Wed, 19 Sep 2012 16:50:26 -0700
7199010: incorrect vector alignment
Summary: Fixed vectors alignment when several arrays are accessed in one loop.
Reviewed-by: roland, twisti
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 n->Opcode() != Op_LoadUI2L &&
183 is_java_primitive(n->as_Mem()->memory_type())) {
184 int align = memory_alignment(n->as_Mem(), 0);
185 if (align != bottom_align) {
186 memops.push(n);
187 }
188 }
189 }
191 Node_List align_to_refs;
192 int best_iv_adjustment = 0;
193 MemNode* best_align_to_mem_ref = NULL;
195 while (memops.size() != 0) {
196 // Find a memory reference to align to.
197 MemNode* mem_ref = find_align_to_ref(memops);
198 if (mem_ref == NULL) break;
199 align_to_refs.push(mem_ref);
200 int iv_adjustment = get_iv_adjustment(mem_ref);
202 if (best_align_to_mem_ref == NULL) {
203 // Set memory reference which is the best from all memory operations
204 // to be used for alignment. The pre-loop trip count is modified to align
205 // this reference to a vector-aligned address.
206 best_align_to_mem_ref = mem_ref;
207 best_iv_adjustment = iv_adjustment;
208 }
210 SWPointer align_to_ref_p(mem_ref, this);
211 // Set alignment relative to "align_to_ref" for all related memory operations.
212 for (int i = memops.size() - 1; i >= 0; i--) {
213 MemNode* s = memops.at(i)->as_Mem();
214 if (isomorphic(s, mem_ref)) {
215 SWPointer p2(s, this);
216 if (p2.comparable(align_to_ref_p)) {
217 int align = memory_alignment(s, iv_adjustment);
218 set_alignment(s, align);
219 }
220 }
221 }
223 // Create initial pack pairs of memory operations for which
224 // alignment is set and vectors will be aligned.
225 bool create_pack = true;
226 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
227 if (!Matcher::misaligned_vectors_ok()) {
228 int vw = vector_width(mem_ref);
229 int vw_best = vector_width(best_align_to_mem_ref);
230 if (vw > vw_best) {
231 // Do not vectorize a memory access with more elements per vector
232 // if unaligned memory access is not allowed because number of
233 // iterations in pre-loop will be not enough to align it.
234 create_pack = false;
235 }
236 }
237 } else {
238 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
239 // Can't allow vectorization of unaligned memory accesses with the
240 // same type since it could be overlapped accesses to the same array.
241 create_pack = false;
242 } else {
243 // Allow independent (different type) unaligned memory operations
244 // if HW supports them.
245 if (!Matcher::misaligned_vectors_ok()) {
246 create_pack = false;
247 } else {
248 // Check if packs of the same memory type but
249 // with a different alignment were created before.
250 for (uint i = 0; i < align_to_refs.size(); i++) {
251 MemNode* mr = align_to_refs.at(i)->as_Mem();
252 if (same_velt_type(mr, mem_ref) &&
253 memory_alignment(mr, iv_adjustment) != 0)
254 create_pack = false;
255 }
256 }
257 }
258 }
259 if (create_pack) {
260 for (uint i = 0; i < memops.size(); i++) {
261 Node* s1 = memops.at(i);
262 int align = alignment(s1);
263 if (align == top_align) continue;
264 for (uint j = 0; j < memops.size(); j++) {
265 Node* s2 = memops.at(j);
266 if (alignment(s2) == top_align) continue;
267 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
268 if (stmts_can_pack(s1, s2, align)) {
269 Node_List* pair = new Node_List();
270 pair->push(s1);
271 pair->push(s2);
272 _packset.append(pair);
273 }
274 }
275 }
276 }
277 } else { // Don't create unaligned pack
278 // First, remove remaining memory ops of the same type from the list.
279 for (int i = memops.size() - 1; i >= 0; i--) {
280 MemNode* s = memops.at(i)->as_Mem();
281 if (same_velt_type(s, mem_ref)) {
282 memops.remove(i);
283 }
284 }
286 // Second, remove already constructed packs of the same type.
287 for (int i = _packset.length() - 1; i >= 0; i--) {
288 Node_List* p = _packset.at(i);
289 MemNode* s = p->at(0)->as_Mem();
290 if (same_velt_type(s, mem_ref)) {
291 remove_pack_at(i);
292 }
293 }
295 // If needed find the best memory reference for loop alignment again.
296 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
297 // Put memory ops from remaining packs back on memops list for
298 // the best alignment search.
299 uint orig_msize = memops.size();
300 for (int i = 0; i < _packset.length(); i++) {
301 Node_List* p = _packset.at(i);
302 MemNode* s = p->at(0)->as_Mem();
303 assert(!same_velt_type(s, mem_ref), "sanity");
304 memops.push(s);
305 }
306 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
307 if (best_align_to_mem_ref == NULL) break;
308 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
309 // Restore list.
310 while (memops.size() > orig_msize)
311 (void)memops.pop();
312 }
313 } // unaligned memory accesses
315 // Remove used mem nodes.
316 for (int i = memops.size() - 1; i >= 0; i--) {
317 MemNode* m = memops.at(i)->as_Mem();
318 if (alignment(m) != top_align) {
319 memops.remove(i);
320 }
321 }
323 } // while (memops.size() != 0
324 set_align_to_ref(best_align_to_mem_ref);
326 #ifndef PRODUCT
327 if (TraceSuperWord) {
328 tty->print_cr("\nAfter find_adjacent_refs");
329 print_packset();
330 }
331 #endif
332 }
334 //------------------------------find_align_to_ref---------------------------
335 // Find a memory reference to align the loop induction variable to.
336 // Looks first at stores then at loads, looking for a memory reference
337 // with the largest number of references similar to it.
338 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
339 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
341 // Count number of comparable memory ops
342 for (uint i = 0; i < memops.size(); i++) {
343 MemNode* s1 = memops.at(i)->as_Mem();
344 SWPointer p1(s1, this);
345 // Discard if pre loop can't align this reference
346 if (!ref_is_alignable(p1)) {
347 *cmp_ct.adr_at(i) = 0;
348 continue;
349 }
350 for (uint j = i+1; j < memops.size(); j++) {
351 MemNode* s2 = memops.at(j)->as_Mem();
352 if (isomorphic(s1, s2)) {
353 SWPointer p2(s2, this);
354 if (p1.comparable(p2)) {
355 (*cmp_ct.adr_at(i))++;
356 (*cmp_ct.adr_at(j))++;
357 }
358 }
359 }
360 }
362 // Find Store (or Load) with the greatest number of "comparable" references,
363 // biggest vector size, smallest data size and smallest iv offset.
364 int max_ct = 0;
365 int max_vw = 0;
366 int max_idx = -1;
367 int min_size = max_jint;
368 int min_iv_offset = max_jint;
369 for (uint j = 0; j < memops.size(); j++) {
370 MemNode* s = memops.at(j)->as_Mem();
371 if (s->is_Store()) {
372 int vw = vector_width_in_bytes(s);
373 assert(vw > 1, "sanity");
374 SWPointer p(s, this);
375 if (cmp_ct.at(j) > max_ct ||
376 cmp_ct.at(j) == max_ct &&
377 (vw > max_vw ||
378 vw == max_vw &&
379 (data_size(s) < min_size ||
380 data_size(s) == min_size &&
381 (p.offset_in_bytes() < min_iv_offset)))) {
382 max_ct = cmp_ct.at(j);
383 max_vw = vw;
384 max_idx = j;
385 min_size = data_size(s);
386 min_iv_offset = p.offset_in_bytes();
387 }
388 }
389 }
390 // If no stores, look at loads
391 if (max_ct == 0) {
392 for (uint j = 0; j < memops.size(); j++) {
393 MemNode* s = memops.at(j)->as_Mem();
394 if (s->is_Load()) {
395 int vw = vector_width_in_bytes(s);
396 assert(vw > 1, "sanity");
397 SWPointer p(s, this);
398 if (cmp_ct.at(j) > max_ct ||
399 cmp_ct.at(j) == max_ct &&
400 (vw > max_vw ||
401 vw == max_vw &&
402 (data_size(s) < min_size ||
403 data_size(s) == min_size &&
404 (p.offset_in_bytes() < min_iv_offset)))) {
405 max_ct = cmp_ct.at(j);
406 max_vw = vw;
407 max_idx = j;
408 min_size = data_size(s);
409 min_iv_offset = p.offset_in_bytes();
410 }
411 }
412 }
413 }
415 #ifdef ASSERT
416 if (TraceSuperWord && Verbose) {
417 tty->print_cr("\nVector memops after find_align_to_refs");
418 for (uint i = 0; i < memops.size(); i++) {
419 MemNode* s = memops.at(i)->as_Mem();
420 s->dump();
421 }
422 }
423 #endif
425 if (max_ct > 0) {
426 #ifdef ASSERT
427 if (TraceSuperWord) {
428 tty->print("\nVector align to node: ");
429 memops.at(max_idx)->as_Mem()->dump();
430 }
431 #endif
432 return memops.at(max_idx)->as_Mem();
433 }
434 return NULL;
435 }
437 //------------------------------ref_is_alignable---------------------------
438 // Can the preloop align the reference to position zero in the vector?
439 bool SuperWord::ref_is_alignable(SWPointer& p) {
440 if (!p.has_iv()) {
441 return true; // no induction variable
442 }
443 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
444 assert(pre_end->stride_is_con(), "pre loop stride is constant");
445 int preloop_stride = pre_end->stride_con();
447 int span = preloop_stride * p.scale_in_bytes();
449 // Stride one accesses are alignable.
