0
|
1 /*
|
|
2 * Copyright 2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 */
|
|
23
|
|
24 #include "incls/_precompiled.incl"
|
|
25 #include "incls/_superword.cpp.incl"
|
|
26
|
|
27 //
|
|
28 // S U P E R W O R D T R A N S F O R M
|
|
29 //=============================================================================
|
|
30
|
|
31 //------------------------------SuperWord---------------------------
|
|
32 SuperWord::SuperWord(PhaseIdealLoop* phase) :
|
|
33 _phase(phase),
|
|
34 _igvn(phase->_igvn),
|
|
35 _arena(phase->C->comp_arena()),
|
|
36 _packset(arena(), 8, 0, NULL), // packs for the current block
|
|
37 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
|
|
38 _block(arena(), 8, 0, NULL), // nodes in current block
|
|
39 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
|
|
40 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
|
|
41 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
|
|
42 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
|
|
43 _align_to_ref(NULL), // memory reference to align vectors to
|
|
44 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
|
|
45 _dg(_arena), // dependence graph
|
|
46 _visited(arena()), // visited node set
|
|
47 _post_visited(arena()), // post visited node set
|
|
48 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
|
|
49 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
|
|
50 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
|
|
51 _lpt(NULL), // loop tree node
|
|
52 _lp(NULL), // LoopNode
|
|
53 _bb(NULL), // basic block
|
|
54 _iv(NULL) // induction var
|
|
55 {}
|
|
56
|
|
57 //------------------------------transform_loop---------------------------
|
|
58 void SuperWord::transform_loop(IdealLoopTree* lpt) {
|
|
59 assert(lpt->_head->is_CountedLoop(), "must be");
|
|
60 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
|
|
61
|
|
62 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
|
|
63
|
|
64 // Check for no control flow in body (other than exit)
|
|
65 Node *cl_exit = cl->loopexit();
|
|
66 if (cl_exit->in(0) != lpt->_head) return;
|
|
67
|
|
68 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
|
|
69 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
|
|
70 if (pre_end == NULL) return;
|
|
71 Node *pre_opaq1 = pre_end->limit();
|
|
72 if (pre_opaq1->Opcode() != Op_Opaque1) return;
|
|
73
|
|
74 // Do vectors exist on this architecture?
|
|
75 if (vector_width_in_bytes() == 0) return;
|
|
76
|
|
77 init(); // initialize data structures
|
|
78
|
|
79 set_lpt(lpt);
|
|
80 set_lp(cl);
|
|
81
|
|
82 // For now, define one block which is the entire loop body
|
|
83 set_bb(cl);
|
|
84
|
|
85 assert(_packset.length() == 0, "packset must be empty");
|
|
86 SLP_extract();
|
|
87 }
|
|
88
|
|
89 //------------------------------SLP_extract---------------------------
|
|
90 // Extract the superword level parallelism
|
|
91 //
|
|
92 // 1) A reverse post-order of nodes in the block is constructed. By scanning
|
|
93 // this list from first to last, all definitions are visited before their uses.
|
|
94 //
|
|
95 // 2) A point-to-point dependence graph is constructed between memory references.
|
|
96 // This simplies the upcoming "independence" checker.
|
|
97 //
|
|
98 // 3) The maximum depth in the node graph from the beginning of the block
|
|
99 // to each node is computed. This is used to prune the graph search
|
|
100 // in the independence checker.
|
|
101 //
|
|
102 // 4) For integer types, the necessary bit width is propagated backwards
|
|
103 // from stores to allow packed operations on byte, char, and short
|
|
104 // integers. This reverses the promotion to type "int" that javac
|
|
105 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
|
|
106 //
|
|
107 // 5) One of the memory references is picked to be an aligned vector reference.
|
|
108 // The pre-loop trip count is adjusted to align this reference in the
|
|
109 // unrolled body.
|
|
110 //
|
|
111 // 6) The initial set of pack pairs is seeded with memory references.
|
|
112 //
|
|
113 // 7) The set of pack pairs is extended by following use->def and def->use links.
|
|
114 //
|
|
115 // 8) The pairs are combined into vector sized packs.
|
|
116 //
|
|
117 // 9) Reorder the memory slices to co-locate members of the memory packs.
|
|
118 //
|
|
119 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
|
|
120 // inserting scalar promotion, vector creation from multiple scalars, and
|
|
121 // extraction of scalar values from vectors.
|
|
122 //
|
|
123 void SuperWord::SLP_extract() {
|
|
124
|
|
125 // Ready the block
|
|
126
|
|
127 construct_bb();
|
|
128
|
|
129 dependence_graph();
|
|
130
|
|
131 compute_max_depth();
|
|
132
|
|
133 compute_vector_element_type();
|
|
134
|
|
135 // Attempt vectorization
|
|
136
|
|
137 find_adjacent_refs();
|
|
138
|
|
139 extend_packlist();
|
|
140
|
|
141 combine_packs();
|
|
142
|
|
143 construct_my_pack_map();
|
|
144
|
|
145 filter_packs();
|
|
146
|
|
147 schedule();
|
|
148
|
|
149 output();
|
|
150 }
|
|
151
|
|
152 //------------------------------find_adjacent_refs---------------------------
|
|
153 // Find the adjacent memory references and create pack pairs for them.
|
|
154 // This is the initial set of packs that will then be extended by
|
|
155 // following use->def and def->use links. The align positions are
|
|
156 // assigned relative to the reference "align_to_ref"
|
|
157 void SuperWord::find_adjacent_refs() {
|
|
158 // Get list of memory operations
|
|
159 Node_List memops;
|
|
160 for (int i = 0; i < _block.length(); i++) {
|
|
161 Node* n = _block.at(i);
|
|
162 if (n->is_Mem() && in_bb(n)) {
|
|
163 int align = memory_alignment(n->as_Mem(), 0);
|
|
164 if (align != bottom_align) {
|
|
165 memops.push(n);
|
|
166 }
|
|
167 }
|
|
168 }
|
|
169 if (memops.size() == 0) return;
|
|
170
|
|
171 // Find a memory reference to align to. The pre-loop trip count
|
|
172 // is modified to align this reference to a vector-aligned address
|
|
173 find_align_to_ref(memops);
|
|
174 if (align_to_ref() == NULL) return;
|
|
175
|
|
176 SWPointer align_to_ref_p(align_to_ref(), this);
|
|
177 int offset = align_to_ref_p.offset_in_bytes();
|
|
178 int scale = align_to_ref_p.scale_in_bytes();
|
|
179 int vw = vector_width_in_bytes();
|
|
180 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
|
|
181 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
|
|
182
|
|
183 #ifndef PRODUCT
|
|
184 if (TraceSuperWord)
|
|
185 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d",
|
|
186 offset, iv_adjustment, align_to_ref_p.memory_size());
|
|
187 #endif
|
|
188
|
|
189 // Set alignment relative to "align_to_ref"
|
|
190 for (int i = memops.size() - 1; i >= 0; i--) {
|
|
191 MemNode* s = memops.at(i)->as_Mem();
|
|
192 SWPointer p2(s, this);
|
|
193 if (p2.comparable(align_to_ref_p)) {
|
|
194 int align = memory_alignment(s, iv_adjustment);
|
|
195 set_alignment(s, align);
|
|
196 } else {
|
|
197 memops.remove(i);
|
|
198 }
|
|
199 }
|
|
200
|
|
201 // Create initial pack pairs of memory operations
|
|
202 for (uint i = 0; i < memops.size(); i++) {
|
|
203 Node* s1 = memops.at(i);
|
|
204 for (uint j = 0; j < memops.size(); j++) {
|
|
205 Node* s2 = memops.at(j);
|
|
206 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
|
|
207 int align = alignment(s1);
|
|
208 if (stmts_can_pack(s1, s2, align)) {
|
|
209 Node_List* pair = new Node_List();
|
|
210 pair->push(s1);
|
|
211 pair->push(s2);
|
|
212 _packset.append(pair);
|
|
213 }
|
|
214 }
|
|
215 }
|
|
216 }
|
|
217
|
|
218 #ifndef PRODUCT
|
|
219 if (TraceSuperWord) {
|
|
220 tty->print_cr("\nAfter find_adjacent_refs");
|
|
221 print_packset();
|
|
222 }
|
|
223 #endif
|
|
224 }
|
|
225
|
|
226 //------------------------------find_align_to_ref---------------------------
|
|
227 // Find a memory reference to align the loop induction variable to.
|
|
228 // Looks first at stores then at loads, looking for a memory reference
|
|
229 // with the largest number of references similar to it.
|
|
230 void SuperWord::find_align_to_ref(Node_List &memops) {
|
|
231 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
|
|
232
|
|
233 // Count number of comparable memory ops
|
|
234 for (uint i = 0; i < memops.size(); i++) {
|
|
235 MemNode* s1 = memops.at(i)->as_Mem();
|
|
236 SWPointer p1(s1, this);
|
|
237 // Discard if pre loop can't align this reference
|
|
238 if (!ref_is_alignable(p1)) {
|
|
239 *cmp_ct.adr_at(i) = 0;
|
|
240 continue;
|
|
241 }
|
|
242 for (uint j = i+1; j < memops.size(); j++) {
|
|
243 MemNode* s2 = memops.at(j)->as_Mem();
|
|
244 if (isomorphic(s1, s2)) {
|
|
245 SWPointer p2(s2, this);
|
|
246 if (p1.comparable(p2)) {
|
|
247 (*cmp_ct.adr_at(i))++;
|
|
248 (*cmp_ct.adr_at(j))++;
|
|
249 }
|
|
250 }
|
|
251 }
|
|
252 }
|
|
253
|
|
254 // Find Store (or Load) with the greatest number of "comparable" references
|
|
255 int max_ct = 0;
|
|
256 int max_idx = -1;
|
|
257 int min_size = max_jint;
|
|
258 int min_iv_offset = max_jint;
|
|
259 for (uint j = 0; j < memops.size(); j++) {
|
|
260 MemNode* s = memops.at(j)->as_Mem();
|
|
261 if (s->is_Store()) {
|
|
262 SWPointer p(s, this);
|
|
263 if (cmp_ct.at(j) > max_ct ||
|
|
264 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
|
|
265 data_size(s) == min_size &&
|
|
266 p.offset_in_bytes() < min_iv_offset)) {
|
|
267 max_ct = cmp_ct.at(j);
|
|
268 max_idx = j;
|
|
269 min_size = data_size(s);
|
|
270 min_iv_offset = p.offset_in_bytes();
|
|
271 }
|
|
272 }
|
|
273 }
|
|
274 // If no stores, look at loads
|
|
275 if (max_ct == 0) {
|
|
276 for (uint j = 0; j < memops.size(); j++) {
|
|
277 MemNode* s = memops.at(j)->as_Mem();
|
|
278 if (s->is_Load()) {
|
|
279 SWPointer p(s, this);
|
|
280 if (cmp_ct.at(j) > max_ct ||
|
|
281 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
|
|
282 data_size(s) == min_size &&
|
|
283 p.offset_in_bytes() < min_iv_offset)) {
|
|
284 max_ct = cmp_ct.at(j);
|
|
285 max_idx = j;
|
|
286 min_size = data_size(s);
|
|
287 min_iv_offset = p.offset_in_bytes();
|
|
288 }
|
|
289 }
|
|
290 }
|
|
291 }
|
|
292
|
|
293 if (max_ct > 0)
|
|
294 set_align_to_ref(memops.at(max_idx)->as_Mem());
|
|
295
|
|
296 #ifndef PRODUCT
|
|
297 if (TraceSuperWord && Verbose) {
|
|
298 tty->print_cr("\nVector memops after find_align_to_refs");
|
|
299 for (uint i = 0; i < memops.size(); i++) {
|
|
300 MemNode* s = memops.at(i)->as_Mem();
|
|
301 s->dump();
|
|
302 }
|
|
303 }
|
|
304 #endif
|
|
305 }
|
|
306
|
|
307 //------------------------------ref_is_alignable---------------------------
|
|
308 // Can the preloop align the reference to position zero in the vector?
