data_size(s1)) {
689 return true;
690 }
691 }
692 }
693 }
694 }
695 }
696 return false;
697 }
699 //------------------------------exists_at---------------------------
700 // Does s exist in a pack at position pos?
701 bool SuperWord::exists_at(Node* s, uint pos) {
702 for (int i = 0; i < _packset.length(); i++) {
703 Node_List* p = _packset.at(i);
704 if (p->at(pos) == s) {
705 return true;
706 }
707 }
708 return false;
709 }
711 //------------------------------are_adjacent_refs---------------------------
712 // Is s1 immediately before s2 in memory?
713 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
714 if (!s1->is_Mem() || !s2->is_Mem()) return false;
715 if (!in_bb(s1) || !in_bb(s2)) return false;
717 // Do not use superword for non-primitives
718 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
719 !is_java_primitive(s2->as_Mem()->memory_type())) {
720 return false;
721 }
723 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
724 // only pack memops that are in the same alias set until that's fixed.
725 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
726 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
727 return false;
728 SWPointer p1(s1->as_Mem(), this);
729 SWPointer p2(s2->as_Mem(), this);
730 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
731 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
732 return diff == data_size(s1);
733 }
735 //------------------------------isomorphic---------------------------
736 // Are s1 and s2 similar?
737 bool SuperWord::isomorphic(Node* s1, Node* s2) {
738 if (s1->Opcode() != s2->Opcode()) return false;
739 if (s1->req() != s2->req()) return false;
740 if (s1->in(0) != s2->in(0)) return false;
741 if (!same_velt_type(s1, s2)) return false;
742 return true;
743 }
745 //------------------------------independent---------------------------
746 // Is there no data path from s1 to s2 or s2 to s1?
747 bool SuperWord::independent(Node* s1, Node* s2) {
748 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
749 int d1 = depth(s1);
750 int d2 = depth(s2);
751 if (d1 == d2) return s1 != s2;
752 Node* deep = d1 > d2 ? s1 : s2;
753 Node* shallow = d1 > d2 ? s2 : s1;
755 visited_clear();
757 return independent_path(shallow, deep);
758 }
760 //------------------------------independent_path------------------------------
761 // Helper for independent
762 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
763 if (dp >= 1000) return false; // stop deep recursion
764 visited_set(deep);
765 int shal_depth = depth(shallow);
766 assert(shal_depth <= depth(deep), "must be");
767 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
768 Node* pred = preds.current();
769 if (in_bb(pred) && !visited_test(pred)) {
770 if (shallow == pred) {
771 return false;
772 }
773 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
774 return false;
775 }
776 }
777 }
778 return true;
779 }
781 //------------------------------set_alignment---------------------------
782 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
783 set_alignment(s1, align);
784 if (align == top_align || align == bottom_align) {
785 set_alignment(s2, align);
786 } else {
787 set_alignment(s2, align + data_size(s1));
788 }
789 }
791 //------------------------------data_size---------------------------
792 int SuperWord::data_size(Node* s) {
793 int bsize = type2aelembytes(velt_basic_type(s));
794 assert(bsize != 0, "valid size");
795 return bsize;
796 }
798 //------------------------------extend_packlist---------------------------
799 // Extend packset by following use->def and def->use links from pack members.
800 void SuperWord::extend_packlist() {
801 bool changed;
802 do {
803 changed = false;
804 for (int i = 0; i < _packset.length(); i++) {
805 Node_List* p = _packset.at(i);
806 changed |= follow_use_defs(p);
807 changed |= follow_def_uses(p);
808 }
809 } while (changed);
811 #ifndef PRODUCT
812 if (TraceSuperWord) {
813 tty->print_cr("\nAfter extend_packlist");
814 print_packset();
815 }
816 #endif
817 }
819 //------------------------------follow_use_defs---------------------------
820 // Extend the packset by visiting operand definitions of nodes in pack p
821 bool SuperWord::follow_use_defs(Node_List* p) {
822 assert(p->size() == 2, "just checking");
823 Node* s1 = p->at(0);
824 Node* s2 = p->at(1);
825 assert(s1->req() == s2->req(), "just checking");
826 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
828 if (s1->is_Load()) return false;
830 int align = alignment(s1);
831 bool changed = false;
832 int start = s1->is_Store() ? MemNode::ValueIn : 1;
833 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
834 for (int j = start; j < end; j++) {
835 Node* t1 = s1->in(j);
836 Node* t2 = s2->in(j);
837 if (!in_bb(t1) || !in_bb(t2))
838 continue;
839 if (stmts_can_pack(t1, t2, align)) {
840 if (est_savings(t1, t2) >= 0) {
841 Node_List* pair = new Node_List();
842 pair->push(t1);
843 pair->push(t2);
844 _packset.append(pair);
845 set_alignment(t1, t2, align);
846 changed = true;
847 }
848 }
849 }
850 return changed;
851 }
853 //------------------------------follow_def_uses---------------------------
854 // Extend the packset by visiting uses of nodes in pack p
855 bool SuperWord::follow_def_uses(Node_List* p) {
856 bool changed = false;
857 Node* s1 = p->at(0);
858 Node* s2 = p->at(1);
859 assert(p->size() == 2, "just checking");
860 assert(s1->req() == s2->req(), "just checking");
861 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
863 if (s1->is_Store()) return false;
865 int align = alignment(s1);
866 int savings = -1;
867 Node* u1 = NULL;
868 Node* u2 = NULL;
869 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
870 Node* t1 = s1->fast_out(i);
871 if (!in_bb(t1)) continue;
872 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
873 Node* t2 = s2->fast_out(j);
874 if (!in_bb(t2)) continue;
875 if (!opnd_positions_match(s1, t1, s2, t2))
876 continue;
877 if (stmts_can_pack(t1, t2, align)) {
878 int my_savings = est_savings(t1, t2);
879 if (my_savings > savings) {
880 savings = my_savings;
881 u1 = t1;
882 u2 = t2;
883 }
884 }
885 }
886 }
887 if (savings >= 0) {
888 Node_List* pair = new Node_List();
889 pair->push(u1);
890 pair->push(u2);
891 _packset.append(pair);
892 set_alignment(u1, u2, align);
893 changed = true;
894 }
895 return changed;
896 }
898 //---------------------------opnd_positions_match-------------------------
899 // Is the use of d1 in u1 at the same operand position as d2 in u2?
900 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
901 uint ct = u1->req();
902 if (ct != u2->req()) return false;
903 uint i1 = 0;
904 uint i2 = 0;
905 do {
906 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
907 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
908 if (i1 != i2) {
909 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
910 // Further analysis relies on operands position matching.
911 u2->swap_edges(i1, i2);
912 } else {
913 return false;
914 }
915 }
916 } while (i1 < ct);
917 return true;
918 }
920 //------------------------------est_savings---------------------------
921 // Estimate the savings from executing s1 and s2 as a pack
922 int SuperWord::est_savings(Node* s1, Node* s2) {
923 int save_in = 2 - 1; // 2 operations per instruction in packed form
925 // inputs
926 for (uint i = 1; i < s1->req(); i++) {
927 Node* x1 = s1->in(i);
928 Node* x2 = s2->in(i);
929 if (x1 != x2) {
930 if (are_adjacent_refs(x1, x2)) {
931 save_in += adjacent_profit(x1, x2);
932 } else if (!in_packset(x1, x2)) {
933 save_in -= pack_cost(2);
934 } else {
935 save_in += unpack_cost(2);
936 }
937 }
938 }
940 // uses of result
941 uint ct = 0;
942 int save_use = 0;
943 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
944 Node* s1_use = s1->fast_out(i);
945 for (int j = 0; j < _packset.length(); j++) {
946 Node_List* p = _packset.at(j);
947 if (p->at(0) == s1_use) {
948 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
949 Node* s2_use = s2->fast_out(k);
950 if (p->at(p->size()-1) == s2_use) {
951 ct++;
952 if (are_adjacent_refs(s1_use, s2_use)) {
953 save_use += adjacent_profit(s1_use, s2_use);
954 }
955 }
956 }
957 }
958 }
959 }
961 if (ct < s1->outcnt()) save_use += unpack_cost(1);
962 if (ct < s2->outcnt()) save_use += unpack_cost(1);
964 return MAX2(save_in, save_use);
965 }
967 //------------------------------costs---------------------------
968 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
969 int SuperWord::pack_cost(int ct) { return ct; }
970 int SuperWord::unpack_cost(int ct) { return ct; }
972 //------------------------------combine_packs---------------------------
973 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
974 void SuperWord::combine_packs() {
975 bool changed = true;
976 // Combine packs regardless max vector size.
