Tue, 23 Oct 2012 13:06:37 -0700
8001183: incorrect results of char vectors right shift operaiton
Summary: do vector right shift operation for small int types only after loads
Reviewed-by: jrose, dlong
1 /*
2 * Copyright (c) 2007, 2012, 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 construct_bb();
148 dependence_graph();
150 compute_max_depth();
152 compute_vector_element_type();
154 // Attempt vectorization
156 find_adjacent_refs();
158 extend_packlist();
160 combine_packs();
162 construct_my_pack_map();
164 filter_packs();
166 schedule();
168 output();
169 }
171 //------------------------------find_adjacent_refs---------------------------
172 // Find the adjacent memory references and create pack pairs for them.
173 // This is the initial set of packs that will then be extended by
174 // following use->def and def->use links. The align positions are
175 // assigned relative to the reference "align_to_ref"
176 void SuperWord::find_adjacent_refs() {
177 // Get list of memory operations
178 Node_List memops;
179 for (int i = 0; i < _block.length(); i++) {
180 Node* n = _block.at(i);
181 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
182 is_java_primitive(n->as_Mem()->memory_type())) {
183 int align = memory_alignment(n->as_Mem(), 0);
184 if (align != bottom_align) {
185 memops.push(n);
186 }
187 }
188 }
190 Node_List align_to_refs;
191 int best_iv_adjustment = 0;
192 MemNode* best_align_to_mem_ref = NULL;
194 while (memops.size() != 0) {
195 // Find a memory reference to align to.
196 MemNode* mem_ref = find_align_to_ref(memops);
197 if (mem_ref == NULL) break;
198 align_to_refs.push(mem_ref);
199 int iv_adjustment = get_iv_adjustment(mem_ref);
201 if (best_align_to_mem_ref == NULL) {
202 // Set memory reference which is the best from all memory operations
203 // to be used for alignment. The pre-loop trip count is modified to align
204 // this reference to a vector-aligned address.
205 best_align_to_mem_ref = mem_ref;
206 best_iv_adjustment = iv_adjustment;
207 }
209 SWPointer align_to_ref_p(mem_ref, this);
210 // Set alignment relative to "align_to_ref" for all related memory operations.
211 for (int i = memops.size() - 1; i >= 0; i--) {
212 MemNode* s = memops.at(i)->as_Mem();
213 if (isomorphic(s, mem_ref)) {
214 SWPointer p2(s, this);
215 if (p2.comparable(align_to_ref_p)) {
216 int align = memory_alignment(s, iv_adjustment);
217 set_alignment(s, align);
218 }
219 }
220 }
222 // Create initial pack pairs of memory operations for which
223 // alignment is set and vectors will be aligned.
224 bool create_pack = true;
225 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
226 if (!Matcher::misaligned_vectors_ok()) {
227 int vw = vector_width(mem_ref);
228 int vw_best = vector_width(best_align_to_mem_ref);
229 if (vw > vw_best) {
230 // Do not vectorize a memory access with more elements per vector
231 // if unaligned memory access is not allowed because number of
232 // iterations in pre-loop will be not enough to align it.
233 create_pack = false;
234 }
235 }
236 } else {
237 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
238 // Can't allow vectorization of unaligned memory accesses with the
239 // same type since it could be overlapped accesses to the same array.
240 create_pack = false;
241 } else {
242 // Allow independent (different type) unaligned memory operations
243 // if HW supports them.
244 if (!Matcher::misaligned_vectors_ok()) {
245 create_pack = false;
246 } else {
247 // Check if packs of the same memory type but
248 // with a different alignment were created before.
249 for (uint i = 0; i < align_to_refs.size(); i++) {
250 MemNode* mr = align_to_refs.at(i)->as_Mem();
251 if (same_velt_type(mr, mem_ref) &&
252 memory_alignment(mr, iv_adjustment) != 0)
253 create_pack = false;
254 }
255 }
256 }
257 }
258 if (create_pack) {
259 for (uint i = 0; i < memops.size(); i++) {
260 Node* s1 = memops.at(i);
261 int align = alignment(s1);
262 if (align == top_align) continue;
263 for (uint j = 0; j < memops.size(); j++) {
264 Node* s2 = memops.at(j);
265 if (alignment(s2) == top_align) continue;
266 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
267 if (stmts_can_pack(s1, s2, align)) {
268 Node_List* pair = new Node_List();
269 pair->push(s1);
270 pair->push(s2);
271 _packset.append(pair);
272 }
273 }
274 }
275 }
276 } else { // Don't create unaligned pack
277 // First, remove remaining memory ops of the same type from the list.
278 for (int i = memops.size() - 1; i >= 0; i--) {
279 MemNode* s = memops.at(i)->as_Mem();
280 if (same_velt_type(s, mem_ref)) {
281 memops.remove(i);
282 }
283 }
285 // Second, remove already constructed packs of the same type.
286 for (int i = _packset.length() - 1; i >= 0; i--) {
287 Node_List* p = _packset.at(i);
288 MemNode* s = p->at(0)->as_Mem();
289 if (same_velt_type(s, mem_ref)) {
290 remove_pack_at(i);
291 }
292 }
294 // If needed find the best memory reference for loop alignment again.
295 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
296 // Put memory ops from remaining packs back on memops list for
297 // the best alignment search.
298 uint orig_msize = memops.size();
299 for (int i = 0; i < _packset.length(); i++) {
300 Node_List* p = _packset.at(i);
301 MemNode* s = p->at(0)->as_Mem();
302 assert(!same_velt_type(s, mem_ref), "sanity");
303 memops.push(s);
304 }
305 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
306 if (best_align_to_mem_ref == NULL) break;
307 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
308 // Restore list.
309 while (memops.size() > orig_msize)
310 (void)memops.pop();
311 }
312 } // unaligned memory accesses
314 // Remove used mem nodes.
315 for (int i = memops.size() - 1; i >= 0; i--) {
316 MemNode* m = memops.at(i)->as_Mem();
317 if (alignment(m) != top_align) {
318 memops.remove(i);
319 }
320 }
322 } // while (memops.size() != 0
323 set_align_to_ref(best_align_to_mem_ref);
325 #ifndef PRODUCT
326 if (TraceSuperWord) {
327 tty->print_cr("\nAfter find_adjacent_refs");
328 print_packset();
329 }
330 #endif
331 }
333 //------------------------------find_align_to_ref---------------------------
334 // Find a memory reference to align the loop induction variable to.
335 // Looks first at stores then at loads, looking for a memory reference
336 // with the largest number of references similar to it.
337 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
338 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
340 // Count number of comparable memory ops
341 for (uint i = 0; i < memops.size(); i++) {
342 MemNode* s1 = memops.at(i)->as_Mem();
343 SWPointer p1(s1, this);
344 // Discard if pre loop can't align this reference
345 if (!ref_is_alignable(p1)) {
346 *cmp_ct.adr_at(i) = 0;
347 continue;
348 }
349 for (uint j = i+1; j < memops.size(); j++) {
350 MemNode* s2 = memops.at(j)->as_Mem();
351 if (isomorphic(s1, s2)) {
352 SWPointer p2(s2, this);
353 if (p1.comparable(p2)) {
354 (*cmp_ct.adr_at(i))++;
355 (*cmp_ct.adr_at(j))++;
356 }
357 }
358 }
359 }
361 // Find Store (or Load) with the greatest number of "comparable" references,
362 // biggest vector size, smallest data size and smallest iv offset.
363 int max_ct = 0;
364 int max_vw = 0;
365 int max_idx = -1;
366 int min_size = max_jint;
367 int min_iv_offset = max_jint;
368 for (uint j = 0; j < memops.size(); j++) {
369 MemNode* s = memops.at(j)->as_Mem();
370 if (s->is_Store()) {
371 int vw = vector_width_in_bytes(s);
372 assert(vw > 1, "sanity");
373 SWPointer p(s, this);
374 if (cmp_ct.at(j) > max_ct ||
375 cmp_ct.at(j) == max_ct &&
376 (vw > max_vw ||
377 vw == max_vw &&
378 (data_size(s) < min_size ||
379 data_size(s) == min_size &&
380 (p.offset_in_bytes() < min_iv_offset)))) {
381 max_ct = cmp_ct.at(j);
382 max_vw = vw;
383 max_idx = j;
384 min_size = data_size(s);
385 min_iv_offset = p.offset_in_bytes();
386 }
387 }
388 }
389 // If no stores, look at loads
390 if (max_ct == 0) {
391 for (uint j = 0; j < memops.size(); j++) {
392 MemNode* s = memops.at(j)->as_Mem();
393 if (s->is_Load()) {
394 int vw = vector_width_in_bytes(s);
395 assert(vw > 1, "sanity");
396 SWPointer p(s, this);
397 if (cmp_ct.at(j) > max_ct ||
398 cmp_ct.at(j) == max_ct &&
399 (vw > max_vw ||
400 vw == max_vw &&
401 (data_size(s) < min_size ||
402 data_size(s) == min_size &&
403 (p.offset_in_bytes() < min_iv_offset)))) {
404 max_ct = cmp_ct.at(j);
405 max_vw = vw;
406 max_idx = j;
407 min_size = data_size(s);
408 min_iv_offset = p.offset_in_bytes();
409 }
410 }
411 }
412 }
414 #ifdef ASSERT
415 if (TraceSuperWord && Verbose) {
416 tty->print_cr("\nVector memops after find_align_to_refs");
417 for (uint i = 0; i < memops.size(); i++) {
418 MemNode* s = memops.at(i)->as_Mem();
419 s->dump();
420 }
421 }
422 #endif
424 if (max_ct > 0) {
425 #ifdef ASSERT
426 if (TraceSuperWord) {
427 tty->print("\nVector align to node: ");
428 memops.at(max_idx)->as_Mem()->dump();
429 }
430 #endif
431 return memops.at(max_idx)->as_Mem();
432 }
433 return NULL;
434 }
436 //------------------------------ref_is_alignable---------------------------
437 // Can the preloop align the reference to position zero in the vector?