|
|
309 bool SuperWord::ref_is_alignable(SWPointer& p) {
|
|
310 if (!p.has_iv()) {
|
|
311 return true; // no induction variable
|
|
312 }
|
|
313 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
|
|
314 assert(pre_end->stride_is_con(), "pre loop stride is constant");
|
|
315 int preloop_stride = pre_end->stride_con();
|
|
316
|
|
317 int span = preloop_stride * p.scale_in_bytes();
|
|
318
|
|
319 // Stride one accesses are alignable.
|
|
320 if (ABS(span) == p.memory_size())
|
|
321 return true;
|
|
322
|
|
323 // If initial offset from start of object is computable,
|
|
324 // compute alignment within the vector.
|
|
325 int vw = vector_width_in_bytes();
|
|
326 if (vw % span == 0) {
|
|
327 Node* init_nd = pre_end->init_trip();
|
|
328 if (init_nd->is_Con() && p.invar() == NULL) {
|
|
329 int init = init_nd->bottom_type()->is_int()->get_con();
|
|
330
|
|
331 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
|
|
332 assert(init_offset >= 0, "positive offset from object start");
|
|
333
|
|
334 if (span > 0) {
|
|
335 return (vw - (init_offset % vw)) % span == 0;
|
|
336 } else {
|
|
337 assert(span < 0, "nonzero stride * scale");
|
|
338 return (init_offset % vw) % -span == 0;
|
|
339 }
|
|
340 }
|
|
341 }
|
|
342 return false;
|
|
343 }
|
|
344
|
|
345 //---------------------------dependence_graph---------------------------
|
|
346 // Construct dependency graph.
|
|
347 // Add dependence edges to load/store nodes for memory dependence
|
|
348 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
|
|
349 void SuperWord::dependence_graph() {
|
|
350 // First, assign a dependence node to each memory node
|
|
351 for (int i = 0; i < _block.length(); i++ ) {
|
|
352 Node *n = _block.at(i);
|
|
353 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
|
|
354 _dg.make_node(n);
|
|
355 }
|
|
356 }
|
|
357
|
|
358 // For each memory slice, create the dependences
|
|
359 for (int i = 0; i < _mem_slice_head.length(); i++) {
|
|
360 Node* n = _mem_slice_head.at(i);
|
|
361 Node* n_tail = _mem_slice_tail.at(i);
|
|
362
|
|
363 // Get slice in predecessor order (last is first)
|
|
364 mem_slice_preds(n_tail, n, _nlist);
|
|
365
|
|
366 // Make the slice dependent on the root
|
|
367 DepMem* slice = _dg.dep(n);
|
|
368 _dg.make_edge(_dg.root(), slice);
|
|
369
|
|
370 // Create a sink for the slice
|
|
371 DepMem* slice_sink = _dg.make_node(NULL);
|
|
372 _dg.make_edge(slice_sink, _dg.tail());
|
|
373
|
|
374 // Now visit each pair of memory ops, creating the edges
|
|
375 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
|
|
376 Node* s1 = _nlist.at(j);
|
|
377
|
|
378 // If no dependency yet, use slice
|
|
379 if (_dg.dep(s1)->in_cnt() == 0) {
|
|
380 _dg.make_edge(slice, s1);
|
|
381 }
|
|
382 SWPointer p1(s1->as_Mem(), this);
|
|
383 bool sink_dependent = true;
|
|
384 for (int k = j - 1; k >= 0; k--) {
|
|
385 Node* s2 = _nlist.at(k);
|
|
386 if (s1->is_Load() && s2->is_Load())
|
|
387 continue;
|
|
388 SWPointer p2(s2->as_Mem(), this);
|
|
389
|
|
390 int cmp = p1.cmp(p2);
|
|
391 if (SuperWordRTDepCheck &&
|
|
392 p1.base() != p2.base() && p1.valid() && p2.valid()) {
|
|
393 // Create a runtime check to disambiguate
|
|
394 OrderedPair pp(p1.base(), p2.base());
|
|
395 _disjoint_ptrs.append_if_missing(pp);
|
|
396 } else if (!SWPointer::not_equal(cmp)) {
|
|
397 // Possibly same address
|
|
398 _dg.make_edge(s1, s2);
|
|
399 sink_dependent = false;
|
|
400 }
|
|
401 }
|
|
402 if (sink_dependent) {
|
|
403 _dg.make_edge(s1, slice_sink);
|
|
404 }
|
|
405 }
|
|
406 #ifndef PRODUCT
|
|
407 if (TraceSuperWord) {
|
|
408 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
|
|
409 for (int q = 0; q < _nlist.length(); q++) {
|
|
410 _dg.print(_nlist.at(q));
|
|
411 }
|
|
412 tty->cr();
|
|
413 }
|
|
414 #endif
|
|
415 _nlist.clear();
|
|
416 }
|
|
417
|
|
418 #ifndef PRODUCT
|
|
419 if (TraceSuperWord) {
|
|
420 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
|
|
421 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
|
|
422 _disjoint_ptrs.at(r).print();
|
|
423 tty->cr();
|
|
424 }
|
|
425 tty->cr();
|
|
426 }
|
|
427 #endif
|
|
428 }
|
|
429
|
|
430 //---------------------------mem_slice_preds---------------------------
|
|
431 // Return a memory slice (node list) in predecessor order starting at "start"
|
|
432 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
|
|
433 assert(preds.length() == 0, "start empty");
|
|
434 Node* n = start;
|
|
435 Node* prev = NULL;
|
|
436 while (true) {
|
|
437 assert(in_bb(n), "must be in block");
|
|
438 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
|
|
439 Node* out = n->fast_out(i);
|
|
440 if (out->is_Load()) {
|
|
441 if (in_bb(out)) {
|
|
442 preds.push(out);
|
|
443 }
|
|
444 } else {
|
|
445 // FIXME
|
|
446 if (out->is_MergeMem() && !in_bb(out)) {
|
|
447 // Either unrolling is causing a memory edge not to disappear,
|
|
448 // or need to run igvn.optimize() again before SLP
|
|
449 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
|
|
450 // Ditto. Not sure what else to check further.
|
|
451 } else if (out->Opcode() == Op_StoreCM && out->in(4) == n) {
|
|
452 // StoreCM has an input edge used as a precedence edge.
|
|
453 // Maybe an issue when oop stores are vectorized.
|
|
454 } else {
|
|
455 assert(out == prev || prev == NULL, "no branches off of store slice");
|
|
456 }
|
|
457 }
|
|
458 }
|
|
459 if (n == stop) break;
|
|
460 preds.push(n);
|
|
461 prev = n;
|
|
462 n = n->in(MemNode::Memory);
|
|
463 }
|
|
464 }
|
|
465
|
|
466 //------------------------------stmts_can_pack---------------------------
|
|
467 // Can s1 and s2 be in a pack with s1 immediately preceeding s2 and
|
|
468 // s1 aligned at "align"
|
|
469 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
|
|
470 if (isomorphic(s1, s2)) {
|
|
471 if (independent(s1, s2)) {
|
|
472 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
|
|
473 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
|
|
474 int s1_align = alignment(s1);
|
|
475 int s2_align = alignment(s2);
|
|
476 if (s1_align == top_align || s1_align == align) {
|
|
477 if (s2_align == top_align || s2_align == align + data_size(s1)) {
|
|
478 return true;
|
|
479 }
|
|
480 }
|
|
481 }
|
|
482 }
|
|
483 }
|
|
484 }
|
|
485 return false;
|
|
486 }
|
|
487
|
|
488 //------------------------------exists_at---------------------------
|
|
489 // Does s exist in a pack at position pos?
|
|
490 bool SuperWord::exists_at(Node* s, uint pos) {
|
|
491 for (int i = 0; i < _packset.length(); i++) {
|
|
492 Node_List* p = _packset.at(i);
|
|
493 if (p->at(pos) == s) {
|
|
494 return true;
|
|
495 }
|
|
496 }
|
|
497 return false;
|
|
498 }
|
|
499
|
|
500 //------------------------------are_adjacent_refs---------------------------
|
|
501 // Is s1 immediately before s2 in memory?
|
|
502 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
|
|
503 if (!s1->is_Mem() || !s2->is_Mem()) return false;
|
|
504 if (!in_bb(s1) || !in_bb(s2)) return false;
|
|
505 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
|
|
506 // only pack memops that are in the same alias set until that's fixed.
|
|
507 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
|
|
508 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
|
|
509 return false;
|
|
510 SWPointer p1(s1->as_Mem(), this);
|
|
511 SWPointer p2(s2->as_Mem(), this);
|
|
512 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
|
|
513 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
|
|
514 return diff == data_size(s1);
|
|
515 }
|
|
516
|
|
517 //------------------------------isomorphic---------------------------
|
|
518 // Are s1 and s2 similar?
|
|
519 bool SuperWord::isomorphic(Node* s1, Node* s2) {
|
|
520 if (s1->Opcode() != s2->Opcode()) return false;
|
|
521 if (s1->req() != s2->req()) return false;
|
|
522 if (s1->in(0) != s2->in(0)) return false;
|
|
523 if (velt_type(s1) != velt_type(s2)) return false;
|
|
524 return true;
|
|
525 }
|
|
526
|
|
527 //------------------------------independent---------------------------
|
|
528 // Is there no data path from s1 to s2 or s2 to s1?