977 while (changed) {
978 changed = false;
979 for (int i = 0; i < _packset.length(); i++) {
980 Node_List* p1 = _packset.at(i);
981 if (p1 == NULL) continue;
982 for (int j = 0; j < _packset.length(); j++) {
983 Node_List* p2 = _packset.at(j);
984 if (p2 == NULL) continue;
985 if (i == j) continue;
986 if (p1->at(p1->size()-1) == p2->at(0)) {
987 for (uint k = 1; k < p2->size(); k++) {
988 p1->push(p2->at(k));
989 }
990 _packset.at_put(j, NULL);
991 changed = true;
992 }
993 }
994 }
995 }
997 // Split packs which have size greater then max vector size.
998 for (int i = 0; i < _packset.length(); i++) {
999 Node_List* p1 = _packset.at(i);
1000 if (p1 != NULL) {
1001 BasicType bt = velt_basic_type(p1->at(0));
1002 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1003 assert(is_power_of_2(max_vlen), "sanity");
1004 uint psize = p1->size();
1005 if (!is_power_of_2(psize)) {
1006 // Skip pack which can't be vector.
1007 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1008 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1009 _packset.at_put(i, NULL);
1010 continue;
1011 }
1012 if (psize > max_vlen) {
1013 Node_List* pack = new Node_List();
1014 for (uint j = 0; j < psize; j++) {
1015 pack->push(p1->at(j));
1016 if (pack->size() >= max_vlen) {
1017 assert(is_power_of_2(pack->size()), "sanity");
1018 _packset.append(pack);
1019 pack = new Node_List();
1020 }
1021 }
1022 _packset.at_put(i, NULL);
1023 }
1024 }
1025 }
1027 // Compress list.
1028 for (int i = _packset.length() - 1; i >= 0; i--) {
1029 Node_List* p1 = _packset.at(i);
1030 if (p1 == NULL) {
1031 _packset.remove_at(i);
1032 }
1033 }
1035 #ifndef PRODUCT
1036 if (TraceSuperWord) {
1037 tty->print_cr("\nAfter combine_packs");
1038 print_packset();
1039 }
1040 #endif
1041 }
1043 //-----------------------------construct_my_pack_map--------------------------
1044 // Construct the map from nodes to packs. Only valid after the
1045 // point where a node is only in one pack (after combine_packs).
1046 void SuperWord::construct_my_pack_map() {
1047 Node_List* rslt = NULL;
1048 for (int i = 0; i < _packset.length(); i++) {
1049 Node_List* p = _packset.at(i);
1050 for (uint j = 0; j < p->size(); j++) {
1051 Node* s = p->at(j);
1052 assert(my_pack(s) == NULL, "only in one pack");
1053 set_my_pack(s, p);
1054 }
1055 }
1056 }
1058 //------------------------------filter_packs---------------------------
1059 // Remove packs that are not implemented or not profitable.
1060 void SuperWord::filter_packs() {
1062 // Remove packs that are not implemented
1063 for (int i = _packset.length() - 1; i >= 0; i--) {
1064 Node_List* pk = _packset.at(i);
1065 bool impl = implemented(pk);
1066 if (!impl) {
1067 #ifndef PRODUCT
1068 if (TraceSuperWord && Verbose) {
1069 tty->print_cr("Unimplemented");
1070 pk->at(0)->dump();
1071 }
1072 #endif
1073 remove_pack_at(i);
1074 }
1075 }
1077 // Remove packs that are not profitable
1078 bool changed;
1079 do {
1080 changed = false;
1081 for (int i = _packset.length() - 1; i >= 0; i--) {
1082 Node_List* pk = _packset.at(i);
1083 bool prof = profitable(pk);
1084 if (!prof) {
1085 #ifndef PRODUCT
1086 if (TraceSuperWord && Verbose) {
1087 tty->print_cr("Unprofitable");
1088 pk->at(0)->dump();
1089 }
1090 #endif
1091 remove_pack_at(i);
1092 changed = true;
1093 }
1094 }
1095 } while (changed);
1097 #ifndef PRODUCT
1098 if (TraceSuperWord) {
1099 tty->print_cr("\nAfter filter_packs");
1100 print_packset();
1101 tty->cr();
1102 }
1103 #endif
1104 }
1106 //------------------------------implemented---------------------------
1107 // Can code be generated for pack p?
1108 bool SuperWord::implemented(Node_List* p) {
1109 Node* p0 = p->at(0);
1110 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1111 }
1113 //------------------------------same_inputs--------------------------
1114 // For pack p, are all idx operands the same?
1115 static bool same_inputs(Node_List* p, int idx) {
1116 Node* p0 = p->at(0);
1117 uint vlen = p->size();
1118 Node* p0_def = p0->in(idx);
1119 for (uint i = 1; i < vlen; i++) {
1120 Node* pi = p->at(i);
1121 Node* pi_def = pi->in(idx);
1122 if (p0_def != pi_def)
1123 return false;
1124 }
1125 return true;
1126 }
1128 //------------------------------profitable---------------------------
1129 // For pack p, are all operands and all uses (with in the block) vector?
1130 bool SuperWord::profitable(Node_List* p) {
1131 Node* p0 = p->at(0);
1132 uint start, end;
1133 VectorNode::vector_operands(p0, &start, &end);
1135 // Return false if some inputs are not vectors or vectors with different
1136 // size or alignment.
1137 // Also, for now, return false if not scalar promotion case when inputs are
1138 // the same. Later, implement PackNode and allow differing, non-vector inputs
1139 // (maybe just the ones from outside the block.)
1140 for (uint i = start; i < end; i++) {
1141 if (!is_vector_use(p0, i))
1142 return false;
1143 }
1144 if (VectorNode::is_shift(p0)) {
1145 // For now, return false if shift count is vector or not scalar promotion
1146 // case (different shift counts) because it is not supported yet.
1147 Node* cnt = p0->in(2);
1148 Node_List* cnt_pk = my_pack(cnt);
1149 if (cnt_pk != NULL)
1150 return false;
1151 if (!same_inputs(p, 2))
1152 return false;
1153 }
1154 if (!p0->is_Store()) {
1155 // For now, return false if not all uses are vector.
1156 // Later, implement ExtractNode and allow non-vector uses (maybe
1157 // just the ones outside the block.)
1158 for (uint i = 0; i < p->size(); i++) {
1159 Node* def = p->at(i);
1160 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1161 Node* use = def->fast_out(j);
1162 for (uint k = 0; k < use->req(); k++) {
1163 Node* n = use->in(k);
1164 if (def == n) {
1165 if (!is_vector_use(use, k)) {
1166 return false;
1167 }
1168 }
1169 }
1170 }
1171 }
1172 }
1173 return true;
1174 }
1176 //------------------------------schedule---------------------------
1177 // Adjust the memory graph for the packed operations
1178 void SuperWord::schedule() {
1180 // Co-locate in the memory graph the members of each memory pack
1181 for (int i = 0; i < _packset.length(); i++) {
1182 co_locate_pack(_packset.at(i));
1183 }
1184 }
1186 //-------------------------------remove_and_insert-------------------
1187 // Remove "current" from its current position in the memory graph and insert
1188 // it after the appropriate insertion point (lip or uip).