438 bool SuperWord::ref_is_alignable(SWPointer& p) {
439 if (!p.has_iv()) {
440 return true; // no induction variable
441 }
442 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
443 assert(pre_end->stride_is_con(), "pre loop stride is constant");
444 int preloop_stride = pre_end->stride_con();
446 int span = preloop_stride * p.scale_in_bytes();
448 // Stride one accesses are alignable.
449 if (ABS(span) == p.memory_size())
450 return true;
452 // If initial offset from start of object is computable,
453 // compute alignment within the vector.
454 int vw = vector_width_in_bytes(p.mem());
455 assert(vw > 1, "sanity");
456 if (vw % span == 0) {
457 Node* init_nd = pre_end->init_trip();
458 if (init_nd->is_Con() && p.invar() == NULL) {
459 int init = init_nd->bottom_type()->is_int()->get_con();
461 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
462 assert(init_offset >= 0, "positive offset from object start");
464 if (span > 0) {
465 return (vw - (init_offset % vw)) % span == 0;
466 } else {
467 assert(span < 0, "nonzero stride * scale");
468 return (init_offset % vw) % -span == 0;
469 }
470 }
471 }
472 return false;
473 }
475 //---------------------------get_iv_adjustment---------------------------
476 // Calculate loop's iv adjustment for this memory ops.
477 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
478 SWPointer align_to_ref_p(mem_ref, this);
479 int offset = align_to_ref_p.offset_in_bytes();
480 int scale = align_to_ref_p.scale_in_bytes();
481 int vw = vector_width_in_bytes(mem_ref);
482 assert(vw > 1, "sanity");
483 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
484 // At least one iteration is executed in pre-loop by default. As result
485 // several iterations are needed to align memory operations in main-loop even
486 // if offset is 0.
487 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
488 int elt_size = align_to_ref_p.memory_size();
489 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
490 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
491 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
493 #ifndef PRODUCT
494 if (TraceSuperWord)
495 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
496 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
497 #endif
498 return iv_adjustment;
499 }
501 //---------------------------dependence_graph---------------------------
502 // Construct dependency graph.
503 // Add dependence edges to load/store nodes for memory dependence
504 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
505 void SuperWord::dependence_graph() {
506 // First, assign a dependence node to each memory node
507 for (int i = 0; i < _block.length(); i++ ) {
508 Node *n = _block.at(i);
509 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
510 _dg.make_node(n);
511 }
512 }
514 // For each memory slice, create the dependences
515 for (int i = 0; i < _mem_slice_head.length(); i++) {
516 Node* n = _mem_slice_head.at(i);
517 Node* n_tail = _mem_slice_tail.at(i);
519 // Get slice in predecessor order (last is first)
520 mem_slice_preds(n_tail, n, _nlist);
522 // Make the slice dependent on the root
523 DepMem* slice = _dg.dep(n);
524 _dg.make_edge(_dg.root(), slice);
526 // Create a sink for the slice
527 DepMem* slice_sink = _dg.make_node(NULL);
528 _dg.make_edge(slice_sink, _dg.tail());
530 // Now visit each pair of memory ops, creating the edges
531 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
532 Node* s1 = _nlist.at(j);
534 // If no dependency yet, use slice
535 if (_dg.dep(s1)->in_cnt() == 0) {
536 _dg.make_edge(slice, s1);
537 }
538 SWPointer p1(s1->as_Mem(), this);
539 bool sink_dependent = true;
540 for (int k = j - 1; k >= 0; k--) {
541 Node* s2 = _nlist.at(k);
542 if (s1->is_Load() && s2->is_Load())
543 continue;
544 SWPointer p2(s2->as_Mem(), this);
546 int cmp = p1.cmp(p2);
547 if (SuperWordRTDepCheck &&
548 p1.base() != p2.base() && p1.valid() && p2.valid()) {
549 // Create a runtime check to disambiguate
550 OrderedPair pp(p1.base(), p2.base());
551 _disjoint_ptrs.append_if_missing(pp);
552 } else if (!SWPointer::not_equal(cmp)) {
553 // Possibly same address
554 _dg.make_edge(s1, s2);
555 sink_dependent = false;
556 }
557 }
558 if (sink_dependent) {
559 _dg.make_edge(s1, slice_sink);
560 }
561 }
562 #ifndef PRODUCT
563 if (TraceSuperWord) {
564 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
565 for (int q = 0; q < _nlist.length(); q++) {
566 _dg.print(_nlist.at(q));
567 }
568 tty->cr();
569 }
570 #endif
571 _nlist.clear();
572 }
574 #ifndef PRODUCT
575 if (TraceSuperWord) {
576 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
577 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
578 _disjoint_ptrs.at(r).print();
579 tty->cr();
580 }
581 tty->cr();
582 }
583 #endif
584 }
586 //---------------------------mem_slice_preds---------------------------
587 // Return a memory slice (node list) in predecessor order starting at "start"
588 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
589 assert(preds.length() == 0, "start empty");
590 Node* n = start;
591 Node* prev = NULL;
592 while (true) {
593 assert(in_bb(n), "must be in block");
594 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
595 Node* out = n->fast_out(i);
596 if (out->is_Load()) {
597 if (in_bb(out)) {
598 preds.push(out);
599 }
600 } else {
601 // FIXME
602 if (out->is_MergeMem() && !in_bb(out)) {
603 // Either unrolling is causing a memory edge not to disappear,
604 // or need to run igvn.optimize() again before SLP
605 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
606 // Ditto. Not sure what else to check further.
607 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
608 // StoreCM has an input edge used as a precedence edge.
609 // Maybe an issue when oop stores are vectorized.
610 } else {
611 assert(out == prev || prev == NULL, "no branches off of store slice");
612 }
613 }
614 }
615 if (n == stop) break;
616 preds.push(n);
617 prev = n;
618 n = n->in(MemNode::Memory);
619 }
620 }
622 //------------------------------stmts_can_pack---------------------------
623 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
624 // s1 aligned at "align"
625 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
627 // Do not use superword for non-primitives
628 BasicType bt1 = velt_basic_type(s1);
629 BasicType bt2 = velt_basic_type(s2);
630 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
631 return false;
632 if (Matcher::max_vector_size(bt1) < 2) {
633 return false; // No vectors for this type
634 }
636 if (isomorphic(s1, s2)) {
637 if (independent(s1, s2)) {
638 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
639 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
640 int s1_align = alignment(s1);
641 int s2_align = alignment(s2);
642 if (s1_align == top_align || s1_align == align) {
643 if (s2_align == top_align || s2_align == align + data_size(s1)) {
644 return true;
645 }
646 }
647 }
648 }
649 }
650 }
651 return false;
652 }
654 //------------------------------exists_at---------------------------
655 // Does s exist in a pack at position pos?
656 bool SuperWord::exists_at(Node* s, uint pos) {
657 for (int i = 0; i < _packset.length(); i++) {
658 Node_List* p = _packset.at(i);
659 if (p->at(pos) == s) {
660 return true;
661 }
662 }
663 return false;
664 }
666 //------------------------------are_adjacent_refs---------------------------
667 // Is s1 immediately before s2 in memory?
668 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
669 if (!s1->is_Mem() || !s2->is_Mem()) return false;
670 if (!in_bb(s1) || !in_bb(s2)) return false;
672 // Do not use superword for non-primitives
673 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
674 !is_java_primitive(s2->as_Mem()->memory_type())) {
675 return false;
676 }
678 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
679 // only pack memops that are in the same alias set until that's fixed.
680 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
681 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
682 return false;
683 SWPointer p1(s1->as_Mem(), this);
684 SWPointer p2(s2->as_Mem(), this);
685 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
686 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
687 return diff == data_size(s1);
688 }
690 //------------------------------isomorphic---------------------------
691 // Are s1 and s2 similar?
692 bool SuperWord::isomorphic(Node* s1, Node* s2) {
693 if (s1->Opcode() != s2->Opcode()) return false;
694 if (s1->req() != s2->req()) return false;
695 if (s1->in(0) != s2->in(0)) return false;
696 if (!same_velt_type(s1, s2)) return false;
697 return true;
698 }
700 //------------------------------independent---------------------------
701 // Is there no data path from s1 to s2 or s2 to s1?
702 bool SuperWord::independent(Node* s1, Node* s2) {
703 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
704 int d1 = depth(s1);
705 int d2 = depth(s2);
706 if (d1 == d2) return s1 != s2;
707 Node* deep = d1 > d2 ? s1 : s2;
708 Node* shallow = d1 > d2 ? s2 : s1;
710 visited_clear();
712 return independent_path(shallow, deep);
713 }
715 //------------------------------independent_path------------------------------
716 // Helper for independent
717 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
718 if (dp >= 1000) return false; // stop deep recursion
719 visited_set(deep);
720 int shal_depth = depth(shallow);
721 assert(shal_depth <= depth(deep), "must be");
722 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
723 Node* pred = preds.current();
724 if (in_bb(pred) && !visited_test(pred)) {
725 if (shallow == pred) {
726 return false;
727 }
728 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
729 return false;
730 }
731 }
732 }
733 return true;
734 }
736 //------------------------------set_alignment---------------------------
737 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
738 set_alignment(s1, align);
739 if (align == top_align || align == bottom_align) {
740 set_alignment(s2, align);
741 } else {
742 set_alignment(s2, align + data_size(s1));
743 }
744 }
746 //------------------------------data_size---------------------------
747 int SuperWord::data_size(Node* s) {
748 int bsize = type2aelembytes(velt_basic_type(s));
749 assert(bsize != 0, "valid size");
750 return bsize;
751 }
753 //------------------------------extend_packlist---------------------------
754 // Extend packset by following use->def and def->use links from pack members.
