1 /*
2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.inline.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/debugInfo.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "compiler/compileBroker.hpp"
32 #include "compiler/compilerDirectives.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "opto/ad.hpp"
36 #include "opto/callnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/machnode.hpp"
40 #include "opto/optoreg.hpp"
41 #include "opto/output.hpp"
42 #include "opto/regalloc.hpp"
43 #include "opto/runtime.hpp"
44 #include "opto/subnode.hpp"
45 #include "opto/type.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "utilities/xmlstream.hpp"
48
49 #ifndef PRODUCT
50 #define DEBUG_ARG(x) , x
51 #else
52 #define DEBUG_ARG(x)
53 #endif
54
55 // Convert Nodes to instruction bits and pass off to the VM
56 void Compile::Output() {
57 // RootNode goes
58 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" );
59
60 // The number of new nodes (mostly MachNop) is proportional to
61 // the number of java calls and inner loops which are aligned.
62 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 +
63 C->inner_loops()*(OptoLoopAlignment-1)),
64 "out of nodes before code generation" ) ) {
65 return;
66 }
67 // Make sure I can find the Start Node
68 Block *entry = _cfg->get_block(1);
69 Block *broot = _cfg->get_root_block();
70
71 const StartNode *start = entry->head()->as_Start();
72
73 // Replace StartNode with prolog
74 Label verified_entry;
75 MachPrologNode* prolog = new MachPrologNode(&verified_entry);
76 entry->map_node(prolog, 0);
77 _cfg->map_node_to_block(prolog, entry);
78 _cfg->unmap_node_from_block(start); // start is no longer in any block
79
80 // Virtual methods need an unverified entry point
81 bool has_value_entry = false;
82 if (is_osr_compilation()) {
83 if (PoisonOSREntry) {
84 // TODO: Should use a ShouldNotReachHereNode...
85 _cfg->insert( broot, 0, new MachBreakpointNode() );
86 }
87 } else {
88 if (_method && !_method->flags().is_static()) {
89 // Insert unvalidated entry point
90 _cfg->insert(broot, 0, new MachUEPNode());
91 }
92 if (_method && _method->get_Method()->has_scalarized_args()) {
93 // Insert value type entry point
94 _cfg->insert(broot, 0, new MachVVEPNode(&verified_entry));
95 has_value_entry = true;
96 }
97 }
98
99 // Break before main entry point
100 if ((_method && C->directive()->BreakAtExecuteOption) ||
101 (OptoBreakpoint && is_method_compilation()) ||
102 (OptoBreakpointOSR && is_osr_compilation()) ||
103 (OptoBreakpointC2R && !_method) ) {
104 // checking for _method means that OptoBreakpoint does not apply to
105 // runtime stubs or frame converters
106 _cfg->insert( entry, 1, new MachBreakpointNode() );
107 }
108
109 // Insert epilogs before every return
110 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
111 Block* block = _cfg->get_block(i);
112 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point?
113 Node* m = block->end();
114 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) {
115 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return);
116 block->add_inst(epilog);
117 _cfg->map_node_to_block(epilog, block);
118 }
119 }
120 }
121
122 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1);
123 blk_starts[0] = 0;
124
125 // Initialize code buffer and process short branches.
126 CodeBuffer* cb = init_buffer(blk_starts);
127
128 if (cb == NULL || failing()) {
129 return;
130 }
131
132 if (has_value_entry) {
133 // We added an entry point for unscalarized value types
134 // Compute offset of "normal" entry point
135 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, 0);
136 uint entry_offset = -1; // will be patched later
137 if (!_method->flags().is_static()) {
138 MachVVEPNode* vvep = (MachVVEPNode*)broot->get_node(0);
139 entry_offset = vvep->size(_regalloc);
140 }
141 _code_offsets.set_value(CodeOffsets::Entry, entry_offset);
142 }
143
144 ScheduleAndBundle();
145
146 #ifndef PRODUCT
147 if (trace_opto_output()) {
148 tty->print("\n---- After ScheduleAndBundle ----\n");
149 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
150 tty->print("\nBB#%03d:\n", i);
151 Block* block = _cfg->get_block(i);
152 for (uint j = 0; j < block->number_of_nodes(); j++) {
153 Node* n = block->get_node(j);
154 OptoReg::Name reg = _regalloc->get_reg_first(n);
155 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : "");
156 n->dump();
157 }
158 }
159 }
160 #endif
161
162 if (failing()) {
163 return;
164 }
165
166 BuildOopMaps();
167
168 if (failing()) {
169 return;
170 }
171
172 fill_buffer(cb, blk_starts);
173 }
174
175 bool Compile::need_stack_bang(int frame_size_in_bytes) const {
176 // Determine if we need to generate a stack overflow check.
177 // Do it if the method is not a stub function and
178 // has java calls or has frame size > vm_page_size/8.
179 // The debug VM checks that deoptimization doesn't trigger an
180 // unexpected stack overflow (compiled method stack banging should
181 // guarantee it doesn't happen) so we always need the stack bang in
182 // a debug VM.
183 return (UseStackBanging && stub_function() == NULL &&
184 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3
185 DEBUG_ONLY(|| true)));
186 }
187
188 bool Compile::need_register_stack_bang() const {
189 // Determine if we need to generate a register stack overflow check.
190 // This is only used on architectures which have split register
191 // and memory stacks (ie. IA64).
192 // Bang if the method is not a stub function and has java calls
193 return (stub_function() == NULL && has_java_calls());
194 }
195
196
197 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top
198 // of a loop. When aligning a loop we need to provide enough instructions
199 // in cpu's fetch buffer to feed decoders. The loop alignment could be
200 // avoided if we have enough instructions in fetch buffer at the head of a loop.
201 // By default, the size is set to 999999 by Block's constructor so that
202 // a loop will be aligned if the size is not reset here.
203 //
204 // Note: Mach instructions could contain several HW instructions
205 // so the size is estimated only.
206 //
207 void Compile::compute_loop_first_inst_sizes() {
208 // The next condition is used to gate the loop alignment optimization.
209 // Don't aligned a loop if there are enough instructions at the head of a loop
210 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad
211 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is
212 // equal to 11 bytes which is the largest address NOP instruction.
213 if (MaxLoopPad < OptoLoopAlignment - 1) {
214 uint last_block = _cfg->number_of_blocks() - 1;
215 for (uint i = 1; i <= last_block; i++) {
216 Block* block = _cfg->get_block(i);
217 // Check the first loop's block which requires an alignment.
218 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) {
219 uint sum_size = 0;
220 uint inst_cnt = NumberOfLoopInstrToAlign;
221 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
222
223 // Check subsequent fallthrough blocks if the loop's first
224 // block(s) does not have enough instructions.
225 Block *nb = block;
226 while(inst_cnt > 0 &&
227 i < last_block &&
228 !_cfg->get_block(i + 1)->has_loop_alignment() &&
229 !nb->has_successor(block)) {
230 i++;
231 nb = _cfg->get_block(i);
232 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc);
233 } // while( inst_cnt > 0 && i < last_block )
234
235 block->set_first_inst_size(sum_size);
236 } // f( b->head()->is_Loop() )
237 } // for( i <= last_block )
238 } // if( MaxLoopPad < OptoLoopAlignment-1 )
239 }
240
241 // The architecture description provides short branch variants for some long
242 // branch instructions. Replace eligible long branches with short branches.
243 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) {
244 // Compute size of each block, method size, and relocation information size
245 uint nblocks = _cfg->number_of_blocks();
246
247 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
248 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
249 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks);
250
251 // Collect worst case block paddings
252 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks);
253 memset(block_worst_case_pad, 0, nblocks * sizeof(int));
254
255 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); )
256 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); )
257
258 bool has_short_branch_candidate = false;
259
260 // Initialize the sizes to 0
261 code_size = 0; // Size in bytes of generated code
262 stub_size = 0; // Size in bytes of all stub entries
263 // Size in bytes of all relocation entries, including those in local stubs.
264 // Start with 2-bytes of reloc info for the unvalidated entry point
265 reloc_size = 1; // Number of relocation entries
266
267 // Make three passes. The first computes pessimistic blk_starts,
268 // relative jmp_offset and reloc_size information. The second performs
269 // short branch substitution using the pessimistic sizing. The
270 // third inserts nops where needed.
271
272 // Step one, perform a pessimistic sizing pass.
273 uint last_call_adr = max_juint;
274 uint last_avoid_back_to_back_adr = max_juint;
275 uint nop_size = (new MachNopNode())->size(_regalloc);
276 for (uint i = 0; i < nblocks; i++) { // For all blocks
277 Block* block = _cfg->get_block(i);
278
279 // During short branch replacement, we store the relative (to blk_starts)
280 // offset of jump in jmp_offset, rather than the absolute offset of jump.
281 // This is so that we do not need to recompute sizes of all nodes when
282 // we compute correct blk_starts in our next sizing pass.
283 jmp_offset[i] = 0;
284 jmp_size[i] = 0;
285 jmp_nidx[i] = -1;
286 DEBUG_ONLY( jmp_target[i] = 0; )
287 DEBUG_ONLY( jmp_rule[i] = 0; )
288
289 // Sum all instruction sizes to compute block size
290 uint last_inst = block->number_of_nodes();
291 uint blk_size = 0;
292 for (uint j = 0; j < last_inst; j++) {
293 Node* nj = block->get_node(j);
294 // Handle machine instruction nodes
295 if (nj->is_Mach()) {
296 MachNode *mach = nj->as_Mach();
297 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding
298 reloc_size += mach->reloc();
299 if (mach->is_MachCall()) {
300 // add size information for trampoline stub
301 // class CallStubImpl is platform-specific and defined in the *.ad files.
302 stub_size += CallStubImpl::size_call_trampoline();
303 reloc_size += CallStubImpl::reloc_call_trampoline();
304
305 MachCallNode *mcall = mach->as_MachCall();
306 // This destination address is NOT PC-relative
307
308 if (mcall->entry_point() != NULL) {
309 mcall->method_set((intptr_t)mcall->entry_point());
310 }
311
312 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) {
313 stub_size += CompiledStaticCall::to_interp_stub_size();
314 reloc_size += CompiledStaticCall::reloc_to_interp_stub();
315 #if INCLUDE_AOT
316 stub_size += CompiledStaticCall::to_aot_stub_size();
317 reloc_size += CompiledStaticCall::reloc_to_aot_stub();
318 #endif
319 }
320 } else if (mach->is_MachSafePoint()) {
321 // If call/safepoint are adjacent, account for possible
322 // nop to disambiguate the two safepoints.
323 // ScheduleAndBundle() can rearrange nodes in a block,
324 // check for all offsets inside this block.
325 if (last_call_adr >= blk_starts[i]) {
326 blk_size += nop_size;
327 }
328 }
329 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
330 // Nop is inserted between "avoid back to back" instructions.
331 // ScheduleAndBundle() can rearrange nodes in a block,
332 // check for all offsets inside this block.
333 if (last_avoid_back_to_back_adr >= blk_starts[i]) {
334 blk_size += nop_size;
335 }
336 }
337 if (mach->may_be_short_branch()) {
338 if (!nj->is_MachBranch()) {
339 #ifndef PRODUCT
340 nj->dump(3);
341 #endif
342 Unimplemented();
343 }
344 assert(jmp_nidx[i] == -1, "block should have only one branch");
345 jmp_offset[i] = blk_size;
346 jmp_size[i] = nj->size(_regalloc);
347 jmp_nidx[i] = j;
348 has_short_branch_candidate = true;
349 }
350 }
351 blk_size += nj->size(_regalloc);
352 // Remember end of call offset
353 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) {
354 last_call_adr = blk_starts[i]+blk_size;
355 }
356 // Remember end of avoid_back_to_back offset
357 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
358 last_avoid_back_to_back_adr = blk_starts[i]+blk_size;
359 }
360 }
361
362 // When the next block starts a loop, we may insert pad NOP
363 // instructions. Since we cannot know our future alignment,
364 // assume the worst.
365 if (i < nblocks - 1) {
366 Block* nb = _cfg->get_block(i + 1);
367 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit();
368 if (max_loop_pad > 0) {
369 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), "");
370 // Adjust last_call_adr and/or last_avoid_back_to_back_adr.
371 // If either is the last instruction in this block, bump by
372 // max_loop_pad in lock-step with blk_size, so sizing
373 // calculations in subsequent blocks still can conservatively
374 // detect that it may the last instruction in this block.
375 if (last_call_adr == blk_starts[i]+blk_size) {
376 last_call_adr += max_loop_pad;
377 }
378 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) {
379 last_avoid_back_to_back_adr += max_loop_pad;
380 }
381 blk_size += max_loop_pad;
382 block_worst_case_pad[i + 1] = max_loop_pad;
383 }
384 }
385
386 // Save block size; update total method size
387 blk_starts[i+1] = blk_starts[i]+blk_size;
388 }
389
390 // Step two, replace eligible long jumps.