1189 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1190 Node *uip, Unique_Node_List &sched_before) {
1191 Node* my_mem = current->in(MemNode::Memory);
1192 bool sched_up = sched_before.member(current);
1194 // remove current_store from its current position in the memmory graph
1195 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1196 Node* use = current->out(i);
1197 if (use->is_Mem()) {
1198 assert(use->in(MemNode::Memory) == current, "must be");
1199 if (use == prev) { // connect prev to my_mem
1200 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1201 --i; //deleted this edge; rescan position
1202 } else if (sched_before.member(use)) {
1203 if (!sched_up) { // Will be moved together with current
1204 _igvn.replace_input_of(use, MemNode::Memory, uip);
1205 --i; //deleted this edge; rescan position
1206 }
1207 } else {
1208 if (sched_up) { // Will be moved together with current
1209 _igvn.replace_input_of(use, MemNode::Memory, lip);
1210 --i; //deleted this edge; rescan position
1211 }
1212 }
1213 }
1214 }
1216 Node *insert_pt = sched_up ? uip : lip;
1218 // all uses of insert_pt's memory state should use current's instead
1219 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1220 Node* use = insert_pt->out(i);
1221 if (use->is_Mem()) {
1222 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1223 _igvn.replace_input_of(use, MemNode::Memory, current);
1224 --i; //deleted this edge; rescan position
1225 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1226 uint pos; //lip (lower insert point) must be the last one in the memory slice
1227 for (pos=1; pos < use->req(); pos++) {
1228 if (use->in(pos) == insert_pt) break;
1229 }
1230 _igvn.replace_input_of(use, pos, current);
1231 --i;
1232 }
1233 }
1235 //connect current to insert_pt
1236 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1237 }
1239 //------------------------------co_locate_pack----------------------------------
1240 // To schedule a store pack, we need to move any sandwiched memory ops either before
1241 // or after the pack, based upon dependence information:
1242 // (1) If any store in the pack depends on the sandwiched memory op, the
1243 // sandwiched memory op must be scheduled BEFORE the pack;
1244 // (2) If a sandwiched memory op depends on any store in the pack, the
1245 // sandwiched memory op must be scheduled AFTER the pack;
1246 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1247 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1248 // scheduled before the pack, memB must also be scheduled before the pack;
1249 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1250 // schedule this store AFTER the pack
1251 // (5) We know there is no dependence cycle, so there in no other case;
1252 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1253 //
1254 // To schedule a load pack, we use the memory state of either the first or the last load in
1255 // the pack, based on the dependence constraint.
1256 void SuperWord::co_locate_pack(Node_List* pk) {
1257 if (pk->at(0)->is_Store()) {
1258 MemNode* first = executed_first(pk)->as_Mem();
1259 MemNode* last = executed_last(pk)->as_Mem();
1260 Unique_Node_List schedule_before_pack;
1261 Unique_Node_List memops;
1263 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1264 MemNode* previous = last;
1265 while (true) {
1266 assert(in_bb(current), "stay in block");
1267 memops.push(previous);
1268 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1269 Node* use = current->out(i);
1270 if (use->is_Mem() && use != previous)
1271 memops.push(use);
1272 }
1273 if (current == first) break;
1274 previous = current;
1275 current = current->in(MemNode::Memory)->as_Mem();
1276 }
1278 // determine which memory operations should be scheduled before the pack
1279 for (uint i = 1; i < memops.size(); i++) {
1280 Node *s1 = memops.at(i);
1281 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1282 for (uint j = 0; j< i; j++) {
1283 Node *s2 = memops.at(j);
1284 if (!independent(s1, s2)) {
1285 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1286 schedule_before_pack.push(s1); // s1 must be scheduled before
1287 Node_List* mem_pk = my_pack(s1);
1288 if (mem_pk != NULL) {
1289 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1290 Node* s = mem_pk->at(ii); // follow partner
1291 if (memops.member(s) && !schedule_before_pack.member(s))
1292 schedule_before_pack.push(s);
1293 }
1294 }
1295 break;
1296 }
1297 }
1298 }
1299 }
1300 }
1302 Node* upper_insert_pt = first->in(MemNode::Memory);
1303 // Following code moves loads connected to upper_insert_pt below aliased stores.
1304 // Collect such loads here and reconnect them back to upper_insert_pt later.
1305 memops.clear();
1306 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1307 Node* use = upper_insert_pt->out(i);
1308 if (use->is_Mem() && !use->is_Store()) {
1309 memops.push(use);
1310 }
1311 }
1313 MemNode* lower_insert_pt = last;
1314 previous = last; //previous store in pk
1315 current = last->in(MemNode::Memory)->as_Mem();
1317 // start scheduling from "last" to "first"
1318 while (true) {
1319 assert(in_bb(current), "stay in block");
1320 assert(in_pack(previous, pk), "previous stays in pack");
1321 Node* my_mem = current->in(MemNode::Memory);
1323 if (in_pack(current, pk)) {
1324 // Forward users of my memory state (except "previous) to my input memory state
1325 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1326 Node* use = current->out(i);
1327 if (use->is_Mem() && use != previous) {
1328 assert(use->in(MemNode::Memory) == current, "must be");
1329 if (schedule_before_pack.member(use)) {
1330 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1331 } else {
1332 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1333 }
1334 --i; // deleted this edge; rescan position
1335 }
1336 }
1337 previous = current;
1338 } else { // !in_pack(current, pk) ==> a sandwiched store
1339 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1340 }
1342 if (current == first) break;
1343 current = my_mem->as_Mem();
1344 } // end while
1346 // Reconnect loads back to upper_insert_pt.
1347 for (uint i = 0; i < memops.size(); i++) {
1348 Node *ld = memops.at(i);
1349 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1350 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1351 }
1352 }
1353 } else if (pk->at(0)->is_Load()) { //load
1354 // all loads in the pack should have the same memory state. By default,
1355 // we use the memory state of the last load. However, if any load could
1356 // not be moved down due to the dependence constraint, we use the memory
1357 // state of the first load.
1358 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1359 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1360 bool schedule_last = true;
1361 for (uint i = 0; i < pk->size(); i++) {
1362 Node* ld = pk->at(i);
1363 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1364 current=current->in(MemNode::Memory)) {
1365 assert(current != first_mem, "corrupted memory graph");
1366 if(current->is_Mem() && !independent(current, ld)){
1367 schedule_last = false; // a later store depends on this load
1368 break;
1369 }
1370 }
1371 }
1373 Node* mem_input = schedule_last ? last_mem : first_mem;
1374 _igvn.hash_delete(mem_input);
1375 // Give each load the same memory state
1376 for (uint i = 0; i < pk->size(); i++) {
1377 LoadNode* ld = pk->at(i)->as_Load();
1378 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1379 }
1380 }
1381 }
1383 //------------------------------output---------------------------
1384 // Convert packs into vector node operations
1385 void SuperWord::output() {
1386 if (_packset.length() == 0) return;
1388 #ifndef PRODUCT
1389 if (TraceLoopOpts) {
1390 tty->print("SuperWord ");
1391 lpt()->dump_head();
1392 }
1393 #endif
1395 // MUST ENSURE main loop's initial value is properly aligned:
1396 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1398 align_initial_loop_index(align_to_ref());
1400 // Insert extract (unpack) operations for scalar uses
1401 for (int i = 0; i < _packset.length(); i++) {
1402 insert_extracts(_packset.at(i));
1403 }
1405 Compile* C = _phase->C;
1406 uint max_vlen_in_bytes = 0;
1407 for (int i = 0; i < _block.length(); i++) {
1408 Node* n = _block.at(i);
1409 Node_List* p = my_pack(n);
1410 if (p && n == executed_last(p)) {
1411 uint vlen = p->size();
1412 uint vlen_in_bytes = 0;
1413 Node* vn = NULL;
1414 Node* low_adr = p->at(0);
1415 Node* first = executed_first(p);
1416 int opc = n->Opcode();
1417 if (n->is_Load()) {
1418 Node* ctl = n->in(MemNode::Control);
1419 Node* mem = first->in(MemNode::Memory);
1420 SWPointer p1(n->as_Mem(), this);
1421 // Identify the memory dependency for the new loadVector node by
1422 // walking up through memory chain.