755 void SuperWord::extend_packlist() {
756 bool changed;
757 do {
758 changed = false;
759 for (int i = 0; i < _packset.length(); i++) {
760 Node_List* p = _packset.at(i);
761 changed |= follow_use_defs(p);
762 changed |= follow_def_uses(p);
763 }
764 } while (changed);
766 #ifndef PRODUCT
767 if (TraceSuperWord) {
768 tty->print_cr("\nAfter extend_packlist");
769 print_packset();
770 }
771 #endif
772 }
774 //------------------------------follow_use_defs---------------------------
775 // Extend the packset by visiting operand definitions of nodes in pack p
776 bool SuperWord::follow_use_defs(Node_List* p) {
777 assert(p->size() == 2, "just checking");
778 Node* s1 = p->at(0);
779 Node* s2 = p->at(1);
780 assert(s1->req() == s2->req(), "just checking");
781 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
783 if (s1->is_Load()) return false;
785 int align = alignment(s1);
786 bool changed = false;
787 int start = s1->is_Store() ? MemNode::ValueIn : 1;
788 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
789 for (int j = start; j < end; j++) {
790 Node* t1 = s1->in(j);
791 Node* t2 = s2->in(j);
792 if (!in_bb(t1) || !in_bb(t2))
793 continue;
794 if (stmts_can_pack(t1, t2, align)) {
795 if (est_savings(t1, t2) >= 0) {
796 Node_List* pair = new Node_List();
797 pair->push(t1);
798 pair->push(t2);
799 _packset.append(pair);
800 set_alignment(t1, t2, align);
801 changed = true;
802 }
803 }
804 }
805 return changed;
806 }
808 //------------------------------follow_def_uses---------------------------
809 // Extend the packset by visiting uses of nodes in pack p
810 bool SuperWord::follow_def_uses(Node_List* p) {
811 bool changed = false;
812 Node* s1 = p->at(0);
813 Node* s2 = p->at(1);
814 assert(p->size() == 2, "just checking");
815 assert(s1->req() == s2->req(), "just checking");
816 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
818 if (s1->is_Store()) return false;
820 int align = alignment(s1);
821 int savings = -1;
822 Node* u1 = NULL;
823 Node* u2 = NULL;
824 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
825 Node* t1 = s1->fast_out(i);
826 if (!in_bb(t1)) continue;
827 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
828 Node* t2 = s2->fast_out(j);
829 if (!in_bb(t2)) continue;
830 if (!opnd_positions_match(s1, t1, s2, t2))
831 continue;
832 if (stmts_can_pack(t1, t2, align)) {
833 int my_savings = est_savings(t1, t2);
834 if (my_savings > savings) {
835 savings = my_savings;
836 u1 = t1;
837 u2 = t2;
838 }
839 }
840 }
841 }
842 if (savings >= 0) {
843 Node_List* pair = new Node_List();
844 pair->push(u1);
845 pair->push(u2);
846 _packset.append(pair);
847 set_alignment(u1, u2, align);
848 changed = true;
849 }
850 return changed;
851 }
853 //---------------------------opnd_positions_match-------------------------
854 // Is the use of d1 in u1 at the same operand position as d2 in u2?
855 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
856 uint ct = u1->req();
857 if (ct != u2->req()) return false;
858 uint i1 = 0;
859 uint i2 = 0;
860 do {
861 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
862 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
863 if (i1 != i2) {
864 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
865 // Further analysis relies on operands position matching.
866 u2->swap_edges(i1, i2);
867 } else {
868 return false;
869 }
870 }
871 } while (i1 < ct);
872 return true;
873 }
875 //------------------------------est_savings---------------------------
876 // Estimate the savings from executing s1 and s2 as a pack
877 int SuperWord::est_savings(Node* s1, Node* s2) {
878 int save_in = 2 - 1; // 2 operations per instruction in packed form
880 // inputs
881 for (uint i = 1; i < s1->req(); i++) {
882 Node* x1 = s1->in(i);
883 Node* x2 = s2->in(i);
884 if (x1 != x2) {
885 if (are_adjacent_refs(x1, x2)) {
886 save_in += adjacent_profit(x1, x2);
887 } else if (!in_packset(x1, x2)) {
888 save_in -= pack_cost(2);
889 } else {
890 save_in += unpack_cost(2);
891 }
892 }
893 }
895 // uses of result
896 uint ct = 0;
897 int save_use = 0;
898 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
899 Node* s1_use = s1->fast_out(i);
900 for (int j = 0; j < _packset.length(); j++) {
901 Node_List* p = _packset.at(j);
902 if (p->at(0) == s1_use) {
903 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
904 Node* s2_use = s2->fast_out(k);
905 if (p->at(p->size()-1) == s2_use) {
906 ct++;
907 if (are_adjacent_refs(s1_use, s2_use)) {
908 save_use += adjacent_profit(s1_use, s2_use);
909 }
910 }
911 }
912 }
913 }
914 }
916 if (ct < s1->outcnt()) save_use += unpack_cost(1);
917 if (ct < s2->outcnt()) save_use += unpack_cost(1);
919 return MAX2(save_in, save_use);
920 }
922 //------------------------------costs---------------------------
923 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
924 int SuperWord::pack_cost(int ct) { return ct; }
925 int SuperWord::unpack_cost(int ct) { return ct; }
927 //------------------------------combine_packs---------------------------
928 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
929 void SuperWord::combine_packs() {
930 bool changed = true;
931 // Combine packs regardless max vector size.
932 while (changed) {
933 changed = false;
934 for (int i = 0; i < _packset.length(); i++) {
935 Node_List* p1 = _packset.at(i);
936 if (p1 == NULL) continue;
937 for (int j = 0; j < _packset.length(); j++) {
938 Node_List* p2 = _packset.at(j);
939 if (p2 == NULL) continue;
940 if (i == j) continue;
941 if (p1->at(p1->size()-1) == p2->at(0)) {
942 for (uint k = 1; k < p2->size(); k++) {
943 p1->push(p2->at(k));
944 }
945 _packset.at_put(j, NULL);
946 changed = true;
947 }
948 }
949 }
950 }
952 // Split packs which have size greater then max vector size.
953 for (int i = 0; i < _packset.length(); i++) {
954 Node_List* p1 = _packset.at(i);
955 if (p1 != NULL) {
956 BasicType bt = velt_basic_type(p1->at(0));
957 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
958 assert(is_power_of_2(max_vlen), "sanity");
959 uint psize = p1->size();
960 if (!is_power_of_2(psize)) {
961 // Skip pack which can't be vector.
962 // case1: for(...) { a[i] = i; } elements values are different (i+x)
963 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
964 _packset.at_put(i, NULL);
965 continue;
966 }
967 if (psize > max_vlen) {
968 Node_List* pack = new Node_List();
969 for (uint j = 0; j < psize; j++) {
970 pack->push(p1->at(j));
971 if (pack->size() >= max_vlen) {
972 assert(is_power_of_2(pack->size()), "sanity");
973 _packset.append(pack);
974 pack = new Node_List();
975 }
976 }
977 _packset.at_put(i, NULL);
978 }
979 }
980 }
982 // Compress list.
983 for (int i = _packset.length() - 1; i >= 0; i--) {
984 Node_List* p1 = _packset.at(i);
985 if (p1 == NULL) {
986 _packset.remove_at(i);
987 }
988 }
990 #ifndef PRODUCT
991 if (TraceSuperWord) {
992 tty->print_cr("\nAfter combine_packs");
993 print_packset();
994 }
995 #endif
996 }
998 //-----------------------------construct_my_pack_map--------------------------
999 // Construct the map from nodes to packs. Only valid after the
1000 // point where a node is only in one pack (after combine_packs).
1001 void SuperWord::construct_my_pack_map() {
1002 Node_List* rslt = NULL;
1003 for (int i = 0; i < _packset.length(); i++) {
1004 Node_List* p = _packset.at(i);
1005 for (uint j = 0; j < p->size(); j++) {
1006 Node* s = p->at(j);
1007 assert(my_pack(s) == NULL, "only in one pack");
1008 set_my_pack(s, p);
1009 }
1010 }
1011 }
1013 //------------------------------filter_packs---------------------------
1014 // Remove packs that are not implemented or not profitable.
1015 void SuperWord::filter_packs() {
1017 // Remove packs that are not implemented
1018 for (int i = _packset.length() - 1; i >= 0; i--) {
1019 Node_List* pk = _packset.at(i);
1020 bool impl = implemented(pk);
1021 if (!impl) {
1022 #ifndef PRODUCT
1023 if (TraceSuperWord && Verbose) {
1024 tty->print_cr("Unimplemented");
1025 pk->at(0)->dump();
1026 }
1027 #endif
1028 remove_pack_at(i);
1029 }
1030 }
1032 // Remove packs that are not profitable
1033 bool changed;
1034 do {
1035 changed = false;
1036 for (int i = _packset.length() - 1; i >= 0; i--) {
1037 Node_List* pk = _packset.at(i);
1038 bool prof = profitable(pk);
1039 if (!prof) {
1040 #ifndef PRODUCT
1041 if (TraceSuperWord && Verbose) {
1042 tty->print_cr("Unprofitable");
1043 pk->at(0)->dump();
1044 }
1045 #endif
1046 remove_pack_at(i);
1047 changed = true;
1048 }
1049 }
1050 } while (changed);
1052 #ifndef PRODUCT
1053 if (TraceSuperWord) {
1054 tty->print_cr("\nAfter filter_packs");
1055 print_packset();
1056 tty->cr();
1057 }
1058 #endif
1059 }
1061 //------------------------------implemented---------------------------
1062 // Can code be generated for pack p?