450 if (ABS(span) == p.memory_size())
451 return true;
453 // If initial offset from start of object is computable,
454 // compute alignment within the vector.
455 int vw = vector_width_in_bytes(p.mem());
456 assert(vw > 1, "sanity");
457 if (vw % span == 0) {
458 Node* init_nd = pre_end->init_trip();
459 if (init_nd->is_Con() && p.invar() == NULL) {
460 int init = init_nd->bottom_type()->is_int()->get_con();
462 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
463 assert(init_offset >= 0, "positive offset from object start");
465 if (span > 0) {
466 return (vw - (init_offset % vw)) % span == 0;
467 } else {
468 assert(span < 0, "nonzero stride * scale");
469 return (init_offset % vw) % -span == 0;
470 }
471 }
472 }
473 return false;
474 }
476 //---------------------------get_iv_adjustment---------------------------
477 // Calculate loop's iv adjustment for this memory ops.
478 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
479 SWPointer align_to_ref_p(mem_ref, this);
480 int offset = align_to_ref_p.offset_in_bytes();
481 int scale = align_to_ref_p.scale_in_bytes();
482 int vw = vector_width_in_bytes(mem_ref);
483 assert(vw > 1, "sanity");
484 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
485 // At least one iteration is executed in pre-loop by default. As result
486 // several iterations are needed to align memory operations in main-loop even
487 // if offset is 0.
488 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
489 int elt_size = align_to_ref_p.memory_size();
490 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
491 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
492 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
494 #ifndef PRODUCT
495 if (TraceSuperWord)
496 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
497 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
498 #endif
499 return iv_adjustment;
500 }
502 //---------------------------dependence_graph---------------------------
503 // Construct dependency graph.
504 // Add dependence edges to load/store nodes for memory dependence
505 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
506 void SuperWord::dependence_graph() {
507 // First, assign a dependence node to each memory node
508 for (int i = 0; i < _block.length(); i++ ) {
509 Node *n = _block.at(i);
510 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
511 _dg.make_node(n);
512 }
513 }
515 // For each memory slice, create the dependences
516 for (int i = 0; i < _mem_slice_head.length(); i++) {
517 Node* n = _mem_slice_head.at(i);
518 Node* n_tail = _mem_slice_tail.at(i);
520 // Get slice in predecessor order (last is first)
521 mem_slice_preds(n_tail, n, _nlist);
523 // Make the slice dependent on the root
524 DepMem* slice = _dg.dep(n);
525 _dg.make_edge(_dg.root(), slice);
527 // Create a sink for the slice
528 DepMem* slice_sink = _dg.make_node(NULL);
529 _dg.make_edge(slice_sink, _dg.tail());
531 // Now visit each pair of memory ops, creating the edges
532 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
533 Node* s1 = _nlist.at(j);
535 // If no dependency yet, use slice
536 if (_dg.dep(s1)->in_cnt() == 0) {
537 _dg.make_edge(slice, s1);
538 }
539 SWPointer p1(s1->as_Mem(), this);
540 bool sink_dependent = true;
541 for (int k = j - 1; k >= 0; k--) {
542 Node* s2 = _nlist.at(k);
543 if (s1->is_Load() && s2->is_Load())
544 continue;
545 SWPointer p2(s2->as_Mem(), this);
547 int cmp = p1.cmp(p2);
548 if (SuperWordRTDepCheck &&
549 p1.base() != p2.base() && p1.valid() && p2.valid()) {
550 // Create a runtime check to disambiguate
551 OrderedPair pp(p1.base(), p2.base());
552 _disjoint_ptrs.append_if_missing(pp);
553 } else if (!SWPointer::not_equal(cmp)) {
554 // Possibly same address
555 _dg.make_edge(s1, s2);
556 sink_dependent = false;
557 }
558 }
559 if (sink_dependent) {
560 _dg.make_edge(s1, slice_sink);
561 }
562 }
563 #ifndef PRODUCT
564 if (TraceSuperWord) {
565 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
566 for (int q = 0; q < _nlist.length(); q++) {
567 _dg.print(_nlist.at(q));
568 }
569 tty->cr();
570 }
571 #endif
572 _nlist.clear();
573 }
575 #ifndef PRODUCT
576 if (TraceSuperWord) {
577 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
578 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
579 _disjoint_ptrs.at(r).print();
580 tty->cr();
581 }
582 tty->cr();
583 }
584 #endif
585 }
587 //---------------------------mem_slice_preds---------------------------
588 // Return a memory slice (node list) in predecessor order starting at "start"
589 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
590 assert(preds.length() == 0, "start empty");
591 Node* n = start;
592 Node* prev = NULL;
593 while (true) {
594 assert(in_bb(n), "must be in block");
595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
596 Node* out = n->fast_out(i);
597 if (out->is_Load()) {
598 if (in_bb(out)) {
599 preds.push(out);
600 }
601 } else {
602 // FIXME
603 if (out->is_MergeMem() && !in_bb(out)) {
604 // Either unrolling is causing a memory edge not to disappear,
605 // or need to run igvn.optimize() again before SLP
606 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
607 // Ditto. Not sure what else to check further.
608 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
609 // StoreCM has an input edge used as a precedence edge.
610 // Maybe an issue when oop stores are vectorized.
611 } else {
612 assert(out == prev || prev == NULL, "no branches off of store slice");
613 }
614 }
615 }
616 if (n == stop) break;
617 preds.push(n);
618 prev = n;
619 n = n->in(MemNode::Memory);
620 }
621 }
623 //------------------------------stmts_can_pack---------------------------
624 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
625 // s1 aligned at "align"
626 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
628 // Do not use superword for non-primitives
629 BasicType bt1 = velt_basic_type(s1);
630 BasicType bt2 = velt_basic_type(s2);
631 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
632 return false;
633 if (Matcher::max_vector_size(bt1) < 2) {
634 return false; // No vectors for this type
635 }
637 if (isomorphic(s1, s2)) {
638 if (independent(s1, s2)) {
639 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
640 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
641 int s1_align = alignment(s1);
642 int s2_align = alignment(s2);
643 if (s1_align == top_align || s1_align == align) {
644 if (s2_align == top_align || s2_align == align + data_size(s1)) {
645 return true;
646 }
647 }
648 }
649 }
650 }
651 }
652 return false;
653 }
655 //------------------------------exists_at---------------------------
656 // Does s exist in a pack at position pos?
657 bool SuperWord::exists_at(Node* s, uint pos) {
658 for (int i = 0; i < _packset.length(); i++) {
659 Node_List* p = _packset.at(i);
660 if (p->at(pos) == s) {
661 return true;
662 }
663 }
664 return false;
665 }
667 //------------------------------are_adjacent_refs---------------------------
668 // Is s1 immediately before s2 in memory?
669 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
670 if (!s1->is_Mem() || !s2->is_Mem()) return false;
671 if (!in_bb(s1) || !in_bb(s2)) return false;
673 // Do not use superword for non-primitives
674 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
675 !is_java_primitive(s2->as_Mem()->memory_type())) {
676 return false;
677 }
679 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
680 // only pack memops that are in the same alias set until that's fixed.
681 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
682 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
683 return false;
684 SWPointer p1(s1->as_Mem(), this);
685 SWPointer p2(s2->as_Mem(), this);
686 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
687 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
688 return diff == data_size(s1);
689 }
691 //------------------------------isomorphic---------------------------
692 // Are s1 and s2 similar?
693 bool SuperWord::isomorphic(Node* s1, Node* s2) {
694 if (s1->Opcode() != s2->Opcode()) return false;
695 if (s1->req() != s2->req()) return false;
696 if (s1->in(0) != s2->in(0)) return false;
697 if (!same_velt_type(s1, s2)) return false;
698 return true;
699 }
701 //------------------------------independent---------------------------
702 // Is there no data path from s1 to s2 or s2 to s1?
703 bool SuperWord::independent(Node* s1, Node* s2) {
704 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
705 int d1 = depth(s1);
706 int d2 = depth(s2);
707 if (d1 == d2) return s1 != s2;
708 Node* deep = d1 > d2 ? s1 : s2;
709 Node* shallow = d1 > d2 ? s2 : s1;
711 visited_clear();
713 return independent_path(shallow, deep);
714 }
716 //------------------------------independent_path------------------------------
717 // Helper for independent
718 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
719 if (dp >= 1000) return false; // stop deep recursion
720 visited_set(deep);
721 int shal_depth = depth(shallow);
722 assert(shal_depth <= depth(deep), "must be");
723 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
724 Node* pred = preds.current();
725 if (in_bb(pred) && !visited_test(pred)) {
726 if (shallow == pred) {
727 return false;
728 }
729 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
730 return false;
731 }
732 }
733 }
734 return true;
735 }
737 //------------------------------set_alignment---------------------------
738 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
739 set_alignment(s1, align);
740 if (align == top_align || align == bottom_align) {
741 set_alignment(s2, align);
742 } else {
743 set_alignment(s2, align + data_size(s1));
744 }
745 }
747 //------------------------------data_size---------------------------
748 int SuperWord::data_size(Node* s) {
749 int bsize = type2aelembytes(velt_basic_type(s));
750 assert(bsize != 0, "valid size");
751 return bsize;
752 }
754 //------------------------------extend_packlist---------------------------
755 // Extend packset by following use->def and def->use links from pack members.