1423 // This is done to give flexibility to the new loadVector node so that
1424 // it can move above independent storeVector nodes.
1425 while (mem->is_StoreVector()) {
1426 SWPointer p2(mem->as_Mem(), this);
1427 int cmp = p1.cmp(p2);
1428 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1429 mem = mem->in(MemNode::Memory);
1430 } else {
1431 break; // dependent memory
1432 }
1433 }
1434 Node* adr = low_adr->in(MemNode::Address);
1435 const TypePtr* atyp = n->adr_type();
1436 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
1437 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1438 } else if (n->is_Store()) {
1439 // Promote value to be stored to vector
1440 Node* val = vector_opd(p, MemNode::ValueIn);
1441 Node* ctl = n->in(MemNode::Control);
1442 Node* mem = first->in(MemNode::Memory);
1443 Node* adr = low_adr->in(MemNode::Address);
1444 const TypePtr* atyp = n->adr_type();
1445 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1446 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1447 } else if (n->req() == 3) {
1448 // Promote operands to vector
1449 Node* in1 = vector_opd(p, 1);
1450 Node* in2 = vector_opd(p, 2);
1451 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1452 // Move invariant vector input into second position to avoid register spilling.
1453 Node* tmp = in1;
1454 in1 = in2;
1455 in2 = tmp;
1456 }
1457 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1458 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1459 } else {
1460 ShouldNotReachHere();
1461 }
1462 assert(vn != NULL, "sanity");
1463 _igvn.register_new_node_with_optimizer(vn);
1464 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1465 for (uint j = 0; j < p->size(); j++) {
1466 Node* pm = p->at(j);
1467 _igvn.replace_node(pm, vn);
1468 }
1469 _igvn._worklist.push(vn);
1471 if (vlen_in_bytes > max_vlen_in_bytes) {
1472 max_vlen_in_bytes = vlen_in_bytes;
1473 }
1474 #ifdef ASSERT
1475 if (TraceNewVectors) {
1476 tty->print("new Vector node: ");
1477 vn->dump();
1478 }
1479 #endif
1480 }
1481 }
1482 C->set_max_vector_size(max_vlen_in_bytes);
1483 }
1485 //------------------------------vector_opd---------------------------
1486 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1487 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1488 Node* p0 = p->at(0);
1489 uint vlen = p->size();
1490 Node* opd = p0->in(opd_idx);
1492 if (same_inputs(p, opd_idx)) {
1493 if (opd->is_Vector() || opd->is_LoadVector()) {
1494 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1495 return opd; // input is matching vector
1496 }
1497 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1498 Compile* C = _phase->C;
1499 Node* cnt = opd;
1500 // Vector instructions do not mask shift count, do it here.
1501 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1502 const TypeInt* t = opd->find_int_type();
1503 if (t != NULL && t->is_con()) {
1504 juint shift = t->get_con();
1505 if (shift > mask) { // Unsigned cmp
1506 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1507 }
1508 } else {
1509 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1510 cnt = ConNode::make(C, TypeInt::make(mask));
1511 _igvn.register_new_node_with_optimizer(cnt);
1512 cnt = new (C) AndINode(opd, cnt);
1513 _igvn.register_new_node_with_optimizer(cnt);
1514 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1515 }
1516 assert(opd->bottom_type()->isa_int(), "int type only");
1517 // Move non constant shift count into vector register.
1518 cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1519 }
1520 if (cnt != opd) {
1521 _igvn.register_new_node_with_optimizer(cnt);
1522 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1523 }
1524 return cnt;
1525 }
1526 assert(!opd->is_StoreVector(), "such vector is not expected here");
1527 // Convert scalar input to vector with the same number of elements as
1528 // p0's vector. Use p0's type because size of operand's container in
1529 // vector should match p0's size regardless operand's size.
1530 const Type* p0_t = velt_type(p0);
1531 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1533 _igvn.register_new_node_with_optimizer(vn);
1534 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1535 #ifdef ASSERT
1536 if (TraceNewVectors) {
1537 tty->print("new Vector node: ");
1538 vn->dump();
1539 }
1540 #endif
1541 return vn;
1542 }
1544 // Insert pack operation
1545 BasicType bt = velt_basic_type(p0);
1546 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1547 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1549 for (uint i = 1; i < vlen; i++) {
1550 Node* pi = p->at(i);
1551 Node* in = pi->in(opd_idx);
1552 assert(my_pack(in) == NULL, "Should already have been unpacked");
1553 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1554 pk->add_opd(in);
1555 }
1556 _igvn.register_new_node_with_optimizer(pk);
1557 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1558 #ifdef ASSERT
1559 if (TraceNewVectors) {
1560 tty->print("new Vector node: ");
1561 pk->dump();
1562 }
1563 #endif
1564 return pk;
1565 }
1567 //------------------------------insert_extracts---------------------------
1568 // If a use of pack p is not a vector use, then replace the
1569 // use with an extract operation.
1570 void SuperWord::insert_extracts(Node_List* p) {
1571 if (p->at(0)->is_Store()) return;
1572 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1574 // Inspect each use of each pack member. For each use that is
1575 // not a vector use, replace the use with an extract operation.
1577 for (uint i = 0; i < p->size(); i++) {
1578 Node* def = p->at(i);
1579 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1580 Node* use = def->fast_out(j);
1581 for (uint k = 0; k < use->req(); k++) {
1582 Node* n = use->in(k);
1583 if (def == n) {
1584 if (!is_vector_use(use, k)) {
1585 _n_idx_list.push(use, k);
1586 }
1587 }
1588 }
1589 }
1590 }
1592 while (_n_idx_list.is_nonempty()) {
1593 Node* use = _n_idx_list.node();
1594 int idx = _n_idx_list.index();
1595 _n_idx_list.pop();
1596 Node* def = use->in(idx);
1598 // Insert extract operation
1599 _igvn.hash_delete(def);
1600 int def_pos = alignment(def) / data_size(def);
1602 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1603 _igvn.register_new_node_with_optimizer(ex);
1604 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1605 _igvn.replace_input_of(use, idx, ex);
1606 _igvn._worklist.push(def);
1608 bb_insert_after(ex, bb_idx(def));
1609 set_velt_type(ex, velt_type(def));
1610 }
1611 }
1613 //------------------------------is_vector_use---------------------------
1614 // Is use->in(u_idx) a vector use?
1615 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1616 Node_List* u_pk = my_pack(use);
1617 if (u_pk == NULL) return false;
1618 Node* def = use->in(u_idx);
1619 Node_List* d_pk = my_pack(def);
1620 if (d_pk == NULL) {
1621 // check for scalar promotion
1622 Node* n = u_pk->at(0)->in(u_idx);
1623 for (uint i = 1; i < u_pk->size(); i++) {
1624 if (u_pk->at(i)->in(u_idx) != n) return false;
1625 }
1626 return true;
1627 }
1628 if (u_pk->size() != d_pk->size())
1629 return false;
1630 for (uint i = 0; i < u_pk->size(); i++) {
1631 Node* ui = u_pk->at(i);
1632 Node* di = d_pk->at(i);
1633 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1634 return false;
1635 }
1636 return true;
1637 }
1639 //------------------------------construct_bb---------------------------
1640 // Construct reverse postorder list of block members
1641 bool SuperWord::construct_bb() {
1642 Node* entry = bb();
1644 assert(_stk.length() == 0, "stk is empty");
1645 assert(_block.length() == 0, "block is empty");
1646 assert(_data_entry.length() == 0, "data_entry is empty");
1647 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1648 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1650 // Find non-control nodes with no inputs from within block,
1651 // create a temporary map from node _idx to bb_idx for use
1652 // by the visited and post_visited sets,
1653 // and count number of nodes in block.