391 bool progress = true;
392 uint last_may_be_short_branch_adr = max_juint;
393 while (has_short_branch_candidate && progress) {
394 progress = false;
395 has_short_branch_candidate = false;
396 int adjust_block_start = 0;
397 for (uint i = 0; i < nblocks; i++) {
398 Block* block = _cfg->get_block(i);
399 int idx = jmp_nidx[i];
400 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach();
401 if (mach != NULL && mach->may_be_short_branch()) {
402 #ifdef ASSERT
403 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity");
404 int j;
405 // Find the branch; ignore trailing NOPs.
406 for (j = block->number_of_nodes()-1; j>=0; j--) {
407 Node* n = block->get_node(j);
408 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con)
409 break;
410 }
411 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity");
412 #endif
413 int br_size = jmp_size[i];
414 int br_offs = blk_starts[i] + jmp_offset[i];
415
416 // This requires the TRUE branch target be in succs[0]
417 uint bnum = block->non_connector_successor(0)->_pre_order;
418 int offset = blk_starts[bnum] - br_offs;
419 if (bnum > i) { // adjust following block's offset
420 offset -= adjust_block_start;
421 }
422
423 // This block can be a loop header, account for the padding
424 // in the previous block.
425 int block_padding = block_worst_case_pad[i];
426 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top");
427 // In the following code a nop could be inserted before
428 // the branch which will increase the backward distance.
429 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr);
430 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block");
431
432 if (needs_padding && offset <= 0)
433 offset -= nop_size;
434
435 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
436 // We've got a winner. Replace this branch.
437 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
438
439 // Update the jmp_size.
440 int new_size = replacement->size(_regalloc);
441 int diff = br_size - new_size;
442 assert(diff >= (int)nop_size, "short_branch size should be smaller");
443 // Conservatively take into account padding between
444 // avoid_back_to_back branches. Previous branch could be
445 // converted into avoid_back_to_back branch during next
446 // rounds.
447 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
448 jmp_offset[i] += nop_size;
449 diff -= nop_size;
450 }
451 adjust_block_start += diff;
452 block->map_node(replacement, idx);
453 mach->subsume_by(replacement, C);
454 mach = replacement;
455 progress = true;
456
457 jmp_size[i] = new_size;
458 DEBUG_ONLY( jmp_target[i] = bnum; );
459 DEBUG_ONLY( jmp_rule[i] = mach->rule(); );
460 } else {
461 // The jump distance is not short, try again during next iteration.
462 has_short_branch_candidate = true;
463 }
464 } // (mach->may_be_short_branch())
465 if (mach != NULL && (mach->may_be_short_branch() ||
466 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) {
467 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i];
468 }
469 blk_starts[i+1] -= adjust_block_start;
470 }
471 }
472
473 #ifdef ASSERT
474 for (uint i = 0; i < nblocks; i++) { // For all blocks
475 if (jmp_target[i] != 0) {
476 int br_size = jmp_size[i];
477 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
478 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
479 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
480 }
481 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp");
482 }
483 }
484 #endif
485
486 // Step 3, compute the offsets of all blocks, will be done in fill_buffer()
487 // after ScheduleAndBundle().
488
489 // ------------------
490 // Compute size for code buffer
491 code_size = blk_starts[nblocks];
492
493 // Relocation records
494 reloc_size += 1; // Relo entry for exception handler
495
496 // Adjust reloc_size to number of record of relocation info
497 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for
498 // a relocation index.
499 // The CodeBuffer will expand the locs array if this estimate is too low.
500 reloc_size *= 10 / sizeof(relocInfo);
501 }
502
503 //------------------------------FillLocArray-----------------------------------
504 // Create a bit of debug info and append it to the array. The mapping is from
505 // Java local or expression stack to constant, register or stack-slot. For
506 // doubles, insert 2 mappings and return 1 (to tell the caller that the next
507 // entry has been taken care of and caller should skip it).
508 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) {
509 // This should never have accepted Bad before
510 assert(OptoReg::is_valid(regnum), "location must be valid");
511 return (OptoReg::is_reg(regnum))
512 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) )
513 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum)));
514 }
515
516
517 ObjectValue*
518 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) {
519 for (int i = 0; i < objs->length(); i++) {
520 assert(objs->at(i)->is_object(), "corrupt object cache");
521 ObjectValue* sv = (ObjectValue*) objs->at(i);
522 if (sv->id() == id) {
523 return sv;
524 }
525 }
526 // Otherwise..
527 return NULL;
528 }
529
530 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
531 ObjectValue* sv ) {
532 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition");
533 objs->append(sv);
534 }
535
536
537 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local,
538 GrowableArray<ScopeValue*> *array,
539 GrowableArray<ScopeValue*> *objs ) {
540 assert( local, "use _top instead of null" );
541 if (array->length() != idx) {
542 assert(array->length() == idx + 1, "Unexpected array count");
543 // Old functionality:
544 // return
545 // New functionality:
546 // Assert if the local is not top. In product mode let the new node
547 // override the old entry.
548 assert(local == top(), "LocArray collision");
549 if (local == top()) {
550 return;
551 }
552 array->pop();
553 }
554 const Type *t = local->bottom_type();
555
556 // Is it a safepoint scalar object node?
557 if (local->is_SafePointScalarObject()) {
558 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
559
560 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
561 if (sv == NULL) {
562 ciKlass* cik = t->is_oopptr()->klass();
563 assert(cik->is_instance_klass() ||
564 cik->is_array_klass(), "Not supported allocation.");
565 sv = new ObjectValue(spobj->_idx,
566 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
567 Compile::set_sv_for_object_node(objs, sv);
568
569 uint first_ind = spobj->first_index(sfpt->jvms());
570 for (uint i = 0; i < spobj->n_fields(); i++) {
571 Node* fld_node = sfpt->in(first_ind+i);
572 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs);
573 }
574 }
575 array->append(sv);
576 return;
577 }
578
579 // Grab the register number for the local
580 OptoReg::Name regnum = _regalloc->get_reg_first(local);
581 if( OptoReg::is_valid(regnum) ) {// Got a register/stack?
582 // Record the double as two float registers.
583 // The register mask for such a value always specifies two adjacent
584 // float registers, with the lower register number even.
585 // Normally, the allocation of high and low words to these registers
586 // is irrelevant, because nearly all operations on register pairs
587 // (e.g., StoreD) treat them as a single unit.
588 // Here, we assume in addition that the words in these two registers
589 // stored "naturally" (by operations like StoreD and double stores
590 // within the interpreter) such that the lower-numbered register
591 // is written to the lower memory address. This may seem like
592 // a machine dependency, but it is not--it is a requirement on
593 // the author of the <arch>.ad file to ensure that, for every
594 // even/odd double-register pair to which a double may be allocated,
595 // the word in the even single-register is stored to the first
596 // memory word. (Note that register numbers are completely
597 // arbitrary, and are not tied to any machine-level encodings.)
598 #ifdef _LP64
599 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) {
600 array->append(new ConstantIntValue((jint)0));
601 array->append(new_loc_value( _regalloc, regnum, Location::dbl ));
602 } else if ( t->base() == Type::Long ) {
603 array->append(new ConstantIntValue((jint)0));
604 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
605 } else if ( t->base() == Type::RawPtr ) {
606 // jsr/ret return address which must be restored into a the full
607 // width 64-bit stack slot.
608 array->append(new_loc_value( _regalloc, regnum, Location::lng ));
609 }
610 #else //_LP64
611 #ifdef SPARC
612 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) {
613 // For SPARC we have to swap high and low words for
614 // long values stored in a single-register (g0-g7).
615 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
616 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
617 } else
618 #endif //SPARC
619 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) {
620 // Repack the double/long as two jints.
621 // The convention the interpreter uses is that the second local
622 // holds the first raw word of the native double representation.
623 // This is actually reasonable, since locals and stack arrays
624 // grow downwards in all implementations.
625 // (If, on some machine, the interpreter's Java locals or stack
626 // were to grow upwards, the embedded doubles would be word-swapped.)
627 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal ));
628 array->append(new_loc_value( _regalloc, regnum , Location::normal ));
629 }
630 #endif //_LP64
631 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) &&
632 OptoReg::is_reg(regnum) ) {
633 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double()
634 ? Location::float_in_dbl : Location::normal ));
635 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) {
636 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long
637 ? Location::int_in_long : Location::normal ));
638 } else if( t->base() == Type::NarrowOop ) {
639 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop ));
640 } else {
641 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal ));
642 }
643 return;
644 }
645
646 // No register. It must be constant data.
647 switch (t->base()) {
648 case Type::Half: // Second half of a double
649 ShouldNotReachHere(); // Caller should skip 2nd halves
650 break;
651 case Type::AnyPtr:
652 array->append(new ConstantOopWriteValue(NULL));
653 break;
654 case Type::AryPtr:
655 case Type::InstPtr: // fall through
656 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding()));
657 break;
658 case Type::NarrowOop:
659 if (t == TypeNarrowOop::NULL_PTR) {
660 array->append(new ConstantOopWriteValue(NULL));
661 } else {
662 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding()));
663 }
664 break;
665 case Type::Int:
666 array->append(new ConstantIntValue(t->is_int()->get_con()));
667 break;
668 case Type::RawPtr:
669 // A return address (T_ADDRESS).
670 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI");
671 #ifdef _LP64
672 // Must be restored to the full-width 64-bit stack slot.
673 array->append(new ConstantLongValue(t->is_ptr()->get_con()));
674 #else
675 array->append(new ConstantIntValue(t->is_ptr()->get_con()));
676 #endif
677 break;
678 case Type::FloatCon: {
679 float f = t->is_float_constant()->getf();
680 array->append(new ConstantIntValue(jint_cast(f)));
681 break;
682 }
683 case Type::DoubleCon: {
684 jdouble d = t->is_double_constant()->getd();
685 #ifdef _LP64
686 array->append(new ConstantIntValue((jint)0));
687 array->append(new ConstantDoubleValue(d));
688 #else
689 // Repack the double as two jints.
690 // The convention the interpreter uses is that the second local
691 // holds the first raw word of the native double representation.
692 // This is actually reasonable, since locals and stack arrays
693 // grow downwards in all implementations.
694 // (If, on some machine, the interpreter's Java locals or stack
695 // were to grow upwards, the embedded doubles would be word-swapped.)
696 jlong_accessor acc;
697 acc.long_value = jlong_cast(d);
698 array->append(new ConstantIntValue(acc.words[1]));
699 array->append(new ConstantIntValue(acc.words[0]));
700 #endif
701 break;
702 }
703 case Type::Long: {
704 jlong d = t->is_long()->get_con();
705 #ifdef _LP64
706 array->append(new ConstantIntValue((jint)0));
707 array->append(new ConstantLongValue(d));
708 #else
709 // Repack the long as two jints.
710 // The convention the interpreter uses is that the second local
711 // holds the first raw word of the native double representation.
712 // This is actually reasonable, since locals and stack arrays
713 // grow downwards in all implementations.
714 // (If, on some machine, the interpreter's Java locals or stack
715 // were to grow upwards, the embedded doubles would be word-swapped.)
716 jlong_accessor acc;
717 acc.long_value = d;
718 array->append(new ConstantIntValue(acc.words[1]));
719 array->append(new ConstantIntValue(acc.words[0]));
720 #endif
721 break;
722 }
723 case Type::Top: // Add an illegal value here
724 array->append(new LocationValue(Location()));
725 break;
726 default:
727 ShouldNotReachHere();
728 break;
729 }
730 }
731
732 // Determine if this node starts a bundle
733 bool Compile::starts_bundle(const Node *n) const {
734 return (_node_bundling_limit > n->_idx &&
735 _node_bundling_base[n->_idx].starts_bundle());
736 }
737
738 //--------------------------Process_OopMap_Node--------------------------------
739 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) {
740
741 // Handle special safepoint nodes for synchronization
742 MachSafePointNode *sfn = mach->as_MachSafePoint();
743 MachCallNode *mcall;
744
745 int safepoint_pc_offset = current_offset;
746 bool is_method_handle_invoke = false;
747 bool return_oop = false;
748 bool return_vt = false;
749
750 // Add the safepoint in the DebugInfoRecorder
751 if( !mach->is_MachCall() ) {
752 mcall = NULL;
753 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map);
754 } else {
755 mcall = mach->as_MachCall();
756
757 // Is the call a MethodHandle call?
758 if (mcall->is_MachCallJava()) {
759 if (mcall->as_MachCallJava()->_method_handle_invoke) {
760 assert(has_method_handle_invokes(), "must have been set during call generation");
761 is_method_handle_invoke = true;
762 }
763 }
764
765 // Check if a call returns an object.