756 void SuperWord::extend_packlist() {
757 bool changed;
758 do {
759 changed = false;
760 for (int i = 0; i < _packset.length(); i++) {
761 Node_List* p = _packset.at(i);
762 changed |= follow_use_defs(p);
763 changed |= follow_def_uses(p);
764 }
765 } while (changed);
767 #ifndef PRODUCT
768 if (TraceSuperWord) {
769 tty->print_cr("\nAfter extend_packlist");
770 print_packset();
771 }
772 #endif
773 }
775 //------------------------------follow_use_defs---------------------------
776 // Extend the packset by visiting operand definitions of nodes in pack p
777 bool SuperWord::follow_use_defs(Node_List* p) {
778 assert(p->size() == 2, "just checking");
779 Node* s1 = p->at(0);
780 Node* s2 = p->at(1);
781 assert(s1->req() == s2->req(), "just checking");
782 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
784 if (s1->is_Load()) return false;
786 int align = alignment(s1);
787 bool changed = false;
788 int start = s1->is_Store() ? MemNode::ValueIn : 1;
789 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
790 for (int j = start; j < end; j++) {
791 Node* t1 = s1->in(j);
792 Node* t2 = s2->in(j);
793 if (!in_bb(t1) || !in_bb(t2))
794 continue;
795 if (stmts_can_pack(t1, t2, align)) {
796 if (est_savings(t1, t2) >= 0) {
797 Node_List* pair = new Node_List();
798 pair->push(t1);
799 pair->push(t2);
800 _packset.append(pair);
801 set_alignment(t1, t2, align);
802 changed = true;
803 }
804 }
805 }
806 return changed;
807 }
809 //------------------------------follow_def_uses---------------------------
810 // Extend the packset by visiting uses of nodes in pack p
811 bool SuperWord::follow_def_uses(Node_List* p) {
812 bool changed = false;
813 Node* s1 = p->at(0);
814 Node* s2 = p->at(1);
815 assert(p->size() == 2, "just checking");
816 assert(s1->req() == s2->req(), "just checking");
817 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
819 if (s1->is_Store()) return false;
821 int align = alignment(s1);
822 int savings = -1;
823 Node* u1 = NULL;
824 Node* u2 = NULL;
825 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
826 Node* t1 = s1->fast_out(i);
827 if (!in_bb(t1)) continue;
828 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
829 Node* t2 = s2->fast_out(j);
830 if (!in_bb(t2)) continue;
831 if (!opnd_positions_match(s1, t1, s2, t2))
832 continue;
833 if (stmts_can_pack(t1, t2, align)) {
834 int my_savings = est_savings(t1, t2);
835 if (my_savings > savings) {
836 savings = my_savings;
837 u1 = t1;
838 u2 = t2;
839 }
840 }
841 }
842 }
843 if (savings >= 0) {
844 Node_List* pair = new Node_List();
845 pair->push(u1);
846 pair->push(u2);
847 _packset.append(pair);
848 set_alignment(u1, u2, align);
849 changed = true;
850 }
851 return changed;
852 }
854 //---------------------------opnd_positions_match-------------------------
855 // Is the use of d1 in u1 at the same operand position as d2 in u2?
856 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
857 uint ct = u1->req();
858 if (ct != u2->req()) return false;
859 uint i1 = 0;
860 uint i2 = 0;
861 do {
862 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
863 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
864 if (i1 != i2) {
865 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
866 // Further analysis relies on operands position matching.
867 u2->swap_edges(i1, i2);
868 } else {
869 return false;
870 }
871 }
872 } while (i1 < ct);
873 return true;
874 }
876 //------------------------------est_savings---------------------------
877 // Estimate the savings from executing s1 and s2 as a pack
878 int SuperWord::est_savings(Node* s1, Node* s2) {
879 int save_in = 2 - 1; // 2 operations per instruction in packed form
881 // inputs
882 for (uint i = 1; i < s1->req(); i++) {
883 Node* x1 = s1->in(i);
884 Node* x2 = s2->in(i);
885 if (x1 != x2) {
886 if (are_adjacent_refs(x1, x2)) {
887 save_in += adjacent_profit(x1, x2);
888 } else if (!in_packset(x1, x2)) {
889 save_in -= pack_cost(2);
890 } else {
891 save_in += unpack_cost(2);
892 }
893 }
894 }
896 // uses of result
897 uint ct = 0;
898 int save_use = 0;
899 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
900 Node* s1_use = s1->fast_out(i);
901 for (int j = 0; j < _packset.length(); j++) {
902 Node_List* p = _packset.at(j);
903 if (p->at(0) == s1_use) {
904 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
905 Node* s2_use = s2->fast_out(k);
906 if (p->at(p->size()-1) == s2_use) {
907 ct++;
908 if (are_adjacent_refs(s1_use, s2_use)) {
909 save_use += adjacent_profit(s1_use, s2_use);
910 }
911 }
912 }
913 }
914 }
915 }
917 if (ct < s1->outcnt()) save_use += unpack_cost(1);
918 if (ct < s2->outcnt()) save_use += unpack_cost(1);
920 return MAX2(save_in, save_use);
921 }
923 //------------------------------costs---------------------------
924 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
925 int SuperWord::pack_cost(int ct) { return ct; }
926 int SuperWord::unpack_cost(int ct) { return ct; }
928 //------------------------------combine_packs---------------------------
929 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
930 void SuperWord::combine_packs() {
931 bool changed = true;
932 // Combine packs regardless max vector size.
933 while (changed) {
934 changed = false;
935 for (int i = 0; i < _packset.length(); i++) {
936 Node_List* p1 = _packset.at(i);
937 if (p1 == NULL) continue;
938 for (int j = 0; j < _packset.length(); j++) {
939 Node_List* p2 = _packset.at(j);
940 if (p2 == NULL) continue;
941 if (i == j) continue;
942 if (p1->at(p1->size()-1) == p2->at(0)) {
943 for (uint k = 1; k < p2->size(); k++) {
944 p1->push(p2->at(k));
945 }
946 _packset.at_put(j, NULL);
947 changed = true;
948 }
949 }
950 }
951 }
953 // Split packs which have size greater then max vector size.
954 for (int i = 0; i < _packset.length(); i++) {
955 Node_List* p1 = _packset.at(i);
956 if (p1 != NULL) {
957 BasicType bt = velt_basic_type(p1->at(0));
958 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
959 assert(is_power_of_2(max_vlen), "sanity");
960 uint psize = p1->size();
961 if (!is_power_of_2(psize)) {
962 // Skip pack which can't be vector.
963 // case1: for(...) { a[i] = i; } elements values are different (i+x)
964 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
965 _packset.at_put(i, NULL);
966 continue;
967 }
968 if (psize > max_vlen) {
969 Node_List* pack = new Node_List();
970 for (uint j = 0; j < psize; j++) {
971 pack->push(p1->at(j));
972 if (pack->size() >= max_vlen) {
973 assert(is_power_of_2(pack->size()), "sanity");
974 _packset.append(pack);
975 pack = new Node_List();
976 }
977 }
978 _packset.at_put(i, NULL);
979 }
980 }
981 }
983 // Compress list.
984 for (int i = _packset.length() - 1; i >= 0; i--) {
985 Node_List* p1 = _packset.at(i);
986 if (p1 == NULL) {
987 _packset.remove_at(i);
988 }
989 }
991 #ifndef PRODUCT
992 if (TraceSuperWord) {
993 tty->print_cr("\nAfter combine_packs");
994 print_packset();
995 }
996 #endif
997 }
999 //-----------------------------construct_my_pack_map--------------------------
1000 // Construct the map from nodes to packs. Only valid after the
1001 // point where a node is only in one pack (after combine_packs).
1002 void SuperWord::construct_my_pack_map() {
1003 Node_List* rslt = NULL;
1004 for (int i = 0; i < _packset.length(); i++) {
1005 Node_List* p = _packset.at(i);
1006 for (uint j = 0; j < p->size(); j++) {
1007 Node* s = p->at(j);
1008 assert(my_pack(s) == NULL, "only in one pack");
1009 set_my_pack(s, p);
1010 }
1011 }
1012 }
1014 //------------------------------filter_packs---------------------------
1015 // Remove packs that are not implemented or not profitable.
1016 void SuperWord::filter_packs() {
1018 // Remove packs that are not implemented
1019 for (int i = _packset.length() - 1; i >= 0; i--) {
1020 Node_List* pk = _packset.at(i);
1021 bool impl = implemented(pk);
1022 if (!impl) {
1023 #ifndef PRODUCT
1024 if (TraceSuperWord && Verbose) {
1025 tty->print_cr("Unimplemented");
1026 pk->at(0)->dump();
1027 }
1028 #endif
1029 remove_pack_at(i);
1030 }
1031 }
1033 // Remove packs that are not profitable
1034 bool changed;
1035 do {
1036 changed = false;
1037 for (int i = _packset.length() - 1; i >= 0; i--) {
1038 Node_List* pk = _packset.at(i);
1039 bool prof = profitable(pk);
1040 if (!prof) {
1041 #ifndef PRODUCT
1042 if (TraceSuperWord && Verbose) {
1043 tty->print_cr("Unprofitable");
1044 pk->at(0)->dump();
1045 }
1046 #endif
1047 remove_pack_at(i);
1048 changed = true;
1049 }
1050 }
1051 } while (changed);
1053 #ifndef PRODUCT
1054 if (TraceSuperWord) {
1055 tty->print_cr("\nAfter filter_packs");
1056 print_packset();
1057 tty->cr();
1058 }
1059 #endif
1060 }
1062 //------------------------------implemented---------------------------
1063 // Can code be generated for pack p?