1654 int bb_ct = 0;
1655 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1656 Node *n = lpt()->_body.at(i);
1657 set_bb_idx(n, i); // Create a temporary map
1658 if (in_bb(n)) {
1659 if (n->is_LoadStore() || n->is_MergeMem() ||
1660 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1661 // Bailout if the loop has LoadStore, MergeMem or data Proj
1662 // nodes. Superword optimization does not work with them.
1663 return false;
1664 }
1665 bb_ct++;
1666 if (!n->is_CFG()) {
1667 bool found = false;
1668 for (uint j = 0; j < n->req(); j++) {
1669 Node* def = n->in(j);
1670 if (def && in_bb(def)) {
1671 found = true;
1672 break;
1673 }
1674 }
1675 if (!found) {
1676 assert(n != entry, "can't be entry");
1677 _data_entry.push(n);
1678 }
1679 }
1680 }
1681 }
1683 // Find memory slices (head and tail)
1684 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1685 Node *n = lp()->fast_out(i);
1686 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1687 Node* n_tail = n->in(LoopNode::LoopBackControl);
1688 if (n_tail != n->in(LoopNode::EntryControl)) {
1689 if (!n_tail->is_Mem()) {
1690 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1691 return false; // Bailout
1692 }
1693 _mem_slice_head.push(n);
1694 _mem_slice_tail.push(n_tail);
1695 }
1696 }
1697 }
1699 // Create an RPO list of nodes in block
1701 visited_clear();
1702 post_visited_clear();
1704 // Push all non-control nodes with no inputs from within block, then control entry
1705 for (int j = 0; j < _data_entry.length(); j++) {
1706 Node* n = _data_entry.at(j);
1707 visited_set(n);
1708 _stk.push(n);
1709 }
1710 visited_set(entry);
1711 _stk.push(entry);
1713 // Do a depth first walk over out edges
1714 int rpo_idx = bb_ct - 1;
1715 int size;
1716 while ((size = _stk.length()) > 0) {
1717 Node* n = _stk.top(); // Leave node on stack
1718 if (!visited_test_set(n)) {
1719 // forward arc in graph
1720 } else if (!post_visited_test(n)) {
1721 // cross or back arc
1722 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1723 Node *use = n->fast_out(i);
1724 if (in_bb(use) && !visited_test(use) &&
1725 // Don't go around backedge
1726 (!use->is_Phi() || n == entry)) {
1727 _stk.push(use);
1728 }
1729 }
1730 if (_stk.length() == size) {
1731 // There were no additional uses, post visit node now
1732 _stk.pop(); // Remove node from stack
1733 assert(rpo_idx >= 0, "");
1734 _block.at_put_grow(rpo_idx, n);
1735 rpo_idx--;
1736 post_visited_set(n);
1737 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1738 }
1739 } else {
1740 _stk.pop(); // Remove post-visited node from stack
1741 }
1742 }
1744 // Create real map of block indices for nodes
1745 for (int j = 0; j < _block.length(); j++) {
1746 Node* n = _block.at(j);
1747 set_bb_idx(n, j);
1748 }
1750 initialize_bb(); // Ensure extra info is allocated.
1752 #ifndef PRODUCT
1753 if (TraceSuperWord) {
1754 print_bb();
1755 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1756 for (int m = 0; m < _data_entry.length(); m++) {
1757 tty->print("%3d ", m);
1758 _data_entry.at(m)->dump();
1759 }
1760 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1761 for (int m = 0; m < _mem_slice_head.length(); m++) {
1762 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1763 tty->print(" "); _mem_slice_tail.at(m)->dump();
1764 }
1765 }
1766 #endif
1767 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1768 return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1769 }
1771 //------------------------------initialize_bb---------------------------
1772 // Initialize per node info
1773 void SuperWord::initialize_bb() {
1774 Node* last = _block.at(_block.length() - 1);
1775 grow_node_info(bb_idx(last));
1776 }
1778 //------------------------------bb_insert_after---------------------------
1779 // Insert n into block after pos
1780 void SuperWord::bb_insert_after(Node* n, int pos) {
1781 int n_pos = pos + 1;
1782 // Make room
1783 for (int i = _block.length() - 1; i >= n_pos; i--) {
1784 _block.at_put_grow(i+1, _block.at(i));
1785 }
1786 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1787 _node_info.at_put_grow(j+1, _node_info.at(j));
1788 }
1789 // Set value
1790 _block.at_put_grow(n_pos, n);
1791 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1792 // Adjust map from node->_idx to _block index
1793 for (int i = n_pos; i < _block.length(); i++) {
1794 set_bb_idx(_block.at(i), i);
1795 }
1796 }
1798 //------------------------------compute_max_depth---------------------------
1799 // Compute max depth for expressions from beginning of block
1800 // Use to prune search paths during test for independence.
1801 void SuperWord::compute_max_depth() {
1802 int ct = 0;
1803 bool again;
1804 do {
1805 again = false;
1806 for (int i = 0; i < _block.length(); i++) {
1807 Node* n = _block.at(i);
1808 if (!n->is_Phi()) {
1809 int d_orig = depth(n);
1810 int d_in = 0;
1811 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1812 Node* pred = preds.current();
1813 if (in_bb(pred)) {
1814 d_in = MAX2(d_in, depth(pred));
1815 }
1816 }
1817 if (d_in + 1 != d_orig) {
1818 set_depth(n, d_in + 1);
1819 again = true;
1820 }
1821 }
1822 }
1823 ct++;
1824 } while (again);
1825 #ifndef PRODUCT
1826 if (TraceSuperWord && Verbose)
1827 tty->print_cr("compute_max_depth iterated: %d times", ct);
1828 #endif
1829 }
1831 //-------------------------compute_vector_element_type-----------------------
1832 // Compute necessary vector element type for expressions
1833 // This propagates backwards a narrower integer type when the
1834 // upper bits of the value are not needed.
1835 // Example: char a,b,c; a = b + c;
1836 // Normally the type of the add is integer, but for packed character
1837 // operations the type of the add needs to be char.
1838 void SuperWord::compute_vector_element_type() {
1839 #ifndef PRODUCT
1840 if (TraceSuperWord && Verbose)
1841 tty->print_cr("\ncompute_velt_type:");
1842 #endif
1844 // Initial type
1845 for (int i = 0; i < _block.length(); i++) {
1846 Node* n = _block.at(i);
1847 set_velt_type(n, container_type(n));
1848 }
1850 // Propagate integer narrowed type backwards through operations
1851 // that don't depend on higher order bits
1852 for (int i = _block.length() - 1; i >= 0; i--) {
1853 Node* n = _block.at(i);
1854 // Only integer types need be examined
1855 const Type* vtn = velt_type(n);
1856 if (vtn->basic_type() == T_INT) {
1857 uint start, end;
1858 VectorNode::vector_operands(n, &start, &end);
1860 for (uint j = start; j < end; j++) {
1861 Node* in = n->in(j);
1862 // Don't propagate through a memory
1863 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1864 data_size(n) < data_size(in)) {
1865 bool same_type = true;
1866 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1867 Node *use = in->fast_out(k);
1868 if (!in_bb(use) || !same_velt_type(use, n)) {
1869 same_type = false;
1870 break;
1871 }
1872 }
1873 if (same_type) {
1874 // For right shifts of small integer types (bool, byte, char, short)
1875 // we need precise information about sign-ness. Only Load nodes have
1876 // this information because Store nodes are the same for signed and
1877 // unsigned values. And any arithmetic operation after a load may
1878 // expand a value to signed Int so such right shifts can't be used
1879 // because vector elements do not have upper bits of Int.
1880 const Type* vt = vtn;
1881 if (VectorNode::is_shift(in)) {
1882 Node* load = in->in(1);
1883 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1884 vt = velt_type(load);
1885 } else if (in->Opcode() != Op_LShiftI) {
1886 // Widen type to Int to avoid creation of right shift vector
1887 // (align + data_size(s1) check in stmts_can_pack() will fail).
