Print this page
rev 1839 : 6961690: load oops from constant table on SPARC
Summary: oops should be loaded from the constant table of an nmethod instead of materializing them with a long code sequence.
Reviewed-by:
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/opto/gcm.cpp
+++ new/src/share/vm/opto/gcm.cpp
1 1 /*
2 - * Copyright (c) 1997, 2009, Oracle and/or its affiliates. All rights reserved.
2 + * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 // Portions of code courtesy of Clifford Click
26 26
27 27 // Optimization - Graph Style
28 28
29 29 #include "incls/_precompiled.incl"
30 30 #include "incls/_gcm.cpp.incl"
31 31
32 32 // To avoid float value underflow
33 33 #define MIN_BLOCK_FREQUENCY 1.e-35f
34 34
35 35 //----------------------------schedule_node_into_block-------------------------
36 36 // Insert node n into block b. Look for projections of n and make sure they
37 37 // are in b also.
38 38 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
39 39 // Set basic block of n, Add n to b,
40 40 _bbs.map(n->_idx, b);
41 41 b->add_inst(n);
42 42
43 43 // After Matching, nearly any old Node may have projections trailing it.
44 44 // These are usually machine-dependent flags. In any case, they might
45 45 // float to another block below this one. Move them up.
46 46 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
47 47 Node* use = n->fast_out(i);
48 48 if (use->is_Proj()) {
49 49 Block* buse = _bbs[use->_idx];
50 50 if (buse != b) { // In wrong block?
51 51 if (buse != NULL)
52 52 buse->find_remove(use); // Remove from wrong block
53 53 _bbs.map(use->_idx, b); // Re-insert in this block
54 54 b->add_inst(use);
55 55 }
56 56 }
57 57 }
58 58 }
|
↓ open down ↓ |
46 lines elided |
↑ open up ↑ |
59 59
60 60 //----------------------------replace_block_proj_ctrl-------------------------
61 61 // Nodes that have is_block_proj() nodes as their control need to use
62 62 // the appropriate Region for their actual block as their control since
63 63 // the projection will be in a predecessor block.
64 64 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
65 65 const Node *in0 = n->in(0);
66 66 assert(in0 != NULL, "Only control-dependent");
67 67 const Node *p = in0->is_block_proj();
68 68 if (p != NULL && p != n) { // Control from a block projection?
69 - assert(!n->pinned() || n->is_SafePointScalarObject(), "only SafePointScalarObject pinned node is expected here");
69 + assert(!n->pinned() || n->is_MachConstantBase() || n->is_SafePointScalarObject(), "only pinned MachConstantBase or SafePointScalarObject node is expected here");
70 70 // Find trailing Region
71 71 Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
72 72 uint j = 0;
73 73 if (pb->_num_succs != 1) { // More then 1 successor?
74 74 // Search for successor
75 75 uint max = pb->_nodes.size();
76 76 assert( max > 1, "" );
77 77 uint start = max - pb->_num_succs;
78 78 // Find which output path belongs to projection
79 79 for (j = start; j < max; j++) {
80 80 if( pb->_nodes[j] == in0 )
81 81 break;
82 82 }
83 83 assert( j < max, "must find" );
84 84 // Change control to match head of successor basic block
85 85 j -= start;
86 86 }
87 87 n->set_req(0, pb->_succs[j]->head());
88 88 }
89 89 }
90 90
91 91
92 92 //------------------------------schedule_pinned_nodes--------------------------
93 93 // Set the basic block for Nodes pinned into blocks
94 94 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
95 95 // Allocate node stack of size C->unique()+8 to avoid frequent realloc
96 96 GrowableArray <Node *> spstack(C->unique()+8);
97 97 spstack.push(_root);
98 98 while ( spstack.is_nonempty() ) {
99 99 Node *n = spstack.pop();
100 100 if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
101 101 if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down!
102 102 assert( n->in(0), "pinned Node must have Control" );
103 103 // Before setting block replace block_proj control edge
104 104 replace_block_proj_ctrl(n);
105 105 Node *input = n->in(0);
106 106 while( !input->is_block_start() )
107 107 input = input->in(0);
108 108 Block *b = _bbs[input->_idx]; // Basic block of controlling input
109 109 schedule_node_into_block(n, b);
110 110 }
111 111 for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs
112 112 if( n->in(i) != NULL )
113 113 spstack.push(n->in(i));
114 114 }
115 115 }
116 116 }
117 117 }
118 118
119 119 #ifdef ASSERT
120 120 // Assert that new input b2 is dominated by all previous inputs.
121 121 // Check this by by seeing that it is dominated by b1, the deepest
122 122 // input observed until b2.
123 123 static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) {
124 124 if (b1 == NULL) return;
125 125 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
126 126 Block* tmp = b2;
127 127 while (tmp != b1 && tmp != NULL) {
128 128 tmp = tmp->_idom;
129 129 }
130 130 if (tmp != b1) {
131 131 // Detected an unschedulable graph. Print some nice stuff and die.
132 132 tty->print_cr("!!! Unschedulable graph !!!");
133 133 for (uint j=0; j<n->len(); j++) { // For all inputs
134 134 Node* inn = n->in(j); // Get input
135 135 if (inn == NULL) continue; // Ignore NULL, missing inputs
136 136 Block* inb = bbs[inn->_idx];
137 137 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
138 138 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
139 139 inn->dump();
140 140 }
141 141 tty->print("Failing node: ");
142 142 n->dump();
143 143 assert(false, "unscheduable graph");
144 144 }
145 145 }
146 146 #endif
147 147
148 148 static Block* find_deepest_input(Node* n, Block_Array &bbs) {
149 149 // Find the last input dominated by all other inputs.
150 150 Block* deepb = NULL; // Deepest block so far
151 151 int deepb_dom_depth = 0;
152 152 for (uint k = 0; k < n->len(); k++) { // For all inputs
153 153 Node* inn = n->in(k); // Get input
154 154 if (inn == NULL) continue; // Ignore NULL, missing inputs
155 155 Block* inb = bbs[inn->_idx];
156 156 assert(inb != NULL, "must already have scheduled this input");
157 157 if (deepb_dom_depth < (int) inb->_dom_depth) {
158 158 // The new inb must be dominated by the previous deepb.
159 159 // The various inputs must be linearly ordered in the dom
160 160 // tree, or else there will not be a unique deepest block.
161 161 DEBUG_ONLY(assert_dom(deepb, inb, n, bbs));
162 162 deepb = inb; // Save deepest block
163 163 deepb_dom_depth = deepb->_dom_depth;
164 164 }
165 165 }
166 166 assert(deepb != NULL, "must be at least one input to n");
167 167 return deepb;
168 168 }
169 169
170 170
171 171 //------------------------------schedule_early---------------------------------
172 172 // Find the earliest Block any instruction can be placed in. Some instructions
173 173 // are pinned into Blocks. Unpinned instructions can appear in last block in
174 174 // which all their inputs occur.
175 175 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
176 176 // Allocate stack with enough space to avoid frequent realloc
177 177 Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
178 178 // roots.push(_root); _root will be processed among C->top() inputs
179 179 roots.push(C->top());
180 180 visited.set(C->top()->_idx);
181 181
182 182 while (roots.size() != 0) {
183 183 // Use local variables nstack_top_n & nstack_top_i to cache values
184 184 // on stack's top.
185 185 Node *nstack_top_n = roots.pop();
186 186 uint nstack_top_i = 0;
187 187 //while_nstack_nonempty:
188 188 while (true) {
189 189 // Get parent node and next input's index from stack's top.
190 190 Node *n = nstack_top_n;
191 191 uint i = nstack_top_i;
192 192
193 193 if (i == 0) {
194 194 // Fixup some control. Constants without control get attached
195 195 // to root and nodes that use is_block_proj() nodes should be attached
196 196 // to the region that starts their block.
197 197 const Node *in0 = n->in(0);
198 198 if (in0 != NULL) { // Control-dependent?
199 199 replace_block_proj_ctrl(n);
200 200 } else { // n->in(0) == NULL
201 201 if (n->req() == 1) { // This guy is a constant with NO inputs?
202 202 n->set_req(0, _root);
203 203 }
204 204 }
205 205 }
206 206
207 207 // First, visit all inputs and force them to get a block. If an
208 208 // input is already in a block we quit following inputs (to avoid
209 209 // cycles). Instead we put that Node on a worklist to be handled
210 210 // later (since IT'S inputs may not have a block yet).
211 211 bool done = true; // Assume all n's inputs will be processed
212 212 while (i < n->len()) { // For all inputs
213 213 Node *in = n->in(i); // Get input
214 214 ++i;
215 215 if (in == NULL) continue; // Ignore NULL, missing inputs
216 216 int is_visited = visited.test_set(in->_idx);
217 217 if (!_bbs.lookup(in->_idx)) { // Missing block selection?
218 218 if (is_visited) {
219 219 // assert( !visited.test(in->_idx), "did not schedule early" );
220 220 return false;
221 221 }
222 222 nstack.push(n, i); // Save parent node and next input's index.
223 223 nstack_top_n = in; // Process current input now.
224 224 nstack_top_i = 0;
225 225 done = false; // Not all n's inputs processed.
226 226 break; // continue while_nstack_nonempty;
227 227 } else if (!is_visited) { // Input not yet visited?
228 228 roots.push(in); // Visit this guy later, using worklist
229 229 }
230 230 }
231 231 if (done) {
232 232 // All of n's inputs have been processed, complete post-processing.
233 233
234 234 // Some instructions are pinned into a block. These include Region,
235 235 // Phi, Start, Return, and other control-dependent instructions and
236 236 // any projections which depend on them.
237 237 if (!n->pinned()) {
238 238 // Set earliest legal block.
239 239 _bbs.map(n->_idx, find_deepest_input(n, _bbs));
240 240 } else {
241 241 assert(_bbs[n->_idx] == _bbs[n->in(0)->_idx], "Pinned Node should be at the same block as its control edge");
242 242 }
243 243
244 244 if (nstack.is_empty()) {
245 245 // Finished all nodes on stack.
246 246 // Process next node on the worklist 'roots'.
247 247 break;
248 248 }
249 249 // Get saved parent node and next input's index.