1064 bool SuperWord::implemented(Node_List* p) {
1065 Node* p0 = p->at(0);
1066 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1067 }
1069 //------------------------------same_inputs--------------------------
1070 // For pack p, are all idx operands the same?
1071 static bool same_inputs(Node_List* p, int idx) {
1072 Node* p0 = p->at(0);
1073 uint vlen = p->size();
1074 Node* p0_def = p0->in(idx);
1075 for (uint i = 1; i < vlen; i++) {
1076 Node* pi = p->at(i);
1077 Node* pi_def = pi->in(idx);
1078 if (p0_def != pi_def)
1079 return false;
1080 }
1081 return true;
1082 }
1084 //------------------------------profitable---------------------------
1085 // For pack p, are all operands and all uses (with in the block) vector?
1086 bool SuperWord::profitable(Node_List* p) {
1087 Node* p0 = p->at(0);
1088 uint start, end;
1089 VectorNode::vector_operands(p0, &start, &end);
1091 // Return false if some input is not vector and inside block
1092 for (uint i = start; i < end; i++) {
1093 if (!is_vector_use(p0, i)) {
1094 // For now, return false if not scalar promotion case (inputs are the same.)
1095 // Later, implement PackNode and allow differing, non-vector inputs
1096 // (maybe just the ones from outside the block.)
1097 if (!same_inputs(p, i)) {
1098 return false;
1099 }
1100 }
1101 }
1102 if (VectorNode::is_shift(p0)) {
1103 // For now, return false if shift count is vector because
1104 // hw does not support it.
1105 if (is_vector_use(p0, 2))
1106 return false;
1107 // For the same reason return false if different shift counts.
1108 if (!same_inputs(p, 2))
1109 return false;
1110 }
1111 if (!p0->is_Store()) {
1112 // For now, return false if not all uses are vector.
1113 // Later, implement ExtractNode and allow non-vector uses (maybe
1114 // just the ones outside the block.)
1115 for (uint i = 0; i < p->size(); i++) {
1116 Node* def = p->at(i);
1117 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1118 Node* use = def->fast_out(j);
1119 for (uint k = 0; k < use->req(); k++) {
1120 Node* n = use->in(k);
1121 if (def == n) {
1122 if (!is_vector_use(use, k)) {
1123 return false;
1124 }
1125 }
1126 }
1127 }
1128 }
1129 }
1130 return true;
1131 }
1133 //------------------------------schedule---------------------------
1134 // Adjust the memory graph for the packed operations
1135 void SuperWord::schedule() {
1137 // Co-locate in the memory graph the members of each memory pack
1138 for (int i = 0; i < _packset.length(); i++) {
1139 co_locate_pack(_packset.at(i));
1140 }
1141 }
1143 //-------------------------------remove_and_insert-------------------
1144 // Remove "current" from its current position in the memory graph and insert
1145 // it after the appropriate insertion point (lip or uip).
1146 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1147 Node *uip, Unique_Node_List &sched_before) {
1148 Node* my_mem = current->in(MemNode::Memory);
1149 bool sched_up = sched_before.member(current);
1151 // remove current_store from its current position in the memmory graph
1152 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1153 Node* use = current->out(i);
1154 if (use->is_Mem()) {
1155 assert(use->in(MemNode::Memory) == current, "must be");
1156 if (use == prev) { // connect prev to my_mem
1157 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1158 --i; //deleted this edge; rescan position
1159 } else if (sched_before.member(use)) {
1160 if (!sched_up) { // Will be moved together with current
1161 _igvn.replace_input_of(use, MemNode::Memory, uip);
1162 --i; //deleted this edge; rescan position
1163 }
1164 } else {
1165 if (sched_up) { // Will be moved together with current
1166 _igvn.replace_input_of(use, MemNode::Memory, lip);
1167 --i; //deleted this edge; rescan position
1168 }
1169 }
1170 }
1171 }
1173 Node *insert_pt = sched_up ? uip : lip;
1175 // all uses of insert_pt's memory state should use current's instead
1176 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1177 Node* use = insert_pt->out(i);
1178 if (use->is_Mem()) {
1179 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1180 _igvn.replace_input_of(use, MemNode::Memory, current);
1181 --i; //deleted this edge; rescan position
1182 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1183 uint pos; //lip (lower insert point) must be the last one in the memory slice
1184 for (pos=1; pos < use->req(); pos++) {
1185 if (use->in(pos) == insert_pt) break;
1186 }
1187 _igvn.replace_input_of(use, pos, current);
1188 --i;
1189 }
1190 }
1192 //connect current to insert_pt
1193 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1194 }
1196 //------------------------------co_locate_pack----------------------------------
1197 // To schedule a store pack, we need to move any sandwiched memory ops either before
1198 // or after the pack, based upon dependence information:
1199 // (1) If any store in the pack depends on the sandwiched memory op, the
1200 // sandwiched memory op must be scheduled BEFORE the pack;
1201 // (2) If a sandwiched memory op depends on any store in the pack, the
1202 // sandwiched memory op must be scheduled AFTER the pack;
1203 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1204 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1205 // scheduled before the pack, memB must also be scheduled before the pack;
1206 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1207 // schedule this store AFTER the pack
1208 // (5) We know there is no dependence cycle, so there in no other case;
1209 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1210 //
1211 // To schedule a load pack, we use the memory state of either the first or the last load in
1212 // the pack, based on the dependence constraint.
1213 void SuperWord::co_locate_pack(Node_List* pk) {
1214 if (pk->at(0)->is_Store()) {
1215 MemNode* first = executed_first(pk)->as_Mem();
1216 MemNode* last = executed_last(pk)->as_Mem();
1217 Unique_Node_List schedule_before_pack;
1218 Unique_Node_List memops;
1220 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1221 MemNode* previous = last;
1222 while (true) {
1223 assert(in_bb(current), "stay in block");
1224 memops.push(previous);
1225 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1226 Node* use = current->out(i);
1227 if (use->is_Mem() && use != previous)
1228 memops.push(use);
1229 }
1230 if (current == first) break;
1231 previous = current;
1232 current = current->in(MemNode::Memory)->as_Mem();
1233 }
1235 // determine which memory operations should be scheduled before the pack
1236 for (uint i = 1; i < memops.size(); i++) {
1237 Node *s1 = memops.at(i);
1238 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1239 for (uint j = 0; j< i; j++) {
1240 Node *s2 = memops.at(j);
1241 if (!independent(s1, s2)) {
1242 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1243 schedule_before_pack.push(s1); // s1 must be scheduled before
1244 Node_List* mem_pk = my_pack(s1);
1245 if (mem_pk != NULL) {
1246 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1247 Node* s = mem_pk->at(ii); // follow partner
1248 if (memops.member(s) && !schedule_before_pack.member(s))
1249 schedule_before_pack.push(s);
1250 }
1251 }
1252 break;
1253 }
1254 }
1255 }
1256 }
1257 }
1259 Node* upper_insert_pt = first->in(MemNode::Memory);
1260 // Following code moves loads connected to upper_insert_pt below aliased stores.
1261 // Collect such loads here and reconnect them back to upper_insert_pt later.
1262 memops.clear();
1263 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1264 Node* use = upper_insert_pt->out(i);
1265 if (!use->is_Store())
1266 memops.push(use);
1267 }
1269 MemNode* lower_insert_pt = last;
1270 previous = last; //previous store in pk
1271 current = last->in(MemNode::Memory)->as_Mem();
1273 // start scheduling from "last" to "first"
1274 while (true) {
1275 assert(in_bb(current), "stay in block");
1276 assert(in_pack(previous, pk), "previous stays in pack");
1277 Node* my_mem = current->in(MemNode::Memory);
1279 if (in_pack(current, pk)) {
1280 // Forward users of my memory state (except "previous) to my input memory state
1281 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1282 Node* use = current->out(i);
1283 if (use->is_Mem() && use != previous) {
1284 assert(use->in(MemNode::Memory) == current, "must be");
1285 if (schedule_before_pack.member(use)) {
1286 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1287 } else {
1288 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1289 }
1290 --i; // deleted this edge; rescan position
1291 }
1292 }
1293 previous = current;
1294 } else { // !in_pack(current, pk) ==> a sandwiched store
1295 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1296 }
1298 if (current == first) break;
1299 current = my_mem->as_Mem();
1300 } // end while
1302 // Reconnect loads back to upper_insert_pt.
1303 for (uint i = 0; i < memops.size(); i++) {
1304 Node *ld = memops.at(i);
1305 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1306 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1307 }
1308 }
1309 } else if (pk->at(0)->is_Load()) { //load
1310 // all loads in the pack should have the same memory state. By default,
1311 // we use the memory state of the last load. However, if any load could
1312 // not be moved down due to the dependence constraint, we use the memory
1313 // state of the first load.