1063 bool SuperWord::implemented(Node_List* p) {
1064 Node* p0 = p->at(0);
1065 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1066 }
1068 //------------------------------same_inputs--------------------------
1069 // For pack p, are all idx operands the same?
1070 static bool same_inputs(Node_List* p, int idx) {
1071 Node* p0 = p->at(0);
1072 uint vlen = p->size();
1073 Node* p0_def = p0->in(idx);
1074 for (uint i = 1; i < vlen; i++) {
1075 Node* pi = p->at(i);
1076 Node* pi_def = pi->in(idx);
1077 if (p0_def != pi_def)
1078 return false;
1079 }
1080 return true;
1081 }
1083 //------------------------------profitable---------------------------
1084 // For pack p, are all operands and all uses (with in the block) vector?
1085 bool SuperWord::profitable(Node_List* p) {
1086 Node* p0 = p->at(0);
1087 uint start, end;
1088 VectorNode::vector_operands(p0, &start, &end);
1090 // Return false if some inputs are not vectors or vectors with different
1091 // size or alignment.
1092 // Also, for now, return false if not scalar promotion case when inputs are
1093 // the same. Later, implement PackNode and allow differing, non-vector inputs
1094 // (maybe just the ones from outside the block.)
1095 for (uint i = start; i < end; i++) {
1096 if (!is_vector_use(p0, i))
1097 return false;
1098 }
1099 if (VectorNode::is_shift(p0)) {
1100 // For now, return false if shift count is vector or not scalar promotion
1101 // case (different shift counts) because it is not supported yet.
1102 Node* cnt = p0->in(2);
1103 Node_List* cnt_pk = my_pack(cnt);
1104 if (cnt_pk != NULL)
1105 return false;
1106 if (!same_inputs(p, 2))
1107 return false;
1108 }
1109 if (!p0->is_Store()) {
1110 // For now, return false if not all uses are vector.
1111 // Later, implement ExtractNode and allow non-vector uses (maybe
1112 // just the ones outside the block.)
1113 for (uint i = 0; i < p->size(); i++) {
1114 Node* def = p->at(i);
1115 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1116 Node* use = def->fast_out(j);
1117 for (uint k = 0; k < use->req(); k++) {
1118 Node* n = use->in(k);
1119 if (def == n) {
1120 if (!is_vector_use(use, k)) {
1121 return false;
1122 }
1123 }
1124 }
1125 }
1126 }
1127 }
1128 return true;
1129 }
1131 //------------------------------schedule---------------------------
1132 // Adjust the memory graph for the packed operations
1133 void SuperWord::schedule() {
1135 // Co-locate in the memory graph the members of each memory pack
1136 for (int i = 0; i < _packset.length(); i++) {
1137 co_locate_pack(_packset.at(i));
1138 }
1139 }
1141 //-------------------------------remove_and_insert-------------------
1142 // Remove "current" from its current position in the memory graph and insert
1143 // it after the appropriate insertion point (lip or uip).
1144 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1145 Node *uip, Unique_Node_List &sched_before) {
1146 Node* my_mem = current->in(MemNode::Memory);
1147 bool sched_up = sched_before.member(current);
1149 // remove current_store from its current position in the memmory graph
1150 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1151 Node* use = current->out(i);
1152 if (use->is_Mem()) {
1153 assert(use->in(MemNode::Memory) == current, "must be");
1154 if (use == prev) { // connect prev to my_mem
1155 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1156 --i; //deleted this edge; rescan position
1157 } else if (sched_before.member(use)) {
1158 if (!sched_up) { // Will be moved together with current
1159 _igvn.replace_input_of(use, MemNode::Memory, uip);
1160 --i; //deleted this edge; rescan position
1161 }
1162 } else {
1163 if (sched_up) { // Will be moved together with current
1164 _igvn.replace_input_of(use, MemNode::Memory, lip);
1165 --i; //deleted this edge; rescan position
1166 }
1167 }
1168 }
1169 }
1171 Node *insert_pt = sched_up ? uip : lip;
1173 // all uses of insert_pt's memory state should use current's instead
1174 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1175 Node* use = insert_pt->out(i);
1176 if (use->is_Mem()) {
1177 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1178 _igvn.replace_input_of(use, MemNode::Memory, current);
1179 --i; //deleted this edge; rescan position
1180 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1181 uint pos; //lip (lower insert point) must be the last one in the memory slice
1182 for (pos=1; pos < use->req(); pos++) {
1183 if (use->in(pos) == insert_pt) break;
1184 }
1185 _igvn.replace_input_of(use, pos, current);
1186 --i;
1187 }
1188 }
1190 //connect current to insert_pt
1191 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1192 }
1194 //------------------------------co_locate_pack----------------------------------
1195 // To schedule a store pack, we need to move any sandwiched memory ops either before
1196 // or after the pack, based upon dependence information:
1197 // (1) If any store in the pack depends on the sandwiched memory op, the
1198 // sandwiched memory op must be scheduled BEFORE the pack;
1199 // (2) If a sandwiched memory op depends on any store in the pack, the
1200 // sandwiched memory op must be scheduled AFTER the pack;
1201 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1202 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1203 // scheduled before the pack, memB must also be scheduled before the pack;
1204 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1205 // schedule this store AFTER the pack
1206 // (5) We know there is no dependence cycle, so there in no other case;
1207 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1208 //
1209 // To schedule a load pack, we use the memory state of either the first or the last load in
1210 // the pack, based on the dependence constraint.
1211 void SuperWord::co_locate_pack(Node_List* pk) {
1212 if (pk->at(0)->is_Store()) {
1213 MemNode* first = executed_first(pk)->as_Mem();
1214 MemNode* last = executed_last(pk)->as_Mem();
1215 Unique_Node_List schedule_before_pack;
1216 Unique_Node_List memops;
1218 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1219 MemNode* previous = last;
1220 while (true) {
1221 assert(in_bb(current), "stay in block");
1222 memops.push(previous);
1223 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1224 Node* use = current->out(i);
1225 if (use->is_Mem() && use != previous)
1226 memops.push(use);
1227 }
1228 if (current == first) break;
1229 previous = current;
1230 current = current->in(MemNode::Memory)->as_Mem();
1231 }
1233 // determine which memory operations should be scheduled before the pack
1234 for (uint i = 1; i < memops.size(); i++) {
1235 Node *s1 = memops.at(i);
1236 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1237 for (uint j = 0; j< i; j++) {
1238 Node *s2 = memops.at(j);
1239 if (!independent(s1, s2)) {
1240 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1241 schedule_before_pack.push(s1); // s1 must be scheduled before
1242 Node_List* mem_pk = my_pack(s1);
1243 if (mem_pk != NULL) {
1244 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1245 Node* s = mem_pk->at(ii); // follow partner
1246 if (memops.member(s) && !schedule_before_pack.member(s))
1247 schedule_before_pack.push(s);
1248 }
1249 }
1250 break;
1251 }
1252 }
1253 }
1254 }
1255 }
1257 Node* upper_insert_pt = first->in(MemNode::Memory);
1258 // Following code moves loads connected to upper_insert_pt below aliased stores.
1259 // Collect such loads here and reconnect them back to upper_insert_pt later.
1260 memops.clear();
1261 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1262 Node* use = upper_insert_pt->out(i);
1263 if (!use->is_Store())
1264 memops.push(use);
1265 }
1267 MemNode* lower_insert_pt = last;
1268 previous = last; //previous store in pk
1269 current = last->in(MemNode::Memory)->as_Mem();
1271 // start scheduling from "last" to "first"
1272 while (true) {
1273 assert(in_bb(current), "stay in block");
1274 assert(in_pack(previous, pk), "previous stays in pack");
1275 Node* my_mem = current->in(MemNode::Memory);
1277 if (in_pack(current, pk)) {
1278 // Forward users of my memory state (except "previous) to my input memory state
1279 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1280 Node* use = current->out(i);
1281 if (use->is_Mem() && use != previous) {
1282 assert(use->in(MemNode::Memory) == current, "must be");
1283 if (schedule_before_pack.member(use)) {
1284 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1285 } else {
1286 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1287 }
1288 --i; // deleted this edge; rescan position
1289 }
1290 }
1291 previous = current;
1292 } else { // !in_pack(current, pk) ==> a sandwiched store
1293 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1294 }
1296 if (current == first) break;
1297 current = my_mem->as_Mem();
1298 } // end while
1300 // Reconnect loads back to upper_insert_pt.
1301 for (uint i = 0; i < memops.size(); i++) {
1302 Node *ld = memops.at(i);
1303 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1304 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1305 }
1306 }
1307 } else if (pk->at(0)->is_Load()) { //load
1308 // all loads in the pack should have the same memory state. By default,
1309 // we use the memory state of the last load. However, if any load could
1310 // not be moved down due to the dependence constraint, we use the memory
1311 // state of the first load.