766 if (mcall->returns_pointer() || mcall->returns_vt()) {
767 return_oop = true;
768 }
769 if (mcall->returns_vt()) {
770 return_vt = true;
771 }
772 safepoint_pc_offset += mcall->ret_addr_offset();
773 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map);
774 }
775
776 // Loop over the JVMState list to add scope information
777 // Do not skip safepoints with a NULL method, they need monitor info
778 JVMState* youngest_jvms = sfn->jvms();
779 int max_depth = youngest_jvms->depth();
780
781 // Allocate the object pool for scalar-replaced objects -- the map from
782 // small-integer keys (which can be recorded in the local and ostack
783 // arrays) to descriptions of the object state.
784 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>();
785
786 // Visit scopes from oldest to youngest.
787 for (int depth = 1; depth <= max_depth; depth++) {
788 JVMState* jvms = youngest_jvms->of_depth(depth);
789 int idx;
790 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
791 // Safepoints that do not have method() set only provide oop-map and monitor info
792 // to support GC; these do not support deoptimization.
793 int num_locs = (method == NULL) ? 0 : jvms->loc_size();
794 int num_exps = (method == NULL) ? 0 : jvms->stk_size();
795 int num_mon = jvms->nof_monitors();
796 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(),
797 "JVMS local count must match that of the method");
798
799 // Add Local and Expression Stack Information
800
801 // Insert locals into the locarray
802 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs);
803 for( idx = 0; idx < num_locs; idx++ ) {
804 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs );
805 }
806
807 // Insert expression stack entries into the exparray
808 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps);
809 for( idx = 0; idx < num_exps; idx++ ) {
810 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs );
811 }
812
813 // Add in mappings of the monitors
814 assert( !method ||
815 !method->is_synchronized() ||
816 method->is_native() ||
817 num_mon > 0 ||
818 !GenerateSynchronizationCode,
819 "monitors must always exist for synchronized methods");
820
821 // Build the growable array of ScopeValues for exp stack
822 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon);
823
824 // Loop over monitors and insert into array
825 for (idx = 0; idx < num_mon; idx++) {
826 // Grab the node that defines this monitor
827 Node* box_node = sfn->monitor_box(jvms, idx);
828 Node* obj_node = sfn->monitor_obj(jvms, idx);
829
830 // Create ScopeValue for object
831 ScopeValue *scval = NULL;
832
833 if (obj_node->is_SafePointScalarObject()) {
834 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject();
835 scval = Compile::sv_for_node_id(objs, spobj->_idx);
836 if (scval == NULL) {
837 const Type *t = spobj->bottom_type();
838 ciKlass* cik = t->is_oopptr()->klass();
839 assert(cik->is_instance_klass() ||
840 cik->is_array_klass(), "Not supported allocation.");
841 ObjectValue* sv = new ObjectValue(spobj->_idx,
842 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding()));
843 Compile::set_sv_for_object_node(objs, sv);
844
845 uint first_ind = spobj->first_index(youngest_jvms);
846 for (uint i = 0; i < spobj->n_fields(); i++) {
847 Node* fld_node = sfn->in(first_ind+i);
848 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs);
849 }
850 scval = sv;
851 }
852 } else if (!obj_node->is_Con()) {
853 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node);
854 if( obj_node->bottom_type()->base() == Type::NarrowOop ) {
855 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop );
856 } else {
857 scval = new_loc_value( _regalloc, obj_reg, Location::oop );
858 }
859 } else {
860 const TypePtr *tp = obj_node->get_ptr_type();
861 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding());
862 }
863
864 OptoReg::Name box_reg = BoxLockNode::reg(box_node);
865 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg));
866 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated());
867 monarray->append(new MonitorValue(scval, basic_lock, eliminated));
868 }
869
870 // We dump the object pool first, since deoptimization reads it in first.
871 debug_info()->dump_object_pool(objs);
872
873 // Build first class objects to pass to scope
874 DebugToken *locvals = debug_info()->create_scope_values(locarray);
875 DebugToken *expvals = debug_info()->create_scope_values(exparray);
876 DebugToken *monvals = debug_info()->create_monitor_values(monarray);
877
878 // Make method available for all Safepoints
879 ciMethod* scope_method = method ? method : _method;
880 // Describe the scope here
881 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI");
882 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest");
883 // Now we can describe the scope.
884 methodHandle null_mh;
885 bool rethrow_exception = false;
886 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, return_vt, locvals, expvals, monvals);
887 } // End jvms loop
888
889 // Mark the end of the scope set.
890 debug_info()->end_safepoint(safepoint_pc_offset);
891 }
892
893
894
895 // A simplified version of Process_OopMap_Node, to handle non-safepoints.
896 class NonSafepointEmitter {
897 Compile* C;
898 JVMState* _pending_jvms;
899 int _pending_offset;
900
901 void emit_non_safepoint();
902
903 public:
904 NonSafepointEmitter(Compile* compile) {
905 this->C = compile;
906 _pending_jvms = NULL;
907 _pending_offset = 0;
908 }
909
910 void observe_instruction(Node* n, int pc_offset) {
911 if (!C->debug_info()->recording_non_safepoints()) return;
912
913 Node_Notes* nn = C->node_notes_at(n->_idx);
914 if (nn == NULL || nn->jvms() == NULL) return;
915 if (_pending_jvms != NULL &&
916 _pending_jvms->same_calls_as(nn->jvms())) {
917 // Repeated JVMS? Stretch it up here.
918 _pending_offset = pc_offset;
919 } else {
920 if (_pending_jvms != NULL &&
921 _pending_offset < pc_offset) {
922 emit_non_safepoint();
923 }
924 _pending_jvms = NULL;
925 if (pc_offset > C->debug_info()->last_pc_offset()) {
926 // This is the only way _pending_jvms can become non-NULL:
927 _pending_jvms = nn->jvms();
928 _pending_offset = pc_offset;
929 }
930 }
931 }
932
933 // Stay out of the way of real safepoints:
934 void observe_safepoint(JVMState* jvms, int pc_offset) {
935 if (_pending_jvms != NULL &&
936 !_pending_jvms->same_calls_as(jvms) &&
937 _pending_offset < pc_offset) {
938 emit_non_safepoint();
939 }
940 _pending_jvms = NULL;
941 }
942
943 void flush_at_end() {
944 if (_pending_jvms != NULL) {
945 emit_non_safepoint();
946 }
947 _pending_jvms = NULL;
948 }
949 };
950
951 void NonSafepointEmitter::emit_non_safepoint() {
952 JVMState* youngest_jvms = _pending_jvms;
953 int pc_offset = _pending_offset;
954
955 // Clear it now:
956 _pending_jvms = NULL;
957
958 DebugInformationRecorder* debug_info = C->debug_info();
959 assert(debug_info->recording_non_safepoints(), "sanity");
960
961 debug_info->add_non_safepoint(pc_offset);
962 int max_depth = youngest_jvms->depth();
963
964 // Visit scopes from oldest to youngest.
965 for (int depth = 1; depth <= max_depth; depth++) {
966 JVMState* jvms = youngest_jvms->of_depth(depth);
967 ciMethod* method = jvms->has_method() ? jvms->method() : NULL;
968 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest");
969 methodHandle null_mh;
970 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute());
971 }
972
973 // Mark the end of the scope set.
974 debug_info->end_non_safepoint(pc_offset);
975 }
976
977 //------------------------------init_buffer------------------------------------
978 CodeBuffer* Compile::init_buffer(uint* blk_starts) {
979
980 // Set the initially allocated size
981 int code_req = initial_code_capacity;
982 int locs_req = initial_locs_capacity;
983 int stub_req = initial_stub_capacity;
984 int const_req = initial_const_capacity;
985
986 int pad_req = NativeCall::instruction_size;
987 // The extra spacing after the code is necessary on some platforms.
988 // Sometimes we need to patch in a jump after the last instruction,
989 // if the nmethod has been deoptimized. (See 4932387, 4894843.)
990
991 // Compute the byte offset where we can store the deopt pc.
992 if (fixed_slots() != 0) {
993 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot));
994 }
995 if (C->needs_stack_repair()) {
996 // Compute the byte offset of the stack increment value
997 _sp_inc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_sp_inc_slot));
998 }
999
1000 // Compute prolog code size
1001 _method_size = 0;
1002 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize;
1003 #if defined(IA64) && !defined(AIX)
1004 if (save_argument_registers()) {
1005 // 4815101: this is a stub with implicit and unknown precision fp args.
1006 // The usual spill mechanism can only generate stfd's in this case, which
1007 // doesn't work if the fp reg to spill contains a single-precision denorm.
1008 // Instead, we hack around the normal spill mechanism using stfspill's and
1009 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate
1010 // space here for the fp arg regs (f8-f15) we're going to thusly spill.
1011 //
1012 // If we ever implement 16-byte 'registers' == stack slots, we can
1013 // get rid of this hack and have SpillCopy generate stfspill/ldffill
1014 // instead of stfd/stfs/ldfd/ldfs.
1015 _frame_slots += 8*(16/BytesPerInt);
1016 }
1017 #endif
1018 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check");
1019
1020 if (has_mach_constant_base_node()) {
1021 uint add_size = 0;
1022 // Fill the constant table.
1023 // Note: This must happen before shorten_branches.
1024 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1025 Block* b = _cfg->get_block(i);
1026
1027 for (uint j = 0; j < b->number_of_nodes(); j++) {
1028 Node* n = b->get_node(j);
1029
1030 // If the node is a MachConstantNode evaluate the constant
1031 // value section.
1032 if (n->is_MachConstant()) {
1033 MachConstantNode* machcon = n->as_MachConstant();
1034 machcon->eval_constant(C);
1035 } else if (n->is_Mach()) {
1036 // On Power there are more nodes that issue constants.
1037 add_size += (n->as_Mach()->ins_num_consts() * 8);
1038 }
1039 }
1040 }
1041
1042 // Calculate the offsets of the constants and the size of the
1043 // constant table (including the padding to the next section).
1044 constant_table().calculate_offsets_and_size();
1045 const_req = constant_table().size() + add_size;
1046 }
1047
1048 // Initialize the space for the BufferBlob used to find and verify
1049 // instruction size in MachNode::emit_size()
1050 init_scratch_buffer_blob(const_req);
1051 if (failing()) return NULL; // Out of memory
1052
1053 // Pre-compute the length of blocks and replace
1054 // long branches with short if machine supports it.
1055 shorten_branches(blk_starts, code_req, locs_req, stub_req);
1056
1057 // nmethod and CodeBuffer count stubs & constants as part of method's code.
1058 // class HandlerImpl is platform-specific and defined in the *.ad files.
1059 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler
1060 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler
1061 stub_req += MAX_stubs_size; // ensure per-stub margin
1062 code_req += MAX_inst_size; // ensure per-instruction margin
1063
1064 if (StressCodeBuffers)
1065 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion
1066
1067 int total_req =
1068 const_req +
1069 code_req +
1070 pad_req +
1071 stub_req +
1072 exception_handler_req +
1073 deopt_handler_req; // deopt handler
1074
1075 if (has_method_handle_invokes())
1076 total_req += deopt_handler_req; // deopt MH handler
1077
1078 CodeBuffer* cb = code_buffer();
1079 cb->initialize(total_req, locs_req);
1080
1081 // Have we run out of code space?
1082 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1083 C->record_failure("CodeCache is full");
1084 return NULL;
1085 }
1086 // Configure the code buffer.
1087 cb->initialize_consts_size(const_req);
1088 cb->initialize_stubs_size(stub_req);
1089 cb->initialize_oop_recorder(env()->oop_recorder());
1090
1091 // fill in the nop array for bundling computations
1092 MachNode *_nop_list[Bundle::_nop_count];
1093 Bundle::initialize_nops(_nop_list);
1094
1095 return cb;
1096 }
1097
1098 //------------------------------fill_buffer------------------------------------
1099 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) {
1100 // blk_starts[] contains offsets calculated during short branches processing,
1101 // offsets should not be increased during following steps.
1102
1103 // Compute the size of first NumberOfLoopInstrToAlign instructions at head
1104 // of a loop. It is used to determine the padding for loop alignment.
1105 compute_loop_first_inst_sizes();
1106
1107 // Create oopmap set.
1108 _oop_map_set = new OopMapSet();
1109
1110 // !!!!! This preserves old handling of oopmaps for now
1111 debug_info()->set_oopmaps(_oop_map_set);
1112
1113 uint nblocks = _cfg->number_of_blocks();
1114 // Count and start of implicit null check instructions
1115 uint inct_cnt = 0;
1116 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1117
1118 // Count and start of calls
1119 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1);
1120
1121 uint return_offset = 0;
1122 int nop_size = (new MachNopNode())->size(_regalloc);
1123
1124 int previous_offset = 0;
1125 int current_offset = 0;
1126 int last_call_offset = -1;
1127 int last_avoid_back_to_back_offset = -1;
1128 #ifdef ASSERT
1129 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks);
1130 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks);
1131 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks);
1132 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks);
1133 #endif
1134
1135 // Create an array of unused labels, one for each basic block, if printing is enabled
1136 #ifndef PRODUCT
1137 int *node_offsets = NULL;
1138 uint node_offset_limit = unique();
1139
1140 if (print_assembly())
1141 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit);
1142 #endif
1143
1144 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily
1145
1146 // Emit the constant table.