1314 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1315 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1316 bool schedule_last = true;
1317 for (uint i = 0; i < pk->size(); i++) {
1318 Node* ld = pk->at(i);
1319 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1320 current=current->in(MemNode::Memory)) {
1321 assert(current != first_mem, "corrupted memory graph");
1322 if(current->is_Mem() && !independent(current, ld)){
1323 schedule_last = false; // a later store depends on this load
1324 break;
1325 }
1326 }
1327 }
1329 Node* mem_input = schedule_last ? last_mem : first_mem;
1330 _igvn.hash_delete(mem_input);
1331 // Give each load the same memory state
1332 for (uint i = 0; i < pk->size(); i++) {
1333 LoadNode* ld = pk->at(i)->as_Load();
1334 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1335 }
1336 }
1337 }
1339 //------------------------------output---------------------------
1340 // Convert packs into vector node operations
1341 void SuperWord::output() {
1342 if (_packset.length() == 0) return;
1344 #ifndef PRODUCT
1345 if (TraceLoopOpts) {
1346 tty->print("SuperWord ");
1347 lpt()->dump_head();
1348 }
1349 #endif
1351 // MUST ENSURE main loop's initial value is properly aligned:
1352 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1354 align_initial_loop_index(align_to_ref());
1356 // Insert extract (unpack) operations for scalar uses
1357 for (int i = 0; i < _packset.length(); i++) {
1358 insert_extracts(_packset.at(i));
1359 }
1361 Compile* C = _phase->C;
1362 uint max_vlen_in_bytes = 0;
1363 for (int i = 0; i < _block.length(); i++) {
1364 Node* n = _block.at(i);
1365 Node_List* p = my_pack(n);
1366 if (p && n == executed_last(p)) {
1367 uint vlen = p->size();
1368 uint vlen_in_bytes = 0;
1369 Node* vn = NULL;
1370 Node* low_adr = p->at(0);
1371 Node* first = executed_first(p);
1372 int opc = n->Opcode();
1373 if (n->is_Load()) {
1374 Node* ctl = n->in(MemNode::Control);
1375 Node* mem = first->in(MemNode::Memory);
1376 Node* adr = low_adr->in(MemNode::Address);
1377 const TypePtr* atyp = n->adr_type();
1378 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1379 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1380 } else if (n->is_Store()) {
1381 // Promote value to be stored to vector
1382 Node* val = vector_opd(p, MemNode::ValueIn);
1383 Node* ctl = n->in(MemNode::Control);
1384 Node* mem = first->in(MemNode::Memory);
1385 Node* adr = low_adr->in(MemNode::Address);
1386 const TypePtr* atyp = n->adr_type();
1387 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1388 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1389 } else if (n->req() == 3) {
1390 // Promote operands to vector
1391 Node* in1 = vector_opd(p, 1);
1392 Node* in2 = vector_opd(p, 2);
1393 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1394 // Move invariant vector input into second position to avoid register spilling.
1395 Node* tmp = in1;
1396 in1 = in2;
1397 in2 = tmp;
1398 }
1399 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1400 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1401 } else {
1402 ShouldNotReachHere();
1403 }
1404 assert(vn != NULL, "sanity");
1405 _phase->_igvn.register_new_node_with_optimizer(vn);
1406 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1407 for (uint j = 0; j < p->size(); j++) {
1408 Node* pm = p->at(j);
1409 _igvn.replace_node(pm, vn);
1410 }
1411 _igvn._worklist.push(vn);
1413 if (vlen_in_bytes > max_vlen_in_bytes) {
1414 max_vlen_in_bytes = vlen_in_bytes;
1415 }
1416 #ifdef ASSERT
1417 if (TraceNewVectors) {
1418 tty->print("new Vector node: ");
1419 vn->dump();
1420 }
1421 #endif
1422 }
1423 }
1424 C->set_max_vector_size(max_vlen_in_bytes);
1425 }
1427 //------------------------------vector_opd---------------------------
1428 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1429 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1430 Node* p0 = p->at(0);
1431 uint vlen = p->size();
1432 Node* opd = p0->in(opd_idx);
1434 if (same_inputs(p, opd_idx)) {
1435 if (opd->is_Vector() || opd->is_LoadVector()) {
1436 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1437 return opd; // input is matching vector
1438 }
1439 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1440 // No vector is needed for shift count.
1441 // Vector instructions do not mask shift count, do it here.
1442 Compile* C = _phase->C;
1443 Node* cnt = opd;
1444 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1445 const TypeInt* t = opd->find_int_type();
1446 if (t != NULL && t->is_con()) {
1447 juint shift = t->get_con();
1448 if (shift > mask) { // Unsigned cmp
1449 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1450 }
1451 } else {
1452 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1453 cnt = ConNode::make(C, TypeInt::make(mask));
1454 _phase->_igvn.register_new_node_with_optimizer(cnt);
1455 cnt = new (C, 3) AndINode(opd, cnt);
1456 _phase->_igvn.register_new_node_with_optimizer(cnt);
1457 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1458 }
1459 assert(opd->bottom_type()->isa_int(), "int type only");
1460 // Move non constant shift count into XMM register.
1461 cnt = new (C, 2) MoveI2FNode(cnt);
1462 }
1463 if (cnt != opd) {
1464 _phase->_igvn.register_new_node_with_optimizer(cnt);
1465 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1466 }
1467 return cnt;
1468 }
1469 assert(!opd->is_StoreVector(), "such vector is not expected here");
1470 // Convert scalar input to vector with the same number of elements as
1471 // p0's vector. Use p0's type because size of operand's container in
1472 // vector should match p0's size regardless operand's size.
1473 const Type* p0_t = velt_type(p0);
1474 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1476 _phase->_igvn.register_new_node_with_optimizer(vn);
1477 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1478 #ifdef ASSERT
1479 if (TraceNewVectors) {
1480 tty->print("new Vector node: ");
1481 vn->dump();
1482 }
1483 #endif
1484 return vn;
1485 }
1487 // Insert pack operation
1488 BasicType bt = velt_basic_type(p0);
1489 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1490 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1492 for (uint i = 1; i < vlen; i++) {
1493 Node* pi = p->at(i);
1494 Node* in = pi->in(opd_idx);
1495 assert(my_pack(in) == NULL, "Should already have been unpacked");
1496 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1497 pk->add_opd(in);
1498 }
1499 _phase->_igvn.register_new_node_with_optimizer(pk);
1500 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1501 #ifdef ASSERT
1502 if (TraceNewVectors) {
1503 tty->print("new Vector node: ");
1504 pk->dump();
1505 }
1506 #endif
1507 return pk;
1508 }
1510 //------------------------------insert_extracts---------------------------
1511 // If a use of pack p is not a vector use, then replace the
1512 // use with an extract operation.
1513 void SuperWord::insert_extracts(Node_List* p) {
1514 if (p->at(0)->is_Store()) return;
1515 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1517 // Inspect each use of each pack member. For each use that is
1518 // not a vector use, replace the use with an extract operation.
1520 for (uint i = 0; i < p->size(); i++) {
1521 Node* def = p->at(i);
1522 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1523 Node* use = def->fast_out(j);
1524 for (uint k = 0; k < use->req(); k++) {
1525 Node* n = use->in(k);
1526 if (def == n) {
1527 if (!is_vector_use(use, k)) {
1528 _n_idx_list.push(use, k);
1529 }
1530 }
1531 }
1532 }
1533 }
1535 while (_n_idx_list.is_nonempty()) {
1536 Node* use = _n_idx_list.node();
1537 int idx = _n_idx_list.index();
1538 _n_idx_list.pop();
1539 Node* def = use->in(idx);
1541 // Insert extract operation
1542 _igvn.hash_delete(def);
1543 int def_pos = alignment(def) / data_size(def);
1545 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1546 _phase->_igvn.register_new_node_with_optimizer(ex);
1547 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1548 _igvn.replace_input_of(use, idx, ex);
1549 _igvn._worklist.push(def);
1551 bb_insert_after(ex, bb_idx(def));
1552 set_velt_type(ex, velt_type(def));
1553 }
1554 }
1556 //------------------------------is_vector_use---------------------------
1557 // Is use->in(u_idx) a vector use?
1558 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1559 Node_List* u_pk = my_pack(use);
1560 if (u_pk == NULL) return false;
1561 Node* def = use->in(u_idx);
1562 Node_List* d_pk = my_pack(def);
1563 if (d_pk == NULL) {
1564 // check for scalar promotion
1565 Node* n = u_pk->at(0)->in(u_idx);
1566 for (uint i = 1; i < u_pk->size(); i++) {
1567 if (u_pk->at(i)->in(u_idx) != n) return false;
1568 }
1569 return true;
1570 }
1571 if (u_pk->size() != d_pk->size())
1572 return false;
1573 for (uint i = 0; i < u_pk->size(); i++) {
1574 Node* ui = u_pk->at(i);
1575 Node* di = d_pk->at(i);
1576 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1577 return false;
1578 }
1579 return true;
1580 }
1582 //------------------------------construct_bb---------------------------
1583 // Construct reverse postorder list of block members
1584 void SuperWord::construct_bb() {
1585 Node* entry = bb();
1587 assert(_stk.length() == 0, "stk is empty");
1588 assert(_block.length() == 0, "block is empty");
1589 assert(_data_entry.length() == 0, "data_entry is empty");
1590 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1591 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1593 // Find non-control nodes with no inputs from within block,
1594 // create a temporary map from node _idx to bb_idx for use
1595 // by the visited and post_visited sets,
1596 // and count number of nodes in block.
