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