1312 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1313 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1314 bool schedule_last = true;
1315 for (uint i = 0; i < pk->size(); i++) {
1316 Node* ld = pk->at(i);
1317 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1318 current=current->in(MemNode::Memory)) {
1319 assert(current != first_mem, "corrupted memory graph");
1320 if(current->is_Mem() && !independent(current, ld)){
1321 schedule_last = false; // a later store depends on this load
1322 break;
1323 }
1324 }
1325 }
1327 Node* mem_input = schedule_last ? last_mem : first_mem;
1328 _igvn.hash_delete(mem_input);
1329 // Give each load the same memory state
1330 for (uint i = 0; i < pk->size(); i++) {
1331 LoadNode* ld = pk->at(i)->as_Load();
1332 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1333 }
1334 }
1335 }
1337 //------------------------------output---------------------------
1338 // Convert packs into vector node operations
1339 void SuperWord::output() {
1340 if (_packset.length() == 0) return;
1342 #ifndef PRODUCT
1343 if (TraceLoopOpts) {
1344 tty->print("SuperWord ");
1345 lpt()->dump_head();
1346 }
1347 #endif
1349 // MUST ENSURE main loop's initial value is properly aligned:
1350 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1352 align_initial_loop_index(align_to_ref());
1354 // Insert extract (unpack) operations for scalar uses
1355 for (int i = 0; i < _packset.length(); i++) {
1356 insert_extracts(_packset.at(i));
1357 }
1359 Compile* C = _phase->C;
1360 uint max_vlen_in_bytes = 0;
1361 for (int i = 0; i < _block.length(); i++) {
1362 Node* n = _block.at(i);
1363 Node_List* p = my_pack(n);
1364 if (p && n == executed_last(p)) {
1365 uint vlen = p->size();
1366 uint vlen_in_bytes = 0;
1367 Node* vn = NULL;
1368 Node* low_adr = p->at(0);
1369 Node* first = executed_first(p);
1370 int opc = n->Opcode();
1371 if (n->is_Load()) {
1372 Node* ctl = n->in(MemNode::Control);
1373 Node* mem = first->in(MemNode::Memory);
1374 Node* adr = low_adr->in(MemNode::Address);
1375 const TypePtr* atyp = n->adr_type();
1376 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1377 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1378 } else if (n->is_Store()) {
1379 // Promote value to be stored to vector
1380 Node* val = vector_opd(p, MemNode::ValueIn);
1381 Node* ctl = n->in(MemNode::Control);
1382 Node* mem = first->in(MemNode::Memory);
1383 Node* adr = low_adr->in(MemNode::Address);
1384 const TypePtr* atyp = n->adr_type();
1385 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1386 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1387 } else if (n->req() == 3) {
1388 // Promote operands to vector
1389 Node* in1 = vector_opd(p, 1);
1390 Node* in2 = vector_opd(p, 2);
1391 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1392 // Move invariant vector input into second position to avoid register spilling.
1393 Node* tmp = in1;
1394 in1 = in2;
1395 in2 = tmp;
1396 }
1397 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1398 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1399 } else {
1400 ShouldNotReachHere();
1401 }
1402 assert(vn != NULL, "sanity");
1403 _igvn.register_new_node_with_optimizer(vn);
1404 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1405 for (uint j = 0; j < p->size(); j++) {
1406 Node* pm = p->at(j);
1407 _igvn.replace_node(pm, vn);
1408 }
1409 _igvn._worklist.push(vn);
1411 if (vlen_in_bytes > max_vlen_in_bytes) {
1412 max_vlen_in_bytes = vlen_in_bytes;
1413 }
1414 #ifdef ASSERT
1415 if (TraceNewVectors) {
1416 tty->print("new Vector node: ");
1417 vn->dump();
1418 }
1419 #endif
1420 }
1421 }
1422 C->set_max_vector_size(max_vlen_in_bytes);
1423 }
1425 //------------------------------vector_opd---------------------------
1426 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1427 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1428 Node* p0 = p->at(0);
1429 uint vlen = p->size();
1430 Node* opd = p0->in(opd_idx);
1432 if (same_inputs(p, opd_idx)) {
1433 if (opd->is_Vector() || opd->is_LoadVector()) {
1434 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1435 return opd; // input is matching vector
1436 }
1437 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1438 Compile* C = _phase->C;
1439 Node* cnt = opd;
1440 // Vector instructions do not mask shift count, do it here.
1441 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1442 const TypeInt* t = opd->find_int_type();
1443 if (t != NULL && t->is_con()) {
1444 juint shift = t->get_con();
1445 if (shift > mask) { // Unsigned cmp
1446 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1447 }
1448 } else {
1449 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1450 cnt = ConNode::make(C, TypeInt::make(mask));
1451 _igvn.register_new_node_with_optimizer(cnt);
1452 cnt = new (C) AndINode(opd, cnt);
1453 _igvn.register_new_node_with_optimizer(cnt);
1454 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1455 }
1456 assert(opd->bottom_type()->isa_int(), "int type only");
1457 // Move non constant shift count into vector register.
1458 cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1459 }
1460 if (cnt != opd) {
1461 _igvn.register_new_node_with_optimizer(cnt);
1462 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1463 }
1464 return cnt;
1465 }
1466 assert(!opd->is_StoreVector(), "such vector is not expected here");
1467 // Convert scalar input to vector with the same number of elements as
1468 // p0's vector. Use p0's type because size of operand's container in
1469 // vector should match p0's size regardless operand's size.
1470 const Type* p0_t = velt_type(p0);
1471 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1473 _igvn.register_new_node_with_optimizer(vn);
1474 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1475 #ifdef ASSERT
1476 if (TraceNewVectors) {
1477 tty->print("new Vector node: ");
1478 vn->dump();
1479 }
1480 #endif
1481 return vn;
1482 }
1484 // Insert pack operation
1485 BasicType bt = velt_basic_type(p0);
1486 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1487 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1489 for (uint i = 1; i < vlen; i++) {
1490 Node* pi = p->at(i);
1491 Node* in = pi->in(opd_idx);
1492 assert(my_pack(in) == NULL, "Should already have been unpacked");
1493 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1494 pk->add_opd(in);
1495 }
1496 _igvn.register_new_node_with_optimizer(pk);
1497 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1498 #ifdef ASSERT
1499 if (TraceNewVectors) {
1500 tty->print("new Vector node: ");
1501 pk->dump();
1502 }
1503 #endif
1504 return pk;
1505 }
1507 //------------------------------insert_extracts---------------------------
1508 // If a use of pack p is not a vector use, then replace the
1509 // use with an extract operation.
1510 void SuperWord::insert_extracts(Node_List* p) {
1511 if (p->at(0)->is_Store()) return;
1512 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1514 // Inspect each use of each pack member. For each use that is
1515 // not a vector use, replace the use with an extract operation.
1517 for (uint i = 0; i < p->size(); i++) {
1518 Node* def = p->at(i);
1519 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1520 Node* use = def->fast_out(j);
1521 for (uint k = 0; k < use->req(); k++) {
1522 Node* n = use->in(k);
1523 if (def == n) {
1524 if (!is_vector_use(use, k)) {
1525 _n_idx_list.push(use, k);
1526 }
1527 }
1528 }
1529 }
1530 }
1532 while (_n_idx_list.is_nonempty()) {
1533 Node* use = _n_idx_list.node();
1534 int idx = _n_idx_list.index();
1535 _n_idx_list.pop();
1536 Node* def = use->in(idx);
1538 // Insert extract operation
1539 _igvn.hash_delete(def);
1540 int def_pos = alignment(def) / data_size(def);
1542 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1543 _igvn.register_new_node_with_optimizer(ex);
1544 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1545 _igvn.replace_input_of(use, idx, ex);
1546 _igvn._worklist.push(def);
1548 bb_insert_after(ex, bb_idx(def));
1549 set_velt_type(ex, velt_type(def));
1550 }
1551 }
1553 //------------------------------is_vector_use---------------------------
1554 // Is use->in(u_idx) a vector use?
1555 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1556 Node_List* u_pk = my_pack(use);
1557 if (u_pk == NULL) return false;
1558 Node* def = use->in(u_idx);
1559 Node_List* d_pk = my_pack(def);
1560 if (d_pk == NULL) {
1561 // check for scalar promotion
1562 Node* n = u_pk->at(0)->in(u_idx);
1563 for (uint i = 1; i < u_pk->size(); i++) {
1564 if (u_pk->at(i)->in(u_idx) != n) return false;
1565 }
1566 return true;
1567 }
1568 if (u_pk->size() != d_pk->size())
1569 return false;
1570 for (uint i = 0; i < u_pk->size(); i++) {
1571 Node* ui = u_pk->at(i);
1572 Node* di = d_pk->at(i);
1573 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1574 return false;
1575 }
1576 return true;
1577 }
1579 //------------------------------construct_bb---------------------------
1580 // Construct reverse postorder list of block members
1581 void SuperWord::construct_bb() {
1582 Node* entry = bb();
1584 assert(_stk.length() == 0, "stk is empty");
1585 assert(_block.length() == 0, "block is empty");
1586 assert(_data_entry.length() == 0, "data_entry is empty");
1587 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1588 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1590 // Find non-control nodes with no inputs from within block,
1591 // create a temporary map from node _idx to bb_idx for use
1592 // by the visited and post_visited sets,
1593 // and count number of nodes in block.