1888 // Note, left shifts work regardless type.
1889 vt = TypeInt::INT;
1890 }
1891 }
1892 set_velt_type(in, vt);
1893 }
1894 }
1895 }
1896 }
1897 }
1898 #ifndef PRODUCT
1899 if (TraceSuperWord && Verbose) {
1900 for (int i = 0; i < _block.length(); i++) {
1901 Node* n = _block.at(i);
1902 velt_type(n)->dump();
1903 tty->print("\t");
1904 n->dump();
1905 }
1906 }
1907 #endif
1908 }
1910 //------------------------------memory_alignment---------------------------
1911 // Alignment within a vector memory reference
1912 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1913 SWPointer p(s, this);
1914 if (!p.valid()) {
1915 return bottom_align;
1916 }
1917 int vw = vector_width_in_bytes(s);
1918 if (vw < 2) {
1919 return bottom_align; // No vectors for this type
1920 }
1921 int offset = p.offset_in_bytes();
1922 offset += iv_adjust*p.memory_size();
1923 int off_rem = offset % vw;
1924 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1925 return off_mod;
1926 }
1928 //---------------------------container_type---------------------------
1929 // Smallest type containing range of values
1930 const Type* SuperWord::container_type(Node* n) {
1931 if (n->is_Mem()) {
1932 BasicType bt = n->as_Mem()->memory_type();
1933 if (n->is_Store() && (bt == T_CHAR)) {
1934 // Use T_SHORT type instead of T_CHAR for stored values because any
1935 // preceding arithmetic operation extends values to signed Int.
1936 bt = T_SHORT;
1937 }
1938 if (n->Opcode() == Op_LoadUB) {
1939 // Adjust type for unsigned byte loads, it is important for right shifts.
1940 // T_BOOLEAN is used because there is no basic type representing type
1941 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1942 // size (one byte) and sign is important.
1943 bt = T_BOOLEAN;
1944 }
1945 return Type::get_const_basic_type(bt);
1946 }
1947 const Type* t = _igvn.type(n);
1948 if (t->basic_type() == T_INT) {
1949 // A narrow type of arithmetic operations will be determined by
1950 // propagating the type of memory operations.
1951 return TypeInt::INT;
1952 }
1953 return t;
1954 }
1956 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1957 const Type* vt1 = velt_type(n1);
1958 const Type* vt2 = velt_type(n2);
1959 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1960 // Compare vectors element sizes for integer types.
1961 return data_size(n1) == data_size(n2);
1962 }
1963 return vt1 == vt2;
1964 }
1966 //------------------------------in_packset---------------------------
1967 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1968 bool SuperWord::in_packset(Node* s1, Node* s2) {
1969 for (int i = 0; i < _packset.length(); i++) {
1970 Node_List* p = _packset.at(i);
1971 assert(p->size() == 2, "must be");
1972 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1973 return true;
1974 }
1975 }
1976 return false;
1977 }
1979 //------------------------------in_pack---------------------------
1980 // Is s in pack p?
1981 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1982 for (uint i = 0; i < p->size(); i++) {
1983 if (p->at(i) == s) {
1984 return p;
1985 }
1986 }
1987 return NULL;
1988 }
1990 //------------------------------remove_pack_at---------------------------
1991 // Remove the pack at position pos in the packset
1992 void SuperWord::remove_pack_at(int pos) {
1993 Node_List* p = _packset.at(pos);
1994 for (uint i = 0; i < p->size(); i++) {
1995 Node* s = p->at(i);
1996 set_my_pack(s, NULL);
1997 }
1998 _packset.remove_at(pos);
1999 }
2001 //------------------------------executed_first---------------------------
2002 // Return the node executed first in pack p. Uses the RPO block list
2003 // to determine order.
2004 Node* SuperWord::executed_first(Node_List* p) {
2005 Node* n = p->at(0);
2006 int n_rpo = bb_idx(n);
2007 for (uint i = 1; i < p->size(); i++) {
2008 Node* s = p->at(i);
2009 int s_rpo = bb_idx(s);
2010 if (s_rpo < n_rpo) {
2011 n = s;
2012 n_rpo = s_rpo;
2013 }
2014 }
2015 return n;
2016 }
2018 //------------------------------executed_last---------------------------
2019 // Return the node executed last in pack p.
2020 Node* SuperWord::executed_last(Node_List* p) {
2021 Node* n = p->at(0);
2022 int n_rpo = bb_idx(n);
2023 for (uint i = 1; i < p->size(); i++) {
2024 Node* s = p->at(i);
2025 int s_rpo = bb_idx(s);
2026 if (s_rpo > n_rpo) {
2027 n = s;
2028 n_rpo = s_rpo;
2029 }
2030 }
2031 return n;
2032 }
2034 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2035 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2036 for (uint i = 0; i < p->size(); i++) {
2037 Node* n = p->at(i);
2038 assert(n->is_Load(), "only meaningful for loads");
2039 if (!n->depends_only_on_test()) {
2040 dep = LoadNode::Pinned;
2041 }
2042 }
2043 return dep;
2044 }
2047 //----------------------------align_initial_loop_index---------------------------
2048 // Adjust pre-loop limit so that in main loop, a load/store reference
2049 // to align_to_ref will be a position zero in the vector.
2050 // (iv + k) mod vector_align == 0
2051 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2052 CountedLoopNode *main_head = lp()->as_CountedLoop();
2053 assert(main_head->is_main_loop(), "");
2054 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2055 assert(pre_end != NULL, "");
2056 Node *pre_opaq1 = pre_end->limit();
2057 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2058 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2059 Node *lim0 = pre_opaq->in(1);
2061 // Where we put new limit calculations
2062 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2064 // Ensure the original loop limit is available from the
2065 // pre-loop Opaque1 node.
2066 Node *orig_limit = pre_opaq->original_loop_limit();
2067 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2069 SWPointer align_to_ref_p(align_to_ref, this);
2070 assert(align_to_ref_p.valid(), "sanity");
2072 // Given:
2073 // lim0 == original pre loop limit
2074 // V == v_align (power of 2)
2075 // invar == extra invariant piece of the address expression
2076 // e == offset [ +/- invar ]
2077 //
2078 // When reassociating expressions involving '%' the basic rules are:
2079 // (a - b) % k == 0 => a % k == b % k
2080 // and:
2081 // (a + b) % k == 0 => a % k == (k - b) % k
2082 //
2083 // For stride > 0 && scale > 0,
2084 // Derive the new pre-loop limit "lim" such that the two constraints:
2085 // (1) lim = lim0 + N (where N is some positive integer < V)
2086 // (2) (e + lim) % V == 0
2087 // are true.