250 250 nstack_top_n = nstack.node();
251 251 nstack_top_i = nstack.index();
252 252 nstack.pop();
253 253 } // if (done)
254 254 } // while (true)
255 255 } // while (roots.size() != 0)
256 256 return true;
257 257 }
258 258
259 259 //------------------------------dom_lca----------------------------------------
260 260 // Find least common ancestor in dominator tree
261 261 // LCA is a current notion of LCA, to be raised above 'this'.
262 262 // As a convenient boundary condition, return 'this' if LCA is NULL.
263 263 // Find the LCA of those two nodes.
264 264 Block* Block::dom_lca(Block* LCA) {
265 265 if (LCA == NULL || LCA == this) return this;
266 266
267 267 Block* anc = this;
268 268 while (anc->_dom_depth > LCA->_dom_depth)
269 269 anc = anc->_idom; // Walk up till anc is as high as LCA
270 270
271 271 while (LCA->_dom_depth > anc->_dom_depth)
272 272 LCA = LCA->_idom; // Walk up till LCA is as high as anc
273 273
274 274 while (LCA != anc) { // Walk both up till they are the same
275 275 LCA = LCA->_idom;
276 276 anc = anc->_idom;
277 277 }
278 278
279 279 return LCA;
280 280 }
281 281
282 282 //--------------------------raise_LCA_above_use--------------------------------
283 283 // We are placing a definition, and have been given a def->use edge.
284 284 // The definition must dominate the use, so move the LCA upward in the
285 285 // dominator tree to dominate the use. If the use is a phi, adjust
286 286 // the LCA only with the phi input paths which actually use this def.
287 287 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) {
288 288 Block* buse = bbs[use->_idx];
289 289 if (buse == NULL) return LCA; // Unused killing Projs have no use block
290 290 if (!use->is_Phi()) return buse->dom_lca(LCA);
291 291 uint pmax = use->req(); // Number of Phi inputs
292 292 // Why does not this loop just break after finding the matching input to
293 293 // the Phi? Well...it's like this. I do not have true def-use/use-def
294 294 // chains. Means I cannot distinguish, from the def-use direction, which
295 295 // of many use-defs lead from the same use to the same def. That is, this
296 296 // Phi might have several uses of the same def. Each use appears in a
297 297 // different predecessor block. But when I enter here, I cannot distinguish
298 298 // which use-def edge I should find the predecessor block for. So I find
299 299 // them all. Means I do a little extra work if a Phi uses the same value
300 300 // more than once.
301 301 for (uint j=1; j<pmax; j++) { // For all inputs
302 302 if (use->in(j) == def) { // Found matching input?
303 303 Block* pred = bbs[buse->pred(j)->_idx];
304 304 LCA = pred->dom_lca(LCA);
305 305 }
306 306 }
307 307 return LCA;
308 308 }
309 309
310 310 //----------------------------raise_LCA_above_marks----------------------------
311 311 // Return a new LCA that dominates LCA and any of its marked predecessors.
312 312 // Search all my parents up to 'early' (exclusive), looking for predecessors
313 313 // which are marked with the given index. Return the LCA (in the dom tree)
314 314 // of all marked blocks. If there are none marked, return the original
315 315 // LCA.
316 316 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark,
317 317 Block* early, Block_Array &bbs) {
318 318 Block_List worklist;
319 319 worklist.push(LCA);
320 320 while (worklist.size() > 0) {
321 321 Block* mid = worklist.pop();
322 322 if (mid == early) continue; // stop searching here
323 323
324 324 // Test and set the visited bit.
325 325 if (mid->raise_LCA_visited() == mark) continue; // already visited
326 326
327 327 // Don't process the current LCA, otherwise the search may terminate early
328 328 if (mid != LCA && mid->raise_LCA_mark() == mark) {
329 329 // Raise the LCA.
330 330 LCA = mid->dom_lca(LCA);
331 331 if (LCA == early) break; // stop searching everywhere
332 332 assert(early->dominates(LCA), "early is high enough");
333 333 // Resume searching at that point, skipping intermediate levels.
334 334 worklist.push(LCA);
335 335 if (LCA == mid)
336 336 continue; // Don't mark as visited to avoid early termination.
337 337 } else {
338 338 // Keep searching through this block's predecessors.
339 339 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
340 340 Block* mid_parent = bbs[ mid->pred(j)->_idx ];
341 341 worklist.push(mid_parent);
342 342 }
343 343 }
344 344 mid->set_raise_LCA_visited(mark);
345 345 }
346 346 return LCA;
347 347 }
348 348
349 349 //--------------------------memory_early_block--------------------------------
350 350 // This is a variation of find_deepest_input, the heart of schedule_early.
351 351 // Find the "early" block for a load, if we considered only memory and
352 352 // address inputs, that is, if other data inputs were ignored.
353 353 //
354 354 // Because a subset of edges are considered, the resulting block will
355 355 // be earlier (at a shallower dom_depth) than the true schedule_early
356 356 // point of the node. We compute this earlier block as a more permissive
357 357 // site for anti-dependency insertion, but only if subsume_loads is enabled.
358 358 static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) {
359 359 Node* base;
360 360 Node* index;
361 361 Node* store = load->in(MemNode::Memory);
362 362 load->as_Mach()->memory_inputs(base, index);
363 363
364 364 assert(base != NodeSentinel && index != NodeSentinel,
365 365 "unexpected base/index inputs");
366 366
367 367 Node* mem_inputs[4];
368 368 int mem_inputs_length = 0;
369 369 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
370 370 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
371 371 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
372 372
373 373 // In the comparision below, add one to account for the control input,
374 374 // which may be null, but always takes up a spot in the in array.
375 375 if (mem_inputs_length + 1 < (int) load->req()) {
376 376 // This "load" has more inputs than just the memory, base and index inputs.
377 377 // For purposes of checking anti-dependences, we need to start
378 378 // from the early block of only the address portion of the instruction,
379 379 // and ignore other blocks that may have factored into the wider
380 380 // schedule_early calculation.
381 381 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
382 382
383 383 Block* deepb = NULL; // Deepest block so far
384 384 int deepb_dom_depth = 0;
385 385 for (int i = 0; i < mem_inputs_length; i++) {
386 386 Block* inb = bbs[mem_inputs[i]->_idx];
387 387 if (deepb_dom_depth < (int) inb->_dom_depth) {
388 388 // The new inb must be dominated by the previous deepb.
389 389 // The various inputs must be linearly ordered in the dom
390 390 // tree, or else there will not be a unique deepest block.
391 391 DEBUG_ONLY(assert_dom(deepb, inb, load, bbs));
392 392 deepb = inb; // Save deepest block
393 393 deepb_dom_depth = deepb->_dom_depth;
394 394 }
395 395 }
396 396 early = deepb;
397 397 }
398 398
399 399 return early;
400 400 }
401 401
402 402 //--------------------------insert_anti_dependences---------------------------
403 403 // A load may need to witness memory that nearby stores can overwrite.
404 404 // For each nearby store, either insert an "anti-dependence" edge
405 405 // from the load to the store, or else move LCA upward to force the
406 406 // load to (eventually) be scheduled in a block above the store.
407 407 //
408 408 // Do not add edges to stores on distinct control-flow paths;
409 409 // only add edges to stores which might interfere.
410 410 //
411 411 // Return the (updated) LCA. There will not be any possibly interfering
412 412 // store between the load's "early block" and the updated LCA.
413 413 // Any stores in the updated LCA will have new precedence edges
414 414 // back to the load. The caller is expected to schedule the load
415 415 // in the LCA, in which case the precedence edges will make LCM
416 416 // preserve anti-dependences. The caller may also hoist the load
417 417 // above the LCA, if it is not the early block.
418 418 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
419 419 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
420 420 assert(LCA != NULL, "");
421 421 DEBUG_ONLY(Block* LCA_orig = LCA);
422 422
423 423 // Compute the alias index. Loads and stores with different alias indices
424 424 // do not need anti-dependence edges.
425 425 uint load_alias_idx = C->get_alias_index(load->adr_type());
426 426 #ifdef ASSERT
427 427 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
428 428 (PrintOpto || VerifyAliases ||
429 429 PrintMiscellaneous && (WizardMode || Verbose))) {
430 430 // Load nodes should not consume all of memory.
431 431 // Reporting a bottom type indicates a bug in adlc.
432 432 // If some particular type of node validly consumes all of memory,
433 433 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
434 434 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
435 435 load->dump(2);
436 436 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
437 437 }
438 438 #endif
439 439 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
440 440 "String compare is only known 'load' that does not conflict with any stores");
441 441 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
442 442 "String equals is a 'load' that does not conflict with any stores");
443 443 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
444 444 "String indexOf is a 'load' that does not conflict with any stores");
445 445 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
446 446 "Arrays equals is a 'load' that do not conflict with any stores");
447 447
448 448 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
449 449 // It is impossible to spoil this load by putting stores before it,
450 450 // because we know that the stores will never update the value
451 451 // which 'load' must witness.
452 452 return LCA;
453 453 }
454 454
455 455 node_idx_t load_index = load->_idx;
456 456
457 457 // Note the earliest legal placement of 'load', as determined by
458 458 // by the unique point in the dom tree where all memory effects
459 459 // and other inputs are first available. (Computed by schedule_early.)
460 460 // For normal loads, 'early' is the shallowest place (dom graph wise)
461 461 // to look for anti-deps between this load and any store.
462 462 Block* early = _bbs[load_index];
463 463
464 464 // If we are subsuming loads, compute an "early" block that only considers
465 465 // memory or address inputs. This block may be different than the
466 466 // schedule_early block in that it could be at an even shallower depth in the
467 467 // dominator tree, and allow for a broader discovery of anti-dependences.
468 468 if (C->subsume_loads()) {
469 469 early = memory_early_block(load, early, _bbs);
470 470 }
471 471
472 472 ResourceArea *area = Thread::current()->resource_area();
473 473 Node_List worklist_mem(area); // prior memory state to store
474 474 Node_List worklist_store(area); // possible-def to explore
475 475 Node_List worklist_visited(area); // visited mergemem nodes
476 476 Node_List non_early_stores(area); // all relevant stores outside of early
477 477 bool must_raise_LCA = false;
478 478
479 479 #ifdef TRACK_PHI_INPUTS
480 480 // %%% This extra checking fails because MergeMem nodes are not GVNed.
481 481 // Provide "phi_inputs" to check if every input to a PhiNode is from the
482 482 // original memory state. This indicates a PhiNode for which should not
483 483 // prevent the load from sinking. For such a block, set_raise_LCA_mark
484 484 // may be overly conservative.