1597 int bb_ct = 0;
1598 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1599 Node *n = lpt()->_body.at(i);
1600 set_bb_idx(n, i); // Create a temporary map
1601 if (in_bb(n)) {
1602 bb_ct++;
1603 if (!n->is_CFG()) {
1604 bool found = false;
1605 for (uint j = 0; j < n->req(); j++) {
1606 Node* def = n->in(j);
1607 if (def && in_bb(def)) {
1608 found = true;
1609 break;
1610 }
1611 }
1612 if (!found) {
1613 assert(n != entry, "can't be entry");
1614 _data_entry.push(n);
1615 }
1616 }
1617 }
1618 }
1620 // Find memory slices (head and tail)
1621 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1622 Node *n = lp()->fast_out(i);
1623 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1624 Node* n_tail = n->in(LoopNode::LoopBackControl);
1625 if (n_tail != n->in(LoopNode::EntryControl)) {
1626 _mem_slice_head.push(n);
1627 _mem_slice_tail.push(n_tail);
1628 }
1629 }
1630 }
1632 // Create an RPO list of nodes in block
1634 visited_clear();
1635 post_visited_clear();
1637 // Push all non-control nodes with no inputs from within block, then control entry
1638 for (int j = 0; j < _data_entry.length(); j++) {
1639 Node* n = _data_entry.at(j);
1640 visited_set(n);
1641 _stk.push(n);
1642 }
1643 visited_set(entry);
1644 _stk.push(entry);
1646 // Do a depth first walk over out edges
1647 int rpo_idx = bb_ct - 1;
1648 int size;
1649 while ((size = _stk.length()) > 0) {
1650 Node* n = _stk.top(); // Leave node on stack
1651 if (!visited_test_set(n)) {
1652 // forward arc in graph
1653 } else if (!post_visited_test(n)) {
1654 // cross or back arc
1655 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1656 Node *use = n->fast_out(i);
1657 if (in_bb(use) && !visited_test(use) &&
1658 // Don't go around backedge
1659 (!use->is_Phi() || n == entry)) {
1660 _stk.push(use);
1661 }
1662 }
1663 if (_stk.length() == size) {
1664 // There were no additional uses, post visit node now
1665 _stk.pop(); // Remove node from stack
1666 assert(rpo_idx >= 0, "");
1667 _block.at_put_grow(rpo_idx, n);
1668 rpo_idx--;
1669 post_visited_set(n);
1670 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1671 }
1672 } else {
1673 _stk.pop(); // Remove post-visited node from stack
1674 }
1675 }
1677 // Create real map of block indices for nodes
1678 for (int j = 0; j < _block.length(); j++) {
1679 Node* n = _block.at(j);
1680 set_bb_idx(n, j);
1681 }
1683 initialize_bb(); // Ensure extra info is allocated.
1685 #ifndef PRODUCT
1686 if (TraceSuperWord) {
1687 print_bb();
1688 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1689 for (int m = 0; m < _data_entry.length(); m++) {
1690 tty->print("%3d ", m);
1691 _data_entry.at(m)->dump();
1692 }
1693 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1694 for (int m = 0; m < _mem_slice_head.length(); m++) {
1695 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1696 tty->print(" "); _mem_slice_tail.at(m)->dump();
1697 }
1698 }
1699 #endif
1700 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1701 }
1703 //------------------------------initialize_bb---------------------------
1704 // Initialize per node info
1705 void SuperWord::initialize_bb() {
1706 Node* last = _block.at(_block.length() - 1);
1707 grow_node_info(bb_idx(last));
1708 }
1710 //------------------------------bb_insert_after---------------------------
1711 // Insert n into block after pos
1712 void SuperWord::bb_insert_after(Node* n, int pos) {
1713 int n_pos = pos + 1;
1714 // Make room
1715 for (int i = _block.length() - 1; i >= n_pos; i--) {
1716 _block.at_put_grow(i+1, _block.at(i));
1717 }
1718 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1719 _node_info.at_put_grow(j+1, _node_info.at(j));
1720 }
1721 // Set value
1722 _block.at_put_grow(n_pos, n);
1723 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1724 // Adjust map from node->_idx to _block index
1725 for (int i = n_pos; i < _block.length(); i++) {
1726 set_bb_idx(_block.at(i), i);
1727 }
1728 }
1730 //------------------------------compute_max_depth---------------------------
1731 // Compute max depth for expressions from beginning of block
1732 // Use to prune search paths during test for independence.
1733 void SuperWord::compute_max_depth() {
1734 int ct = 0;
1735 bool again;
1736 do {
1737 again = false;
1738 for (int i = 0; i < _block.length(); i++) {
1739 Node* n = _block.at(i);
1740 if (!n->is_Phi()) {
1741 int d_orig = depth(n);
1742 int d_in = 0;
1743 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1744 Node* pred = preds.current();
1745 if (in_bb(pred)) {
1746 d_in = MAX2(d_in, depth(pred));
1747 }
1748 }
1749 if (d_in + 1 != d_orig) {
1750 set_depth(n, d_in + 1);
1751 again = true;
1752 }
1753 }
1754 }
1755 ct++;
1756 } while (again);
1757 #ifndef PRODUCT
1758 if (TraceSuperWord && Verbose)
1759 tty->print_cr("compute_max_depth iterated: %d times", ct);
1760 #endif
1761 }
1763 //-------------------------compute_vector_element_type-----------------------
1764 // Compute necessary vector element type for expressions
1765 // This propagates backwards a narrower integer type when the
1766 // upper bits of the value are not needed.
1767 // Example: char a,b,c; a = b + c;
1768 // Normally the type of the add is integer, but for packed character
1769 // operations the type of the add needs to be char.
1770 void SuperWord::compute_vector_element_type() {
1771 #ifndef PRODUCT
1772 if (TraceSuperWord && Verbose)
1773 tty->print_cr("\ncompute_velt_type:");
1774 #endif
1776 // Initial type
1777 for (int i = 0; i < _block.length(); i++) {
1778 Node* n = _block.at(i);
1779 set_velt_type(n, container_type(n));
1780 }
1782 // Propagate narrowed type backwards through operations
1783 // that don't depend on higher order bits
1784 for (int i = _block.length() - 1; i >= 0; i--) {
1785 Node* n = _block.at(i);
1786 // Only integer types need be examined
1787 const Type* vt = velt_type(n);
1788 if (vt->basic_type() == T_INT) {
1789 uint start, end;
1790 VectorNode::vector_operands(n, &start, &end);
1791 const Type* vt = velt_type(n);
1793 for (uint j = start; j < end; j++) {
1794 Node* in = n->in(j);
1795 // Don't propagate through a memory
1796 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1797 data_size(n) < data_size(in)) {
1798 bool same_type = true;
1799 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1800 Node *use = in->fast_out(k);
1801 if (!in_bb(use) || !same_velt_type(use, n)) {
1802 same_type = false;
1803 break;
1804 }
1805 }
1806 if (same_type) {
1807 set_velt_type(in, vt);
1808 }
1809 }
1810 }
1811 }
1812 }
1813 #ifndef PRODUCT
1814 if (TraceSuperWord && Verbose) {
1815 for (int i = 0; i < _block.length(); i++) {
1816 Node* n = _block.at(i);
1817 velt_type(n)->dump();
1818 tty->print("\t");
1819 n->dump();
1820 }
1821 }
1822 #endif
1823 }
1825 //------------------------------memory_alignment---------------------------
1826 // Alignment within a vector memory reference
1827 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1828 SWPointer p(s, this);
1829 if (!p.valid()) {
1830 return bottom_align;
1831 }
1832 int vw = vector_width_in_bytes(s);
1833 if (vw < 2) {
1834 return bottom_align; // No vectors for this type
1835 }
1836 int offset = p.offset_in_bytes();
1837 offset += iv_adjust*p.memory_size();
1838 int off_rem = offset % vw;
1839 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1840 return off_mod;
1841 }
1843 //---------------------------container_type---------------------------
1844 // Smallest type containing range of values
1845 const Type* SuperWord::container_type(Node* n) {
1846 if (n->is_Mem()) {
1847 return Type::get_const_basic_type(n->as_Mem()->memory_type());
1848 }
1849 const Type* t = _igvn.type(n);
1850 if (t->basic_type() == T_INT) {
1851 // A narrow type of arithmetic operations will be determined by
1852 // propagating the type of memory operations.
1853 return TypeInt::INT;
1854 }
1855 return t;
1856 }
1858 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1859 const Type* vt1 = velt_type(n1);
1860 const Type* vt2 = velt_type(n2);
1861 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1862 // Compare vectors element sizes for integer types.
1863 return data_size(n1) == data_size(n2);
1864 }
1865 return vt1 == vt2;
1866 }
1868 //------------------------------in_packset---------------------------
1869 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1870 bool SuperWord::in_packset(Node* s1, Node* s2) {
1871 for (int i = 0; i < _packset.length(); i++) {
1872 Node_List* p = _packset.at(i);
1873 assert(p->size() == 2, "must be");
1874 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1875 return true;
1876 }
1877 }
1878 return false;
1879 }
1881 //------------------------------in_pack---------------------------
1882 // Is s in pack p?
1883 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1884 for (uint i = 0; i < p->size(); i++) {
1885 if (p->at(i) == s) {
1886 return p;
1887 }
1888 }
1889 return NULL;
1890 }
1892 //------------------------------remove_pack_at---------------------------
1893 // Remove the pack at position pos in the packset
1894 void SuperWord::remove_pack_at(int pos) {
1895 Node_List* p = _packset.at(pos);
1896 for (uint i = 0; i < p->size(); i++) {
1897 Node* s = p->at(i);
1898 set_my_pack(s, NULL);
1899 }
1900 _packset.remove_at(pos);
1901 }
1903 //------------------------------executed_first---------------------------
1904 // Return the node executed first in pack p. Uses the RPO block list
1905 // to determine order.
1906 Node* SuperWord::executed_first(Node_List* p) {
1907 Node* n = p->at(0);
1908 int n_rpo = bb_idx(n);
1909 for (uint i = 1; i < p->size(); i++) {
1910 Node* s = p->at(i);
1911 int s_rpo = bb_idx(s);
1912 if (s_rpo < n_rpo) {
1913 n = s;
1914 n_rpo = s_rpo;
1915 }
1916 }
1917 return n;
1918 }
1920 //------------------------------executed_last---------------------------
1921 // Return the node executed last in pack p.