2088 //
2089 // Substituting (1) into (2),
2090 // (e + lim0 + N) % V == 0
2091 // solve for N:
2092 // N = (V - (e + lim0)) % V
2093 // substitute back into (1), so that new limit
2094 // lim = lim0 + (V - (e + lim0)) % V
2095 //
2096 // For stride > 0 && scale < 0
2097 // Constraints:
2098 // lim = lim0 + N
2099 // (e - lim) % V == 0
2100 // Solving for lim:
2101 // (e - lim0 - N) % V == 0
2102 // N = (e - lim0) % V
2103 // lim = lim0 + (e - lim0) % V
2104 //
2105 // For stride < 0 && scale > 0
2106 // Constraints:
2107 // lim = lim0 - N
2108 // (e + lim) % V == 0
2109 // Solving for lim:
2110 // (e + lim0 - N) % V == 0
2111 // N = (e + lim0) % V
2112 // lim = lim0 - (e + lim0) % V
2113 //
2114 // For stride < 0 && scale < 0
2115 // Constraints:
2116 // lim = lim0 - N
2117 // (e - lim) % V == 0
2118 // Solving for lim:
2119 // (e - lim0 + N) % V == 0
2120 // N = (V - (e - lim0)) % V
2121 // lim = lim0 - (V - (e - lim0)) % V
2123 int vw = vector_width_in_bytes(align_to_ref);
2124 int stride = iv_stride();
2125 int scale = align_to_ref_p.scale_in_bytes();
2126 int elt_size = align_to_ref_p.memory_size();
2127 int v_align = vw / elt_size;
2128 assert(v_align > 1, "sanity");
2129 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2130 Node *offsn = _igvn.intcon(offset);
2132 Node *e = offsn;
2133 if (align_to_ref_p.invar() != NULL) {
2134 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2135 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2136 Node* aref = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2137 _igvn.register_new_node_with_optimizer(aref);
2138 _phase->set_ctrl(aref, pre_ctrl);
2139 if (align_to_ref_p.negate_invar()) {
2140 e = new (_phase->C) SubINode(e, aref);
2141 } else {
2142 e = new (_phase->C) AddINode(e, aref);
2143 }
2144 _igvn.register_new_node_with_optimizer(e);
2145 _phase->set_ctrl(e, pre_ctrl);
2146 }
2147 if (vw > ObjectAlignmentInBytes) {
2148 // incorporate base e +/- base && Mask >>> log2(elt)
2149 Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2150 _igvn.register_new_node_with_optimizer(xbase);
2151 #ifdef _LP64
2152 xbase = new (_phase->C) ConvL2INode(xbase);
2153 _igvn.register_new_node_with_optimizer(xbase);
2154 #endif
2155 Node* mask = _igvn.intcon(vw-1);
2156 Node* masked_xbase = new (_phase->C) AndINode(xbase, mask);
2157 _igvn.register_new_node_with_optimizer(masked_xbase);
2158 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2159 Node* bref = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2160 _igvn.register_new_node_with_optimizer(bref);
2161 _phase->set_ctrl(bref, pre_ctrl);
2162 e = new (_phase->C) AddINode(e, bref);
2163 _igvn.register_new_node_with_optimizer(e);
2164 _phase->set_ctrl(e, pre_ctrl);
2165 }
2167 // compute e +/- lim0
2168 if (scale < 0) {
2169 e = new (_phase->C) SubINode(e, lim0);
2170 } else {
2171 e = new (_phase->C) AddINode(e, lim0);
2172 }
2173 _igvn.register_new_node_with_optimizer(e);
2174 _phase->set_ctrl(e, pre_ctrl);
2176 if (stride * scale > 0) {
2177 // compute V - (e +/- lim0)
2178 Node* va = _igvn.intcon(v_align);
2179 e = new (_phase->C) SubINode(va, e);
2180 _igvn.register_new_node_with_optimizer(e);
2181 _phase->set_ctrl(e, pre_ctrl);
2182 }
2183 // compute N = (exp) % V
2184 Node* va_msk = _igvn.intcon(v_align - 1);
2185 Node* N = new (_phase->C) AndINode(e, va_msk);
2186 _igvn.register_new_node_with_optimizer(N);
2187 _phase->set_ctrl(N, pre_ctrl);
2189 // substitute back into (1), so that new limit
2190 // lim = lim0 + N
2191 Node* lim;
2192 if (stride < 0) {
2193 lim = new (_phase->C) SubINode(lim0, N);
2194 } else {
2195 lim = new (_phase->C) AddINode(lim0, N);
2196 }
2197 _igvn.register_new_node_with_optimizer(lim);
2198 _phase->set_ctrl(lim, pre_ctrl);
2199 Node* constrained =
2200 (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2201 : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2202 _igvn.register_new_node_with_optimizer(constrained);
2203 _phase->set_ctrl(constrained, pre_ctrl);
2204 _igvn.hash_delete(pre_opaq);
2205 pre_opaq->set_req(1, constrained);
2206 }
2208 //----------------------------get_pre_loop_end---------------------------
2209 // Find pre loop end from main loop. Returns null if none.
2210 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2211 Node *ctrl = cl->in(LoopNode::EntryControl);
2212 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2213 Node *iffm = ctrl->in(0);
2214 if (!iffm->is_If()) return NULL;
2215 Node *p_f = iffm->in(0);
2216 if (!p_f->is_IfFalse()) return NULL;
2217 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2218 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2219 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2220 return pre_end;
2221 }
2224 //------------------------------init---------------------------
2225 void SuperWord::init() {
2226 _dg.init();
2227 _packset.clear();
2228 _disjoint_ptrs.clear();
2229 _block.clear();
2230 _data_entry.clear();
2231 _mem_slice_head.clear();
2232 _mem_slice_tail.clear();
2233 _node_info.clear();
2234 _align_to_ref = NULL;
2235 _lpt = NULL;
2236 _lp = NULL;
2237 _bb = NULL;
2238 _iv = NULL;
2239 }
2241 //------------------------------print_packset---------------------------
2242 void SuperWord::print_packset() {
2243 #ifndef PRODUCT
2244 tty->print_cr("packset");
2245 for (int i = 0; i < _packset.length(); i++) {
2246 tty->print_cr("Pack: %d", i);
2247 Node_List* p = _packset.at(i);
2248 print_pack(p);
2249 }
2250 #endif
2251 }
2253 //------------------------------print_pack---------------------------
2254 void SuperWord::print_pack(Node_List* p) {
2255 for (uint i = 0; i < p->size(); i++) {
2256 print_stmt(p->at(i));
2257 }
2258 }
2260 //------------------------------print_bb---------------------------
2261 void SuperWord::print_bb() {
2262 #ifndef PRODUCT
2263 tty->print_cr("\nBlock");
2264 for (int i = 0; i < _block.length(); i++) {
2265 Node* n = _block.at(i);
2266 tty->print("%d ", i);
2267 if (n) {
2268 n->dump();
2269 }
2270 }
2271 #endif
2272 }
2274 //------------------------------print_stmt---------------------------
2275 void SuperWord::print_stmt(Node* s) {
2276 #ifndef PRODUCT
2277 tty->print(" align: %d \t", alignment(s));
2278 s->dump();
2279 #endif
2280 }
2282 //------------------------------blank---------------------------
2283 char* SuperWord::blank(uint depth) {
2284 static char blanks[101];
2285 assert(depth < 101, "too deep");
2286 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2287 blanks[depth] = '\0';
2288 return blanks;
2289 }
2292 //==============================SWPointer===========================
2294 //----------------------------SWPointer------------------------
2295 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2296 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2297 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2299 Node* adr = mem->in(MemNode::Address);
2300 if (!adr->is_AddP()) {
2301 assert(!valid(), "too complex");
2302 return;
2303 }
2304 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2305 Node* base = adr->in(AddPNode::Base);
2306 // The base address should be loop invariant
2307 if (!invariant(base)) {
2308 assert(!valid(), "base address is loop variant");
2309 return;
2310 }
2311 //unsafe reference could not be aligned appropriately without runtime checking
2312 if (base == NULL || base->bottom_type() == Type::TOP) {
2313 assert(!valid(), "unsafe access");
2314 return;
2315 }
2316 for (int i = 0; i < 3; i++) {
2317 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2318 assert(!valid(), "too complex");
2319 return;
2320 }
2321 adr = adr->in(AddPNode::Address);
2322 if (base == adr || !adr->is_AddP()) {
2323 break; // stop looking at addp's
2324 }
2325 }
2326 _base = base;
2327 _adr = adr;
2328 assert(valid(), "Usable");
2329 }
2331 // Following is used to create a temporary object during
2332 // the pattern match of an address expression.