|
|
529 bool SuperWord::independent(Node* s1, Node* s2) {
|
|
530 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
|
|
531 int d1 = depth(s1);
|
|
532 int d2 = depth(s2);
|
|
533 if (d1 == d2) return s1 != s2;
|
|
534 Node* deep = d1 > d2 ? s1 : s2;
|
|
535 Node* shallow = d1 > d2 ? s2 : s1;
|
|
536
|
|
537 visited_clear();
|
|
538
|
|
539 return independent_path(shallow, deep);
|
|
540 }
|
|
541
|
|
542 //------------------------------independent_path------------------------------
|
|
543 // Helper for independent
|
|
544 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
|
|
545 if (dp >= 1000) return false; // stop deep recursion
|
|
546 visited_set(deep);
|
|
547 int shal_depth = depth(shallow);
|
|
548 assert(shal_depth <= depth(deep), "must be");
|
|
549 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
|
|
550 Node* pred = preds.current();
|
|
551 if (in_bb(pred) && !visited_test(pred)) {
|
|
552 if (shallow == pred) {
|
|
553 return false;
|
|
554 }
|
|
555 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
|
|
556 return false;
|
|
557 }
|
|
558 }
|
|
559 }
|
|
560 return true;
|
|
561 }
|
|
562
|
|
563 //------------------------------set_alignment---------------------------
|
|
564 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
|
|
565 set_alignment(s1, align);
|
|
566 set_alignment(s2, align + data_size(s1));
|
|
567 }
|
|
568
|
|
569 //------------------------------data_size---------------------------
|
|
570 int SuperWord::data_size(Node* s) {
|
|
571 const Type* t = velt_type(s);
|
|
572 BasicType bt = t->array_element_basic_type();
|
|
573 int bsize = type2aelembytes[bt];
|
|
574 assert(bsize != 0, "valid size");
|
|
575 return bsize;
|
|
576 }
|
|
577
|
|
578 //------------------------------extend_packlist---------------------------
|
|
579 // Extend packset by following use->def and def->use links from pack members.
|
|
580 void SuperWord::extend_packlist() {
|
|
581 bool changed;
|
|
582 do {
|
|
583 changed = false;
|
|
584 for (int i = 0; i < _packset.length(); i++) {
|
|
585 Node_List* p = _packset.at(i);
|
|
586 changed |= follow_use_defs(p);
|
|
587 changed |= follow_def_uses(p);
|
|
588 }
|
|
589 } while (changed);
|
|
590
|
|
591 #ifndef PRODUCT
|
|
592 if (TraceSuperWord) {
|
|
593 tty->print_cr("\nAfter extend_packlist");
|
|
594 print_packset();
|
|
595 }
|
|
596 #endif
|
|
597 }
|
|
598
|
|
599 //------------------------------follow_use_defs---------------------------
|
|
600 // Extend the packset by visiting operand definitions of nodes in pack p
|
|
601 bool SuperWord::follow_use_defs(Node_List* p) {
|
|
602 Node* s1 = p->at(0);
|
|
603 Node* s2 = p->at(1);
|
|
604 assert(p->size() == 2, "just checking");
|
|
605 assert(s1->req() == s2->req(), "just checking");
|
|
606 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
|
|
607
|
|
608 if (s1->is_Load()) return false;
|
|
609
|
|
610 int align = alignment(s1);
|
|
611 bool changed = false;
|
|
612 int start = s1->is_Store() ? MemNode::ValueIn : 1;
|
|
613 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
|
|
614 for (int j = start; j < end; j++) {
|
|
615 Node* t1 = s1->in(j);
|
|
616 Node* t2 = s2->in(j);
|
|
617 if (!in_bb(t1) || !in_bb(t2))
|
|
618 continue;
|
|
619 if (stmts_can_pack(t1, t2, align)) {
|
|
620 if (est_savings(t1, t2) >= 0) {
|
|
621 Node_List* pair = new Node_List();
|
|
622 pair->push(t1);
|
|
623 pair->push(t2);
|
|
624 _packset.append(pair);
|
|
625 set_alignment(t1, t2, align);
|
|
626 changed = true;
|
|
627 }
|
|
628 }
|
|
629 }
|
|
630 return changed;
|
|
631 }
|
|
632
|
|
633 //------------------------------follow_def_uses---------------------------
|
|
634 // Extend the packset by visiting uses of nodes in pack p
|
|
635 bool SuperWord::follow_def_uses(Node_List* p) {
|
|
636 bool changed = false;
|
|
637 Node* s1 = p->at(0);
|
|
638 Node* s2 = p->at(1);
|
|
639 assert(p->size() == 2, "just checking");
|
|
640 assert(s1->req() == s2->req(), "just checking");
|
|
641 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
|
|
642
|
|
643 if (s1->is_Store()) return false;
|
|
644
|
|
645 int align = alignment(s1);
|
|
646 int savings = -1;
|
|
647 Node* u1 = NULL;
|
|
648 Node* u2 = NULL;
|
|
649 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
|
|
650 Node* t1 = s1->fast_out(i);
|
|
651 if (!in_bb(t1)) continue;
|
|
652 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
|
|
653 Node* t2 = s2->fast_out(j);
|
|
654 if (!in_bb(t2)) continue;
|
|
655 if (!opnd_positions_match(s1, t1, s2, t2))
|
|
656 continue;
|
|
657 if (stmts_can_pack(t1, t2, align)) {
|
|
658 int my_savings = est_savings(t1, t2);
|
|
659 if (my_savings > savings) {
|
|
660 savings = my_savings;
|
|
661 u1 = t1;
|
|
662 u2 = t2;
|
|
663 }
|
|
664 }
|
|
665 }
|
|
666 }
|
|
667 if (savings >= 0) {
|
|
668 Node_List* pair = new Node_List();
|
|
669 pair->push(u1);
|
|
670 pair->push(u2);
|
|
671 _packset.append(pair);
|
|
672 set_alignment(u1, u2, align);
|
|
673 changed = true;
|
|
674 }
|
|
675 return changed;
|
|
676 }
|
|
677
|
|
678 //---------------------------opnd_positions_match-------------------------
|
|
679 // Is the use of d1 in u1 at the same operand position as d2 in u2?
|
|
680 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
|
|
681 uint ct = u1->req();
|
|
682 if (ct != u2->req()) return false;
|
|
683 uint i1 = 0;
|
|
684 uint i2 = 0;
|
|
685 do {
|
|
686 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
|
|
687 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
|
|
688 if (i1 != i2) {
|
|
689 return false;
|
|
690 }
|
|
691 } while (i1 < ct);
|
|
692 return true;
|
|
693 }
|
|
694
|
|
695 //------------------------------est_savings---------------------------
|
|
696 // Estimate the savings from executing s1 and s2 as a pack
|
|
697 int SuperWord::est_savings(Node* s1, Node* s2) {
|
|
698 int save = 2 - 1; // 2 operations per instruction in packed form
|
|
699
|
|
700 // inputs
|
|
701 for (uint i = 1; i < s1->req(); i++) {
|
|
702 Node* x1 = s1->in(i);
|
|
703 Node* x2 = s2->in(i);
|
|
704 if (x1 != x2) {
|
|
705 if (are_adjacent_refs(x1, x2)) {
|
|
706 save += adjacent_profit(x1, x2);
|
|
707 } else if (!in_packset(x1, x2)) {
|
|
708 save -= pack_cost(2);
|
|
709 } else {
|
|
710 save += unpack_cost(2);
|
|
711 }
|
|
712 }
|
|
713 }
|
|
714
|
|
715 // uses of result
|
|
716 uint ct = 0;
|
|
717 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
|
|
718 Node* s1_use = s1->fast_out(i);
|
|
719 for (int j = 0; j < _packset.length(); j++) {
|
|
720 Node_List* p = _packset.at(j);
|
|
721 if (p->at(0) == s1_use) {
|
|
722 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
|
|
723 Node* s2_use = s2->fast_out(k);
|
|
724 if (p->at(p->size()-1) == s2_use) {
|
|
725 ct++;
|
|
726 if (are_adjacent_refs(s1_use, s2_use)) {
|
|
727 save += adjacent_profit(s1_use, s2_use);
|
|
728 }
|
|
729 }
|
|
730 }
|
|
731 }
|
|
732 }
|
|
733 }
|
|
734
|
|
735 if (ct < s1->outcnt()) save += unpack_cost(1);
|
|
736 if (ct < s2->outcnt()) save += unpack_cost(1);
|
|
737
|
|
738 return save;
|
|
739 }
|
|
740
|
|
741 //------------------------------costs---------------------------
|
|
742 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
|
|
743 int SuperWord::pack_cost(int ct) { return ct; }
|
|
744 int SuperWord::unpack_cost(int ct) { return ct; }
|
|
745
|
|
746 //------------------------------combine_packs---------------------------
|
|
747 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
|
|
748 void SuperWord::combine_packs() {
|
|
749 bool changed;
|
|
750 do {
|
|
751 changed = false;
|
|
752 for (int i = 0; i < _packset.length(); i++) {
|
|
753 Node_List* p1 = _packset.at(i);
|
|
754 if (p1 == NULL) continue;
|
|
755 for (int j = 0; j < _packset.length(); j++) {
|
|
756 Node_List* p2 = _packset.at(j);
|
|
757 if (p2 == NULL) continue;
|
|
758 if (p1->at(p1->size()-1) == p2->at(0)) {
|
|
759 for (uint k = 1; k < p2->size(); k++) {
|
|
760 p1->push(p2->at(k));
|
|
761 }
|
|
762 _packset.at_put(j, NULL);
|
|
763 changed = true;
|
|
764 }
|
|
765 }
|
|
766 }
|
|
767 } while (changed);
|
|
768
|
|
769 for (int i = _packset.length() - 1; i >= 0; i--) {
|
|
770 Node_List* p1 = _packset.at(i);
|
|
771 if (p1 == NULL) {
|
|
772 _packset.remove_at(i);
|
|
773 }
|
|
774 }
|
|
775
|
|
776 #ifndef PRODUCT
|
|
777 if (TraceSuperWord) {
|
|
778 tty->print_cr("\nAfter combine_packs");
|
|
779 print_packset();
|
|
780 }
|
|
781 #endif
|
|
782 }
|
|
783
|
|
784 //-----------------------------construct_my_pack_map--------------------------
|
|
785 // Construct the map from nodes to packs. Only valid after the
|
|
786 // point where a node is only in one pack (after combine_packs).
|
|
787 void SuperWord::construct_my_pack_map() {
|
|
788 Node_List* rslt = NULL;
|
|
789 for (int i = 0; i < _packset.length(); i++) {
|
|
790 Node_List* p = _packset.at(i);
|
|
791 for (uint j = 0; j < p->size(); j++) {
|
|
792 Node* s = p->at(j);
|
|
793 assert(my_pack(s) == NULL, "only in one pack");
|
|
794 set_my_pack(s, p);
|
|
795 }
|
|
796 }
|
|
797 }
|
|
798
|
|
799 //------------------------------filter_packs---------------------------
|
|
800 // Remove packs that are not implemented or not profitable.