1594 int bb_ct = 0;
1595 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1596 Node *n = lpt()->_body.at(i);
1597 set_bb_idx(n, i); // Create a temporary map
1598 if (in_bb(n)) {
1599 bb_ct++;
1600 if (!n->is_CFG()) {
1601 bool found = false;
1602 for (uint j = 0; j < n->req(); j++) {
1603 Node* def = n->in(j);
1604 if (def && in_bb(def)) {
1605 found = true;
1606 break;
1607 }
1608 }
1609 if (!found) {
1610 assert(n != entry, "can't be entry");
1611 _data_entry.push(n);
1612 }
1613 }
1614 }
1615 }
1617 // Find memory slices (head and tail)
1618 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1619 Node *n = lp()->fast_out(i);
1620 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1621 Node* n_tail = n->in(LoopNode::LoopBackControl);
1622 if (n_tail != n->in(LoopNode::EntryControl)) {
1623 _mem_slice_head.push(n);
1624 _mem_slice_tail.push(n_tail);
1625 }
1626 }
1627 }
1629 // Create an RPO list of nodes in block
1631 visited_clear();
1632 post_visited_clear();
1634 // Push all non-control nodes with no inputs from within block, then control entry
1635 for (int j = 0; j < _data_entry.length(); j++) {
1636 Node* n = _data_entry.at(j);
1637 visited_set(n);
1638 _stk.push(n);
1639 }
1640 visited_set(entry);
1641 _stk.push(entry);
1643 // Do a depth first walk over out edges
1644 int rpo_idx = bb_ct - 1;
1645 int size;
1646 while ((size = _stk.length()) > 0) {
1647 Node* n = _stk.top(); // Leave node on stack
1648 if (!visited_test_set(n)) {
1649 // forward arc in graph
1650 } else if (!post_visited_test(n)) {
1651 // cross or back arc
1652 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1653 Node *use = n->fast_out(i);
1654 if (in_bb(use) && !visited_test(use) &&
1655 // Don't go around backedge
1656 (!use->is_Phi() || n == entry)) {
1657 _stk.push(use);
1658 }
1659 }
1660 if (_stk.length() == size) {
1661 // There were no additional uses, post visit node now
1662 _stk.pop(); // Remove node from stack
1663 assert(rpo_idx >= 0, "");
1664 _block.at_put_grow(rpo_idx, n);
1665 rpo_idx--;
1666 post_visited_set(n);
1667 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1668 }
1669 } else {
1670 _stk.pop(); // Remove post-visited node from stack
1671 }
1672 }
1674 // Create real map of block indices for nodes
1675 for (int j = 0; j < _block.length(); j++) {
1676 Node* n = _block.at(j);
1677 set_bb_idx(n, j);
1678 }
1680 initialize_bb(); // Ensure extra info is allocated.
1682 #ifndef PRODUCT
1683 if (TraceSuperWord) {
1684 print_bb();
1685 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1686 for (int m = 0; m < _data_entry.length(); m++) {
1687 tty->print("%3d ", m);
1688 _data_entry.at(m)->dump();
1689 }
1690 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1691 for (int m = 0; m < _mem_slice_head.length(); m++) {
1692 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1693 tty->print(" "); _mem_slice_tail.at(m)->dump();
1694 }
1695 }
1696 #endif
1697 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1698 }
1700 //------------------------------initialize_bb---------------------------
1701 // Initialize per node info
1702 void SuperWord::initialize_bb() {
1703 Node* last = _block.at(_block.length() - 1);
1704 grow_node_info(bb_idx(last));
1705 }
1707 //------------------------------bb_insert_after---------------------------
1708 // Insert n into block after pos
1709 void SuperWord::bb_insert_after(Node* n, int pos) {
1710 int n_pos = pos + 1;
1711 // Make room
1712 for (int i = _block.length() - 1; i >= n_pos; i--) {
1713 _block.at_put_grow(i+1, _block.at(i));
1714 }
1715 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1716 _node_info.at_put_grow(j+1, _node_info.at(j));
1717 }
1718 // Set value
1719 _block.at_put_grow(n_pos, n);
1720 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1721 // Adjust map from node->_idx to _block index
1722 for (int i = n_pos; i < _block.length(); i++) {
1723 set_bb_idx(_block.at(i), i);
1724 }
1725 }
1727 //------------------------------compute_max_depth---------------------------
1728 // Compute max depth for expressions from beginning of block
1729 // Use to prune search paths during test for independence.
1730 void SuperWord::compute_max_depth() {
1731 int ct = 0;
1732 bool again;
1733 do {
1734 again = false;
1735 for (int i = 0; i < _block.length(); i++) {
1736 Node* n = _block.at(i);
1737 if (!n->is_Phi()) {
1738 int d_orig = depth(n);
1739 int d_in = 0;
1740 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1741 Node* pred = preds.current();
1742 if (in_bb(pred)) {
1743 d_in = MAX2(d_in, depth(pred));
1744 }
1745 }
1746 if (d_in + 1 != d_orig) {
1747 set_depth(n, d_in + 1);
1748 again = true;
1749 }
1750 }
1751 }
1752 ct++;
1753 } while (again);
1754 #ifndef PRODUCT
1755 if (TraceSuperWord && Verbose)
1756 tty->print_cr("compute_max_depth iterated: %d times", ct);
1757 #endif
1758 }
1760 //-------------------------compute_vector_element_type-----------------------
1761 // Compute necessary vector element type for expressions
1762 // This propagates backwards a narrower integer type when the
1763 // upper bits of the value are not needed.
1764 // Example: char a,b,c; a = b + c;
1765 // Normally the type of the add is integer, but for packed character
1766 // operations the type of the add needs to be char.
1767 void SuperWord::compute_vector_element_type() {
1768 #ifndef PRODUCT
1769 if (TraceSuperWord && Verbose)
1770 tty->print_cr("\ncompute_velt_type:");
1771 #endif
1773 // Initial type
1774 for (int i = 0; i < _block.length(); i++) {
1775 Node* n = _block.at(i);
1776 set_velt_type(n, container_type(n));
1777 }
1779 // Propagate integer narrowed type backwards through operations
1780 // that don't depend on higher order bits
1781 for (int i = _block.length() - 1; i >= 0; i--) {
1782 Node* n = _block.at(i);
1783 // Only integer types need be examined
1784 const Type* vtn = velt_type(n);
1785 if (vtn->basic_type() == T_INT) {
1786 uint start, end;
1787 VectorNode::vector_operands(n, &start, &end);
1789 for (uint j = start; j < end; j++) {
1790 Node* in = n->in(j);
1791 // Don't propagate through a memory
1792 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1793 data_size(n) < data_size(in)) {
1794 bool same_type = true;
1795 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1796 Node *use = in->fast_out(k);
1797 if (!in_bb(use) || !same_velt_type(use, n)) {
1798 same_type = false;
1799 break;
1800 }
1801 }
1802 if (same_type) {
1803 // For right shifts of small integer types (bool, byte, char, short)
1804 // we need precise information about sign-ness. Only Load nodes have
1805 // this information because Store nodes are the same for signed and
1806 // unsigned values. And any arithmetic operation after a load may
1807 // expand a value to signed Int so such right shifts can't be used
1808 // because vector elements do not have upper bits of Int.
1809 const Type* vt = vtn;
1810 if (VectorNode::is_shift(in)) {
1811 Node* load = in->in(1);
1812 if (load->is_Load() && (velt_type(load)->basic_type() == T_INT)) {
1813 vt = velt_type(load);
1814 } else if (in->Opcode() != Op_LShiftI) {
1815 // Widen type to Int to avoid creation of right shift vector
1816 // (align + data_size(s1) check in stmts_can_pack() will fail).
1817 // Note, left shifts work regardless type.
1818 vt = TypeInt::INT;
1819 }
1820 }
1821 set_velt_type(in, vt);
1822 }
1823 }
1824 }
1825 }
1826 }
1827 #ifndef PRODUCT
1828 if (TraceSuperWord && Verbose) {
1829 for (int i = 0; i < _block.length(); i++) {
1830 Node* n = _block.at(i);
1831 velt_type(n)->dump();
1832 tty->print("\t");
1833 n->dump();
1834 }
1835 }
1836 #endif
1837 }
1839 //------------------------------memory_alignment---------------------------
1840 // Alignment within a vector memory reference
1841 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1842 SWPointer p(s, this);
1843 if (!p.valid()) {
1844 return bottom_align;
1845 }
1846 int vw = vector_width_in_bytes(s);
1847 if (vw < 2) {
1848 return bottom_align; // No vectors for this type
1849 }
1850 int offset = p.offset_in_bytes();
1851 offset += iv_adjust*p.memory_size();
1852 int off_rem = offset % vw;
1853 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1854 return off_mod;
1855 }
1857 //---------------------------container_type---------------------------
1858 // Smallest type containing range of values
1859 const Type* SuperWord::container_type(Node* n) {
1860 if (n->is_Mem()) {
1861 BasicType bt = n->as_Mem()->memory_type();
1862 if (n->is_Store() && (bt == T_CHAR)) {
1863 // Use T_SHORT type instead of T_CHAR for stored values because any
1864 // preceding arithmetic operation extends values to signed Int.
1865 bt = T_SHORT;
1866 }
1867 if (n->Opcode() == Op_LoadUB) {
1868 // Adjust type for unsigned byte loads, it is important for right shifts.
1869 // T_BOOLEAN is used because there is no basic type representing type
1870 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1871 // size (one byte) and sign is important.
1872 bt = T_BOOLEAN;
1873 }
1874 return Type::get_const_basic_type(bt);
1875 }
1876 const Type* t = _igvn.type(n);
1877 if (t->basic_type() == T_INT) {
1878 // A narrow type of arithmetic operations will be determined by
1879 // propagating the type of memory operations.
1880 return TypeInt::INT;
1881 }
1882 return t;
1883 }
1885 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1886 const Type* vt1 = velt_type(n1);
1887 const Type* vt2 = velt_type(n2);
1888 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1889 // Compare vectors element sizes for integer types.
1890 return data_size(n1) == data_size(n2);
1891 }
1892 return vt1 == vt2;
1893 }
1895 //------------------------------in_packset---------------------------
1896 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1897 bool SuperWord::in_packset(Node* s1, Node* s2) {
1898 for (int i = 0; i < _packset.length(); i++) {
1899 Node_List* p = _packset.at(i);
1900 assert(p->size() == 2, "must be");
1901 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1902 return true;
1903 }
1904 }
1905 return false;
1906 }
1908 //------------------------------in_pack---------------------------
1909 // Is s in pack p?
1910 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1911 for (uint i = 0; i < p->size(); i++) {
1912 if (p->at(i) == s) {
1913 return p;
1914 }
1915 }
1916 return NULL;
1917 }
1919 //------------------------------remove_pack_at---------------------------
1920 // Remove the pack at position pos in the packset
1921 void SuperWord::remove_pack_at(int pos) {
1922 Node_List* p = _packset.at(pos);
1923 for (uint i = 0; i < p->size(); i++) {
1924 Node* s = p->at(i);
1925 set_my_pack(s, NULL);
1926 }
1927 _packset.remove_at(pos);
1928 }
1930 //------------------------------executed_first---------------------------
1931 // Return the node executed first in pack p. Uses the RPO block list
1932 // to determine order.