1922 Node* SuperWord::executed_last(Node_List* p) {
1923 Node* n = p->at(0);
1924 int n_rpo = bb_idx(n);
1925 for (uint i = 1; i < p->size(); i++) {
1926 Node* s = p->at(i);
1927 int s_rpo = bb_idx(s);
1928 if (s_rpo > n_rpo) {
1929 n = s;
1930 n_rpo = s_rpo;
1931 }
1932 }
1933 return n;
1934 }
1936 //----------------------------align_initial_loop_index---------------------------
1937 // Adjust pre-loop limit so that in main loop, a load/store reference
1938 // to align_to_ref will be a position zero in the vector.
1939 // (iv + k) mod vector_align == 0
1940 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1941 CountedLoopNode *main_head = lp()->as_CountedLoop();
1942 assert(main_head->is_main_loop(), "");
1943 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1944 assert(pre_end != NULL, "");
1945 Node *pre_opaq1 = pre_end->limit();
1946 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1947 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1948 Node *lim0 = pre_opaq->in(1);
1950 // Where we put new limit calculations
1951 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1953 // Ensure the original loop limit is available from the
1954 // pre-loop Opaque1 node.
1955 Node *orig_limit = pre_opaq->original_loop_limit();
1956 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1958 SWPointer align_to_ref_p(align_to_ref, this);
1959 assert(align_to_ref_p.valid(), "sanity");
1961 // Given:
1962 // lim0 == original pre loop limit
1963 // V == v_align (power of 2)
1964 // invar == extra invariant piece of the address expression
1965 // e == offset [ +/- invar ]
1966 //
1967 // When reassociating expressions involving '%' the basic rules are:
1968 // (a - b) % k == 0 => a % k == b % k
1969 // and:
1970 // (a + b) % k == 0 => a % k == (k - b) % k
1971 //
1972 // For stride > 0 && scale > 0,
1973 // Derive the new pre-loop limit "lim" such that the two constraints:
1974 // (1) lim = lim0 + N (where N is some positive integer < V)
1975 // (2) (e + lim) % V == 0
1976 // are true.
1977 //
1978 // Substituting (1) into (2),
1979 // (e + lim0 + N) % V == 0
1980 // solve for N:
1981 // N = (V - (e + lim0)) % V
1982 // substitute back into (1), so that new limit
1983 // lim = lim0 + (V - (e + lim0)) % V
1984 //
1985 // For stride > 0 && scale < 0
1986 // Constraints:
1987 // lim = lim0 + N
1988 // (e - lim) % V == 0
1989 // Solving for lim:
1990 // (e - lim0 - N) % V == 0
1991 // N = (e - lim0) % V
1992 // lim = lim0 + (e - lim0) % V
1993 //
1994 // For stride < 0 && scale > 0
1995 // Constraints:
1996 // lim = lim0 - N
1997 // (e + lim) % V == 0
1998 // Solving for lim:
1999 // (e + lim0 - N) % V == 0
2000 // N = (e + lim0) % V
2001 // lim = lim0 - (e + lim0) % V
2002 //
2003 // For stride < 0 && scale < 0
2004 // Constraints:
2005 // lim = lim0 - N
2006 // (e - lim) % V == 0
2007 // Solving for lim:
2008 // (e - lim0 + N) % V == 0
2009 // N = (V - (e - lim0)) % V
2010 // lim = lim0 - (V - (e - lim0)) % V
2012 int vw = vector_width_in_bytes(align_to_ref);
2013 int stride = iv_stride();
2014 int scale = align_to_ref_p.scale_in_bytes();
2015 int elt_size = align_to_ref_p.memory_size();
2016 int v_align = vw / elt_size;
2017 assert(v_align > 1, "sanity");
2018 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2019 Node *offsn = _igvn.intcon(offset);
2021 Node *e = offsn;
2022 if (align_to_ref_p.invar() != NULL) {
2023 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2024 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2025 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
2026 _phase->_igvn.register_new_node_with_optimizer(aref);
2027 _phase->set_ctrl(aref, pre_ctrl);
2028 if (align_to_ref_p.negate_invar()) {
2029 e = new (_phase->C, 3) SubINode(e, aref);
2030 } else {
2031 e = new (_phase->C, 3) AddINode(e, aref);
2032 }
2033 _phase->_igvn.register_new_node_with_optimizer(e);
2034 _phase->set_ctrl(e, pre_ctrl);
2035 }
2036 if (vw > ObjectAlignmentInBytes) {
2037 // incorporate base e +/- base && Mask >>> log2(elt)
2038 Node* xbase = new(_phase->C, 2) CastP2XNode(NULL, align_to_ref_p.base());
2039 _phase->_igvn.register_new_node_with_optimizer(xbase);
2040 #ifdef _LP64
2041 xbase = new (_phase->C, 2) ConvL2INode(xbase);
2042 _phase->_igvn.register_new_node_with_optimizer(xbase);
2043 #endif
2044 Node* mask = _igvn.intcon(vw-1);
2045 Node* masked_xbase = new (_phase->C, 3) AndINode(xbase, mask);
2046 _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
2047 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2048 Node* bref = new (_phase->C, 3) URShiftINode(masked_xbase, log2_elt);
2049 _phase->_igvn.register_new_node_with_optimizer(bref);
2050 _phase->set_ctrl(bref, pre_ctrl);
2051 e = new (_phase->C, 3) AddINode(e, bref);
2052 _phase->_igvn.register_new_node_with_optimizer(e);
2053 _phase->set_ctrl(e, pre_ctrl);
2054 }
2056 // compute e +/- lim0
2057 if (scale < 0) {
2058 e = new (_phase->C, 3) SubINode(e, lim0);
2059 } else {
2060 e = new (_phase->C, 3) AddINode(e, lim0);
2061 }
2062 _phase->_igvn.register_new_node_with_optimizer(e);
2063 _phase->set_ctrl(e, pre_ctrl);
2065 if (stride * scale > 0) {
2066 // compute V - (e +/- lim0)
2067 Node* va = _igvn.intcon(v_align);
2068 e = new (_phase->C, 3) SubINode(va, e);
2069 _phase->_igvn.register_new_node_with_optimizer(e);
2070 _phase->set_ctrl(e, pre_ctrl);
2071 }
2072 // compute N = (exp) % V
2073 Node* va_msk = _igvn.intcon(v_align - 1);
2074 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
2075 _phase->_igvn.register_new_node_with_optimizer(N);
2076 _phase->set_ctrl(N, pre_ctrl);
2078 // substitute back into (1), so that new limit
2079 // lim = lim0 + N
2080 Node* lim;
2081 if (stride < 0) {
2082 lim = new (_phase->C, 3) SubINode(lim0, N);
2083 } else {
2084 lim = new (_phase->C, 3) AddINode(lim0, N);
2085 }
2086 _phase->_igvn.register_new_node_with_optimizer(lim);
2087 _phase->set_ctrl(lim, pre_ctrl);
2088 Node* constrained =
2089 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
2090 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
2091 _phase->_igvn.register_new_node_with_optimizer(constrained);
2092 _phase->set_ctrl(constrained, pre_ctrl);
2093 _igvn.hash_delete(pre_opaq);
2094 pre_opaq->set_req(1, constrained);
2095 }
2097 //----------------------------get_pre_loop_end---------------------------
2098 // Find pre loop end from main loop. Returns null if none.
2099 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2100 Node *ctrl = cl->in(LoopNode::EntryControl);
2101 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2102 Node *iffm = ctrl->in(0);
2103 if (!iffm->is_If()) return NULL;
2104 Node *p_f = iffm->in(0);
2105 if (!p_f->is_IfFalse()) return NULL;
2106 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2107 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2108 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2109 return pre_end;
2110 }
2113 //------------------------------init---------------------------
2114 void SuperWord::init() {
2115 _dg.init();
2116 _packset.clear();
2117 _disjoint_ptrs.clear();
2118 _block.clear();
2119 _data_entry.clear();
2120 _mem_slice_head.clear();
2121 _mem_slice_tail.clear();
2122 _node_info.clear();
2123 _align_to_ref = NULL;
2124 _lpt = NULL;
2125 _lp = NULL;
2126 _bb = NULL;
2127 _iv = NULL;
2128 }
2130 //------------------------------print_packset---------------------------
2131 void SuperWord::print_packset() {
2132 #ifndef PRODUCT
2133 tty->print_cr("packset");
2134 for (int i = 0; i < _packset.length(); i++) {
2135 tty->print_cr("Pack: %d", i);
2136 Node_List* p = _packset.at(i);
2137 print_pack(p);
2138 }
2139 #endif
2140 }
2142 //------------------------------print_pack---------------------------
2143 void SuperWord::print_pack(Node_List* p) {
2144 for (uint i = 0; i < p->size(); i++) {
2145 print_stmt(p->at(i));
2146 }
2147 }
2149 //------------------------------print_bb---------------------------
2150 void SuperWord::print_bb() {
2151 #ifndef PRODUCT
2152 tty->print_cr("\nBlock");
2153 for (int i = 0; i < _block.length(); i++) {
2154 Node* n = _block.at(i);
2155 tty->print("%d ", i);
2156 if (n) {
2157 n->dump();
2158 }
2159 }
2160 #endif
2161 }
2163 //------------------------------print_stmt---------------------------
2164 void SuperWord::print_stmt(Node* s) {
2165 #ifndef PRODUCT
2166 tty->print(" align: %d \t", alignment(s));
2167 s->dump();
2168 #endif
2169 }
2171 //------------------------------blank---------------------------
2172 char* SuperWord::blank(uint depth) {
2173 static char blanks[101];
2174 assert(depth < 101, "too deep");
2175 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2176 blanks[depth] = '\0';
2177 return blanks;
2178 }
2181 //==============================SWPointer===========================
2183 //----------------------------SWPointer------------------------
2184 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2185 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2186 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2188 Node* adr = mem->in(MemNode::Address);
2189 if (!adr->is_AddP()) {
2190 assert(!valid(), "too complex");
2191 return;
2192 }
2193 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2194 Node* base = adr->in(AddPNode::Base);
2195 //unsafe reference could not be aligned appropriately without runtime checking
2196 if (base == NULL || base->bottom_type() == Type::TOP) {
2197 assert(!valid(), "unsafe access");
2198 return;
2199 }
2200 for (int i = 0; i < 3; i++) {
2201 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2202 assert(!valid(), "too complex");
2203 return;
2204 }
2205 adr = adr->in(AddPNode::Address);
2206 if (base == adr || !adr->is_AddP()) {
2207 break; // stop looking at addp's
2208 }
2209 }
2210 _base = base;
2211 _adr = adr;
2212 assert(valid(), "Usable");
2213 }
2215 // Following is used to create a temporary object during
2216 // the pattern match of an address expression.