485 485 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
486 486 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
487 487 #endif
488 488
489 489 // 'load' uses some memory state; look for users of the same state.
490 490 // Recurse through MergeMem nodes to the stores that use them.
491 491
492 492 // Each of these stores is a possible definition of memory
493 493 // that 'load' needs to use. We need to force 'load'
494 494 // to occur before each such store. When the store is in
495 495 // the same block as 'load', we insert an anti-dependence
496 496 // edge load->store.
497 497
498 498 // The relevant stores "nearby" the load consist of a tree rooted
499 499 // at initial_mem, with internal nodes of type MergeMem.
500 500 // Therefore, the branches visited by the worklist are of this form:
501 501 // initial_mem -> (MergeMem ->)* store
502 502 // The anti-dependence constraints apply only to the fringe of this tree.
503 503
504 504 Node* initial_mem = load->in(MemNode::Memory);
505 505 worklist_store.push(initial_mem);
506 506 worklist_visited.push(initial_mem);
507 507 worklist_mem.push(NULL);
508 508 while (worklist_store.size() > 0) {
509 509 // Examine a nearby store to see if it might interfere with our load.
510 510 Node* mem = worklist_mem.pop();
511 511 Node* store = worklist_store.pop();
512 512 uint op = store->Opcode();
513 513
514 514 // MergeMems do not directly have anti-deps.
515 515 // Treat them as internal nodes in a forward tree of memory states,
516 516 // the leaves of which are each a 'possible-def'.
517 517 if (store == initial_mem // root (exclusive) of tree we are searching
518 518 || op == Op_MergeMem // internal node of tree we are searching
519 519 ) {
520 520 mem = store; // It's not a possibly interfering store.
521 521 if (store == initial_mem)
522 522 initial_mem = NULL; // only process initial memory once
523 523
524 524 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
525 525 store = mem->fast_out(i);
526 526 if (store->is_MergeMem()) {
527 527 // Be sure we don't get into combinatorial problems.
528 528 // (Allow phis to be repeated; they can merge two relevant states.)
529 529 uint j = worklist_visited.size();
530 530 for (; j > 0; j--) {
531 531 if (worklist_visited.at(j-1) == store) break;
532 532 }
533 533 if (j > 0) continue; // already on work list; do not repeat
534 534 worklist_visited.push(store);
535 535 }
536 536 worklist_mem.push(mem);
537 537 worklist_store.push(store);
538 538 }
539 539 continue;
540 540 }
541 541
542 542 if (op == Op_MachProj || op == Op_Catch) continue;
543 543 if (store->needs_anti_dependence_check()) continue; // not really a store
544 544
545 545 // Compute the alias index. Loads and stores with different alias
546 546 // indices do not need anti-dependence edges. Wide MemBar's are
547 547 // anti-dependent on everything (except immutable memories).
548 548 const TypePtr* adr_type = store->adr_type();
549 549 if (!C->can_alias(adr_type, load_alias_idx)) continue;
550 550
551 551 // Most slow-path runtime calls do NOT modify Java memory, but
552 552 // they can block and so write Raw memory.
553 553 if (store->is_Mach()) {
554 554 MachNode* mstore = store->as_Mach();
555 555 if (load_alias_idx != Compile::AliasIdxRaw) {
556 556 // Check for call into the runtime using the Java calling
557 557 // convention (and from there into a wrapper); it has no
558 558 // _method. Can't do this optimization for Native calls because
559 559 // they CAN write to Java memory.
560 560 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
561 561 assert(mstore->is_MachSafePoint(), "");
562 562 MachSafePointNode* ms = (MachSafePointNode*) mstore;
563 563 assert(ms->is_MachCallJava(), "");
564 564 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
565 565 if (mcj->_method == NULL) {
566 566 // These runtime calls do not write to Java visible memory
567 567 // (other than Raw) and so do not require anti-dependence edges.
568 568 continue;
569 569 }
570 570 }
571 571 // Same for SafePoints: they read/write Raw but only read otherwise.
572 572 // This is basically a workaround for SafePoints only defining control
573 573 // instead of control + memory.
574 574 if (mstore->ideal_Opcode() == Op_SafePoint)
575 575 continue;
576 576 } else {
577 577 // Some raw memory, such as the load of "top" at an allocation,
578 578 // can be control dependent on the previous safepoint. See
579 579 // comments in GraphKit::allocate_heap() about control input.
580 580 // Inserting an anti-dep between such a safepoint and a use
581 581 // creates a cycle, and will cause a subsequent failure in
582 582 // local scheduling. (BugId 4919904)
583 583 // (%%% How can a control input be a safepoint and not a projection??)
584 584 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
585 585 continue;
586 586 }
587 587 }
588 588
589 589 // Identify a block that the current load must be above,
590 590 // or else observe that 'store' is all the way up in the
591 591 // earliest legal block for 'load'. In the latter case,
592 592 // immediately insert an anti-dependence edge.
593 593 Block* store_block = _bbs[store->_idx];
594 594 assert(store_block != NULL, "unused killing projections skipped above");
595 595
596 596 if (store->is_Phi()) {
597 597 // 'load' uses memory which is one (or more) of the Phi's inputs.
598 598 // It must be scheduled not before the Phi, but rather before
599 599 // each of the relevant Phi inputs.
600 600 //
601 601 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
602 602 // we mark each corresponding predecessor block and do a combined
603 603 // hoisting operation later (raise_LCA_above_marks).
604 604 //
605 605 // Do not assert(store_block != early, "Phi merging memory after access")
606 606 // PhiNode may be at start of block 'early' with backedge to 'early'
607 607 DEBUG_ONLY(bool found_match = false);
608 608 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
609 609 if (store->in(j) == mem) { // Found matching input?
610 610 DEBUG_ONLY(found_match = true);
611 611 Block* pred_block = _bbs[store_block->pred(j)->_idx];
612 612 if (pred_block != early) {
613 613 // If any predecessor of the Phi matches the load's "early block",
614 614 // we do not need a precedence edge between the Phi and 'load'
615 615 // since the load will be forced into a block preceding the Phi.
616 616 pred_block->set_raise_LCA_mark(load_index);
617 617 assert(!LCA_orig->dominates(pred_block) ||
618 618 early->dominates(pred_block), "early is high enough");
619 619 must_raise_LCA = true;
620 620 } else {
621 621 // anti-dependent upon PHI pinned below 'early', no edge needed
622 622 LCA = early; // but can not schedule below 'early'
623 623 }
624 624 }
625 625 }
626 626 assert(found_match, "no worklist bug");
627 627 #ifdef TRACK_PHI_INPUTS
628 628 #ifdef ASSERT
629 629 // This assert asks about correct handling of PhiNodes, which may not
630 630 // have all input edges directly from 'mem'. See BugId 4621264
631 631 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
632 632 // Increment by exactly one even if there are multiple copies of 'mem'
633 633 // coming into the phi, because we will run this block several times
634 634 // if there are several copies of 'mem'. (That's how DU iterators work.)
635 635 phi_inputs.at_put(store->_idx, num_mem_inputs);
636 636 assert(PhiNode::Input + num_mem_inputs < store->req(),
637 637 "Expect at least one phi input will not be from original memory state");
638 638 #endif //ASSERT
639 639 #endif //TRACK_PHI_INPUTS
640 640 } else if (store_block != early) {
641 641 // 'store' is between the current LCA and earliest possible block.
642 642 // Label its block, and decide later on how to raise the LCA
643 643 // to include the effect on LCA of this store.
644 644 // If this store's block gets chosen as the raised LCA, we
645 645 // will find him on the non_early_stores list and stick him
646 646 // with a precedence edge.
647 647 // (But, don't bother if LCA is already raised all the way.)
648 648 if (LCA != early) {
649 649 store_block->set_raise_LCA_mark(load_index);
650 650 must_raise_LCA = true;
651 651 non_early_stores.push(store);
652 652 }
653 653 } else {
654 654 // Found a possibly-interfering store in the load's 'early' block.
655 655 // This means 'load' cannot sink at all in the dominator tree.
656 656 // Add an anti-dep edge, and squeeze 'load' into the highest block.
657 657 assert(store != load->in(0), "dependence cycle found");
658 658 if (verify) {
659 659 assert(store->find_edge(load) != -1, "missing precedence edge");
660 660 } else {
661 661 store->add_prec(load);
662 662 }
663 663 LCA = early;
664 664 // This turns off the process of gathering non_early_stores.
665 665 }
666 666 }
667 667 // (Worklist is now empty; all nearby stores have been visited.)
668 668
669 669 // Finished if 'load' must be scheduled in its 'early' block.
670 670 // If we found any stores there, they have already been given
671 671 // precedence edges.
672 672 if (LCA == early) return LCA;
673 673
674 674 // We get here only if there are no possibly-interfering stores
675 675 // in the load's 'early' block. Move LCA up above all predecessors
676 676 // which contain stores we have noted.
677 677 //
678 678 // The raised LCA block can be a home to such interfering stores,
679 679 // but its predecessors must not contain any such stores.
680 680 //
681 681 // The raised LCA will be a lower bound for placing the load,
682 682 // preventing the load from sinking past any block containing
683 683 // a store that may invalidate the memory state required by 'load'.
684 684 if (must_raise_LCA)
685 685 LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs);
686 686 if (LCA == early) return LCA;
687 687
688 688 // Insert anti-dependence edges from 'load' to each store
689 689 // in the non-early LCA block.
690 690 // Mine the non_early_stores list for such stores.
691 691 if (LCA->raise_LCA_mark() == load_index) {
692 692 while (non_early_stores.size() > 0) {
693 693 Node* store = non_early_stores.pop();
694 694 Block* store_block = _bbs[store->_idx];
695 695 if (store_block == LCA) {
696 696 // add anti_dependence from store to load in its own block
697 697 assert(store != load->in(0), "dependence cycle found");
698 698 if (verify) {
699 699 assert(store->find_edge(load) != -1, "missing precedence edge");
700 700 } else {
701 701 store->add_prec(load);
702 702 }
703 703 } else {
704 704 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
705 705 // Any other stores we found must be either inside the new LCA
706 706 // or else outside the original LCA. In the latter case, they
707 707 // did not interfere with any use of 'load'.