1933 Node* SuperWord::executed_first(Node_List* p) {
1934 Node* n = p->at(0);
1935 int n_rpo = bb_idx(n);
1936 for (uint i = 1; i < p->size(); i++) {
1937 Node* s = p->at(i);
1938 int s_rpo = bb_idx(s);
1939 if (s_rpo < n_rpo) {
1940 n = s;
1941 n_rpo = s_rpo;
1942 }
1943 }
1944 return n;
1945 }
1947 //------------------------------executed_last---------------------------
1948 // Return the node executed last in pack p.
1949 Node* SuperWord::executed_last(Node_List* p) {
1950 Node* n = p->at(0);
1951 int n_rpo = bb_idx(n);
1952 for (uint i = 1; i < p->size(); i++) {
1953 Node* s = p->at(i);
1954 int s_rpo = bb_idx(s);
1955 if (s_rpo > n_rpo) {
1956 n = s;
1957 n_rpo = s_rpo;
1958 }
1959 }
1960 return n;
1961 }
1963 //----------------------------align_initial_loop_index---------------------------
1964 // Adjust pre-loop limit so that in main loop, a load/store reference
1965 // to align_to_ref will be a position zero in the vector.
1966 // (iv + k) mod vector_align == 0
1967 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1968 CountedLoopNode *main_head = lp()->as_CountedLoop();
1969 assert(main_head->is_main_loop(), "");
1970 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1971 assert(pre_end != NULL, "");
1972 Node *pre_opaq1 = pre_end->limit();
1973 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1974 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1975 Node *lim0 = pre_opaq->in(1);
1977 // Where we put new limit calculations
1978 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1980 // Ensure the original loop limit is available from the
1981 // pre-loop Opaque1 node.
1982 Node *orig_limit = pre_opaq->original_loop_limit();
1983 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1985 SWPointer align_to_ref_p(align_to_ref, this);
1986 assert(align_to_ref_p.valid(), "sanity");
1988 // Given:
1989 // lim0 == original pre loop limit
1990 // V == v_align (power of 2)
1991 // invar == extra invariant piece of the address expression
1992 // e == offset [ +/- invar ]
1993 //
1994 // When reassociating expressions involving '%' the basic rules are:
1995 // (a - b) % k == 0 => a % k == b % k
1996 // and:
1997 // (a + b) % k == 0 => a % k == (k - b) % k
1998 //
1999 // For stride > 0 && scale > 0,
2000 // Derive the new pre-loop limit "lim" such that the two constraints:
2001 // (1) lim = lim0 + N (where N is some positive integer < V)
2002 // (2) (e + lim) % V == 0
2003 // are true.
2004 //
2005 // Substituting (1) into (2),
2006 // (e + lim0 + N) % V == 0
2007 // solve for N:
2008 // N = (V - (e + lim0)) % V
2009 // substitute back into (1), so that new limit
2010 // lim = lim0 + (V - (e + lim0)) % V
2011 //
2012 // For stride > 0 && scale < 0
2013 // Constraints:
2014 // lim = lim0 + N
2015 // (e - lim) % V == 0
2016 // Solving for lim:
2017 // (e - lim0 - N) % V == 0
2018 // N = (e - lim0) % V
2019 // lim = lim0 + (e - lim0) % V
2020 //
2021 // For stride < 0 && scale > 0
2022 // Constraints:
2023 // lim = lim0 - N
2024 // (e + lim) % V == 0
2025 // Solving for lim:
2026 // (e + lim0 - N) % V == 0
2027 // N = (e + lim0) % V
2028 // lim = lim0 - (e + lim0) % V
2029 //
2030 // For stride < 0 && scale < 0
2031 // Constraints:
2032 // lim = lim0 - N
2033 // (e - lim) % V == 0
2034 // Solving for lim:
2035 // (e - lim0 + N) % V == 0
2036 // N = (V - (e - lim0)) % V
2037 // lim = lim0 - (V - (e - lim0)) % V
2039 int vw = vector_width_in_bytes(align_to_ref);
2040 int stride = iv_stride();
2041 int scale = align_to_ref_p.scale_in_bytes();
2042 int elt_size = align_to_ref_p.memory_size();
2043 int v_align = vw / elt_size;
2044 assert(v_align > 1, "sanity");
2045 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2046 Node *offsn = _igvn.intcon(offset);
2048 Node *e = offsn;
2049 if (align_to_ref_p.invar() != NULL) {
2050 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2051 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2052 Node* aref = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2053 _igvn.register_new_node_with_optimizer(aref);
2054 _phase->set_ctrl(aref, pre_ctrl);
2055 if (align_to_ref_p.negate_invar()) {
2056 e = new (_phase->C) SubINode(e, aref);
2057 } else {
2058 e = new (_phase->C) AddINode(e, aref);
2059 }
2060 _igvn.register_new_node_with_optimizer(e);
2061 _phase->set_ctrl(e, pre_ctrl);
2062 }
2063 if (vw > ObjectAlignmentInBytes) {
2064 // incorporate base e +/- base && Mask >>> log2(elt)
2065 Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2066 _igvn.register_new_node_with_optimizer(xbase);
2067 #ifdef _LP64
2068 xbase = new (_phase->C) ConvL2INode(xbase);
2069 _igvn.register_new_node_with_optimizer(xbase);
2070 #endif
2071 Node* mask = _igvn.intcon(vw-1);
2072 Node* masked_xbase = new (_phase->C) AndINode(xbase, mask);
2073 _igvn.register_new_node_with_optimizer(masked_xbase);
2074 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2075 Node* bref = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2076 _igvn.register_new_node_with_optimizer(bref);
2077 _phase->set_ctrl(bref, pre_ctrl);
2078 e = new (_phase->C) AddINode(e, bref);
2079 _igvn.register_new_node_with_optimizer(e);
2080 _phase->set_ctrl(e, pre_ctrl);
2081 }
2083 // compute e +/- lim0
2084 if (scale < 0) {
2085 e = new (_phase->C) SubINode(e, lim0);
2086 } else {
2087 e = new (_phase->C) AddINode(e, lim0);
2088 }
2089 _igvn.register_new_node_with_optimizer(e);
2090 _phase->set_ctrl(e, pre_ctrl);
2092 if (stride * scale > 0) {
2093 // compute V - (e +/- lim0)
2094 Node* va = _igvn.intcon(v_align);
2095 e = new (_phase->C) SubINode(va, e);
2096 _igvn.register_new_node_with_optimizer(e);
2097 _phase->set_ctrl(e, pre_ctrl);
2098 }
2099 // compute N = (exp) % V
2100 Node* va_msk = _igvn.intcon(v_align - 1);
2101 Node* N = new (_phase->C) AndINode(e, va_msk);
2102 _igvn.register_new_node_with_optimizer(N);
2103 _phase->set_ctrl(N, pre_ctrl);
2105 // substitute back into (1), so that new limit
2106 // lim = lim0 + N
2107 Node* lim;
2108 if (stride < 0) {
2109 lim = new (_phase->C) SubINode(lim0, N);
2110 } else {
2111 lim = new (_phase->C) AddINode(lim0, N);
2112 }
2113 _igvn.register_new_node_with_optimizer(lim);
2114 _phase->set_ctrl(lim, pre_ctrl);
2115 Node* constrained =
2116 (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2117 : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2118 _igvn.register_new_node_with_optimizer(constrained);
2119 _phase->set_ctrl(constrained, pre_ctrl);
2120 _igvn.hash_delete(pre_opaq);
2121 pre_opaq->set_req(1, constrained);
2122 }
2124 //----------------------------get_pre_loop_end---------------------------
2125 // Find pre loop end from main loop. Returns null if none.
2126 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2127 Node *ctrl = cl->in(LoopNode::EntryControl);
2128 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2129 Node *iffm = ctrl->in(0);
2130 if (!iffm->is_If()) return NULL;
2131 Node *p_f = iffm->in(0);
2132 if (!p_f->is_IfFalse()) return NULL;
2133 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2134 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2135 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2136 return pre_end;
2137 }
2140 //------------------------------init---------------------------
2141 void SuperWord::init() {
2142 _dg.init();
2143 _packset.clear();
2144 _disjoint_ptrs.clear();
2145 _block.clear();
2146 _data_entry.clear();
2147 _mem_slice_head.clear();
2148 _mem_slice_tail.clear();
2149 _node_info.clear();
2150 _align_to_ref = NULL;
2151 _lpt = NULL;
2152 _lp = NULL;
2153 _bb = NULL;
2154 _iv = NULL;
2155 }
2157 //------------------------------print_packset---------------------------
2158 void SuperWord::print_packset() {
2159 #ifndef PRODUCT
2160 tty->print_cr("packset");
2161 for (int i = 0; i < _packset.length(); i++) {
2162 tty->print_cr("Pack: %d", i);
2163 Node_List* p = _packset.at(i);
2164 print_pack(p);
2165 }
2166 #endif
2167 }
2169 //------------------------------print_pack---------------------------
2170 void SuperWord::print_pack(Node_List* p) {
2171 for (uint i = 0; i < p->size(); i++) {
2172 print_stmt(p->at(i));
2173 }
2174 }
2176 //------------------------------print_bb---------------------------
2177 void SuperWord::print_bb() {
2178 #ifndef PRODUCT
2179 tty->print_cr("\nBlock");
2180 for (int i = 0; i < _block.length(); i++) {
2181 Node* n = _block.at(i);
2182 tty->print("%d ", i);
2183 if (n) {
2184 n->dump();
2185 }
2186 }
2187 #endif
2188 }
2190 //------------------------------print_stmt---------------------------
2191 void SuperWord::print_stmt(Node* s) {
2192 #ifndef PRODUCT
2193 tty->print(" align: %d \t", alignment(s));
2194 s->dump();
2195 #endif
2196 }
2198 //------------------------------blank---------------------------
2199 char* SuperWord::blank(uint depth) {
2200 static char blanks[101];
2201 assert(depth < 101, "too deep");
2202 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2203 blanks[depth] = '\0';
2204 return blanks;
2205 }
2208 //==============================SWPointer===========================
2210 //----------------------------SWPointer------------------------
2211 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2212 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2213 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2215 Node* adr = mem->in(MemNode::Address);
2216 if (!adr->is_AddP()) {
2217 assert(!valid(), "too complex");
2218 return;
2219 }
2220 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2221 Node* base = adr->in(AddPNode::Base);
2222 //unsafe reference could not be aligned appropriately without runtime checking
2223 if (base == NULL || base->bottom_type() == Type::TOP) {
2224 assert(!valid(), "unsafe access");
2225 return;
2226 }
2227 for (int i = 0; i < 3; i++) {
2228 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2229 assert(!valid(), "too complex");
2230 return;
2231 }
2232 adr = adr->in(AddPNode::Address);
2233 if (base == adr || !adr->is_AddP()) {
2234 break; // stop looking at addp's
2235 }
2236 }
2237 _base = base;
2238 _adr = adr;
2239 assert(valid(), "Usable");
2240 }
2242 // Following is used to create a temporary object during
2243 // the pattern match of an address expression.