2217 SWPointer::SWPointer(SWPointer* p) :
2218 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2219 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2221 //------------------------scaled_iv_plus_offset--------------------
2222 // Match: k*iv + offset
2223 // where: k is a constant that maybe zero, and
2224 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2225 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2226 if (scaled_iv(n)) {
2227 return true;
2228 }
2229 if (offset_plus_k(n)) {
2230 return true;
2231 }
2232 int opc = n->Opcode();
2233 if (opc == Op_AddI) {
2234 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2235 return true;
2236 }
2237 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2238 return true;
2239 }
2240 } else if (opc == Op_SubI) {
2241 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2242 return true;
2243 }
2244 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2245 _scale *= -1;
2246 return true;
2247 }
2248 }
2249 return false;
2250 }
2252 //----------------------------scaled_iv------------------------
2253 // Match: k*iv where k is a constant that's not zero
2254 bool SWPointer::scaled_iv(Node* n) {
2255 if (_scale != 0) {
2256 return false; // already found a scale
2257 }
2258 if (n == iv()) {
2259 _scale = 1;
2260 return true;
2261 }
2262 int opc = n->Opcode();
2263 if (opc == Op_MulI) {
2264 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2265 _scale = n->in(2)->get_int();
2266 return true;
2267 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2268 _scale = n->in(1)->get_int();
2269 return true;
2270 }
2271 } else if (opc == Op_LShiftI) {
2272 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2273 _scale = 1 << n->in(2)->get_int();
2274 return true;
2275 }
2276 } else if (opc == Op_ConvI2L) {
2277 if (scaled_iv_plus_offset(n->in(1))) {
2278 return true;
2279 }
2280 } else if (opc == Op_LShiftL) {
2281 if (!has_iv() && _invar == NULL) {
2282 // Need to preserve the current _offset value, so
2283 // create a temporary object for this expression subtree.
2284 // Hacky, so should re-engineer the address pattern match.
2285 SWPointer tmp(this);
2286 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2287 if (tmp._invar == NULL) {
2288 int mult = 1 << n->in(2)->get_int();
2289 _scale = tmp._scale * mult;
2290 _offset += tmp._offset * mult;
2291 return true;
2292 }
2293 }
2294 }
2295 }
2296 return false;
2297 }
2299 //----------------------------offset_plus_k------------------------
2300 // Match: offset is (k [+/- invariant])
2301 // where k maybe zero and invariant is optional, but not both.
2302 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2303 int opc = n->Opcode();
2304 if (opc == Op_ConI) {
2305 _offset += negate ? -(n->get_int()) : n->get_int();
2306 return true;
2307 } else if (opc == Op_ConL) {
2308 // Okay if value fits into an int
2309 const TypeLong* t = n->find_long_type();
2310 if (t->higher_equal(TypeLong::INT)) {
2311 jlong loff = n->get_long();
2312 jint off = (jint)loff;
2313 _offset += negate ? -off : loff;
2314 return true;
2315 }
2316 return false;
2317 }
2318 if (_invar != NULL) return false; // already have an invariant
2319 if (opc == Op_AddI) {
2320 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2321 _negate_invar = negate;
2322 _invar = n->in(1);
2323 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2324 return true;
2325 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2326 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2327 _negate_invar = negate;
2328 _invar = n->in(2);
2329 return true;
2330 }
2331 }
2332 if (opc == Op_SubI) {
2333 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2334 _negate_invar = negate;
2335 _invar = n->in(1);
2336 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2337 return true;
2338 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2339 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2340 _negate_invar = !negate;
2341 _invar = n->in(2);
2342 return true;
2343 }
2344 }
2345 if (invariant(n)) {
2346 _negate_invar = negate;
2347 _invar = n;
2348 return true;
2349 }
2350 return false;
2351 }
2353 //----------------------------print------------------------
2354 void SWPointer::print() {
2355 #ifndef PRODUCT
2356 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2357 _base != NULL ? _base->_idx : 0,
2358 _adr != NULL ? _adr->_idx : 0,
2359 _scale, _offset,
2360 _negate_invar?'-':'+',
2361 _invar != NULL ? _invar->_idx : 0);
2362 #endif
2363 }
2365 // ========================= OrderedPair =====================
2367 const OrderedPair OrderedPair::initial;
2369 // ========================= SWNodeInfo =====================
2371 const SWNodeInfo SWNodeInfo::initial;
2374 // ============================ DepGraph ===========================
2376 //------------------------------make_node---------------------------
2377 // Make a new dependence graph node for an ideal node.
2378 DepMem* DepGraph::make_node(Node* node) {
2379 DepMem* m = new (_arena) DepMem(node);
2380 if (node != NULL) {
2381 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2382 _map.at_put_grow(node->_idx, m);
2383 }
2384 return m;
2385 }
2387 //------------------------------make_edge---------------------------
2388 // Make a new dependence graph edge from dpred -> dsucc
2389 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2390 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2391 dpred->set_out_head(e);
2392 dsucc->set_in_head(e);
2393 return e;
2394 }
2396 // ========================== DepMem ========================
2398 //------------------------------in_cnt---------------------------
2399 int DepMem::in_cnt() {
2400 int ct = 0;
2401 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2402 return ct;
2403 }
2405 //------------------------------out_cnt---------------------------
2406 int DepMem::out_cnt() {
2407 int ct = 0;
2408 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2409 return ct;
2410 }
2412 //------------------------------print-----------------------------
2413 void DepMem::print() {
2414 #ifndef PRODUCT
2415 tty->print(" DepNode %d (", _node->_idx);
2416 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2417 Node* pred = p->pred()->node();
2418 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2419 }
2420 tty->print(") [");
2421 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2422 Node* succ = s->succ()->node();
2423 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2424 }
2425 tty->print_cr(" ]");
2426 #endif
2427 }
2429 // =========================== DepEdge =========================
2431 //------------------------------DepPreds---------------------------
2432 void DepEdge::print() {
2433 #ifndef PRODUCT
2434 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2435 #endif
2436 }
2438 // =========================== DepPreds =========================
2439 // Iterator over predecessor edges in the dependence graph.
2441 //------------------------------DepPreds---------------------------
2442 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2443 _n = n;
2444 _done = false;
2445 if (_n->is_Store() || _n->is_Load()) {
2446 _next_idx = MemNode::Address;
2447 _end_idx = n->req();
2448 _dep_next = dg.dep(_n)->in_head();
2449 } else if (_n->is_Mem()) {
2450 _next_idx = 0;
2451 _end_idx = 0;
2452 _dep_next = dg.dep(_n)->in_head();
2453 } else {
2454 _next_idx = 1;
2455 _end_idx = _n->req();
2456 _dep_next = NULL;
2457 }
2458 next();
2459 }
2461 //------------------------------next---------------------------
2462 void DepPreds::next() {
2463 if (_dep_next != NULL) {
2464 _current = _dep_next->pred()->node();
2465 _dep_next = _dep_next->next_in();
2466 } else if (_next_idx < _end_idx) {
2467 _current = _n->in(_next_idx++);
2468 } else {
2469 _done = true;
2470 }
2471 }
2473 // =========================== DepSuccs =========================
2474 // Iterator over successor edges in the dependence graph.
2476 //------------------------------DepSuccs---------------------------
2477 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2478 _n = n;
2479 _done = false;
2480 if (_n->is_Load()) {
2481 _next_idx = 0;
2482 _end_idx = _n->outcnt();
2483 _dep_next = dg.dep(_n)->out_head();
2484 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2485 _next_idx = 0;
2486 _end_idx = 0;
2487 _dep_next = dg.dep(_n)->out_head();
2488 } else {
2489 _next_idx = 0;
2490 _end_idx = _n->outcnt();
2491 _dep_next = NULL;
2492 }
2493 next();
2494 }
2496 //-------------------------------next---------------------------
2497 void DepSuccs::next() {
2498 if (_dep_next != NULL) {
2499 _current = _dep_next->succ()->node();
2500 _dep_next = _dep_next->next_out();
2501 } else if (_next_idx < _end_idx) {
2502 _current = _n->raw_out(_next_idx++);
2503 } else {
2504 _done = true;
2505 }
2506 }