1147 if (has_mach_constant_base_node()) {
1148 constant_table().emit(*cb);
1149 }
1150
1151 // Create an array of labels, one for each basic block
1152 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1);
1153 for (uint i=0; i <= nblocks; i++) {
1154 blk_labels[i].init();
1155 }
1156
1157 // ------------------
1158 // Now fill in the code buffer
1159 Node *delay_slot = NULL;
1160
1161 for (uint i = 0; i < nblocks; i++) {
1162 Block* block = _cfg->get_block(i);
1163 Node* head = block->head();
1164
1165 // If this block needs to start aligned (i.e, can be reached other
1166 // than by falling-thru from the previous block), then force the
1167 // start of a new bundle.
1168 if (Pipeline::requires_bundling() && starts_bundle(head)) {
1169 cb->flush_bundle(true);
1170 }
1171
1172 #ifdef ASSERT
1173 if (!block->is_connector()) {
1174 stringStream st;
1175 block->dump_head(_cfg, &st);
1176 MacroAssembler(cb).block_comment(st.as_string());
1177 }
1178 jmp_target[i] = 0;
1179 jmp_offset[i] = 0;
1180 jmp_size[i] = 0;
1181 jmp_rule[i] = 0;
1182 #endif
1183 int blk_offset = current_offset;
1184
1185 // Define the label at the beginning of the basic block
1186 MacroAssembler(cb).bind(blk_labels[block->_pre_order]);
1187
1188 uint last_inst = block->number_of_nodes();
1189
1190 // Emit block normally, except for last instruction.
1191 // Emit means "dump code bits into code buffer".
1192 for (uint j = 0; j<last_inst; j++) {
1193
1194 // Get the node
1195 Node* n = block->get_node(j);
1196
1197 // See if delay slots are supported
1198 if (valid_bundle_info(n) &&
1199 node_bundling(n)->used_in_unconditional_delay()) {
1200 assert(delay_slot == NULL, "no use of delay slot node");
1201 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size");
1202
1203 delay_slot = n;
1204 continue;
1205 }
1206
1207 // If this starts a new instruction group, then flush the current one
1208 // (but allow split bundles)
1209 if (Pipeline::requires_bundling() && starts_bundle(n))
1210 cb->flush_bundle(false);
1211
1212 // Special handling for SafePoint/Call Nodes
1213 bool is_mcall = false;
1214 if (n->is_Mach()) {
1215 MachNode *mach = n->as_Mach();
1216 is_mcall = n->is_MachCall();
1217 bool is_sfn = n->is_MachSafePoint();
1218
1219 // If this requires all previous instructions be flushed, then do so
1220 if (is_sfn || is_mcall || mach->alignment_required() != 1) {
1221 cb->flush_bundle(true);
1222 current_offset = cb->insts_size();
1223 }
1224
1225 // A padding may be needed again since a previous instruction
1226 // could be moved to delay slot.
1227
1228 // align the instruction if necessary
1229 int padding = mach->compute_padding(current_offset);
1230 // Make sure safepoint node for polling is distinct from a call's
1231 // return by adding a nop if needed.
1232 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) {
1233 padding = nop_size;
1234 }
1235 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) &&
1236 current_offset == last_avoid_back_to_back_offset) {
1237 // Avoid back to back some instructions.
1238 padding = nop_size;
1239 }
1240
1241 if (padding > 0) {
1242 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size");
1243 int nops_cnt = padding / nop_size;
1244 MachNode *nop = new MachNopNode(nops_cnt);
1245 block->insert_node(nop, j++);
1246 last_inst++;
1247 _cfg->map_node_to_block(nop, block);
1248 // Ensure enough space.
1249 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1250 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1251 C->record_failure("CodeCache is full");
1252 return;
1253 }
1254 nop->emit(*cb, _regalloc);
1255 cb->flush_bundle(true);
1256 current_offset = cb->insts_size();
1257 }
1258
1259 // Remember the start of the last call in a basic block
1260 if (is_mcall) {
1261 MachCallNode *mcall = mach->as_MachCall();
1262
1263 if (mcall->entry_point() != NULL) {
1264 // This destination address is NOT PC-relative
1265 mcall->method_set((intptr_t)mcall->entry_point());
1266 }
1267
1268 // Save the return address
1269 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset();
1270
1271 if (mcall->is_MachCallLeaf()) {
1272 is_mcall = false;
1273 is_sfn = false;
1274 }
1275 }
1276
1277 // sfn will be valid whenever mcall is valid now because of inheritance
1278 if (is_sfn || is_mcall) {
1279
1280 // Handle special safepoint nodes for synchronization
1281 if (!is_mcall) {
1282 MachSafePointNode *sfn = mach->as_MachSafePoint();
1283 // !!!!! Stubs only need an oopmap right now, so bail out
1284 if (sfn->jvms()->method() == NULL) {
1285 // Write the oopmap directly to the code blob??!!
1286 continue;
1287 }
1288 } // End synchronization
1289
1290 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1291 current_offset);
1292 Process_OopMap_Node(mach, current_offset);
1293 } // End if safepoint
1294
1295 // If this is a null check, then add the start of the previous instruction to the list
1296 else if( mach->is_MachNullCheck() ) {
1297 inct_starts[inct_cnt++] = previous_offset;
1298 }
1299
1300 // If this is a branch, then fill in the label with the target BB's label
1301 else if (mach->is_MachBranch()) {
1302 // This requires the TRUE branch target be in succs[0]
1303 uint block_num = block->non_connector_successor(0)->_pre_order;
1304
1305 // Try to replace long branch if delay slot is not used,
1306 // it is mostly for back branches since forward branch's
1307 // distance is not updated yet.
1308 bool delay_slot_is_used = valid_bundle_info(n) &&
1309 node_bundling(n)->use_unconditional_delay();
1310 if (!delay_slot_is_used && mach->may_be_short_branch()) {
1311 assert(delay_slot == NULL, "not expecting delay slot node");
1312 int br_size = n->size(_regalloc);
1313 int offset = blk_starts[block_num] - current_offset;
1314 if (block_num >= i) {
1315 // Current and following block's offset are not
1316 // finalized yet, adjust distance by the difference
1317 // between calculated and final offsets of current block.
1318 offset -= (blk_starts[i] - blk_offset);
1319 }
1320 // In the following code a nop could be inserted before
1321 // the branch which will increase the backward distance.
1322 bool needs_padding = (current_offset == last_avoid_back_to_back_offset);
1323 if (needs_padding && offset <= 0)
1324 offset -= nop_size;
1325
1326 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) {
1327 // We've got a winner. Replace this branch.
1328 MachNode* replacement = mach->as_MachBranch()->short_branch_version();
1329
1330 // Update the jmp_size.
1331 int new_size = replacement->size(_regalloc);
1332 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller");
1333 // Insert padding between avoid_back_to_back branches.
1334 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) {
1335 MachNode *nop = new MachNopNode();
1336 block->insert_node(nop, j++);
1337 _cfg->map_node_to_block(nop, block);
1338 last_inst++;
1339 nop->emit(*cb, _regalloc);
1340 cb->flush_bundle(true);
1341 current_offset = cb->insts_size();
1342 }
1343 #ifdef ASSERT
1344 jmp_target[i] = block_num;
1345 jmp_offset[i] = current_offset - blk_offset;
1346 jmp_size[i] = new_size;
1347 jmp_rule[i] = mach->rule();
1348 #endif
1349 block->map_node(replacement, j);
1350 mach->subsume_by(replacement, C);
1351 n = replacement;
1352 mach = replacement;
1353 }
1354 }
1355 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num );
1356 } else if (mach->ideal_Opcode() == Op_Jump) {
1357 for (uint h = 0; h < block->_num_succs; h++) {
1358 Block* succs_block = block->_succs[h];
1359 for (uint j = 1; j < succs_block->num_preds(); j++) {
1360 Node* jpn = succs_block->pred(j);
1361 if (jpn->is_JumpProj() && jpn->in(0) == mach) {
1362 uint block_num = succs_block->non_connector()->_pre_order;
1363 Label *blkLabel = &blk_labels[block_num];
1364 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel);
1365 }
1366 }
1367 }
1368 }
1369 #ifdef ASSERT
1370 // Check that oop-store precedes the card-mark
1371 else if (mach->ideal_Opcode() == Op_StoreCM) {
1372 uint storeCM_idx = j;
1373 int count = 0;
1374 for (uint prec = mach->req(); prec < mach->len(); prec++) {
1375 Node *oop_store = mach->in(prec); // Precedence edge
1376 if (oop_store == NULL) continue;
1377 count++;
1378 uint i4;
1379 for (i4 = 0; i4 < last_inst; ++i4) {
1380 if (block->get_node(i4) == oop_store) {
1381 break;
1382 }
1383 }
1384 // Note: This test can provide a false failure if other precedence
1385 // edges have been added to the storeCMNode.
1386 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store");
1387 }
1388 assert(count > 0, "storeCM expects at least one precedence edge");
1389 }
1390 #endif
1391 else if (!n->is_Proj()) {
1392 // Remember the beginning of the previous instruction, in case
1393 // it's followed by a flag-kill and a null-check. Happens on
1394 // Intel all the time, with add-to-memory kind of opcodes.
1395 previous_offset = current_offset;
1396 }
1397
1398 // Not an else-if!
1399 // If this is a trap based cmp then add its offset to the list.
1400 if (mach->is_TrapBasedCheckNode()) {
1401 inct_starts[inct_cnt++] = current_offset;
1402 }
1403 }
1404
1405 // Verify that there is sufficient space remaining
1406 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size);
1407 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1408 C->record_failure("CodeCache is full");
1409 return;
1410 }
1411
1412 // Save the offset for the listing
1413 #ifndef PRODUCT
1414 if (node_offsets && n->_idx < node_offset_limit)
1415 node_offsets[n->_idx] = cb->insts_size();
1416 #endif
1417
1418 // "Normal" instruction case
1419 DEBUG_ONLY( uint instr_offset = cb->insts_size(); )
1420 n->emit(*cb, _regalloc);
1421 current_offset = cb->insts_size();
1422
1423 // Above we only verified that there is enough space in the instruction section.
1424 // However, the instruction may emit stubs that cause code buffer expansion.
1425 // Bail out here if expansion failed due to a lack of code cache space.
1426 if (failing()) {
1427 return;
1428 }
1429
1430 #ifdef ASSERT
1431 if (n->size(_regalloc) < (current_offset-instr_offset)) {
1432 n->dump();
1433 assert(false, "wrong size of mach node");
1434 }
1435 #endif
1436 non_safepoints.observe_instruction(n, current_offset);
1437
1438 // mcall is last "call" that can be a safepoint
1439 // record it so we can see if a poll will directly follow it
1440 // in which case we'll need a pad to make the PcDesc sites unique
1441 // see 5010568. This can be slightly inaccurate but conservative
1442 // in the case that return address is not actually at current_offset.
1443 // This is a small price to pay.
1444
1445 if (is_mcall) {
1446 last_call_offset = current_offset;
1447 }
1448
1449 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) {
1450 // Avoid back to back some instructions.
1451 last_avoid_back_to_back_offset = current_offset;
1452 }
1453
1454 // See if this instruction has a delay slot
1455 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1456 guarantee(delay_slot != NULL, "expecting delay slot node");
1457
1458 // Back up 1 instruction
1459 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size());
1460
1461 // Save the offset for the listing
1462 #ifndef PRODUCT
1463 if (node_offsets && delay_slot->_idx < node_offset_limit)
1464 node_offsets[delay_slot->_idx] = cb->insts_size();
1465 #endif
1466
1467 // Support a SafePoint in the delay slot
1468 if (delay_slot->is_MachSafePoint()) {
1469 MachNode *mach = delay_slot->as_Mach();
1470 // !!!!! Stubs only need an oopmap right now, so bail out
1471 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) {
1472 // Write the oopmap directly to the code blob??!!
1473 delay_slot = NULL;
1474 continue;
1475 }
1476
1477 int adjusted_offset = current_offset - Pipeline::instr_unit_size();
1478 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(),
1479 adjusted_offset);
1480 // Generate an OopMap entry
1481 Process_OopMap_Node(mach, adjusted_offset);
1482 }
1483
1484 // Insert the delay slot instruction
1485 delay_slot->emit(*cb, _regalloc);
1486
1487 // Don't reuse it
1488 delay_slot = NULL;
1489 }
1490
1491 } // End for all instructions in block
1492
1493 // If the next block is the top of a loop, pad this block out to align
1494 // the loop top a little. Helps prevent pipe stalls at loop back branches.
1495 if (i < nblocks-1) {
1496 Block *nb = _cfg->get_block(i + 1);
1497 int padding = nb->alignment_padding(current_offset);
1498 if( padding > 0 ) {
1499 MachNode *nop = new MachNopNode(padding / nop_size);
1500 block->insert_node(nop, block->number_of_nodes());
1501 _cfg->map_node_to_block(nop, block);
1502 nop->emit(*cb, _regalloc);
1503 current_offset = cb->insts_size();
1504 }
1505 }
1506 // Verify that the distance for generated before forward
1507 // short branches is still valid.
1508 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size");
1509
1510 // Save new block start offset
1511 blk_starts[i] = blk_offset;
1512 } // End of for all blocks
1513 blk_starts[nblocks] = current_offset;
1514
1515 non_safepoints.flush_at_end();
1516
1517 // Offset too large?
1518 if (failing()) return;
1519
1520 // Define a pseudo-label at the end of the code
1521 MacroAssembler(cb).bind( blk_labels[nblocks] );
1522
1523 // Compute the size of the first block
1524 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos();
1525
1526 #ifdef ASSERT
1527 for (uint i = 0; i < nblocks; i++) { // For all blocks
1528 if (jmp_target[i] != 0) {
1529 int br_size = jmp_size[i];
1530 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]);
1531 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) {
1532 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]);
1533 assert(false, "Displacement too large for short jmp");
1534 }
1535 }
1536 }
1537 #endif
1538
1539 #ifndef PRODUCT
1540 // Information on the size of the method, without the extraneous code
1541 Scheduling::increment_method_size(cb->insts_size());
1542 #endif
1543
1544 // ------------------
1545 // Fill in exception table entries.
1546 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels);
1547
1548 // Only java methods have exception handlers and deopt handlers
1549 // class HandlerImpl is platform-specific and defined in the *.ad files.
1550 if (_method) {
1551 // Emit the exception handler code.
1552 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb));
1553 if (failing()) {
1554 return; // CodeBuffer::expand failed
1555 }
1556 // Emit the deopt handler code.
1557 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb));
1558
1559 // Emit the MethodHandle deopt handler code (if required).
1560 if (has_method_handle_invokes() && !failing()) {
1561 // We can use the same code as for the normal deopt handler, we
1562 // just need a different entry point address.
1563 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb));
1564 }
1565 }
1566
1567 // One last check for failed CodeBuffer::expand:
1568 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) {
1569 C->record_failure("CodeCache is full");
1570 return;
1571 }
1572
1573 #ifndef PRODUCT
1574 // Dump the assembly code, including basic-block numbers
1575 if (print_assembly()) {
1576 ttyLocker ttyl; // keep the following output all in one block
1577 if (!VMThread::should_terminate()) { // test this under the tty lock
1578 // This output goes directly to the tty, not the compiler log.
1579 // To enable tools to match it up with the compilation activity,
1580 // be sure to tag this tty output with the compile ID.
1581 if (xtty != NULL) {
1582 xtty->head("opto_assembly compile_id='%d'%s", compile_id(),
1583 is_osr_compilation() ? " compile_kind='osr'" :
1584 "");
1585 }
1586 if (method() != NULL) {
1587 method()->print_metadata();
1588 }
1589 dump_asm(node_offsets, node_offset_limit);
1590 if (xtty != NULL) {
1591 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock.
1592 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done
1593 // thread safe
1594 ttyLocker ttyl2;
1595 xtty->tail("opto_assembly");
1596 }
1597 }
1598 }
1599 #endif
1600
1601 }
1602
1603 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) {
1604 _inc_table.set_size(cnt);
1605
1606 uint inct_cnt = 0;
1607 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
1608 Block* block = _cfg->get_block(i);
1609 Node *n = NULL;
1610 int j;
1611
1612 // Find the branch; ignore trailing NOPs.
1613 for (j = block->number_of_nodes() - 1; j >= 0; j--) {
1614 n = block->get_node(j);
1615 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) {
1616 break;
1617 }
1618 }
1619
1620 // If we didn't find anything, continue
1621 if (j < 0) {
1622 continue;
1623 }
1624
1625 // Compute ExceptionHandlerTable subtable entry and add it
1626 // (skip empty blocks)
1627 if (n->is_Catch()) {
1628
1629 // Get the offset of the return from the call
1630 uint call_return = call_returns[block->_pre_order];
1631 #ifdef ASSERT
1632 assert( call_return > 0, "no call seen for this basic block" );
1633 while (block->get_node(--j)->is_MachProj()) ;
1634 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
1635 #endif
1636 // last instruction is a CatchNode, find it's CatchProjNodes
1637 int nof_succs = block->_num_succs;
1638 // allocate space
1639 GrowableArray<intptr_t> handler_bcis(nof_succs);
1640 GrowableArray<intptr_t> handler_pcos(nof_succs);
1641 // iterate through all successors
1642 for (int j = 0; j < nof_succs; j++) {
1643 Block* s = block->_succs[j];
1644 bool found_p = false;
1645 for (uint k = 1; k < s->num_preds(); k++) {
1646 Node* pk = s->pred(k);
1647 if (pk->is_CatchProj() && pk->in(0) == n) {
1648 const CatchProjNode* p = pk->as_CatchProj();
1649 found_p = true;
1650 // add the corresponding handler bci & pco information
1651 if (p->_con != CatchProjNode::fall_through_index) {
1652 // p leads to an exception handler (and is not fall through)
1653 assert(s == _cfg->get_block(s->_pre_order), "bad numbering");
1654 // no duplicates, please
1655 if (!handler_bcis.contains(p->handler_bci())) {
1656 uint block_num = s->non_connector()->_pre_order;
1657 handler_bcis.append(p->handler_bci());
1658 handler_pcos.append(blk_labels[block_num].loc_pos());
1659 }
1660 }
1661 }
1662 }
1663 assert(found_p, "no matching predecessor found");
1664 // Note: Due to empty block removal, one block may have
1665 // several CatchProj inputs, from the same Catch.
1666 }
1667
1668 // Set the offset of the return from the call
1669 assert(handler_bcis.find(-1) != -1, "must have default handler");
1670 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos);
1671 continue;
1672 }
1673
1674 // Handle implicit null exception table updates
1675 if (n->is_MachNullCheck()) {
1676 uint block_num = block->non_connector_successor(0)->_pre_order;
1677 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1678 continue;
1679 }
1680 // Handle implicit exception table updates: trap instructions.
1681 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) {
1682 uint block_num = block->non_connector_successor(0)->_pre_order;
1683 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos());
1684 continue;
1685 }
1686 } // End of for all blocks fill in exception table entries
1687 }
1688
1689 // Static Variables
1690 #ifndef PRODUCT
1691 uint Scheduling::_total_nop_size = 0;
1692 uint Scheduling::_total_method_size = 0;
1693 uint Scheduling::_total_branches = 0;
1694 uint Scheduling::_total_unconditional_delays = 0;
1695 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1];
1696 #endif
1697
1698 // Initializer for class Scheduling
1699
1700 Scheduling::Scheduling(Arena *arena, Compile &compile)
1701 : _arena(arena),
1702 _cfg(compile.cfg()),
1703 _regalloc(compile.regalloc()),
1704 _scheduled(arena),
1705 _available(arena),
1706 _reg_node(arena),
1707 _pinch_free_list(arena),
1708 _next_node(NULL),
1709 _bundle_instr_count(0),
1710 _bundle_cycle_number(0),
1711 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0])
1712 #ifndef PRODUCT
1713 , _branches(0)
1714 , _unconditional_delays(0)
1715 #endif
1716 {
1717 // Create a MachNopNode
1718 _nop = new MachNopNode();
1719
1720 // Now that the nops are in the array, save the count
1721 // (but allow entries for the nops)
1722 _node_bundling_limit = compile.unique();
1723 uint node_max = _regalloc->node_regs_max_index();
1724
1725 compile.set_node_bundling_limit(_node_bundling_limit);
1726
1727 // This one is persistent within the Compile class
1728 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max);
1729
1730 // Allocate space for fixed-size arrays
1731 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1732 _uses = NEW_ARENA_ARRAY(arena, short, node_max);
1733 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max);
1734
1735 // Clear the arrays
1736 memset(_node_bundling_base, 0, node_max * sizeof(Bundle));
1737 memset(_node_latency, 0, node_max * sizeof(unsigned short));
1738 memset(_uses, 0, node_max * sizeof(short));
1739 memset(_current_latency, 0, node_max * sizeof(unsigned short));
1740
1741 // Clear the bundling information
1742 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements));
1743
1744 // Get the last node
1745 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1);
1746
1747 _next_node = block->get_node(block->number_of_nodes() - 1);
1748 }
1749
1750 #ifndef PRODUCT
1751 // Scheduling destructor
1752 Scheduling::~Scheduling() {
1753 _total_branches += _branches;
1754 _total_unconditional_delays += _unconditional_delays;
1755 }
1756 #endif
1757
1758 // Step ahead "i" cycles
1759 void Scheduling::step(uint i) {
1760
1761 Bundle *bundle = node_bundling(_next_node);
1762 bundle->set_starts_bundle();
1763
1764 // Update the bundle record, but leave the flags information alone
1765 if (_bundle_instr_count > 0) {
1766 bundle->set_instr_count(_bundle_instr_count);
1767 bundle->set_resources_used(_bundle_use.resourcesUsed());
1768 }
1769
1770 // Update the state information
1771 _bundle_instr_count = 0;
1772 _bundle_cycle_number += i;
1773 _bundle_use.step(i);
1774 }
1775
1776 void Scheduling::step_and_clear() {
1777 Bundle *bundle = node_bundling(_next_node);
1778 bundle->set_starts_bundle();
1779
1780 // Update the bundle record
1781 if (_bundle_instr_count > 0) {
1782 bundle->set_instr_count(_bundle_instr_count);
1783 bundle->set_resources_used(_bundle_use.resourcesUsed());
1784
1785 _bundle_cycle_number += 1;
1786 }
1787
1788 // Clear the bundling information
1789 _bundle_instr_count = 0;
1790 _bundle_use.reset();
1791
1792 memcpy(_bundle_use_elements,
1793 Pipeline_Use::elaborated_elements,
1794 sizeof(Pipeline_Use::elaborated_elements));
1795 }
1796
1797 // Perform instruction scheduling and bundling over the sequence of
1798 // instructions in backwards order.
1799 void Compile::ScheduleAndBundle() {
1800
1801 // Don't optimize this if it isn't a method
1802 if (!_method)
1803 return;
1804
1805 // Don't optimize this if scheduling is disabled
1806 if (!do_scheduling())
1807 return;
1808
1809 // Scheduling code works only with pairs (16 bytes) maximum.
1810 if (max_vector_size() > 16)
1811 return;
1812
1813 TracePhase tp("isched", &timers[_t_instrSched]);
1814
1815 // Create a data structure for all the scheduling information
1816 Scheduling scheduling(Thread::current()->resource_area(), *this);
1817
1818 // Walk backwards over each basic block, computing the needed alignment
1819 // Walk over all the basic blocks
1820 scheduling.DoScheduling();
1821 }
1822
1823 // Compute the latency of all the instructions. This is fairly simple,
1824 // because we already have a legal ordering. Walk over the instructions
1825 // from first to last, and compute the latency of the instruction based
1826 // on the latency of the preceding instruction(s).
1827 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) {
1828 #ifndef PRODUCT
1829 if (_cfg->C->trace_opto_output())
1830 tty->print("# -> ComputeLocalLatenciesForward\n");
1831 #endif
1832
1833 // Walk over all the schedulable instructions
1834 for( uint j=_bb_start; j < _bb_end; j++ ) {
1835
1836 // This is a kludge, forcing all latency calculations to start at 1.
1837 // Used to allow latency 0 to force an instruction to the beginning
1838 // of the bb
1839 uint latency = 1;
1840 Node *use = bb->get_node(j);
1841 uint nlen = use->len();
1842
1843 // Walk over all the inputs
1844 for ( uint k=0; k < nlen; k++ ) {
1845 Node *def = use->in(k);
1846 if (!def)
1847 continue;
1848
1849 uint l = _node_latency[def->_idx] + use->latency(k);
1850 if (latency < l)
1851 latency = l;
1852 }
1853
1854 _node_latency[use->_idx] = latency;
1855
1856 #ifndef PRODUCT
1857 if (_cfg->C->trace_opto_output()) {
1858 tty->print("# latency %4d: ", latency);
1859 use->dump();
1860 }
1861 #endif
1862 }
1863
1864 #ifndef PRODUCT
1865 if (_cfg->C->trace_opto_output())
1866 tty->print("# <- ComputeLocalLatenciesForward\n");
1867 #endif
1868
1869 } // end ComputeLocalLatenciesForward
1870
1871 // See if this node fits into the present instruction bundle
1872 bool Scheduling::NodeFitsInBundle(Node *n) {
1873 uint n_idx = n->_idx;
1874
1875 // If this is the unconditional delay instruction, then it fits
1876 if (n == _unconditional_delay_slot) {
1877 #ifndef PRODUCT
1878 if (_cfg->C->trace_opto_output())
1879 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx);
1880 #endif
1881 return (true);
1882 }
1883
1884 // If the node cannot be scheduled this cycle, skip it
1885 if (_current_latency[n_idx] > _bundle_cycle_number) {
1886 #ifndef PRODUCT
1887 if (_cfg->C->trace_opto_output())
1888 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n",
1889 n->_idx, _current_latency[n_idx], _bundle_cycle_number);
1890 #endif
1891 return (false);
1892 }
1893
1894 const Pipeline *node_pipeline = n->pipeline();
1895
1896 uint instruction_count = node_pipeline->instructionCount();
1897 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
1898 instruction_count = 0;
1899 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
1900 instruction_count++;
1901
1902 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) {
1903 #ifndef PRODUCT
1904 if (_cfg->C->trace_opto_output())
1905 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n",
1906 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle);
1907 #endif
1908 return (false);
1909 }
1910
1911 // Don't allow non-machine nodes to be handled this way
1912 if (!n->is_Mach() && instruction_count == 0)
1913 return (false);
1914
1915 // See if there is any overlap
1916 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse());
1917
1918 if (delay > 0) {
1919 #ifndef PRODUCT
1920 if (_cfg->C->trace_opto_output())
1921 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx);
1922 #endif
1923 return false;
1924 }
1925
1926 #ifndef PRODUCT
1927 if (_cfg->C->trace_opto_output())
1928 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx);
1929 #endif
1930
1931 return true;
1932 }
1933
1934 Node * Scheduling::ChooseNodeToBundle() {
1935 uint siz = _available.size();
1936
1937 if (siz == 0) {
1938
1939 #ifndef PRODUCT
1940 if (_cfg->C->trace_opto_output())
1941 tty->print("# ChooseNodeToBundle: NULL\n");
1942 #endif
1943 return (NULL);
1944 }
1945
1946 // Fast path, if only 1 instruction in the bundle
1947 if (siz == 1) {
1948 #ifndef PRODUCT
1949 if (_cfg->C->trace_opto_output()) {
1950 tty->print("# ChooseNodeToBundle (only 1): ");
1951 _available[0]->dump();
1952 }
1953 #endif
1954 return (_available[0]);
1955 }
1956
1957 // Don't bother, if the bundle is already full
1958 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) {
1959 for ( uint i = 0; i < siz; i++ ) {
1960 Node *n = _available[i];
1961
1962 // Skip projections, we'll handle them another way
1963 if (n->is_Proj())
1964 continue;
1965
1966 // This presupposed that instructions are inserted into the
1967 // available list in a legality order; i.e. instructions that
1968 // must be inserted first are at the head of the list
1969 if (NodeFitsInBundle(n)) {
1970 #ifndef PRODUCT
1971 if (_cfg->C->trace_opto_output()) {
1972 tty->print("# ChooseNodeToBundle: ");
1973 n->dump();
1974 }
1975 #endif
1976 return (n);
1977 }
1978 }
1979 }
1980
1981 // Nothing fits in this bundle, choose the highest priority
1982 #ifndef PRODUCT
1983 if (_cfg->C->trace_opto_output()) {
1984 tty->print("# ChooseNodeToBundle: ");
1985 _available[0]->dump();
1986 }
1987 #endif
1988
1989 return _available[0];
1990 }
1991
1992 void Scheduling::AddNodeToAvailableList(Node *n) {
1993 assert( !n->is_Proj(), "projections never directly made available" );
1994 #ifndef PRODUCT
1995 if (_cfg->C->trace_opto_output()) {
1996 tty->print("# AddNodeToAvailableList: ");
1997 n->dump();
1998 }
1999 #endif
2000
2001 int latency = _current_latency[n->_idx];
2002
2003 // Insert in latency order (insertion sort)
2004 uint i;
2005 for ( i=0; i < _available.size(); i++ )
2006 if (_current_latency[_available[i]->_idx] > latency)
2007 break;
2008
2009 // Special Check for compares following branches
2010 if( n->is_Mach() && _scheduled.size() > 0 ) {
2011 int op = n->as_Mach()->ideal_Opcode();
2012 Node *last = _scheduled[0];
2013 if( last->is_MachIf() && last->in(1) == n &&
2014 ( op == Op_CmpI ||
2015 op == Op_CmpU ||
2016 op == Op_CmpUL ||
2017 op == Op_CmpP ||
2018 op == Op_CmpF ||
2019 op == Op_CmpD ||
2020 op == Op_CmpL ) ) {
2021
2022 // Recalculate position, moving to front of same latency
2023 for ( i=0 ; i < _available.size(); i++ )
2024 if (_current_latency[_available[i]->_idx] >= latency)
2025 break;
2026 }
2027 }
2028
2029 // Insert the node in the available list
2030 _available.insert(i, n);
2031
2032 #ifndef PRODUCT
2033 if (_cfg->C->trace_opto_output())
2034 dump_available();
2035 #endif
2036 }
2037
2038 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) {
2039 for ( uint i=0; i < n->len(); i++ ) {
2040 Node *def = n->in(i);
2041 if (!def) continue;
2042 if( def->is_Proj() ) // If this is a machine projection, then
2043 def = def->in(0); // propagate usage thru to the base instruction
2044
2045 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local
2046 continue;
2047 }
2048
2049 // Compute the latency
2050 uint l = _bundle_cycle_number + n->latency(i);
2051 if (_current_latency[def->_idx] < l)
2052 _current_latency[def->_idx] = l;
2053
2054 // If this does not have uses then schedule it
2055 if ((--_uses[def->_idx]) == 0)
2056 AddNodeToAvailableList(def);
2057 }
2058 }
2059
2060 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) {
2061 #ifndef PRODUCT
2062 if (_cfg->C->trace_opto_output()) {
2063 tty->print("# AddNodeToBundle: ");
2064 n->dump();
2065 }
2066 #endif
2067
2068 // Remove this from the available list
2069 uint i;
2070 for (i = 0; i < _available.size(); i++)
2071 if (_available[i] == n)
2072 break;
2073 assert(i < _available.size(), "entry in _available list not found");
2074 _available.remove(i);
2075
2076 // See if this fits in the current bundle
2077 const Pipeline *node_pipeline = n->pipeline();
2078 const Pipeline_Use& node_usage = node_pipeline->resourceUse();
2079
2080 // Check for instructions to be placed in the delay slot. We
2081 // do this before we actually schedule the current instruction,
2082 // because the delay slot follows the current instruction.
2083 if (Pipeline::_branch_has_delay_slot &&
2084 node_pipeline->hasBranchDelay() &&
2085 !_unconditional_delay_slot) {
2086
2087 uint siz = _available.size();
2088
2089 // Conditional branches can support an instruction that
2090 // is unconditionally executed and not dependent by the
2091 // branch, OR a conditionally executed instruction if
2092 // the branch is taken. In practice, this means that
2093 // the first instruction at the branch target is
2094 // copied to the delay slot, and the branch goes to
2095 // the instruction after that at the branch target
2096 if ( n->is_MachBranch() ) {
2097
2098 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" );
2099 assert( !n->is_Catch(), "should not look for delay slot for Catch" );
2100
2101 #ifndef PRODUCT
2102 _branches++;
2103 #endif
2104
2105 // At least 1 instruction is on the available list
2106 // that is not dependent on the branch
2107 for (uint i = 0; i < siz; i++) {
2108 Node *d = _available[i];
2109 const Pipeline *avail_pipeline = d->pipeline();
2110
2111 // Don't allow safepoints in the branch shadow, that will
2112 // cause a number of difficulties
2113 if ( avail_pipeline->instructionCount() == 1 &&
2114 !avail_pipeline->hasMultipleBundles() &&
2115 !avail_pipeline->hasBranchDelay() &&
2116 Pipeline::instr_has_unit_size() &&
2117 d->size(_regalloc) == Pipeline::instr_unit_size() &&
2118 NodeFitsInBundle(d) &&
2119 !node_bundling(d)->used_in_delay()) {
2120
2121 if (d->is_Mach() && !d->is_MachSafePoint()) {
2122 // A node that fits in the delay slot was found, so we need to
2123 // set the appropriate bits in the bundle pipeline information so
2124 // that it correctly indicates resource usage. Later, when we
2125 // attempt to add this instruction to the bundle, we will skip
2126 // setting the resource usage.
2127 _unconditional_delay_slot = d;
2128 node_bundling(n)->set_use_unconditional_delay();
2129 node_bundling(d)->set_used_in_unconditional_delay();
2130 _bundle_use.add_usage(avail_pipeline->resourceUse());
2131 _current_latency[d->_idx] = _bundle_cycle_number;
2132 _next_node = d;
2133 ++_bundle_instr_count;
2134 #ifndef PRODUCT
2135 _unconditional_delays++;
2136 #endif
2137 break;
2138 }
2139 }
2140 }
2141 }
2142
2143 // No delay slot, add a nop to the usage
2144 if (!_unconditional_delay_slot) {
2145 // See if adding an instruction in the delay slot will overflow
2146 // the bundle.
2147 if (!NodeFitsInBundle(_nop)) {
2148 #ifndef PRODUCT
2149 if (_cfg->C->trace_opto_output())
2150 tty->print("# *** STEP(1 instruction for delay slot) ***\n");
2151 #endif
2152 step(1);
2153 }
2154
2155 _bundle_use.add_usage(_nop->pipeline()->resourceUse());
2156 _next_node = _nop;
2157 ++_bundle_instr_count;
2158 }
2159
2160 // See if the instruction in the delay slot requires a
2161 // step of the bundles
2162 if (!NodeFitsInBundle(n)) {
2163 #ifndef PRODUCT
2164 if (_cfg->C->trace_opto_output())
2165 tty->print("# *** STEP(branch won't fit) ***\n");
2166 #endif
2167 // Update the state information
2168 _bundle_instr_count = 0;
2169 _bundle_cycle_number += 1;
2170 _bundle_use.step(1);
2171 }
2172 }
2173
2174 // Get the number of instructions
2175 uint instruction_count = node_pipeline->instructionCount();
2176 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0)
2177 instruction_count = 0;
2178
2179 // Compute the latency information
2180 uint delay = 0;
2181
2182 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) {
2183 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number;
2184 if (relative_latency < 0)
2185 relative_latency = 0;
2186
2187 delay = _bundle_use.full_latency(relative_latency, node_usage);
2188
2189 // Does not fit in this bundle, start a new one
2190 if (delay > 0) {
2191 step(delay);
2192
2193 #ifndef PRODUCT
2194 if (_cfg->C->trace_opto_output())
2195 tty->print("# *** STEP(%d) ***\n", delay);
2196 #endif
2197 }
2198 }
2199
2200 // If this was placed in the delay slot, ignore it
2201 if (n != _unconditional_delay_slot) {
2202
2203 if (delay == 0) {
2204 if (node_pipeline->hasMultipleBundles()) {
2205 #ifndef PRODUCT
2206 if (_cfg->C->trace_opto_output())
2207 tty->print("# *** STEP(multiple instructions) ***\n");
2208 #endif
2209 step(1);
2210 }
2211
2212 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) {
2213 #ifndef PRODUCT
2214 if (_cfg->C->trace_opto_output())
2215 tty->print("# *** STEP(%d >= %d instructions) ***\n",
2216 instruction_count + _bundle_instr_count,
2217 Pipeline::_max_instrs_per_cycle);
2218 #endif
2219 step(1);
2220 }
2221 }
2222
2223 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot)
2224 _bundle_instr_count++;
2225
2226 // Set the node's latency
2227 _current_latency[n->_idx] = _bundle_cycle_number;
2228
2229 // Now merge the functional unit information
2230 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode())
2231 _bundle_use.add_usage(node_usage);
2232
2233 // Increment the number of instructions in this bundle
2234 _bundle_instr_count += instruction_count;
2235
2236 // Remember this node for later
2237 if (n->is_Mach())
2238 _next_node = n;
2239 }
2240
2241 // It's possible to have a BoxLock in the graph and in the _bbs mapping but
2242 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks.
2243 // 'Schedule' them (basically ignore in the schedule) but do not insert them
2244 // into the block. All other scheduled nodes get put in the schedule here.
2245 int op = n->Opcode();
2246 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR
2247 (op != Op_Node && // Not an unused antidepedence node and
2248 // not an unallocated boxlock
2249 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) {
2250
2251 // Push any trailing projections
2252 if( bb->get_node(bb->number_of_nodes()-1) != n ) {
2253 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2254 Node *foi = n->fast_out(i);
2255 if( foi->is_Proj() )
2256 _scheduled.push(foi);
2257 }
2258 }
2259
2260 // Put the instruction in the schedule list
2261 _scheduled.push(n);
2262 }
2263
2264 #ifndef PRODUCT
2265 if (_cfg->C->trace_opto_output())
2266 dump_available();
2267 #endif
2268
2269 // Walk all the definitions, decrementing use counts, and
2270 // if a definition has a 0 use count, place it in the available list.
2271 DecrementUseCounts(n,bb);
2272 }
2273
2274 // This method sets the use count within a basic block. We will ignore all
2275 // uses outside the current basic block. As we are doing a backwards walk,
2276 // any node we reach that has a use count of 0 may be scheduled. This also
2277 // avoids the problem of cyclic references from phi nodes, as long as phi
2278 // nodes are at the front of the basic block. This method also initializes
2279 // the available list to the set of instructions that have no uses within this
2280 // basic block.
2281 void Scheduling::ComputeUseCount(const Block *bb) {
2282 #ifndef PRODUCT
2283 if (_cfg->C->trace_opto_output())
2284 tty->print("# -> ComputeUseCount\n");
2285 #endif
2286
2287 // Clear the list of available and scheduled instructions, just in case
2288 _available.clear();
2289 _scheduled.clear();
2290
2291 // No delay slot specified
2292 _unconditional_delay_slot = NULL;
2293
2294 #ifdef ASSERT
2295 for( uint i=0; i < bb->number_of_nodes(); i++ )
2296 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" );
2297 #endif
2298
2299 // Force the _uses count to never go to zero for unscheduable pieces
2300 // of the block
2301 for( uint k = 0; k < _bb_start; k++ )
2302 _uses[bb->get_node(k)->_idx] = 1;
2303 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ )
2304 _uses[bb->get_node(l)->_idx] = 1;
2305
2306 // Iterate backwards over the instructions in the block. Don't count the
2307 // branch projections at end or the block header instructions.
2308 for( uint j = _bb_end-1; j >= _bb_start; j-- ) {
2309 Node *n = bb->get_node(j);
2310 if( n->is_Proj() ) continue; // Projections handled another way
2311
2312 // Account for all uses
2313 for ( uint k = 0; k < n->len(); k++ ) {
2314 Node *inp = n->in(k);
2315 if (!inp) continue;
2316 assert(inp != n, "no cycles allowed" );
2317 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use?
2318 if (inp->is_Proj()) { // Skip through Proj's
2319 inp = inp->in(0);
2320 }
2321 ++_uses[inp->_idx]; // Count 1 block-local use
2322 }
2323 }
2324
2325 // If this instruction has a 0 use count, then it is available
2326 if (!_uses[n->_idx]) {
2327 _current_latency[n->_idx] = _bundle_cycle_number;
2328 AddNodeToAvailableList(n);
2329 }
2330
2331 #ifndef PRODUCT
2332 if (_cfg->C->trace_opto_output()) {
2333 tty->print("# uses: %3d: ", _uses[n->_idx]);
2334 n->dump();
2335 }
2336 #endif
2337 }
2338
2339 #ifndef PRODUCT
2340 if (_cfg->C->trace_opto_output())
2341 tty->print("# <- ComputeUseCount\n");
2342 #endif
2343 }
2344
2345 // This routine performs scheduling on each basic block in reverse order,
2346 // using instruction latencies and taking into account function unit
2347 // availability.
2348 void Scheduling::DoScheduling() {
2349 #ifndef PRODUCT
2350 if (_cfg->C->trace_opto_output())
2351 tty->print("# -> DoScheduling\n");
2352 #endif
2353
2354 Block *succ_bb = NULL;
2355 Block *bb;
2356
2357 // Walk over all the basic blocks in reverse order
2358 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) {
2359 bb = _cfg->get_block(i);
2360
2361 #ifndef PRODUCT
2362 if (_cfg->C->trace_opto_output()) {
2363 tty->print("# Schedule BB#%03d (initial)\n", i);
2364 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2365 bb->get_node(j)->dump();
2366 }
2367 }
2368 #endif
2369
2370 // On the head node, skip processing
2371 if (bb == _cfg->get_root_block()) {
2372 continue;
2373 }
2374
2375 // Skip empty, connector blocks
2376 if (bb->is_connector())
2377 continue;
2378
2379 // If the following block is not the sole successor of
2380 // this one, then reset the pipeline information
2381 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) {
2382 #ifndef PRODUCT
2383 if (_cfg->C->trace_opto_output()) {
2384 tty->print("*** bundle start of next BB, node %d, for %d instructions\n",
2385 _next_node->_idx, _bundle_instr_count);
2386 }
2387 #endif
2388 step_and_clear();
2389 }
2390
2391 // Leave untouched the starting instruction, any Phis, a CreateEx node
2392 // or Top. bb->get_node(_bb_start) is the first schedulable instruction.
2393 _bb_end = bb->number_of_nodes()-1;
2394 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) {
2395 Node *n = bb->get_node(_bb_start);
2396 // Things not matched, like Phinodes and ProjNodes don't get scheduled.
2397 // Also, MachIdealNodes do not get scheduled
2398 if( !n->is_Mach() ) continue; // Skip non-machine nodes
2399 MachNode *mach = n->as_Mach();
2400 int iop = mach->ideal_Opcode();
2401 if( iop == Op_CreateEx ) continue; // CreateEx is pinned
2402 if( iop == Op_Con ) continue; // Do not schedule Top
2403 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes
2404 mach->pipeline() == MachNode::pipeline_class() &&
2405 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc
2406 continue;
2407 break; // Funny loop structure to be sure...
2408 }
2409 // Compute last "interesting" instruction in block - last instruction we
2410 // might schedule. _bb_end points just after last schedulable inst. We
2411 // normally schedule conditional branches (despite them being forced last
2412 // in the block), because they have delay slots we can fill. Calls all
2413 // have their delay slots filled in the template expansions, so we don't
2414 // bother scheduling them.
2415 Node *last = bb->get_node(_bb_end);
2416 // Ignore trailing NOPs.
2417 while (_bb_end > 0 && last->is_Mach() &&
2418 last->as_Mach()->ideal_Opcode() == Op_Con) {
2419 last = bb->get_node(--_bb_end);
2420 }
2421 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, "");
2422 if( last->is_Catch() ||
2423 // Exclude unreachable path case when Halt node is in a separate block.
2424 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) {
2425 // There must be a prior call. Skip it.
2426 while( !bb->get_node(--_bb_end)->is_MachCall() ) {
2427 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" );
2428 }
2429 } else if( last->is_MachNullCheck() ) {
2430 // Backup so the last null-checked memory instruction is
2431 // outside the schedulable range. Skip over the nullcheck,
2432 // projection, and the memory nodes.
2433 Node *mem = last->in(1);
2434 do {
2435 _bb_end--;
2436 } while (mem != bb->get_node(_bb_end));
2437 } else {
2438 // Set _bb_end to point after last schedulable inst.
2439 _bb_end++;
2440 }
2441
2442 assert( _bb_start <= _bb_end, "inverted block ends" );
2443
2444 // Compute the register antidependencies for the basic block
2445 ComputeRegisterAntidependencies(bb);
2446 if (_cfg->C->failing()) return; // too many D-U pinch points
2447
2448 // Compute intra-bb latencies for the nodes
2449 ComputeLocalLatenciesForward(bb);
2450
2451 // Compute the usage within the block, and set the list of all nodes
2452 // in the block that have no uses within the block.
2453 ComputeUseCount(bb);
2454
2455 // Schedule the remaining instructions in the block
2456 while ( _available.size() > 0 ) {
2457 Node *n = ChooseNodeToBundle();
2458 guarantee(n != NULL, "no nodes available");
2459 AddNodeToBundle(n,bb);
2460 }
2461
2462 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" );
2463 #ifdef ASSERT
2464 for( uint l = _bb_start; l < _bb_end; l++ ) {
2465 Node *n = bb->get_node(l);
2466 uint m;
2467 for( m = 0; m < _bb_end-_bb_start; m++ )
2468 if( _scheduled[m] == n )
2469 break;
2470 assert( m < _bb_end-_bb_start, "instruction missing in schedule" );
2471 }
2472 #endif
2473
2474 // Now copy the instructions (in reverse order) back to the block
2475 for ( uint k = _bb_start; k < _bb_end; k++ )
2476 bb->map_node(_scheduled[_bb_end-k-1], k);
2477
2478 #ifndef PRODUCT
2479 if (_cfg->C->trace_opto_output()) {
2480 tty->print("# Schedule BB#%03d (final)\n", i);
2481 uint current = 0;
2482 for (uint j = 0; j < bb->number_of_nodes(); j++) {
2483 Node *n = bb->get_node(j);
2484 if( valid_bundle_info(n) ) {
2485 Bundle *bundle = node_bundling(n);
2486 if (bundle->instr_count() > 0 || bundle->flags() > 0) {
2487 tty->print("*** Bundle: ");
2488 bundle->dump();
2489 }
2490 n->dump();
2491 }
2492 }
2493 }
2494 #endif
2495 #ifdef ASSERT
2496 verify_good_schedule(bb,"after block local scheduling");
2497 #endif
2498 }
2499
2500 #ifndef PRODUCT
2501 if (_cfg->C->trace_opto_output())
2502 tty->print("# <- DoScheduling\n");
2503 #endif
2504
2505 // Record final node-bundling array location
2506 _regalloc->C->set_node_bundling_base(_node_bundling_base);
2507
2508 } // end DoScheduling
2509
2510 // Verify that no live-range used in the block is killed in the block by a
2511 // wrong DEF. This doesn't verify live-ranges that span blocks.
2512
2513 // Check for edge existence. Used to avoid adding redundant precedence edges.
2514 static bool edge_from_to( Node *from, Node *to ) {
2515 for( uint i=0; i<from->len(); i++ )
2516 if( from->in(i) == to )
2517 return true;
2518 return false;
2519 }
2520
2521 #ifdef ASSERT
2522 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) {
2523 // Check for bad kills
2524 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow
2525 Node *prior_use = _reg_node[def];
2526 if( prior_use && !edge_from_to(prior_use,n) ) {
2527 tty->print("%s = ",OptoReg::as_VMReg(def)->name());
2528 n->dump();
2529 tty->print_cr("...");
2530 prior_use->dump();
2531 assert(edge_from_to(prior_use,n), "%s", msg);
2532 }
2533 _reg_node.map(def,NULL); // Kill live USEs
2534 }
2535 }
2536
2537 void Scheduling::verify_good_schedule( Block *b, const char *msg ) {
2538
2539 // Zap to something reasonable for the verify code
2540 _reg_node.clear();
2541
2542 // Walk over the block backwards. Check to make sure each DEF doesn't
2543 // kill a live value (other than the one it's supposed to). Add each
2544 // USE to the live set.
2545 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) {
2546 Node *n = b->get_node(i);
2547 int n_op = n->Opcode();
2548 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) {
2549 // Fat-proj kills a slew of registers
2550 RegMask rm = n->out_RegMask();// Make local copy
2551 while( rm.is_NotEmpty() ) {
2552 OptoReg::Name kill = rm.find_first_elem();
2553 rm.Remove(kill);
2554 verify_do_def( n, kill, msg );
2555 }
2556 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes
2557 // Get DEF'd registers the normal way
2558 verify_do_def( n, _regalloc->get_reg_first(n), msg );
2559 verify_do_def( n, _regalloc->get_reg_second(n), msg );
2560 }
2561
2562 // Now make all USEs live
2563 for( uint i=1; i<n->req(); i++ ) {
2564 Node *def = n->in(i);
2565 assert(def != 0, "input edge required");
2566 OptoReg::Name reg_lo = _regalloc->get_reg_first(def);
2567 OptoReg::Name reg_hi = _regalloc->get_reg_second(def);
2568 if( OptoReg::is_valid(reg_lo) ) {
2569 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg);
2570 _reg_node.map(reg_lo,n);
2571 }
2572 if( OptoReg::is_valid(reg_hi) ) {
2573 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg);
2574 _reg_node.map(reg_hi,n);
2575 }
2576 }
2577
2578 }
2579
2580 // Zap to something reasonable for the Antidependence code
2581 _reg_node.clear();
2582 }
2583 #endif
2584
2585 // Conditionally add precedence edges. Avoid putting edges on Projs.
2586 static void add_prec_edge_from_to( Node *from, Node *to ) {
2587 if( from->is_Proj() ) { // Put precedence edge on Proj's input
2588 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" );
2589 from = from->in(0);
2590 }
2591 if( from != to && // No cycles (for things like LD L0,[L0+4] )
2592 !edge_from_to( from, to ) ) // Avoid duplicate edge
2593 from->add_prec(to);
2594 }
2595
2596 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) {
2597 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow
2598 return;
2599
2600 Node *pinch = _reg_node[def_reg]; // Get pinch point
2601 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet?
2602 is_def ) { // Check for a true def (not a kill)
2603 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point
2604 return;
2605 }
2606
2607 Node *kill = def; // Rename 'def' to more descriptive 'kill'
2608 debug_only( def = (Node*)((intptr_t)0xdeadbeef); )
2609
2610 // After some number of kills there _may_ be a later def
2611 Node *later_def = NULL;
2612
2613 // Finding a kill requires a real pinch-point.
2614 // Check for not already having a pinch-point.
2615 // Pinch points are Op_Node's.
2616 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point?
2617 later_def = pinch; // Must be def/kill as optimistic pinch-point
2618 if ( _pinch_free_list.size() > 0) {
2619 pinch = _pinch_free_list.pop();
2620 } else {
2621 pinch = new Node(1); // Pinch point to-be
2622 }
2623 if (pinch->_idx >= _regalloc->node_regs_max_index()) {
2624 _cfg->C->record_method_not_compilable("too many D-U pinch points");
2625 return;
2626 }
2627 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init)
2628 _reg_node.map(def_reg,pinch); // Record pinch-point
2629 //_regalloc->set_bad(pinch->_idx); // Already initialized this way.
2630 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill
2631 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call
2632 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch
2633 later_def = NULL; // and no later def
2634 }
2635 pinch->set_req(0,later_def); // Hook later def so we can find it
2636 } else { // Else have valid pinch point
2637 if( pinch->in(0) ) // If there is a later-def
2638 later_def = pinch->in(0); // Get it
2639 }
2640
2641 // Add output-dependence edge from later def to kill
2642 if( later_def ) // If there is some original def
2643 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill
2644
2645 // See if current kill is also a use, and so is forced to be the pinch-point.
2646 if( pinch->Opcode() == Op_Node ) {
2647 Node *uses = kill->is_Proj() ? kill->in(0) : kill;
2648 for( uint i=1; i<uses->req(); i++ ) {
2649 if( _regalloc->get_reg_first(uses->in(i)) == def_reg ||
2650 _regalloc->get_reg_second(uses->in(i)) == def_reg ) {
2651 // Yes, found a use/kill pinch-point
2652 pinch->set_req(0,NULL); //
2653 pinch->replace_by(kill); // Move anti-dep edges up
2654 pinch = kill;
2655 _reg_node.map(def_reg,pinch);
2656 return;
2657 }
2658 }
2659 }
2660
2661 // Add edge from kill to pinch-point
2662 add_prec_edge_from_to(kill,pinch);
2663 }
2664
2665 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) {
2666 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow
2667 return;
2668 Node *pinch = _reg_node[use_reg]; // Get pinch point
2669 // Check for no later def_reg/kill in block
2670 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b &&
2671 // Use has to be block-local as well
2672 _cfg->get_block_for_node(use) == b) {
2673 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?)
2674 pinch->req() == 1 ) { // pinch not yet in block?
2675 pinch->del_req(0); // yank pointer to later-def, also set flag
2676 // Insert the pinch-point in the block just after the last use
2677 b->insert_node(pinch, b->find_node(use) + 1);
2678 _bb_end++; // Increase size scheduled region in block
2679 }
2680
2681 add_prec_edge_from_to(pinch,use);
2682 }
2683 }
2684
2685 // We insert antidependences between the reads and following write of
2686 // allocated registers to prevent illegal code motion. Hopefully, the
2687 // number of added references should be fairly small, especially as we
2688 // are only adding references within the current basic block.
2689 void Scheduling::ComputeRegisterAntidependencies(Block *b) {
2690
2691 #ifdef ASSERT
2692 verify_good_schedule(b,"before block local scheduling");
2693 #endif
2694
2695 // A valid schedule, for each register independently, is an endless cycle
2696 // of: a def, then some uses (connected to the def by true dependencies),
2697 // then some kills (defs with no uses), finally the cycle repeats with a new
2698 // def. The uses are allowed to float relative to each other, as are the
2699 // kills. No use is allowed to slide past a kill (or def). This requires
2700 // antidependencies between all uses of a single def and all kills that
2701 // follow, up to the next def. More edges are redundant, because later defs
2702 // & kills are already serialized with true or antidependencies. To keep
2703 // the edge count down, we add a 'pinch point' node if there's more than
2704 // one use or more than one kill/def.
2705
2706 // We add dependencies in one bottom-up pass.
2707
2708 // For each instruction we handle it's DEFs/KILLs, then it's USEs.
2709
2710 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this
2711 // register. If not, we record the DEF/KILL in _reg_node, the
2712 // register-to-def mapping. If there is a prior DEF/KILL, we insert a
2713 // "pinch point", a new Node that's in the graph but not in the block.
2714 // We put edges from the prior and current DEF/KILLs to the pinch point.
2715 // We put the pinch point in _reg_node. If there's already a pinch point
2716 // we merely add an edge from the current DEF/KILL to the pinch point.
2717
2718 // After doing the DEF/KILLs, we handle USEs. For each used register, we
2719 // put an edge from the pinch point to the USE.
2720
2721 // To be expedient, the _reg_node array is pre-allocated for the whole
2722 // compilation. _reg_node is lazily initialized; it either contains a NULL,
2723 // or a valid def/kill/pinch-point, or a leftover node from some prior
2724 // block. Leftover node from some prior block is treated like a NULL (no
2725 // prior def, so no anti-dependence needed). Valid def is distinguished by
2726 // it being in the current block.
2727 bool fat_proj_seen = false;
2728 uint last_safept = _bb_end-1;
2729 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL;
2730 Node* last_safept_node = end_node;
2731 for( uint i = _bb_end-1; i >= _bb_start; i-- ) {
2732 Node *n = b->get_node(i);
2733 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges
2734 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) {
2735 // Fat-proj kills a slew of registers
2736 // This can add edges to 'n' and obscure whether or not it was a def,
2737 // hence the is_def flag.
2738 fat_proj_seen = true;
2739 RegMask rm = n->out_RegMask();// Make local copy
2740 while( rm.is_NotEmpty() ) {
2741 OptoReg::Name kill = rm.find_first_elem();
2742 rm.Remove(kill);
2743 anti_do_def( b, n, kill, is_def );
2744 }
2745 } else {
2746 // Get DEF'd registers the normal way
2747 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def );
2748 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def );
2749 }
2750
2751 // Kill projections on a branch should appear to occur on the
2752 // branch, not afterwards, so grab the masks from the projections
2753 // and process them.
2754 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) {
2755 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2756 Node* use = n->fast_out(i);
2757 if (use->is_Proj()) {
2758 RegMask rm = use->out_RegMask();// Make local copy
2759 while( rm.is_NotEmpty() ) {
2760 OptoReg::Name kill = rm.find_first_elem();
2761 rm.Remove(kill);
2762 anti_do_def( b, n, kill, false );
2763 }
2764 }
2765 }
2766 }
2767
2768 // Check each register used by this instruction for a following DEF/KILL
2769 // that must occur afterward and requires an anti-dependence edge.
2770 for( uint j=0; j<n->req(); j++ ) {
2771 Node *def = n->in(j);
2772 if( def ) {
2773 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" );
2774 anti_do_use( b, n, _regalloc->get_reg_first(def) );
2775 anti_do_use( b, n, _regalloc->get_reg_second(def) );
2776 }
2777 }
2778 // Do not allow defs of new derived values to float above GC
2779 // points unless the base is definitely available at the GC point.
2780
2781 Node *m = b->get_node(i);
2782
2783 // Add precedence edge from following safepoint to use of derived pointer
2784 if( last_safept_node != end_node &&
2785 m != last_safept_node) {
2786 for (uint k = 1; k < m->req(); k++) {
2787 const Type *t = m->in(k)->bottom_type();
2788 if( t->isa_oop_ptr() &&
2789 t->is_ptr()->offset() != 0 ) {
2790 last_safept_node->add_prec( m );
2791 break;
2792 }
2793 }
2794 }
2795
2796 if( n->jvms() ) { // Precedence edge from derived to safept
2797 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use()
2798 if( b->get_node(last_safept) != last_safept_node ) {
2799 last_safept = b->find_node(last_safept_node);
2800 }
2801 for( uint j=last_safept; j > i; j-- ) {
2802 Node *mach = b->get_node(j);
2803 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP )
2804 mach->add_prec( n );
2805 }
2806 last_safept = i;
2807 last_safept_node = m;
2808 }
2809 }
2810
2811 if (fat_proj_seen) {
2812 // Garbage collect pinch nodes that were not consumed.
2813 // They are usually created by a fat kill MachProj for a call.
2814 garbage_collect_pinch_nodes();
2815 }
2816 }
2817
2818 // Garbage collect pinch nodes for reuse by other blocks.
2819 //
2820 // The block scheduler's insertion of anti-dependence
2821 // edges creates many pinch nodes when the block contains
2822 // 2 or more Calls. A pinch node is used to prevent a
2823 // combinatorial explosion of edges. If a set of kills for a
2824 // register is anti-dependent on a set of uses (or defs), rather
2825 // than adding an edge in the graph between each pair of kill
2826 // and use (or def), a pinch is inserted between them:
2827 //
2828 // use1 use2 use3
2829 // \ | /
2830 // \ | /
2831 // pinch
2832 // / | \
2833 // / | \
2834 // kill1 kill2 kill3
2835 //
2836 // One pinch node is created per register killed when
2837 // the second call is encountered during a backwards pass
2838 // over the block. Most of these pinch nodes are never
2839 // wired into the graph because the register is never
2840 // used or def'ed in the block.
2841 //
2842 void Scheduling::garbage_collect_pinch_nodes() {
2843 #ifndef PRODUCT
2844 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:");
2845 #endif
2846 int trace_cnt = 0;
2847 for (uint k = 0; k < _reg_node.Size(); k++) {
2848 Node* pinch = _reg_node[k];
2849 if ((pinch != NULL) && pinch->Opcode() == Op_Node &&
2850 // no predecence input edges
2851 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) {
2852 cleanup_pinch(pinch);
2853 _pinch_free_list.push(pinch);
2854 _reg_node.map(k, NULL);
2855 #ifndef PRODUCT
2856 if (_cfg->C->trace_opto_output()) {
2857 trace_cnt++;
2858 if (trace_cnt > 40) {
2859 tty->print("\n");
2860 trace_cnt = 0;
2861 }
2862 tty->print(" %d", pinch->_idx);
2863 }
2864 #endif
2865 }
2866 }
2867 #ifndef PRODUCT
2868 if (_cfg->C->trace_opto_output()) tty->print("\n");
2869 #endif
2870 }
2871
2872 // Clean up a pinch node for reuse.
2873 void Scheduling::cleanup_pinch( Node *pinch ) {
2874 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking");
2875
2876 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) {
2877 Node* use = pinch->last_out(i);
2878 uint uses_found = 0;
2879 for (uint j = use->req(); j < use->len(); j++) {
2880 if (use->in(j) == pinch) {
2881 use->rm_prec(j);
2882 uses_found++;
2883 }
2884 }
2885 assert(uses_found > 0, "must be a precedence edge");
2886 i -= uses_found; // we deleted 1 or more copies of this edge
2887 }
2888 // May have a later_def entry
2889 pinch->set_req(0, NULL);
2890 }
2891
2892 #ifndef PRODUCT
2893
2894 void Scheduling::dump_available() const {
2895 tty->print("#Availist ");
2896 for (uint i = 0; i < _available.size(); i++)
2897 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]);
2898 tty->cr();
2899 }
2900
2901 // Print Scheduling Statistics
2902 void Scheduling::print_statistics() {
2903 // Print the size added by nops for bundling
2904 tty->print("Nops added %d bytes to total of %d bytes",
2905 _total_nop_size, _total_method_size);
2906 if (_total_method_size > 0)
2907 tty->print(", for %.2f%%",
2908 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0);
2909 tty->print("\n");
2910
2911 // Print the number of branch shadows filled
2912 if (Pipeline::_branch_has_delay_slot) {
2913 tty->print("Of %d branches, %d had unconditional delay slots filled",
2914 _total_branches, _total_unconditional_delays);
2915 if (_total_branches > 0)
2916 tty->print(", for %.2f%%",
2917 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0);
2918 tty->print("\n");
2919 }
2920
2921 uint total_instructions = 0, total_bundles = 0;
2922
2923 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) {
2924 uint bundle_count = _total_instructions_per_bundle[i];
2925 total_instructions += bundle_count * i;
2926 total_bundles += bundle_count;
2927 }
2928
2929 if (total_bundles > 0)
2930 tty->print("Average ILP (excluding nops) is %.2f\n",
2931 ((double)total_instructions) / ((double)total_bundles));
2932 }
2933 #endif