708 708 assert(LCA->dominates(store_block)
709 709 || !LCA_orig->dominates(store_block), "no stray stores");
710 710 }
711 711 }
712 712 }
713 713
714 714 // Return the highest block containing stores; any stores
715 715 // within that block have been given anti-dependence edges.
716 716 return LCA;
717 717 }
718 718
719 719 // This class is used to iterate backwards over the nodes in the graph.
720 720
721 721 class Node_Backward_Iterator {
722 722
723 723 private:
724 724 Node_Backward_Iterator();
725 725
726 726 public:
727 727 // Constructor for the iterator
728 728 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs);
729 729
730 730 // Postincrement operator to iterate over the nodes
731 731 Node *next();
732 732
733 733 private:
734 734 VectorSet &_visited;
735 735 Node_List &_stack;
736 736 Block_Array &_bbs;
737 737 };
738 738
739 739 // Constructor for the Node_Backward_Iterator
740 740 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs )
741 741 : _visited(visited), _stack(stack), _bbs(bbs) {
742 742 // The stack should contain exactly the root
743 743 stack.clear();
744 744 stack.push(root);
745 745
746 746 // Clear the visited bits
747 747 visited.Clear();
748 748 }
749 749
750 750 // Iterator for the Node_Backward_Iterator
751 751 Node *Node_Backward_Iterator::next() {
752 752
753 753 // If the _stack is empty, then just return NULL: finished.
754 754 if ( !_stack.size() )
755 755 return NULL;
756 756
757 757 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
758 758 // made stateless, so I do not need to record the index 'i' on my _stack.
759 759 // Instead I visit all users each time, scanning for unvisited users.
760 760 // I visit unvisited not-anti-dependence users first, then anti-dependent
761 761 // children next.
762 762 Node *self = _stack.pop();
763 763
764 764 // I cycle here when I am entering a deeper level of recursion.
765 765 // The key variable 'self' was set prior to jumping here.
766 766 while( 1 ) {
767 767
768 768 _visited.set(self->_idx);
769 769
770 770 // Now schedule all uses as late as possible.
771 771 uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx;
772 772 uint src_rpo = _bbs[src]->_rpo;
773 773
774 774 // Schedule all nodes in a post-order visit
775 775 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
776 776
777 777 // Scan for unvisited nodes
778 778 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
779 779 // For all uses, schedule late
780 780 Node* n = self->fast_out(i); // Use
781 781
782 782 // Skip already visited children
783 783 if ( _visited.test(n->_idx) )
784 784 continue;
785 785
786 786 // do not traverse backward control edges
787 787 Node *use = n->is_Proj() ? n->in(0) : n;
788 788 uint use_rpo = _bbs[use->_idx]->_rpo;
789 789
790 790 if ( use_rpo < src_rpo )
791 791 continue;
792 792
793 793 // Phi nodes always precede uses in a basic block
794 794 if ( use_rpo == src_rpo && use->is_Phi() )
795 795 continue;
796 796
797 797 unvisited = n; // Found unvisited
798 798
799 799 // Check for possible-anti-dependent
800 800 if( !n->needs_anti_dependence_check() )
801 801 break; // Not visited, not anti-dep; schedule it NOW
802 802 }
803 803
804 804 // Did I find an unvisited not-anti-dependent Node?
805 805 if ( !unvisited )
806 806 break; // All done with children; post-visit 'self'
807 807
808 808 // Visit the unvisited Node. Contains the obvious push to
809 809 // indicate I'm entering a deeper level of recursion. I push the
810 810 // old state onto the _stack and set a new state and loop (recurse).
811 811 _stack.push(self);
812 812 self = unvisited;
813 813 } // End recursion loop
814 814
815 815 return self;
816 816 }
817 817
818 818 //------------------------------ComputeLatenciesBackwards----------------------
819 819 // Compute the latency of all the instructions.
820 820 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
821 821 #ifndef PRODUCT
822 822 if (trace_opto_pipelining())
823 823 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
824 824 #endif
825 825
826 826 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
827 827 Node *n;
828 828
829 829 // Walk over all the nodes from last to first
830 830 while (n = iter.next()) {
831 831 // Set the latency for the definitions of this instruction
832 832 partial_latency_of_defs(n);
833 833 }
834 834 } // end ComputeLatenciesBackwards
835 835
836 836 //------------------------------partial_latency_of_defs------------------------
837 837 // Compute the latency impact of this node on all defs. This computes
838 838 // a number that increases as we approach the beginning of the routine.
839 839 void PhaseCFG::partial_latency_of_defs(Node *n) {
840 840 // Set the latency for this instruction
841 841 #ifndef PRODUCT
842 842 if (trace_opto_pipelining()) {
843 843 tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
844 844 n->_idx, _node_latency->at_grow(n->_idx));
845 845 dump();
846 846 }
847 847 #endif
848 848
849 849 if (n->is_Proj())
850 850 n = n->in(0);
851 851
852 852 if (n->is_Root())
853 853 return;
854 854
855 855 uint nlen = n->len();
856 856 uint use_latency = _node_latency->at_grow(n->_idx);
857 857 uint use_pre_order = _bbs[n->_idx]->_pre_order;
858 858
859 859 for ( uint j=0; j<nlen; j++ ) {
860 860 Node *def = n->in(j);
861 861
862 862 if (!def || def == n)
863 863 continue;
864 864
865 865 // Walk backwards thru projections
866 866 if (def->is_Proj())
867 867 def = def->in(0);
868 868
869 869 #ifndef PRODUCT
870 870 if (trace_opto_pipelining()) {
871 871 tty->print("# in(%2d): ", j);
872 872 def->dump();
873 873 }
874 874 #endif
875 875
876 876 // If the defining block is not known, assume it is ok
877 877 Block *def_block = _bbs[def->_idx];
878 878 uint def_pre_order = def_block ? def_block->_pre_order : 0;
879 879
880 880 if ( (use_pre_order < def_pre_order) ||
881 881 (use_pre_order == def_pre_order && n->is_Phi()) )
882 882 continue;
883 883
884 884 uint delta_latency = n->latency(j);
885 885 uint current_latency = delta_latency + use_latency;
886 886
887 887 if (_node_latency->at_grow(def->_idx) < current_latency) {
888 888 _node_latency->at_put_grow(def->_idx, current_latency);
889 889 }
890 890
891 891 #ifndef PRODUCT
892 892 if (trace_opto_pipelining()) {
893 893 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
894 894 use_latency, j, delta_latency, current_latency, def->_idx,
895 895 _node_latency->at_grow(def->_idx));
896 896 }
897 897 #endif
898 898 }
899 899 }
900 900
901 901 //------------------------------latency_from_use-------------------------------
902 902 // Compute the latency of a specific use
903 903 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
904 904 // If self-reference, return no latency
905 905 if (use == n || use->is_Root())
906 906 return 0;
907 907
908 908 uint def_pre_order = _bbs[def->_idx]->_pre_order;
909 909 uint latency = 0;
910 910
911 911 // If the use is not a projection, then it is simple...
912 912 if (!use->is_Proj()) {
913 913 #ifndef PRODUCT
914 914 if (trace_opto_pipelining()) {
915 915 tty->print("# out(): ");
916 916 use->dump();
917 917 }
918 918 #endif
919 919
920 920 uint use_pre_order = _bbs[use->_idx]->_pre_order;
921 921
922 922 if (use_pre_order < def_pre_order)
923 923 return 0;
924 924
925 925 if (use_pre_order == def_pre_order && use->is_Phi())
926 926 return 0;
927 927
928 928 uint nlen = use->len();
929 929 uint nl = _node_latency->at_grow(use->_idx);
930 930
931 931 for ( uint j=0; j<nlen; j++ ) {
932 932 if (use->in(j) == n) {
933 933 // Change this if we want local latencies
934 934 uint ul = use->latency(j);
935 935 uint l = ul + nl;
936 936 if (latency < l) latency = l;
937 937 #ifndef PRODUCT
938 938 if (trace_opto_pipelining()) {
939 939 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
940 940 nl, j, ul, l, latency);
941 941 }
942 942 #endif
943 943 }
944 944 }
945 945 } else {
946 946 // This is a projection, just grab the latency of the use(s)
947 947 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
948 948 uint l = latency_from_use(use, def, use->fast_out(j));
949 949 if (latency < l) latency = l;
950 950 }
951 951 }
952 952
953 953 return latency;
954 954 }
955 955
956 956 //------------------------------latency_from_uses------------------------------
957 957 // Compute the latency of this instruction relative to all of it's uses.
958 958 // This computes a number that increases as we approach the beginning of the
959 959 // routine.
960 960 void PhaseCFG::latency_from_uses(Node *n) {
961 961 // Set the latency for this instruction
962 962 #ifndef PRODUCT
963 963 if (trace_opto_pipelining()) {
964 964 tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
965 965 n->_idx, _node_latency->at_grow(n->_idx));
966 966 dump();
967 967 }
968 968 #endif
969 969 uint latency=0;
970 970 const Node *def = n->is_Proj() ? n->in(0): n;
971 971
972 972 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
973 973 uint l = latency_from_use(n, def, n->fast_out(i));
974 974
975 975 if (latency < l) latency = l;
976 976 }
977 977
978 978 _node_latency->at_put_grow(n->_idx, latency);
979 979 }
980 980
981 981 //------------------------------hoist_to_cheaper_block-------------------------
982 982 // Pick a block for node self, between early and LCA, that is a cheaper
983 983 // alternative to LCA.
984 984 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
985 985 const double delta = 1+PROB_UNLIKELY_MAG(4);
986 986 Block* least = LCA;
987 987 double least_freq = least->_freq;
988 988 uint target = _node_latency->at_grow(self->_idx);
989 989 uint start_latency = _node_latency->at_grow(LCA->_nodes[0]->_idx);
990 990 uint end_latency = _node_latency->at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
991 991 bool in_latency = (target <= start_latency);
992 992 const Block* root_block = _bbs[_root->_idx];
993 993
994 994 // Turn off latency scheduling if scheduling is just plain off
995 995 if (!C->do_scheduling())
996 996 in_latency = true;
997 997
998 998 // Do not hoist (to cover latency) instructions which target a
999 999 // single register. Hoisting stretches the live range of the
1000 1000 // single register and may force spilling.