2333 SWPointer::SWPointer(SWPointer* p) :
2334 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2335 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2337 //------------------------scaled_iv_plus_offset--------------------
2338 // Match: k*iv + offset
2339 // where: k is a constant that maybe zero, and
2340 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2341 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2342 if (scaled_iv(n)) {
2343 return true;
2344 }
2345 if (offset_plus_k(n)) {
2346 return true;
2347 }
2348 int opc = n->Opcode();
2349 if (opc == Op_AddI) {
2350 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2351 return true;
2352 }
2353 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2354 return true;
2355 }
2356 } else if (opc == Op_SubI) {
2357 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2358 return true;
2359 }
2360 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2361 _scale *= -1;
2362 return true;
2363 }
2364 }
2365 return false;
2366 }
2368 //----------------------------scaled_iv------------------------
2369 // Match: k*iv where k is a constant that's not zero
2370 bool SWPointer::scaled_iv(Node* n) {
2371 if (_scale != 0) {
2372 return false; // already found a scale
2373 }
2374 if (n == iv()) {
2375 _scale = 1;
2376 return true;
2377 }
2378 int opc = n->Opcode();
2379 if (opc == Op_MulI) {
2380 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2381 _scale = n->in(2)->get_int();
2382 return true;
2383 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2384 _scale = n->in(1)->get_int();
2385 return true;
2386 }
2387 } else if (opc == Op_LShiftI) {
2388 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2389 _scale = 1 << n->in(2)->get_int();
2390 return true;
2391 }
2392 } else if (opc == Op_ConvI2L) {
2393 if (n->in(1)->Opcode() == Op_CastII &&
2394 n->in(1)->as_CastII()->has_range_check()) {
2395 // Skip range check dependent CastII nodes
2396 n = n->in(1);
2397 }
2398 if (scaled_iv_plus_offset(n->in(1))) {
2399 return true;
2400 }
2401 } else if (opc == Op_LShiftL) {
2402 if (!has_iv() && _invar == NULL) {
2403 // Need to preserve the current _offset value, so
2404 // create a temporary object for this expression subtree.
2405 // Hacky, so should re-engineer the address pattern match.
2406 SWPointer tmp(this);
2407 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2408 if (tmp._invar == NULL) {
2409 int mult = 1 << n->in(2)->get_int();
2410 _scale = tmp._scale * mult;
2411 _offset += tmp._offset * mult;
2412 return true;
2413 }
2414 }
2415 }
2416 }
2417 return false;
2418 }
2420 //----------------------------offset_plus_k------------------------
2421 // Match: offset is (k [+/- invariant])
2422 // where k maybe zero and invariant is optional, but not both.
2423 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2424 int opc = n->Opcode();
2425 if (opc == Op_ConI) {
2426 _offset += negate ? -(n->get_int()) : n->get_int();
2427 return true;
2428 } else if (opc == Op_ConL) {
2429 // Okay if value fits into an int
2430 const TypeLong* t = n->find_long_type();
2431 if (t->higher_equal(TypeLong::INT)) {
2432 jlong loff = n->get_long();
2433 jint off = (jint)loff;
2434 _offset += negate ? -off : loff;
2435 return true;
2436 }
2437 return false;
2438 }
2439 if (_invar != NULL) return false; // already have an invariant
2440 if (opc == Op_AddI) {
2441 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2442 _negate_invar = negate;
2443 _invar = n->in(1);
2444 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2445 return true;
2446 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2447 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2448 _negate_invar = negate;
2449 _invar = n->in(2);
2450 return true;
2451 }
2452 }
2453 if (opc == Op_SubI) {
2454 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2455 _negate_invar = negate;
2456 _invar = n->in(1);
2457 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2458 return true;
2459 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2460 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2461 _negate_invar = !negate;
2462 _invar = n->in(2);
2463 return true;
2464 }
2465 }
2466 if (invariant(n)) {
2467 _negate_invar = negate;
2468 _invar = n;
2469 return true;
2470 }
2471 return false;
2472 }
2474 //----------------------------print------------------------
2475 void SWPointer::print() {
2476 #ifndef PRODUCT
2477 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2478 _base != NULL ? _base->_idx : 0,
2479 _adr != NULL ? _adr->_idx : 0,
2480 _scale, _offset,
2481 _negate_invar?'-':'+',
2482 _invar != NULL ? _invar->_idx : 0);
2483 #endif
2484 }
2486 // ========================= OrderedPair =====================
2488 const OrderedPair OrderedPair::initial;
2490 // ========================= SWNodeInfo =====================
2492 const SWNodeInfo SWNodeInfo::initial;
2495 // ============================ DepGraph ===========================
2497 //------------------------------make_node---------------------------
2498 // Make a new dependence graph node for an ideal node.
2499 DepMem* DepGraph::make_node(Node* node) {
2500 DepMem* m = new (_arena) DepMem(node);
2501 if (node != NULL) {
2502 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2503 _map.at_put_grow(node->_idx, m);
2504 }
2505 return m;
2506 }
2508 //------------------------------make_edge---------------------------
2509 // Make a new dependence graph edge from dpred -> dsucc
2510 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2511 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2512 dpred->set_out_head(e);
2513 dsucc->set_in_head(e);
2514 return e;
2515 }
2517 // ========================== DepMem ========================
2519 //------------------------------in_cnt---------------------------
2520 int DepMem::in_cnt() {
2521 int ct = 0;
2522 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2523 return ct;
2524 }
2526 //------------------------------out_cnt---------------------------
2527 int DepMem::out_cnt() {
2528 int ct = 0;
2529 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2530 return ct;
2531 }
2533 //------------------------------print-----------------------------
2534 void DepMem::print() {
2535 #ifndef PRODUCT
2536 tty->print(" DepNode %d (", _node->_idx);
2537 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2538 Node* pred = p->pred()->node();
2539 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2540 }
2541 tty->print(") [");
2542 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2543 Node* succ = s->succ()->node();
2544 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2545 }
2546 tty->print_cr(" ]");
2547 #endif
2548 }
2550 // =========================== DepEdge =========================
2552 //------------------------------DepPreds---------------------------
2553 void DepEdge::print() {
2554 #ifndef PRODUCT
2555 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2556 #endif
2557 }
2559 // =========================== DepPreds =========================
2560 // Iterator over predecessor edges in the dependence graph.
2562 //------------------------------DepPreds---------------------------
2563 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2564 _n = n;
2565 _done = false;
2566 if (_n->is_Store() || _n->is_Load()) {
2567 _next_idx = MemNode::Address;
2568 _end_idx = n->req();
2569 _dep_next = dg.dep(_n)->in_head();
2570 } else if (_n->is_Mem()) {
2571 _next_idx = 0;
2572 _end_idx = 0;
2573 _dep_next = dg.dep(_n)->in_head();
2574 } else {
2575 _next_idx = 1;
2576 _end_idx = _n->req();
2577 _dep_next = NULL;
2578 }
2579 next();
2580 }
2582 //------------------------------next---------------------------
2583 void DepPreds::next() {
2584 if (_dep_next != NULL) {
2585 _current = _dep_next->pred()->node();
2586 _dep_next = _dep_next->next_in();
2587 } else if (_next_idx < _end_idx) {
2588 _current = _n->in(_next_idx++);
2589 } else {
2590 _done = true;
2591 }
2592 }
2594 // =========================== DepSuccs =========================
2595 // Iterator over successor edges in the dependence graph.
2597 //------------------------------DepSuccs---------------------------
2598 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2599 _n = n;
2600 _done = false;
2601 if (_n->is_Load()) {
2602 _next_idx = 0;
2603 _end_idx = _n->outcnt();
2604 _dep_next = dg.dep(_n)->out_head();
2605 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2606 _next_idx = 0;
2607 _end_idx = 0;
2608 _dep_next = dg.dep(_n)->out_head();
2609 } else {
2610 _next_idx = 0;
2611 _end_idx = _n->outcnt();
2612 _dep_next = NULL;
2613 }
2614 next();
2615 }
2617 //-------------------------------next---------------------------
2618 void DepSuccs::next() {
2619 if (_dep_next != NULL) {
2620 _current = _dep_next->succ()->node();
2621 _dep_next = _dep_next->next_out();
2622 } else if (_next_idx < _end_idx) {
2623 _current = _n->raw_out(_next_idx++);
2624 } else {
2625 _done = true;
2626 }
2627 }