|
|
801 void SuperWord::filter_packs() {
|
|
802
|
|
803 // Remove packs that are not implemented
|
|
804 for (int i = _packset.length() - 1; i >= 0; i--) {
|
|
805 Node_List* pk = _packset.at(i);
|
|
806 bool impl = implemented(pk);
|
|
807 if (!impl) {
|
|
808 #ifndef PRODUCT
|
|
809 if (TraceSuperWord && Verbose) {
|
|
810 tty->print_cr("Unimplemented");
|
|
811 pk->at(0)->dump();
|
|
812 }
|
|
813 #endif
|
|
814 remove_pack_at(i);
|
|
815 }
|
|
816 }
|
|
817
|
|
818 // Remove packs that are not profitable
|
|
819 bool changed;
|
|
820 do {
|
|
821 changed = false;
|
|
822 for (int i = _packset.length() - 1; i >= 0; i--) {
|
|
823 Node_List* pk = _packset.at(i);
|
|
824 bool prof = profitable(pk);
|
|
825 if (!prof) {
|
|
826 #ifndef PRODUCT
|
|
827 if (TraceSuperWord && Verbose) {
|
|
828 tty->print_cr("Unprofitable");
|
|
829 pk->at(0)->dump();
|
|
830 }
|
|
831 #endif
|
|
832 remove_pack_at(i);
|
|
833 changed = true;
|
|
834 }
|
|
835 }
|
|
836 } while (changed);
|
|
837
|
|
838 #ifndef PRODUCT
|
|
839 if (TraceSuperWord) {
|
|
840 tty->print_cr("\nAfter filter_packs");
|
|
841 print_packset();
|
|
842 tty->cr();
|
|
843 }
|
|
844 #endif
|
|
845 }
|
|
846
|
|
847 //------------------------------implemented---------------------------
|
|
848 // Can code be generated for pack p?
|
|
849 bool SuperWord::implemented(Node_List* p) {
|
|
850 Node* p0 = p->at(0);
|
|
851 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
|
|
852 return vopc > 0 && Matcher::has_match_rule(vopc);
|
|
853 }
|
|
854
|
|
855 //------------------------------profitable---------------------------
|
|
856 // For pack p, are all operands and all uses (with in the block) vector?
|
|
857 bool SuperWord::profitable(Node_List* p) {
|
|
858 Node* p0 = p->at(0);
|
|
859 uint start, end;
|
|
860 vector_opd_range(p0, &start, &end);
|
|
861
|
|
862 // Return false if some input is not vector and inside block
|
|
863 for (uint i = start; i < end; i++) {
|
|
864 if (!is_vector_use(p0, i)) {
|
|
865 // For now, return false if not scalar promotion case (inputs are the same.)
|
|
866 // Later, implement PackNode and allow differring, non-vector inputs
|
|
867 // (maybe just the ones from outside the block.)
|
|
868 Node* p0_def = p0->in(i);
|
|
869 for (uint j = 1; j < p->size(); j++) {
|
|
870 Node* use = p->at(j);
|
|
871 Node* def = use->in(i);
|
|
872 if (p0_def != def)
|
|
873 return false;
|
|
874 }
|
|
875 }
|
|
876 }
|
|
877 if (!p0->is_Store()) {
|
|
878 // For now, return false if not all uses are vector.
|
|
879 // Later, implement ExtractNode and allow non-vector uses (maybe
|
|
880 // just the ones outside the block.)
|
|
881 for (uint i = 0; i < p->size(); i++) {
|
|
882 Node* def = p->at(i);
|
|
883 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
|
|
884 Node* use = def->fast_out(j);
|
|
885 for (uint k = 0; k < use->req(); k++) {
|
|
886 Node* n = use->in(k);
|
|
887 if (def == n) {
|
|
888 if (!is_vector_use(use, k)) {
|
|
889 return false;
|
|
890 }
|
|
891 }
|
|
892 }
|
|
893 }
|
|
894 }
|
|
895 }
|
|
896 return true;
|
|
897 }
|
|
898
|
|
899 //------------------------------schedule---------------------------
|
|
900 // Adjust the memory graph for the packed operations
|
|
901 void SuperWord::schedule() {
|
|
902
|
|
903 // Co-locate in the memory graph the members of each memory pack
|
|
904 for (int i = 0; i < _packset.length(); i++) {
|
|
905 co_locate_pack(_packset.at(i));
|
|
906 }
|
|
907 }
|
|
908
|
|
909 //------------------------------co_locate_pack---------------------------
|
|
910 // Within a pack, move stores down to the last executed store,
|
|
911 // and move loads up to the first executed load.
|
|
912 void SuperWord::co_locate_pack(Node_List* pk) {
|
|
913 if (pk->at(0)->is_Store()) {
|
|
914 // Push Stores down towards last executed pack member
|
|
915 MemNode* first = executed_first(pk)->as_Mem();
|
|
916 MemNode* last = executed_last(pk)->as_Mem();
|
|
917 MemNode* insert_pt = last;
|
|
918 MemNode* current = last->in(MemNode::Memory)->as_Mem();
|
|
919 while (true) {
|
|
920 assert(in_bb(current), "stay in block");
|
|
921 Node* my_mem = current->in(MemNode::Memory);
|
|
922 if (in_pack(current, pk)) {
|
|
923 // Forward users of my memory state to my input memory state
|
|
924 _igvn.hash_delete(current);
|
|
925 _igvn.hash_delete(my_mem);
|
|
926 for (DUIterator i = current->outs(); current->has_out(i); i++) {
|
|
927 Node* use = current->out(i);
|
|
928 if (use->is_Mem()) {
|
|
929 assert(use->in(MemNode::Memory) == current, "must be");
|
|
930 _igvn.hash_delete(use);
|
|
931 use->set_req(MemNode::Memory, my_mem);
|
|
932 _igvn._worklist.push(use);
|
|
933 --i; // deleted this edge; rescan position
|
|
934 }
|
|
935 }
|
|
936 // put current immediately before insert_pt
|
|
937 current->set_req(MemNode::Memory, insert_pt->in(MemNode::Memory));
|
|
938 _igvn.hash_delete(insert_pt);
|
|
939 insert_pt->set_req(MemNode::Memory, current);
|
|
940 _igvn._worklist.push(insert_pt);
|
|
941 _igvn._worklist.push(current);
|
|
942 insert_pt = current;
|
|
943 }
|
|
944 if (current == first) break;
|
|
945 current = my_mem->as_Mem();
|
|
946 }
|
|
947 } else if (pk->at(0)->is_Load()) {
|
|
948 // Pull Loads up towards first executed pack member
|
|
949 LoadNode* first = executed_first(pk)->as_Load();
|
|
950 Node* first_mem = first->in(MemNode::Memory);
|
|
951 _igvn.hash_delete(first_mem);
|
|
952 // Give each load same memory state as first
|
|
953 for (uint i = 0; i < pk->size(); i++) {
|
|
954 LoadNode* ld = pk->at(i)->as_Load();
|
|
955 _igvn.hash_delete(ld);
|
|
956 ld->set_req(MemNode::Memory, first_mem);
|
|
957 _igvn._worklist.push(ld);
|
|
958 }
|
|
959 }
|
|
960 }
|
|
961
|
|
962 //------------------------------output---------------------------
|
|
963 // Convert packs into vector node operations
|
|
964 void SuperWord::output() {
|
|
965 if (_packset.length() == 0) return;
|
|
966
|
|
967 // MUST ENSURE main loop's initial value is properly aligned:
|
|
968 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
|
|
969
|
|
970 align_initial_loop_index(align_to_ref());
|
|
971
|
|
972 // Insert extract (unpack) operations for scalar uses
|
|
973 for (int i = 0; i < _packset.length(); i++) {
|
|
974 insert_extracts(_packset.at(i));
|
|
975 }
|
|
976
|
|
977 for (int i = 0; i < _block.length(); i++) {
|
|
978 Node* n = _block.at(i);
|
|
979 Node_List* p = my_pack(n);
|
|
980 if (p && n == executed_last(p)) {
|
|
981 uint vlen = p->size();
|
|
982 Node* vn = NULL;
|
|
983 Node* low_adr = p->at(0);
|
|
984 Node* first = executed_first(p);
|
|
985 if (n->is_Load()) {
|
|
986 int opc = n->Opcode();
|
|
987 Node* ctl = n->in(MemNode::Control);
|
|
988 Node* mem = first->in(MemNode::Memory);
|
|
989 Node* adr = low_adr->in(MemNode::Address);
|
|
990 const TypePtr* atyp = n->adr_type();
|
|
991 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
|
|
992
|
|
993 } else if (n->is_Store()) {
|
|
994 // Promote value to be stored to vector
|
|
995 VectorNode* val = vector_opd(p, MemNode::ValueIn);
|
|
996
|
|
997 int opc = n->Opcode();
|
|
998 Node* ctl = n->in(MemNode::Control);
|
|
999 Node* mem = first->in(MemNode::Memory);
|
|
1000 Node* adr = low_adr->in(MemNode::Address);
|
|
1001 const TypePtr* atyp = n->adr_type();
|
|
1002 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
|
|
1003
|
|
1004 } else if (n->req() == 3) {
|
|
1005 // Promote operands to vector
|
|
1006 Node* in1 = vector_opd(p, 1);
|
|
1007 Node* in2 = vector_opd(p, 2);
|
|
1008 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
|
|
1009
|
|
1010 } else {
|
|
1011 ShouldNotReachHere();
|
|
1012 }
|
|
1013
|
|
1014 _phase->_igvn.register_new_node_with_optimizer(vn);
|
|
1015 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
|
|
1016 for (uint j = 0; j < p->size(); j++) {
|
|
1017 Node* pm = p->at(j);
|
|
1018 _igvn.hash_delete(pm);
|
|
1019 _igvn.subsume_node(pm, vn);
|
|
1020 }
|
|
1021 _igvn._worklist.push(vn);
|
|
1022 }
|
|
1023 }
|
|
1024 }
|
|
1025
|
|
1026 //------------------------------vector_opd---------------------------
|
|
1027 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
|
|
1028 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
|
|
1029 Node* p0 = p->at(0);
|
|
1030 uint vlen = p->size();
|
|
1031 Node* opd = p0->in(opd_idx);
|
|
1032
|
|
1033 bool same_opd = true;
|
|
1034 for (uint i = 1; i < vlen; i++) {
|
|
1035 Node* pi = p->at(i);
|
|
1036 Node* in = pi->in(opd_idx);
|
|
1037 if (opd != in) {
|
|
1038 same_opd = false;
|
|
1039 break;
|
|
1040 }
|
|
1041 }
|
|
1042
|
|
1043 if (same_opd) {
|
|
1044 if (opd->is_Vector()) {
|
|
1045 return (VectorNode*)opd; // input is matching vector
|
|
1046 }
|
|
1047 // Convert scalar input to vector. Use p0's type because it's container
|
|
1048 // maybe smaller than the operand's container.