1001 1001 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1002 1002 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1003 1003 in_latency = true;
1004 1004
1005 1005 #ifndef PRODUCT
1006 1006 if (trace_opto_pipelining()) {
1007 1007 tty->print("# Find cheaper block for latency %d: ",
1008 1008 _node_latency->at_grow(self->_idx));
1009 1009 self->dump();
1010 1010 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1011 1011 LCA->_pre_order,
1012 1012 LCA->_nodes[0]->_idx,
1013 1013 start_latency,
1014 1014 LCA->_nodes[LCA->end_idx()]->_idx,
1015 1015 end_latency,
1016 1016 least_freq);
1017 1017 }
1018 1018 #endif
1019 1019
1020 1020 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1021 1021 // the earliest legal location. Capture the least execution frequency.
1022 1022 while (LCA != early) {
1023 1023 LCA = LCA->_idom; // Follow up the dominator tree
1024 1024
1025 1025 if (LCA == NULL) {
1026 1026 // Bailout without retry
1027 1027 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1028 1028 return least;
1029 1029 }
1030 1030
1031 1031 // Don't hoist machine instructions to the root basic block
1032 1032 if (mach && LCA == root_block)
1033 1033 break;
1034 1034
1035 1035 uint start_lat = _node_latency->at_grow(LCA->_nodes[0]->_idx);
1036 1036 uint end_idx = LCA->end_idx();
1037 1037 uint end_lat = _node_latency->at_grow(LCA->_nodes[end_idx]->_idx);
1038 1038 double LCA_freq = LCA->_freq;
1039 1039 #ifndef PRODUCT
1040 1040 if (trace_opto_pipelining()) {
1041 1041 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1042 1042 LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
1043 1043 }
1044 1044 #endif
1045 1045 if (LCA_freq < least_freq || // Better Frequency
1046 1046 ( !in_latency && // No block containing latency
1047 1047 LCA_freq < least_freq * delta && // No worse frequency
1048 1048 target >= end_lat && // within latency range
1049 1049 !self->is_iteratively_computed() ) // But don't hoist IV increments
1050 1050 // because they may end up above other uses of their phi forcing
1051 1051 // their result register to be different from their input.
1052 1052 ) {
1053 1053 least = LCA; // Found cheaper block
1054 1054 least_freq = LCA_freq;
1055 1055 start_latency = start_lat;
1056 1056 end_latency = end_lat;
1057 1057 if (target <= start_lat)
1058 1058 in_latency = true;
1059 1059 }
1060 1060 }
1061 1061
1062 1062 #ifndef PRODUCT
1063 1063 if (trace_opto_pipelining()) {
1064 1064 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1065 1065 least->_pre_order, start_latency, least_freq);
1066 1066 }
1067 1067 #endif
1068 1068
1069 1069 // See if the latency needs to be updated
1070 1070 if (target < end_latency) {
1071 1071 #ifndef PRODUCT
1072 1072 if (trace_opto_pipelining()) {
1073 1073 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1074 1074 }
1075 1075 #endif
1076 1076 _node_latency->at_put_grow(self->_idx, end_latency);
1077 1077 partial_latency_of_defs(self);
1078 1078 }
1079 1079
1080 1080 return least;
1081 1081 }
1082 1082
1083 1083
1084 1084 //------------------------------schedule_late-----------------------------------
1085 1085 // Now schedule all codes as LATE as possible. This is the LCA in the
1086 1086 // dominator tree of all USES of a value. Pick the block with the least
1087 1087 // loop nesting depth that is lowest in the dominator tree.
1088 1088 extern const char must_clone[];
1089 1089 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1090 1090 #ifndef PRODUCT
1091 1091 if (trace_opto_pipelining())
1092 1092 tty->print("\n#---- schedule_late ----\n");
1093 1093 #endif
1094 1094
1095 1095 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
1096 1096 Node *self;
1097 1097
1098 1098 // Walk over all the nodes from last to first
1099 1099 while (self = iter.next()) {
1100 1100 Block* early = _bbs[self->_idx]; // Earliest legal placement
1101 1101
1102 1102 if (self->is_top()) {
1103 1103 // Top node goes in bb #2 with other constants.
1104 1104 // It must be special-cased, because it has no out edges.
1105 1105 early->add_inst(self);
1106 1106 continue;
1107 1107 }
1108 1108
1109 1109 // No uses, just terminate
1110 1110 if (self->outcnt() == 0) {
1111 1111 assert(self->Opcode() == Op_MachProj, "sanity");
1112 1112 continue; // Must be a dead machine projection
1113 1113 }
1114 1114
1115 1115 // If node is pinned in the block, then no scheduling can be done.
1116 1116 if( self->pinned() ) // Pinned in block?
1117 1117 continue;
1118 1118
1119 1119 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1120 1120 if (mach) {
1121 1121 switch (mach->ideal_Opcode()) {
1122 1122 case Op_CreateEx:
1123 1123 // Don't move exception creation
1124 1124 early->add_inst(self);
1125 1125 continue;
1126 1126 break;
1127 1127 case Op_CheckCastPP:
1128 1128 // Don't move CheckCastPP nodes away from their input, if the input
1129 1129 // is a rawptr (5071820).
1130 1130 Node *def = self->in(1);
1131 1131 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1132 1132 early->add_inst(self);
1133 1133 #ifdef ASSERT
1134 1134 _raw_oops.push(def);
1135 1135 #endif
1136 1136 continue;
1137 1137 }
1138 1138 break;
1139 1139 }
1140 1140 }
1141 1141
1142 1142 // Gather LCA of all uses
1143 1143 Block *LCA = NULL;
1144 1144 {
1145 1145 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1146 1146 // For all uses, find LCA
1147 1147 Node* use = self->fast_out(i);
1148 1148 LCA = raise_LCA_above_use(LCA, use, self, _bbs);
1149 1149 }
1150 1150 } // (Hide defs of imax, i from rest of block.)
1151 1151
1152 1152 // Place temps in the block of their use. This isn't a
1153 1153 // requirement for correctness but it reduces useless
1154 1154 // interference between temps and other nodes.
1155 1155 if (mach != NULL && mach->is_MachTemp()) {
1156 1156 _bbs.map(self->_idx, LCA);
1157 1157 LCA->add_inst(self);
1158 1158 continue;
1159 1159 }
1160 1160
1161 1161 // Check if 'self' could be anti-dependent on memory
1162 1162 if (self->needs_anti_dependence_check()) {
1163 1163 // Hoist LCA above possible-defs and insert anti-dependences to
1164 1164 // defs in new LCA block.
1165 1165 LCA = insert_anti_dependences(LCA, self);
1166 1166 }
1167 1167
1168 1168 if (early->_dom_depth > LCA->_dom_depth) {
1169 1169 // Somehow the LCA has moved above the earliest legal point.
1170 1170 // (One way this can happen is via memory_early_block.)
1171 1171 if (C->subsume_loads() == true && !C->failing()) {
1172 1172 // Retry with subsume_loads == false
1173 1173 // If this is the first failure, the sentinel string will "stick"
1174 1174 // to the Compile object, and the C2Compiler will see it and retry.
1175 1175 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1176 1176 } else {
1177 1177 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1178 1178 C->record_method_not_compilable("late schedule failed: incorrect graph");
1179 1179 }
1180 1180 return;
1181 1181 }
1182 1182
1183 1183 // If there is no opportunity to hoist, then we're done.
1184 1184 bool try_to_hoist = (LCA != early);
1185 1185
1186 1186 // Must clone guys stay next to use; no hoisting allowed.
1187 1187 // Also cannot hoist guys that alter memory or are otherwise not
1188 1188 // allocatable (hoisting can make a value live longer, leading to
1189 1189 // anti and output dependency problems which are normally resolved
1190 1190 // by the register allocator giving everyone a different register).
1191 1191 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1192 1192 try_to_hoist = false;
1193 1193
1194 1194 Block* late = NULL;
1195 1195 if (try_to_hoist) {
1196 1196 // Now find the block with the least execution frequency.
1197 1197 // Start at the latest schedule and work up to the earliest schedule
1198 1198 // in the dominator tree. Thus the Node will dominate all its uses.
1199 1199 late = hoist_to_cheaper_block(LCA, early, self);
1200 1200 } else {
1201 1201 // Just use the LCA of the uses.
1202 1202 late = LCA;
1203 1203 }
1204 1204
1205 1205 // Put the node into target block
1206 1206 schedule_node_into_block(self, late);
1207 1207
1208 1208 #ifdef ASSERT
1209 1209 if (self->needs_anti_dependence_check()) {
1210 1210 // since precedence edges are only inserted when we're sure they
1211 1211 // are needed make sure that after placement in a block we don't
1212 1212 // need any new precedence edges.
1213 1213 verify_anti_dependences(late, self);
1214 1214 }
1215 1215 #endif
1216 1216 } // Loop until all nodes have been visited
1217 1217
1218 1218 } // end ScheduleLate
1219 1219
1220 1220 //------------------------------GlobalCodeMotion-------------------------------
1221 1221 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
1222 1222 ResourceMark rm;
1223 1223
1224 1224 #ifndef PRODUCT
1225 1225 if (trace_opto_pipelining()) {
1226 1226 tty->print("\n---- Start GlobalCodeMotion ----\n");
1227 1227 }
1228 1228 #endif
1229 1229
1230 1230 // Initialize the bbs.map for things on the proj_list
1231 1231 uint i;
1232 1232 for( i=0; i < proj_list.size(); i++ )
1233 1233 _bbs.map(proj_list[i]->_idx, NULL);
1234 1234
1235 1235 // Set the basic block for Nodes pinned into blocks
1236 1236 Arena *a = Thread::current()->resource_area();
1237 1237 VectorSet visited(a);
1238 1238 schedule_pinned_nodes( visited );
1239 1239
1240 1240 // Find the earliest Block any instruction can be placed in. Some
1241 1241 // instructions are pinned into Blocks. Unpinned instructions can
1242 1242 // appear in last block in which all their inputs occur.
1243 1243 visited.Clear();
1244 1244 Node_List stack(a);
1245 1245 stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
1246 1246 if (!schedule_early(visited, stack)) {
1247 1247 // Bailout without retry
1248 1248 C->record_method_not_compilable("early schedule failed");
1249 1249 return;
1250 1250 }
1251 1251
1252 1252 // Build Def-Use edges.