2244 SWPointer::SWPointer(SWPointer* p) :
2245 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2246 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2248 //------------------------scaled_iv_plus_offset--------------------
2249 // Match: k*iv + offset
2250 // where: k is a constant that maybe zero, and
2251 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2252 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2253 if (scaled_iv(n)) {
2254 return true;
2255 }
2256 if (offset_plus_k(n)) {
2257 return true;
2258 }
2259 int opc = n->Opcode();
2260 if (opc == Op_AddI) {
2261 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2262 return true;
2263 }
2264 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2265 return true;
2266 }
2267 } else if (opc == Op_SubI) {
2268 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2269 return true;
2270 }
2271 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2272 _scale *= -1;
2273 return true;
2274 }
2275 }
2276 return false;
2277 }
2279 //----------------------------scaled_iv------------------------
2280 // Match: k*iv where k is a constant that's not zero
2281 bool SWPointer::scaled_iv(Node* n) {
2282 if (_scale != 0) {
2283 return false; // already found a scale
2284 }
2285 if (n == iv()) {
2286 _scale = 1;
2287 return true;
2288 }
2289 int opc = n->Opcode();
2290 if (opc == Op_MulI) {
2291 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2292 _scale = n->in(2)->get_int();
2293 return true;
2294 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2295 _scale = n->in(1)->get_int();
2296 return true;
2297 }
2298 } else if (opc == Op_LShiftI) {
2299 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2300 _scale = 1 << n->in(2)->get_int();
2301 return true;
2302 }
2303 } else if (opc == Op_ConvI2L) {
2304 if (scaled_iv_plus_offset(n->in(1))) {
2305 return true;
2306 }
2307 } else if (opc == Op_LShiftL) {
2308 if (!has_iv() && _invar == NULL) {
2309 // Need to preserve the current _offset value, so
2310 // create a temporary object for this expression subtree.
2311 // Hacky, so should re-engineer the address pattern match.
2312 SWPointer tmp(this);
2313 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2314 if (tmp._invar == NULL) {
2315 int mult = 1 << n->in(2)->get_int();
2316 _scale = tmp._scale * mult;
2317 _offset += tmp._offset * mult;
2318 return true;
2319 }
2320 }
2321 }
2322 }
2323 return false;
2324 }
2326 //----------------------------offset_plus_k------------------------
2327 // Match: offset is (k [+/- invariant])
2328 // where k maybe zero and invariant is optional, but not both.
2329 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2330 int opc = n->Opcode();
2331 if (opc == Op_ConI) {
2332 _offset += negate ? -(n->get_int()) : n->get_int();
2333 return true;
2334 } else if (opc == Op_ConL) {
2335 // Okay if value fits into an int
2336 const TypeLong* t = n->find_long_type();
2337 if (t->higher_equal(TypeLong::INT)) {
2338 jlong loff = n->get_long();
2339 jint off = (jint)loff;
2340 _offset += negate ? -off : loff;
2341 return true;
2342 }
2343 return false;
2344 }
2345 if (_invar != NULL) return false; // already have an invariant
2346 if (opc == Op_AddI) {
2347 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2348 _negate_invar = negate;
2349 _invar = n->in(1);
2350 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2351 return true;
2352 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2353 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2354 _negate_invar = negate;
2355 _invar = n->in(2);
2356 return true;
2357 }
2358 }
2359 if (opc == Op_SubI) {
2360 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2361 _negate_invar = negate;
2362 _invar = n->in(1);
2363 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2364 return true;
2365 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2366 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2367 _negate_invar = !negate;
2368 _invar = n->in(2);
2369 return true;
2370 }
2371 }
2372 if (invariant(n)) {
2373 _negate_invar = negate;
2374 _invar = n;
2375 return true;
2376 }
2377 return false;
2378 }
2380 //----------------------------print------------------------
2381 void SWPointer::print() {
2382 #ifndef PRODUCT
2383 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2384 _base != NULL ? _base->_idx : 0,
2385 _adr != NULL ? _adr->_idx : 0,
2386 _scale, _offset,
2387 _negate_invar?'-':'+',
2388 _invar != NULL ? _invar->_idx : 0);
2389 #endif
2390 }
2392 // ========================= OrderedPair =====================
2394 const OrderedPair OrderedPair::initial;
2396 // ========================= SWNodeInfo =====================
2398 const SWNodeInfo SWNodeInfo::initial;
2401 // ============================ DepGraph ===========================
2403 //------------------------------make_node---------------------------
2404 // Make a new dependence graph node for an ideal node.
2405 DepMem* DepGraph::make_node(Node* node) {
2406 DepMem* m = new (_arena) DepMem(node);
2407 if (node != NULL) {
2408 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2409 _map.at_put_grow(node->_idx, m);
2410 }
2411 return m;
2412 }
2414 //------------------------------make_edge---------------------------
2415 // Make a new dependence graph edge from dpred -> dsucc
2416 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2417 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2418 dpred->set_out_head(e);
2419 dsucc->set_in_head(e);
2420 return e;
2421 }
2423 // ========================== DepMem ========================
2425 //------------------------------in_cnt---------------------------
2426 int DepMem::in_cnt() {
2427 int ct = 0;
2428 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2429 return ct;
2430 }
2432 //------------------------------out_cnt---------------------------
2433 int DepMem::out_cnt() {
2434 int ct = 0;
2435 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2436 return ct;
2437 }
2439 //------------------------------print-----------------------------
2440 void DepMem::print() {
2441 #ifndef PRODUCT
2442 tty->print(" DepNode %d (", _node->_idx);
2443 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2444 Node* pred = p->pred()->node();
2445 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2446 }
2447 tty->print(") [");
2448 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2449 Node* succ = s->succ()->node();
2450 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2451 }
2452 tty->print_cr(" ]");
2453 #endif
2454 }
2456 // =========================== DepEdge =========================
2458 //------------------------------DepPreds---------------------------
2459 void DepEdge::print() {
2460 #ifndef PRODUCT
2461 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2462 #endif
2463 }
2465 // =========================== DepPreds =========================
2466 // Iterator over predecessor edges in the dependence graph.
2468 //------------------------------DepPreds---------------------------
2469 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2470 _n = n;
2471 _done = false;
2472 if (_n->is_Store() || _n->is_Load()) {
2473 _next_idx = MemNode::Address;
2474 _end_idx = n->req();
2475 _dep_next = dg.dep(_n)->in_head();
2476 } else if (_n->is_Mem()) {
2477 _next_idx = 0;
2478 _end_idx = 0;
2479 _dep_next = dg.dep(_n)->in_head();
2480 } else {
2481 _next_idx = 1;
2482 _end_idx = _n->req();
2483 _dep_next = NULL;
2484 }
2485 next();
2486 }
2488 //------------------------------next---------------------------
2489 void DepPreds::next() {
2490 if (_dep_next != NULL) {
2491 _current = _dep_next->pred()->node();
2492 _dep_next = _dep_next->next_in();
2493 } else if (_next_idx < _end_idx) {
2494 _current = _n->in(_next_idx++);
2495 } else {
2496 _done = true;
2497 }
2498 }
2500 // =========================== DepSuccs =========================
2501 // Iterator over successor edges in the dependence graph.
2503 //------------------------------DepSuccs---------------------------
2504 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2505 _n = n;
2506 _done = false;
2507 if (_n->is_Load()) {
2508 _next_idx = 0;
2509 _end_idx = _n->outcnt();
2510 _dep_next = dg.dep(_n)->out_head();
2511 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2512 _next_idx = 0;
2513 _end_idx = 0;
2514 _dep_next = dg.dep(_n)->out_head();
2515 } else {
2516 _next_idx = 0;
2517 _end_idx = _n->outcnt();
2518 _dep_next = NULL;
2519 }
2520 next();
2521 }
2523 //-------------------------------next---------------------------
2524 void DepSuccs::next() {
2525 if (_dep_next != NULL) {
2526 _current = _dep_next->succ()->node();
2527 _dep_next = _dep_next->next_out();
2528 } else if (_next_idx < _end_idx) {
2529 _current = _n->raw_out(_next_idx++);
2530 } else {
2531 _done = true;
2532 }
2533 }