|
|
1049 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
|
|
1050 const Type* p0_t = velt_type(p0);
|
|
1051 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
|
|
1052 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
|
|
1053
|
|
1054 _phase->_igvn.register_new_node_with_optimizer(vn);
|
|
1055 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
|
|
1056 return vn;
|
|
1057 }
|
|
1058
|
|
1059 // Insert pack operation
|
|
1060 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
|
|
1061 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
|
|
1062
|
|
1063 for (uint i = 1; i < vlen; i++) {
|
|
1064 Node* pi = p->at(i);
|
|
1065 Node* in = pi->in(opd_idx);
|
|
1066 assert(my_pack(in) == NULL, "Should already have been unpacked");
|
|
1067 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
|
|
1068 pk->add_opd(in);
|
|
1069 }
|
|
1070 _phase->_igvn.register_new_node_with_optimizer(pk);
|
|
1071 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
|
|
1072 return pk;
|
|
1073 }
|
|
1074
|
|
1075 //------------------------------insert_extracts---------------------------
|
|
1076 // If a use of pack p is not a vector use, then replace the
|
|
1077 // use with an extract operation.
|
|
1078 void SuperWord::insert_extracts(Node_List* p) {
|
|
1079 if (p->at(0)->is_Store()) return;
|
|
1080 assert(_n_idx_list.is_empty(), "empty (node,index) list");
|
|
1081
|
|
1082 // Inspect each use of each pack member. For each use that is
|
|
1083 // not a vector use, replace the use with an extract operation.
|
|
1084
|
|
1085 for (uint i = 0; i < p->size(); i++) {
|
|
1086 Node* def = p->at(i);
|
|
1087 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
|
|
1088 Node* use = def->fast_out(j);
|
|
1089 for (uint k = 0; k < use->req(); k++) {
|
|
1090 Node* n = use->in(k);
|
|
1091 if (def == n) {
|
|
1092 if (!is_vector_use(use, k)) {
|
|
1093 _n_idx_list.push(use, k);
|
|
1094 }
|
|
1095 }
|
|
1096 }
|
|
1097 }
|
|
1098 }
|
|
1099
|
|
1100 while (_n_idx_list.is_nonempty()) {
|
|
1101 Node* use = _n_idx_list.node();
|
|
1102 int idx = _n_idx_list.index();
|
|
1103 _n_idx_list.pop();
|
|
1104 Node* def = use->in(idx);
|
|
1105
|
|
1106 // Insert extract operation
|
|
1107 _igvn.hash_delete(def);
|
|
1108 _igvn.hash_delete(use);
|
|
1109 int def_pos = alignment(def) / data_size(def);
|
|
1110 const Type* def_t = velt_type(def);
|
|
1111
|
|
1112 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
|
|
1113 _phase->_igvn.register_new_node_with_optimizer(ex);
|
|
1114 _phase->set_ctrl(ex, _phase->get_ctrl(def));
|
|
1115 use->set_req(idx, ex);
|
|
1116 _igvn._worklist.push(def);
|
|
1117 _igvn._worklist.push(use);
|
|
1118
|
|
1119 bb_insert_after(ex, bb_idx(def));
|
|
1120 set_velt_type(ex, def_t);
|
|
1121 }
|
|
1122 }
|
|
1123
|
|
1124 //------------------------------is_vector_use---------------------------
|
|
1125 // Is use->in(u_idx) a vector use?
|
|
1126 bool SuperWord::is_vector_use(Node* use, int u_idx) {
|
|
1127 Node_List* u_pk = my_pack(use);
|
|
1128 if (u_pk == NULL) return false;
|
|
1129 Node* def = use->in(u_idx);
|
|
1130 Node_List* d_pk = my_pack(def);
|
|
1131 if (d_pk == NULL) {
|
|
1132 // check for scalar promotion
|
|
1133 Node* n = u_pk->at(0)->in(u_idx);
|
|
1134 for (uint i = 1; i < u_pk->size(); i++) {
|
|
1135 if (u_pk->at(i)->in(u_idx) != n) return false;
|
|
1136 }
|
|
1137 return true;
|
|
1138 }
|
|
1139 if (u_pk->size() != d_pk->size())
|
|
1140 return false;
|
|
1141 for (uint i = 0; i < u_pk->size(); i++) {
|
|
1142 Node* ui = u_pk->at(i);
|
|
1143 Node* di = d_pk->at(i);
|
|
1144 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
|
|
1145 return false;
|
|
1146 }
|
|
1147 return true;
|
|
1148 }
|
|
1149
|
|
1150 //------------------------------construct_bb---------------------------
|
|
1151 // Construct reverse postorder list of block members
|
|
1152 void SuperWord::construct_bb() {
|
|
1153 Node* entry = bb();
|
|
1154
|
|
1155 assert(_stk.length() == 0, "stk is empty");
|
|
1156 assert(_block.length() == 0, "block is empty");
|
|
1157 assert(_data_entry.length() == 0, "data_entry is empty");
|
|
1158 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
|
|
1159 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
|
|
1160
|
|
1161 // Find non-control nodes with no inputs from within block,
|
|
1162 // create a temporary map from node _idx to bb_idx for use
|
|
1163 // by the visited and post_visited sets,
|
|
1164 // and count number of nodes in block.
|
|
1165 int bb_ct = 0;
|
|
1166 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
|
|
1167 Node *n = lpt()->_body.at(i);
|
|
1168 set_bb_idx(n, i); // Create a temporary map
|
|
1169 if (in_bb(n)) {
|
|
1170 bb_ct++;
|
|
1171 if (!n->is_CFG()) {
|
|
1172 bool found = false;
|
|
1173 for (uint j = 0; j < n->req(); j++) {
|
|
1174 Node* def = n->in(j);
|
|
1175 if (def && in_bb(def)) {
|
|
1176 found = true;
|
|
1177 break;
|
|
1178 }
|
|
1179 }
|
|
1180 if (!found) {
|
|
1181 assert(n != entry, "can't be entry");
|
|
1182 _data_entry.push(n);
|
|
1183 }
|
|
1184 }
|
|
1185 }
|
|
1186 }
|
|
1187
|
|
1188 // Find memory slices (head and tail)
|
|
1189 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
|
|
1190 Node *n = lp()->fast_out(i);
|
|
1191 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
|
|
1192 Node* n_tail = n->in(LoopNode::LoopBackControl);
|
|
1193 _mem_slice_head.push(n);
|
|
1194 _mem_slice_tail.push(n_tail);
|
|
1195 }
|
|
1196 }
|
|
1197
|
|
1198 // Create an RPO list of nodes in block
|
|
1199
|
|
1200 visited_clear();
|
|
1201 post_visited_clear();
|
|
1202
|
|
1203 // Push all non-control nodes with no inputs from within block, then control entry
|
|
1204 for (int j = 0; j < _data_entry.length(); j++) {
|
|
1205 Node* n = _data_entry.at(j);
|
|
1206 visited_set(n);
|
|
1207 _stk.push(n);
|
|
1208 }
|
|
1209 visited_set(entry);
|
|
1210 _stk.push(entry);
|
|
1211
|
|
1212 // Do a depth first walk over out edges
|
|
1213 int rpo_idx = bb_ct - 1;
|
|
1214 int size;
|
|
1215 while ((size = _stk.length()) > 0) {
|
|
1216 Node* n = _stk.top(); // Leave node on stack
|
|
1217 if (!visited_test_set(n)) {
|
|
1218 // forward arc in graph
|
|
1219 } else if (!post_visited_test(n)) {
|
|
1220 // cross or back arc
|
|
1221 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
|
|
1222 Node *use = n->fast_out(i);
|
|
1223 if (in_bb(use) && !visited_test(use) &&
|
|
1224 // Don't go around backedge
|
|
1225 (!use->is_Phi() || n == entry)) {
|
|
1226 _stk.push(use);
|
|
1227 }
|
|
1228 }
|
|
1229 if (_stk.length() == size) {
|
|
1230 // There were no additional uses, post visit node now
|
|
1231 _stk.pop(); // Remove node from stack
|
|
1232 assert(rpo_idx >= 0, "");
|
|
1233 _block.at_put_grow(rpo_idx, n);
|
|
1234 rpo_idx--;
|
|
1235 post_visited_set(n);
|
|
1236 assert(rpo_idx >= 0 || _stk.is_empty(), "");
|
|
1237 }
|
|
1238 } else {
|
|
1239 _stk.pop(); // Remove post-visited node from stack
|
|
1240 }
|
|
1241 }
|
|
1242
|
|
1243 // Create real map of block indices for nodes
|
|
1244 for (int j = 0; j < _block.length(); j++) {
|
|
1245 Node* n = _block.at(j);
|
|
1246 set_bb_idx(n, j);
|
|
1247 }
|
|
1248
|
|
1249 initialize_bb(); // Ensure extra info is allocated.
|
|
1250
|
|
1251 #ifndef PRODUCT
|
|
1252 if (TraceSuperWord) {
|
|
1253 print_bb();
|
|
1254 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
|
|
1255 for (int m = 0; m < _data_entry.length(); m++) {
|
|
1256 tty->print("%3d ", m);
|
|
1257 _data_entry.at(m)->dump();
|
|
1258 }
|
|
1259 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
|
|
1260 for (int m = 0; m < _mem_slice_head.length(); m++) {
|
|
1261 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
|
|
1262 tty->print(" "); _mem_slice_tail.at(m)->dump();
|
|
1263 }
|
|
1264 }
|
|
1265 #endif
|
|
1266 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
|
|
1267 }
|
|
1268
|
|
1269 //------------------------------initialize_bb---------------------------
|
|
1270 // Initialize per node info
|
|
1271 void SuperWord::initialize_bb() {
|
|
1272 Node* last = _block.at(_block.length() - 1);
|
|
1273 grow_node_info(bb_idx(last));
|
|
1274 }
|
|
1275
|
|
1276 //------------------------------bb_insert_after---------------------------
|
|
1277 // Insert n into block after pos
|
|
1278 void SuperWord::bb_insert_after(Node* n, int pos) {
|
|
1279 int n_pos = pos + 1;
|
|
1280 // Make room
|
|
1281 for (int i = _block.length() - 1; i >= n_pos; i--) {
|
|
1282 _block.at_put_grow(i+1, _block.at(i));
|
|
1283 }
|
|
1284 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
|
|
1285 _node_info.at_put_grow(j+1, _node_info.at(j));
|
|
1286 }
|
|
1287 // Set value
|
|
1288 _block.at_put_grow(n_pos, n);
|
|
1289 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
|
|
1290 // Adjust map from node->_idx to _block index
|
|
1291 for (int i = n_pos; i < _block.length(); i++) {
|
|
1292 set_bb_idx(_block.at(i), i);
|
|
1293 }
|
|
1294 }
|
|
1295
|
|
1296 //------------------------------compute_max_depth---------------------------
|
|
1297 // Compute max depth for expressions from beginning of block
|
|
1298 // Use to prune search paths during test for independence.
|
|
1299 void SuperWord::compute_max_depth() {
|
|
1300 int ct = 0;
|
|
1301 bool again;
|
|
1302 do {
|
|
1303 again = false;
|
|
1304 for (int i = 0; i < _block.length(); i++) {
|
|
1305 Node* n = _block.at(i);
|
|
1306 if (!n->is_Phi()) {
|
|
1307 int d_orig = depth(n);
|
|
1308 int d_in = 0;
|
|
1309 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
|
|
1310 Node* pred = preds.current();
|
|
1311 if (in_bb(pred)) {
|
|
1312 d_in = MAX2(d_in, depth(pred));
|
|
1313 }
|
|
1314 }
|
|
1315 if (d_in + 1 != d_orig) {
|
|
1316 set_depth(n, d_in + 1);
|
|
1317 again = true;
|
|
1318 }
|
|
1319 }
|
|
1320 }
|
|
1321 ct++;
|
|
1322 } while (again);
|
|
1323 #ifndef PRODUCT
|
|
1324 if (TraceSuperWord && Verbose)
|
|
1325 tty->print_cr("compute_max_depth iterated: %d times", ct);
|
|
1326 #endif
|
|
1327 }
|
|
1328
|
|
1329 //-------------------------compute_vector_element_type-----------------------
|
|
1330 // Compute necessary vector element type for expressions
|
|
1331 // This propagates backwards a narrower integer type when the
|
|
1332 // upper bits of the value are not needed.
|
|
1333 // Example: char a,b,c; a = b + c;
|
|
1334 // Normally the type of the add is integer, but for packed character
|
|
1335 // operations the type of the add needs to be char.
|
|
1336 void SuperWord::compute_vector_element_type() {
|
|
1337 #ifndef PRODUCT
|
|
1338 if (TraceSuperWord && Verbose)
|
|
1339 tty->print_cr("\ncompute_velt_type:");
|
|
1340 #endif
|
|
1341
|
|
1342 // Initial type
|
|
1343 for (int i = 0; i < _block.length(); i++) {
|
|
1344 Node* n = _block.at(i);
|
|
1345 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
|
|
1346 : _igvn.type(n);
|
|
1347 const Type* vt = container_type(t);
|
|
1348 set_velt_type(n, vt);
|
|
1349 }
|
|
1350
|
|
1351 // Propagate narrowed type backwards through operations
|
|
1352 // that don't depend on higher order bits
|
|
1353 for (int i = _block.length() - 1; i >= 0; i--) {
|
|
1354 Node* n = _block.at(i);
|
|
1355 // Only integer types need be examined
|
|
1356 if (n->bottom_type()->isa_int()) {
|
|
1357 uint start, end;
|
|
1358 vector_opd_range(n, &start, &end);
|
|
1359 const Type* vt = velt_type(n);
|
|
1360
|
|
1361 for (uint j = start; j < end; j++) {
|
|
1362 Node* in = n->in(j);
|
|
1363 // Don't propagate through a type conversion
|
|
1364 if (n->bottom_type() != in->bottom_type())
|
|
1365 continue;
|
|
1366 switch(in->Opcode()) {
|
|
1367 case Op_AddI: case Op_AddL:
|
|
1368 case Op_SubI: case Op_SubL:
|
|
1369 case Op_MulI: case Op_MulL:
|
|
1370 case Op_AndI: case Op_AndL:
|
|
1371 case Op_OrI: case Op_OrL:
|
|
1372 case Op_XorI: case Op_XorL:
|
|
1373 case Op_LShiftI: case Op_LShiftL:
|
|
1374 case Op_CMoveI: case Op_CMoveL:
|
|
1375 if (in_bb(in)) {
|
|
1376 bool same_type = true;
|
|
1377 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
|
|
1378 Node *use = in->fast_out(k);
|
|
1379 if (!in_bb(use) || velt_type(use) != vt) {
|
|
1380 same_type = false;
|
|
1381 break;
|
|
1382 }
|
|
1383 }
|
|
1384 if (same_type) {
|
|
1385 set_velt_type(in, vt);
|
|
1386 }
|
|
1387 }
|
|
1388 }
|
|
1389 }
|
|
1390 }
|
|
1391 }
|
|
1392 #ifndef PRODUCT
|
|
1393 if (TraceSuperWord && Verbose) {
|
|
1394 for (int i = 0; i < _block.length(); i++) {
|
|
1395 Node* n = _block.at(i);
|
|
1396 velt_type(n)->dump();
|
|
1397 tty->print("\t");
|
|
1398 n->dump();
|
|
1399 }
|
|
1400 }
|
|
1401 #endif
|
|
1402 }
|
|
1403
|
|
1404 //------------------------------memory_alignment---------------------------
|
|
1405 // Alignment within a vector memory reference
|
|
1406 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
|
|
1407 SWPointer p(s, this);
|
|
1408 if (!p.valid()) {
|
|
1409 return bottom_align;
|
|
1410 }
|
|
1411 int offset = p.offset_in_bytes();
|
|
1412 offset += iv_adjust_in_bytes;
|
|
1413 int off_rem = offset % vector_width_in_bytes();
|
|
1414 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
|
|
1415 return off_mod;
|
|
1416 }
|
|
1417
|
|
1418 //---------------------------container_type---------------------------
|
|
1419 // Smallest type containing range of values
|
|
1420 const Type* SuperWord::container_type(const Type* t) {
|
|
1421 if (t->isa_aryptr()) {
|
|
1422 t = t->is_aryptr()->elem();
|
|
1423 }
|
|
1424 if (t->basic_type() == T_INT) {
|
|
1425 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
|
|
1426 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
|
|
1427 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
|
|
1428 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
|
|
1429 return TypeInt::INT;
|
|
1430 }
|
|
1431 return t;
|
|
1432 }
|
|
1433
|
|
1434 //-------------------------vector_opd_range-----------------------
|
|
1435 // (Start, end] half-open range defining which operands are vector
|
|
1436 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
|
|
1437 switch (n->Opcode()) {
|
|
1438 case Op_LoadB: case Op_LoadC:
|
|
1439 case Op_LoadI: case Op_LoadL:
|
|
1440 case Op_LoadF: case Op_LoadD:
|
|
1441 case Op_LoadP:
|
|
1442 *start = 0;
|
|
1443 *end = 0;
|
|
1444 return;
|
|
1445 case Op_StoreB: case Op_StoreC:
|
|
1446 case Op_StoreI: case Op_StoreL:
|
|
1447 case Op_StoreF: case Op_StoreD:
|
|
1448 case Op_StoreP:
|
|
1449 *start = MemNode::ValueIn;
|
|
1450 *end = *start + 1;
|
|
1451 return;
|
|
1452 case Op_LShiftI: case Op_LShiftL:
|
|
1453 *start = 1;
|
|
1454 *end = 2;
|
|
1455 return;
|
|
1456 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
|
|
1457 *start = 2;
|
|
1458 *end = n->req();
|
|
1459 return;
|
|
1460 }
|
|
1461 *start = 1;
|
|
1462 *end = n->req(); // default is all operands
|
|
1463 }
|
|
1464
|
|
1465 //------------------------------in_packset---------------------------
|
|
1466 // Are s1 and s2 in a pack pair and ordered as s1,s2?
|
|
1467 bool SuperWord::in_packset(Node* s1, Node* s2) {
|
|
1468 for (int i = 0; i < _packset.length(); i++) {
|
|
1469 Node_List* p = _packset.at(i);
|
|
1470 assert(p->size() == 2, "must be");
|
|
1471 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
|
|
1472 return true;
|
|
1473 }
|
|
1474 }
|
|
1475 return false;
|
|
1476 }
|
|
1477
|
|
1478 //------------------------------in_pack---------------------------
|
|
1479 // Is s in pack p?
|
|
1480 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
|
|
1481 for (uint i = 0; i < p->size(); i++) {
|
|
1482 if (p->at(i) == s) {
|
|
1483 return p;
|
|
1484 }
|
|
1485 }
|
|
1486 return NULL;
|
|
1487 }
|
|
1488
|
|
1489 //------------------------------remove_pack_at---------------------------
|
|
1490 // Remove the pack at position pos in the packset
|
|
1491 void SuperWord::remove_pack_at(int pos) {
|
|
1492 Node_List* p = _packset.at(pos);
|
|
1493 for (uint i = 0; i < p->size(); i++) {
|
|
1494 Node* s = p->at(i);
|
|
1495 set_my_pack(s, NULL);
|
|
1496 }
|
|
1497 _packset.remove_at(pos);
|
|
1498 }
|
|
1499
|
|
1500 //------------------------------executed_first---------------------------
|
|
1501 // Return the node executed first in pack p. Uses the RPO block list
|
|
1502 // to determine order.
|
|
1503 Node* SuperWord::executed_first(Node_List* p) {
|
|
1504 Node* n = p->at(0);
|
|
1505 int n_rpo = bb_idx(n);
|
|
1506 for (uint i = 1; i < p->size(); i++) {
|
|
1507 Node* s = p->at(i);
|
|
1508 int s_rpo = bb_idx(s);
|
|
1509 if (s_rpo < n_rpo) {
|
|
1510 n = s;
|
|
1511 n_rpo = s_rpo;
|
|
1512 }
|
|
1513 }
|
|
1514 return n;
|
|
1515 }
|
|
1516
|
|
1517 //------------------------------executed_last---------------------------
|
|
1518 // Return the node executed last in pack p.
|
|
1519 Node* SuperWord::executed_last(Node_List* p) {
|
|
1520 Node* n = p->at(0);
|
|
1521 int n_rpo = bb_idx(n);
|
|
1522 for (uint i = 1; i < p->size(); i++) {
|
|
1523 Node* s = p->at(i);
|
|
1524 int s_rpo = bb_idx(s);
|
|
1525 if (s_rpo > n_rpo) {
|
|
1526 n = s;
|
|
1527 n_rpo = s_rpo;
|
|
1528 }
|
|
1529 }
|
|
1530 return n;
|
|
1531 }
|
|
1532
|
|
1533 //----------------------------align_initial_loop_index---------------------------
|
|
1534 // Adjust pre-loop limit so that in main loop, a load/store reference
|
|
1535 // to align_to_ref will be a position zero in the vector.
|
|
1536 // (iv + k) mod vector_align == 0
|
|
1537 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
|
|
1538 CountedLoopNode *main_head = lp()->as_CountedLoop();
|
|
1539 assert(main_head->is_main_loop(), "");
|
|
1540 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
|
|
1541 assert(pre_end != NULL, "");
|
|
1542 Node *pre_opaq1 = pre_end->limit();
|
|
1543 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
|
|
1544 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
|
|
1545 Node *pre_limit = pre_opaq->in(1);
|
|
1546
|
|
1547 // Where we put new limit calculations
|
|
1548 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
|
|
1549
|
|
1550 // Ensure the original loop limit is available from the
|
|
1551 // pre-loop Opaque1 node.
|
|
1552 Node *orig_limit = pre_opaq->original_loop_limit();
|
|
1553 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
|
|
1554
|
|
1555 SWPointer align_to_ref_p(align_to_ref, this);
|
|
1556
|
|
1557 // Let l0 == original pre_limit, l == new pre_limit, V == v_align
|
|
1558 //
|
|
1559 // For stride > 0
|
|
1560 // Need l such that l > l0 && (l+k)%V == 0
|
|
1561 // Find n such that l = (l0 + n)
|
|
1562 // (l0 + n + k) % V == 0
|
|
1563 // n = [V - (l0 + k)%V]%V
|
|
1564 // new limit = l0 + [V - (l0 + k)%V]%V
|
|
1565 // For stride < 0
|
|
1566 // Need l such that l < l0 && (l+k)%V == 0
|
|
1567 // Find n such that l = (l0 - n)
|
|
1568 // (l0 - n + k) % V == 0
|
|
1569 // n = (l0 + k)%V
|
|
1570 // new limit = l0 - (l0 + k)%V
|
|
1571
|
|
1572 int elt_size = align_to_ref_p.memory_size();
|
|
1573 int v_align = vector_width_in_bytes() / elt_size;
|
|
1574 int k = align_to_ref_p.offset_in_bytes() / elt_size;
|
|
1575
|
|
1576 Node *kn = _igvn.intcon(k);
|
|
1577 Node *limk = new (_phase->C, 3) AddINode(pre_limit, kn);
|
|
1578 _phase->_igvn.register_new_node_with_optimizer(limk);
|
|
1579 _phase->set_ctrl(limk, pre_ctrl);
|
|
1580 if (align_to_ref_p.invar() != NULL) {
|
|
1581 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
|
|
1582 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
|
|
1583 _phase->_igvn.register_new_node_with_optimizer(aref);
|
|
1584 _phase->set_ctrl(aref, pre_ctrl);
|
|
1585 if (!align_to_ref_p.negate_invar()) {
|
|
1586 limk = new (_phase->C, 3) AddINode(limk, aref);
|
|
1587 } else {
|
|
1588 limk = new (_phase->C, 3) SubINode(limk, aref);
|
|
1589 }
|
|
1590 _phase->_igvn.register_new_node_with_optimizer(limk);
|
|
1591 _phase->set_ctrl(limk, pre_ctrl);
|
|
1592 }
|
|
1593 Node* va_msk = _igvn.intcon(v_align - 1);
|
|
1594 Node* n = new (_phase->C, 3) AndINode(limk, va_msk);
|
|
1595 _phase->_igvn.register_new_node_with_optimizer(n);
|
|
1596 _phase->set_ctrl(n, pre_ctrl);
|
|
1597 Node* newlim;
|
|
1598 if (iv_stride() > 0) {
|
|
1599 Node* va = _igvn.intcon(v_align);
|
|
1600 Node* adj = new (_phase->C, 3) SubINode(va, n);
|
|
1601 _phase->_igvn.register_new_node_with_optimizer(adj);
|
|
1602 _phase->set_ctrl(adj, pre_ctrl);
|
|
1603 Node* adj2 = new (_phase->C, 3) AndINode(adj, va_msk);
|
|
1604 _phase->_igvn.register_new_node_with_optimizer(adj2);
|
|
1605 _phase->set_ctrl(adj2, pre_ctrl);
|
|
1606 newlim = new (_phase->C, 3) AddINode(pre_limit, adj2);
|
|
1607 } else {
|
|
1608 newlim = new (_phase->C, 3) SubINode(pre_limit, n);
|
|
1609 }
|
|
1610 _phase->_igvn.register_new_node_with_optimizer(newlim);
|
|
1611 _phase->set_ctrl(newlim, pre_ctrl);
|
|
1612 Node* constrained =
|
|
1613 (iv_stride() > 0) ? (Node*) new (_phase->C,3) MinINode(newlim, orig_limit)
|
|
1614 : (Node*) new (_phase->C,3) MaxINode(newlim, orig_limit);
|
|
1615 _phase->_igvn.register_new_node_with_optimizer(constrained);
|
|
1616 _phase->set_ctrl(constrained, pre_ctrl);
|
|
1617 _igvn.hash_delete(pre_opaq);
|
|
1618 pre_opaq->set_req(1, constrained);
|
|
1619 }
|
|
1620
|
|
1621 //----------------------------get_pre_loop_end---------------------------
|
|
1622 // Find pre loop end from main loop. Returns null if none.
|
|
1623 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
|
|
1624 Node *ctrl = cl->in(LoopNode::EntryControl);
|
|
1625 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
|
|
1626 Node *iffm = ctrl->in(0);
|
|
1627 if (!iffm->is_If()) return NULL;
|
|
1628 Node *p_f = iffm->in(0);
|
|
1629 if (!p_f->is_IfFalse()) return NULL;
|
|
1630 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
|
|
1631 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
|
|
1632 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
|
|
1633 return pre_end;
|
|
1634 }
|
|
1635
|
|
1636
|
|
1637 //------------------------------init---------------------------
|
|
1638 void SuperWord::init() {
|
|
1639 _dg.init();
|
|
1640 _packset.clear();
|
|
1641 _disjoint_ptrs.clear();
|
|
1642 _block.clear();
|
|
1643 _data_entry.clear();
|
|
1644 _mem_slice_head.clear();
|
|
1645 _mem_slice_tail.clear();
|
|
1646 _node_info.clear();
|
|
1647 _align_to_ref = NULL;
|
|
1648 _lpt = NULL;
|
|
1649 _lp = NULL;
|
|
1650 _bb = NULL;
|
|
1651 _iv = NULL;
|
|
1652 }
|
|
1653
|
|
1654 //------------------------------print_packset---------------------------
|
|
1655 void SuperWord::print_packset() {
|
|
1656 #ifndef PRODUCT
|
|
1657 tty->print_cr("packset");
|
|
1658 for (int i = 0; i < _packset.length(); i++) {
|
|
1659 tty->print_cr("Pack: %d", i);
|
|
1660 Node_List* p = _packset.at(i);
|
|
1661 print_pack(p);
|
|
1662 }
|
|
1663 #endif
|
|
1664 }
|
|
1665
|
|
1666 //------------------------------print_pack---------------------------
|
|
1667 void SuperWord::print_pack(Node_List* p) {
|
|
1668 for (uint i = 0; i < p->size(); i++) {
|
|
1669 print_stmt(p->at(i));
|
|
1670 }
|
|
1671 }
|
|
1672
|
|
1673 //------------------------------print_bb---------------------------
|
|
1674 void SuperWord::print_bb() {
|
|
1675 #ifndef PRODUCT
|
|
1676 tty->print_cr("\nBlock");
|
|
1677 for (int i = 0; i < _block.length(); i++) {
|
|
1678 Node* n = _block.at(i);
|
|
1679 tty->print("%d ", i);
|
|
1680 if (n) {
|
|
1681 n->dump();
|
|
1682 }
|
|
1683 }
|
|
1684 #endif
|
|
1685 }
|
|
1686
|
|
1687 //------------------------------print_stmt---------------------------
|
|
1688 void SuperWord::print_stmt(Node* s) {
|
|
1689 #ifndef PRODUCT
|
|
1690 tty->print(" align: %d \t", alignment(s));
|
|
1691 s->dump();
|
|
1692 #endif
|
|
1693 }
|
|
1694
|
|
1695 //------------------------------blank---------------------------
|
|
1696 char* SuperWord::blank(uint depth) {
|
|
1697 static char blanks[101];
|
|
1698 assert(depth < 101, "too deep");
|
|
1699 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
|
|
1700 blanks[depth] = '\0';
|
|
1701 return blanks;
|
|
1702 }
|
|
1703
|
|
1704
|
|
1705 //==============================SWPointer===========================
|
|
1706
|
|
1707 //----------------------------SWPointer------------------------
|
|
1708 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
|
|
1709 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
|
|
1710 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
|
|
1711
|
|
1712 Node* adr = mem->in(MemNode::Address);
|
|
1713 if (!adr->is_AddP()) {
|
|
1714 assert(!valid(), "too complex");
|
|
1715 return;
|
|
1716 }
|
|
1717 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
|
|
1718 Node* base = adr->in(AddPNode::Base);
|
|
1719 for (int i = 0; i < 3; i++) {
|
|
1720 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
|
|
1721 assert(!valid(), "too complex");
|
|
1722 return;
|
|
1723 }
|
|
1724 adr = adr->in(AddPNode::Address);
|
|
1725 if (base == adr || !adr->is_AddP()) {
|
|
1726 break; // stop looking at addp's
|
|
1727 }
|
|
1728 }
|
|
1729 _base = base;
|
|
1730 _adr = adr;
|
|
1731 assert(valid(), "Usable");
|
|
1732 }
|
|
1733
|
|
1734 // Following is used to create a temporary object during
|
|
1735 // the pattern match of an address expression.
|
|
1736 SWPointer::SWPointer(SWPointer* p) :
|
|
1737 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
|
|
1738 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
|
|
1739
|
|
1740 //------------------------scaled_iv_plus_offset--------------------
|
|
1741 // Match: k*iv + offset
|
|
1742 // where: k is a constant that maybe zero, and
|
|
1743 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
|
|
1744 bool SWPointer::scaled_iv_plus_offset(Node* n) {
|
|
1745 if (scaled_iv(n)) {
|
|
1746 return true;
|
|
1747 }
|
|
1748 if (offset_plus_k(n)) {
|
|
1749 return true;
|
|
1750 }
|
|
1751 int opc = n->Opcode();
|
|
1752 if (opc == Op_AddI) {
|
|
1753 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
|
|
1754 return true;
|
|
1755 }
|
|
1756 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
|
|
1757 return true;
|
|
1758 }
|
|
1759 } else if (opc == Op_SubI) {
|
|
1760 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
|
|
1761 return true;
|
|
1762 }
|
|
1763 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
|
|
1764 _scale *= -1;
|
|
1765 return true;
|
|
1766 }
|
|
1767 }
|
|
1768 return false;
|
|
1769 }
|
|
1770
|
|
1771 //----------------------------scaled_iv------------------------
|
|
1772 // Match: k*iv where k is a constant that's not zero
|
|
1773 bool SWPointer::scaled_iv(Node* n) {
|
|
1774 if (_scale != 0) {
|
|
1775 return false; // already found a scale
|
|
1776 }
|
|
1777 if (n == iv()) {
|
|
1778 _scale = 1;
|
|
1779 return true;
|
|
1780 }
|
|
1781 int opc = n->Opcode();
|
|
1782 if (opc == Op_MulI) {
|
|
1783 if (n->in(1) == iv() && n->in(2)->is_Con()) {
|
|
1784 _scale = n->in(2)->get_int();
|
|
1785 return true;
|
|
1786 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
|
|
1787 _scale = n->in(1)->get_int();
|
|
1788 return true;
|
|
1789 }
|
|
1790 } else if (opc == Op_LShiftI) {
|
|
1791 if (n->in(1) == iv() && n->in(2)->is_Con()) {
|
|
1792 _scale = 1 << n->in(2)->get_int();
|
|
1793 return true;
|
|
1794 }
|
|
1795 } else if (opc == Op_ConvI2L) {
|
|
1796 if (scaled_iv_plus_offset(n->in(1))) {
|
|
1797 return true;
|
|
1798 }
|
|
1799 } else if (opc == Op_LShiftL) {
|
|
1800 if (!has_iv() && _invar == NULL) {
|
|
1801 // Need to preserve the current _offset value, so
|
|
1802 // create a temporary object for this expression subtree.
|
|
1803 // Hacky, so should re-engineer the address pattern match.
|
|
1804 SWPointer tmp(this);
|
|
1805 if (tmp.scaled_iv_plus_offset(n->in(1))) {
|
|
1806 if (tmp._invar == NULL) {
|
|
1807 int mult = 1 << n->in(2)->get_int();
|
|
1808 _scale = tmp._scale * mult;
|
|
1809 _offset += tmp._offset * mult;
|
|
1810 return true;
|
|
1811 }
|
|
1812 }
|
|
1813 }
|
|
1814 }
|
|
1815 return false;
|
|
1816 }
|
|
1817
|
|
1818 //----------------------------offset_plus_k------------------------
|
|
1819 // Match: offset is (k [+/- invariant])
|
|
1820 // where k maybe zero and invariant is optional, but not both.
|
|
1821 bool SWPointer::offset_plus_k(Node* n, bool negate) {
|
|
1822 int opc = n->Opcode();
|
|
1823 if (opc == Op_ConI) {
|
|
1824 _offset += negate ? -(n->get_int()) : n->get_int();
|
|
1825 return true;
|
|
1826 } else if (opc == Op_ConL) {
|
|
1827 // Okay if value fits into an int
|
|
1828 const TypeLong* t = n->find_long_type();
|
|
1829 if (t->higher_equal(TypeLong::INT)) {
|
|
1830 jlong loff = n->get_long();
|
|
1831 jint off = (jint)loff;
|
|
1832 _offset += negate ? -off : loff;
|
|
1833 return true;
|
|
1834 }
|
|
1835 return false;
|
|
1836 }
|
|
1837 if (_invar != NULL) return false; // already have an invariant
|
|
1838 if (opc == Op_AddI) {
|
|
1839 if (n->in(2)->is_Con() && invariant(n->in(1))) {
|
|
1840 _negate_invar = negate;
|
|
1841 _invar = n->in(1);
|
|
1842 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
|
|
1843 return true;
|
|
1844 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
|
|
1845 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
|
|
1846 _negate_invar = negate;
|
|
1847 _invar = n->in(2);
|
|
1848 return true;
|
|
1849 }
|
|
1850 }
|
|
1851 if (opc == Op_SubI) {
|
|
1852 if (n->in(2)->is_Con() && invariant(n->in(1))) {
|
|
1853 _negate_invar = negate;
|
|
1854 _invar = n->in(1);
|
|
1855 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
|
|
1856 return true;
|
|
1857 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
|
|
1858 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
|
|
1859 _negate_invar = !negate;
|
|
1860 _invar = n->in(2);
|
|
1861 return true;
|
|
1862 }
|
|
1863 }
|
|
1864 if (invariant(n)) {
|
|
1865 _negate_invar = negate;
|
|
1866 _invar = n;
|
|
1867 return true;
|
|
1868 }
|
|
1869 return false;
|
|
1870 }
|
|
1871
|
|
1872 //----------------------------print------------------------
|
|
1873 void SWPointer::print() {
|
|
1874 #ifndef PRODUCT
|
|
1875 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
|
|
1876 _base != NULL ? _base->_idx : 0,
|
|
1877 _adr != NULL ? _adr->_idx : 0,
|
|
1878 _scale, _offset,
|
|
1879 _negate_invar?'-':'+',
|
|
1880 _invar != NULL ? _invar->_idx : 0);
|
|
1881 #endif
|
|
1882 }
|
|
1883
|
|
1884 // ========================= OrderedPair =====================
|
|
1885
|
|
1886 const OrderedPair OrderedPair::initial;
|
|
1887
|
|
1888 // ========================= SWNodeInfo =====================
|
|
1889
|
|
1890 const SWNodeInfo SWNodeInfo::initial;
|
|
1891
|
|
1892
|
|
1893 // ============================ DepGraph ===========================
|
|
1894
|
|
1895 //------------------------------make_node---------------------------
|
|
1896 // Make a new dependence graph node for an ideal node.
|
|
1897 DepMem* DepGraph::make_node(Node* node) {
|
|
1898 DepMem* m = new (_arena) DepMem(node);
|
|
1899 if (node != NULL) {
|
|
1900 assert(_map.at_grow(node->_idx) == NULL, "one init only");
|
|
1901 _map.at_put_grow(node->_idx, m);
|
|
1902 }
|
|
1903 return m;
|
|
1904 }
|
|
1905
|
|
1906 //------------------------------make_edge---------------------------
|
|
1907 // Make a new dependence graph edge from dpred -> dsucc
|
|
1908 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
|
|
1909 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
|
|
1910 dpred->set_out_head(e);
|
|
1911 dsucc->set_in_head(e);
|
|
1912 return e;
|
|
1913 }
|
|
1914
|
|
1915 // ========================== DepMem ========================
|
|
1916
|
|
1917 //------------------------------in_cnt---------------------------
|
|
1918 int DepMem::in_cnt() {
|
|
1919 int ct = 0;
|
|
1920 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
|
|
1921 return ct;
|
|
1922 }
|
|
1923
|
|
1924 //------------------------------out_cnt---------------------------
|
|
1925 int DepMem::out_cnt() {
|
|
1926 int ct = 0;
|
|
1927 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
|
|
1928 return ct;
|
|
1929 }
|
|
1930
|
|
1931 //------------------------------print-----------------------------
|
|
1932 void DepMem::print() {
|
|
1933 #ifndef PRODUCT
|
|
1934 tty->print(" DepNode %d (", _node->_idx);
|
|
1935 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
|
|
1936 Node* pred = p->pred()->node();
|
|
1937 tty->print(" %d", pred != NULL ? pred->_idx : 0);
|
|
1938 }
|
|
1939 tty->print(") [");
|
|
1940 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
|
|
1941 Node* succ = s->succ()->node();
|
|
1942 tty->print(" %d", succ != NULL ? succ->_idx : 0);
|
|
1943 }
|
|
1944 tty->print_cr(" ]");
|
|
1945 #endif
|
|
1946 }
|
|
1947
|
|
1948 // =========================== DepEdge =========================
|
|
1949
|
|
1950 //------------------------------DepPreds---------------------------
|
|
1951 void DepEdge::print() {
|
|
1952 #ifndef PRODUCT
|
|
1953 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
|
|
1954 #endif
|
|
1955 }
|
|
1956
|
|
1957 // =========================== DepPreds =========================
|
|
1958 // Iterator over predecessor edges in the dependence graph.
|
|
1959
|
|
1960 //------------------------------DepPreds---------------------------
|
|
1961 DepPreds::DepPreds(Node* n, DepGraph& dg) {
|
|
1962 _n = n;
|
|
1963 _done = false;
|
|
1964 if (_n->is_Store() || _n->is_Load()) {
|
|
1965 _next_idx = MemNode::Address;
|
|
1966 _end_idx = n->req();
|
|
1967 _dep_next = dg.dep(_n)->in_head();
|
|
1968 } else if (_n->is_Mem()) {
|
|
1969 _next_idx = 0;
|
|
1970 _end_idx = 0;
|
|
1971 _dep_next = dg.dep(_n)->in_head();
|
|
1972 } else {
|
|
1973 _next_idx = 1;
|
|
1974 _end_idx = _n->req();
|
|
1975 _dep_next = NULL;
|
|
1976 }
|
|
1977 next();
|
|
1978 }
|
|
1979
|
|
1980 //------------------------------next---------------------------
|
|
1981 void DepPreds::next() {
|
|
1982 if (_dep_next != NULL) {
|
|
1983 _current = _dep_next->pred()->node();
|
|
1984 _dep_next = _dep_next->next_in();
|
|
1985 } else if (_next_idx < _end_idx) {
|
|
1986 _current = _n->in(_next_idx++);
|
|
1987 } else {
|
|
1988 _done = true;
|
|
1989 }
|
|
1990 }
|
|
1991
|
|
1992 // =========================== DepSuccs =========================
|
|
1993 // Iterator over successor edges in the dependence graph.
|
|
1994
|
|
1995 //------------------------------DepSuccs---------------------------
|
|
1996 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
|
|
1997 _n = n;
|
|
1998 _done = false;
|
|
1999 if (_n->is_Load()) {
|
|
2000 _next_idx = 0;
|
|
2001 _end_idx = _n->outcnt();
|
|
2002 _dep_next = dg.dep(_n)->out_head();
|
|
2003 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
|
|
2004 _next_idx = 0;
|
|
2005 _end_idx = 0;
|
|
2006 _dep_next = dg.dep(_n)->out_head();
|
|
2007 } else {
|
|
2008 _next_idx = 0;
|
|
2009 _end_idx = _n->outcnt();
|
|
2010 _dep_next = NULL;
|
|
2011 }
|
|
2012 next();
|
|
2013 }
|
|
2014
|
|
2015 //-------------------------------next---------------------------
|
|
2016 void DepSuccs::next() {
|
|
2017 if (_dep_next != NULL) {
|
|
2018 _current = _dep_next->succ()->node();
|
|
2019 _dep_next = _dep_next->next_out();
|
|
2020 } else if (_next_idx < _end_idx) {
|
|
2021 _current = _n->raw_out(_next_idx++);
|
|
2022 } else {
|
|
2023 _done = true;
|
|
2024 }
|
|
2025 }
|