1253 1253 proj_list.push(_root); // Add real root as another root
1254 1254 proj_list.pop();
1255 1255
1256 1256 // Compute the latency information (via backwards walk) for all the
1257 1257 // instructions in the graph
1258 1258 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1259 1259
1260 1260 if( C->do_scheduling() )
1261 1261 ComputeLatenciesBackwards(visited, stack);
1262 1262
1263 1263 // Now schedule all codes as LATE as possible. This is the LCA in the
1264 1264 // dominator tree of all USES of a value. Pick the block with the least
1265 1265 // loop nesting depth that is lowest in the dominator tree.
1266 1266 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1267 1267 schedule_late(visited, stack);
1268 1268 if( C->failing() ) {
1269 1269 // schedule_late fails only when graph is incorrect.
1270 1270 assert(!VerifyGraphEdges, "verification should have failed");
1271 1271 return;
1272 1272 }
1273 1273
1274 1274 unique = C->unique();
1275 1275
1276 1276 #ifndef PRODUCT
1277 1277 if (trace_opto_pipelining()) {
1278 1278 tty->print("\n---- Detect implicit null checks ----\n");
1279 1279 }
1280 1280 #endif
1281 1281
1282 1282 // Detect implicit-null-check opportunities. Basically, find NULL checks
1283 1283 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1284 1284 // I can generate a memory op if there is not one nearby.
1285 1285 if (C->is_method_compilation()) {
1286 1286 // Don't do it for natives, adapters, or runtime stubs
1287 1287 int allowed_reasons = 0;
1288 1288 // ...and don't do it when there have been too many traps, globally.
1289 1289 for (int reason = (int)Deoptimization::Reason_none+1;
1290 1290 reason < Compile::trapHistLength; reason++) {
1291 1291 assert(reason < BitsPerInt, "recode bit map");
1292 1292 if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
1293 1293 allowed_reasons |= nth_bit(reason);
1294 1294 }
1295 1295 // By reversing the loop direction we get a very minor gain on mpegaudio.
1296 1296 // Feel free to revert to a forward loop for clarity.
1297 1297 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1298 1298 for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
1299 1299 Node *proj = matcher._null_check_tests[i ];
1300 1300 Node *val = matcher._null_check_tests[i+1];
1301 1301 _bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons);
1302 1302 // The implicit_null_check will only perform the transformation
1303 1303 // if the null branch is truly uncommon, *and* it leads to an
1304 1304 // uncommon trap. Combined with the too_many_traps guards
1305 1305 // above, this prevents SEGV storms reported in 6366351,
1306 1306 // by recompiling offending methods without this optimization.
1307 1307 }
1308 1308 }
1309 1309
1310 1310 #ifndef PRODUCT
1311 1311 if (trace_opto_pipelining()) {
1312 1312 tty->print("\n---- Start Local Scheduling ----\n");
1313 1313 }
1314 1314 #endif
1315 1315
1316 1316 // Schedule locally. Right now a simple topological sort.
1317 1317 // Later, do a real latency aware scheduler.
1318 1318 int *ready_cnt = NEW_RESOURCE_ARRAY(int,C->unique());
1319 1319 memset( ready_cnt, -1, C->unique() * sizeof(int) );
1320 1320 visited.Clear();
1321 1321 for (i = 0; i < _num_blocks; i++) {
1322 1322 if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
1323 1323 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1324 1324 C->record_method_not_compilable("local schedule failed");
1325 1325 }
1326 1326 return;
1327 1327 }
1328 1328 }
1329 1329
1330 1330 // If we inserted any instructions between a Call and his CatchNode,
1331 1331 // clone the instructions on all paths below the Catch.
1332 1332 for( i=0; i < _num_blocks; i++ )
1333 1333 _blocks[i]->call_catch_cleanup(_bbs);
1334 1334
1335 1335 #ifndef PRODUCT
1336 1336 if (trace_opto_pipelining()) {
1337 1337 tty->print("\n---- After GlobalCodeMotion ----\n");
1338 1338 for (uint i = 0; i < _num_blocks; i++) {
1339 1339 _blocks[i]->dump();
1340 1340 }
1341 1341 }
1342 1342 #endif
1343 1343 // Dead.
1344 1344 _node_latency = (GrowableArray<uint> *)0xdeadbeef;
1345 1345 }
1346 1346
1347 1347
1348 1348 //------------------------------Estimate_Block_Frequency-----------------------
1349 1349 // Estimate block frequencies based on IfNode probabilities.
1350 1350 void PhaseCFG::Estimate_Block_Frequency() {
1351 1351
1352 1352 // Force conditional branches leading to uncommon traps to be unlikely,
1353 1353 // not because we get to the uncommon_trap with less relative frequency,
1354 1354 // but because an uncommon_trap typically causes a deopt, so we only get
1355 1355 // there once.
1356 1356 if (C->do_freq_based_layout()) {
1357 1357 Block_List worklist;
1358 1358 Block* root_blk = _blocks[0];
1359 1359 for (uint i = 1; i < root_blk->num_preds(); i++) {
1360 1360 Block *pb = _bbs[root_blk->pred(i)->_idx];
1361 1361 if (pb->has_uncommon_code()) {
1362 1362 worklist.push(pb);
1363 1363 }
1364 1364 }
1365 1365 while (worklist.size() > 0) {
1366 1366 Block* uct = worklist.pop();
1367 1367 if (uct == _broot) continue;
1368 1368 for (uint i = 1; i < uct->num_preds(); i++) {
1369 1369 Block *pb = _bbs[uct->pred(i)->_idx];
1370 1370 if (pb->_num_succs == 1) {
1371 1371 worklist.push(pb);
1372 1372 } else if (pb->num_fall_throughs() == 2) {
1373 1373 pb->update_uncommon_branch(uct);
1374 1374 }
1375 1375 }
1376 1376 }
1377 1377 }
1378 1378
1379 1379 // Create the loop tree and calculate loop depth.
1380 1380 _root_loop = create_loop_tree();
1381 1381 _root_loop->compute_loop_depth(0);
1382 1382
1383 1383 // Compute block frequency of each block, relative to a single loop entry.
1384 1384 _root_loop->compute_freq();
1385 1385
1386 1386 // Adjust all frequencies to be relative to a single method entry
1387 1387 _root_loop->_freq = 1.0;
1388 1388 _root_loop->scale_freq();
1389 1389
1390 1390 // Save outmost loop frequency for LRG frequency threshold
1391 1391 _outer_loop_freq = _root_loop->outer_loop_freq();
1392 1392
1393 1393 // force paths ending at uncommon traps to be infrequent
1394 1394 if (!C->do_freq_based_layout()) {
1395 1395 Block_List worklist;
1396 1396 Block* root_blk = _blocks[0];
1397 1397 for (uint i = 1; i < root_blk->num_preds(); i++) {
1398 1398 Block *pb = _bbs[root_blk->pred(i)->_idx];
1399 1399 if (pb->has_uncommon_code()) {
1400 1400 worklist.push(pb);
1401 1401 }
1402 1402 }
1403 1403 while (worklist.size() > 0) {
1404 1404 Block* uct = worklist.pop();
1405 1405 uct->_freq = PROB_MIN;
1406 1406 for (uint i = 1; i < uct->num_preds(); i++) {
1407 1407 Block *pb = _bbs[uct->pred(i)->_idx];
1408 1408 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1409 1409 worklist.push(pb);
1410 1410 }
1411 1411 }
1412 1412 }
1413 1413 }
1414 1414
1415 1415 #ifdef ASSERT
1416 1416 for (uint i = 0; i < _num_blocks; i++ ) {
1417 1417 Block *b = _blocks[i];
1418 1418 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1419 1419 }
1420 1420 #endif
1421 1421
1422 1422 #ifndef PRODUCT
1423 1423 if (PrintCFGBlockFreq) {
1424 1424 tty->print_cr("CFG Block Frequencies");
1425 1425 _root_loop->dump_tree();
1426 1426 if (Verbose) {
1427 1427 tty->print_cr("PhaseCFG dump");
1428 1428 dump();
1429 1429 tty->print_cr("Node dump");
1430 1430 _root->dump(99999);
1431 1431 }
1432 1432 }
1433 1433 #endif
1434 1434 }
1435 1435
1436 1436 //----------------------------create_loop_tree--------------------------------
1437 1437 // Create a loop tree from the CFG
1438 1438 CFGLoop* PhaseCFG::create_loop_tree() {
1439 1439
1440 1440 #ifdef ASSERT
1441 1441 assert( _blocks[0] == _broot, "" );
1442 1442 for (uint i = 0; i < _num_blocks; i++ ) {
1443 1443 Block *b = _blocks[i];
1444 1444 // Check that _loop field are clear...we could clear them if not.
1445 1445 assert(b->_loop == NULL, "clear _loop expected");
1446 1446 // Sanity check that the RPO numbering is reflected in the _blocks array.
1447 1447 // It doesn't have to be for the loop tree to be built, but if it is not,
1448 1448 // then the blocks have been reordered since dom graph building...which
1449 1449 // may question the RPO numbering
1450 1450 assert(b->_rpo == i, "unexpected reverse post order number");
1451 1451 }
1452 1452 #endif
1453 1453
1454 1454 int idct = 0;
1455 1455 CFGLoop* root_loop = new CFGLoop(idct++);
1456 1456
1457 1457 Block_List worklist;
1458 1458
1459 1459 // Assign blocks to loops
1460 1460 for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
1461 1461 Block *b = _blocks[i];
1462 1462
1463 1463 if (b->head()->is_Loop()) {
1464 1464 Block* loop_head = b;
1465 1465 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1466 1466 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1467 1467 Block* tail = _bbs[tail_n->_idx];
1468 1468
1469 1469 // Defensively filter out Loop nodes for non-single-entry loops.
1470 1470 // For all reasonable loops, the head occurs before the tail in RPO.
1471 1471 if (i <= tail->_rpo) {
1472 1472
1473 1473 // The tail and (recursive) predecessors of the tail
1474 1474 // are made members of a new loop.
1475 1475
1476 1476 assert(worklist.size() == 0, "nonempty worklist");
1477 1477 CFGLoop* nloop = new CFGLoop(idct++);
1478 1478 assert(loop_head->_loop == NULL, "just checking");
1479 1479 loop_head->_loop = nloop;
1480 1480 // Add to nloop so push_pred() will skip over inner loops
1481 1481 nloop->add_member(loop_head);
1482 1482 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs);
1483 1483
1484 1484 while (worklist.size() > 0) {
1485 1485 Block* member = worklist.pop();
1486 1486 if (member != loop_head) {
1487 1487 for (uint j = 1; j < member->num_preds(); j++) {
1488 1488 nloop->push_pred(member, j, worklist, _bbs);
1489 1489 }
1490 1490 }
1491 1491 }
1492 1492 }
1493 1493 }
1494 1494 }
1495 1495
1496 1496 // Create a member list for each loop consisting
1497 1497 // of both blocks and (immediate child) loops.
1498 1498 for (uint i = 0; i < _num_blocks; i++) {
1499 1499 Block *b = _blocks[i];
1500 1500 CFGLoop* lp = b->_loop;
1501 1501 if (lp == NULL) {
1502 1502 // Not assigned to a loop. Add it to the method's pseudo loop.
1503 1503 b->_loop = root_loop;
1504 1504 lp = root_loop;
1505 1505 }
1506 1506 if (lp == root_loop || b != lp->head()) { // loop heads are already members
1507 1507 lp->add_member(b);
1508 1508 }
1509 1509 if (lp != root_loop) {
1510 1510 if (lp->parent() == NULL) {
1511 1511 // Not a nested loop. Make it a child of the method's pseudo loop.
1512 1512 root_loop->add_nested_loop(lp);
1513 1513 }
1514 1514 if (b == lp->head()) {
1515 1515 // Add nested loop to member list of parent loop.
1516 1516 lp->parent()->add_member(lp);
1517 1517 }
1518 1518 }
1519 1519 }
1520 1520
1521 1521 return root_loop;
1522 1522 }
1523 1523
1524 1524 //------------------------------push_pred--------------------------------------
1525 1525 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) {
1526 1526 Node* pred_n = blk->pred(i);
1527 1527 Block* pred = node_to_blk[pred_n->_idx];
1528 1528 CFGLoop *pred_loop = pred->_loop;
1529 1529 if (pred_loop == NULL) {
1530 1530 // Filter out blocks for non-single-entry loops.
1531 1531 // For all reasonable loops, the head occurs before the tail in RPO.
1532 1532 if (pred->_rpo > head()->_rpo) {
1533 1533 pred->_loop = this;
1534 1534 worklist.push(pred);
1535 1535 }
1536 1536 } else if (pred_loop != this) {
1537 1537 // Nested loop.
1538 1538 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1539 1539 pred_loop = pred_loop->_parent;
1540 1540 }
1541 1541 // Make pred's loop be a child
1542 1542 if (pred_loop->_parent == NULL) {
1543 1543 add_nested_loop(pred_loop);
1544 1544 // Continue with loop entry predecessor.
1545 1545 Block* pred_head = pred_loop->head();
1546 1546 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1547 1547 assert(pred_head != head(), "loop head in only one loop");
1548 1548 push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk);
1549 1549 } else {
1550 1550 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1551 1551 }
1552 1552 }
1553 1553 }
1554 1554
1555 1555 //------------------------------add_nested_loop--------------------------------
1556 1556 // Make cl a child of the current loop in the loop tree.
1557 1557 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1558 1558 assert(_parent == NULL, "no parent yet");
1559 1559 assert(cl != this, "not my own parent");
1560 1560 cl->_parent = this;
1561 1561 CFGLoop* ch = _child;
1562 1562 if (ch == NULL) {
1563 1563 _child = cl;
1564 1564 } else {
1565 1565 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1566 1566 ch->_sibling = cl;
1567 1567 }
1568 1568 }
1569 1569
1570 1570 //------------------------------compute_loop_depth-----------------------------
1571 1571 // Store the loop depth in each CFGLoop object.
1572 1572 // Recursively walk the children to do the same for them.
1573 1573 void CFGLoop::compute_loop_depth(int depth) {
1574 1574 _depth = depth;
1575 1575 CFGLoop* ch = _child;
1576 1576 while (ch != NULL) {
1577 1577 ch->compute_loop_depth(depth + 1);
1578 1578 ch = ch->_sibling;
1579 1579 }
1580 1580 }
1581 1581
1582 1582 //------------------------------compute_freq-----------------------------------
1583 1583 // Compute the frequency of each block and loop, relative to a single entry
1584 1584 // into the dominating loop head.
1585 1585 void CFGLoop::compute_freq() {
1586 1586 // Bottom up traversal of loop tree (visit inner loops first.)
1587 1587 // Set loop head frequency to 1.0, then transitively
1588 1588 // compute frequency for all successors in the loop,
1589 1589 // as well as for each exit edge. Inner loops are
1590 1590 // treated as single blocks with loop exit targets
1591 1591 // as the successor blocks.
1592 1592
1593 1593 // Nested loops first
1594 1594 CFGLoop* ch = _child;
1595 1595 while (ch != NULL) {
1596 1596 ch->compute_freq();
1597 1597 ch = ch->_sibling;
1598 1598 }
1599 1599 assert (_members.length() > 0, "no empty loops");
1600 1600 Block* hd = head();
1601 1601 hd->_freq = 1.0f;
1602 1602 for (int i = 0; i < _members.length(); i++) {
1603 1603 CFGElement* s = _members.at(i);
1604 1604 float freq = s->_freq;
1605 1605 if (s->is_block()) {
1606 1606 Block* b = s->as_Block();
1607 1607 for (uint j = 0; j < b->_num_succs; j++) {
1608 1608 Block* sb = b->_succs[j];
1609 1609 update_succ_freq(sb, freq * b->succ_prob(j));
1610 1610 }
1611 1611 } else {
1612 1612 CFGLoop* lp = s->as_CFGLoop();
1613 1613 assert(lp->_parent == this, "immediate child");
1614 1614 for (int k = 0; k < lp->_exits.length(); k++) {
1615 1615 Block* eb = lp->_exits.at(k).get_target();
1616 1616 float prob = lp->_exits.at(k).get_prob();
1617 1617 update_succ_freq(eb, freq * prob);
1618 1618 }
1619 1619 }
1620 1620 }
1621 1621
1622 1622 // For all loops other than the outer, "method" loop,
1623 1623 // sum and normalize the exit probability. The "method" loop
1624 1624 // should keep the initial exit probability of 1, so that
1625 1625 // inner blocks do not get erroneously scaled.
1626 1626 if (_depth != 0) {
1627 1627 // Total the exit probabilities for this loop.
1628 1628 float exits_sum = 0.0f;
1629 1629 for (int i = 0; i < _exits.length(); i++) {
1630 1630 exits_sum += _exits.at(i).get_prob();
1631 1631 }
1632 1632
1633 1633 // Normalize the exit probabilities. Until now, the
1634 1634 // probabilities estimate the possibility of exit per
1635 1635 // a single loop iteration; afterward, they estimate
1636 1636 // the probability of exit per loop entry.
1637 1637 for (int i = 0; i < _exits.length(); i++) {
1638 1638 Block* et = _exits.at(i).get_target();
1639 1639 float new_prob = 0.0f;
1640 1640 if (_exits.at(i).get_prob() > 0.0f) {
1641 1641 new_prob = _exits.at(i).get_prob() / exits_sum;
1642 1642 }
1643 1643 BlockProbPair bpp(et, new_prob);
1644 1644 _exits.at_put(i, bpp);
1645 1645 }
1646 1646
1647 1647 // Save the total, but guard against unreasonable probability,
1648 1648 // as the value is used to estimate the loop trip count.
1649 1649 // An infinite trip count would blur relative block
1650 1650 // frequencies.
1651 1651 if (exits_sum > 1.0f) exits_sum = 1.0;
1652 1652 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1653 1653 _exit_prob = exits_sum;
1654 1654 }
1655 1655 }
1656 1656
1657 1657 //------------------------------succ_prob-------------------------------------
1658 1658 // Determine the probability of reaching successor 'i' from the receiver block.
1659 1659 float Block::succ_prob(uint i) {
1660 1660 int eidx = end_idx();
1661 1661 Node *n = _nodes[eidx]; // Get ending Node
1662 1662
1663 1663 int op = n->Opcode();
1664 1664 if (n->is_Mach()) {
1665 1665 if (n->is_MachNullCheck()) {
1666 1666 // Can only reach here if called after lcm. The original Op_If is gone,
1667 1667 // so we attempt to infer the probability from one or both of the
1668 1668 // successor blocks.
1669 1669 assert(_num_succs == 2, "expecting 2 successors of a null check");
1670 1670 // If either successor has only one predecessor, then the
1671 1671 // probability estimate can be derived using the
1672 1672 // relative frequency of the successor and this block.
1673 1673 if (_succs[i]->num_preds() == 2) {
1674 1674 return _succs[i]->_freq / _freq;
1675 1675 } else if (_succs[1-i]->num_preds() == 2) {
1676 1676 return 1 - (_succs[1-i]->_freq / _freq);
1677 1677 } else {
1678 1678 // Estimate using both successor frequencies
1679 1679 float freq = _succs[i]->_freq;
1680 1680 return freq / (freq + _succs[1-i]->_freq);
1681 1681 }
1682 1682 }
1683 1683 op = n->as_Mach()->ideal_Opcode();
1684 1684 }
1685 1685
1686 1686
1687 1687 // Switch on branch type
1688 1688 switch( op ) {
1689 1689 case Op_CountedLoopEnd:
1690 1690 case Op_If: {
1691 1691 assert (i < 2, "just checking");
1692 1692 // Conditionals pass on only part of their frequency
1693 1693 float prob = n->as_MachIf()->_prob;
1694 1694 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1695 1695 // If succ[i] is the FALSE branch, invert path info
1696 1696 if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
1697 1697 return 1.0f - prob; // not taken
1698 1698 } else {
1699 1699 return prob; // taken
1700 1700 }
1701 1701 }
1702 1702
1703 1703 case Op_Jump:
1704 1704 // Divide the frequency between all successors evenly
1705 1705 return 1.0f/_num_succs;
1706 1706
1707 1707 case Op_Catch: {
1708 1708 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1709 1709 if (ci->_con == CatchProjNode::fall_through_index) {
1710 1710 // Fall-thru path gets the lion's share.
1711 1711 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1712 1712 } else {
1713 1713 // Presume exceptional paths are equally unlikely
1714 1714 return PROB_UNLIKELY_MAG(5);
1715 1715 }
1716 1716 }
1717 1717
1718 1718 case Op_Root:
1719 1719 case Op_Goto:
1720 1720 // Pass frequency straight thru to target
1721 1721 return 1.0f;
1722 1722
1723 1723 case Op_NeverBranch:
1724 1724 return 0.0f;
1725 1725
1726 1726 case Op_TailCall:
1727 1727 case Op_TailJump:
1728 1728 case Op_Return:
1729 1729 case Op_Halt:
1730 1730 case Op_Rethrow:
1731 1731 // Do not push out freq to root block
1732 1732 return 0.0f;
1733 1733
1734 1734 default:
1735 1735 ShouldNotReachHere();
1736 1736 }
1737 1737
1738 1738 return 0.0f;
1739 1739 }
1740 1740
1741 1741 //------------------------------num_fall_throughs-----------------------------
1742 1742 // Return the number of fall-through candidates for a block
1743 1743 int Block::num_fall_throughs() {
1744 1744 int eidx = end_idx();
1745 1745 Node *n = _nodes[eidx]; // Get ending Node
1746 1746
1747 1747 int op = n->Opcode();
1748 1748 if (n->is_Mach()) {
1749 1749 if (n->is_MachNullCheck()) {
1750 1750 // In theory, either side can fall-thru, for simplicity sake,
1751 1751 // let's say only the false branch can now.
1752 1752 return 1;
1753 1753 }
1754 1754 op = n->as_Mach()->ideal_Opcode();
1755 1755 }
1756 1756
1757 1757 // Switch on branch type
1758 1758 switch( op ) {
1759 1759 case Op_CountedLoopEnd:
1760 1760 case Op_If:
1761 1761 return 2;
1762 1762
1763 1763 case Op_Root:
1764 1764 case Op_Goto:
1765 1765 return 1;
1766 1766
1767 1767 case Op_Catch: {
1768 1768 for (uint i = 0; i < _num_succs; i++) {
1769 1769 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1770 1770 if (ci->_con == CatchProjNode::fall_through_index) {
1771 1771 return 1;
1772 1772 }
1773 1773 }
1774 1774 return 0;
1775 1775 }
1776 1776
1777 1777 case Op_Jump:
1778 1778 case Op_NeverBranch:
1779 1779 case Op_TailCall:
1780 1780 case Op_TailJump:
1781 1781 case Op_Return:
1782 1782 case Op_Halt:
1783 1783 case Op_Rethrow:
1784 1784 return 0;
1785 1785
1786 1786 default:
1787 1787 ShouldNotReachHere();
1788 1788 }
1789 1789
1790 1790 return 0;
1791 1791 }
1792 1792
1793 1793 //------------------------------succ_fall_through-----------------------------
1794 1794 // Return true if a specific successor could be fall-through target.
1795 1795 bool Block::succ_fall_through(uint i) {
1796 1796 int eidx = end_idx();
1797 1797 Node *n = _nodes[eidx]; // Get ending Node
1798 1798
1799 1799 int op = n->Opcode();
1800 1800 if (n->is_Mach()) {
1801 1801 if (n->is_MachNullCheck()) {
1802 1802 // In theory, either side can fall-thru, for simplicity sake,
1803 1803 // let's say only the false branch can now.
1804 1804 return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse;
1805 1805 }
1806 1806 op = n->as_Mach()->ideal_Opcode();
1807 1807 }
1808 1808
1809 1809 // Switch on branch type
1810 1810 switch( op ) {
1811 1811 case Op_CountedLoopEnd:
1812 1812 case Op_If:
1813 1813 case Op_Root:
1814 1814 case Op_Goto:
1815 1815 return true;
1816 1816
1817 1817 case Op_Catch: {
1818 1818 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1819 1819 return ci->_con == CatchProjNode::fall_through_index;
1820 1820 }
1821 1821
1822 1822 case Op_Jump:
1823 1823 case Op_NeverBranch:
1824 1824 case Op_TailCall:
1825 1825 case Op_TailJump:
1826 1826 case Op_Return:
1827 1827 case Op_Halt:
1828 1828 case Op_Rethrow:
1829 1829 return false;
1830 1830
1831 1831 default:
1832 1832 ShouldNotReachHere();
1833 1833 }
1834 1834
1835 1835 return false;
1836 1836 }
1837 1837
1838 1838 //------------------------------update_uncommon_branch------------------------
1839 1839 // Update the probability of a two-branch to be uncommon
1840 1840 void Block::update_uncommon_branch(Block* ub) {
1841 1841 int eidx = end_idx();
1842 1842 Node *n = _nodes[eidx]; // Get ending Node
1843 1843
1844 1844 int op = n->as_Mach()->ideal_Opcode();
1845 1845
1846 1846 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
1847 1847 assert(num_fall_throughs() == 2, "must be a two way branch block");
1848 1848
1849 1849 // Which successor is ub?
1850 1850 uint s;
1851 1851 for (s = 0; s <_num_succs; s++) {
1852 1852 if (_succs[s] == ub) break;
1853 1853 }
1854 1854 assert(s < 2, "uncommon successor must be found");
1855 1855
1856 1856 // If ub is the true path, make the proability small, else
1857 1857 // ub is the false path, and make the probability large
1858 1858 bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse);
1859 1859
1860 1860 // Get existing probability
1861 1861 float p = n->as_MachIf()->_prob;
1862 1862
1863 1863 if (invert) p = 1.0 - p;
1864 1864 if (p > PROB_MIN) {
1865 1865 p = PROB_MIN;
1866 1866 }
1867 1867 if (invert) p = 1.0 - p;
1868 1868
1869 1869 n->as_MachIf()->_prob = p;
1870 1870 }
1871 1871
1872 1872 //------------------------------update_succ_freq-------------------------------
1873 1873 // Update the appropriate frequency associated with block 'b', a successor of
1874 1874 // a block in this loop.
1875 1875 void CFGLoop::update_succ_freq(Block* b, float freq) {
1876 1876 if (b->_loop == this) {
1877 1877 if (b == head()) {
1878 1878 // back branch within the loop
1879 1879 // Do nothing now, the loop carried frequency will be
1880 1880 // adjust later in scale_freq().
1881 1881 } else {
1882 1882 // simple branch within the loop
1883 1883 b->_freq += freq;
1884 1884 }
1885 1885 } else if (!in_loop_nest(b)) {
1886 1886 // branch is exit from this loop
1887 1887 BlockProbPair bpp(b, freq);
1888 1888 _exits.append(bpp);
1889 1889 } else {
1890 1890 // branch into nested loop
1891 1891 CFGLoop* ch = b->_loop;
1892 1892 ch->_freq += freq;
1893 1893 }
1894 1894 }
1895 1895
1896 1896 //------------------------------in_loop_nest-----------------------------------
1897 1897 // Determine if block b is in the receiver's loop nest.
1898 1898 bool CFGLoop::in_loop_nest(Block* b) {
1899 1899 int depth = _depth;
1900 1900 CFGLoop* b_loop = b->_loop;
1901 1901 int b_depth = b_loop->_depth;
1902 1902 if (depth == b_depth) {
1903 1903 return true;
1904 1904 }
1905 1905 while (b_depth > depth) {
1906 1906 b_loop = b_loop->_parent;
1907 1907 b_depth = b_loop->_depth;
1908 1908 }
1909 1909 return b_loop == this;
1910 1910 }
1911 1911
1912 1912 //------------------------------scale_freq-------------------------------------
1913 1913 // Scale frequency of loops and blocks by trip counts from outer loops
1914 1914 // Do a top down traversal of loop tree (visit outer loops first.)
1915 1915 void CFGLoop::scale_freq() {
1916 1916 float loop_freq = _freq * trip_count();
1917 1917 _freq = loop_freq;
1918 1918 for (int i = 0; i < _members.length(); i++) {
1919 1919 CFGElement* s = _members.at(i);
1920 1920 float block_freq = s->_freq * loop_freq;
1921 1921 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
1922 1922 block_freq = MIN_BLOCK_FREQUENCY;
1923 1923 s->_freq = block_freq;
1924 1924 }
1925 1925 CFGLoop* ch = _child;
1926 1926 while (ch != NULL) {
1927 1927 ch->scale_freq();
1928 1928 ch = ch->_sibling;
1929 1929 }
1930 1930 }
1931 1931
1932 1932 // Frequency of outer loop
1933 1933 float CFGLoop::outer_loop_freq() const {
1934 1934 if (_child != NULL) {
1935 1935 return _child->_freq;
1936 1936 }
1937 1937 return _freq;
1938 1938 }
1939 1939
1940 1940 #ifndef PRODUCT
1941 1941 //------------------------------dump_tree--------------------------------------
1942 1942 void CFGLoop::dump_tree() const {
1943 1943 dump();
1944 1944 if (_child != NULL) _child->dump_tree();
1945 1945 if (_sibling != NULL) _sibling->dump_tree();
1946 1946 }
1947 1947
1948 1948 //------------------------------dump-------------------------------------------
1949 1949 void CFGLoop::dump() const {
1950 1950 for (int i = 0; i < _depth; i++) tty->print(" ");
1951 1951 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
1952 1952 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
1953 1953 for (int i = 0; i < _depth; i++) tty->print(" ");
1954 1954 tty->print(" members:", _id);
1955 1955 int k = 0;
1956 1956 for (int i = 0; i < _members.length(); i++) {
1957 1957 if (k++ >= 6) {
1958 1958 tty->print("\n ");
1959 1959 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1960 1960 k = 0;
1961 1961 }
1962 1962 CFGElement *s = _members.at(i);
1963 1963 if (s->is_block()) {
1964 1964 Block *b = s->as_Block();
1965 1965 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
1966 1966 } else {
1967 1967 CFGLoop* lp = s->as_CFGLoop();
1968 1968 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
1969 1969 }
1970 1970 }
1971 1971 tty->print("\n");
1972 1972 for (int i = 0; i < _depth; i++) tty->print(" ");
1973 1973 tty->print(" exits: ");
1974 1974 k = 0;
1975 1975 for (int i = 0; i < _exits.length(); i++) {
1976 1976 if (k++ >= 7) {
1977 1977 tty->print("\n ");
1978 1978 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1979 1979 k = 0;
1980 1980 }
1981 1981 Block *blk = _exits.at(i).get_target();
1982 1982 float prob = _exits.at(i).get_prob();
1983 1983 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
1984 1984 }
1985 1985 tty->print("\n");
1986 1986 }
1987 1987 #endif
|
↓ open down ↓ |
1908 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX