Print this page
rev 1838 : 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/output.cpp
+++ new/src/share/vm/opto/output.cpp
1 1 /*
2 2 * Copyright (c) 1998, 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 #include "incls/_precompiled.incl"
26 26 #include "incls/_output.cpp.incl"
27 27
28 28 extern uint size_java_to_interp();
29 29 extern uint reloc_java_to_interp();
30 30 extern uint size_exception_handler();
31 31 extern uint size_deopt_handler();
32 32
33 33 #ifndef PRODUCT
34 34 #define DEBUG_ARG(x) , x
35 35 #else
36 36 #define DEBUG_ARG(x)
37 37 #endif
↓ open down ↓ |
37 lines elided |
↑ open up ↑ |
38 38
39 39 extern int emit_exception_handler(CodeBuffer &cbuf);
40 40 extern int emit_deopt_handler(CodeBuffer &cbuf);
41 41
42 42 //------------------------------Output-----------------------------------------
43 43 // Convert Nodes to instruction bits and pass off to the VM
44 44 void Compile::Output() {
45 45 // RootNode goes
46 46 assert( _cfg->_broot->_nodes.size() == 0, "" );
47 47
48 - // Initialize the space for the BufferBlob used to find and verify
49 - // instruction size in MachNode::emit_size()
50 - init_scratch_buffer_blob();
51 - if (failing()) return; // Out of memory
52 -
53 48 // The number of new nodes (mostly MachNop) is proportional to
54 49 // the number of java calls and inner loops which are aligned.
55 50 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
56 51 C->inner_loops()*(OptoLoopAlignment-1)),
57 52 "out of nodes before code generation" ) ) {
58 53 return;
59 54 }
60 55 // Make sure I can find the Start Node
61 56 Block_Array& bbs = _cfg->_bbs;
62 57 Block *entry = _cfg->_blocks[1];
63 58 Block *broot = _cfg->_broot;
64 59
65 60 const StartNode *start = entry->_nodes[0]->as_Start();
66 61
67 62 // Replace StartNode with prolog
68 63 MachPrologNode *prolog = new (this) MachPrologNode();
69 64 entry->_nodes.map( 0, prolog );
70 65 bbs.map( prolog->_idx, entry );
71 66 bbs.map( start->_idx, NULL ); // start is no longer in any block
72 67
73 68 // Virtual methods need an unverified entry point
74 69
75 70 if( is_osr_compilation() ) {
76 71 if( PoisonOSREntry ) {
77 72 // TODO: Should use a ShouldNotReachHereNode...
78 73 _cfg->insert( broot, 0, new (this) MachBreakpointNode() );
79 74 }
80 75 } else {
81 76 if( _method && !_method->flags().is_static() ) {
82 77 // Insert unvalidated entry point
83 78 _cfg->insert( broot, 0, new (this) MachUEPNode() );
84 79 }
85 80
86 81 }
87 82
88 83
89 84 // Break before main entry point
90 85 if( (_method && _method->break_at_execute())
91 86 #ifndef PRODUCT
92 87 ||(OptoBreakpoint && is_method_compilation())
93 88 ||(OptoBreakpointOSR && is_osr_compilation())
94 89 ||(OptoBreakpointC2R && !_method)
95 90 #endif
96 91 ) {
97 92 // checking for _method means that OptoBreakpoint does not apply to
98 93 // runtime stubs or frame converters
99 94 _cfg->insert( entry, 1, new (this) MachBreakpointNode() );
100 95 }
101 96
102 97 // Insert epilogs before every return
103 98 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
104 99 Block *b = _cfg->_blocks[i];
105 100 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point?
106 101 Node *m = b->end();
107 102 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) {
108 103 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
109 104 b->add_inst( epilog );
110 105 bbs.map(epilog->_idx, b);
111 106 //_regalloc->set_bad(epilog->_idx); // Already initialized this way.
112 107 }
113 108 }
114 109 }
115 110
116 111 # ifdef ENABLE_ZAP_DEAD_LOCALS
117 112 if ( ZapDeadCompiledLocals ) Insert_zap_nodes();
118 113 # endif
119 114
120 115 ScheduleAndBundle();
121 116
122 117 #ifndef PRODUCT
123 118 if (trace_opto_output()) {
124 119 tty->print("\n---- After ScheduleAndBundle ----\n");
125 120 for (uint i = 0; i < _cfg->_num_blocks; i++) {
126 121 tty->print("\nBB#%03d:\n", i);
127 122 Block *bb = _cfg->_blocks[i];
128 123 for (uint j = 0; j < bb->_nodes.size(); j++) {
129 124 Node *n = bb->_nodes[j];
130 125 OptoReg::Name reg = _regalloc->get_reg_first(n);
131 126 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
132 127 n->dump();
133 128 }
134 129 }
135 130 }
136 131 #endif
137 132
138 133 if (failing()) return;
139 134
140 135 BuildOopMaps();
141 136
142 137 if (failing()) return;
143 138
144 139 Fill_buffer();
145 140 }
146 141
147 142 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
148 143 // Determine if we need to generate a stack overflow check.
149 144 // Do it if the method is not a stub function and
150 145 // has java calls or has frame size > vm_page_size/8.
151 146 return (stub_function() == NULL &&
152 147 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3));
153 148 }
154 149
155 150 bool Compile::need_register_stack_bang() const {
156 151 // Determine if we need to generate a register stack overflow check.
157 152 // This is only used on architectures which have split register
158 153 // and memory stacks (ie. IA64).
159 154 // Bang if the method is not a stub function and has java calls
160 155 return (stub_function() == NULL && has_java_calls());
161 156 }
162 157
163 158 # ifdef ENABLE_ZAP_DEAD_LOCALS
164 159
165 160
166 161 // In order to catch compiler oop-map bugs, we have implemented
167 162 // a debugging mode called ZapDeadCompilerLocals.
168 163 // This mode causes the compiler to insert a call to a runtime routine,
169 164 // "zap_dead_locals", right before each place in compiled code
170 165 // that could potentially be a gc-point (i.e., a safepoint or oop map point).
171 166 // The runtime routine checks that locations mapped as oops are really
172 167 // oops, that locations mapped as values do not look like oops,
173 168 // and that locations mapped as dead are not used later
174 169 // (by zapping them to an invalid address).
175 170
176 171 int Compile::_CompiledZap_count = 0;
177 172
178 173 void Compile::Insert_zap_nodes() {
179 174 bool skip = false;
180 175
181 176
182 177 // Dink with static counts because code code without the extra
183 178 // runtime calls is MUCH faster for debugging purposes
184 179
185 180 if ( CompileZapFirst == 0 ) ; // nothing special
186 181 else if ( CompileZapFirst > CompiledZap_count() ) skip = true;
187 182 else if ( CompileZapFirst == CompiledZap_count() )
188 183 warning("starting zap compilation after skipping");
189 184
190 185 if ( CompileZapLast == -1 ) ; // nothing special
191 186 else if ( CompileZapLast < CompiledZap_count() ) skip = true;
192 187 else if ( CompileZapLast == CompiledZap_count() )
193 188 warning("about to compile last zap");
194 189
195 190 ++_CompiledZap_count; // counts skipped zaps, too
196 191
197 192 if ( skip ) return;
198 193
199 194
200 195 if ( _method == NULL )
201 196 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care
202 197
203 198 // Insert call to zap runtime stub before every node with an oop map
204 199 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
205 200 Block *b = _cfg->_blocks[i];
206 201 for ( uint j = 0; j < b->_nodes.size(); ++j ) {
207 202 Node *n = b->_nodes[j];
208 203
209 204 // Determining if we should insert a zap-a-lot node in output.
210 205 // We do that for all nodes that has oopmap info, except for calls
211 206 // to allocation. Calls to allocation passes in the old top-of-eden pointer
212 207 // and expect the C code to reset it. Hence, there can be no safepoints between
213 208 // the inlined-allocation and the call to new_Java, etc.
214 209 // We also cannot zap monitor calls, as they must hold the microlock
215 210 // during the call to Zap, which also wants to grab the microlock.
216 211 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL);
217 212 if ( insert ) { // it is MachSafePoint
218 213 if ( !n->is_MachCall() ) {
219 214 insert = false;
220 215 } else if ( n->is_MachCall() ) {
221 216 MachCallNode* call = n->as_MachCall();
222 217 if (call->entry_point() == OptoRuntime::new_instance_Java() ||
223 218 call->entry_point() == OptoRuntime::new_array_Java() ||
224 219 call->entry_point() == OptoRuntime::multianewarray2_Java() ||
225 220 call->entry_point() == OptoRuntime::multianewarray3_Java() ||
226 221 call->entry_point() == OptoRuntime::multianewarray4_Java() ||
227 222 call->entry_point() == OptoRuntime::multianewarray5_Java() ||
228 223 call->entry_point() == OptoRuntime::slow_arraycopy_Java() ||
229 224 call->entry_point() == OptoRuntime::complete_monitor_locking_Java()
230 225 ) {
231 226 insert = false;
232 227 }
233 228 }
234 229 if (insert) {
235 230 Node *zap = call_zap_node(n->as_MachSafePoint(), i);
236 231 b->_nodes.insert( j, zap );
237 232 _cfg->_bbs.map( zap->_idx, b );
238 233 ++j;
239 234 }
240 235 }
241 236 }
242 237 }
243 238 }
244 239
245 240
246 241 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) {
247 242 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type();
248 243 CallStaticJavaNode* ideal_node =
249 244 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf,
250 245 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()),
251 246 "call zap dead locals stub", 0, TypePtr::BOTTOM);
252 247 // We need to copy the OopMap from the site we're zapping at.
253 248 // We have to make a copy, because the zap site might not be
254 249 // a call site, and zap_dead is a call site.
255 250 OopMap* clone = node_to_check->oop_map()->deep_copy();
256 251
257 252 // Add the cloned OopMap to the zap node
258 253 ideal_node->set_oop_map(clone);
259 254 return _matcher->match_sfpt(ideal_node);
260 255 }
261 256
262 257 //------------------------------is_node_getting_a_safepoint--------------------
263 258 bool Compile::is_node_getting_a_safepoint( Node* n) {
264 259 // This code duplicates the logic prior to the call of add_safepoint
265 260 // below in this file.
266 261 if( n->is_MachSafePoint() ) return true;
267 262 return false;
268 263 }
269 264
270 265 # endif // ENABLE_ZAP_DEAD_LOCALS
271 266
272 267 //------------------------------compute_loop_first_inst_sizes------------------
273 268 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
274 269 // of a loop. When aligning a loop we need to provide enough instructions
275 270 // in cpu's fetch buffer to feed decoders. The loop alignment could be
276 271 // avoided if we have enough instructions in fetch buffer at the head of a loop.
277 272 // By default, the size is set to 999999 by Block's constructor so that
278 273 // a loop will be aligned if the size is not reset here.
279 274 //
280 275 // Note: Mach instructions could contain several HW instructions
281 276 // so the size is estimated only.
282 277 //
283 278 void Compile::compute_loop_first_inst_sizes() {
284 279 // The next condition is used to gate the loop alignment optimization.
285 280 // Don't aligned a loop if there are enough instructions at the head of a loop
286 281 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
287 282 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
288 283 // equal to 11 bytes which is the largest address NOP instruction.
289 284 if( MaxLoopPad < OptoLoopAlignment-1 ) {
290 285 uint last_block = _cfg->_num_blocks-1;
291 286 for( uint i=1; i <= last_block; i++ ) {
292 287 Block *b = _cfg->_blocks[i];
293 288 // Check the first loop's block which requires an alignment.
294 289 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) {
295 290 uint sum_size = 0;
296 291 uint inst_cnt = NumberOfLoopInstrToAlign;
297 292 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
298 293
299 294 // Check subsequent fallthrough blocks if the loop's first
300 295 // block(s) does not have enough instructions.
301 296 Block *nb = b;
302 297 while( inst_cnt > 0 &&
303 298 i < last_block &&
304 299 !_cfg->_blocks[i+1]->has_loop_alignment() &&
305 300 !nb->has_successor(b) ) {
306 301 i++;
307 302 nb = _cfg->_blocks[i];
308 303 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
309 304 } // while( inst_cnt > 0 && i < last_block )
↓ open down ↓ |
247 lines elided |
↑ open up ↑ |
310 305
311 306 b->set_first_inst_size(sum_size);
312 307 } // f( b->head()->is_Loop() )
313 308 } // for( i <= last_block )
314 309 } // if( MaxLoopPad < OptoLoopAlignment-1 )
315 310 }
316 311
317 312 //----------------------Shorten_branches---------------------------------------
318 313 // The architecture description provides short branch variants for some long
319 314 // branch instructions. Replace eligible long branches with short branches.
320 -void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size) {
315 +void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size) {
321 316
322 317 // fill in the nop array for bundling computations
323 318 MachNode *_nop_list[Bundle::_nop_count];
324 319 Bundle::initialize_nops(_nop_list, this);
325 320
326 321 // ------------------
327 322 // Compute size of each block, method size, and relocation information size
328 323 uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks);
329 324 uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1);
330 325 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); )
331 326 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); )
332 327 blk_starts[0] = 0;
333 328
334 329 // Initialize the sizes to 0
335 330 code_size = 0; // Size in bytes of generated code
336 331 stub_size = 0; // Size in bytes of all stub entries
337 332 // Size in bytes of all relocation entries, including those in local stubs.
338 333 // Start with 2-bytes of reloc info for the unvalidated entry point
339 334 reloc_size = 1; // Number of relocation entries
340 - const_size = 0; // size of fp constants in words
341 335
342 336 // Make three passes. The first computes pessimistic blk_starts,
343 - // relative jmp_end, reloc_size and const_size information.
344 - // The second performs short branch substitution using the pessimistic
345 - // sizing. The third inserts nops where needed.
337 + // relative jmp_end and reloc_size information. The second performs
338 + // short branch substitution using the pessimistic sizing. The
339 + // third inserts nops where needed.
346 340
347 341 Node *nj; // tmp
348 342
349 343 // Step one, perform a pessimistic sizing pass.
350 344 uint i;
351 345 uint min_offset_from_last_call = 1; // init to a positive value
352 346 uint nop_size = (new (this) MachNopNode())->size(_regalloc);
353 347 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
354 348 Block *b = _cfg->_blocks[i];
355 349
356 350 // Sum all instruction sizes to compute block size
357 351 uint last_inst = b->_nodes.size();
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
358 352 uint blk_size = 0;
359 353 for( uint j = 0; j<last_inst; j++ ) {
360 354 nj = b->_nodes[j];
361 355 uint inst_size = nj->size(_regalloc);
362 356 blk_size += inst_size;
363 357 // Handle machine instruction nodes
364 358 if( nj->is_Mach() ) {
365 359 MachNode *mach = nj->as_Mach();
366 360 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
367 361 reloc_size += mach->reloc();
368 - const_size += mach->const_size();
369 362 if( mach->is_MachCall() ) {
370 363 MachCallNode *mcall = mach->as_MachCall();
371 364 // This destination address is NOT PC-relative
372 365
373 366 mcall->method_set((intptr_t)mcall->entry_point());
374 367
375 368 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) {
376 369 stub_size += size_java_to_interp();
377 370 reloc_size += reloc_java_to_interp();
378 371 }
379 372 } else if (mach->is_MachSafePoint()) {
380 373 // If call/safepoint are adjacent, account for possible
381 374 // nop to disambiguate the two safepoints.
382 375 if (min_offset_from_last_call == 0) {
383 376 blk_size += nop_size;
384 377 }
385 - } else if (mach->ideal_Opcode() == Op_Jump) {
386 - const_size += b->_num_succs; // Address table size
387 - // The size is valid even for 64 bit since it is
388 - // multiplied by 2*jintSize on this method exit.
389 378 }
390 379 }
391 380 min_offset_from_last_call += inst_size;
392 381 // Remember end of call offset
393 382 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) {
394 383 min_offset_from_last_call = 0;
395 384 }
396 385 }
397 386
398 387 // During short branch replacement, we store the relative (to blk_starts)
399 388 // end of jump in jmp_end, rather than the absolute end of jump. This
400 389 // is so that we do not need to recompute sizes of all nodes when we compute
401 390 // correct blk_starts in our next sizing pass.
402 391 jmp_end[i] = blk_size;
403 392 DEBUG_ONLY( jmp_target[i] = 0; )
404 393
405 394 // When the next block starts a loop, we may insert pad NOP
406 395 // instructions. Since we cannot know our future alignment,
407 396 // assume the worst.
408 397 if( i<_cfg->_num_blocks-1 ) {
409 398 Block *nb = _cfg->_blocks[i+1];
410 399 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
411 400 if( max_loop_pad > 0 ) {
412 401 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
413 402 blk_size += max_loop_pad;
414 403 }
415 404 }
416 405
417 406 // Save block size; update total method size
418 407 blk_starts[i+1] = blk_starts[i]+blk_size;
419 408 }
420 409
421 410 // Step two, replace eligible long jumps.
422 411
423 412 // Note: this will only get the long branches within short branch
424 413 // range. Another pass might detect more branches that became
425 414 // candidates because the shortening in the first pass exposed
426 415 // more opportunities. Unfortunately, this would require
427 416 // recomputing the starting and ending positions for the blocks
428 417 for( i=0; i<_cfg->_num_blocks; i++ ) {
429 418 Block *b = _cfg->_blocks[i];
430 419
431 420 int j;
432 421 // Find the branch; ignore trailing NOPs.
433 422 for( j = b->_nodes.size()-1; j>=0; j-- ) {
434 423 nj = b->_nodes[j];
435 424 if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con )
436 425 break;
437 426 }
438 427
439 428 if (j >= 0) {
440 429 if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) {
441 430 MachNode *mach = nj->as_Mach();
442 431 // This requires the TRUE branch target be in succs[0]
443 432 uint bnum = b->non_connector_successor(0)->_pre_order;
444 433 uintptr_t target = blk_starts[bnum];
445 434 if( mach->is_pc_relative() ) {
446 435 int offset = target-(blk_starts[i] + jmp_end[i]);
447 436 if (_matcher->is_short_branch_offset(mach->rule(), offset)) {
448 437 // We've got a winner. Replace this branch.
449 438 MachNode* replacement = mach->short_branch_version(this);
450 439 b->_nodes.map(j, replacement);
451 440 mach->subsume_by(replacement);
452 441
453 442 // Update the jmp_end size to save time in our
454 443 // next pass.
455 444 jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc));
456 445 DEBUG_ONLY( jmp_target[i] = bnum; );
457 446 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
458 447 }
459 448 } else {
460 449 #ifndef PRODUCT
461 450 mach->dump(3);
462 451 #endif
463 452 Unimplemented();
464 453 }
465 454 }
466 455 }
467 456 }
468 457
469 458 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
470 459 // of a loop. It is used to determine the padding for loop alignment.
471 460 compute_loop_first_inst_sizes();
472 461
473 462 // Step 3, compute the offsets of all the labels
474 463 uint last_call_adr = max_uint;
475 464 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
476 465 // copy the offset of the beginning to the corresponding label
477 466 assert(labels[i].is_unused(), "cannot patch at this point");
478 467 labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS);
479 468
480 469 // insert padding for any instructions that need it
481 470 Block *b = _cfg->_blocks[i];
482 471 uint last_inst = b->_nodes.size();
483 472 uint adr = blk_starts[i];
484 473 for( uint j = 0; j<last_inst; j++ ) {
485 474 nj = b->_nodes[j];
486 475 if( nj->is_Mach() ) {
487 476 int padding = nj->as_Mach()->compute_padding(adr);
488 477 // If call/safepoint are adjacent insert a nop (5010568)
489 478 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() &&
490 479 adr == last_call_adr ) {
491 480 padding = nop_size;
492 481 }
493 482 if(padding > 0) {
494 483 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
495 484 int nops_cnt = padding / nop_size;
496 485 MachNode *nop = new (this) MachNopNode(nops_cnt);
497 486 b->_nodes.insert(j++, nop);
498 487 _cfg->_bbs.map( nop->_idx, b );
499 488 adr += padding;
500 489 last_inst++;
501 490 }
502 491 }
503 492 adr += nj->size(_regalloc);
504 493
505 494 // Remember end of call offset
506 495 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) {
507 496 last_call_adr = adr;
508 497 }
509 498 }
510 499
511 500 if ( i != _cfg->_num_blocks-1) {
512 501 // Get the size of the block
513 502 uint blk_size = adr - blk_starts[i];
514 503
515 504 // When the next block is the top of a loop, we may insert pad NOP
516 505 // instructions.
517 506 Block *nb = _cfg->_blocks[i+1];
518 507 int current_offset = blk_starts[i] + blk_size;
519 508 current_offset += nb->alignment_padding(current_offset);
520 509 // Save block size; update total method size
521 510 blk_starts[i+1] = current_offset;
522 511 }
523 512 }
524 513
525 514 #ifdef ASSERT
526 515 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks
527 516 if( jmp_target[i] != 0 ) {
528 517 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]);
529 518 if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) {
530 519 tty->print_cr("target (%d) - jmp_end(%d) = offset (%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_end[i], offset, i, jmp_target[i]);
531 520 }
532 521 assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp");
533 522 }
534 523 }
535 524 #endif
536 525
537 526 // ------------------
538 527 // Compute size for code buffer
↓ open down ↓ |
140 lines elided |
↑ open up ↑ |
539 528 code_size = blk_starts[i-1] + jmp_end[i-1];
540 529
541 530 // Relocation records
542 531 reloc_size += 1; // Relo entry for exception handler
543 532
544 533 // Adjust reloc_size to number of record of relocation info
545 534 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
546 535 // a relocation index.
547 536 // The CodeBuffer will expand the locs array if this estimate is too low.
548 537 reloc_size *= 10 / sizeof(relocInfo);
549 -
550 - // Adjust const_size to number of bytes
551 - const_size *= 2*jintSize; // both float and double take two words per entry
552 -
553 538 }
554 539
555 540 //------------------------------FillLocArray-----------------------------------
556 541 // Create a bit of debug info and append it to the array. The mapping is from
557 542 // Java local or expression stack to constant, register or stack-slot. For
558 543 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
559 544 // entry has been taken care of and caller should skip it).
560 545 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
561 546 // This should never have accepted Bad before
562 547 assert(OptoReg::is_valid(regnum), "location must be valid");
563 548 return (OptoReg::is_reg(regnum))
564 549 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
565 550 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
566 551 }
567 552
568 553
569 554 ObjectValue*
570 555 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
571 556 for (int i = 0; i < objs->length(); i++) {
572 557 assert(objs->at(i)->is_object(), "corrupt object cache");
573 558 ObjectValue* sv = (ObjectValue*) objs->at(i);
574 559 if (sv->id() == id) {
575 560 return sv;
576 561 }
577 562 }
578 563 // Otherwise..
579 564 return NULL;
580 565 }
581 566
582 567 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
583 568 ObjectValue* sv ) {
584 569 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
585 570 objs->append(sv);
586 571 }
587 572
588 573
589 574 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
590 575 GrowableArray<ScopeValue*> *array,
591 576 GrowableArray<ScopeValue*> *objs ) {
592 577 assert( local, "use _top instead of null" );
593 578 if (array->length() != idx) {
594 579 assert(array->length() == idx + 1, "Unexpected array count");
595 580 // Old functionality:
596 581 // return
597 582 // New functionality:
598 583 // Assert if the local is not top. In product mode let the new node
599 584 // override the old entry.
600 585 assert(local == top(), "LocArray collision");
601 586 if (local == top()) {
602 587 return;
603 588 }
604 589 array->pop();
605 590 }
606 591 const Type *t = local->bottom_type();
607 592
608 593 // Is it a safepoint scalar object node?
609 594 if (local->is_SafePointScalarObject()) {
610 595 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
611 596
612 597 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
613 598 if (sv == NULL) {
614 599 ciKlass* cik = t->is_oopptr()->klass();
615 600 assert(cik->is_instance_klass() ||
616 601 cik->is_array_klass(), "Not supported allocation.");
617 602 sv = new ObjectValue(spobj->_idx,
618 603 new ConstantOopWriteValue(cik->constant_encoding()));
619 604 Compile::set_sv_for_object_node(objs, sv);
620 605
621 606 uint first_ind = spobj->first_index();
622 607 for (uint i = 0; i < spobj->n_fields(); i++) {
623 608 Node* fld_node = sfpt->in(first_ind+i);
624 609 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
625 610 }
626 611 }
627 612 array->append(sv);
628 613 return;
629 614 }
630 615
631 616 // Grab the register number for the local
632 617 OptoReg::Name regnum = _regalloc->get_reg_first(local);
633 618 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
634 619 // Record the double as two float registers.
635 620 // The register mask for such a value always specifies two adjacent
636 621 // float registers, with the lower register number even.
637 622 // Normally, the allocation of high and low words to these registers
638 623 // is irrelevant, because nearly all operations on register pairs
639 624 // (e.g., StoreD) treat them as a single unit.
640 625 // Here, we assume in addition that the words in these two registers
641 626 // stored "naturally" (by operations like StoreD and double stores
642 627 // within the interpreter) such that the lower-numbered register
643 628 // is written to the lower memory address. This may seem like
644 629 // a machine dependency, but it is not--it is a requirement on
645 630 // the author of the <arch>.ad file to ensure that, for every
646 631 // even/odd double-register pair to which a double may be allocated,
647 632 // the word in the even single-register is stored to the first
648 633 // memory word. (Note that register numbers are completely
649 634 // arbitrary, and are not tied to any machine-level encodings.)
650 635 #ifdef _LP64
651 636 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
652 637 array->append(new ConstantIntValue(0));
653 638 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
654 639 } else if ( t->base() == Type::Long ) {
655 640 array->append(new ConstantIntValue(0));
656 641 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
657 642 } else if ( t->base() == Type::RawPtr ) {
658 643 // jsr/ret return address which must be restored into a the full
659 644 // width 64-bit stack slot.
660 645 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
661 646 }
662 647 #else //_LP64
663 648 #ifdef SPARC
664 649 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
665 650 // For SPARC we have to swap high and low words for
666 651 // long values stored in a single-register (g0-g7).
667 652 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
668 653 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
669 654 } else
670 655 #endif //SPARC
671 656 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
672 657 // Repack the double/long as two jints.
673 658 // The convention the interpreter uses is that the second local
674 659 // holds the first raw word of the native double representation.
675 660 // This is actually reasonable, since locals and stack arrays
676 661 // grow downwards in all implementations.
677 662 // (If, on some machine, the interpreter's Java locals or stack
678 663 // were to grow upwards, the embedded doubles would be word-swapped.)
679 664 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
680 665 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
681 666 }
682 667 #endif //_LP64
683 668 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
684 669 OptoReg::is_reg(regnum) ) {
685 670 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
686 671 ? Location::float_in_dbl : Location::normal ));
687 672 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
688 673 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
689 674 ? Location::int_in_long : Location::normal ));
690 675 } else if( t->base() == Type::NarrowOop ) {
691 676 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
692 677 } else {
693 678 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
694 679 }
695 680 return;
696 681 }
697 682
698 683 // No register. It must be constant data.
699 684 switch (t->base()) {
700 685 case Type::Half: // Second half of a double
701 686 ShouldNotReachHere(); // Caller should skip 2nd halves
702 687 break;
703 688 case Type::AnyPtr:
704 689 array->append(new ConstantOopWriteValue(NULL));
705 690 break;
706 691 case Type::AryPtr:
707 692 case Type::InstPtr:
708 693 case Type::KlassPtr: // fall through
709 694 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
710 695 break;
711 696 case Type::NarrowOop:
712 697 if (t == TypeNarrowOop::NULL_PTR) {
713 698 array->append(new ConstantOopWriteValue(NULL));
714 699 } else {
715 700 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
716 701 }
717 702 break;
718 703 case Type::Int:
719 704 array->append(new ConstantIntValue(t->is_int()->get_con()));
720 705 break;
721 706 case Type::RawPtr:
722 707 // A return address (T_ADDRESS).
723 708 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
724 709 #ifdef _LP64
725 710 // Must be restored to the full-width 64-bit stack slot.
726 711 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
727 712 #else
728 713 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
729 714 #endif
730 715 break;
731 716 case Type::FloatCon: {
732 717 float f = t->is_float_constant()->getf();
733 718 array->append(new ConstantIntValue(jint_cast(f)));
734 719 break;
735 720 }
736 721 case Type::DoubleCon: {
737 722 jdouble d = t->is_double_constant()->getd();
738 723 #ifdef _LP64
739 724 array->append(new ConstantIntValue(0));
740 725 array->append(new ConstantDoubleValue(d));
741 726 #else
742 727 // Repack the double as two jints.
743 728 // The convention the interpreter uses is that the second local
744 729 // holds the first raw word of the native double representation.
745 730 // This is actually reasonable, since locals and stack arrays
746 731 // grow downwards in all implementations.
747 732 // (If, on some machine, the interpreter's Java locals or stack
748 733 // were to grow upwards, the embedded doubles would be word-swapped.)
749 734 jint *dp = (jint*)&d;
750 735 array->append(new ConstantIntValue(dp[1]));
751 736 array->append(new ConstantIntValue(dp[0]));
752 737 #endif
753 738 break;
754 739 }
755 740 case Type::Long: {
756 741 jlong d = t->is_long()->get_con();
757 742 #ifdef _LP64
758 743 array->append(new ConstantIntValue(0));
759 744 array->append(new ConstantLongValue(d));
760 745 #else
761 746 // Repack the long as two jints.
762 747 // The convention the interpreter uses is that the second local
763 748 // holds the first raw word of the native double representation.
764 749 // This is actually reasonable, since locals and stack arrays
765 750 // grow downwards in all implementations.
766 751 // (If, on some machine, the interpreter's Java locals or stack
767 752 // were to grow upwards, the embedded doubles would be word-swapped.)
768 753 jint *dp = (jint*)&d;
769 754 array->append(new ConstantIntValue(dp[1]));
770 755 array->append(new ConstantIntValue(dp[0]));
771 756 #endif
772 757 break;
773 758 }
774 759 case Type::Top: // Add an illegal value here
775 760 array->append(new LocationValue(Location()));
776 761 break;
777 762 default:
778 763 ShouldNotReachHere();
779 764 break;
780 765 }
781 766 }
782 767
783 768 // Determine if this node starts a bundle
784 769 bool Compile::starts_bundle(const Node *n) const {
785 770 return (_node_bundling_limit > n->_idx &&
786 771 _node_bundling_base[n->_idx].starts_bundle());
787 772 }
788 773
789 774 //--------------------------Process_OopMap_Node--------------------------------
790 775 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
791 776
792 777 // Handle special safepoint nodes for synchronization
793 778 MachSafePointNode *sfn = mach->as_MachSafePoint();
794 779 MachCallNode *mcall;
795 780
796 781 #ifdef ENABLE_ZAP_DEAD_LOCALS
797 782 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative");
798 783 #endif
799 784
800 785 int safepoint_pc_offset = current_offset;
801 786 bool is_method_handle_invoke = false;
802 787 bool return_oop = false;
803 788
804 789 // Add the safepoint in the DebugInfoRecorder
805 790 if( !mach->is_MachCall() ) {
806 791 mcall = NULL;
807 792 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
808 793 } else {
809 794 mcall = mach->as_MachCall();
810 795
811 796 // Is the call a MethodHandle call?
812 797 if (mcall->is_MachCallJava()) {
813 798 if (mcall->as_MachCallJava()->_method_handle_invoke) {
814 799 assert(has_method_handle_invokes(), "must have been set during call generation");
815 800 is_method_handle_invoke = true;
816 801 }
817 802 }
818 803
819 804 // Check if a call returns an object.
820 805 if (mcall->return_value_is_used() &&
821 806 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) {
822 807 return_oop = true;
823 808 }
824 809 safepoint_pc_offset += mcall->ret_addr_offset();
825 810 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
826 811 }
827 812
828 813 // Loop over the JVMState list to add scope information
829 814 // Do not skip safepoints with a NULL method, they need monitor info
830 815 JVMState* youngest_jvms = sfn->jvms();
831 816 int max_depth = youngest_jvms->depth();
832 817
833 818 // Allocate the object pool for scalar-replaced objects -- the map from
834 819 // small-integer keys (which can be recorded in the local and ostack
835 820 // arrays) to descriptions of the object state.
836 821 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
837 822
838 823 // Visit scopes from oldest to youngest.
839 824 for (int depth = 1; depth <= max_depth; depth++) {
840 825 JVMState* jvms = youngest_jvms->of_depth(depth);
841 826 int idx;
842 827 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
843 828 // Safepoints that do not have method() set only provide oop-map and monitor info
844 829 // to support GC; these do not support deoptimization.
845 830 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
846 831 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
847 832 int num_mon = jvms->nof_monitors();
848 833 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
849 834 "JVMS local count must match that of the method");
850 835
851 836 // Add Local and Expression Stack Information
852 837
853 838 // Insert locals into the locarray
854 839 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
855 840 for( idx = 0; idx < num_locs; idx++ ) {
856 841 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
857 842 }
858 843
859 844 // Insert expression stack entries into the exparray
860 845 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
861 846 for( idx = 0; idx < num_exps; idx++ ) {
862 847 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
863 848 }
864 849
865 850 // Add in mappings of the monitors
866 851 assert( !method ||
867 852 !method->is_synchronized() ||
868 853 method->is_native() ||
869 854 num_mon > 0 ||
870 855 !GenerateSynchronizationCode,
871 856 "monitors must always exist for synchronized methods");
872 857
873 858 // Build the growable array of ScopeValues for exp stack
874 859 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
875 860
876 861 // Loop over monitors and insert into array
877 862 for(idx = 0; idx < num_mon; idx++) {
878 863 // Grab the node that defines this monitor
879 864 Node* box_node = sfn->monitor_box(jvms, idx);
880 865 Node* obj_node = sfn->monitor_obj(jvms, idx);
881 866
882 867 // Create ScopeValue for object
883 868 ScopeValue *scval = NULL;
884 869
885 870 if( obj_node->is_SafePointScalarObject() ) {
886 871 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
887 872 scval = Compile::sv_for_node_id(objs, spobj->_idx);
888 873 if (scval == NULL) {
889 874 const Type *t = obj_node->bottom_type();
890 875 ciKlass* cik = t->is_oopptr()->klass();
891 876 assert(cik->is_instance_klass() ||
892 877 cik->is_array_klass(), "Not supported allocation.");
893 878 ObjectValue* sv = new ObjectValue(spobj->_idx,
894 879 new ConstantOopWriteValue(cik->constant_encoding()));
895 880 Compile::set_sv_for_object_node(objs, sv);
896 881
897 882 uint first_ind = spobj->first_index();
898 883 for (uint i = 0; i < spobj->n_fields(); i++) {
899 884 Node* fld_node = sfn->in(first_ind+i);
900 885 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
901 886 }
902 887 scval = sv;
903 888 }
904 889 } else if( !obj_node->is_Con() ) {
905 890 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
906 891 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
907 892 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
908 893 } else {
909 894 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
910 895 }
911 896 } else {
912 897 const TypePtr *tp = obj_node->bottom_type()->make_ptr();
913 898 scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->constant_encoding());
914 899 }
915 900
916 901 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node);
917 902 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
918 903 while( !box_node->is_BoxLock() ) box_node = box_node->in(1);
919 904 monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated()));
920 905 }
921 906
922 907 // We dump the object pool first, since deoptimization reads it in first.
923 908 debug_info()->dump_object_pool(objs);
924 909
925 910 // Build first class objects to pass to scope
926 911 DebugToken *locvals = debug_info()->create_scope_values(locarray);
927 912 DebugToken *expvals = debug_info()->create_scope_values(exparray);
928 913 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
929 914
930 915 // Make method available for all Safepoints
931 916 ciMethod* scope_method = method ? method : _method;
932 917 // Describe the scope here
933 918 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
934 919 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
935 920 // Now we can describe the scope.
936 921 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals);
937 922 } // End jvms loop
938 923
939 924 // Mark the end of the scope set.
940 925 debug_info()->end_safepoint(safepoint_pc_offset);
941 926 }
942 927
943 928
944 929
945 930 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
946 931 class NonSafepointEmitter {
947 932 Compile* C;
948 933 JVMState* _pending_jvms;
949 934 int _pending_offset;
950 935
951 936 void emit_non_safepoint();
952 937
953 938 public:
954 939 NonSafepointEmitter(Compile* compile) {
955 940 this->C = compile;
956 941 _pending_jvms = NULL;
957 942 _pending_offset = 0;
958 943 }
959 944
960 945 void observe_instruction(Node* n, int pc_offset) {
961 946 if (!C->debug_info()->recording_non_safepoints()) return;
962 947
963 948 Node_Notes* nn = C->node_notes_at(n->_idx);
964 949 if (nn == NULL || nn->jvms() == NULL) return;
965 950 if (_pending_jvms != NULL &&
966 951 _pending_jvms->same_calls_as(nn->jvms())) {
967 952 // Repeated JVMS? Stretch it up here.
968 953 _pending_offset = pc_offset;
969 954 } else {
970 955 if (_pending_jvms != NULL &&
971 956 _pending_offset < pc_offset) {
972 957 emit_non_safepoint();
973 958 }
974 959 _pending_jvms = NULL;
975 960 if (pc_offset > C->debug_info()->last_pc_offset()) {
976 961 // This is the only way _pending_jvms can become non-NULL:
977 962 _pending_jvms = nn->jvms();
978 963 _pending_offset = pc_offset;
979 964 }
980 965 }
981 966 }
982 967
983 968 // Stay out of the way of real safepoints:
984 969 void observe_safepoint(JVMState* jvms, int pc_offset) {
985 970 if (_pending_jvms != NULL &&
986 971 !_pending_jvms->same_calls_as(jvms) &&
987 972 _pending_offset < pc_offset) {
988 973 emit_non_safepoint();
989 974 }
990 975 _pending_jvms = NULL;
991 976 }
992 977
993 978 void flush_at_end() {
994 979 if (_pending_jvms != NULL) {
995 980 emit_non_safepoint();
996 981 }
997 982 _pending_jvms = NULL;
998 983 }
999 984 };
1000 985
1001 986 void NonSafepointEmitter::emit_non_safepoint() {
1002 987 JVMState* youngest_jvms = _pending_jvms;
1003 988 int pc_offset = _pending_offset;
1004 989
1005 990 // Clear it now:
1006 991 _pending_jvms = NULL;
1007 992
1008 993 DebugInformationRecorder* debug_info = C->debug_info();
1009 994 assert(debug_info->recording_non_safepoints(), "sanity");
1010 995
1011 996 debug_info->add_non_safepoint(pc_offset);
1012 997 int max_depth = youngest_jvms->depth();
1013 998
1014 999 // Visit scopes from oldest to youngest.
1015 1000 for (int depth = 1; depth <= max_depth; depth++) {
1016 1001 JVMState* jvms = youngest_jvms->of_depth(depth);
1017 1002 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
1018 1003 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
1019 1004 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute());
1020 1005 }
1021 1006
1022 1007 // Mark the end of the scope set.
1023 1008 debug_info->end_non_safepoint(pc_offset);
1024 1009 }
1025 1010
1026 1011
1027 1012
1028 1013 // helper for Fill_buffer bailout logic
1029 1014 static void turn_off_compiler(Compile* C) {
1030 1015 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) {
1031 1016 // Do not turn off compilation if a single giant method has
1032 1017 // blown the code cache size.
1033 1018 C->record_failure("excessive request to CodeCache");
1034 1019 } else {
1035 1020 // Let CompilerBroker disable further compilations.
1036 1021 C->record_failure("CodeCache is full");
1037 1022 }
1038 1023 }
1039 1024
1040 1025
1041 1026 //------------------------------Fill_buffer------------------------------------
1042 1027 void Compile::Fill_buffer() {
1043 1028
1044 1029 // Set the initially allocated size
1045 1030 int code_req = initial_code_capacity;
1046 1031 int locs_req = initial_locs_capacity;
1047 1032 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity;
1048 1033 int const_req = initial_const_capacity;
1049 1034 bool labels_not_set = true;
1050 1035
1051 1036 int pad_req = NativeCall::instruction_size;
1052 1037 // The extra spacing after the code is necessary on some platforms.
1053 1038 // Sometimes we need to patch in a jump after the last instruction,
1054 1039 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
1055 1040
1056 1041 uint i;
1057 1042 // Compute the byte offset where we can store the deopt pc.
1058 1043 if (fixed_slots() != 0) {
1059 1044 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
1060 1045 }
1061 1046
1062 1047 // Compute prolog code size
1063 1048 _method_size = 0;
1064 1049 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1065 1050 #ifdef IA64
1066 1051 if (save_argument_registers()) {
1067 1052 // 4815101: this is a stub with implicit and unknown precision fp args.
1068 1053 // The usual spill mechanism can only generate stfd's in this case, which
1069 1054 // doesn't work if the fp reg to spill contains a single-precision denorm.
1070 1055 // Instead, we hack around the normal spill mechanism using stfspill's and
1071 1056 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1072 1057 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1073 1058 //
1074 1059 // If we ever implement 16-byte 'registers' == stack slots, we can
1075 1060 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1076 1061 // instead of stfd/stfs/ldfd/ldfs.
1077 1062 _frame_slots += 8*(16/BytesPerInt);
1078 1063 }
↓ open down ↓ |
516 lines elided |
↑ open up ↑ |
1079 1064 #endif
1080 1065 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" );
1081 1066
1082 1067 // Create an array of unused labels, one for each basic block
1083 1068 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1);
1084 1069
1085 1070 for( i=0; i <= _cfg->_num_blocks; i++ ) {
1086 1071 blk_labels[i].init();
1087 1072 }
1088 1073
1074 + if (has_mach_constant_base_node()) {
1075 + // Fill the constant table.
1076 + // Note: This must happen before Shorten_branches.
1077 + for (i = 0; i < _cfg->_num_blocks; i++) {
1078 + Block* b = _cfg->_blocks[i];
1079 +
1080 + for (uint j = 0; j < b->_nodes.size(); j++) {
1081 + Node* n = b->_nodes[j];
1082 +
1083 + if (n->is_Mach()) {
1084 + MachNode *mach = n->as_Mach();
1085 +
1086 + // If the MachNode is a MachConstantNode evaluate the
1087 + // constant value section.
1088 + if (mach->is_MachConstant()) {
1089 + MachConstantNode* machcon = mach->as_MachConstant();
1090 + machcon->eval_constant();
1091 + }
1092 + }
1093 + }
1094 + }
1095 +
1096 + // Calculate the size of the constant table (including the padding
1097 + // to the next section).
1098 + const_req = mach_constant_base_node()->calculate_constant_table_size();
1099 + }
1100 +
1101 + // Initialize the space for the BufferBlob used to find and verify
1102 + // instruction size in MachNode::emit_size()
1103 + init_scratch_buffer_blob(const_req);
1104 + if (failing()) return; // Out of memory
1105 +
1089 1106 // If this machine supports different size branch offsets, then pre-compute
1090 1107 // the length of the blocks
1091 1108 if( _matcher->is_short_branch_offset(-1, 0) ) {
1092 - Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req);
1109 + Shorten_branches(blk_labels, code_req, locs_req, stub_req);
1093 1110 labels_not_set = false;
1094 1111 }
1095 1112
1096 1113 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1097 1114 int exception_handler_req = size_exception_handler();
1098 1115 int deopt_handler_req = size_deopt_handler();
1099 1116 exception_handler_req += MAX_stubs_size; // add marginal slop for handler
1100 1117 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler
1101 1118 stub_req += MAX_stubs_size; // ensure per-stub margin
1102 1119 code_req += MAX_inst_size; // ensure per-instruction margin
1103 1120
1104 1121 if (StressCodeBuffers)
1105 1122 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1106 1123
1107 1124 int total_req =
1125 + const_req +
1108 1126 code_req +
1109 1127 pad_req +
1110 1128 stub_req +
1111 1129 exception_handler_req +
1112 - deopt_handler_req + // deopt handler
1113 - const_req;
1130 + deopt_handler_req; // deopt handler
1114 1131
1115 1132 if (has_method_handle_invokes())
1116 1133 total_req += deopt_handler_req; // deopt MH handler
1117 1134
1118 1135 CodeBuffer* cb = code_buffer();
1119 1136 cb->initialize(total_req, locs_req);
1120 1137
1121 1138 // Have we run out of code space?
1122 1139 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1123 1140 turn_off_compiler(this);
1124 1141 return;
1125 1142 }
1126 1143 // Configure the code buffer.
1127 1144 cb->initialize_consts_size(const_req);
1128 1145 cb->initialize_stubs_size(stub_req);
1129 1146 cb->initialize_oop_recorder(env()->oop_recorder());
1130 1147
1131 1148 // fill in the nop array for bundling computations
1132 1149 MachNode *_nop_list[Bundle::_nop_count];
1133 1150 Bundle::initialize_nops(_nop_list, this);
1134 1151
1135 1152 // Create oopmap set.
1136 1153 _oop_map_set = new OopMapSet();
1137 1154
1138 1155 // !!!!! This preserves old handling of oopmaps for now
1139 1156 debug_info()->set_oopmaps(_oop_map_set);
1140 1157
1141 1158 // Count and start of implicit null check instructions
1142 1159 uint inct_cnt = 0;
1143 1160 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1);
1144 1161
1145 1162 // Count and start of calls
1146 1163 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1);
1147 1164
1148 1165 uint return_offset = 0;
1149 1166 int nop_size = (new (this) MachNopNode())->size(_regalloc);
1150 1167
1151 1168 int previous_offset = 0;
1152 1169 int current_offset = 0;
1153 1170 int last_call_offset = -1;
1154 1171
1155 1172 // Create an array of unused labels, one for each basic block, if printing is enabled
1156 1173 #ifndef PRODUCT
1157 1174 int *node_offsets = NULL;
1158 1175 uint node_offset_limit = unique();
1159 1176
1160 1177
1161 1178 if ( print_assembly() )
1162 1179 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1163 1180 #endif
1164 1181
1165 1182 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1166 1183
1167 1184 // ------------------
1168 1185 // Now fill in the code buffer
1169 1186 Node *delay_slot = NULL;
1170 1187
1171 1188 for( i=0; i < _cfg->_num_blocks; i++ ) {
1172 1189 Block *b = _cfg->_blocks[i];
↓ open down ↓ |
49 lines elided |
↑ open up ↑ |
1173 1190
1174 1191 Node *head = b->head();
1175 1192
1176 1193 // If this block needs to start aligned (i.e, can be reached other
1177 1194 // than by falling-thru from the previous block), then force the
1178 1195 // start of a new bundle.
1179 1196 if( Pipeline::requires_bundling() && starts_bundle(head) )
1180 1197 cb->flush_bundle(true);
1181 1198
1182 1199 // Define the label at the beginning of the basic block
1183 - if( labels_not_set )
1184 - MacroAssembler(cb).bind( blk_labels[b->_pre_order] );
1185 -
1186 - else
1187 - assert( blk_labels[b->_pre_order].loc_pos() == cb->insts_size(),
1188 - "label position does not match code offset" );
1200 + if (labels_not_set) {
1201 + MacroAssembler(cb).bind(blk_labels[b->_pre_order]);
1202 + } else {
1203 + assert(blk_labels[b->_pre_order].loc_pos() == cb->insts_size(),
1204 + err_msg("label position does not match code offset: %d != %d",
1205 + blk_labels[b->_pre_order].loc_pos(), cb->insts_size()));
1206 + }
1189 1207
1190 1208 uint last_inst = b->_nodes.size();
1191 1209
1192 1210 // Emit block normally, except for last instruction.
1193 1211 // Emit means "dump code bits into code buffer".
1194 1212 for( uint j = 0; j<last_inst; j++ ) {
1195 1213
1196 1214 // Get the node
1197 1215 Node* n = b->_nodes[j];
1198 1216
1199 1217 // See if delay slots are supported
1200 1218 if (valid_bundle_info(n) &&
1201 1219 node_bundling(n)->used_in_unconditional_delay()) {
1202 1220 assert(delay_slot == NULL, "no use of delay slot node");
1203 1221 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1204 1222
1205 1223 delay_slot = n;
1206 1224 continue;
1207 1225 }
1208 1226
1209 1227 // If this starts a new instruction group, then flush the current one
1210 1228 // (but allow split bundles)
1211 1229 if( Pipeline::requires_bundling() && starts_bundle(n) )
1212 1230 cb->flush_bundle(false);
1213 1231
1214 1232 // The following logic is duplicated in the code ifdeffed for
1215 1233 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It
1216 1234 // should be factored out. Or maybe dispersed to the nodes?
1217 1235
1218 1236 // Special handling for SafePoint/Call Nodes
1219 1237 bool is_mcall = false;
1220 1238 if( n->is_Mach() ) {
1221 1239 MachNode *mach = n->as_Mach();
1222 1240 is_mcall = n->is_MachCall();
1223 1241 bool is_sfn = n->is_MachSafePoint();
1224 1242
1225 1243 // If this requires all previous instructions be flushed, then do so
1226 1244 if( is_sfn || is_mcall || mach->alignment_required() != 1) {
1227 1245 cb->flush_bundle(true);
1228 1246 current_offset = cb->insts_size();
1229 1247 }
1230 1248
1231 1249 // align the instruction if necessary
1232 1250 int padding = mach->compute_padding(current_offset);
1233 1251 // Make sure safepoint node for polling is distinct from a call's
1234 1252 // return by adding a nop if needed.
1235 1253 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) {
1236 1254 padding = nop_size;
1237 1255 }
1238 1256 assert( labels_not_set || padding == 0, "instruction should already be aligned");
1239 1257
1240 1258 if(padding > 0) {
1241 1259 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1242 1260 int nops_cnt = padding / nop_size;
1243 1261 MachNode *nop = new (this) MachNopNode(nops_cnt);
1244 1262 b->_nodes.insert(j++, nop);
1245 1263 last_inst++;
1246 1264 _cfg->_bbs.map( nop->_idx, b );
1247 1265 nop->emit(*cb, _regalloc);
1248 1266 cb->flush_bundle(true);
1249 1267 current_offset = cb->insts_size();
1250 1268 }
1251 1269
1252 1270 // Remember the start of the last call in a basic block
1253 1271 if (is_mcall) {
1254 1272 MachCallNode *mcall = mach->as_MachCall();
1255 1273
1256 1274 // This destination address is NOT PC-relative
1257 1275 mcall->method_set((intptr_t)mcall->entry_point());
1258 1276
1259 1277 // Save the return address
1260 1278 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset();
1261 1279
1262 1280 if (!mcall->is_safepoint_node()) {
1263 1281 is_mcall = false;
1264 1282 is_sfn = false;
1265 1283 }
1266 1284 }
1267 1285
1268 1286 // sfn will be valid whenever mcall is valid now because of inheritance
1269 1287 if( is_sfn || is_mcall ) {
1270 1288
1271 1289 // Handle special safepoint nodes for synchronization
1272 1290 if( !is_mcall ) {
1273 1291 MachSafePointNode *sfn = mach->as_MachSafePoint();
1274 1292 // !!!!! Stubs only need an oopmap right now, so bail out
1275 1293 if( sfn->jvms()->method() == NULL) {
1276 1294 // Write the oopmap directly to the code blob??!!
1277 1295 # ifdef ENABLE_ZAP_DEAD_LOCALS
1278 1296 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive");
1279 1297 # endif
1280 1298 continue;
1281 1299 }
1282 1300 } // End synchronization
1283 1301
1284 1302 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1285 1303 current_offset);
1286 1304 Process_OopMap_Node(mach, current_offset);
1287 1305 } // End if safepoint
1288 1306
1289 1307 // If this is a null check, then add the start of the previous instruction to the list
1290 1308 else if( mach->is_MachNullCheck() ) {
1291 1309 inct_starts[inct_cnt++] = previous_offset;
1292 1310 }
1293 1311
1294 1312 // If this is a branch, then fill in the label with the target BB's label
1295 1313 else if ( mach->is_Branch() ) {
1296 1314
1297 1315 if ( mach->ideal_Opcode() == Op_Jump ) {
1298 1316 for (uint h = 0; h < b->_num_succs; h++ ) {
1299 1317 Block* succs_block = b->_succs[h];
1300 1318 for (uint j = 1; j < succs_block->num_preds(); j++) {
1301 1319 Node* jpn = succs_block->pred(j);
1302 1320 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) {
1303 1321 uint block_num = succs_block->non_connector()->_pre_order;
1304 1322 Label *blkLabel = &blk_labels[block_num];
1305 1323 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1306 1324 }
1307 1325 }
1308 1326 }
1309 1327 } else {
1310 1328 // For Branchs
1311 1329 // This requires the TRUE branch target be in succs[0]
1312 1330 uint block_num = b->non_connector_successor(0)->_pre_order;
1313 1331 mach->label_set( blk_labels[block_num], block_num );
1314 1332 }
1315 1333 }
1316 1334
1317 1335 #ifdef ASSERT
1318 1336 // Check that oop-store precedes the card-mark
1319 1337 else if( mach->ideal_Opcode() == Op_StoreCM ) {
1320 1338 uint storeCM_idx = j;
1321 1339 Node *oop_store = mach->in(mach->_cnt); // First precedence edge
1322 1340 assert( oop_store != NULL, "storeCM expects a precedence edge");
1323 1341 uint i4;
1324 1342 for( i4 = 0; i4 < last_inst; ++i4 ) {
1325 1343 if( b->_nodes[i4] == oop_store ) break;
1326 1344 }
1327 1345 // Note: This test can provide a false failure if other precedence
1328 1346 // edges have been added to the storeCMNode.
1329 1347 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1330 1348 }
1331 1349 #endif
1332 1350
1333 1351 else if( !n->is_Proj() ) {
1334 1352 // Remember the beginning of the previous instruction, in case
1335 1353 // it's followed by a flag-kill and a null-check. Happens on
1336 1354 // Intel all the time, with add-to-memory kind of opcodes.
1337 1355 previous_offset = current_offset;
1338 1356 }
1339 1357 }
1340 1358
1341 1359 // Verify that there is sufficient space remaining
1342 1360 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1343 1361 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1344 1362 turn_off_compiler(this);
1345 1363 return;
1346 1364 }
1347 1365
1348 1366 // Save the offset for the listing
1349 1367 #ifndef PRODUCT
1350 1368 if( node_offsets && n->_idx < node_offset_limit )
1351 1369 node_offsets[n->_idx] = cb->insts_size();
1352 1370 #endif
1353 1371
1354 1372 // "Normal" instruction case
1355 1373 n->emit(*cb, _regalloc);
1356 1374 current_offset = cb->insts_size();
1357 1375 non_safepoints.observe_instruction(n, current_offset);
1358 1376
1359 1377 // mcall is last "call" that can be a safepoint
1360 1378 // record it so we can see if a poll will directly follow it
1361 1379 // in which case we'll need a pad to make the PcDesc sites unique
1362 1380 // see 5010568. This can be slightly inaccurate but conservative
1363 1381 // in the case that return address is not actually at current_offset.
1364 1382 // This is a small price to pay.
1365 1383
1366 1384 if (is_mcall) {
1367 1385 last_call_offset = current_offset;
1368 1386 }
1369 1387
1370 1388 // See if this instruction has a delay slot
1371 1389 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1372 1390 assert(delay_slot != NULL, "expecting delay slot node");
1373 1391
1374 1392 // Back up 1 instruction
1375 1393 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1376 1394
1377 1395 // Save the offset for the listing
1378 1396 #ifndef PRODUCT
1379 1397 if( node_offsets && delay_slot->_idx < node_offset_limit )
1380 1398 node_offsets[delay_slot->_idx] = cb->insts_size();
1381 1399 #endif
1382 1400
1383 1401 // Support a SafePoint in the delay slot
1384 1402 if( delay_slot->is_MachSafePoint() ) {
1385 1403 MachNode *mach = delay_slot->as_Mach();
1386 1404 // !!!!! Stubs only need an oopmap right now, so bail out
1387 1405 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) {
1388 1406 // Write the oopmap directly to the code blob??!!
1389 1407 # ifdef ENABLE_ZAP_DEAD_LOCALS
1390 1408 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive");
1391 1409 # endif
1392 1410 delay_slot = NULL;
1393 1411 continue;
1394 1412 }
1395 1413
1396 1414 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1397 1415 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1398 1416 adjusted_offset);
1399 1417 // Generate an OopMap entry
1400 1418 Process_OopMap_Node(mach, adjusted_offset);
1401 1419 }
1402 1420
1403 1421 // Insert the delay slot instruction
1404 1422 delay_slot->emit(*cb, _regalloc);
1405 1423
1406 1424 // Don't reuse it
1407 1425 delay_slot = NULL;
1408 1426 }
1409 1427
1410 1428 } // End for all instructions in block
1411 1429
1412 1430 // If the next block is the top of a loop, pad this block out to align
1413 1431 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1414 1432 if( i<_cfg->_num_blocks-1 ) {
1415 1433 Block *nb = _cfg->_blocks[i+1];
1416 1434 uint padding = nb->alignment_padding(current_offset);
1417 1435 if( padding > 0 ) {
1418 1436 MachNode *nop = new (this) MachNopNode(padding / nop_size);
1419 1437 b->_nodes.insert( b->_nodes.size(), nop );
1420 1438 _cfg->_bbs.map( nop->_idx, b );
1421 1439 nop->emit(*cb, _regalloc);
1422 1440 current_offset = cb->insts_size();
1423 1441 }
1424 1442 }
1425 1443
1426 1444 } // End of for all blocks
1427 1445
1428 1446 non_safepoints.flush_at_end();
1429 1447
1430 1448 // Offset too large?
1431 1449 if (failing()) return;
1432 1450
1433 1451 // Define a pseudo-label at the end of the code
1434 1452 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] );
1435 1453
1436 1454 // Compute the size of the first block
1437 1455 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1438 1456
1439 1457 assert(cb->insts_size() < 500000, "method is unreasonably large");
1440 1458
1441 1459 // ------------------
1442 1460
1443 1461 #ifndef PRODUCT
1444 1462 // Information on the size of the method, without the extraneous code
1445 1463 Scheduling::increment_method_size(cb->insts_size());
1446 1464 #endif
1447 1465
1448 1466 // ------------------
1449 1467 // Fill in exception table entries.
1450 1468 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1451 1469
1452 1470 // Only java methods have exception handlers and deopt handlers
1453 1471 if (_method) {
1454 1472 // Emit the exception handler code.
1455 1473 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb));
1456 1474 // Emit the deopt handler code.
1457 1475 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb));
1458 1476
1459 1477 // Emit the MethodHandle deopt handler code (if required).
1460 1478 if (has_method_handle_invokes()) {
1461 1479 // We can use the same code as for the normal deopt handler, we
1462 1480 // just need a different entry point address.
1463 1481 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb));
1464 1482 }
1465 1483 }
1466 1484
1467 1485 // One last check for failed CodeBuffer::expand:
1468 1486 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1469 1487 turn_off_compiler(this);
1470 1488 return;
1471 1489 }
1472 1490
1473 1491 #ifndef PRODUCT
1474 1492 // Dump the assembly code, including basic-block numbers
1475 1493 if (print_assembly()) {
1476 1494 ttyLocker ttyl; // keep the following output all in one block
1477 1495 if (!VMThread::should_terminate()) { // test this under the tty lock
1478 1496 // This output goes directly to the tty, not the compiler log.
1479 1497 // To enable tools to match it up with the compilation activity,
1480 1498 // be sure to tag this tty output with the compile ID.
1481 1499 if (xtty != NULL) {
1482 1500 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1483 1501 is_osr_compilation() ? " compile_kind='osr'" :
1484 1502 "");
1485 1503 }
1486 1504 if (method() != NULL) {
1487 1505 method()->print_oop();
1488 1506 print_codes();
1489 1507 }
1490 1508 dump_asm(node_offsets, node_offset_limit);
1491 1509 if (xtty != NULL) {
1492 1510 xtty->tail("opto_assembly");
1493 1511 }
1494 1512 }
1495 1513 }
1496 1514 #endif
1497 1515
1498 1516 }
1499 1517
1500 1518 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1501 1519 _inc_table.set_size(cnt);
1502 1520
1503 1521 uint inct_cnt = 0;
1504 1522 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1505 1523 Block *b = _cfg->_blocks[i];
1506 1524 Node *n = NULL;
1507 1525 int j;
1508 1526
1509 1527 // Find the branch; ignore trailing NOPs.
1510 1528 for( j = b->_nodes.size()-1; j>=0; j-- ) {
1511 1529 n = b->_nodes[j];
1512 1530 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con )
1513 1531 break;
1514 1532 }
1515 1533
1516 1534 // If we didn't find anything, continue
1517 1535 if( j < 0 ) continue;
1518 1536
1519 1537 // Compute ExceptionHandlerTable subtable entry and add it
1520 1538 // (skip empty blocks)
1521 1539 if( n->is_Catch() ) {
1522 1540
1523 1541 // Get the offset of the return from the call
1524 1542 uint call_return = call_returns[b->_pre_order];
1525 1543 #ifdef ASSERT
1526 1544 assert( call_return > 0, "no call seen for this basic block" );
1527 1545 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ;
1528 1546 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" );
1529 1547 #endif
1530 1548 // last instruction is a CatchNode, find it's CatchProjNodes
1531 1549 int nof_succs = b->_num_succs;
1532 1550 // allocate space
1533 1551 GrowableArray<intptr_t> handler_bcis(nof_succs);
1534 1552 GrowableArray<intptr_t> handler_pcos(nof_succs);
1535 1553 // iterate through all successors
1536 1554 for (int j = 0; j < nof_succs; j++) {
1537 1555 Block* s = b->_succs[j];
1538 1556 bool found_p = false;
1539 1557 for( uint k = 1; k < s->num_preds(); k++ ) {
1540 1558 Node *pk = s->pred(k);
1541 1559 if( pk->is_CatchProj() && pk->in(0) == n ) {
1542 1560 const CatchProjNode* p = pk->as_CatchProj();
1543 1561 found_p = true;
1544 1562 // add the corresponding handler bci & pco information
1545 1563 if( p->_con != CatchProjNode::fall_through_index ) {
1546 1564 // p leads to an exception handler (and is not fall through)
1547 1565 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering");
1548 1566 // no duplicates, please
1549 1567 if( !handler_bcis.contains(p->handler_bci()) ) {
1550 1568 uint block_num = s->non_connector()->_pre_order;
1551 1569 handler_bcis.append(p->handler_bci());
1552 1570 handler_pcos.append(blk_labels[block_num].loc_pos());
1553 1571 }
1554 1572 }
1555 1573 }
1556 1574 }
1557 1575 assert(found_p, "no matching predecessor found");
1558 1576 // Note: Due to empty block removal, one block may have
1559 1577 // several CatchProj inputs, from the same Catch.
1560 1578 }
1561 1579
1562 1580 // Set the offset of the return from the call
1563 1581 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1564 1582 continue;
1565 1583 }
1566 1584
1567 1585 // Handle implicit null exception table updates
1568 1586 if( n->is_MachNullCheck() ) {
1569 1587 uint block_num = b->non_connector_successor(0)->_pre_order;
1570 1588 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() );
1571 1589 continue;
1572 1590 }
1573 1591 } // End of for all blocks fill in exception table entries
1574 1592 }
1575 1593
1576 1594 // Static Variables
1577 1595 #ifndef PRODUCT
1578 1596 uint Scheduling::_total_nop_size = 0;
1579 1597 uint Scheduling::_total_method_size = 0;
1580 1598 uint Scheduling::_total_branches = 0;
1581 1599 uint Scheduling::_total_unconditional_delays = 0;
1582 1600 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1583 1601 #endif
1584 1602
1585 1603 // Initializer for class Scheduling
1586 1604
1587 1605 Scheduling::Scheduling(Arena *arena, Compile &compile)
1588 1606 : _arena(arena),
1589 1607 _cfg(compile.cfg()),
1590 1608 _bbs(compile.cfg()->_bbs),
1591 1609 _regalloc(compile.regalloc()),
1592 1610 _reg_node(arena),
1593 1611 _bundle_instr_count(0),
1594 1612 _bundle_cycle_number(0),
1595 1613 _scheduled(arena),
1596 1614 _available(arena),
1597 1615 _next_node(NULL),
1598 1616 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]),
1599 1617 _pinch_free_list(arena)
1600 1618 #ifndef PRODUCT
1601 1619 , _branches(0)
1602 1620 , _unconditional_delays(0)
1603 1621 #endif
1604 1622 {
1605 1623 // Create a MachNopNode
1606 1624 _nop = new (&compile) MachNopNode();
1607 1625
1608 1626 // Now that the nops are in the array, save the count
1609 1627 // (but allow entries for the nops)
1610 1628 _node_bundling_limit = compile.unique();
1611 1629 uint node_max = _regalloc->node_regs_max_index();
1612 1630
1613 1631 compile.set_node_bundling_limit(_node_bundling_limit);
1614 1632
1615 1633 // This one is persistent within the Compile class
1616 1634 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1617 1635
1618 1636 // Allocate space for fixed-size arrays
1619 1637 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1620 1638 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1621 1639 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1622 1640
1623 1641 // Clear the arrays
1624 1642 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1625 1643 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1626 1644 memset(_uses, 0, node_max * sizeof(short));
1627 1645 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1628 1646
1629 1647 // Clear the bundling information
1630 1648 memcpy(_bundle_use_elements,
1631 1649 Pipeline_Use::elaborated_elements,
1632 1650 sizeof(Pipeline_Use::elaborated_elements));
1633 1651
1634 1652 // Get the last node
1635 1653 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1];
1636 1654
1637 1655 _next_node = bb->_nodes[bb->_nodes.size()-1];
1638 1656 }
1639 1657
1640 1658 #ifndef PRODUCT
1641 1659 // Scheduling destructor
1642 1660 Scheduling::~Scheduling() {
1643 1661 _total_branches += _branches;
1644 1662 _total_unconditional_delays += _unconditional_delays;
1645 1663 }
1646 1664 #endif
1647 1665
1648 1666 // Step ahead "i" cycles
1649 1667 void Scheduling::step(uint i) {
1650 1668
1651 1669 Bundle *bundle = node_bundling(_next_node);
1652 1670 bundle->set_starts_bundle();
1653 1671
1654 1672 // Update the bundle record, but leave the flags information alone
1655 1673 if (_bundle_instr_count > 0) {
1656 1674 bundle->set_instr_count(_bundle_instr_count);
1657 1675 bundle->set_resources_used(_bundle_use.resourcesUsed());
1658 1676 }
1659 1677
1660 1678 // Update the state information
1661 1679 _bundle_instr_count = 0;
1662 1680 _bundle_cycle_number += i;
1663 1681 _bundle_use.step(i);
1664 1682 }
1665 1683
1666 1684 void Scheduling::step_and_clear() {
1667 1685 Bundle *bundle = node_bundling(_next_node);
1668 1686 bundle->set_starts_bundle();
1669 1687
1670 1688 // Update the bundle record
1671 1689 if (_bundle_instr_count > 0) {
1672 1690 bundle->set_instr_count(_bundle_instr_count);
1673 1691 bundle->set_resources_used(_bundle_use.resourcesUsed());
1674 1692
1675 1693 _bundle_cycle_number += 1;
1676 1694 }
1677 1695
1678 1696 // Clear the bundling information
1679 1697 _bundle_instr_count = 0;
1680 1698 _bundle_use.reset();
1681 1699
1682 1700 memcpy(_bundle_use_elements,
1683 1701 Pipeline_Use::elaborated_elements,
1684 1702 sizeof(Pipeline_Use::elaborated_elements));
1685 1703 }
1686 1704
1687 1705 //------------------------------ScheduleAndBundle------------------------------
1688 1706 // Perform instruction scheduling and bundling over the sequence of
1689 1707 // instructions in backwards order.
1690 1708 void Compile::ScheduleAndBundle() {
1691 1709
1692 1710 // Don't optimize this if it isn't a method
1693 1711 if (!_method)
1694 1712 return;
↓ open down ↓ |
496 lines elided |
↑ open up ↑ |
1695 1713
1696 1714 // Don't optimize this if scheduling is disabled
1697 1715 if (!do_scheduling())
1698 1716 return;
1699 1717
1700 1718 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); )
1701 1719
1702 1720 // Create a data structure for all the scheduling information
1703 1721 Scheduling scheduling(Thread::current()->resource_area(), *this);
1704 1722
1723 + // Initialize the space for the BufferBlob used to find and verify
1724 + // instruction size in MachNode::emit_size()
1725 + init_scratch_buffer_blob(MAX_const_size);
1726 + if (failing()) return; // Out of memory
1727 +
1705 1728 // Walk backwards over each basic block, computing the needed alignment
1706 1729 // Walk over all the basic blocks
1707 1730 scheduling.DoScheduling();
1731 +
1732 + // Clear the BufferBlob used for scheduling.
1733 + clear_scratch_buffer_blob();
1708 1734 }
1709 1735
1710 1736 //------------------------------ComputeLocalLatenciesForward-------------------
1711 1737 // Compute the latency of all the instructions. This is fairly simple,
1712 1738 // because we already have a legal ordering. Walk over the instructions
1713 1739 // from first to last, and compute the latency of the instruction based
1714 1740 // on the latency of the preceding instruction(s).
1715 1741 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1716 1742 #ifndef PRODUCT
1717 1743 if (_cfg->C->trace_opto_output())
1718 1744 tty->print("# -> ComputeLocalLatenciesForward\n");
1719 1745 #endif
1720 1746
1721 1747 // Walk over all the schedulable instructions
1722 1748 for( uint j=_bb_start; j < _bb_end; j++ ) {
1723 1749
1724 1750 // This is a kludge, forcing all latency calculations to start at 1.
1725 1751 // Used to allow latency 0 to force an instruction to the beginning
1726 1752 // of the bb
1727 1753 uint latency = 1;
1728 1754 Node *use = bb->_nodes[j];
1729 1755 uint nlen = use->len();
1730 1756
1731 1757 // Walk over all the inputs
1732 1758 for ( uint k=0; k < nlen; k++ ) {
1733 1759 Node *def = use->in(k);
1734 1760 if (!def)
1735 1761 continue;
1736 1762
1737 1763 uint l = _node_latency[def->_idx] + use->latency(k);
1738 1764 if (latency < l)
1739 1765 latency = l;
1740 1766 }
1741 1767
1742 1768 _node_latency[use->_idx] = latency;
1743 1769
1744 1770 #ifndef PRODUCT
1745 1771 if (_cfg->C->trace_opto_output()) {
1746 1772 tty->print("# latency %4d: ", latency);
1747 1773 use->dump();
1748 1774 }
1749 1775 #endif
1750 1776 }
1751 1777
1752 1778 #ifndef PRODUCT
1753 1779 if (_cfg->C->trace_opto_output())
1754 1780 tty->print("# <- ComputeLocalLatenciesForward\n");
1755 1781 #endif
1756 1782
1757 1783 } // end ComputeLocalLatenciesForward
1758 1784
1759 1785 // See if this node fits into the present instruction bundle
1760 1786 bool Scheduling::NodeFitsInBundle(Node *n) {
1761 1787 uint n_idx = n->_idx;
1762 1788
1763 1789 // If this is the unconditional delay instruction, then it fits
1764 1790 if (n == _unconditional_delay_slot) {
1765 1791 #ifndef PRODUCT
1766 1792 if (_cfg->C->trace_opto_output())
1767 1793 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1768 1794 #endif
1769 1795 return (true);
1770 1796 }
1771 1797
1772 1798 // If the node cannot be scheduled this cycle, skip it
1773 1799 if (_current_latency[n_idx] > _bundle_cycle_number) {
1774 1800 #ifndef PRODUCT
1775 1801 if (_cfg->C->trace_opto_output())
1776 1802 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1777 1803 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1778 1804 #endif
1779 1805 return (false);
1780 1806 }
1781 1807
1782 1808 const Pipeline *node_pipeline = n->pipeline();
1783 1809
1784 1810 uint instruction_count = node_pipeline->instructionCount();
1785 1811 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1786 1812 instruction_count = 0;
1787 1813 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1788 1814 instruction_count++;
1789 1815
1790 1816 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1791 1817 #ifndef PRODUCT
1792 1818 if (_cfg->C->trace_opto_output())
1793 1819 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1794 1820 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1795 1821 #endif
1796 1822 return (false);
1797 1823 }
1798 1824
1799 1825 // Don't allow non-machine nodes to be handled this way
1800 1826 if (!n->is_Mach() && instruction_count == 0)
1801 1827 return (false);
1802 1828
1803 1829 // See if there is any overlap
1804 1830 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1805 1831
1806 1832 if (delay > 0) {
1807 1833 #ifndef PRODUCT
1808 1834 if (_cfg->C->trace_opto_output())
1809 1835 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1810 1836 #endif
1811 1837 return false;
1812 1838 }
1813 1839
1814 1840 #ifndef PRODUCT
1815 1841 if (_cfg->C->trace_opto_output())
1816 1842 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1817 1843 #endif
1818 1844
1819 1845 return true;
1820 1846 }
1821 1847
1822 1848 Node * Scheduling::ChooseNodeToBundle() {
1823 1849 uint siz = _available.size();
1824 1850
1825 1851 if (siz == 0) {
1826 1852
1827 1853 #ifndef PRODUCT
1828 1854 if (_cfg->C->trace_opto_output())
1829 1855 tty->print("# ChooseNodeToBundle: NULL\n");
1830 1856 #endif
1831 1857 return (NULL);
1832 1858 }
1833 1859
1834 1860 // Fast path, if only 1 instruction in the bundle
1835 1861 if (siz == 1) {
1836 1862 #ifndef PRODUCT
1837 1863 if (_cfg->C->trace_opto_output()) {
1838 1864 tty->print("# ChooseNodeToBundle (only 1): ");
1839 1865 _available[0]->dump();
1840 1866 }
1841 1867 #endif
1842 1868 return (_available[0]);
1843 1869 }
1844 1870
1845 1871 // Don't bother, if the bundle is already full
1846 1872 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1847 1873 for ( uint i = 0; i < siz; i++ ) {
1848 1874 Node *n = _available[i];
1849 1875
1850 1876 // Skip projections, we'll handle them another way
1851 1877 if (n->is_Proj())
1852 1878 continue;
1853 1879
1854 1880 // This presupposed that instructions are inserted into the
1855 1881 // available list in a legality order; i.e. instructions that
1856 1882 // must be inserted first are at the head of the list
1857 1883 if (NodeFitsInBundle(n)) {
1858 1884 #ifndef PRODUCT
1859 1885 if (_cfg->C->trace_opto_output()) {
1860 1886 tty->print("# ChooseNodeToBundle: ");
1861 1887 n->dump();
1862 1888 }
1863 1889 #endif
1864 1890 return (n);
1865 1891 }
1866 1892 }
1867 1893 }
1868 1894
1869 1895 // Nothing fits in this bundle, choose the highest priority
1870 1896 #ifndef PRODUCT
1871 1897 if (_cfg->C->trace_opto_output()) {
1872 1898 tty->print("# ChooseNodeToBundle: ");
1873 1899 _available[0]->dump();
1874 1900 }
1875 1901 #endif
1876 1902
1877 1903 return _available[0];
1878 1904 }
1879 1905
1880 1906 //------------------------------AddNodeToAvailableList-------------------------
1881 1907 void Scheduling::AddNodeToAvailableList(Node *n) {
1882 1908 assert( !n->is_Proj(), "projections never directly made available" );
1883 1909 #ifndef PRODUCT
1884 1910 if (_cfg->C->trace_opto_output()) {
1885 1911 tty->print("# AddNodeToAvailableList: ");
1886 1912 n->dump();
1887 1913 }
1888 1914 #endif
1889 1915
1890 1916 int latency = _current_latency[n->_idx];
1891 1917
1892 1918 // Insert in latency order (insertion sort)
1893 1919 uint i;
1894 1920 for ( i=0; i < _available.size(); i++ )
1895 1921 if (_current_latency[_available[i]->_idx] > latency)
1896 1922 break;
1897 1923
1898 1924 // Special Check for compares following branches
1899 1925 if( n->is_Mach() && _scheduled.size() > 0 ) {
1900 1926 int op = n->as_Mach()->ideal_Opcode();
1901 1927 Node *last = _scheduled[0];
1902 1928 if( last->is_MachIf() && last->in(1) == n &&
1903 1929 ( op == Op_CmpI ||
1904 1930 op == Op_CmpU ||
1905 1931 op == Op_CmpP ||
1906 1932 op == Op_CmpF ||
1907 1933 op == Op_CmpD ||
1908 1934 op == Op_CmpL ) ) {
1909 1935
1910 1936 // Recalculate position, moving to front of same latency
1911 1937 for ( i=0 ; i < _available.size(); i++ )
1912 1938 if (_current_latency[_available[i]->_idx] >= latency)
1913 1939 break;
1914 1940 }
1915 1941 }
1916 1942
1917 1943 // Insert the node in the available list
1918 1944 _available.insert(i, n);
1919 1945
1920 1946 #ifndef PRODUCT
1921 1947 if (_cfg->C->trace_opto_output())
1922 1948 dump_available();
1923 1949 #endif
1924 1950 }
1925 1951
1926 1952 //------------------------------DecrementUseCounts-----------------------------
1927 1953 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
1928 1954 for ( uint i=0; i < n->len(); i++ ) {
1929 1955 Node *def = n->in(i);
1930 1956 if (!def) continue;
1931 1957 if( def->is_Proj() ) // If this is a machine projection, then
1932 1958 def = def->in(0); // propagate usage thru to the base instruction
1933 1959
1934 1960 if( _bbs[def->_idx] != bb ) // Ignore if not block-local
1935 1961 continue;
1936 1962
1937 1963 // Compute the latency
1938 1964 uint l = _bundle_cycle_number + n->latency(i);
1939 1965 if (_current_latency[def->_idx] < l)
1940 1966 _current_latency[def->_idx] = l;
1941 1967
1942 1968 // If this does not have uses then schedule it
1943 1969 if ((--_uses[def->_idx]) == 0)
1944 1970 AddNodeToAvailableList(def);
1945 1971 }
1946 1972 }
1947 1973
1948 1974 //------------------------------AddNodeToBundle--------------------------------
1949 1975 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
1950 1976 #ifndef PRODUCT
1951 1977 if (_cfg->C->trace_opto_output()) {
1952 1978 tty->print("# AddNodeToBundle: ");
1953 1979 n->dump();
1954 1980 }
1955 1981 #endif
1956 1982
1957 1983 // Remove this from the available list
1958 1984 uint i;
1959 1985 for (i = 0; i < _available.size(); i++)
1960 1986 if (_available[i] == n)
1961 1987 break;
1962 1988 assert(i < _available.size(), "entry in _available list not found");
1963 1989 _available.remove(i);
1964 1990
1965 1991 // See if this fits in the current bundle
1966 1992 const Pipeline *node_pipeline = n->pipeline();
1967 1993 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
1968 1994
1969 1995 // Check for instructions to be placed in the delay slot. We
1970 1996 // do this before we actually schedule the current instruction,
1971 1997 // because the delay slot follows the current instruction.
1972 1998 if (Pipeline::_branch_has_delay_slot &&
1973 1999 node_pipeline->hasBranchDelay() &&
1974 2000 !_unconditional_delay_slot) {
1975 2001
1976 2002 uint siz = _available.size();
1977 2003
1978 2004 // Conditional branches can support an instruction that
1979 2005 // is unconditionally executed and not dependent by the
1980 2006 // branch, OR a conditionally executed instruction if
1981 2007 // the branch is taken. In practice, this means that
1982 2008 // the first instruction at the branch target is
1983 2009 // copied to the delay slot, and the branch goes to
1984 2010 // the instruction after that at the branch target
1985 2011 if ( n->is_Mach() && n->is_Branch() ) {
1986 2012
1987 2013 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
1988 2014 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
1989 2015
1990 2016 #ifndef PRODUCT
1991 2017 _branches++;
1992 2018 #endif
1993 2019
1994 2020 // At least 1 instruction is on the available list
1995 2021 // that is not dependent on the branch
1996 2022 for (uint i = 0; i < siz; i++) {
1997 2023 Node *d = _available[i];
1998 2024 const Pipeline *avail_pipeline = d->pipeline();
1999 2025
2000 2026 // Don't allow safepoints in the branch shadow, that will
2001 2027 // cause a number of difficulties
2002 2028 if ( avail_pipeline->instructionCount() == 1 &&
2003 2029 !avail_pipeline->hasMultipleBundles() &&
2004 2030 !avail_pipeline->hasBranchDelay() &&
2005 2031 Pipeline::instr_has_unit_size() &&
2006 2032 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2007 2033 NodeFitsInBundle(d) &&
2008 2034 !node_bundling(d)->used_in_delay()) {
2009 2035
2010 2036 if (d->is_Mach() && !d->is_MachSafePoint()) {
2011 2037 // A node that fits in the delay slot was found, so we need to
2012 2038 // set the appropriate bits in the bundle pipeline information so
2013 2039 // that it correctly indicates resource usage. Later, when we
2014 2040 // attempt to add this instruction to the bundle, we will skip
2015 2041 // setting the resource usage.
2016 2042 _unconditional_delay_slot = d;
2017 2043 node_bundling(n)->set_use_unconditional_delay();
2018 2044 node_bundling(d)->set_used_in_unconditional_delay();
2019 2045 _bundle_use.add_usage(avail_pipeline->resourceUse());
2020 2046 _current_latency[d->_idx] = _bundle_cycle_number;
2021 2047 _next_node = d;
2022 2048 ++_bundle_instr_count;
2023 2049 #ifndef PRODUCT
2024 2050 _unconditional_delays++;
2025 2051 #endif
2026 2052 break;
2027 2053 }
2028 2054 }
2029 2055 }
2030 2056 }
2031 2057
2032 2058 // No delay slot, add a nop to the usage
2033 2059 if (!_unconditional_delay_slot) {
2034 2060 // See if adding an instruction in the delay slot will overflow
2035 2061 // the bundle.
2036 2062 if (!NodeFitsInBundle(_nop)) {
2037 2063 #ifndef PRODUCT
2038 2064 if (_cfg->C->trace_opto_output())
2039 2065 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2040 2066 #endif
2041 2067 step(1);
2042 2068 }
2043 2069
2044 2070 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2045 2071 _next_node = _nop;
2046 2072 ++_bundle_instr_count;
2047 2073 }
2048 2074
2049 2075 // See if the instruction in the delay slot requires a
2050 2076 // step of the bundles
2051 2077 if (!NodeFitsInBundle(n)) {
2052 2078 #ifndef PRODUCT
2053 2079 if (_cfg->C->trace_opto_output())
2054 2080 tty->print("# *** STEP(branch won't fit) ***\n");
2055 2081 #endif
2056 2082 // Update the state information
2057 2083 _bundle_instr_count = 0;
2058 2084 _bundle_cycle_number += 1;
2059 2085 _bundle_use.step(1);
2060 2086 }
2061 2087 }
2062 2088
2063 2089 // Get the number of instructions
2064 2090 uint instruction_count = node_pipeline->instructionCount();
2065 2091 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2066 2092 instruction_count = 0;
2067 2093
2068 2094 // Compute the latency information
2069 2095 uint delay = 0;
2070 2096
2071 2097 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2072 2098 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2073 2099 if (relative_latency < 0)
2074 2100 relative_latency = 0;
2075 2101
2076 2102 delay = _bundle_use.full_latency(relative_latency, node_usage);
2077 2103
2078 2104 // Does not fit in this bundle, start a new one
2079 2105 if (delay > 0) {
2080 2106 step(delay);
2081 2107
2082 2108 #ifndef PRODUCT
2083 2109 if (_cfg->C->trace_opto_output())
2084 2110 tty->print("# *** STEP(%d) ***\n", delay);
2085 2111 #endif
2086 2112 }
2087 2113 }
2088 2114
2089 2115 // If this was placed in the delay slot, ignore it
2090 2116 if (n != _unconditional_delay_slot) {
2091 2117
2092 2118 if (delay == 0) {
2093 2119 if (node_pipeline->hasMultipleBundles()) {
2094 2120 #ifndef PRODUCT
2095 2121 if (_cfg->C->trace_opto_output())
2096 2122 tty->print("# *** STEP(multiple instructions) ***\n");
2097 2123 #endif
2098 2124 step(1);
2099 2125 }
2100 2126
2101 2127 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2102 2128 #ifndef PRODUCT
2103 2129 if (_cfg->C->trace_opto_output())
2104 2130 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2105 2131 instruction_count + _bundle_instr_count,
2106 2132 Pipeline::_max_instrs_per_cycle);
2107 2133 #endif
2108 2134 step(1);
2109 2135 }
2110 2136 }
2111 2137
2112 2138 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2113 2139 _bundle_instr_count++;
2114 2140
2115 2141 // Set the node's latency
2116 2142 _current_latency[n->_idx] = _bundle_cycle_number;
2117 2143
2118 2144 // Now merge the functional unit information
2119 2145 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2120 2146 _bundle_use.add_usage(node_usage);
2121 2147
2122 2148 // Increment the number of instructions in this bundle
2123 2149 _bundle_instr_count += instruction_count;
2124 2150
2125 2151 // Remember this node for later
2126 2152 if (n->is_Mach())
2127 2153 _next_node = n;
2128 2154 }
2129 2155
2130 2156 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2131 2157 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2132 2158 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2133 2159 // into the block. All other scheduled nodes get put in the schedule here.
2134 2160 int op = n->Opcode();
2135 2161 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2136 2162 (op != Op_Node && // Not an unused antidepedence node and
2137 2163 // not an unallocated boxlock
2138 2164 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2139 2165
2140 2166 // Push any trailing projections
2141 2167 if( bb->_nodes[bb->_nodes.size()-1] != n ) {
2142 2168 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2143 2169 Node *foi = n->fast_out(i);
2144 2170 if( foi->is_Proj() )
2145 2171 _scheduled.push(foi);
2146 2172 }
2147 2173 }
2148 2174
2149 2175 // Put the instruction in the schedule list
2150 2176 _scheduled.push(n);
2151 2177 }
2152 2178
2153 2179 #ifndef PRODUCT
2154 2180 if (_cfg->C->trace_opto_output())
2155 2181 dump_available();
2156 2182 #endif
2157 2183
2158 2184 // Walk all the definitions, decrementing use counts, and
2159 2185 // if a definition has a 0 use count, place it in the available list.
2160 2186 DecrementUseCounts(n,bb);
2161 2187 }
2162 2188
2163 2189 //------------------------------ComputeUseCount--------------------------------
2164 2190 // This method sets the use count within a basic block. We will ignore all
2165 2191 // uses outside the current basic block. As we are doing a backwards walk,
2166 2192 // any node we reach that has a use count of 0 may be scheduled. This also
2167 2193 // avoids the problem of cyclic references from phi nodes, as long as phi
2168 2194 // nodes are at the front of the basic block. This method also initializes
2169 2195 // the available list to the set of instructions that have no uses within this
2170 2196 // basic block.
2171 2197 void Scheduling::ComputeUseCount(const Block *bb) {
2172 2198 #ifndef PRODUCT
2173 2199 if (_cfg->C->trace_opto_output())
2174 2200 tty->print("# -> ComputeUseCount\n");
2175 2201 #endif
2176 2202
2177 2203 // Clear the list of available and scheduled instructions, just in case
2178 2204 _available.clear();
2179 2205 _scheduled.clear();
2180 2206
2181 2207 // No delay slot specified
2182 2208 _unconditional_delay_slot = NULL;
2183 2209
2184 2210 #ifdef ASSERT
2185 2211 for( uint i=0; i < bb->_nodes.size(); i++ )
2186 2212 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" );
2187 2213 #endif
2188 2214
2189 2215 // Force the _uses count to never go to zero for unscheduable pieces
2190 2216 // of the block
2191 2217 for( uint k = 0; k < _bb_start; k++ )
2192 2218 _uses[bb->_nodes[k]->_idx] = 1;
2193 2219 for( uint l = _bb_end; l < bb->_nodes.size(); l++ )
2194 2220 _uses[bb->_nodes[l]->_idx] = 1;
2195 2221
2196 2222 // Iterate backwards over the instructions in the block. Don't count the
2197 2223 // branch projections at end or the block header instructions.
2198 2224 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2199 2225 Node *n = bb->_nodes[j];
2200 2226 if( n->is_Proj() ) continue; // Projections handled another way
2201 2227
2202 2228 // Account for all uses
2203 2229 for ( uint k = 0; k < n->len(); k++ ) {
2204 2230 Node *inp = n->in(k);
2205 2231 if (!inp) continue;
2206 2232 assert(inp != n, "no cycles allowed" );
2207 2233 if( _bbs[inp->_idx] == bb ) { // Block-local use?
2208 2234 if( inp->is_Proj() ) // Skip through Proj's
2209 2235 inp = inp->in(0);
2210 2236 ++_uses[inp->_idx]; // Count 1 block-local use
2211 2237 }
2212 2238 }
2213 2239
2214 2240 // If this instruction has a 0 use count, then it is available
2215 2241 if (!_uses[n->_idx]) {
2216 2242 _current_latency[n->_idx] = _bundle_cycle_number;
2217 2243 AddNodeToAvailableList(n);
2218 2244 }
2219 2245
2220 2246 #ifndef PRODUCT
2221 2247 if (_cfg->C->trace_opto_output()) {
2222 2248 tty->print("# uses: %3d: ", _uses[n->_idx]);
2223 2249 n->dump();
2224 2250 }
2225 2251 #endif
2226 2252 }
2227 2253
2228 2254 #ifndef PRODUCT
2229 2255 if (_cfg->C->trace_opto_output())
2230 2256 tty->print("# <- ComputeUseCount\n");
2231 2257 #endif
2232 2258 }
2233 2259
2234 2260 // This routine performs scheduling on each basic block in reverse order,
2235 2261 // using instruction latencies and taking into account function unit
2236 2262 // availability.
2237 2263 void Scheduling::DoScheduling() {
2238 2264 #ifndef PRODUCT
2239 2265 if (_cfg->C->trace_opto_output())
2240 2266 tty->print("# -> DoScheduling\n");
2241 2267 #endif
2242 2268
2243 2269 Block *succ_bb = NULL;
2244 2270 Block *bb;
2245 2271
2246 2272 // Walk over all the basic blocks in reverse order
2247 2273 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) {
2248 2274 bb = _cfg->_blocks[i];
2249 2275
2250 2276 #ifndef PRODUCT
2251 2277 if (_cfg->C->trace_opto_output()) {
2252 2278 tty->print("# Schedule BB#%03d (initial)\n", i);
2253 2279 for (uint j = 0; j < bb->_nodes.size(); j++)
2254 2280 bb->_nodes[j]->dump();
2255 2281 }
2256 2282 #endif
2257 2283
2258 2284 // On the head node, skip processing
2259 2285 if( bb == _cfg->_broot )
2260 2286 continue;
2261 2287
2262 2288 // Skip empty, connector blocks
2263 2289 if (bb->is_connector())
2264 2290 continue;
2265 2291
2266 2292 // If the following block is not the sole successor of
2267 2293 // this one, then reset the pipeline information
2268 2294 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2269 2295 #ifndef PRODUCT
2270 2296 if (_cfg->C->trace_opto_output()) {
2271 2297 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2272 2298 _next_node->_idx, _bundle_instr_count);
2273 2299 }
2274 2300 #endif
2275 2301 step_and_clear();
2276 2302 }
2277 2303
2278 2304 // Leave untouched the starting instruction, any Phis, a CreateEx node
2279 2305 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction.
2280 2306 _bb_end = bb->_nodes.size()-1;
2281 2307 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2282 2308 Node *n = bb->_nodes[_bb_start];
2283 2309 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2284 2310 // Also, MachIdealNodes do not get scheduled
2285 2311 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2286 2312 MachNode *mach = n->as_Mach();
2287 2313 int iop = mach->ideal_Opcode();
2288 2314 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2289 2315 if( iop == Op_Con ) continue; // Do not schedule Top
2290 2316 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2291 2317 mach->pipeline() == MachNode::pipeline_class() &&
2292 2318 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc
2293 2319 continue;
2294 2320 break; // Funny loop structure to be sure...
2295 2321 }
2296 2322 // Compute last "interesting" instruction in block - last instruction we
2297 2323 // might schedule. _bb_end points just after last schedulable inst. We
2298 2324 // normally schedule conditional branches (despite them being forced last
2299 2325 // in the block), because they have delay slots we can fill. Calls all
2300 2326 // have their delay slots filled in the template expansions, so we don't
2301 2327 // bother scheduling them.
2302 2328 Node *last = bb->_nodes[_bb_end];
2303 2329 if( last->is_Catch() ||
2304 2330 // Exclude unreachable path case when Halt node is in a separate block.
2305 2331 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2306 2332 // There must be a prior call. Skip it.
2307 2333 while( !bb->_nodes[--_bb_end]->is_Call() ) {
2308 2334 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" );
2309 2335 }
2310 2336 } else if( last->is_MachNullCheck() ) {
2311 2337 // Backup so the last null-checked memory instruction is
2312 2338 // outside the schedulable range. Skip over the nullcheck,
2313 2339 // projection, and the memory nodes.
2314 2340 Node *mem = last->in(1);
2315 2341 do {
2316 2342 _bb_end--;
2317 2343 } while (mem != bb->_nodes[_bb_end]);
2318 2344 } else {
2319 2345 // Set _bb_end to point after last schedulable inst.
2320 2346 _bb_end++;
2321 2347 }
2322 2348
2323 2349 assert( _bb_start <= _bb_end, "inverted block ends" );
2324 2350
2325 2351 // Compute the register antidependencies for the basic block
2326 2352 ComputeRegisterAntidependencies(bb);
2327 2353 if (_cfg->C->failing()) return; // too many D-U pinch points
2328 2354
2329 2355 // Compute intra-bb latencies for the nodes
2330 2356 ComputeLocalLatenciesForward(bb);
2331 2357
2332 2358 // Compute the usage within the block, and set the list of all nodes
2333 2359 // in the block that have no uses within the block.
2334 2360 ComputeUseCount(bb);
2335 2361
2336 2362 // Schedule the remaining instructions in the block
2337 2363 while ( _available.size() > 0 ) {
2338 2364 Node *n = ChooseNodeToBundle();
2339 2365 AddNodeToBundle(n,bb);
2340 2366 }
2341 2367
2342 2368 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2343 2369 #ifdef ASSERT
2344 2370 for( uint l = _bb_start; l < _bb_end; l++ ) {
2345 2371 Node *n = bb->_nodes[l];
2346 2372 uint m;
2347 2373 for( m = 0; m < _bb_end-_bb_start; m++ )
2348 2374 if( _scheduled[m] == n )
2349 2375 break;
2350 2376 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2351 2377 }
2352 2378 #endif
2353 2379
2354 2380 // Now copy the instructions (in reverse order) back to the block
2355 2381 for ( uint k = _bb_start; k < _bb_end; k++ )
2356 2382 bb->_nodes.map(k, _scheduled[_bb_end-k-1]);
2357 2383
2358 2384 #ifndef PRODUCT
2359 2385 if (_cfg->C->trace_opto_output()) {
2360 2386 tty->print("# Schedule BB#%03d (final)\n", i);
2361 2387 uint current = 0;
2362 2388 for (uint j = 0; j < bb->_nodes.size(); j++) {
2363 2389 Node *n = bb->_nodes[j];
2364 2390 if( valid_bundle_info(n) ) {
2365 2391 Bundle *bundle = node_bundling(n);
2366 2392 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2367 2393 tty->print("*** Bundle: ");
2368 2394 bundle->dump();
2369 2395 }
2370 2396 n->dump();
2371 2397 }
2372 2398 }
2373 2399 }
2374 2400 #endif
2375 2401 #ifdef ASSERT
2376 2402 verify_good_schedule(bb,"after block local scheduling");
2377 2403 #endif
2378 2404 }
2379 2405
2380 2406 #ifndef PRODUCT
2381 2407 if (_cfg->C->trace_opto_output())
2382 2408 tty->print("# <- DoScheduling\n");
2383 2409 #endif
2384 2410
2385 2411 // Record final node-bundling array location
2386 2412 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2387 2413
2388 2414 } // end DoScheduling
2389 2415
2390 2416 //------------------------------verify_good_schedule---------------------------
2391 2417 // Verify that no live-range used in the block is killed in the block by a
2392 2418 // wrong DEF. This doesn't verify live-ranges that span blocks.
2393 2419
2394 2420 // Check for edge existence. Used to avoid adding redundant precedence edges.
2395 2421 static bool edge_from_to( Node *from, Node *to ) {
2396 2422 for( uint i=0; i<from->len(); i++ )
2397 2423 if( from->in(i) == to )
2398 2424 return true;
2399 2425 return false;
2400 2426 }
2401 2427
2402 2428 #ifdef ASSERT
2403 2429 //------------------------------verify_do_def----------------------------------
2404 2430 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2405 2431 // Check for bad kills
2406 2432 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2407 2433 Node *prior_use = _reg_node[def];
2408 2434 if( prior_use && !edge_from_to(prior_use,n) ) {
2409 2435 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2410 2436 n->dump();
2411 2437 tty->print_cr("...");
2412 2438 prior_use->dump();
2413 2439 assert(edge_from_to(prior_use,n),msg);
2414 2440 }
2415 2441 _reg_node.map(def,NULL); // Kill live USEs
2416 2442 }
2417 2443 }
2418 2444
2419 2445 //------------------------------verify_good_schedule---------------------------
2420 2446 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2421 2447
2422 2448 // Zap to something reasonable for the verify code
2423 2449 _reg_node.clear();
2424 2450
2425 2451 // Walk over the block backwards. Check to make sure each DEF doesn't
2426 2452 // kill a live value (other than the one it's supposed to). Add each
2427 2453 // USE to the live set.
2428 2454 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) {
2429 2455 Node *n = b->_nodes[i];
2430 2456 int n_op = n->Opcode();
2431 2457 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2432 2458 // Fat-proj kills a slew of registers
2433 2459 RegMask rm = n->out_RegMask();// Make local copy
2434 2460 while( rm.is_NotEmpty() ) {
2435 2461 OptoReg::Name kill = rm.find_first_elem();
2436 2462 rm.Remove(kill);
2437 2463 verify_do_def( n, kill, msg );
2438 2464 }
2439 2465 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2440 2466 // Get DEF'd registers the normal way
2441 2467 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2442 2468 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2443 2469 }
2444 2470
2445 2471 // Now make all USEs live
2446 2472 for( uint i=1; i<n->req(); i++ ) {
2447 2473 Node *def = n->in(i);
2448 2474 assert(def != 0, "input edge required");
2449 2475 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2450 2476 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2451 2477 if( OptoReg::is_valid(reg_lo) ) {
2452 2478 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg);
2453 2479 _reg_node.map(reg_lo,n);
2454 2480 }
2455 2481 if( OptoReg::is_valid(reg_hi) ) {
2456 2482 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg);
2457 2483 _reg_node.map(reg_hi,n);
2458 2484 }
2459 2485 }
2460 2486
2461 2487 }
2462 2488
2463 2489 // Zap to something reasonable for the Antidependence code
2464 2490 _reg_node.clear();
2465 2491 }
2466 2492 #endif
2467 2493
2468 2494 // Conditionally add precedence edges. Avoid putting edges on Projs.
2469 2495 static void add_prec_edge_from_to( Node *from, Node *to ) {
2470 2496 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2471 2497 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2472 2498 from = from->in(0);
2473 2499 }
2474 2500 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2475 2501 !edge_from_to( from, to ) ) // Avoid duplicate edge
2476 2502 from->add_prec(to);
2477 2503 }
2478 2504
2479 2505 //------------------------------anti_do_def------------------------------------
2480 2506 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2481 2507 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2482 2508 return;
2483 2509
2484 2510 Node *pinch = _reg_node[def_reg]; // Get pinch point
2485 2511 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet?
2486 2512 is_def ) { // Check for a true def (not a kill)
2487 2513 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2488 2514 return;
2489 2515 }
2490 2516
2491 2517 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2492 2518 debug_only( def = (Node*)0xdeadbeef; )
2493 2519
2494 2520 // After some number of kills there _may_ be a later def
2495 2521 Node *later_def = NULL;
2496 2522
2497 2523 // Finding a kill requires a real pinch-point.
2498 2524 // Check for not already having a pinch-point.
2499 2525 // Pinch points are Op_Node's.
2500 2526 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2501 2527 later_def = pinch; // Must be def/kill as optimistic pinch-point
2502 2528 if ( _pinch_free_list.size() > 0) {
2503 2529 pinch = _pinch_free_list.pop();
2504 2530 } else {
2505 2531 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be
2506 2532 }
2507 2533 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2508 2534 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2509 2535 return;
2510 2536 }
2511 2537 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init)
2512 2538 _reg_node.map(def_reg,pinch); // Record pinch-point
2513 2539 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2514 2540 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2515 2541 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2516 2542 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2517 2543 later_def = NULL; // and no later def
2518 2544 }
2519 2545 pinch->set_req(0,later_def); // Hook later def so we can find it
2520 2546 } else { // Else have valid pinch point
2521 2547 if( pinch->in(0) ) // If there is a later-def
2522 2548 later_def = pinch->in(0); // Get it
2523 2549 }
2524 2550
2525 2551 // Add output-dependence edge from later def to kill
2526 2552 if( later_def ) // If there is some original def
2527 2553 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2528 2554
2529 2555 // See if current kill is also a use, and so is forced to be the pinch-point.
2530 2556 if( pinch->Opcode() == Op_Node ) {
2531 2557 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2532 2558 for( uint i=1; i<uses->req(); i++ ) {
2533 2559 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2534 2560 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2535 2561 // Yes, found a use/kill pinch-point
2536 2562 pinch->set_req(0,NULL); //
2537 2563 pinch->replace_by(kill); // Move anti-dep edges up
2538 2564 pinch = kill;
2539 2565 _reg_node.map(def_reg,pinch);
2540 2566 return;
2541 2567 }
2542 2568 }
2543 2569 }
2544 2570
2545 2571 // Add edge from kill to pinch-point
2546 2572 add_prec_edge_from_to(kill,pinch);
2547 2573 }
2548 2574
2549 2575 //------------------------------anti_do_use------------------------------------
2550 2576 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2551 2577 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2552 2578 return;
2553 2579 Node *pinch = _reg_node[use_reg]; // Get pinch point
2554 2580 // Check for no later def_reg/kill in block
2555 2581 if( pinch && _bbs[pinch->_idx] == b &&
2556 2582 // Use has to be block-local as well
2557 2583 _bbs[use->_idx] == b ) {
2558 2584 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2559 2585 pinch->req() == 1 ) { // pinch not yet in block?
2560 2586 pinch->del_req(0); // yank pointer to later-def, also set flag
2561 2587 // Insert the pinch-point in the block just after the last use
2562 2588 b->_nodes.insert(b->find_node(use)+1,pinch);
2563 2589 _bb_end++; // Increase size scheduled region in block
2564 2590 }
2565 2591
2566 2592 add_prec_edge_from_to(pinch,use);
2567 2593 }
2568 2594 }
2569 2595
2570 2596 //------------------------------ComputeRegisterAntidependences-----------------
2571 2597 // We insert antidependences between the reads and following write of
2572 2598 // allocated registers to prevent illegal code motion. Hopefully, the
2573 2599 // number of added references should be fairly small, especially as we
2574 2600 // are only adding references within the current basic block.
2575 2601 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2576 2602
2577 2603 #ifdef ASSERT
2578 2604 verify_good_schedule(b,"before block local scheduling");
2579 2605 #endif
2580 2606
2581 2607 // A valid schedule, for each register independently, is an endless cycle
2582 2608 // of: a def, then some uses (connected to the def by true dependencies),
2583 2609 // then some kills (defs with no uses), finally the cycle repeats with a new
2584 2610 // def. The uses are allowed to float relative to each other, as are the
2585 2611 // kills. No use is allowed to slide past a kill (or def). This requires
2586 2612 // antidependencies between all uses of a single def and all kills that
2587 2613 // follow, up to the next def. More edges are redundant, because later defs
2588 2614 // & kills are already serialized with true or antidependencies. To keep
2589 2615 // the edge count down, we add a 'pinch point' node if there's more than
2590 2616 // one use or more than one kill/def.
2591 2617
2592 2618 // We add dependencies in one bottom-up pass.
2593 2619
2594 2620 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2595 2621
2596 2622 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2597 2623 // register. If not, we record the DEF/KILL in _reg_node, the
2598 2624 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2599 2625 // "pinch point", a new Node that's in the graph but not in the block.
2600 2626 // We put edges from the prior and current DEF/KILLs to the pinch point.
2601 2627 // We put the pinch point in _reg_node. If there's already a pinch point
2602 2628 // we merely add an edge from the current DEF/KILL to the pinch point.
2603 2629
2604 2630 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2605 2631 // put an edge from the pinch point to the USE.
2606 2632
2607 2633 // To be expedient, the _reg_node array is pre-allocated for the whole
2608 2634 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2609 2635 // or a valid def/kill/pinch-point, or a leftover node from some prior
2610 2636 // block. Leftover node from some prior block is treated like a NULL (no
2611 2637 // prior def, so no anti-dependence needed). Valid def is distinguished by
2612 2638 // it being in the current block.
2613 2639 bool fat_proj_seen = false;
2614 2640 uint last_safept = _bb_end-1;
2615 2641 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL;
2616 2642 Node* last_safept_node = end_node;
2617 2643 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2618 2644 Node *n = b->_nodes[i];
2619 2645 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2620 2646 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2621 2647 // Fat-proj kills a slew of registers
2622 2648 // This can add edges to 'n' and obscure whether or not it was a def,
2623 2649 // hence the is_def flag.
2624 2650 fat_proj_seen = true;
2625 2651 RegMask rm = n->out_RegMask();// Make local copy
2626 2652 while( rm.is_NotEmpty() ) {
2627 2653 OptoReg::Name kill = rm.find_first_elem();
2628 2654 rm.Remove(kill);
2629 2655 anti_do_def( b, n, kill, is_def );
2630 2656 }
2631 2657 } else {
2632 2658 // Get DEF'd registers the normal way
2633 2659 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2634 2660 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2635 2661 }
2636 2662
2637 2663 // Check each register used by this instruction for a following DEF/KILL
2638 2664 // that must occur afterward and requires an anti-dependence edge.
2639 2665 for( uint j=0; j<n->req(); j++ ) {
2640 2666 Node *def = n->in(j);
2641 2667 if( def ) {
2642 2668 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" );
2643 2669 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2644 2670 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2645 2671 }
2646 2672 }
2647 2673 // Do not allow defs of new derived values to float above GC
2648 2674 // points unless the base is definitely available at the GC point.
2649 2675
2650 2676 Node *m = b->_nodes[i];
2651 2677
2652 2678 // Add precedence edge from following safepoint to use of derived pointer
2653 2679 if( last_safept_node != end_node &&
2654 2680 m != last_safept_node) {
2655 2681 for (uint k = 1; k < m->req(); k++) {
2656 2682 const Type *t = m->in(k)->bottom_type();
2657 2683 if( t->isa_oop_ptr() &&
2658 2684 t->is_ptr()->offset() != 0 ) {
2659 2685 last_safept_node->add_prec( m );
2660 2686 break;
2661 2687 }
2662 2688 }
2663 2689 }
2664 2690
2665 2691 if( n->jvms() ) { // Precedence edge from derived to safept
2666 2692 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2667 2693 if( b->_nodes[last_safept] != last_safept_node ) {
2668 2694 last_safept = b->find_node(last_safept_node);
2669 2695 }
2670 2696 for( uint j=last_safept; j > i; j-- ) {
2671 2697 Node *mach = b->_nodes[j];
2672 2698 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2673 2699 mach->add_prec( n );
2674 2700 }
2675 2701 last_safept = i;
2676 2702 last_safept_node = m;
2677 2703 }
2678 2704 }
2679 2705
2680 2706 if (fat_proj_seen) {
2681 2707 // Garbage collect pinch nodes that were not consumed.
2682 2708 // They are usually created by a fat kill MachProj for a call.
2683 2709 garbage_collect_pinch_nodes();
2684 2710 }
2685 2711 }
2686 2712
2687 2713 //------------------------------garbage_collect_pinch_nodes-------------------------------
2688 2714
2689 2715 // Garbage collect pinch nodes for reuse by other blocks.
2690 2716 //
2691 2717 // The block scheduler's insertion of anti-dependence
2692 2718 // edges creates many pinch nodes when the block contains
2693 2719 // 2 or more Calls. A pinch node is used to prevent a
2694 2720 // combinatorial explosion of edges. If a set of kills for a
2695 2721 // register is anti-dependent on a set of uses (or defs), rather
2696 2722 // than adding an edge in the graph between each pair of kill
2697 2723 // and use (or def), a pinch is inserted between them:
2698 2724 //
2699 2725 // use1 use2 use3
2700 2726 // \ | /
2701 2727 // \ | /
2702 2728 // pinch
2703 2729 // / | \
2704 2730 // / | \
2705 2731 // kill1 kill2 kill3
2706 2732 //
2707 2733 // One pinch node is created per register killed when
2708 2734 // the second call is encountered during a backwards pass
2709 2735 // over the block. Most of these pinch nodes are never
2710 2736 // wired into the graph because the register is never
2711 2737 // used or def'ed in the block.
2712 2738 //
2713 2739 void Scheduling::garbage_collect_pinch_nodes() {
2714 2740 #ifndef PRODUCT
2715 2741 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2716 2742 #endif
2717 2743 int trace_cnt = 0;
2718 2744 for (uint k = 0; k < _reg_node.Size(); k++) {
2719 2745 Node* pinch = _reg_node[k];
2720 2746 if (pinch != NULL && pinch->Opcode() == Op_Node &&
2721 2747 // no predecence input edges
2722 2748 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2723 2749 cleanup_pinch(pinch);
2724 2750 _pinch_free_list.push(pinch);
2725 2751 _reg_node.map(k, NULL);
2726 2752 #ifndef PRODUCT
2727 2753 if (_cfg->C->trace_opto_output()) {
2728 2754 trace_cnt++;
2729 2755 if (trace_cnt > 40) {
2730 2756 tty->print("\n");
2731 2757 trace_cnt = 0;
2732 2758 }
2733 2759 tty->print(" %d", pinch->_idx);
2734 2760 }
2735 2761 #endif
2736 2762 }
2737 2763 }
2738 2764 #ifndef PRODUCT
2739 2765 if (_cfg->C->trace_opto_output()) tty->print("\n");
2740 2766 #endif
2741 2767 }
2742 2768
2743 2769 // Clean up a pinch node for reuse.
2744 2770 void Scheduling::cleanup_pinch( Node *pinch ) {
2745 2771 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2746 2772
2747 2773 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2748 2774 Node* use = pinch->last_out(i);
2749 2775 uint uses_found = 0;
2750 2776 for (uint j = use->req(); j < use->len(); j++) {
2751 2777 if (use->in(j) == pinch) {
2752 2778 use->rm_prec(j);
2753 2779 uses_found++;
2754 2780 }
2755 2781 }
2756 2782 assert(uses_found > 0, "must be a precedence edge");
2757 2783 i -= uses_found; // we deleted 1 or more copies of this edge
2758 2784 }
2759 2785 // May have a later_def entry
2760 2786 pinch->set_req(0, NULL);
2761 2787 }
2762 2788
2763 2789 //------------------------------print_statistics-------------------------------
2764 2790 #ifndef PRODUCT
2765 2791
2766 2792 void Scheduling::dump_available() const {
2767 2793 tty->print("#Availist ");
2768 2794 for (uint i = 0; i < _available.size(); i++)
2769 2795 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2770 2796 tty->cr();
2771 2797 }
2772 2798
2773 2799 // Print Scheduling Statistics
2774 2800 void Scheduling::print_statistics() {
2775 2801 // Print the size added by nops for bundling
2776 2802 tty->print("Nops added %d bytes to total of %d bytes",
2777 2803 _total_nop_size, _total_method_size);
2778 2804 if (_total_method_size > 0)
2779 2805 tty->print(", for %.2f%%",
2780 2806 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2781 2807 tty->print("\n");
2782 2808
2783 2809 // Print the number of branch shadows filled
2784 2810 if (Pipeline::_branch_has_delay_slot) {
2785 2811 tty->print("Of %d branches, %d had unconditional delay slots filled",
2786 2812 _total_branches, _total_unconditional_delays);
2787 2813 if (_total_branches > 0)
2788 2814 tty->print(", for %.2f%%",
2789 2815 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2790 2816 tty->print("\n");
2791 2817 }
2792 2818
2793 2819 uint total_instructions = 0, total_bundles = 0;
2794 2820
2795 2821 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2796 2822 uint bundle_count = _total_instructions_per_bundle[i];
2797 2823 total_instructions += bundle_count * i;
2798 2824 total_bundles += bundle_count;
2799 2825 }
2800 2826
2801 2827 if (total_bundles > 0)
2802 2828 tty->print("Average ILP (excluding nops) is %.2f\n",
2803 2829 ((double)total_instructions) / ((double)total_bundles));
2804 2830 }
2805 2831 #endif
↓ open down ↓ |
1088 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX